We introduce GRAM, a generative framework for recursive reasoning. Unlike deterministic recursive models such as HRM, TRM, and Looped Transformers, GRAM treats reasoning as a stochastic latent trajectory, allowing the model to explore multiple possible solution paths in parallel. This enables inference-time scaling not only by running deeper recursion, but also by sampling wider sets of reasoning trajectories. The same formulation supports both conditional reasoning, p(y|x), and unconditional generation, p(x). With only 10M parameters, GRAM achieves strong results on Sudoku-Extreme, ARC-AGI, and multi-solution tasks such as N-Queens.</p>\n","updatedAt":"2026-05-21T01:54:29.422Z","author":{"_id":"64c37ee87d89024360937d81","avatarUrl":"/avatars/cc9733b0862bbdca5e00f61a7ff7bb94.svg","fullname":"mingyu jo","name":"jojo0217","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":5,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8837345838546753},"editors":["jojo0217"],"editorAvatarUrls":["/avatars/cc9733b0862bbdca5e00f61a7ff7bb94.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.19376","authors":[{"_id":"6a0e64d1164dbbc68a26c410","name":"Junyeob Baek","hidden":false},{"_id":"6a0e64d1164dbbc68a26c411","name":"Mingyu Jo","hidden":false},{"_id":"6a0e64d1164dbbc68a26c412","name":"Minsu Kim","hidden":false},{"_id":"6a0e64d1164dbbc68a26c413","name":"Mengye Ren","hidden":false},{"_id":"6a0e64d1164dbbc68a26c414","name":"Yoshua Bengio","hidden":false},{"_id":"6a0e64d1164dbbc68a26c415","name":"Sungjin Ahn","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/64c37ee87d89024360937d81/HwM12lgDELM3_XYU9xTLh.gif","https://cdn-uploads.huggingface.co/production/uploads/64c37ee87d89024360937d81/UavA4IbIO37JXscFN7y73.gif"],"publishedAt":"2026-05-20T00:00:00.000Z","submittedOnDailyAt":"2026-05-21T00:00:00.000Z","title":"Generative Recursive Reasoning","submittedOnDailyBy":{"_id":"64c37ee87d89024360937d81","avatarUrl":"/avatars/cc9733b0862bbdca5e00f61a7ff7bb94.svg","isPro":false,"fullname":"mingyu jo","user":"jojo0217","type":"user","name":"jojo0217"},"summary":"How should future neural reasoning systems implement extended computation? Recursive Reasoning Models (RRMs) offer a promising alternative to autoregressive sequence extension by performing iterative latent-state refinement with shared transition functions. Yet existing RRMs are largely deterministic, following a single latent trajectory and converging to a single prediction. We introduce Generative Recursive reAsoning Models (GRAM), a framework that turns recursive latent reasoning into probabilistic multi-trajectory computation. GRAM models reasoning as a stochastic latent trajectory, enabling multiple hypotheses, alternative solution strategies, and inference-time scaling through both recursive depth and parallel trajectory sampling. This yields a latent-variable generative model supporting conditional reasoning via p_θ(y mid x) and, with fixed or absent inputs, unconditional generation via p_θ(x). Trained with amortized variational inference, GRAM improves over deterministic recurrent and recursive baselines on structured reasoning and multi-solution constraint satisfaction tasks, while demonstrating an unconditional generation capability. https://ahn-ml.github.io/gram-website","upvotes":14,"discussionId":"6a0e64d1164dbbc68a26c416","projectPage":"https://ahn-ml.github.io/gram-website/","ai_summary":"Generative Recursive reAsoning Models (GRAM) introduce probabilistic multi-trajectory computation for neural reasoning systems, enabling multiple hypotheses and parallel inference through stochastic latent trajectories.","ai_keywords":["recursive reasoning models","latent-state refinement","shared transition functions","probabilistic multi-trajectory computation","stochastic latent trajectory","conditional reasoning","unconditional generation","amortized variational inference","structured reasoning","constraint satisfaction"]},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"64c37ee87d89024360937d81","avatarUrl":"/avatars/cc9733b0862bbdca5e00f61a7ff7bb94.svg","isPro":false,"fullname":"mingyu