**Which skill is most worth evaluating next?**\n\nInstead of exhaustively trying and refining candidate skills, COBRA-Skills maintains a population of skills and uses:\n\n- a neural predictor to estimate skill utility,\n- **LinearUCB** to balance exploration and exploitation,\n- target-agent execution feedback to update the bandit,\n- scheduled evolution operators (regeneration, rollout mutation, and crossover) to evolve the skill population.\n\n### Results\n\nWe evaluate COBRA-Skills on **6 diverse agent benchmarks × 3 target models**, covering search QA, spreadsheets, document understanding, mathematical reasoning, social reasoning, and embodied tasks.\n\nCOBRA-Skills achieves:\n\n- 🏆 the best average performance among the compared methods,\n- 💰 **~55–58% lower optimization cost** than SkillOpt,\n- 📊 only **50 optimization examples per benchmark**,\n- 🔧 consistent effectiveness under **Codex and Claude Code** harnesses,\n- 🤖 strong performance even when the **target model itself** generates and refines skills.\n\nAn interesting finding is that the cost reduction does not mainly come from reducing target-agent executions. A large part comes from avoiding repeated LLM-based trajectory analysis and skill rewriting.\n\n📄 Paper: https://arxiv.org/abs/2609.11682 \n💻 Code: https://github.com/Jerry-LuP/COBRA-Skills\n\nFeedback and discussions are very welcome!","html":"<h1 class=\"relative group flex items-baseline\">\n\t<a id=\"cobra-skills-contextual-bandits-for-efficient-agent-skill-optimization-🚀\" class=\"block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full\" href=\"#cobra-skills-contextual-bandits-for-efficient-agent-skill-optimization-🚀\" rel=\"nofollow\">\n\t\t<span class=\"header-link\"><svg class=\"text-gray-500 hover:text-black dark:hover:text-gray-200 w-4\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" aria-hidden=\"true\" role=\"img\" width=\"1em\" height=\"1em\" preserveAspectRatio=\"xMidYMid meet\" viewBox=\"0 0 256 256\"><path d=\"M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z\" fill=\"currentColor\"></path></svg></span>\n\t</a>\n\t<span>\n\t\tCOBRA-Skills: Contextual Bandits for Efficient Agent Skill Optimization 🚀\n\t</span>\n</h1>\n<p>How can we optimize Agent Skills without repeatedly spending large amounts of computation on weak candidates and costly LLM-based refinement?</p>\n<p><strong>COBRA-Skills</strong> treats skill optimization as a sequential budget-allocation problem:</p>\n<blockquote>\n<p><strong>Which skill is most worth evaluating next?</strong></p>\n</blockquote>\n<p>Instead of exhaustively trying and refining candidate skills, COBRA-Skills maintains a population of skills and uses:</p>\n<ul>\n<li>a neural predictor to estimate skill utility,</li>\n<li><strong>LinearUCB</strong> to balance exploration and exploitation,</li>\n<li>target-agent execution feedback to update the bandit,</li>\n<li>scheduled evolution operators (regeneration, rollout mutation, and crossover) to evolve the skill population.</li>\n</ul>\n<h3 class=\"relative group flex items-baseline\">\n\t<a id=\"results\" class=\"block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full\" href=\"#results\" rel=\"nofollow\">\n\t\t<span class=\"header-link\"><svg class=\"text-gray-500 hover:text-black dark:hover:text-gray-200 w-4\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" aria-hidden=\"true\" role=\"img\" width=\"1em\" height=\"1em\" preserveAspectRatio=\"xMidYMid meet\" viewBox=\"0 0 256 256\"><path d=\"M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z\" fill=\"currentColor\"></path></svg></span>\n\t</a>\n\t<span>\n\t\tResults\n\t</span>\n</h3>\n<p>We evaluate COBRA-Skills on <strong>6 diverse agent benchmarks × 3 target models</strong>, covering search QA, spreadsheets, document understanding, mathematical reasoning, social reasoning, and embodied tasks.