Unifying Stable Optimization and Reference Regularization in RLHF

TL;DR

This paper proposes a unified regularization method for RLHF that balances preventing reward hacking and ensuring stable policy updates, leading to improved alignment and stability across benchmarks.

Contribution

It introduces a novel unified regularization approach that explicitly balances reward hacking prevention and stable optimization in RLHF, improving performance and simplicity.

Findings

Outperforms existing RLHF methods on multiple benchmarks.

Enhances alignment quality and training stability.

Simplifies implementation compared to separate regularization strategies.

Abstract

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 ($\pi_0$) to mitigate reward hacking, and policy ratio clipping towards the current policy ($\pi_t$) to promote stable alignment. However, the implicit trade-off arising from simultaneously regularizing towards both $\pi_0$ and $\pi_t$ 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.