https://huggingface.co/collections/KOREAson/ko-reason-1009-68e7a925c3c41417dee4db3c

\n","updatedAt":"2025-10-09T12:29:47.276Z","author":{"_id":"60d3e619b8448e1785bbda2a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/60d3e619b8448e1785bbda2a/q2re5u1HNwsCCyIMtid_I.jpeg","fullname":"GUIJIN SON","name":"amphora","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":78,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.38656100630760193},"editors":["amphora"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/60d3e619b8448e1785bbda2a/q2re5u1HNwsCCyIMtid_I.jpeg"],"reactions":[],"isReport":false}},{"id":"68e86296cddede4ab0bd94f7","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":318,"isUserFollowing":false},"createdAt":"2025-10-10T01:34:14.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"This is an automated message from the [Librarian Bot](https://huggingface.co/librarian-bots). I found the following papers similar to this paper. \n\nThe following papers were recommended by the Semantic Scholar API \n\n* [Long Chain-of-Thought Reasoning Across Languages](https://huggingface.co/papers/2508.14828) (2025)\n* [MobileLLM-R1: Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes](https://huggingface.co/papers/2509.24945) (2025)\n* [Aligning Multilingual Reasoning with Verifiable Semantics from a High-Resource Expert Model](https://huggingface.co/papers/2509.25543) (2025)\n* [Code-Switching In-Context Learning for Cross-Lingual Transfer of Large Language Models](https://huggingface.co/papers/2510.05678) (2025)\n* [mR3: Multilingual Rubric-Agnostic Reward Reasoning Models](https://huggingface.co/papers/2510.01146) (2025)\n* [SoT: Structured-of-Thought Prompting Guides Multilingual Reasoning in Large Language Models](https://huggingface.co/papers/2510.02648) (2025)\n* [Luth: Efficient French Specialization for Small Language Models and Cross-Lingual Transfer](https://huggingface.co/papers/2510.05846) (2025)\n\n\n Please give a thumbs up to this comment if you found it helpful!\n\n If you want recommendations for any Paper on Hugging Face checkout [this](https://huggingface.co/spaces/librarian-bots/recommend_similar_papers) Space\n\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: `@librarian-bot recommend`","html":"

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: \n\n@librarian-bot\n\t recommend

