SkillReason: AI Skill Retrieval Framework for Implicit User Requests

To tackle the issue of retrieving suitable skills for large language model agents when user requests are vague and brief, researchers have developed SkillReason-Bench. This extensive cross-domain benchmark includes 3,729 queries and a retrieval corpus of 61,228 skills across nine domains, addressing the limitations of existing benchmarks that inadequately cover such implicit requests. Additionally, the team introduces SkillReason, a two-stage framework leveraging chain-of-thought reasoning for skill retrieval supervision. In the first stage, capability reasoning traces from a more advanced teacher offer explicit guidance through contrastive learning, retrieval distribution alignment, and language modeling. The goal is to improve agents' abilities to deduce necessary capabilities and execution steps based solely on task objectives. The paper can be found on arXiv with the identifier 2608.08640.