</p>\n<p>COBRA-Skills achieves:</p>\n<ul>\n<li>🏆 the best average performance among the compared methods,</li>\n<li>💰 <strong>~55–58% lower optimization cost</strong> than SkillOpt,</li>\n<li>📊 only <strong>50 optimization examples per benchmark</strong>,</li>\n<li>🔧 consistent effectiveness under <strong>Codex and Claude Code</strong> harnesses,</li>\n<li>🤖 strong performance even when the <strong>target model itself</strong> generates and refines skills.</li>\n</ul>\n<p>An interesting finding is that the cost reduction does not mainly come from reducing target-agent executions. A large part comes from avoiding repeated LLM-based trajectory analysis and skill rewriting.</p>\n<p>📄 Paper: <a href=\"https://arxiv.org/abs/2609.11682\" rel=\"nofollow\">https://arxiv.org/abs/2609.11682</a><br>💻 Code: <a href=\"https://github.com/Jerry-LuP/COBRA-Skills\" rel=\"nofollow\">https://github.com/Jerry-LuP/COBRA-Skills</a></p>\n<p>Feedback and discussions are very welcome!</p>\n","updatedAt":"2026-09-14T02:09:05.673Z","author":{"_id":"67208288846173fe57b31427","avatarUrl":"/avatars/920dfb18cf30d98e17d6a7f3843e663f.svg","fullname":"Jerry smith","name":"Lpc206","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":2,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8400620222091675},"editors":["Lpc206"],"editorAvatarUrls":["/avatars/920dfb18cf30d98e17d6a7f3843e663f.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2609.11682","authors":[{"_id":"6aa3669547a406da7901e6bd","user":{"_id":"67208288846173fe57b31427","avatarUrl":"/avatars/920dfb18cf30d98e17d6a7f3843e663f.svg","isPro":false,"fullname":"Jerry smith","user":"Lpc206","type":"user","name":"Lpc206"},"name":"Pingchen Lu","status":"claimed_verified","statusLastChangedAt":"2026-09-11T09:53:31.646Z","hidden":false},{"_id":"6aa3669547a406da7901e6be","name":"Xiangyi Wang","hidden":false},{"_id":"6aa3669547a406da7901e6bf","name":"Xiang Li","hidden":false},{"_id":"6aa3669547a406da7901e6c0","name":"Jie Mao","hidden":false},{"_id":"6aa3669547a406da7901e6c1","name":"Zikun Qu","hidden":false},{"_id":"6aa3669547a406da7901e6c2","name":"Junfeng Luo","hidden":false},{"_id":"6aa3669547a406da7901e6c3","name":"Yao Shu","hidden":false},{"_id":"6aa3669547a406da7901e6c4","name":"Bryan Kian Hsiang Low","hidden":false},{"_id":"6aa3669547a406da7901e6c5","name":"Zhongxiang Dai","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/67208288846173fe57b31427/Ikw82bmRnjsXzLWN3Wgbs.png","https://cdn-uploads.huggingface.co/production/uploads/67208288846173fe57b31427/VzFHMnI6TKRR5QMXOOZmX.png","https://cdn-uploads.huggingface.co/production/uploads/67208288846173fe57b31427/-XtSd8Kz8u-TX2P3tSd9s.png","https://cdn-uploads.huggingface.co/production/uploads/67208288846173fe57b31427/Sk6VCaoRexRBQ4aFYLUtB.png","https://cdn-uploads.huggingface.co/production/uploads/67208288846173fe57b31427/Ay0CxxJVGIPpxjN8EjG-i.png","https://cdn-uploads.huggingface.co/production/uploads/67208288846173fe57b31427/QisCBkx-rfXrKkdSo_otH.png","https://cdn-uploads.huggingface.co/production/uploads/67208288846173fe57b31427/H-LmlKDksmisVH5kDCz6D.png"],"publishedAt":"2026-09-10T00:00:00.000Z","submittedOnDailyAt":"2026-09-14T00:00:00.000Z","title":"COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization","submittedOnDailyBy":{"_id":"67208288846173fe57b31427","avatarUrl":"/avatars/920dfb18cf30d98e17d6a7f3843e663f.svg","isPro":false,"fullname":"Jerry