https://ursa-math.github.io/

\n","updatedAt":"2025-01-09T07:18:53.452Z","author":{"_id":"6548956fe49bd8d58e8adf0e","avatarUrl":"/avatars/24c54b14c253c87d4b7438193f16ce28.svg","fullname":"Ruilin","name":"Antimage01","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.40194031596183777},"editors":["Antimage01"],"editorAvatarUrls":["/avatars/24c54b14c253c87d4b7438193f16ce28.svg"],"reactions":[],"isReport":false}},{"id":"6780791ec377ac7f3f9c669c","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":317,"isUserFollowing":false},"createdAt":"2025-01-10T01:34:22.000Z","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* [AtomThink: A Slow Thinking Framework for Multimodal Mathematical Reasoning](https://huggingface.co/papers/2411.11930) (2024)\n* [Beyond Examples: High-level Automated Reasoning Paradigm in In-Context Learning via MCTS](https://huggingface.co/papers/2411.18478) (2024)\n* [LLaVA-CoT: Let Vision Language Models Reason Step-by-Step](https://huggingface.co/papers/2411.10440) (2024)\n* [System-2 Mathematical Reasoning via Enriched Instruction Tuning](https://huggingface.co/papers/2412.16964) (2024)\n* [SRA-MCTS: Self-driven Reasoning Augmentation with Monte Carlo Tree Search for Code Generation](https://huggingface.co/papers/2411.11053) (2024)\n* [Mars-PO: Multi-Agent Reasoning System Preference Optimization](https://huggingface.co/papers/2411.19039) (2024)\n* [BoostStep: Boosting mathematical capability of Large Language Models via improved single-step reasoning](https://huggingface.co/papers/2501.03226) (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":"2025-01-10T01:34:22.136Z","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":317,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.7108842730522156},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2501.04686","authors":[{"_id":"677f62ca407edfda4408dd07","user":{"_id":"6548956fe49bd8d58e8adf0e","avatarUrl":"/avatars/24c54b14c253c87d4b7438193f16ce28.svg","isPro":false,"fullname":"Ruilin","user":"Antimage01","type":"user"},"name":"Ruilin Luo","status":"claimed_verified","statusLastChangedAt":"2025-01-09T10:06:31.575Z","hidden":false},{"_id":"677f62ca407edfda4408dd08","user":{"_id":"64ae0f825d48838462023c9b","avatarUrl":"/avatars/d3f348c1428376aea490339e94d4c239.svg","isPro":false,"fullname":"Zheng Zhuofan","user":"fun6668","type":"user"},"name":"Zhuofan Zheng","status":"admin_assigned","statusLastChangedAt":"2025-01-09T11:09:45.251Z","hidden":false},{"_id":"677f62ca407edfda4408dd09","name":"Yifan Wang","hidden":false},{"_id":"677f62ca407edfda4408dd0a","user":{"_id":"66b1eb17652012ddfb59e41e","avatarUrl":"/avatars/b1c8ce44e7a7d789a3eb2e29e2ddfebe.svg","isPro":false,"fullname":"Yiyao Yu","user":"yiyaoyu","type":"user"},"name":"Yiyao Yu","status":"admin_assigned","statusLastChangedAt":"2025-01-09T20:19:06.927Z","hidden":false},{"_id":"677f62ca407edfda4408dd0b","name":"Xinzhe Ni","hidden":false},{"_id":"677f62ca407edfda4408dd0c","user":{"_id":"64292eb375bcc24c5e52c011","avatarUrl":"/avatars/c8cb03ca35ca12d8831be5f4e8547d54.svg","isPro":false,"fullname":"czl","user":"Lin1557","type":"user"},"name":"Zicheng Lin","status":"claimed_verified","statusLastChangedAt":"2025-01-13T09:09:56.183Z","hidden":false},{"_id":"677f62ca407edfda4408dd0d","name":"Jin Zeng","hidden":false},{"_id":"677f62ca407edfda4408dd0e","user":{"_id":"64ca1fe838837b12d5e529b7","avatarUrl":"/avatars/44a3ad9e59318784ac531993b5f69f6b.svg","isPro":false,"fullname":"Yujiu Yang","user":"Thu-redrobot","type":"user"},"name":"Yujiu Yang","status":"admin_assigned","statusLastChangedAt":"2025-01-09T20:02:25.883Z","hidden":false}],"publishedAt":"2025-01-08T18:49:41.000Z","submittedOnDailyAt":"2025-01-09T03:59:54.445Z","title":"URSA: Understanding and Verifying Chain-of-thought Reasoning in\n Multimodal Mathematics","submittedOnDailyBy":{"_id":"64292eb375bcc24c5e52c011","avatarUrl":"/avatars/c8cb03ca35ca12d8831be5f4e8547d54.svg","isPro":false,"fullname":"czl","user":"Lin1557","type":"user"},"summary":"Chain-of-thought (CoT) reasoning has been widely applied in the mathematical\nreasoning of Large Language Models (LLMs). Recently, the introduction of\nderivative process supervision on CoT trajectories has sparked discussions on\nenhancing scaling capabilities during test time, thereby