MetaSkill-Evolve: A Recursive Two-Timescale Framework for Self-Improving LLM Agents
By
[Submitted on 6 Jul 2026]
Summary
This paper introduces MetaSkill-Evolve, a two-timescale framework for recursive self-improvement of LLM agents. Unlike prior self-evolution methods that only improve task skills (what the agent does) while keeping the improvement procedure fixed, MetaSkill-Evolve makes skill improvement recursive: both task skills and meta-skills (which govern how improvement happens) evolve over time. The meta-skill has five components parameterizing the Analyzer, Retriever, Allocator, Proposer, and Evolver agents. Task skills evolve on a fast loop while meta-skills evolve on a slower loop under the same pipeline applied to itself, requiring no additional model or objective. The framework outperforms 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.
Source
Twitter / XMetaSkill-Evolve: A Recursive Two-Timescale Framework for Self-Improving LLM Agentsarxiv.orgKey quotes
· 4 pulledYet 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,ε) 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.
You might also wanna read
LifeSkill: A Reinforcement Learning Framework for Online Lifelong Learning in LLM Agents
This paper introduces LifeSkill, a two-stage reinforcement learning framework for online lifelong learning in Large Language Model (LLM) age
arxiv.org·13d ago
EvoTrainer: A Framework for Co-Evolving LLM Policies and Training Harnesses in Autonomous Agentic Reinforcement Learning
The article introduces EvoTrainer, an autonomous training framework for LLMs that goes beyond traditional recipe search by co-evolving both
arxiv.org·1mo ago
Survey of Self-Evolving AI Agents: Bridging Foundation Models and Lifelong Adaptability
The article surveys the emerging field of self-evolving AI agents, which aim to bridge the static capabilities of foundation models with the
arxiv.org·10mo ago
SkillsBench: A Benchmark for Evaluating AI Agent Skills Across Diverse Tasks
SkillsBench is a new benchmark for evaluating how well AI agent skills work across diverse tasks. The benchmark includes 86 tasks across 11
arxiv.org·4mo ago
The Evolution of AI: From Static Benchmarks to Inference-Time Search for Autonomous Agents
The article explores the shift from traditional AI benchmarking to inference-time search as the future of AI development. It discusses how c
adlrocha.substack.com·6mo ago
OpenEvolve: Combining LLMs with Evolutionary Search for Algorithm Discovery
OpenEvolve is an open-source evolutionary coding agent that integrates large language models (LLMs) into a quality-diversity search framewor
algorithmicsuperintelligence.ai·6mo ago
Comments
Sign in to join the conversation.
No comments yet. Be the first.