Towards Pareto-Efficient RLHF: Paying Attention to a Few High-Reward Samples with Reward Dropout

摘要

Recently, leveraging reinforcement learning (RL) to fine-tune language models (LMs), known as reinforcement learning from human feedback (RLHF), has become an important research topic. However, there is still a lack of theoretical understanding of how RLHF works, the conditions under which it succeeds or fails, and whether it guarantees optimization of both likelihood \beta(\cdot) and reward R(\cdot) objectives. To address these issues, we consider RLHF as a bi-objective problem that has the nature of a Pareto optimization, present a Pareto improvement condition that is necessary to obtain Pareto-efficient policies, and propose a simple yet powerful method named reward dropout that guarantees a Pareto improvement. To demonstrate the performance of reward dropout, two benchmark datasets commonly used in text style transfer tasks were utilized in our study: sentiment and topic datasets sourced from Yelp and AG_News, respectively. Our experiments highlight that paying attention to a few samples with higher rewards leads to greater Pareto improvements regardless of model size. We also demonstrate that the effect of reward dropout is generalizable and most effective with non-pretrained target models, saving the effort of pretraining.