The Flip Side of RLHF: On-Policy Feedback for Reward Model Self-Supervised Improvement

Hugging Face Daily Papers Papers

Summary

The SAVE framework improves reward model training by using value functions to grade on-policy responses and update models through contrastive objectives, achieving outperforming results across six benchmarks.

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.
Original Article
View Cached Full Text

Cached at: 06/01/26, 07:18 AM

Paper page - The Flip Side of RLHF: On-Policy Feedback for Reward Model Self-Supervised Improvement

Source: https://huggingface.co/papers/2605.30888

Abstract

SAVE framework improves reward model training by using value functions to grade on-policy responses and update models through contrastive objectives.

Building strongreward 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 gradeson-policy responsesas feedback by using thevalue functionfor on-policy RM training. SAVE naturally converts the reward-gradedon-policy responsesinto supervision with a prompt-specific value head as an adaptive anchor. It computes RM advantages and filters ambiguous samples to update the RM via acontrastive 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.

View arXiv pageView PDFAdd to collection

Get this paper in your agent:

hf papers read 2605\.30888

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2605.30888 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2605.30888 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2605.30888 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Reward Models Can Be Too Sensitive (22 minute read)

TLDR AI

This paper argues that reward models in RL are often oversensitive, assigning different scores to equally good responses, and proposes a training-free discretization algorithm using Monte Carlo dropout to reduce oversensitivity, improving policy quality.

Verifier-Induced Support Reshaping in On-Policy Optimization

arXiv cs.LG

The paper introduces verifier-induced support reshaping, showing that on-policy RL with verifiable rewards can improve the current objective while making successful behaviors for later objectives too rare to sample. Experiments across math reasoning and instruction following demonstrate that endpoint improvements do not guarantee future trainability.