Tag
Proposes Reliability-Guided Preference Optimization (RGPO) to handle inconsistent human feedback in LLM alignment by estimating annotator reliability and dynamically modulating training based on consensus, achieving superior performance over standard RLHF methods.
Proposes Partially Adjudicated Design-Based Supervised Learning (PA-DSL), a method that corrects noisy human labels using a small set of adjudicated cases to debias automated classifiers, achieving nominal coverage and reducing RMSE by 10-17% in experiments.
This paper proposes a bilevel optimization framework for Direct Preference Optimization under noisy preference labels, introducing a metadata-free meta-reweighting method that uses central-difference approximation and LoRA fine-tuning to improve alignment performance.
MOLAR proposes a noise-aware framework for learning multimodal molecular representations from noisy labels by separating clean-property inference from observed label noise, outperforming baselines on molecular benchmarks.
This paper studies symmetrization of loss functions for robust training under label noise, introducing SGCE and alpha-MAE loss functions that interpolate between multi-class unhinged loss and Mean Absolute Error, with theoretical guarantees and competitive empirical performance.