rsi.watch

Wangcheng Tao, Zhanglu Yan, Han Wu and Weng-Fai Wong · 2026-10-26

Skill∞: Recursive Self-Improvement for Agent Skill Optimization

“which makes skill optimization recursively self-improving: a meta agent follows a meta-skill to optimize task skills and optimizes the meta-skill itself using the same procedure.”
Who
Wangcheng Tao, Zhanglu Yan, Han Wu and Weng-Fai Wong
Date
2026-10-26
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ risk of bias. The paper does not say whether the skill optimization tasks used to refine the meta-skill are kept apart from the 83 SkillsBench test tasks. “For self-improvement, the meta agent refines the seed meta-skill in Appendix G on skill optimization tasks drawn from different domains, with the budgets given in Appendix B.”
  • ↓ imprecise. No spread across the runs is reported. Some domains are small: mathematics has 5 tasks and cybersecurity 7. “All reported results are averaged over 5 independent runs.”
how we rate evidence
Replication
No rerun found. The paper says code will be released on GitHub. We found no rerun by another group. Searched up to 2026-10-28. all claims
Checked
2026-10-28, against the source page
Links
Source Paper

Also on the page

  • The meta-skill is a skill package that defines the meta agent's steps: diagnose execution traces, propose candidates, request independent reviews and allocate the search budget. The same steps are used to rewrite the meta-skill. Scoring is outside what self-improvement can change, and the backbone model stays fixed.
  • The meta-skill is refined on a set of skill optimization tasks, then reused to optimize the 83 SkillsBench task skills. With Codex CLI and GPT-6 Luna, the pass rate is 60.7% with the refined meta-skill and 55.9% with the seed meta-skill, against 56.4% for human-curated skills and 54.5% for SkillEvolver. By domain, the refined meta-skill scores lower than the seed in three of eight, for example 47.3% against 52.7% in one.
  • On KernelBench, on three level-3 problems, the mean speedup is 1.62, against 1.18 for SkillEvolver. With GPT-5.6 Luna the SkillsBench pass rate is 66.5% against 61.2% for human-curated skills; with OpenCode it is 56.9% against 53.3%.
  • The authors' limits: gains are "bounded by the expressiveness of skill packages", both levels need automatic verifiers, and the meta agent follows the meta-skill as instructions: "No tool enforces this adherence".

In the self-improving AI catalog

  • Skill∞: what it changes, whether the change lasts, and how it was tested

How it got here

Earlier work in this catalog that Skill∞ cites, oldest first. Under each: where that work sits on the map, then where Skill∞ sits (changes in bold), and what Skill∞'s own paper says about it. Its whole family on the citation map

Cites Good, the intelligence explosion 1965, Schmidhuber, the Gödel machine 2003, Yudkowsky, seed AI 2007, Voyager 2023, Promptbreeder 2023, STOP 2023, ADAS 2024, Gödel Agent 2024, Darwin Gödel Machine 2025, Hyperagents 2026

← Ledger Report an error