smith","user":"Lpc206","type":"user","name":"Lpc206"},"summary":"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.","upvotes":26,"discussionId":"6aa3669547a406da7901e6c6","githubRepo":"https://github.com/Jerry-LuP/COBRA-Skills","githubRepoAddedBy":"user","ai_summary":"COBRA-Skills improves LLM agent skill optimization by using contextual-bandit prioritization and evidence-based evolution to cut evaluation costs while maintaining high performance.","ai_keywords":["COBRA-Skills","contextual-bandit-guided prioritization","evidence-grounded skill evolution","budgeted sequential optimization","skill optimization","LLM agents"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":5,"organization":{"_id":"6223644d0129f2097d69a407","name":"CUHKSZ","fullname":"Chinese University of Hong Kong, Shenzhen","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1646486592158-6108ae87823007eaf0c7bd1e.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"67208288846173fe57b31427","avatarUrl":"/avatars/920dfb18cf30d98e17d6a7f3843e663f.svg","isPro":false,"fullname":"Jerry smith","user":"Lpc206","type":"user"},{"_id":"6a795012189047f82009f3b3","avatarUrl":"/avatars/3367920f104ca45305944f6bc1d6b8df.svg","isPro":false,"fullname":"Zhongxiang Dai","user":"dzxagent","type":"user"},{"_id":"66440e86bfe15e84d369cb03","avatarUrl":"/avatars/d15b3b3831bc74138206071612169f64.svg","isPro":false,"fullname":"Xinyuan Xie","user":"SatsukiVie","type":"user"},{"_id":"6aa75f1d4a6cd3767abed18a","avatarUrl":"/avatars/3a5334b2873e31bae92f7d6e2fc8241e.svg","isPro":false,"fullname":"Kong Zedong","user":"dongdong666","type":"user"},{"_id":"6aa75e97d26a08997320615f","avatarUrl":"/avatars/549248a87cd5fc71ee2b800d71f7f91b.svg","isPro":false,"fullname":"Yan Lu","user":"Salt1222","type":"user"},{"_id":"6aa75f77c77c725f6846090b","avatarUrl":"/avatars/15aa6e38b0a6c58c1ba81ce43c52416a.svg","isPro":false,"fullname":"Cai ruyin","user":"yymtdykx","type":"user"},{"_id":"6aa76030dd4d0e0172461c23","avatarUrl":"/avatars/802485e594fd0a0cc447003a97853822.svg","isPro":false,"fullname":"Xiaoyu Nie","user":"Nxy-207","type":"user"},{"_id":"69fd548a154c7dfe3c9bea2e","avatarUrl":"/avatars/9f3610aa39d4ab9b01d203f80dbd1ee2.svg","isPro":false,"fullname":"L","user":"lzx666YAYAYA","type":"user"},{"_id":"6aa7640e051d3d79e9910de5","avatarUrl":"/avatars/19187167bdc43616cee60ca4fcd16663.svg","isPro":false,"fullname":"王一川","user":"Wangyichuan","type":"user"},{"_id":"6aa765df83fe5c75d66855a4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/tcn_NZiwFhAaM-iZ5oHNN.png","isPro":false,"fullname":"vincent","user":"E23Q","type":"user"},{"_id":"67f369756285f17584396bb0","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/J6e1rXahcyUFtmI152nMF.png","isPro":false,"fullname":"李喆","user":"learning-ljj","type":"user"},{"_id":"6aa76f782e4c51b0c95784cc","avatarUrl":"/avatars/43856d09e3cff816405ee582462b633e.svg","isPro":false,"fullname":"mrj369","user":"mrj369","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6223644d0129f2097d69a407","name":"CUHKSZ","fullname":"Chinese University of Hong Kong, Shenzhen","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1646486592158-6108ae87823007eaf0c7bd1e.png"},"query":{}}">

Cite arxiv.org/abs/2609.11682 in a model README.md to link it from this page.

Cite arxiv.org/abs/2609.11682 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2609.11682 in a Space README.md to link it from this page.

Add this paper to a collection to link it from this page.