Skill adaptation improves weak models in harness self-evolution
Original title: Evolving to Match Capabilities: Improving Harness Self-Evolution via Skill Adaptation
Across 103 task-skill pairs, skills that raise GPT-5.6-luna by 31.1 points lower Gemma-4-E4B by 8.8.
The authors argue harness self-evolution helps LLM agents only when skills match the executor's ability. Skills that help stronger models can hurt weaker ones, often because stated requirements are not followed. A fixed meta-skill, applied by external rewriter Kimi-K3 without execution evidence, raises three weak executors by 5.1 to 14.1 points. In a five-round loop, the meta-skill adds 2.0 to 5.1 points across seven executors from Gemma-4-E4B to Kimi-K2.5.
- The study covers seven models and three benchmarks.
- In 86% of failed weak-model runs, the skill already stated the requirement.
- Diagnoses cover tool use, context, instruction following, and verification.
- Each model revises its skill from trajectories and verifier feedback.
From the original article
1 min read at huggingface.co
Harness self-evolution offers a practical route for LLM agents to improve from execution experience without updating model parameters. Its effectiveness depends on whether models can successfully act on the guidance provided by their harnesses.
By Zhihao Yang, Haojin Wang, Juncheng Wu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.