Skill∞ Beats Human-Curated Skills 60.7% to 56.4% on SkillsBench
TL;DR
- Skill∞ reports a 60.7% pass rate on 83 SkillsBench tasks, versus 56.4% for human-curated skills and 54.5% for SkillEvolver.
- The meta-agent uses the same procedure to optimize task skills and to rewrite its own meta-skill, with skills packaged as reusable units across domains.
- The optimization loop runs through skill diagnosis, candidate proposal, independent review requests, and search budget allocation.
Skill∞, a system in which a meta-agent rewrites its own skill optimizer, scored 60.7% on 83 SkillsBench tasks. Human-curated skills scored 56.4%. SkillEvolver, the named baseline, scored 54.5%.
The authors open plainly: "Recursive self-improvement (RSI), in which a system also improves the way it improves, is a long-standing aspiration on the path to AGI."
Skills in the system work as standard, reusable packages across diverse domains. The optimization loop runs through skill diagnosis, candidate proposal, independent review requests, and search budget allocation. The same procedure the meta-agent uses to tune task skills is then applied to the meta-skill itself.
The 17-page, 5-figure preprint, authored by Wangcheng Tao, Zhanglu Yan, Han Wu and Weng-Fai Wong, was submitted on October 26. The abstract publishes no per-domain pass rate, no compute cost, and no indication of which reviewer model handles the independent review step. It lands the same week as Renmin's survey on trust risks of self-improving models.
Originally reported by arxiv.org
Read the original article →Original headline: Skill∞ Lets a Meta-Agent Rewrite Its Own Skill Optimizer, Beats Human-Curated Skills on SkillsBench