Vision-language models can produce fluent answers that are visually wrong: a single unsupported object, chart value, or intermediate inference can invalidate an otherwise plausible response. We argue that this is a credit-assignment failure in multimodal post-training. Scalar outcome rewards say whether an answer is acceptable, but not which visual facts are grounded, which reasoning steps are valid, or which instruction constraints are missed. We introduce Visual Rubrics-Based Reinforcement Learning, which decomposes reference responses into atomic propositions and scores generated answers along Visual Faithfulness (VF), Reasoning Consistency (RC), and Instruction Following (IF). The resulting rubric items provide structured partial credit and localize rubric credit when supporting evidence spans are available. We first obtain an SFT checkpoint by fine-tuning Qwen3-VL-8B-Instruct on the public OpenMMReasoner-SFT-874K corpus, adapting OpenMMReasoner’s cold-start data recipe. We construct V-Rubrics 50K, a 50,248-example training set from 17 visually grounded sources, by applying rule-based filters before deriving example difficulty from rejection-sampling scores and then annotating every example with Gemini-3-Pro under the same structured prompt and protocol. The same SFT checkpoint supplies the rejection-sampling rollouts and initializes two GRPO variants using either scalar answer credit or component-wise, prefix-localized rubric credit. Rubric-based GRPO improves over both the shared SFT baseline and answer-only GRPO, with the largest gains on knowledge-oriented and visually grounded reasoning benchmarks. The results show rubrics as a useful reward abstraction for visual post-training.</p>\n","updatedAt":"2026-08-27T03:25:13.807Z","author":{"_id":"6658d01c6f1a71ba56d6c273","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/tc4nZrMuZQLfgt5aVxtH4.jpeg","fullname":"Shulin Tian","name":"shulin16","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":5,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8681389093399048},"editors":["shulin16"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/tc4nZrMuZQLfgt5aVxtH4.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.25580","authors":[{"_id":"6a8f92092c24e8c5fab328af","name":"Shulin Tian","hidden":false},{"_id":"6a8f92092c24e8c5fab328b0","name":"Minglun Li","hidden":false},{"_id":"6a8f92092c24e8c5fab328b1","name":"Yuhao Dong","hidden":false},{"_id":"6a8f92092c24e8c5fab328b2","name":"Hao Ding","hidden":false},{"_id":"6a8f92092c24e8c5fab328b3","name":"Jiarui Yao","hidden":false},{"_id":"6a8f92092c24e8c5fab328b4","user":{"_id":"64b4a717aa03b6520839e9b8","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64b4a717aa03b6520839e9b8/Rt3ERG-6BVEA4hAwOz0_I.jpeg","isPro":false,"fullname":"Haiwen Diao","user":"Paranioar","type":"user","name":"Paranioar"},"name":"Haiwen Diao","status":"claimed_verified","statusLastChangedAt":"2026-08-27T08:45:04.901Z","hidden":false},{"_id":"6a8f92092c24e8c5fab328b5","name":"Jingkang Yang","hidden":false},{"_id":"6a8f92092c24e8c5fab328b6","name":"Hongyuan Zhu","hidden":false},{"_id":"6a8f92092c24e8c5fab328b7","name":"Ziwei Liu","hidden":false}],"publishedAt":"2026-08-26T00:00:00.000Z","submittedOnDailyAt":"2026-08-27T00:00:00.000Z","title":"V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning","submittedOnDailyBy":{"_id":"6658d01c6f1a71ba56d6c273","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/tc4nZrMuZQLfgt5aVxtH4.jpeg","isPro":false,"fullname":"Shulin Tian","user":"shulin16","type":"user","name":"shulin16"},"summary":"Vision-language models can produce fluent answers that are insufficiently grounded in the visual evidence: a single unsupported object, chart value, or intermediate inference can undermine an otherwise plausible response. We argue that this is a credit-assignment failure in multimodal post-training. Scalar outcome rewards indicate whether an answer is acceptable, but do not identify which visual facts are grounded, which reasoning steps are valid, or which instruction constraints are missed. We introduce Visual Rubrics-Based Reinforcement Learning, which decomposes reference responses into atomic propositions and scores generated answers along Visual Faithfulness (VF), Reasoning Consistency (RC), and Instruction Following (IF). The resulting rubric items provide structured partial credit and localize rubric credit when supporting evidence spans are available. We first obtain an SFT checkpoint by fine-tuning Qwen3-VL-8B-Instruct on the public OpenMMReasoner-SFT-874K corpus, adapting OpenMMReasoner's cold-start data recipe. We construct V-Rubrics 50K, a 