Mitigating Reward Hacking in RLHF via Advantage Sign Robustness
This addresses reward hacking in RLHF for AI alignment, but it is incremental as it builds on existing policy optimization methods.
The paper tackled reward hacking in RLHF by proposing Sign-Certified Policy Optimization, which down-weights non-robust completions to preserve advantage signs, resulting in better win rates and reduced reward hacking on TL;DR summarization and AlpacaFarm benchmarks.
Reward models (RMs) used in reinforcement learning from human feedback (RLHF) are vulnerable to reward hacking: as the policy maximizes a learned proxy reward, true quality plateaus or degrades. We make the assumption that reward hacking is often caused by flipped advantage signs: instead of reducing the likelihood of a bad response, a flipped sign causes the update to increase it. By considering an adversarial perturbation in the RM parameter space, we can derive a certified sign-preservation radius, which is the smallest perturbation that can flip the advantage sign during policy optimization. Based on this formulation, we propose Sign-Certified Policy Optimization (SignCert-PO), down-weighting non-robust completions in the policy gradient update. Unlike prior approaches that require multiple RMs or access to the RM training data, SignCert-PO is lightweight and operates purely at the policy optimization stage using only the RM parameters and on-policy completions. On TL;DR summarization and AlpacaFarm benchmarks, SignCert-PO consistently achieves a better win rate than baselines and reduces reward hacking.