@librarian-bot\n\t

\n","updatedAt":"2024-11-16T08:14:19.722Z","author":{"_id":"62f5c24eea5bd6b1abc8e151","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1660273191881-noauth.jpeg","fullname":"Hunar Batra","name":"hunarbatra","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":4,"isUserFollowing":false}},"numEdits":0,"editors":["hunarbatra"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1660273191881-noauth.jpeg"],"reactions":[],"isReport":false}},{"id":"67391f5621972b9d6f86800a","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":"2024-11-16T22:40:22.000Z","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* [Improve Vision Language Model Chain-of-thought Reasoning](https://huggingface.co/papers/2410.16198) (2024)\n* [ReGenesis: LLMs can Grow into Reasoning Generalists via Self-Improvement](https://huggingface.co/papers/2410.02108) (2024)\n* [Distill Visual Chart Reasoning Ability from LLMs to MLLMs](https://huggingface.co/papers/2410.18798) (2024)\n* [Self-Correction is More than Refinement: A Learning Framework for Visual and Language Reasoning Tasks](https://huggingface.co/papers/2410.04055) (2024)\n* [Thought-Path Contrastive Learning via Premise-Oriented Data Augmentation for Logical Reading Comprehension](https://huggingface.co/papers/2409.14495) (2024)\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":"2024-11-16T22:40:22.203Z","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":0.778429388999939},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2411.00855","authors":[{"_id":"6731e7538f847a34213daead","user":{"_id":"63340dbbd92c5842ae71d1e9","avatarUrl":"/avatars/3a3182996bd41b526dcbfa8687d91963.svg","isPro":false,"fullname":"Kanzhi Cheng","user":"cckevinn","type":"user"},"name":"Kanzhi Cheng","status":"claimed_verified","statusLastChangedAt":"2025-01-07T08:43:05.098Z","hidden":false},{"_id":"6731e7538f847a34213daeae","name":"Yantao Li","hidden":false},{"_id":"6731e7538f847a34213daeaf","user":{"_id":"656d73ed0bbc114fe6449704","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/656d73ed0bbc114fe6449704/gpteBU9GmKSHRVkRBUHld.png","isPro":false,"fullname":"Symbol-LLM","user":"Symbol-LLM","type":"user"},"name":"Fangzhi Xu","status":"extracted_confirmed","statusLastChangedAt":"2025-01-03T06:58:46.108Z","hidden":false},{"_id":"6731e7538f847a34213daeb0","name":"Jianbing Zhang","hidden":false},{"_id":"6731e7538f847a34213daeb1","name":"Hao Zhou","hidden":false},{"_id":"6731e7538f847a34213daeb2","name":"Yang Liu","hidden":false}],"publishedAt":"2024-10-30T14:45:00.000Z","title":"Vision-Language Models Can Self-Improve Reasoning via Reflection","summary":"Chain-of-thought (CoT) has proven to improve the reasoning capability of\nlarge language models (LLMs). However, due to the complexity of multimodal\nscenarios and the difficulty in collecting high-quality CoT data, CoT reasoning\nin multimodal LLMs has been largely overlooked. To this end, we propose a\nsimple yet effective self-training framework, R3V, which iteratively enhances\nthe model's Vision-language Reasoning by Reflecting on CoT Rationales. Our\nframework consists of two interleaved parts: (1) iteratively bootstrapping\npositive and negative solutions for reasoning datasets, and (2) reflection on\nrationale for learning from mistakes. Specifically, we introduce the\nself-refine and self-select losses, enabling the model to refine flawed\nrationale and derive the correct answer by comparing rationale candidates.\nExperiments on a wide range of vision-language tasks show that R3V consistently\nimproves multimodal LLM reasoning, achieving a relative improvement of 23 to 60\npercent over GPT-distilled baselines. Additionally, our approach supports\nself-reflection on generated solutions, further boosting performance through\ntest-time computation.","upvotes":5,"discussionId":"6731e7548f847a34213daf07","ai_summary":"R3V, a self-training framework, improves multimodal LLMs' reasoning capabilities by refining and selecting rationales, particularly effective in vision-language tasks.","ai_keywords":["Vision-language Reasoning","CoT Rationales","self-refine","self-select losses","self-reflection"]},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"656d73ed0bbc114fe6449704","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/656d73ed0bbc114fe6449704/gpteBU9GmKSHRVkRBUHld.png","isPro":false,"fullname":"Symbol-LLM","user":"Symbol-LLM","type":"user"},{"_id":"67164616c6f6570d4f5da4f1","avatarUrl":"/avatars/368398ba3cd879a78f7171bf97d7f9d9.svg","isPro":false,"fullname":"leelin","user":"leelin-us","type":"user"},{"_id":"63340dbbd92c5842ae71d1e9","avatarUrl":"/avatars/3a3182996bd41b526dcbfa8687d91963.svg","isPro":false,"fullname":"Kanzhi Cheng","user":"cckevinn","type":"user"},{"_id":"6640bbd0220cfa8cbfdce080","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6640bbd0220cfa8cbfdce080/wiAHUu5ewawyipNs0YFBR.png","isPro":true,"fullname":"John Smith","user":"John6666","type":"user"},{"_id":"64bbe9b236eb058cd9d6a5b9","avatarUrl":"/avatars/c7c01a3fa8809e73800392679abff6d5.svg","isPro":false,"fullname":"Kai Zuberbühler","user":"kaizuberbuehler","type":"user"}],"acceptLanguages":["*"]}">

Papers

arxiv:2411.00855

Vision-Language Models Can Self-Improve Reasoning via Reflection

Published on Oct 30, 2024

Upvote

Authors:

Kanzhi Cheng ,

Fangzhi Xu ,

Abstract

R3V, a self-training framework, improves multimodal LLMs' reasoning capabilities by refining and selecting rationales, particularly effective in vision-language tasks.

AI-generated summary

Chain-of-thought (CoT) has proven to improve the reasoning capability of large language models (LLMs). However, due to the complexity of multimodal scenarios and the difficulty in collecting high-quality CoT data, CoT reasoning in multimodal LLMs has been largely overlooked. To this end, we propose a simple yet effective self-training framework, R3V, which iteratively enhances the model's Vision-language Reasoning by Reflecting on CoT Rationales. Our framework consists of two interleaved parts: (1) iteratively bootstrapping positive and negative solutions for reasoning datasets, and (2) reflection on rationale for learning from mistakes. Specifically, we introduce the self-refine and self-select losses, enabling the model to refine flawed rationale and derive the correct answer by comparing rationale candidates. Experiments on a wide range of vision-language tasks show that R3V consistently improves multimodal LLM reasoning, achieving a relative improvement of 23 to 60 percent over GPT-distilled baselines. Additionally, our approach supports self-reflection on generated solutions, further boosting performance through test-time computation.

View arXiv page View PDF Add to collection