Evolving the improver — not just the skill — accounts for all of ALFWorld's gain and half of SealQA's

Notes on MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution (arXiv:2607.05297) — Zefeng Wang, Minxi Yan, Jinhe Bi, Sikuan Yan, Volker Tresp, Yunpu Ma · July 2026

Note published July 24, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

MetaSkill-Evolve doesn't stop at letting an agent revise its own skills — it lets the agent revise the machinery that does the revising. Each search branch pairs a task skill with a meta-skill: five components (Analyzer, Retriever, Allocator, Proposer, Evolver) that define how rewriting happens in the first place. Task skills update every iteration; the meta-skill updates only every H iterations, on a slower clock, and it does so by running that same five-agent pipeline on itself — one level of recursion, no separate model or training objective. Frontier selection then decides which branch to expand next, weighing the skill's current utility against an estimate of meta-productivity (how much gain the meta-skill has recently been producing) and a cooling term that keeps any one branch from dominating the search.

All five pipeline agents run on one frozen backbone, so whatever gains show up have to come from the skill and meta-skill text itself, not extra model capacity. On OfficeQA, SealQA, and ALFWorld, this two-timescale setup beats every baseline tested — no skill at all, a static skill, and single-level evolution that updates the skill but never the improver — lifting held-out accuracy by +23.54, +16.09, and +1.92 points. Break out just the slow loop's contribution and the picture sharpens: meta-skill updates alone account for +6.38, +8.05, and +1.92 of those points. On ALFWorld that's the entire gain — skill-only evolution left accuracy almost exactly where it started. Which component of the meta-skill matters most isn't fixed either: drop the Allocator (the piece deciding how many candidate edits to try) and OfficeQA suffers most; drop the Proposer (the piece drafting the edits) and SealQA and ALFWorld suffer most. Timing matters too — the best gap between meta-updates is every 2 iterations, and stretching that to every 8 can cost up to 9.1 accuracy points as the meta-skill goes stale against the faster task-skill loop.