Skill Routing Fails Because It Matches the Task and Ignores the User: SkillFeed Gains 35.1 Points Where the Profile Decides
arXiv 2608.28241 (2026-08-28, cs.AI) formulates personalized skill routing as profile-conditioned retrieval, showing that a task-only router will pick a semantically plausible but unsuitable skill when two users with incompatible constraints issue an identical request. It introduces a profile-counterfactual benchmark that holds the task fixed while changing the user profile to change the reference skill, and SkillFeed, a progressive retrieve-and-rerank framework that establishes task-skill alignment first and then learns profile-conditioned discrimination. SkillFeed reaches 75.1% top-1 retrieval accuracy on SkillFeed-Bench, a 23.1-point improvement over the pretrained routing baseline, with a 35.1-point gain specifically on queries where the profile changes the correct skill.
↳ Follow the thread