boosting the potential\nof these models. However, in multimodal mathematical reasoning, the scarcity of\nhigh-quality CoT training data has hindered existing models from achieving\nhigh-precision CoT reasoning and has limited the realization of reasoning\npotential during test time. In this work, we propose a three-module synthesis\nstrategy that integrates CoT distillation, trajectory-format rewriting, and\nformat unification. It results in a high-quality CoT reasoning instruction\nfine-tuning dataset in multimodal mathematics, MMathCoT-1M. We comprehensively\nvalidate the state-of-the-art (SOTA) performance of the trained URSA-7B model\non multiple multimodal mathematical benchmarks. For test-time scaling, we\nintroduce a data synthesis strategy that automatically generates process\nannotation datasets, known as DualMath-1.1M, focusing on both interpretation\nand logic. By further training URSA-7B on DualMath-1.1M, we transition from CoT\nreasoning capabilities to robust supervision abilities. The trained URSA-RM-7B\nacts as a verifier, effectively enhancing the performance of URSA-7B at test\ntime. URSA-RM-7B also demonstrates excellent out-of-distribution (OOD)\nverifying capabilities, showcasing its generalization. Model weights, training\ndata and code will be open-sourced.","upvotes":53,"discussionId":"677f62cb407edfda4408dd5c","projectPage":"https://ursa-math.github.io/","githubRepo":"https://github.com/URSA-MATH/URSA-MATH","githubRepoAddedBy":"user","ai_summary":"A three-module synthesis strategy and data generation approach improve Chain-of-Thought reasoning in multimodal mathematical tasks, boosting model performance and out-of-distribution verification capabilities.","ai_keywords":["Chain-of-Thought","reasoning","Large Language Models","derivative process supervision","multimodal mathematical reasoning","CoT distillation","trajectory-format rewriting","format unification","multimodal mathematics","fine-tuning dataset","state-of-the-art","benchmarks","process annotation datasets","robust supervision","verification capabilities","out-of-distribution","generalization"],"githubStars":129},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"64292eb375bcc24c5e52c011","avatarUrl":"/avatars/c8cb03ca35ca12d8831be5f4e8547d54.svg","isPro":false,"fullname":"czl","user":"Lin1557","type":"user"},{"_id":"650be23ec4e52db6a4db63ef","avatarUrl":"/avatars/03af548029b38bee49ec295fefe74f9a.svg","isPro":false,"fullname":"Haoling Li","user":"Ringo1110","type":"user"},{"_id":"66f6627e08be8ab9ab0833f2","avatarUrl":"/avatars/871a0862c22e0e4f8829144953f81d85.svg","isPro":false,"fullname":"mi","user":"mxy123","type":"user"},{"_id":"66b1eb17652012ddfb59e41e","avatarUrl":"/avatars/b1c8ce44e7a7d789a3eb2e29e2ddfebe.svg","isPro":false,"fullname":"Yiyao Yu","user":"yiyaoyu","type":"user"},{"_id":"6535e56ef551a245bbf7e4ed","avatarUrl":"/avatars/540bde030bc61362f4f2ad507a063596.svg","isPro":false,"fullname":"zhr","user":"finduzzz","type":"user"},{"_id":"643d49cf482011f5f2be28fb","avatarUrl":"/avatars/1b2e0213bd53a8260889cded82de47d8.svg","isPro":false,"fullname":"jin","user":"zengjin","type":"user"},{"_id":"64b75db417570fdff9b43fd9","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64b75db417570fdff9b43fd9/-NcNL_QzuiAAvsBIQG3xo.png","isPro":false,"fullname":"Cheng YANG","user":"Cheng-YANG","type":"user"},{"_id":"677f6e9ce3ea3b483ac870f8","avatarUrl":"/avatars/3ad2f744f706e4a29f0cf4b0f7cae662.svg","isPro":false,"fullname":"chan","user":"Lemonade11","type":"user"},{"_id":"666fea4dcb82a3b014edbeeb","avatarUrl":"/avatars/2d17bf2c4cabd8e968f159cf983f490e.svg","isPro":false,"fullname":"wong","user":"efanwong","type":"user"},{"_id":"659bb6ab6d20ab21b0bf2456","avatarUrl":"/avatars/b97bcfafcf961797faa23d5ca55b9261.svg","isPro":false,"fullname":"Xinzhe","user":"Thmars","type":"user"},{"_id":"6683a05e74fb1736a4b7c934","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6683a05e74fb1736a4b7c934/eiz6qlqIUjAWGy5zfg8Cs.jpeg","isPro":false,"fullname":"QRQ","user":"RichardQRQ","type":"user"},{"_id":"620783f24e28382272337ba4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620783f24e28382272337ba4/zkUveQPNiDfYjgGhuFErj.jpeg","isPro":false,"fullname":"GuoLiangTang","user":"Tommy930","type":"user"}],"acceptLanguages":["*"],"dailyPaperRank":0}">