jo","user":"jojo0217","type":"user"},{"_id":"67b4ae695bad3d37468f61e2","avatarUrl":"/avatars/170d08f17a9d795330adec5aaf951271.svg","isPro":false,"fullname":"SeungJu Back","user":"gnwsb","type":"user"},{"_id":"659f4d7bb62804e6f43319ce","avatarUrl":"/avatars/b6ce28db5670411b42458d9edb28854c.svg","isPro":false,"fullname":"Junyeong Park","user":"frechele","type":"user"},{"_id":"64ad860cb7d86b40fd866676","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/LyC0DffhiQZnILhh7Z5-Z.jpeg","isPro":false,"fullname":"Junyeob Baek","user":"cun-bjy","type":"user"},{"_id":"68bf80cb4c4e9b239413f11b","avatarUrl":"/avatars/5add1af26cc77aea9a70073d9813f5c2.svg","isPro":false,"fullname":"Dongwoo Lee","user":"dongwoolee00","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"},{"_id":"68a04c2e8c001e44b4c63bd9","avatarUrl":"/avatars/891949d2300efad97eb29b6ba8add49f.svg","isPro":false,"fullname":"Taegu","user":"prgmti1","type":"user"},{"_id":"6463554dd2044cd1d7c6e0bf","avatarUrl":"/avatars/d7653623117268c545a7063fec69664b.svg","isPro":false,"fullname":"Bingzheng Wei","user":"Bingzheng","type":"user"},{"_id":"66d8512c54209e9101811e8e","avatarUrl":"/avatars/62dfd8e6261108f2508efe678d5a2a57.svg","isPro":false,"fullname":"M Saad Salman","user":"MSS444","type":"user"},{"_id":"68b2a4157f881fc640ba7d80","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/lMTgr3pe7pOHtMe7bVF7F.png","isPro":false,"fullname":"khtsly","user":"khtsly","type":"user"},{"_id":"60eeedbf50b60c406afc1291","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1649111275459-60eeedbf50b60c406afc1291.png","isPro":false,"fullname":"Samuel Arcadinho","user":"SSamDav","type":"user"},{"_id":"6a0b2f731f2a5f058ce5c724","avatarUrl":"/avatars/0e2a406f3c8ae3d17e7e67139aca4553.svg","isPro":false,"fullname":"Tab Bab","user":"Tabbab321","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.19376.md"}">
Generative Recursive Reasoning
Abstract
Generative Recursive reAsoning Models (GRAM) introduce probabilistic multi-trajectory computation for neural reasoning systems, enabling multiple hypotheses and parallel inference through stochastic latent trajectories.
AI-generated summary
How should future neural reasoning systems implement extended computation? Recursive Reasoning Models (RRMs) offer a promising alternative to autoregressive sequence extension by performing iterative latent-state refinement with shared transition functions. Yet existing RRMs are largely deterministic, following a single latent trajectory and converging to a single prediction. We introduce Generative Recursive reAsoning Models (GRAM), a framework that turns recursive latent reasoning into probabilistic multi-trajectory computation. GRAM models reasoning as a stochastic latent trajectory, enabling multiple hypotheses, alternative solution strategies, and inference-time scaling through both recursive depth and parallel trajectory sampling. This yields a latent-variable generative model supporting conditional reasoning via p_θ(y mid x) and, with fixed or absent inputs, unconditional generation via p_θ(x). Trained with amortized variational inference, GRAM improves over deterministic recurrent and recursive baselines on structured reasoning and multi-solution constraint satisfaction tasks, while demonstrating an unconditional generation capability. https://ahn-ml.github.io/gram-website
Community
We introduce GRAM, a generative framework for recursive reasoning. Unlike deterministic recursive models such as HRM, TRM, and Looped Transformers, GRAM treats reasoning as a stochastic latent trajectory, allowing the model to explore multiple possible solution paths in parallel. This enables inference-time scaling not only by running deeper recursion, but also by sampling wider sets of reasoning trajectories. The same formulation supports both conditional reasoning, p(y|x), and unconditional generation, p(x). With only 10M parameters, GRAM achieves strong results on Sudoku-Extreme, ARC-AGI, and multi-solution tasks such as N-Queens.
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/2605.19376 in a model README.md to link it from this page.
Cite arxiv.org/abs/2605.19376 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2605.19376 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.