ODIN: Disentangled Reward Mitigates Hacking in RLHF

TL;DR

This paper addresses reward hacking in RLHF by proposing a disentangled reward model that separates length from content, leading to more reliable evaluations and improved policy quality.

Contribution

It introduces a joint training method with two heads to decorrelate reward from length, effectively mitigating reward hacking in RLHF.

Findings

Almost eliminates reward correlation with length

Significantly improves policy performance

Provides a new evaluation protocol for length bias

Abstract

In this work, we study the issue of reward hacking on the response length, a challenge emerging in Reinforcement Learning from Human Feedback (RLHF) on LLMs. A well-formatted, verbose but less helpful response from the LLMs can often deceive LLMs or even human evaluators to achieve high scores. The same issue also holds for some reward models in RL. To address the challenges in both training and evaluation, we establish a more reliable evaluation protocol for comparing different training configurations, which inspects the trade-off between LLM evaluation score and response length obtained by varying training hyperparameters. Based on this evaluation, we conduct large-scale studies, where the results shed insights into the efficacy of hyperparameters and tricks used in RL on mitigating length bias. We further propose to improve the reward model by jointly training two linear heads on shared feature representations to predict the rewards, one trained to correlate with length, and the other trained to decorrelate with length and therefore focus more on the actual content. We then discard the length head in RL to prevent reward hacking on length. Experiments demonstrate that our approach almost eliminates the reward correlation with length, and improves the obtained policy by a significant margin.