New AI Framework Distills Skills from External Knowledge.

Large Language Model (LLM)-based agents are increasingly being developed to solve complex, real-world professional tasks through reusable skills. However, existing methods for skill acquisition are often limited by the model's internal parametric knowledge or its own generated trajectories. This means that domain-specific conventions and standard procedures, which frequently lie outside the agent's pre-existing knowledge base, are difficult to incorporate, hindering true self-evolution in expert domains. To overcome this limitation, researchers have introduced Search2Skill, a new framework designed to enable agents to distill skills from external knowledge sources. Search2Skill operates by automatically identifying an agent's current capability gaps. Once identified, it intelligently searches external sources for relevant information and then distills this retrieved evidence into structured, reusable skills. This process allows agents to learn beyond their initial knowledge boundaries. The framework is optimized using a rubric-based reinforcement learning scheme. This scheme jointly improves three critical aspects: when to initiate a search, how to conduct an effective search, and how to generate high-quality skills from the retrieved information. Extensive experiments across eight expert-level domains and three benchmarks demonstrate that Search2Skill consistently outperforms both search-augmented and trajectory-based skill-learning baselines. Further analysis confirms that the gains stem from the abstraction of skills rather than just raw retrieved evidence, and that these acquired skills are transferable across different model scales.