Overview
SkillNEO improves reusable agent skills through iterative execution and revision. It separates the skill chosen for use from the versions kept for further development, allowing useful partial repairs to survive even when they initially lower task reward. Comparisons between related execution traces guide subsequent edits.
The study evaluates skill evolution on 86 SkillsBench tasks across eight domains and examines reuse by stronger models. Development and final evaluation use the same task instances.