Introducing Skill-RM: A Unified Framework for Reward Modeling in LLMs

Researchers have developed the Skill Reward Model (Skill-RM), a novel framework designed to unify heterogeneous criteria used in reward modeling for large language models (LLMs). Traditional reward evaluation methods use diverse criteria such as rule-based verifiers and complex rubrics, making integration challenging. Skill-RM reformulates reward evaluation as a reusable Reward-Evaluation Skill, allowing for dynamic selection and aggregation of evidence based on input requirements. This approach enhances consistency and transparency in reward models. Experimental results indicate that Skill-RM outperforms conventional judge baselines across various benchmarks and applications, including reinforcement learning. The project code is available at GitHub.