Hugging Face Daily Papers · June 17, 2026 · 4 min read

LoopCoder-v2: Only Loop Once for Efficient Test-Time Computation Scaling

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

🤗 Model: <a href=\"https://huggingface.co/Multilingual-Multimodal-NLP/LoopCoder-V2\">https://huggingface.co/Multilingual-Multimodal-NLP/LoopCoder-V2</a> 💻 Code: <a href=\"https://github.com/CSJianYang/LoopCoder\" rel=\"nofollow\">https://github.com/CSJianYang/LoopCoder</a> 📄Paper: <a href=\"https://arxiv.org/abs/2606.18023\" rel=\"nofollow\">https://arxiv.org/abs/2606.18023</a>\n","updatedAt":"2026-06-17T06:39:12.839Z","author":{"_id":"69c67b2fa994b07915a6e083","avatarUrl":"/avatars/3d0fd966df540d34095d2c84ce449180.svg","fullname":"wei zhang","name":"zwpride","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.6228488087654114},"editors":["zwpride"],"editorAvatarUrls":["/avatars/3d0fd966df540d34095d2c84ce449180.svg"],"reactions":[],"isReport":false}},{"id":"6a32af36be102880f9ab5324","author":{"_id":"65243980050781c16f234f1f","avatarUrl":"/avatars/743a009681d5d554c27e04300db9f267.svg","fullname":"Avi","name":"avahal","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":6,"isUserFollowing":false},"createdAt":"2026-06-17T14:29:10.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"Interesting breakdown of this paper on arXivLens: https://arxivlens.com/PaperView/Details/loopcoder-v2-only-loop-once-for-efficient-test-time-computation-scaling-6079-1d2399ff\nCovers the executive summary, detailed methodology, and practical applications.","html":"Interesting breakdown of this paper on arXivLens: <a href=\"https://arxivlens.com/PaperView/Details/loopcoder-v2-only-loop-once-for-efficient-test-time-computation-scaling-6079-1d2399ff\" rel=\"nofollow\">https://arxivlens.com/PaperView/Details/loopcoder-v2-only-loop-once-for-efficient-test-time-computation-scaling-6079-1d2399ff</a> Covers the executive summary, detailed methodology, and practical applications.\n","updatedAt":"2026-06-17T14:29:10.578Z","author":{"_id":"65243980050781c16f234f1f","avatarUrl":"/avatars/743a009681d5d554c27e04300db9f267.svg","fullname":"Avi","name":"avahal","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":6,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7079436182975769},"editors":["avahal"],"editorAvatarUrls":["/avatars/743a009681d5d554c27e04300db9f267.svg"],"reactions":[],"isReport":false},"replies":[{"id":"6a32e35bd6472d7e7a3f98e5","author":{"_id":"661ab1f1fa3b144a381fa454","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/661ab1f1fa3b144a381fa454/IlpZBb9NCjo7ntFwMIH53.png","fullname":"Urro","name":"urroxyz","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":12,"isUserFollowing":false},"createdAt":"2026-06-17T18:11:39.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"@adamm-hf ","html":"<a href=\"/adamm-hf\">@adamm-hf</a> \n","updatedAt":"2026-06-17T18:11:39.245Z","author":{"_id":"661ab1f1fa3b144a381fa454","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/661ab1f1fa3b144a381fa454/IlpZBb9NCjo7ntFwMIH53.png","fullname":"Urro","name":"urroxyz","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":12,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"ru","probability":0.15853583812713623},"editors":["urroxyz"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/661ab1f1fa3b144a381fa454/IlpZBb9NCjo7ntFwMIH53.png"],"reactions":[],"isReport":false,"parentCommentId":"6a32af36be102880f9ab5324"}}]}],"primaryEmailConfirmed":false,"paper":{"id":"2606.18023","authors":[{"_id":"6a3201f3bc818ff14e453d61","name":"Jian Yang","hidden":false},{"_id":"6a3201f3bc818ff14e453d62","name":"Shawn Guo","hidden":false},{"_id":"6a3201f3bc818ff14e453d63","name":"Wei Zhang","hidden":false},{"_id":"6a3201f3bc818ff14e453d64","user":{"_id":"64ab99dcb76bfd863eba64c1","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64ab99dcb76bfd863eba64c1/UBXwDPx17X-gl-SzBPvrc.jpeg","isPro":false,"fullname":"TY.Zheng","user":"aaabiao","type":"user","name":"aaabiao"},"name":"Tianyu Zheng","status":"claimed_verified","statusLastChangedAt":"2026-06-17T11:21:10.407Z","hidden":false},{"_id":"6a3201f3bc818ff14e453d65","name":"Yaxin Du","hidden":false},{"_id":"6a3201f3bc818ff14e453d66","name":"Haau-Sing Li","hidden":false},{"_id":"6a3201f3bc818ff14e453d67","user":{"_id":"66a8e2538407031e388c501f","avatarUrl":"/avatars/d16d51f7b1e111efd6d0985995b614be.svg","isPro":false,"fullname":"wjj","user":"wuyuverse","type":"user","name":"wuyuverse"},"name":"Jiajun Wu","status":"claimed_verified","statusLastChangedAt":"2026-06-17T11:21:12.507Z","hidden":false},{"_id":"6a3201f3bc818ff14e453d68","name":"Yue Song","hidden":false},{"_id":"6a3201f3bc818ff14e453d69","name":"Yan Xing","hidden":false},{"_id":"6a3201f3bc818ff14e453d6a","name":"Qingsong Cai","hidden":false},{"_id":"6a3201f3bc818ff14e453d6b","name":"Zelong Huang","hidden":false},{"_id":"6a3201f3bc818ff14e453d6c","name":"Chuan Hao","hidden":false},{"_id":"6a3201f3bc818ff14e453d6d","name":"Ran