COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization
Pingchen Lu, Xiangyi Wang, Xiang Li, Jie Mao, Zikun Qu, Junfeng Luo, Yao Shu, Bryan Kian Hsiang Low and 1 others
The Chinese University of Hong Kong, Shenzhen · Tianjin University · The Hong Kong University of Science and Technology (Guangzhou) · National University of Singapore
Published September 10, 2026 · on the feed September 14, 2026 · ▲ 44 · arXiv 2609.11682
The abstract
Large language model (LLM) agents can benefit from reusable skills distilled from prior task experience, yet existing skill optimization methods often rely on costly execution-based evaluation and substantial task data. We introduce COBRA-Skills, an efficient framework that formulates skill optimization as budgeted sequential optimization over a dynamically evolving candidate space. COBRA-Skills couples contextual-bandit-guided prioritization with evidence-grounded skill evolution, selectively allocating evaluations to promising or informative candidates while continually refining the skill population from execution feedback. Across six heterogeneous agent benchmarks and three target models, COBRA-Skills consistently achieves the strongest average performance among compared methods, while reducing optimization cost by 55--58\% relative to SkillOpt and using only 50 unique optimization examples per benchmark. Further analyses show that COBRA-Skills remains robust to changes in the agent harness and performs effectively when the target model itself is used for skill generation and refinement.
The Brief
2:04A short podcast summary of the paper. Written and voiced by AI.
Transcript
COBRA-Skills. Contextual Bandit-Guided Evolution for Agent Skill Optimization. By Pingchen Lu and eight others, from the Chinese University of Hong Kong, Shenzhen, Tianjin University, and two other institutions. Published on the tenth of September. Its code repository is worked out: about one hundred and sixty source files, with a readme and a licence.
Agents built on large language models can benefit from reusable skills distilled from earlier task experience. Existing ways of optimizing those skills tend to be expensive, though. They often rely on costly evaluation by execution, and on substantial amounts of task data. Finding out whether a candidate skill actually helps means running the agent with it, and doing that for every candidate, on many tasks, adds up quickly.
COBRA-Skills treats skill optimization as budgeted sequential optimization over a candidate space that keeps evolving as it runs. The budget is the number of evaluations it can afford, and the question at each step is which candidate to spend the next one on. Two parts are coupled: prioritization guided by a contextual bandit, and skill evolution grounded in evidence.
The bandit decides where evaluations go, allocating them selectively to candidates that look promising or informative instead of spreading them evenly. Meanwhile the skill population is continually refined from execution feedback. Evaluations are spent where they are most likely to teach something, and the candidates being evaluated keep improving as that feedback arrives.
Across six heterogeneous agent benchmarks and three target models, COBRA-Skills consistently achieves the strongest average performance of the methods compared. It reduces optimization cost by fifty-five to fifty-eight percent relative to SkillOpt, and uses only fifty unique optimization examples per benchmark. Further analyses show it remains robust to changes in the agent harness, and that it performs well when the target model itself generates and refines the skills.
Economy is the contribution here. Evaluation is the costly step in improving skills, and COBRA-Skills treats each one as a decision to be made deliberately, rather than a fee to be paid for every candidate.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Every quest starts with a wall. Here it is the price of finding out whether a skill helps.
You are improving the written skills an LLM agent reads before each task. The only way to know whether a candidate skill helps is to run the agent with it on real tasks.
The road ahead
- 1 · The problem ~6 min Finding out whether a written skill helps an agent means running the agent with it, and those runs are costly and the task data is scarce.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Treat skill search as spending a fixed number of evaluations, let a contextual bandit pick each next skill to test, and keep refining the candidates.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Each round, score all candidates with the bandit, run the top one, update the predictor, and periodically swap weak skills for evidence-built ones.
- 6 · The experiment ~6 min You’ll guess: Compared with SkillOpt, by how much did COBRA-Skills reduce the cost of optimizing agent skills?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.