Skill-infinity makes agent skill optimization self-improving
Original title: Skill^{infty}: Recursive Self-Improvement for Agent Skill Optimization
On 83 SkillsBench tasks the authors report 60.7% pass rate, above curated skills and SkillEvolver.
Skill^infty treats skill optimization as recursive self-improvement: a meta agent follows a meta-skill to improve task skills and applies the same procedure to the meta-skill. The meta agent diagnoses traces, proposes candidates, requests reviews, and allocates search budget, and the meta-skill can rewrite those steps. The meta-skill is trained across optimization tasks, then reused on task skills in other domains, with both levels loaded as standard skill packages. On 83 SkillsBench tasks, the paper reports a 60.7% pass rate versus 56.4% for human-curated skills and 54.5% for SkillEvolver.
Unconfirmed: The pass rates are the paper's own SkillsBench figures, not independently verified.
- Optimization is agent-led rather than a fixed hand-designed pipeline.
- The meta-skill is decoupled from task skills and reused across domains.
- Skills at both levels use native skill discovery.
- Compared methods are human-curated skills and SkillEvolver.
From the original article
1 min read at huggingface.co
Recursive self-improvement (RSI), in which a system also improves the way it improves, is a long-standing aspiration on the path to AGI. Language agents increasingly store procedural knowledge in reusable skills. Optimizing skills offers a direct way to achieve RSI, because the optimization procedure itself can be encoded as a skill.
By Wangcheng Tao, Zhanglu Yan, Han Wu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.