SkillRAE Paper: Retrieval-Augmented Execution Beats All Baselines for Agent Skills
Researchers at CUHK Shenzhen published SkillRAE, which proposes that skill retrieval for agents requires a fundamentally different mathematical space than document RAG—because skills are executable operators, not passive text. SkillRAE builds an offline three-tier community→skill→sub-unit graph, then applies dual top-down (macro keyword) + bottom-up (embedded sub-unit) retrieval, and compiles a "fully resolved blueprint" before invocation. This beats vanilla retrieval, skill-router baselines, and self-generated skill approaches across Skill-Bench and Agent-Skill-OS.
Why It Matters
The insight that LLMs cannot reliably resolve cross-skill dependencies at inference time—requiring a compilation pass—reframes agent skill design from "give the LLM a folder of MDs" to a structured offline build + structured retrieval problem. High practical relevance for production agentic systems.