project page: <a href=\"https://thrillcrazyer.github.io/LoopUS\" rel=\"nofollow\">https://thrillcrazyer.github.io/LoopUS</a></p>\n","updatedAt":"2026-05-13T02:20:40.968Z","author":{"_id":"64ba75761d0a5a5760874197","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64ba75761d0a5a5760874197/5V3jxX-vim-wMaU2Zjz5h.jpeg","fullname":"TaekHyunPark","name":"Thrillcrazyer","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":4,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.8645671606063843},"editors":["Thrillcrazyer"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/64ba75761d0a5a5760874197/5V3jxX-vim-wMaU2Zjz5h.jpeg"],"reactions":[],"isReport":false}},{"id":"6a0481e76ec5a78f4ad7a1ec","author":{"_id":"661ab1f1fa3b144a381fa454","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/661ab1f1fa3b144a381fa454/IlpZBb9NCjo7ntFwMIH53.png","fullname":"Urro","name":"urroxyz","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":9,"isUserFollowing":false},"createdAt":"2026-05-13T13:51:35.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"I love this type of stuff.","html":"<p>I love this type of stuff.</p>\n","updatedAt":"2026-05-13T13:51:35.052Z","author":{"_id":"661ab1f1fa3b144a381fa454","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/661ab1f1fa3b144a381fa454/IlpZBb9NCjo7ntFwMIH53.png","fullname":"Urro","name":"urroxyz","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":9,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8693033456802368},"editors":["urroxyz"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/661ab1f1fa3b144a381fa454/IlpZBb9NCjo7ntFwMIH53.png"],"reactions":[{"reaction":"❤️","users":["Thrillcrazyer"],"count":1}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.11011","authors":[{"_id":"6a03d68086b054ce2fa40cf0","user":{"_id":"64ba75761d0a5a5760874197","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64ba75761d0a5a5760874197/5V3jxX-vim-wMaU2Zjz5h.jpeg","isPro":false,"fullname":"TaekHyunPark","user":"Thrillcrazyer","type":"user","name":"Thrillcrazyer"},"name":"Taekhyun Park","status":"claimed_verified","statusLastChangedAt":"2026-05-13T07:51:08.707Z","hidden":false},{"_id":"6a03d68086b054ce2fa40cf1","name":"Yongjae Lee","hidden":false},{"_id":"6a03d68086b054ce2fa40cf2","name":"Dohee Kim","hidden":false},{"_id":"6a03d68086b054ce2fa40cf3","name":"Hyerim Bae","hidden":false}],"publishedAt":"2026-05-10T00:00:00.000Z","submittedOnDailyAt":"2026-05-13T00:00:00.000Z","title":"LoopUS: Recasting Pretrained LLMs into Looped Latent Refinement Models","submittedOnDailyBy":{"_id":"64ba75761d0a5a5760874197","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64ba75761d0a5a5760874197/5V3jxX-vim-wMaU2Zjz5h.jpeg","isPro":false,"fullname":"TaekHyunPark","user":"Thrillcrazyer","type":"user","name":"Thrillcrazyer"},"summary":"Looped computation shows promise in improving the reasoning-oriented performance of LLMs by scaling test-time compute. However, existing approaches typically require either training recurrent models from scratch or applying disruptive retrofits, which involve substantial computational costs and may compromise pretrained capabilities. To address these limitations, we introduce Looped Depth Up-Scaling (LoopUS), a post-training framework that converts a standard pretrained LLM into a looped architecture. As a key technical contribution, LoopUS recasts the pretrained LLM into an encoder, a looped reasoning block, and a decoder. It operationalizes this latent-refinement architecture through four core components: (1) block decomposition, guided by staged representation dynamics; (2) an input-dependent selective gate to mitigate hidden-state drift; (3) random deep supervision for memory-efficient learning over long recursive horizons; and (4) a confidence head for adaptive early exiting. Collectively, these mechanisms transform a standard non-looped model into a looped form while stabilizing it against both computational bottlenecks and representation collapse. Through stable latent looping, LoopUS improves reasoning-oriented performance without extending the generated traces or requiring recurrent training from scratch. For more details, see https://thrillcrazyer.github.io/LoopUS","upvotes":8,"discussionId":"6a03d68086b054ce2fa40cf4","projectPage":"https://thrillcrazyer.github.io/LoopUS","githubRepo":"https://github.com/Thrillcrazyer/LoopUS","githubRepoAddedBy":"user","ai_summary":"LoopUS is a post-training framework that transforms pretrained LLMs into looped architectures for improved reasoning performance through latent-refinement and adaptive early exiting mechanisms.","ai_keywords":["looped computation","LLMs","post-training framework","encoder","looped reasoning block","decoder","block decomposition","staged representation dynamics","input-dependent selective gate","hidden-state drift","random deep supervision","memory-efficient learning","confidence head","adaptive early exiting","latent-refinement architecture","representation collapse"],"githubStars":2,"organization":{"_id":"6902caeadf78e6ca12c2a398","name":"BAELABPNU","fullname":"BIGDATA ANALYTICS ENGINEERING LAB, Pusan National University, Busan, Korea","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6902c9cb9427990a4948a33e/L59exY-PO66fQXk3lNwQ-.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"64ba75761d0a5a5760874197","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64ba75761d0a5a5760874197/5V3jxX-vim-wMaU2Zjz5h.jpeg","isPro":false,"fullname":"TaekHyunPark","user":"Thrillcrazyer","type":"user"},{"_id":"6902c9cb9427990a4948a33e","avatarUrl":"/avatars/b66536939cde8cc9a4877ded2d99434a.svg","isPro":false,"fullname":"Hyerim BAE","user":"hrbae","type":"user"},{"_id":"68c2af6e6f98def012e1c890","avatarUrl":"/avatars/55fe446471c3e70df16cd1bc44d95b04.svg","isPro":false,"fullname":"Kyunghun Lee","user":"DVDLuv","type":"user"},{"_id":"6a03dbaff6af1506edc14042","avatarUrl":"/avatars/81a0c9d695bc0165ba639042b730f00c.svg","isPro":false,"fullname":"KIMJONGMIN","user":"Zomezome","type":"user"},{"_id":"6902eb791cbf187eb732c7cc","avatarUrl":"/avatars/fbafb51f3e242ea72e607ca359850d6a.svg","isPro":false,"fullname":"SEO JUNHYUK","user":"junsj","type":"user"},{"_id":"697edf1faea9927d62855d2c","avatarUrl":"/avatars/15c9588429737287bdc7a452450919e0.svg","isPro":false,"fullname":"MinGyun Kang","user":"nneans","type":"user"},{"_id":"6902e7d99427990a494a9e05","avatarUrl":"/avatars/18fdaa0d48172124114ef3d74679926a.svg","isPro":false,"fullname":"somyeong kim","user":"somyeongkim","type":"user"},{"_id":"661ab1f1fa3b144a381fa454","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/661ab1f1fa3b144a381fa454/IlpZBb9NCjo7ntFwMIH53.png","isPro":true,"fullname":"Urro","user":"urroxyz","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6902caeadf78e6ca12c2a398","name":"BAELABPNU","fullname":"BIGDATA ANALYTICS ENGINEERING LAB, Pusan National University, Busan, Korea","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6902c9cb9427990a4948a33e/L59exY-PO66fQXk3lNwQ-.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.11011.md"}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2605.11011 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2605.11011 in a Space README.md to link it from this page.

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