LLM Alignment Through Successive Policy Re-weighting (SPR)

Reinforcement learning (RL) is serving as the cornerstone of aligning large language models (LLMs) to human behavior, by providing an appealing formulation and a suite of effective algorithms for learning behavior strategies through interacting with the underlying environment. Current paradigm of RL-based methods for LLM alignment, such as reinforcement learning with human feedback (RLHF) involves utilizing a reward function learned from extensive offline datasets to expediate the online training of reinforcement learning. The reward function learned is then used for policy optimization to obtain an improved policy (i.e. the LLM). Despite the success of RL approaches in aligning LLM with offline datasets, there are significant computational/limit of resources concern on applying RL-based methods for LLMs. For example, standard RLHF requires simultaneously loading four models to the computing unit. In this paper, we develop a novel policy optimization algorithm named Successive Policy Re-weighting (SPR), matching the peak memory consumption of standard supervised fine-tune (SFT). Further, SPR can leverage both offline and online datasets to expediate online training and improve the sample efficiency. Specifically, SPR leverages a supervised learning subroutine to achieve policy improvement through re-weighting the policy according to the importance/performance of executed actions. Such simple and effective method is computationally inexpensive, requiring loading only one model at each update step, matching the computational cost of standard supervised fine-tuning procedure. Experimental results show that the proposed method can significantly outperform benchmark algorithms and accelerate the online training with available offline dataset.