COBRA-Skills reduces the tested skill-optimization budget
The reported 55–58% reduction is relative to SkillOpt during optimization, not a serving-cost guarantee.
COBRA-Skills frames skill refinement as budgeted sequential optimization. It combines contextual-bandit prioritization with evidence-grounded skill evolution, allocating evaluations to promising or informative candidates while updating skills from execution feedback.
Across six benchmarks and three target models, the authors report 55–58% lower optimization cost than SkillOpt while using 50 unique optimization examples per benchmark. The framework’s reward prediction and bandit choice guide which evaluations are purchased; this is different from making each future production request cheaper.
The initial arXiv submission is September 10. September 14 is this article’s date, not a new source publication date. For adoption, compare optimization budget and held-out performance at the same evaluation allowance, then separately measure inference overhead when the chosen skill is served.
Editorial correction (2026-10-13 UTC): Separated source publication, optimization budget and serving cost.