Capability Pages method improves LLM agent skill retrieval by 2.94 points

As LLM agents rely on larger skill libraries, they struggle to pick the right reusable skill when candidates share topics and vocabulary but implement different capabilities. This paper from authors Zifei Wang, Wei Wen, Qiang Ji, and Ruizhi Qiao formalizes the problem by defining a skill's capability as its 'executable region'—the set of queries it can solve. They argue that a skill document is a lossy observation of that region, so retrieval errors are baked into the text itself, not just the retriever. To fix this, they propose Capability Pages: cluster-contrastive skill representations with a positive trigger, a negative boundary, and a discriminative body. An offline compiler analyzes neighboring skills to write these fields, enabling better candidate recall and rejection of confusable alternatives.

On SRA-Bench, which includes 26,262 skills and 5,400 questions from six datasets, Capability Pages improved Recall@10 for all five tested retrievers, with a mean gain of 2.94 points. Adding the negative boundary to candidate cards boosted end-to-end task success by 3.62 points on average across four executors and six datasets. In a transfer evaluation on Chinese SSL-SkillDiscovery, the method reached 73.07% MRR@50 using the same encoder across conditions. Crucially, Capability Pages require no modification to online models—they improve routing purely by rewriting the offline skill library, making them a drop-in enhancement for existing agent systems.