Hugging Face Daily PapersYongqi Tong, Pan Wang, Hang Wang1 min readpaperadvanced
Grounded Skill Synthesis from Code at Scale for Agentic Intelligence
Summary
Code2Skill is an automated pipeline that mines popular GitHub repositories to extract verifiable, implementation‑anchored procedural “skills”. It builds a bank of ~1 M skill records (atomic ops, workflows, patterns) with provenance metadata, verifies each via blind reconstruction, and shows that augmenting LLM‑based agents with these skills yields an average 11.7% performance lift across 72 proto…
- Source code can serve as a grounding source for skill synthesis without requiring any prior environment interaction.
- Code2Skill processes 19,769 actively maintained GitHub repos, producing 1,006,822 accepted skill records with rich metadata (workflow, boundary, provenance, source evidence).
- Verification combines source‑body‑blind reconstruction and source‑aware comparison to ensure the extracted skill matches the original implementation.
- In downstream tasks, models that retrieve CodeSkillBank skills improve by 11.7% on average over matched baselines and beat them in 57 out of 72 protocol‑matched evaluations.
Agentic AI systems need reusable procedural knowledge to generalize beyond their training data. Traditional skill acquisition either requires costly environment interaction (trajectory‑based) or relies on unverified textual descriptions (document‑derived). By grounding skills in executable code, Co…
8/10