Tao","hidden":false},{"_id":"6a3201f3bc818ff14e453d6e","name":"Xianglong Liu","hidden":false},{"_id":"6a3201f3bc818ff14e453d6f","name":"Wayne Xin Zhao","hidden":false},{"_id":"6a3201f3bc818ff14e453d70","name":"Mingjie Tang","hidden":false},{"_id":"6a3201f3bc818ff14e453d71","name":"Weifeng Lv","hidden":false},{"_id":"6a3201f3bc818ff14e453d72","name":"Ming Zhou","hidden":false},{"_id":"6a3201f3bc818ff14e453d73","name":"Bryan Dai","hidden":false}],"publishedAt":"2026-06-16T00:00:00.000Z","submittedOnDailyAt":"2026-06-17T00:00:00.000Z","title":"LoopCoder-v2: Only Loop Once for Efficient Test-Time Computation Scaling","submittedOnDailyBy":{"_id":"6039478ab3ecf716b1a5fd4d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6039478ab3ecf716b1a5fd4d/_Thy4E7taiSYBLKxEKJbT.jpeg","isPro":true,"fullname":"taesiri","user":"taesiri","type":"user","name":"taesiri"},"summary":"Looped Transformers scale latent computation by repeatedly applying shared blocks, but sequential looping increases latency and KV-cache memory with the loop count. Parallel loop Transformers (PLT) alleviate this cost through cross-loop position offsets (CLP) and shared-KV gated sliding-window attention, making loop count a practical design choice. We therefore study PLT loop-count selection through a gain--cost view: an extra loop may refine representations, but CLP also introduces a positional mismatch at each loop boundary. We instantiate this study by training LoopCoder-v2, a family of 7B PLT coders with different loop counts, from scratch on 18T tokens, followed by matched instruction tuning and evaluation. Empirically, the two-loop variant delivers broad gains over the non-looped baseline across code generation, code reasoning, agentic software engineering, and tool-use benchmarks, improving SWE-bench Verified from 43.0 to 64.4 points and Multi-SWE from 14.0 to 31.0 points. In contrast, variants with three or more loops regress, revealing a strongly non-monotonic loop-count effect. Our diagnostics show that loop 2 provides the main productive refinement, while later loops yield diminishing, oscillatory updates and reduced representational diversity. Because the CLP-induced mismatch remains roughly fixed as refinement gains shrink, the offset cost increasingly dominates. This gain--cost trade-off explains PLT's saturation at two loops and provides diagnostics for loop-count selection.","upvotes":113,"discussionId":"6a3201f3bc818ff14e453d74","ai_summary":"Parallel loop Transformers achieve better code generation performance with two loops due to refined representations, while additional loops cause diminishing returns and increased positional mismatch costs.","ai_keywords":["Looped Transformers","parallel loop Transformers","cross-loop position offsets","shared-KV gated sliding-window attention","loop-count selection","LoopCoder-v2","instruction tuning","SWE-bench","Multi-SWE"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"},{"_id":"69c67b2fa994b07915a6e083","avatarUrl":"/avatars/3d0fd966df540d34095d2c84ce449180.svg","isPro":false,"fullname":"wei zhang","user":"zwpride","type":"user"},{"_id":"66a8e2538407031e388c501f","avatarUrl":"/avatars/d16d51f7b1e111efd6d0985995b614be.svg","isPro":false,"fullname":"wjj","user":"wuyuverse","type":"user"},{"_id":"69df2fc26bd856fb9ebb0190","avatarUrl":"/avatars/3f51018cad4743064d06145c394e0eb3.svg","isPro":false,"fullname":"whw","user":"whw06","type":"user"},{"_id":"6382252f54421460665ec501","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6382252f54421460665ec501/gW9fev3T5QPcNq4f9hqB1.jpeg","isPro":false,"fullname":"Yizhi Li","user":"yizhilll","type":"user"},{"_id":"67400450e64f7fba857477b5","avatarUrl":"/avatars/53e780769acd142d72b0578cd6111984.svg","isPro":false,"fullname":"Junhang Cheng","user":"cjhCoder7","type":"user"},{"_id":"64dc39d27f749b6e34702b81","avatarUrl":"/avatars/3db6db301831b838dd172937ef7653df.svg","isPro":false,"fullname":"Du","user":"Dorothydu","type":"user"},{"_id":"64ab99dcb76bfd863eba64c1","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64ab99dcb76bfd863eba64c1/UBXwDPx17X-gl-SzBPvrc.jpeg","isPro":false,"fullname":"TY.Zheng","user":"aaabiao","type":"user"},{"_id":"695f604f980c20dbb2104cb6","avatarUrl":"/avatars/c4cd1039ec6aeb4b4c14c4d83fd4eeeb.svg","isPro":false,"fullname":"x","user":"xhxlb-12138","type":"user"},{"_id":"5ee6cd27464d0272c8b24545","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/5ee6cd27464d0272c8b24545/aSSDBKB84SH4gFuj64ovT.png","isPro":true,"fullname":"Haau-Sing Li","user":"lhaausing","type":"user"},{"_id":"648e5a70df53671f33e94d52","avatarUrl":"/avatars/ea196b6cb1350accc61925cb0875d437.svg","isPro":true,"fullname":"Hongxin Li","user":"HongxinLi","type":"user"},{"_id":"695b11e6a934de056c706ece","avatarUrl":"/avatars/b0655262ceaf3a74cdb7a4ce031b9405.svg","isPro":false,"fullname":"yan xing","user":"yxing-bj","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":1,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2606/2606.18023.md","query":{}}">

