MetaRM: Shifted Distributions Alignment via Meta-Learning
This addresses a critical bottleneck in RLHF for language model alignment, offering a solution to improve training stability and generalization, though it is incremental as it builds on existing meta-learning techniques.
The paper tackles the problem of reward model (RM) performance degradation due to distribution shift in Reinforcement Learning from Human Feedback (RLHF), introducing MetaRM, a meta-learning method that improves the RM's ability to distinguish responses, with experiments showing significant enhancements in iterative optimization and out-of-distribution sample differentiation.
The success of Reinforcement Learning from Human Feedback (RLHF) in language model alignment is critically dependent on the capability of the reward model (RM). However, as the training process progresses, the output distribution of the policy model shifts, leading to the RM's reduced ability to distinguish between responses. This issue is further compounded when the RM, trained on a specific data distribution, struggles to generalize to examples outside of that distribution. These two issues can be united as a challenge posed by the shifted distribution of the environment. To surmount this challenge, we introduce MetaRM, a method leveraging meta-learning to align the RM with the shifted environment distribution. MetaRM is designed to train the RM by minimizing data loss, particularly for data that can improve the differentiation ability to examples of the shifted target distribution. Extensive experiments demonstrate that MetaRM significantly improves the RM's distinguishing ability in iterative RLHF optimization, and also provides the capacity to identify subtle differences in out-of-distribution samples.