Researchers have a new way to stop AI models from asking bad questions.
SearchSkill is a training framework that teaches language models to plan search queries using a library of reusable "skills" - structured strategies that guide how a model decides what to look up and when. Instead of generating a search query in one undifferentiated step, a model first selects a skill from a maintained "SkillBank," then produces a query conditioned on that skill. The SkillBank is not static: the system identifies recurring failure patterns, refines the bank accordingly, and rebuilds affected training examples before a second round of supervised fine-tuning. The result is a two-stage recipe meant to match how the model will actually behave at inference time.
The framing matters because most retrieval-augmented systems treat search as a black box - the model queries, gets results, reasons over them. When queries are vague or copied verbatim from the question, retrieval budgets get wasted and the reasoning that follows degrades. SearchSkill's benchmark results show fewer copied first queries, more tightly scoped multi-hop questions, and higher exact-match scores - all within a constrained search budget.
The broader context is a crowded race to make large language models better at using external tools. Most approaches either fine-tune on tool-use examples or rely on prompting tricks. SearchSkill's angle - making the query planning step explicit and learnable - is a lighter-weight option than end-to-end reinforcement learning approaches like those underlying some commercial retrieval agents, though whether it holds up outside controlled benchmarks remains the usual open question.