=== GO difficulty clustering ===
valid with difficulty: 192

--- overall distribution ---
   35  evaluation_circularity
   32  long_horizon_or_planning
   27  data_or_label_scarcity
   23  uncategorized
   22  perception_grounding
   18  feedback_or_reward_sparse
   17  no_difficulty_extracted
    5  embodiment_or_action
    5  safety_or_alignment
    4  cost_or_latency
    3  hallucination_or_factuality
    1  generalization

--- by domain ---
  agent_tool_execution   long_horizon_or_planning=16 evaluation_circularity=13 uncategorized=12 feedback_or_reward_sparse=9
  dialogue_instruction   uncategorized=3 no_difficulty_extracted=2 evaluation_circularity=2 data_or_label_scarcity=1
  domain_science         evaluation_circularity=8 uncategorized=5 data_or_label_scarcity=5 long_horizon_or_planning=3
  embodied_agent         data_or_label_scarcity=10 perception_grounding=9 long_horizon_or_planning=5 evaluation_circularity=4
  game_simulation        long_horizon_or_planning=5 data_or_label_scarcity=4 evaluation_circularity=2 feedback_or_reward_sparse=2
  nlp_factuality         evaluation_circularity=4 data_or_label_scarcity=3 perception_grounding=2 long_horizon_or_planning=1
  other                  evaluation_circularity=2 no_difficulty_extracted=1 embodiment_or_action=1
  state_estimation       hallucination_or_factuality=1
  visual_perception      perception_grounding=5 long_horizon_or_planning=2 feedback_or_reward_sparse=1 data_or_label_scarcity=1

--- uncategorized (23) ---
  DG-000182 | LLM-based agents struggle with deep protocol-level logic bugs involving complex state-dependent beha
  DG-000363 | Configurations rely on execution failure signals and lack proactive semantic reasoning, making them 
  DG-000539 | The principal limitation concerns fusion beats, where absolute performance remains poor due to the e
  DG-000720 | Rare or unprecedented failure modes without historical analogues bypass the contrastive distillation
  DG-000908 | Standalone LLM reasoning fails in industrial root cause analysis due to the lack of runtime context 
  DG-001032 | VC generators encode full language semantics (e.g., integer overflow, pointer arithmetic) in low-lev
  DG-001069 | The provided text does not contain an explicit Limitations, Discussion, or Conclusion section with s
  DG-001279 | Two gaps separate current AI from reliable model repair: solver interaction, as API models restore o
  DG-001334 | Natural language policy specifications inevitably contain ambiguities and logical or semantic gaps t
  DG-001505 | The environment's inherent randomness (e.g., barbarian spawns) drives large deviations that multiple
  DG-001535 | Tool integration needs to be done in a way that can be prioritized and interpreted by the agents as 
  DG-001563 | User personas evolve over time as routines, environments, and interaction patterns change, introduci
  DG-001578 | Aggressive false positive reduction can suppress true vulnerabilities, and effectiveness is highly d
  DG-002037 | Limitations of an LLM's real-world understanding make it challenging to sample questions that effect
  DG-002297 | Existing automatic summarization methods fail to accurately identify relevant studies and generate h
  DG-002569 | Identifying the clinical necessity to activate the retrieval mechanism and achieving synergy between
  DG-002628 | The selected sections provided do not contain a Limitations, Discussion, or Conclusion section from 
  DG-002637 | Long-tail and distracting information questions remain challenging, and retrieval interference—where
  DG-002802 | The framework currently lacks autonomous employment of unspecified counters and creativity, and futu
  DG-003136 | Capturing the full multidimensional complexity of human identity in LLM agents without oversimplific
  DG-003184 | Different test-takers may be recommended different questions, raising concerns about fairness in the
  DG-003219 | The 15 developer profiles may not represent the full diversity of software developers globally, bein
  DG-003857 | LLM-based agents may need several iterations to find the correct conditions when faced with challeng

--- category keyword tables (go) ---
  no_difficulty_extracted: ['does not include a limitations', 'does not include limitations', 'does not contain a limitations', 'does not contain limitations', 'does not explicitly discuss limitations', 'does not explicitly discuss', 'does not explicitly state', 'does not mention', 'focusing in', 'consists mainly of prompt templates', 'consists only of references', 'only contains references', 'not present in the text', 'does not report', 'does not include an explicit limitations', 'provided only references', 'does not include an explicit']
  feedback_or_reward_sparse: ['reward', 'feedback', 'sparse', 'scarce', 'limited signal', 'delayed', 'credit assignment', 'temporal', 'bootstrapp']
  data_or_label_scarcity: ['data', 'label', 'annotation', 'dataset', 'human-labeled', 'curation', 'collect', 'sample efficiency', 'few-shot']
  long_horizon_or_planning: ['long-horizon', 'long-horizon', 'multi-step', 'compos', 'plan', 'sequential', 'task', 'step', 'hierarch']
  evaluation_circularity: ['evaluat', 'circular', 'verif', 'benchmark', 'metric', 'oracle', 'self-', 'metacognit', 'hallucin', 'confid']
  perception_grounding: ['visual', 'percept', 'vision', '3d', 'segment', 'localiz', 'image', 'spatial', 'multimodal', 'scene', 'object', 'detect', 'grounding']
  cost_or_latency: ['cost', 'latency', 'comput', 'efficien', 'expens', 'token', 'inference', 'resource', 'throughput']
  generalization: ['generaliz', 'novel', 'unseen', 'out-of-distribution', 'transfer', 'adaptation', 'robustness', 'distribution shift']
  hallucination_or_factuality: ['hallucin', 'factual', 'inaccura', 'fabricat', 'fidelity', 'truth', 'claim', 'verificat']
  embodiment_or_action: ['embod', 'afford', 'action space', 'physical', 'motion', 'control', 'manipulation', 'actuator', 'dynamics', 'grasp']
  safety_or_alignment: ['safety', 'alignment', 'harm', 'bias', 'ethic', 'jailbreak', 'attack', 'adversarial', 'poison']