Back to feed
arXiv cs.CL·

Beyond Rubrics: Exploration-Guided Evaluation Skills for Reward Modeling

Signal
78
Hype
25
In three linesEval-Skill, an exploration-guided method, synthesizes reusable evaluation skills for reward modeling without rigid rubrics. Trained on 100 cases per domain, the system progressively generates workflows and principles injected directly into the judge's context. On RewardBench 2, gains of +13.44% (Qwen3-8B) and +18.51% (DeepSeek-V4-Flash).
Read source
Your take?
Reinforcement learningEvalsReasoning

Summary generated by Claude — human-verified