MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Zefeng Wang, Minxi Yan, Jinhe Bi, Sikuan Yan, Volker Tresp, Yunpu Ma
• Venue: arXiv: cs.AI
• Published: 2026-07-06
• Collected: 2026-07-07
• URL: http://arxiv.org/abs/2607.05297v1
• PDF: https://arxiv.org/pdf/2607.05297v1
• DOI: UNKNOWN

Plain-Language Brief

MetaSkill-Evolve is a two-timescale framework that allows self-improving language models to improve not only their task skills but also the improvement procedure itself. The authors claim that this recursive approach outperforms other methods on three agentic benchmarks, improving test accuracy.

What the Document Says

• Recent LLM agents are used for long-horizon, open-ended tasks and external skills.
• A fixed, hand-authored skill is rarely optimal and cannot adapt to diverse tasks.
• Self-improving agents improve by rewriting their own skill files from execution traces.
• MetaSkill-Evolve introduces a recursive framework for agentic skill improvement.
• The framework consists of five components: Analyzer, Retriever, Allocator, Proposer, and Evolver agents.
• These agents share a single frozen backbone.
• MetaSkill-Evolve outperforms no-skill, static-skill, and single-level evolution baselines on three agentic benchmarks.

Key Concepts Mentioned

• Language models
• Self-improving agents
• Recursive skill improvement
• Two-timescale framework
• Analyzer
• Retriever
• Allocator
• Proposer
• Evolver

Claims Made by the Document

• Self-improving agents can yield meaningful gains on challenging benchmarks.
• MetaSkill-Evolve outperforms no-skill, static-skill, and single-level evolution baselines.

Evidence Used

• Improved held-out test accuracy over the raw backbone is reported.

Methods / Approach

• The authors present a two-timescale framework called MetaSkill-Evolve for recursive skill improvement.
• They use five agents (Analyzer, Retriever, Allocator, Proposer, and Evolver) that share a single frozen backbone.

Limitations Stated by the Authors

The paper does not discuss potential limitations of the MetaSkill-Evolve approach.

Limitations Noted by the Agent

Agent interpretation — not from the source.

The agent may struggle with interpretability issues due to recursive skill improvement.

Open Questions

• What are the potential pitfalls and challenges of recursive skill improvement?
• How does the recursive approach compare to other approaches in terms of computational efficiency?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The paper focuses on the benefits and performance of MetaSkill-Evolve, but it does not discuss its limitations or potential drawbacks.

Raw Abstract / Excerpt

Recent LLM agents tackle increasingly long-horizon, open-ended tasks, and external skills, reusable procedural knowledge supplied to the agent, further extend this capability. However, a fixed, hand-authored skill is rarely optimal, and cannot adapt to the diversity of tasks an agent encounters. Self-improving agents address this by rewriting their own skill files from execution traces, yielding meaningful gains on challenging benchmarks. Yet such self-evolution remains non-recursive: it improves only the task skill (what the agent does) while the improvement procedure (how it improves) is authored once and held fixed. We introduce MetaSkill-Evolve, a two-timescale framework that makes agentic skill improvement recursive: every branch carries both a task skill $s$ and a branch-local meta-skill $m=(ψ,σ,α,π,\varepsilon)$ whose five components parameterise the Analyzer, Retriever, Allocator, Proposer, and Evolver agents of the improvement pipeline. Task skills evolve on a fast loop while the meta-skill evolves on a slower one under the same pipeline applied to itself, with no additional model or objective. With all five pipeline agents sharing a single frozen backbone, MetaSkill-Evolve outperforms no-skill, static-skill, and single-level evolution baselines on three agentic benchmarks (OfficeQA, SealQA, ALFWorld), improving held-out test accuracy over the raw backbone by +23.54, +16.09, and +1.92 points respectively.