\begin{table}[t]
\caption{Representative LNL methods mapped to assumptions, observable signals, pipeline roles, resource budgets, and practical boundaries.}\label{tab:pipeline}
\fontsize{6.7}{7.4}\selectfont
\setlength{\tabcolsep}{2pt}
\renewcommand{\arraystretch}{0.88}
\begin{tabular}{@{}L{0.15\textwidth}L{0.22\textwidth}L{0.18\textwidth}L{0.20\textwidth}L{0.15\textwidth}@{}}
\toprule
Method & Main assumption and observable signal & Principal pipeline role & Information or resource budget & Main boundary \\
\midrule
Forward correction \citep{patrini2017loss} & Class-conditional corruption represented by an estimable transition matrix & Noise characterization and loss correction & Known or estimated transition probabilities & Sensitive to transition estimation and instance-dependent or open-set noise \\
GCE \citep{zhang2018generalized} & Bounded influence of high-loss examples improves tolerance to mislabeled targets & Optimization control through a robust loss & No clean subset required; loss hyperparameter must be chosen & Does not explicitly identify or repair noisy labels \\
Co-teaching \citep{han2018coteaching} & Clean examples tend to have smaller early losses; peer disagreement limits self-confirmation & Reliable-signal isolation & Two networks and a keep-rate schedule, commonly linked to a noise-rate estimate & Hard clean examples may be discarded; both networks can share bias \\
DivideMix \citep{li2020dividemix} & Per-sample loss mixtures can separate mostly clean from noisy examples & Isolation, semi-supervised reuse, and target refinement & Two networks, mixture fitting, augmentation, and iterative pseudo-labels & Separation can collapse under structured hard noise or poorly calibrated confidence \\
DISC \citep{li2023disc} & Instance-specific temporal prediction consistency reveals changing reliability & Dynamic selection and target correction & History of predictions and instance-wise thresholds & More state and schedule sensitivity; evidence is protocol dependent \\
Structural labels \citep{kim2024structural} & Neighborhood structure carries supervision that is less brittle than one-hot labels & Representation preservation and target refinement & Reliable feature neighborhoods and structural-label construction & Unreliable neighborhoods can propagate early representation errors \\
VLM-LNL \citep{wei2024vlmnl} & A pre-trained vision--language model provides external semantic evidence about label compatibility & Noise detection and representation-guided correction & Access to a suitable foundation model and its label vocabulary & Can inherit pre-training bias, domain mismatch, and taxonomy mismatch \\
Open-world prototype methods \citep{pan2025openworld,zhang2025multiprototype} & Prototype distance and cluster structure distinguish clean, mislabeled, and unknown samples & Noise characterization, rejection, and representation preservation & Feature clustering or prototype estimation with an open-world decision rule & Unknown discovery is sensitive to representation quality and class overlap \\
\botrule
\end{tabular}
\renewcommand{\arraystretch}{1}
\end{table}