Papers

arxiv:2606.18023

LoopCoder-v2: Only Loop Once for Efficient Test-Time Computation Scaling

Published on Jun 16

· Submitted by

taesiri on Jun 17

#1 Paper of the day

Upvote

113

Authors:

Tianyu Zheng ,

Jiajun Wu ,

Abstract

Parallel loop Transformers achieve better code generation performance with two loops due to refined representations, while additional loops cause diminishing returns and increased positional mismatch costs.

Generated by Qwen/Qwen2.5-Coder-32B-Instruct

Looped Transformers scale latent computation by repeatedly applying shared blocks, but sequential looping increases latency and KV-cache memory with the loop count. Parallel loop Transformers (PLT) alleviate this cost through cross-loop position offsets (CLP) and shared-KV gated sliding-window attention, making loop count a practical design choice. We therefore study PLT loop-count selection through a gain--cost view: an extra loop may refine representations, but CLP also introduces a positional mismatch at each loop boundary. We instantiate this study by training LoopCoder-v2, a family of 7B PLT coders with different loop counts, from scratch on 18T tokens, followed by matched instruction tuning and evaluation. Empirically, the two-loop variant delivers broad gains over the non-looped baseline across code generation, code reasoning, agentic software engineering, and tool-use benchmarks, improving SWE-bench Verified from 43.0 to 64.4 points and Multi-SWE from 14.0 to 31.0 points. In contrast, variants with three or more loops regress, revealing a strongly non-monotonic loop-count effect. Our diagnostics show that loop 2 provides the main productive refinement, while later loops yield diminishing, oscillatory updates and reduced representational diversity. Because the CLP-induced mismatch remains roughly fixed as refinement gains shrink, the offset cost increasingly dominates. This gain--cost trade-off explains PLT's saturation at two loops and provides diagnostics for loop-count selection.