Excited to share LoopArena, a benchmark for Loop Engineering that measures how well a model can guide a fixed coding agent through long-running software development tasks. LoopArena evaluates this ability at three complementary levels, from individual control decisions to task slices and complete tasks. Our results show that reliable long-horizon loop control remains challenging, while the lower-cost Type II setting reduces estimated inference cost by 64.4% and preserves a similar model ordering to full-task evaluation. We welcome feedback on the paper, benchmark, and evaluation protocol!</p>\n","updatedAt":"2026-08-31T11:19:21.134Z","author":{"_id":"6773bcaa675a971ddf1e81dd","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/a8VUwZYXd7O_mq_zFvXMh.png","fullname":"Yi Wang","name":"CokeWang","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8669617772102356},"editors":["CokeWang"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/a8VUwZYXd7O_mq_zFvXMh.png"],"reactions":[],"isReport":false}},{"id":"6a956f6ea55cf629fadb7683","author":{"_id":"658412f93a84a40185adaf37","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/658412f93a84a40185adaf37/FKXH7e1jj09KO1v-B5sER.jpeg","fullname":"Aamer Mihaysi","name":"O96a","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false},"createdAt":"2026-08-31T12:11:26.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"End-to-end scores on loop benchmarks have the same attribution problem I hit in prod every week: a run succeeds and I can't tell if the loop's guidance carried it or the model just happened to be good. So do I fix the loop or swap the model? That's the real debugging question, and a single final score can't answer it — a strong model will paper over a weak controller every time. I'd want per-controller failure breakdowns: where did the loop misdirect, where did the model fail to follow. Until the benchmark separates those, the number tells me less than the logs do.","html":"<p>End-to-end scores on loop benchmarks have the same attribution problem I hit in prod every week: a run succeeds and I can't tell if the loop's guidance carried it or the model just happened to be good. So do I fix the loop or swap the model? That's the real debugging question, and a single final score can't answer it — a strong model will paper over a weak controller every time. I'd want per-controller failure breakdowns: where did the loop misdirect, where did the model fail to follow. Until the benchmark separates those, the number tells me less than the logs do.</p>\n","updatedAt":"2026-08-31T12:11:26.688Z","author":{"_id":"658412f93a84a40185adaf37","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/658412f93a84a40185adaf37/FKXH7e1jj09KO1v-B5sER.jpeg","fullname":"Aamer Mihaysi","name":"O96a","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.9430707693099976},"editors":["O96a"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/658412f93a84a40185adaf37/FKXH7e1jj09KO1v-B5sER.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.28281","authors":[{"_id":"6a94ecd5073195fee5157233","user":{"_id":"6773bcaa675a971ddf1e81dd","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/a8VUwZYXd7O_mq_zFvXMh.png","isPro":false,"fullname":"Yi Wang","user":"CokeWang","type":"user","name":"CokeWang"},"name":"Yi Wang","status":"claimed_verified","statusLastChangedAt":"2026-08-31T08:38:31.285Z","hidden":false},{"_id":"6a94ecd5073195fee5157234","name":"Haopeng Zhang","hidden":false},{"_id":"6a94ecd5073195fee5157235","name":"Chengxiang Huang","hidden":false},{"_id":"6a94ecd5073195fee5157236","name":"Rui Dai","hidden":false},{"_id":"6a94ecd5073195fee5157237","name":"Kaikui Liu","hidden":false},{"_id":"6a94ecd5073195fee5157238","name":"Piotr Koniusz","hidden":false},{"_id":"6a94ecd5073195fee5157239","name":"Xiangxiang Chu","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/6773bcaa675a971ddf1e81dd/CfPlR-glQZ4PWXh6rQQjl.mp4"],"publishedAt":"2026-08-28T00:00:00.000Z","submittedOnDailyAt":"2026-08-31T00:00:00.000Z","title":"LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering","submittedOnDailyBy":{"_id":"6773bcaa675a971ddf1e81dd","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/a8VUwZYXd7O_mq_zFvXMh.png","isPro":false,"fullname":"Yi