Skill++ paper: routing-aware skill refinement lifts agent routing from 36.4% to 76.3%
dair_ai · x · 2026-11-03
dair-ai highlights Skill++: Towards Lifelong Agents via Routing-Aware Skill Refinement (arXiv; UIUC, MBZUAI, Recursal and others).
Problem: LLM agent skill libraries grow continually as users add new skills. Routers pick skills only by name and description, so a new skill with similar metadata can displace an older one and break previously correct routing — a class-incremental continual learning problem for skill routing.
Method: Skill++ rewrites only the name and description of each incoming skill: the new description states what the skill supports, how it differs from its closest existing skill, and what it excludes. No router retraining and no labeled data required.
Results: As a library grows to 731 skills over ten sessions, the router selects every required skill for 76.3% of queries versus 36.4% with original metadata; substitution errors nearly vanish and crowding errors halve. It outperforms five compared approaches across three benchmarks and also improves routing in the Codex CLI agent harness.
More from coding & agent
- 开发者开源 Cosmion：让 Claude Code 等 Agent 一键生成图片 — TheAverageJedi · 2026-11-03
- ComfyUI 新节点按语义分段 conditioning，缓解 H3 长视频循环 — Commercial-Ad-3345 · 2026-11-03
- AI Slop Addons Flood World of Warcraft, Forcing Warcraft Logs to Rework Its API — 404 Media · 2026-11-03
- Hot take: the old-school monolithic web form was perfect for one-shotting agents — aarthir · 2026-11-03
- 为常关闭的 Mac 应用做 MCP server：一位开发者的踩坑长文 — roaringfantong · 2026-11-03
- 20-year language dream handed to Opus: 200 pages of notes turned into an executable plan — pbaylies · 2026-11-03