Optimal Transport for LLM Reward Modeling from Noisy Preference

TL;DR

This paper introduces SelectiveRM, a novel optimal transport-based framework for reward modeling in RLHF that effectively handles noisy preferences by aligning distributions and selectively excluding outliers.

Contribution

The paper presents a new optimal transport framework with a joint consistency discrepancy and mass relaxation, improving reward modeling from noisy preference data.

Findings

SelectiveRM outperforms state-of-the-art baselines on multiple benchmarks.

Theoretical analysis shows tighter bounds on unobserved clean risk.

The approach effectively excludes noisy outliers while aligning model predictions with preferences.

Abstract

Reward models are fundamental to Reinforcement Learning from Human Feedback (RLHF), yet real-world datasets are inevitably corrupted by noisy preference. Conventional training objectives tend to overfit these errors, while existing denoising approaches often rely on homogeneous noise assumptions that fail to capture the complexity of linguistic preferences. To handle these challenges, we propose SelectiveRM, a framework grounded in optimal transport. We first devise a Joint Consistency Discrepancy to align the distribution of model predictions with preference data. Furthermore, to address the limitation of strict mass conservation which compels the model to fit outliers, we incorporate a Mass Relaxation mechanism via partial transport. This enables the autonomous exclusion of samples with noisy preference that contradict semantic consistency. Theoretically, we demonstrate that SelectiveRM optimizes a tighter upper bound on the unobserved clean risk. Extensive experiments validate that our approach significantly outperforms state-of-the-art baselines across diverse benchmarks.