Paper ID: 2408.09385

Offline RLHF Methods Need More Accurate Supervision Signals

Shiqi Wang, Zhengze Zhang, Rui Zhao, Fei Tan, Cam Tu Nguyen

With the rapid advances in Large Language Models (LLMs), aligning LLMs with human preferences become increasingly important. Although Reinforcement Learning with Human Feedback (RLHF) proves effective, it is complicated and highly resource-intensive. As such, offline RLHF has been introduced as an alternative solution, which directly optimizes LLMs with ranking losses on a fixed preference dataset. Current offline RLHF only captures the ``ordinal relationship'' between responses, overlooking the crucial aspect of ``how much'' one is preferred over the others. To address this issue, we propose a simple yet effective solution called \textbf{R}eward \textbf{D}ifference \textbf{O}ptimization, shorted as \textbf{RDO}. Specifically, we introduce {\it reward difference coefficients} to reweigh sample pairs in offline RLHF. We then develop a {\it difference model} involving rich interactions between a pair of responses for predicting these difference coefficients. Experiments with 7B LLMs on the HH and TL;DR datasets substantiate the effectiveness of our method in both automatic metrics and human evaluation, thereby highlighting its potential for aligning LLMs with human intent and values.

Submitted: Aug 18, 2024