SearchSkill: Teaching LLMs to Use Search Tools with Evolving Skill Banks
SearchSkill is an innovative framework that enhances the utilization of search tools by language models, focusing on explicit query planning through reusable search skills. It features a dynamic SkillBank that adapts by learning from recurring failure patterns and adjusts impacted trajectories prior to supervised training. The two-phase SFT approach aligns the training process with the inference-time protocol, which involves selecting skills and executing them based on those skills. SearchSkill demonstrates improved exact match performance on knowledge tasks across both open-source and closed-source models.
Key facts
- SearchSkill is a framework for teaching LLMs to use search tools.
- It makes query planning explicit through reusable search skills.
- The model selects a skill, then generates a search or answer action conditioned on the selected skill card.
- The SkillBank evolves by expanding or refining from recurrent failure patterns.
- Affected trajectories are reconstructed before supervised training.
- The two-stage SFT recipe aligns training with inference-time protocol.
- SearchSkill improves exact match on knowledge tasks across open-source and closed-source models.
- The paper is published on arXiv with ID 2605.09038.
Entities
Institutions
- arXiv