Papers

arxiv:2501.04686

URSA: Understanding and Verifying Chain-of-thought Reasoning in Multimodal Mathematics

Published on Jan 8, 2025

· Submitted by

czl on Jan 9, 2025

Upvote

Authors:

Ruilin Luo ,

Zhuofan Zheng ,

Yiyao Yu ,

Zicheng Lin ,

Yujiu Yang

Abstract

A three-module synthesis strategy and data generation approach improve Chain-of-Thought reasoning in multimodal mathematical tasks, boosting model performance and out-of-distribution verification capabilities.

AI-generated summary

Chain-of-thought (CoT) reasoning has been widely applied in the mathematical reasoning of Large Language Models (LLMs). Recently, the introduction of derivative process supervision on CoT trajectories has sparked discussions on enhancing scaling capabilities during test time, thereby boosting the potential of these models. However, in multimodal mathematical reasoning, the scarcity of high-quality CoT training data has hindered existing models from achieving high-precision CoT reasoning and has limited the realization of reasoning potential during test time. In this work, we propose a three-module synthesis strategy that integrates CoT distillation, trajectory-format rewriting, and format unification. It results in a high-quality CoT reasoning instruction fine-tuning dataset in multimodal mathematics, MMathCoT-1M. We comprehensively validate the state-of-the-art (SOTA) performance of the trained URSA-7B model on multiple multimodal mathematical benchmarks. For test-time scaling, we introduce a data synthesis strategy that automatically generates process annotation datasets, known as DualMath-1.1M, focusing on both interpretation and logic. By further training URSA-7B on DualMath-1.1M, we transition from CoT reasoning capabilities to robust supervision abilities. The trained URSA-RM-7B acts as a verifier, effectively enhancing the performance of URSA-7B at test time. URSA-RM-7B also demonstrates excellent out-of-distribution (OOD) verifying capabilities, showcasing its generalization. Model weights, training data and code will be open-sourced.

View arXiv page View PDF Project page GitHub 129 Add to collection

Community

Lin1557

Paper author Paper submitter Jan 9, 2025

TL;DR: Our work focuses on multimodal mathematical reasoning capabilities. We contribute the MMathCoT-1M and DualMath-1.1M datasets through a three-module high-quality CoT data synthesis and dual-view process label automation. The URSA-7B model, fine-tuned on MMathCoT-1M, achieves SOTA performance among models of the same size on multiple multimodal mathematics benchmarks. Furthermore, the URSA-RM-7B, trained on DualMath-1.1M, is the first contribution of a small-sized reward model in the domain of multimodal mathematics, with its test-time scaling effectiveness being greatly validated.