The Flip Side of RLHF: On-Policy Feedback for Reward Model Self-Supervised Improvement
This work provides a method for improving reward models without relying on expensive human or judge model annotations, which is significant for researchers and practitioners in language model alignment.
This paper addresses the challenge of building robust reward models (RMs) for language model alignment, which is hindered by the cost of preference data and the evolution of policies. The proposed SAVE framework uses on-policy responses, graded by a value function, to self-supervise RM improvement, achieving superior results across six benchmarks and three RL algorithms.
Building strong reward models (RMs) for language model alignment is bottlenecked by the cost and difficulty of acquiring diverse and reliable preference data from human annotation or judge models. It is dramatically worse as the policy evolves beyond the static RM training. Therefore, we propose SAVE (Self-supervised reward model improvement via Value-Anchored On-policy feedback), a framework that grades on-policy responses as feedback by using the value function for on-policy RM training. SAVE naturally converts the reward-graded on-policy responses into supervision with a prompt-specific value head as an adaptive anchor. It computes RM advantages and filters ambiguous samples to update the RM via a contrastive objective. The effectiveness of SAVE for enhancing RM training is strongly validated through rigorous empirical evaluation across six diverse benchmarks. It achieves outperforming results across all datasets while maintaining consistent improvements across three RL algorithms (GRPO, RLOO, GSPO) and different policy backbones.