Skip to content

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

🕒 Published (v1): 2026-07-06 16:40 UTC · Source: Arxiv · link

Why this paper was selected

Two-timescale meta-skill evolution for recursive self-improvement; Volker Tresp; novel agent architecture

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MetaSkill-Evolve introduces a two-timescale recursive self-improvement framework where LLM agents evolve not just their task-solving skill files but also the five-agent improvement procedure (meta-skill) that rewrites those skill files. A single frozen Gemma-4 31B backbone drives all agents, so gains come purely from evolved skills. On OfficeQA and SealQA the system outperforms single-level evolution by +6.38 and +8.05 accuracy points respectively.

Problem

Existing self-improving agent systems (e.g., EvoSkill, GEPA, SkillWeaver) evolve the task skill \(s\) — what the agent does — while the improvement procedure (how failures are diagnosed, edits scoped, search allocated, experience retrieved) is authored once and held fixed throughout. This creates a rigid meta-level that cannot revise its own failure-diagnosis logic; once repeated failures share a diagnosis style the fixed meta-process cannot detect, evolution stalls. The gap is the absence of recursive self-improvement: no system previously treated the improvement operator itself as an evolvable object.

Method

Every branch carries state \(b = (s, m, h)\): a task skill \(s\), a branch-local meta-skill \(m = (\psi, \sigma, \alpha, \pi, \varepsilon)\), and iteration history \(h\). The five components parameterize a specialist pipeline agent each: - \(\psi\) (Analyzer): diagnoses failures and tags them \(\phi\) - \(\sigma\) (Retriever): selects same- and cross-branch inspiration nodes matching \(\phi\) - \(\alpha\) (Allocator): sets per-iteration child budget \(K \in [1, K_{\max}]\) - \(\pi\) (Proposer): proposes edit \(\delta\) conditioned on \((f, a, I)\) - \(\varepsilon\) (Evolver): writes \(\delta\) to disk and verifies with a before/after hash check

Fast timescale (every iteration): the worst-scoring training example drives the five-agent pipeline to produce \(K\) child task skills; each child is scored on \(D_{\text{val}}\) and committed to a SQLite DAG only if \(\Delta U_v > 0\).

Slow timescale (every \(H\) fast iterations): the empirical meta-productivity \(\hat{P}(m|s) = \frac{1}{|H|}\sum_{u \in H} \Delta U_u\) over the last \(H\) descendants is folded into a synthetic meta-failure trace \(f_m\), which is passed to the same five-agent pipeline — now targeting the meta-skill files. All five files are co-edited per round (Proposers sequential, Evolvers parallel) with accumulating children (child \(k+1\) reads files as written by child \(k\)).

Frontier selection scores candidate parents by: $\(v^* = \arg\max_{v \in F} \left( \eta_1 U_v + \eta_2 \hat{P}_v + \eta_3 N_v \right)\)$ where \(N_v = 1/(1 + \text{times\_selected}_v)\) prevents budget monopolies. The full search history is a DAG persisted in SQLite, enabling cross-branch retrieval and provenance tracking.

Key Contributions

  • Two-timescale co-evolution: task skills evolve at every iteration; meta-skill evolves every \(H\) iterations, decoupling single-example noise from meta-level updates.
  • Five-component typed meta-skill \(m = (\psi, \sigma, \alpha, \pi, \varepsilon)\): each component is a Markdown SKILL.md file consumable by the same pipeline, requiring no additional model or objective.
  • Recursive self-improvement via architectural reuse: the same five-agent pipeline that refines \(s\) is applied to \(m\), a bounded one-level recursion that adds no model capacity.
  • Meta-aware frontier selection: the composite score \(\eta_1 U_v + \eta_2 \hat{P}_v + \eta_3 N_v\) distinguishes productive branches from high-utility-but-stagnated ones.
  • Cross-branch meta-retrieval: the Retriever \(\sigma\) can surface meta-failure patterns from other lineages, propagating escape strategies without shared global state.

Results

  • OfficeQA: MetaSkill-Evolve 55.32% vs. No-Skill 31.78% (+23.54 pts), Static Skill 36.09% (+19.23 pts), Single-Level 48.94% (+6.38 pts).
  • SealQA: 45.26% vs. No-Skill 29.17% (+16.09 pts), Static Skill 29.41% (+15.85 pts), Single-Level 37.21% (+8.05 pts).
  • ALFWorld: 94.23% vs. No-Skill 92.31% (+1.92 pts); backbone near ceiling; static skill slightly regresses (−1.93 pts); single-level evolution merely recovers the No-Skill baseline.
  • On both QA benchmarks the ordering No-Skill → Static → Single-Level → MetaSkill-Evolve is strictly monotonic, confirming each design choice (skill artifact, task-level evolution, meta-level evolution) independently pays off.
  • Meta-update horizon ablation: default \(H=2\) is best on all three benchmarks; increasing \(H\) to 4 or 8 degrades OfficeQA accuracy by up to 9.1 points.
  • All five meta-skill components contribute: no single-component ablation matches the full system.

Limitations

  • One-level recursion only: the meta-skill improves the improvement procedure but a fixed higher-order procedure governs meta-skill updates; true infinite regress is not achieved.
  • Computational cost scales with the five-agent pipeline applied twice per meta-update cycle; total iterations = \(3H\) in the horizon sweep, making wall-clock costs non-trivial.
  • ALFWorld near-ceiling results (backbone already at 92.31%) limit demonstration of meta-skill impact in embodied settings; the benchmark may be saturated.
  • Evaluated on a single backbone (Gemma-4 31B) and three benchmarks; generalization to other backbone families or more diverse long-horizon tasks is unverified.
  • Meta-productivity estimate \(\hat{P}_v\) is empirical and noisy for nodes with few children; the paper does not analyze variance or sample efficiency of this signal.
  • Text of the paper is truncated; some ablation details (Table 3, App. E) and the full horizon analysis paragraph are cut off.

Relevance to Agentic AI / LLM Agents

MetaSkill-Evolve is a direct advance on the growing line of skill-based self-improving agents (SkillWeaver, EvoSkill, GEPA), but is the first to make the improvement procedure itself a learnable, branch-local artifact rather than a global constant. For researchers tracking agentic self-improvement, the key insight is that separating fast and slow timescales — with meta-productivity \(\hat{P}(m|s)\) as the slow-loop signal — enables a practically bounded form of recursive self-improvement without any additional model training or objective. The SQLite-backed DAG with cross-branch meta-retrieval is also a concrete architectural pattern for multi-lineage evolutionary search that could generalize to other agentic optimization contexts. The result that meta-skill adaptation accounts for the entire end-to-end gain on ALFWorld (where task-skill evolution saturates) highlights meta-level plasticity as a critical ingredient for agents operating near performance ceilings.