Beyond Rubrics: Exploration-Guided Evaluation Skills for Reward Modeling

The article presents Eval-Skill, an exploration-guided evaluation method for reward modeling that synthesizes domain-specific evaluation skills without the need for rigid rubrics. By utilizing only 100 cases per domain, Eval-Skill enhances judge performance across multiple benchmarks, achieving notable improvements such as +13.44% for Qwen3-8B and +18.51% for DeepSeek-V4-Flash on RewardBench 2. This approach allows for efficient context evolution in LLM-based evaluation, offering practitioners a new paradigm that reduces inference overhead while improving adaptability and performance.