Wang","user":"CokeWang","type":"user","name":"CokeWang"},"summary":"Loop Engineering is emerging as a practice for organizing development work around coding agents. Instead of writing each prompt by hand, practitioners design loops that monitor progress, assign work, run checks, and decide what the agent should do next. Even with a capable coding agent, a loop may trust a stale progress note, skip needed verification, spend its budget in the wrong direction, or stop before the task is safe to submit. Yet the final outcome of one end-to-end run cannot tell whether success or failure reflects the loop's guidance or the coding agent's ability to carry out the task. We introduce LoopArena, a benchmark for evaluating how well one model can guide a separate coding agent through a long-running task. The model under evaluation is the Controller: after each coding round, it receives a structured summary of the run and instructs a separate, fixed coding agent, the Worker, on what to do or verify next, or decides whether to stop. LoopArena evaluates this ability in three complementary settings that differ in execution scope and cost. Type I scores next-step Loop Contract selection through execution-validated questions without running the Worker at evaluation time. Type II executes repeated control over a selected slice of a full task, while Type III evaluates the paired full task from its original state. On full tasks, the best observed Strict Success Rate is 24.69\\%, leaving substantial room for improvement in long-horizon loop control. Across Controllers, the paired reduction in estimated inference cost averages 64.4\\%, and Type II produces a similar ordering under the main Core criterion (Spearman's \\(ρ=0.9747\\)). We release the benchmark data and evaluation code at https://github.com/AMAP-ML/LoopArena .","upvotes":78,"discussionId":"6a94ecd5073195fee515723a","projectPage":"https://amap-ml.github.io/LoopArena/","githubRepo":"https://github.com/AMAP-ML/LoopArena","githubRepoAddedBy":"user","ai_summary":"LoopArena benchmarks how well a controller model guides a separate coding agent through long tasks, revealing low strict success rates and significant cost reductions.","ai_keywords":["Loop Engineering","coding agents","LoopArena","Controller","Worker","Loop Contract","Strict Success Rate","inference cost","Spearman's ρ"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":67,"organization":{"_id":"67d11771890254196d3174e5","name":"GD-ML","fullname":"AMAP-ML","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/67d116c47be76de1a40873ca/s5ukAx9E36ZZIKvbpBRi4.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"661de9defdbc9c247f159d15","avatarUrl":"/avatars/38e21e78327cc908201122405c48f41b.svg","isPro":false,"fullname":"Rui Dai","user":"DerryD","type":"user"},{"_id":"6773bcaa675a971ddf1e81dd","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/a8VUwZYXd7O_mq_zFvXMh.png","isPro":false,"fullname":"Yi Wang","user":"CokeWang","type":"user"},{"_id":"6717ace13fc94b77906d2f35","avatarUrl":"/avatars/a9df1130e76b7aee36e0b0362b01106b.svg","isPro":false,"fullname":"myu","user":"senmengg","type":"user"},{"_id":"650baea0534285d49a5a80de","avatarUrl":"/avatars/c40496b562f9c6e6f21439970d57bab0.svg","isPro":false,"fullname":"mist","user":"mist2234","type":"user"},{"_id":"676becb8317c6dbe689474f0","avatarUrl":"/avatars/39209fce8ed644426d36e1a67a192fc2.svg","isPro":false,"fullname":"bill","user":"hunchteller","type":"user"},{"_id":"64be128a2e66dc7b8bd8459d","avatarUrl":"/avatars/ac5a5246dc19dd35bbd89d7fc492cba5.svg","isPro":false,"fullname":"Rui Chen","user":"ruichen9618","type":"user"},{"_id":"645e0282412d9e9cd044c764","avatarUrl":"/avatars/1eebe4e89e151611ea96c38483549bd0.svg","isPro":false,"fullname":"gao","user":"bingjie","type":"user"},{"_id":"689d4a717fb1c6267bb59acc","avatarUrl":"/avatars/13b5117a0108e40cbffa8114c112cfad.svg","isPro":false,"fullname":"peter","user":"peterlrm","type":"user"},{"_id":"6a6a0475b2e0394c7b820203","avatarUrl":"/avatars/8ac69680f692330dd0e72e1234e53cc7.svg","isPro":false,"fullname":"Swph","user":"Swphhh","type":"user"},{"_id":"6a68922db550eeafa5ce441d","avatarUrl":"/avatars/76a229f93631a393f603d7d68596691e.svg","isPro":false,"fullname":"feng","user":"xiaoe112233","type":"user"},{"_id":"65003db8bef9b594656f8fa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/65003db8bef9b594656f8fa7/L6cvPOAeBRnFnIQwWxYyf.png","isPro":false,"fullname":"Hailang