COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization
COBRA-Skills claims stronger agent-skill optimization at roughly half the evaluation cost of a prior method.
The paper frames skill selection and refinement as a budgeted sequential optimization problem, using contextual bandits to decide which candidate skills are worth evaluating. Its evolution step is grounded in execution feedback, so the skill pool changes as results come in. Across six agent benchmarks and three target models, the authors report the best average performance among compared methods. They also report a 55-58% cost reduction versus SkillOpt while using 50 unique optimization examples per benchmark. HF Daily Papers' note
The paper frames skill selection and refinement as a budgeted sequential optimization problem, using contextual bandits to decide which candidate skills are worth evaluating. Its evolution step is grounded in execution feedback, so the skill pool changes as results come in. Across six agent benchmarks and three target models, the authors report the best average performance among compared methods. They also report a 55-58% cost reduction versus SkillOpt while using 50 unique optimization examples per benchmark. HF Daily Papers' note
score 5