Hugging Face Daily Papers · May 13, 2026 · 3 min read

Teaching Language Models to Think in Code

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

Like Read original ↗

arxiv: <a href=\"https://arxiv.org/abs/2605.07237\" rel=\"nofollow\">https://arxiv.org/abs/2605.07237</a><br>Code and models will be released soon.</p>\n","updatedAt":"2026-05-13T01:40:43.339Z","author":{"_id":"6306df0ed37ce67e0e53e3f1","avatarUrl":"/avatars/9e20d43941169e6207f85f9bcc25a0de.svg","fullname":"Hyeon Hwang","name":"Hyeoni","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8235577344894409},"editors":["Hyeoni"],"editorAvatarUrls":["/avatars/9e20d43941169e6207f85f9bcc25a0de.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.07237","authors":[{"_id":"6a0288c2b823258e76123371","user":{"_id":"6306df0ed37ce67e0e53e3f1","avatarUrl":"/avatars/9e20d43941169e6207f85f9bcc25a0de.svg","isPro":false,"fullname":"Hyeon Hwang","user":"Hyeoni","type":"user","name":"Hyeoni"},"name":"Hyeon Hwang","status":"claimed_verified","statusLastChangedAt":"2026-05-12T08:02:56.288Z","hidden":false},{"_id":"6a0288c2b823258e76123372","name":"Jiwoo Lee","hidden":false},{"_id":"6a0288c2b823258e76123373","name":"Jaewoo Kang","hidden":false}],"publishedAt":"2026-05-11T00:00:00.000Z","submittedOnDailyAt":"2026-05-13T00:00:00.000Z","title":"Teaching Language Models to Think in Code","submittedOnDailyBy":{"_id":"6306df0ed37ce67e0e53e3f1","avatarUrl":"/avatars/9e20d43941169e6207f85f9bcc25a0de.svg","isPro":false,"fullname":"Hyeon Hwang","user":"Hyeoni","type":"user","name":"Hyeoni"},"summary":"Tool-integrated reasoning (TIR) has emerged as a dominant paradigm for mathematical problem solving in language models, combining natural language (NL) reasoning with code execution. However, this interleaved setup has three key limitations: code often acts as a post-hoc verifier, intermediate NL computations are error-prone, and NL and code play overlapping rather than clearly distinct roles. We propose ThinC (Thinking in Code), a framework in which code itself serves as the reasoner rather than as a tool invoked by NL. A ThinC trajectory begins with a brief NL planning step, after which all reasoning unfolds through code blocks connected only by their execution outputs. We distill 12.2k code-centric trajectories from a teacher model and train ThinC-1.7B and ThinC-4B with supervised fine-tuning followed by reinforcement learning. ThinC-4B consistently outperforms every TIR baseline on five competition-level math benchmarks and even surpasses the much larger Qwen3-235B-A22B-Thinking. Further analysis shows that ThinC reasons through code: 99.2% of its final answers are grounded in interpreter output, and the model recovers reliably from code execution failures without intermediate NL reasoning. Our code and models will be released soon.","upvotes":17,"discussionId":"6a0288c3b823258e76123374","ai_summary":"ThinC framework enables mathematical problem solving where code serves as the primary reasoning mechanism instead of a verification tool, demonstrating superior performance on math benchmarks.","ai_keywords":["tool-integrated reasoning","language models","code execution","supervised fine-tuning","reinforcement learning","mathematical problem solving","code-centric trajectories","interpreter output"],"organization":{"_id":"6621bc39e774284ec1742ab8","name":"KoreaUniversity","fullname":"Korea University","avatar":"https://www.gravatar.com/avatar/cc5d7875c605f40a27c72e42e6aa2857?d=retro&size=100"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"60f8435644e75317cc02ed51","avatarUrl":"/avatars/68b7fc077fe2bda6607b1c470add8140.svg","isPro":false,"fullname":"Jungwoo Park","user":"affjljoo3581","type":"user"},{"_id":"67348f009551fdc242064ef4","avatarUrl":"/avatars/38023d6d3c2ea12434ed55aca7ca1c3e.svg","isPro":false,"fullname":"Jueon Park","user":"bioai96","type":"user"},{"_id":"65c485309f6c4689fbbe890c","avatarUrl":"/avatars/c872fbfb4e9e399f17bf96a98aef994d.svg","isPro":false,"fullname":"Jiwoo lee","user":"Jamie-Jiwoo-98","type":"user"},{"_id":"6306df0ed37ce67e0e53e3f1","avatarUrl":"/avatars/9e20d43941169e6207f85f9bcc25a0de.svg","isPro":false,"fullname":"Hyeon Hwang","user":"Hyeoni","type":"user"},{"_id":"68121036912c2103dd3ba6cd","avatarUrl":"/avatars/2ac7eacb486bf1993f7465299282cd3c.svg","isPro":false,"fullname":"Taeyun Roh","user":"txxnrd","type":"user"},{"_id":"656a22fa801ed9952f432e69","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/656a22fa801ed9952f432e69/w8T3rqAvxCu8xlh049jhh.webp","isPro":false,"fullname":"Kyochul Jang","user":"OfficerChul","type":"user"},{"_id":"664b233490135abe9bbbb72a","avatarUrl":"/avatars/d2175e2e2956ae1b61093cd424bc903e.svg","isPro":false,"fullname":"Hoonick Lee","user":"hoonick","type":"user"},{"_id":"631c386bc73939ffc0716a37","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1662793811119-noauth.jpeg","isPro":false,"fullname":"SeongWan Kim","user":"idgmatrix","type":"user"},{"_id":"63f1de31f4e30ffd2bcd626b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63f1de31f4e30ffd2bcd626b/aPHgcUj0NN68_fIEym3HE.jpeg","isPro":false,"fullname":"Suhyeong Park","user":"Codingchild","type":"user"},{"_id":"689431c991b338f15ee89b51","avatarUrl":"/avatars/f3d19eb7dc9dc086c4758e3fe778121d.svg","isPro":false,"fullname":"sanghoon lee","user":"sanghoonAIGEN","type":"user"},{"_id":"6686d82da7812dde115c0fad","avatarUrl":"/avatars/38d64823c807e41abb5dce3d8b38285b.svg","isPro":false,"fullname":"Hyunjin Seo","user":"bellaseo72","type":"user"},{"_id":"66b08b4073063b0cceec914f","avatarUrl":"/avatars/02b1446b9af2ef1504d78f08272b630c.svg","isPro":false,"fullname":"Minbyul Jeong","user":"Minstar","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6621bc39e774284ec1742ab8","name":"KoreaUniversity","fullname":"Korea University","avatar":"https://www.gravatar.com/avatar/cc5d7875c605f40a27c72e42e6aa2857?d=retro&size=100"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.07237.md"}">

