Fine-Tuning Language Models with Advantage-Induced Policy Alignment
This addresses a key bottleneck in reinforcement learning from human feedback for language models, offering a more stable and efficient method, though it appears incremental as it builds on existing RLHF techniques.
The paper tackles the issues of mode collapse, instability, and poor sample efficiency in proximal policy optimization (PPO) for aligning language models with human preferences by introducing Advantage-Induced Policy Alignment (APA), which consistently outperforms PPO by a large margin in language tasks.
Reinforcement learning from human feedback (RLHF) has emerged as a reliable approach to aligning large language models (LLMs) to human preferences. Among the plethora of RLHF techniques, proximal policy optimization (PPO) is of the most widely used methods. Despite its popularity, however, PPO may suffer from mode collapse, instability, and poor sample efficiency. We show that these issues can be alleviated by a novel algorithm that we refer to as Advantage-Induced Policy Alignment (APA), which leverages a squared error loss function based on the estimated advantages. We demonstrate empirically that APA consistently outperforms PPO in language tasks by a large margin, when a separate reward model is employed as the evaluator. In addition, compared with PPO, APA offers a more stable form of control over the deviation from the model's initial policy, ensuring that the model improves its performance without collapsing to deterministic output. In addition to empirical results, we also provide a theoretical justification supporting the design of our loss function.