\n","updatedAt":"2025-10-10T01:34:14.933Z","author":{"_id":"63d3e0e8ff1384ce6c5dd17d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg","fullname":"Librarian Bot (Bot)","name":"librarian-bot","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":318,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7009985446929932},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2510.04230","authors":[{"_id":"68e7a5f744180c42eca8afa7","name":"Guijin Son","hidden":false},{"_id":"68e7a5f744180c42eca8afa8","name":"Donghun Yang","hidden":false},{"_id":"68e7a5f744180c42eca8afa9","name":"Hitesh Laxmichand Patel","hidden":false},{"_id":"68e7a5f744180c42eca8afaa","name":"Amit Agarwal","hidden":false},{"_id":"68e7a5f744180c42eca8afab","user":{"_id":"63e087b6a98d931aa90c1b9c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63e087b6a98d931aa90c1b9c/4ZnfL0U8rrj3cNhj7WTgo.jpeg","isPro":false,"fullname":"Hyunwoo Ko","user":"Cartinoe5930","type":"user"},"name":"Hyunwoo Ko","status":"claimed_verified","statusLastChangedAt":"2025-10-10T09:13:34.356Z","hidden":false},{"_id":"68e7a5f744180c42eca8afac","name":"Chanuk Lim","hidden":false},{"_id":"68e7a5f744180c42eca8afad","name":"Srikant Panda","hidden":false},{"_id":"68e7a5f744180c42eca8afae","name":"Minhyuk Kim","hidden":false},{"_id":"68e7a5f744180c42eca8afaf","name":"Nikunj Drolia","hidden":false},{"_id":"68e7a5f744180c42eca8afb0","user":{"_id":"66120647cac232c1507e13da","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66120647cac232c1507e13da/iOBrsyr2mVdVLAFF6pZt6.png","isPro":false,"fullname":"DasolChoi","user":"Dasool","type":"user"},"name":"Dasol Choi","status":"claimed_verified","statusLastChangedAt":"2025-10-10T09:13:43.834Z","hidden":false},{"_id":"68e7a5f744180c42eca8afb1","name":"Kyong-Ha Lee","hidden":false},{"_id":"68e7a5f744180c42eca8afb2","name":"Youngjae Yu","hidden":false}],"publishedAt":"2025-10-05T14:39:41.000Z","submittedOnDailyAt":"2025-10-09T10:59:47.271Z","title":"Pushing on Multilingual Reasoning Models with Language-Mixed\n Chain-of-Thought","submittedOnDailyBy":{"_id":"60d3e619b8448e1785bbda2a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/60d3e619b8448e1785bbda2a/q2re5u1HNwsCCyIMtid_I.jpeg","isPro":true,"fullname":"GUIJIN SON","user":"amphora","type":"user"},"summary":"Recent frontier models employ long chain-of-thought reasoning to explore\nsolution spaces in context and achieve stonger performance. While many works\nstudy distillation to build smaller yet capable models, most focus on English\nand little is known about language-specific reasoning. To bridge this gap, we\nfirst introduct **Language-Mixed CoT**, a reasoning schema that switches\nbetween English and a target language, using English as an anchor to excel in\nreasoning while minimizing translation artificats. As a Korean case study, we\ncurate **Yi-Sang**: 5.79M native-Korean prompts from web Q&A, exams, STEM, and\ncode; 3.7M long reasoning traces generated from Qwen3-32B; and a targeted 260k\nhigh-yield subset. We train ninve models (4B-35B) across six families (Qwen2.5,\nLlama-3.1, Gemma-3, etc). Our best model, **KO-REAson-35B**, achieves\nstate-of-the-art performance, with the highest overall average score (64.0 \\pm\n25), ranking first on 5/9 benchmarks and second on the remainder. Samller and\nmid-sized models also benefit substantially, with an average improvement of\n+18.6 points across teh evaluated nine benchmarks. Ablations show\n**Language-Mixed CoT** is more effective than monolingual CoT, also resulting\nin cross-lingual and mult-modal performance gains. We release our data-curation\npipeline, evaluation system, datasets, and models to advance research on\nlanguage-specific reasoning. Data and model collection:\nhttps://huggingface.co/KOREAson.","upvotes":27,"discussionId":"68e7a5f744180c42eca8afb3","ai_summary":"A language-mixed chain-of-thought reasoning approach improves performance in Korean-specific tasks by switching between English and Korean, achieving state-of-the-art results across various benchmarks.","ai_keywords":["chain-of-thought reasoning","Language-Mixed CoT","monolingual CoT","Yi-Sang","Qwen3-32B","KO-REAson-35B","cross-lingual","mult-modal"],"organization":{"_id":"685cc6f7126d8787fb3ed2fc","name":"KOREAson","fullname":"KO-REAson","avatar":"https://cdn-uploads.huggingface.co/production/uploads/60d3e619b8448e1785bbda2a/ER77qol4JwxeaxzA4A_i7.jpeg"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"60d3e619b8448e1785bbda2a","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/60d3e619b8448e1785bbda2a/q2re5u1HNwsCCyIMtid_I.jpeg","isPro":true,"fullname":"GUIJIN SON","user":"amphora","type":"user"},{"_id":"65019c5420cfd12a88a74078","avatarUrl":"/avatars/fdcfe7be656e9c5aa7236e3b076c7897.svg","isPro":false,"fullname":"Taegyeong Lee","user":"taegyeonglee","type":"user"},{"_id":"64f48de2e43e0d518c174301","avatarUrl":"/avatars/4e41e56c053209dcd19567f9c2a6550a.svg","isPro":false,"fullname":"HyunjunJeon","user":"de-jhj","type":"user"},{"_id":"640eff5af2d7c41a1e9ba9d4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/640eff5af2d7c41a1e9ba9d4/NcZga_7TbMOW4qhUrE7q1.jpeg","isPro":false,"fullname":"YOUNG SOOK SONG","user":"songys","type":"user"},{"_id":"647c4a2692182942d7c2e698","avatarUrl":"/avatars/bcddf5fe49aa092a2645f70812108348.svg","isPro":false,"fullname":"HWANCHANG","user":"HwanChang0106","type":"user"},{"_id":"642b0c2fecec03b4464a1d9b","avatarUrl":"/avatars/76c794c7c17503ae9a0d5336395bd132.svg","isPro":false,"fullname":"KuKu","user":"dragonkue","type":"user"},{"_id":"662c3305982b04e6a0b354b4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/662c3305982b04e6a0b354b4/QIcRfgGIefBVj9COuWEFu.jpeg","isPro":false,"fullname":"Seunghyeok Hong","user":"shongdr","type":"user"},{"_id":"63e087b6a98d931aa90c1b9c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63e087b6a98d931aa90c1b9c/4ZnfL0U8rrj3cNhj7WTgo.jpeg","isPro":false,"fullname":"Hyunwoo Ko","user":"Cartinoe5930","type":"user"},{"_id":"6152b4b9ecf3ca6ab820e325","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6152b4b9ecf3ca6ab820e325/tH8AMuUYUezkI1ssm9fVp.jpeg","isPro":true,"fullname":"Yohan Na","user":"nayohan","type":"user"},{"_id":"6415c043486c7c9a5d151583","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6415c043486c7c9a5d151583/fUdYFh6iVh57swCkBEy-y.jpeg","isPro":false,"fullname":"Jiwoo Hong","user":"JW17","type":"user"},{"_id":"664ab123f7e9d961cf598157","avatarUrl":"/avatars/c9d7e1b230f14c1158bc8ff6b917edac.svg","isPro":false,"fullname":"Minhyuk Kim","user":"torchtorchkimtorch","type":"user"},{"_id":"665ec951b246e4e792ccf89d","avatarUrl":"/avatars/34043b930ab46426de22bce3e3eecbab.svg","isPro":false,"fullname":"kyubeenhan","user":"kyubeen","type":"user"}],"acceptLanguages":["*"],"dailyPaperRank":0,"organization":{"_id":"685cc6f7126d8787fb3ed2fc","name":"KOREAson","fullname":"KO-REAson","avatar":"https://cdn-uploads.huggingface.co/production/uploads/60d3e619b8448e1785bbda2a/ER77qol4JwxeaxzA4A_i7.jpeg"}}">

