Reproduce “SkillOpt: Executive Strategy for Self-Evolving Agent Skills” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yifan Yang, Ziyang Gong, Weiquan Huang, Qihao Yang, Ziwei Zhou, Zisu Huang, Yan Li, Xuemei Gao and 7 more
SkillOpt trains a natural-language skill document while keeping the target agent and harness fixed. It uses scored task traces, bounded edits, optimizer feedback, and a held-out acceptance check to select procedural changes. Across 52 model–benchmark–harness settings, it was best or tied-best; on GPT–5.5 direct chat, average held-out accuracy rose from 58.8 to 82.3, but results remain setting-specific.
Key insights
- Agent skill documents can be effectively optimized as an external, trainable state for frozen LLM agents, mimicking the rigor of weight-space optimization.
- Systematic textual edits, bounded by a 'textual learning rate' and strictly validated on held-out data, lead to stable and reliable improvements in agent performance.
- Optimized skills encode generalizable procedural knowledge, demonstrating reusability across different models, execution environments, and related tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.