Reproduce “SkillAlchemy: Open-World Agent Skill Creation” 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
Hengjun Wang, Shuyue Wei, Boyi Liu, Jun Yang, Yongxin Tong
SkillAlchemy creates skills from sources by identifying behavior-changing requirements, testing what evidence supports, and admitting procedures only at justified scopes before compiling them into packages. On SkillsBench v1.1, its pooled avg@5 pass rate—task success averaged over five runs—was 55.8%, versus 35.9% without skills and 54.4% for human-curated skills; the 1.4-point lead is descriptive, with a 12-point Media deficit.
Key insights
- Framing skill creation as a 'source-grounded procedure-admission problem' allows for systematically evaluating and admitting procedural knowledge from heterogeneous open-world sources.
- Implicit operational requirements can be effectively discovered through contrastive evidence acquisition, where paired research targets highlight how variations in context alter procedural behavior.
- Admitting procedures based on supporting and conflicting evidence, while explicitly defining their reusable scope, is crucial for generating robust, generalizable, and consistent agent skills.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.