Huang","user":"lerogo","type":"user"},{"_id":"695b26295e8bad3be68b15f4","avatarUrl":"/avatars/2d4f0988dfa72f3388824ef2f377a546.svg","isPro":false,"fullname":"Fortune Wang","user":"FortuneV2","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"67d11771890254196d3174e5","name":"GD-ML","fullname":"AMAP-ML","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/67d116c47be76de1a40873ca/s5ukAx9E36ZZIKvbpBRi4.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.28281.md","query":{}}">
LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering
Abstract
LoopArena benchmarks how well a controller model guides a separate coding agent through long tasks, revealing low strict success rates and significant cost reductions.
Loop Engineering is emerging as a practice for organizing development work around coding agents. Instead of writing each prompt by hand, practitioners design loops that monitor progress, assign work, run checks, and decide what the agent should do next. Even with a capable coding agent, a loop may trust a stale progress note, skip needed verification, spend its budget in the wrong direction, or stop before the task is safe to submit. Yet the final outcome of one end-to-end run cannot tell whether success or failure reflects the loop's guidance or the coding agent's ability to carry out the task. We introduce LoopArena, a benchmark for evaluating how well one model can guide a separate coding agent through a long-running task. The model under evaluation is the Controller: after each coding round, it receives a structured summary of the run and instructs a separate, fixed coding agent, the Worker, on what to do or verify next, or decides whether to stop. LoopArena evaluates this ability in three complementary settings that differ in execution scope and cost. Type I scores next-step Loop Contract selection through execution-validated questions without running the Worker at evaluation time. Type II executes repeated control over a selected slice of a full task, while Type III evaluates the paired full task from its original state. On full tasks, the best observed Strict Success Rate is 24.69\%, leaving substantial room for improvement in long-horizon loop control. Across Controllers, the paired reduction in estimated inference cost averages 64.4\%, and Type II produces a similar ordering under the main Core criterion (Spearman's \(ρ=0.9747\)). We release the benchmark data and evaluation code at https://github.com/AMAP-ML/LoopArena .
Community
Excited to share LoopArena, a benchmark for Loop Engineering that measures how well a model can guide a fixed coding agent through long-running software development tasks. LoopArena evaluates this ability at three complementary levels, from individual control decisions to task slices and complete tasks. Our results show that reliable long-horizon loop control remains challenging, while the lower-cost Type II setting reduces estimated inference cost by 64.4% and preserves a similar model ordering to full-task evaluation. We welcome feedback on the paper, benchmark, and evaluation protocol!
End-to-end scores on loop benchmarks have the same attribution problem I hit in prod every week: a run succeeds and I can't tell if the loop's guidance carried it or the model just happened to be good. So do I fix the loop or swap the model? That's the real debugging question, and a single final score can't answer it — a strong model will paper over a weak controller every time. I'd want per-controller failure breakdowns: where did the loop misdirect, where did the model fail to follow. Until the benchmark separates those, the number tells me less than the logs do.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2608.28281 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.28281 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.28281 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.