Unifying Stable Optimization and Reference Regularization in RLHF
This work addresses core issues in RLHF for AI alignment, offering a more stable and effective method, though it appears incremental as it builds on existing regularization strategies.
The paper tackled the challenges of reward hacking and stable optimization in Reinforcement Learning from Human Feedback (RLHF) by introducing a unified regularization approach that balances these objectives, resulting in improved alignment performance and stability across diverse benchmarks.
Reinforcement Learning from Human Feedback (RLHF) has advanced alignment capabilities significantly but remains hindered by two core challenges: \textbf{reward hacking} and \textbf{stable optimization}. Current solutions independently address these issues through separate regularization strategies, specifically a KL-divergence penalty against a supervised fine-tuned model ($π_0$) to mitigate reward hacking, and policy ratio clipping towards the current policy ($π_t$) to promote stable alignment. However, the implicit trade-off arising from simultaneously regularizing towards both $π_0$ and $π_t$ remains under-explored. In this paper, we introduce a unified regularization approach that explicitly balances the objectives of preventing reward hacking and maintaining stable policy updates. Our simple yet principled alignment objective yields a weighted supervised fine-tuning loss with a superior trade-off, which demonstrably improves both alignment results and implementation complexity. Extensive experiments across diverse benchmarks validate that our method consistently outperforms RLHF and online preference learning methods, achieving enhanced alignment performance and stability.