SearchSkill: Enhancing LLM Search Capabilities with Evolving Skill Banks
Researchers have introduced SearchSkill, a novel framework designed to improve how Large Language Models (LLMs) utilize search tools for open-domain question answering. The study addresses the common issue where models generate broad or copied queries that waste retrieval resources and hinder reasoning. SearchSkill makes query planning explicit by employing reusable search skills. In this process, the model first selects a specific skill from an evolving 'SkillBank' and then generates search or answer actions based on that selection. The SkillBank dynamically expands and refines itself by analyzing recurrent failure patterns and reconstructing affected trajectories before supervised fine-tuning. This two-stage training approach aligns with inference-time protocols, resulting in more atomic, hop-focused queries and fewer redundant initial searches. Evaluations across various open-source and closed-source models demonstrate that SearchSkill significantly improves exact match scores on knowledge-intensive benchmarks while operating within a limited search budget. The findings suggest that explicit, skill-conditioned query planning offers a lightweight and effective alternative to treating search as an undifferentiated action, enhancing both efficiency and accuracy in AI-driven information retrieval.
Wire timeline
SearchSkill: Enhancing LLM Search Capabilities with Evolving Skill Banks
Researchers have introduced SearchSkill, a novel framework designed to improve how Large Language Models (LLMs) utilize search tools for open-domain question answering. The study addresses the common issue where models generate broad or copied queries that waste retrieval resources and hinder reasoning. SearchSkill makes query planning explicit by employing reusable search skills. In this process, the model first selects a specific skill from an evolving 'SkillBank' and then generates search or answer actions based on that selection. The SkillBank dynamically expands and refines itself by analyzing recurrent failure patterns and reconstructing affected trajectories before supervised fine-tuning. This two-stage training approach aligns with inference-time protocols, resulting in more atomic, hop-focused queries and fewer redundant initial searches. Evaluations across various open-source and closed-source models demonstrate that SearchSkill significantly improves exact match scores on knowledge-intensive benchmarks while operating within a limited search budget. The findings suggest that explicit, skill-conditioned query planning offers a lightweight and effective alternative to treating search as an undifferentiated action, enhancing both efficiency and accuracy in AI-driven information retrieval.
cs.AI updates on arXiv.org