HiSME lets LLM agents evolve the way they evolve skills, without changing weights

机器之心 · wechat · 2026-07-23

A Tsinghua University and Huawei team introduces HiSME, a lightweight hierarchy for skill meta-evolving in LLM agents.

Core idea

Agents first turn execution traces into reusable skills.
HiSME adds a second layer: it also evolves the process that generates and maintains those skills.
The system keeps everything in text space, without updating model weights.

How it works

Extractor finds candidate skills from single trajectories.
Refactorer abstracts shared structure across multiple traces or existing skills.
Refiner / filter repair, disable, or keep skills based on usefulness and reuse.
A metarefiner summarizes feedback into meta-skills that are written back into the prompts of the skill pipeline.

Results

Evaluated on BFCL-v3 multi-turn tool use and MineDojo.
It outperforms no-skill baselines, static skill evolution, and other test-time learning methods such as AWM and Memento.
In MineDojo, better skills reduce wasted exploration and token cost.
A meta-test shows that frozen meta-skills transfer well to new static runs.

Why it matters

The paper frames agent optimization as something broader than model fine-tuning: prompts, tool routing, memory, evaluators, and skill maintenance can all become evolvable components.

Original post →

HiSME lets LLM agents evolve the way they evolve skills, without changing weights

More from coding & agent