50,248-example training set from 17 visually grounded sources, by applying rule-based filters before deriving example difficulty from rejection-sampling scores and then annotating every example with Gemini-3-Pro under the same structured prompt and protocol. We train our model based on the same SFT checkpoint using component-wise, prefix-localized rubric credit. Experiments show that our rubricbased GRPO improves over both the shared SFT baseline and answer-only GRPO, with the largest gains on knowledge-oriented and visually grounded reasoning benchmarks. The results show rubrics as a useful reward abstraction for visual post-training.","upvotes":12,"discussionId":"6a8f92092c24e8c5fab328b8","projectPage":"https://shulin16.github.io/v-rubrics/","githubRepo":"https://github.com/shulin16/v-rubrics","githubRepoAddedBy":"user","ai_summary":"Visual Rubrics-Based Reinforcement Learning improves vision-language model grounding by scoring answers on visual faithfulness, reasoning consistency, and instruction following using structured partial credit.","ai_keywords":["Visual Rubrics-Based Reinforcement Learning","Visual Faithfulness","Reasoning Consistency","Instruction Following","rubric credit","GRPO","Qwen3-VL-8B-Instruct","OpenMMReasoner","V-Rubrics 50K"],"ai_summary_model":"thinkingmachines/Inkling-Small","githubStars":4,"organization":{"_id":"6371470aafbe42caa5a76208","name":"nanyang-technological-university-singapore","fullname":"Nanyang Technological University Singapore","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/637146c5afbe42caa5a75e1b/sZyHSA1AQaAS4nrGan682.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"64b4a717aa03b6520839e9b8","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64b4a717aa03b6520839e9b8/Rt3ERG-6BVEA4hAwOz0_I.jpeg","isPro":false,"fullname":"Haiwen Diao","user":"Paranioar","type":"user"},{"_id":"6658d01c6f1a71ba56d6c273","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/tc4nZrMuZQLfgt5aVxtH4.jpeg","isPro":false,"fullname":"Shulin Tian","user":"shulin16","type":"user"},{"_id":"66f8689725464a7989b75845","avatarUrl":"/avatars/43a61a528c5779103eaf5687ba44ee14.svg","isPro":false,"fullname":"Jiarui Yao","user":"FlippyDora","type":"user"},{"_id":"64f7f5b54101c731ca84ae05","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/64f7f5b54101c731ca84ae05/13DwdxOo3tWbxKDLd44B9.jpeg","isPro":false,"fullname":"Shuai Liu","user":"Choiszt","type":"user"},{"_id":"6696b81167c22a79a15ebaef","avatarUrl":"/avatars/57ce0329c4a2c46481818bc99c1d7f17.svg","isPro":false,"fullname":"Seongheon Park","user":"sam121796","type":"user"},{"_id":"66d347eebb76fb26eedb256e","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66d347eebb76fb26eedb256e/iCPF7GkmZu--XCsWzoucl.jpeg","isPro":false,"fullname":"tianqi liu","user":"tqliu","type":"user"},{"_id":"66aa94cbd59743aa4a65646f","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66aa94cbd59743aa4a65646f/Z6BKgT_oYcfJgGHQNm9K3.png","isPro":false,"fullname":"Runmao Yao","user":"yaorunmao","type":"user"},{"_id":"652965773a416e1f2173443b","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/652965773a416e1f2173443b/y9MB8YgHzbwCXAc4EI9T3.jpeg","isPro":false,"fullname":"Yuhao Dong","user":"THUdyh","type":"user"},{"_id":"6a410559d8aa1518d8ac8d74","avatarUrl":"/avatars/9d2efbb1bf482574c3a3b87657b75c31.svg","isPro":false,"fullname":"Black","user":"arist116","type":"user"},{"_id":"62ab1ac1d48b4d8b048a3473","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1656826685333-62ab1ac1d48b4d8b048a3473.png","isPro":false,"fullname":"Ziwei Liu","user":"liuziwei7","type":"user"},{"_id":"652d06833b5997ed71ce5c46","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/652d06833b5997ed71ce5c46/O_D6bpa5mGxLA7uCjmVCG.jpeg","isPro":false,"fullname":"Zhongang Cai","user":"caizhongang","type":"user"},{"_id":"60efe7fa0d920bc7805cada5","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/60efe7fa0d920bc7805cada5/2LBrJBjSCOP5ilZIpWLHl.png","isPro":false,"fullname":"Ziqi Huang","user":"Ziqi","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6371470aafbe42caa5a76208","name":"nanyang-technological-university-singapore","fullname":"Nanyang Technological University Singapore","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/637146c5afbe42caa5a75e1b/sZyHSA1AQaAS4nrGan682.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.25580.md","query":{}}">
V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning
Abstract
Visual Rubrics-Based Reinforcement Learning improves vision-language model grounding by scoring answers on visual faithfulness, reasoning consistency, and instruction following using structured partial credit.
