SkillSight: Calibrating Generic Content Bias for Skill Retrieval

SkillSight is a training-free retrieval framework for large skill libraries used by LLM agents. The authors argue that recurring descriptive patterns across skill documents create generic background signals that inflate dense similarities and obscure discriminative evidence, particularly among structurally similar hard negatives. SkillSight applies Semantic Background Calibration using an IDF-derived background subspace and Lexical Evidence Calibration to reduce shared-token influence. On SRA-Bench and SkillBench-Supp, it reportedly improves Recall@10 by up to 20.21 percentage points over the original dense retriever and runs up to 1,248 times faster than a Dense + Reranker baseline. End-to-end results cover three agent models.