Reproduce “COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization” on OpenResearch

About this paper

Pingchen Lu, Xiangyi Wang, Xiang Li, Jie Mao, Zikun Qu, Junfeng Luo, Yao Shu, Bryan Kian Hsiang Low and 1 more

COBRA-Skills improves language-model agents by selectively testing reusable skills and periodically evolving candidates from execution evidence. It prioritizes candidates using predicted reward and uncertainty, then updates the population through regeneration, mutation, and crossover. On six benchmarks, Qwen3.6-35B-A3B averages 73.5%, versus 69.6% for SkillOpt and 60.4% without skills; its reported token-priced cost is lower, not necessarily total computation.

Key insights

• Skill optimization can be framed as contextual bandit search over semantic instruction representations, allowing evaluation resources to favor both high predicted utility and poorly explored candidates.
• Adaptive candidate selection and population evolution provide complementary benefits: removing either component lowers performance from the full method’s 73.5% average on the Qwen evaluation.
• The framework relies on target-agent execution outcomes rather than the teaching model’s subjective judgment, and skills can be generated by the target model itself with only a modest performance decrease.

Summary by alphaXiv.