Alibaba’s SkillRouter paper
Alibaba published SkillRouter, a 1.2B‑parameter retrieve‑rerank pipeline that indexes ~80K skills and shows full skill text is critical for selection — a reminder that scale and text richness matter in agent skill routing. The architecture highlights retrieval first, then rerank, for accurate skill selection in large agent ecosystems. (x.com)
SKILLROUTER reports a 74.0% top-1 routing accuracy on its benchmark. (arxiv.org) Ablation tests in the paper show removing the skill body produces a 29–44 percentage-point degradation across retrieval methods. (arxiv.org) Cross-encoder attention analysis in the study finds 91.7% of attention mass concentrated on the skill body field during reranking. (arxiv.org) The evaluation dataset used 75 expert-verified queries, and the authors explicitly note that the small query set limits statistical power and that generalization to other skill ecosystems is unverified. (arxiv.org) The paper states the proposed system compares favorably to larger zero-shot baselines while remaining deployable on consumer hardware. (arxiv.org) The jax-agent/skill-router README describes a build-time scan that produces a compact frontmatter index of roughly 50 tokens per skill and lazy-loads full skill content on demand at session time. (github.com) The authors frame their contribution as a challenge to common “progressive disclosure” designs that expose only names and descriptions, arguing that implementation-level content materially alters routing decisions. (arxiv.org)