Papers

arxiv:2510.04230

Pushing on Multilingual Reasoning Models with Language-Mixed Chain-of-Thought

Published on Oct 5, 2025

· Submitted by

GUIJIN SON on Oct 9, 2025

KO-REAson

Upvote

Authors:

Hyunwoo Ko ,

Dasol Choi ,

Abstract

A language-mixed chain-of-thought reasoning approach improves performance in Korean-specific tasks by switching between English and Korean, achieving state-of-the-art results across various benchmarks.

AI-generated summary

Recent frontier models employ long chain-of-thought reasoning to explore solution spaces in context and achieve stonger performance. While many works study distillation to build smaller yet capable models, most focus on English and little is known about language-specific reasoning. To bridge this gap, we first introduct **Language-Mixed CoT**, a reasoning schema that switches between English and a target language, using English as an anchor to excel in reasoning while minimizing translation artificats. As a Korean case study, we curate **Yi-Sang**: 5.79M native-Korean prompts from web Q&A, exams, STEM, and code; 3.7M long reasoning traces generated from Qwen3-32B; and a targeted 260k high-yield subset. We train ninve models (4B-35B) across six families (Qwen2.5, Llama-3.1, Gemma-3, etc). Our best model, **KO-REAson-35B**, achieves state-of-the-art performance, with the highest overall average score (64.0 \pm 25), ranking first on 5/9 benchmarks and second on the remainder. Samller and mid-sized models also benefit substantially, with an average improvement of +18.6 points across teh evaluated nine benchmarks. Ablations show **Language-Mixed CoT** is more effective than monolingual CoT, also resulting in cross-lingual and mult-modal performance gains. We release our data-curation pipeline, evaluation system, datasets, and models to advance research on language-specific reasoning. Data and model collection: https://huggingface.co/KOREAson.