Search2Skill: Rubric-Based RL for Skill Distillation Beyond Knowledge Boundaries

The recently introduced framework, Search2Skill, outlined in arXiv paper 2608.05245, seeks to improve LLM-based agents by extracting reusable skills from external resources, addressing the shortcomings of self-evolving techniques that depend only on the model's internal knowledge. This framework automatically detects gaps in abilities, explores external sources for solutions, and organizes the acquired information into reusable skills. It employs a rubric-based reinforcement learning method that enhances the processes of searching, determining when to search, and skill generation. The abstract of the paper emphasizes that current self-evolving skill approaches are limited by the model's parametric knowledge, while Search2Skill transcends these limitations by integrating domain conventions and standard practices that are often difficult for the agent to derive independently. Although experiments are referenced, specific details are not provided.