Vision-language models can produce fluent answers that are insufficiently grounded in the visual evidence: a single unsupported object, chart value, or intermediate inference can undermine an otherwise plausible response. We argue that this is a credit-assignment failure in multimodal post-training. Scalar outcome rewards indicate whether an answer is acceptable, but do not identify which visual facts are grounded, which reasoning steps are valid, or which instruction constraints are missed. We introduce Visual Rubrics-Based Reinforcement Learning, which decomposes reference responses into atomic propositions and scores generated answers along Visual Faithfulness (VF), Reasoning Consistency (RC), and Instruction Following (IF). The resulting rubric items provide structured partial credit and localize rubric credit when supporting evidence spans are available. We first obtain an SFT checkpoint by fine-tuning Qwen3-VL-8B-Instruct on the public OpenMMReasoner-SFT-874K corpus, adapting OpenMMReasoner's cold-start data recipe. We construct V-Rubrics 50K, a 50,248-example training set from 17 visually grounded sources, by applying rule-based filters before deriving example difficulty from rejection-sampling scores and then annotating every example with Gemini-3-Pro under the same structured prompt and protocol. We train our model based on the same SFT checkpoint using component-wise, prefix-localized rubric credit. Experiments show that our rubricbased GRPO improves over both the shared SFT baseline and answer-only GRPO, with the largest gains on knowledge-oriented and visually grounded reasoning benchmarks. The results show rubrics as a useful reward abstraction for visual post-training.
Community
Vision-language models can produce fluent answers that are visually wrong: a single unsupported object, chart value, or intermediate inference can invalidate an otherwise plausible response. We argue that this is a credit-assignment failure in multimodal post-training. Scalar outcome rewards say whether an answer is acceptable, but not which visual facts are grounded, which reasoning steps are valid, or which instruction constraints are missed. We introduce Visual Rubrics-Based Reinforcement Learning, which decomposes reference responses into atomic propositions and scores generated answers along Visual Faithfulness (VF), Reasoning Consistency (RC), and Instruction Following (IF). The resulting rubric items provide structured partial credit and localize rubric credit when supporting evidence spans are available. We first obtain an SFT checkpoint by fine-tuning Qwen3-VL-8B-Instruct on the public OpenMMReasoner-SFT-874K corpus, adapting OpenMMReasoner’s cold-start data recipe. We construct V-Rubrics 50K, a 50,248-example training set from 17 visually grounded sources, by applying rule-based filters before deriving example difficulty from rejection-sampling scores and then annotating every example with Gemini-3-Pro under the same structured prompt and protocol. The same SFT checkpoint supplies the rejection-sampling rollouts and initializes two GRPO variants using either scalar answer credit or component-wise, prefix-localized rubric credit. Rubric-based GRPO improves over both the shared SFT baseline and answer-only GRPO, with the largest gains on knowledge-oriented and visually grounded reasoning benchmarks. The results show rubrics as a useful reward abstraction for visual post-training.
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/2608.25580 in a model README.md to link it from this page.
Cite arxiv.org/abs/2608.25580 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2608.25580 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.