Zefeng Wang, Minxi Yan, Jinhe Bi, Sikuan Yan, Volker Tresp, Yunpu Ma
A two-timescale framework that recursively evolves both an LLM agent's task skill and the meta-skill governing its own improvement procedure.
LLM agents use external, reusable skills for long-horizon tasks, but fixed, hand-authored skills are suboptimal and cannot adapt. Existing self-improving agents only evolve the task skill, while the improvement procedure itself remains fixed and non-recursive.
MetaSkill-Evolve introduces a framework where each branch carries a task skill (s) and a branch-local meta-skill (m). The meta-skill parameterizes the five agents (Analyzer, Retriever, Allocator, Proposer, Evolver) of the improvement pipeline. Task skills evolve on a fast loop, while the meta-skill evolves on a slower loop under the same pipeline applied to itself, sharing a single frozen backbone model without additional objectives.
Outperforms no-skill, static-skill, and single-level evolution baselines on three agentic benchmarks (OfficeQA, SealQA, ALFWorld), improving held-out test accuracy by +23.54, +16.09, and +1.92 points respectively over the raw backbone. The work provides a novel methodological contribution for recursively extending an agent's self-improvement capability.