Cohere Labs presents Tiny Aya L2-Thinker</p>\n<p>A 3.35B multilingual model that reasons in the language of the prompt instead of defaulting to English, reaching 93%+ in-language reasoning across 60 languages while keeping accuracy strong and reasoning efficient.</p>\n<p>Weights and multilingual reasoning data in 44 languages are open on HuggingFace:</p>\n<p>Tiny Aya L2-Thinker: <a href=\"https://huggingface.co/CohereLabs/tiny-aya-l2-thinker\">https://huggingface.co/CohereLabs/tiny-aya-l2-thinker</a><br>Tiny Aya En-Thinker: <a href=\"https://huggingface.co/CohereLabs/tiny-aya-en-thinker\">https://huggingface.co/CohereLabs/tiny-aya-en-thinker</a><br>Data: <a href=\"https://huggingface.co/datasets/CohereLabs/tiny-aya-l2-thinker-multilingual-reasoning\">https://huggingface.co/datasets/CohereLabs/tiny-aya-l2-thinker-multilingual-reasoning</a></p>\n","updatedAt":"2026-09-11T17:51:14.442Z","author":{"_id":"67896868bd21dd0c757c1e65","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/-5G7C2VjMHOBLZSifP3N8.png","fullname":"Mehrnaz M","name":"Mhrnz","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7117433547973633},"editors":["Mhrnz"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/-5G7C2VjMHOBLZSifP3N8.png"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2609.10445","authors":[{"_id":"6aa3c20147a406da7901e93b","user":{"_id":"67896868bd21dd0c757c1e65","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/-5G7C2VjMHOBLZSifP3N8.png","isPro":false,"fullname":"Mehrnaz M","user":"Mhrnz","type":"user","name":"Mhrnz"},"name":"Mehrnaz Mofakhami","status":"claimed_verified","statusLastChangedAt":"2026-09-11T12:32:33.356Z","hidden":false},{"_id":"6aa3c20147a406da7901e93c","name":"Ananya Sahu","hidden":false},{"_id":"6aa3c20147a406da7901e93d","name":"Alejandro R. Salamanca","hidden":false},{"_id":"6aa3c20147a406da7901e93e","name":"Daniel D'souza","hidden":false},{"_id":"6aa3c20147a406da7901e93f","name":"Alexandre Berard","hidden":false},{"_id":"6aa3c20147a406da7901e940","name":"Thomas Euyang","hidden":false},{"_id":"6aa3c20147a406da7901e941","name":"Marzieh Fadaee","hidden":false},{"_id":"6aa3c20147a406da7901e942","name":"Julia Kreutzer","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/67896868bd21dd0c757c1e65/6JG5iNxKk3ju2BhKmKD5C.png","https://cdn-uploads.huggingface.co/production/uploads/67896868bd21dd0c757c1e65/6DVt_KsYjf8DsnnylsMAA.png"],"publishedAt":"2026-09-09T00:00:00.000Z","submittedOnDailyAt":"2026-09-11T00:00:00.000Z","title":"Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning","submittedOnDailyBy":{"_id":"67896868bd21dd0c757c1e65","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/-5G7C2VjMHOBLZSifP3N8.png","isPro":false,"fullname":"Mehrnaz M","user":"Mhrnz","type":"user","name":"Mhrnz"},"summary":"Reasoning language models have made substantial advances on a variety of complex tasks, yet their capabilities remain overwhelmingly English-centric: models primarily reason in English regardless of the language they are prompted in. This is inaccessible for non-English-speaking users, risks losing the intent of the original question, and forgoes knowledge more readily expressed in the target language. In this work, we advance L2 reasoning, the ability of a model to reason consistently in the language of the user's prompt, thus building an in-language bridge between the prompt and the answer. We approach this problem from a data-centric angle, investigating how to optimize data composition and scheduling in SFT for reasoning generalization. Building Tiny Aya L2-Thinker at 3.35B scale, we achieve an L2 reasoning rate above 93% across 60 languages on 6 benchmarks spanning math, commonsense reasoning, instruction following, open-ended generation, and cultural reasoning while keeping performance strong. We show the path to generalizing L2 reasoning to held-out languages goes through broader language coverage, readily available multilingual non-reasoning data, and a sufficient English reasoning backbone. These findings indicate that reasoning is a language-agnostic behavior that can be transferred across typologically diverse languages through careful data mixing and without requiring reasoning supervision in every target language. We release our model weights and multilingual reasoning data to support further research on accessible, in-language reasoning.","upvotes":1,"discussionId":"6aa3c20147a406da7901e943","projectPage":"https://cohere.com/research/papers/building-multilingual-bridges-2026-09-10","ai_summary":"Optimized supervised fine-tuning data composition enables reasoning models to consistently process and respond in diverse non-English languages without requiring reasoning supervision in each target language.","ai_keywords":["reasoning language models","L2 reasoning","supervised fine-tuning","SFT","reasoning generalization","multilingual reasoning","data scheduling","data composition"],"ai_summary_model":"thinkingmachines/Inkling-Small","organization":{"_id":"640ca1c93623f6a56ddab373","name":"CohereLabs","fullname":"Cohere Labs","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1678549441248-5e70f6048ce3c604d78fe133.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"67896868bd21dd0c757c1e65","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/-5G7C2VjMHOBLZSifP3N8.png","isPro":false,"fullname":"Mehrnaz M","user":"Mhrnz","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"640ca1c93623f6a56ddab373","name":"CohereLabs","fullname":"Cohere Labs","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/1678549441248-5e70f6048ce3c604d78fe133.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2609/2609.10445.md","query":{}}">
Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning
Abstract
Optimized supervised fine-tuning data composition enables reasoning models to consistently process and respond in diverse non-English languages without requiring reasoning supervision in each target language.
Reasoning language models have made substantial advances on a variety of complex tasks, yet their capabilities remain overwhelmingly English-centric: models primarily reason in English regardless of the language they are prompted in. This is inaccessible for non-English-speaking users, risks losing the intent of the original question, and forgoes knowledge more readily expressed in the target language. In this work, we advance L2 reasoning, the ability of a model to reason consistently in the language of the user's prompt, thus building an in-language bridge between the prompt and the answer. We approach this problem from a data-centric angle, investigating how to optimize data composition and scheduling in SFT for reasoning generalization. Building Tiny Aya L2-Thinker at 3.35B scale, we achieve an L2 reasoning rate above 93% across 60 languages on 6 benchmarks spanning math, commonsense reasoning, instruction following, open-ended generation, and cultural reasoning while keeping performance strong. We show the path to generalizing L2 reasoning to held-out languages goes through broader language coverage, readily available multilingual non-reasoning data, and a sufficient English reasoning backbone. These findings indicate that reasoning is a language-agnostic behavior that can be transferred across typologically diverse languages through careful data mixing and without requiring reasoning supervision in every target language. We release our model weights and multilingual reasoning data to support further research on accessible, in-language reasoning.
Community
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/2609.10445 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.