Papers

arxiv:2605.07237

Teaching Language Models to Think in Code

Published on May 11

· Submitted by

Hyeon Hwang on May 13

Korea University

Upvote

Authors:

Hyeon Hwang ,

Abstract

ThinC framework enables mathematical problem solving where code serves as the primary reasoning mechanism instead of a verification tool, demonstrating superior performance on math benchmarks.

AI-generated summary

Tool-integrated reasoning (TIR) has emerged as a dominant paradigm for mathematical problem solving in language models, combining natural language (NL) reasoning with code execution. However, this interleaved setup has three key limitations: code often acts as a post-hoc verifier, intermediate NL computations are error-prone, and NL and code play overlapping rather than clearly distinct roles. We propose ThinC (Thinking in Code), a framework in which code itself serves as the reasoner rather than as a tool invoked by NL. A ThinC trajectory begins with a brief NL planning step, after which all reasoning unfolds through code blocks connected only by their execution outputs. We distill 12.2k code-centric trajectories from a teacher model and train ThinC-1.7B and ThinC-4B with supervised fine-tuning followed by reinforcement learning. ThinC-4B consistently outperforms every TIR baseline on five competition-level math benchmarks and even surpasses the much larger Qwen3-235B-A22B-Thinking. Further analysis shows that ThinC reasons through code: 99.2% of its final answers are grounded in interpreter output, and the model recovers reliably from code execution failures without intermediate NL reasoning. Our code and models will be released soon.