Grounded Skill Synthesis from Code at Scale for Agentic Intelligence
Code2Skill turns GitHub repositories into a million-record skill bank for agents, with reported gains across matched evaluations.
The paper describes an automated pipeline that extracts grounded agent skills from selected code units, then verifies them through reconstruction and comparison. Run over 19,769 popular maintained GitHub repositories, it produced 1,006,822 accepted CodeSkillBank records. Models using retrieved skills improved 11.7% on average over matched baselines across 72 evaluations, according to the authors. The paper also reports a 93.50% pass rate for skills synthesized from tested AI-generated code, slightly above the 93.00% figure for human-written code. HF Daily Papers' note
The paper describes an automated pipeline that extracts grounded agent skills from selected code units, then verifies them through reconstruction and comparison. Run over 19,769 popular maintained GitHub repositories, it produced 1,006,822 accepted CodeSkillBank records. Models using retrieved skills improved 11.7% on average over matched baselines across 72 evaluations, according to the authors. The paper also reports a 93.50% pass rate for skills synthesized from tested AI-generated code, slightly above the 93.00% figure for human-written code. HF Daily Papers' note
score 4