SearchSkill: Teaching LLMs to Use Search Tools with Evolving Skill Banks

2026-05-09

AI Summary

This work addresses the inefficiency of large language models in open-domain question answering, where they often generate redundant or suboptimal search queries that waste retrieval resources and impair reasoning. The authors propose a skill-conditioned query planning framework that explicitly models reusable search skills, guiding the model to first select an appropriate skill before generating either a query or a final answer. They further introduce a dynamically evolving SkillBank mechanism that updates the skill repository based on failure patterns, enhanced by trajectory reconstruction and a two-stage supervised fine-tuning strategy. Evaluated on both open- and closed-source models, the approach significantly improves exact match accuracy on knowledge-intensive QA tasks, reduces initial-query duplication, and promotes atomic, targeted retrieval under constrained budgets, thereby achieving higher answer accuracy.

Abstract

Teaching language models to use search tools is not only a question of whether they search, but also of whether they issue good queries. This is especially important in open-domain question answering, where broad or copied queries often waste retrieval budget and derail later reasoning. We propose \Ours, a framework that makes query planning explicit through reusable search skills. At each step, the model first selects a skill, then generates a search or answer action conditioned on the selected skill card. The skill inventory itself is not fixed: SearchSkill maintains an evolving SkillBank, expands or refines it from recurrent failure patterns, and reconstructs affected trajectories before supervised training. The resulting two-stage SFT recipe aligns training with the inference-time protocol of skill selection followed by skill-grounded execution. Across open-source and closed-source models, SearchSkill improves exact match on knowledge-intensive QA benchmarks and yields better retrieval behavior, including fewer copied first queries, more atomic hop-focused queries, and more correct answers within a small search budget. These results suggest that explicit skill-conditioned query planning is a lightweight alternative to treating search as an undifferentiated action.

Problem

• search tools
• query planning
• open-domain question answering
• retrieval efficiency
• language models

Innovation

• search skill
• skill bank
• query planning
• retrieval-augmented reasoning
• supervised fine-tuning