We introduce TRACE, a taxonomy-guided environment containing 1,000 deterministic visual-reasoning tasks across 11 domains.</p>\n<p>Training Qwen2.5-VL-3B and Qwen2.5-VL-7B on 64,000 TRACE instances improves their macro-average performance across 24 external benchmarks by 3.51 and 4.06 percentage points, respectively.</p>\n<p>We release the full research stack, including the task generator, dataset, trained checkpoints, training and evaluation code, and run artifacts.</p>\n<p>💻 Code: <a href=\"https://github.com/maveryn/trace\" rel=\"nofollow\">https://github.com/maveryn/trace</a><br>📄 Paper: <a href=\"https://huggingface.co/papers/2607.19790\">https://huggingface.co/papers/2607.19790</a><br>🌐 Project page: <a href=\"https://maveryn.github.io/trace/\" rel=\"nofollow\">https://maveryn.github.io/trace/</a><br>🤗 Hugging Face collection: <a href=\"https://huggingface.co/collections/maveryn/trace\">https://huggingface.co/collections/maveryn/trace</a></p>\n<p>Contributions are welcome—please feel free to open an issue or submit a pull request!</p>\n","updatedAt":"2026-07-23T04:57:59.163Z","author":{"_id":"69bc258a4122dc633a350fd7","avatarUrl":"/avatars/df3e39b1f907721d6dc0f9705eed1d46.svg","fullname":"maveryn","name":"maveryn","type":"user","isPro":true,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":3,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8215938210487366},"editors":["maveryn"],"editorAvatarUrls":["/avatars/df3e39b1f907721d6dc0f9705eed1d46.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.19790","authors":[{"_id":"6a619684c3792c34f5a04165","name":"Md Tanvirul Alam","hidden":false}],"mediaUrls":["https://cdn-uploads.huggingface.co/production/uploads/69bc258a4122dc633a350fd7/Y8YZ_mWsa_7YdY5WA5BOs.png","https://cdn-uploads.huggingface.co/production/uploads/69bc258a4122dc633a350fd7/cXmfIGiZSQT2gYzR5LqAC.png"],"publishedAt":"2026-07-22T00:00:00.000Z","submittedOnDailyAt":"2026-07-23T00:00:00.000Z","title":"Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning","submittedOnDailyBy":{"_id":"69bc258a4122dc633a350fd7","avatarUrl":"/avatars/df3e39b1f907721d6dc0f9705eed1d46.svg","isPro":true,"fullname":"maveryn","user":"maveryn","type":"user","name":"maveryn"},"summary":"Reinforcement learning with verifiable rewards (RLVR) has substantially improved language-model reasoning, yet its extension to vision-language models remains constrained by the lack of training data that are simultaneously broad, exactly verifiable, and reproducible. We introduce Trace, a taxonomy-guided environment for multidomain visual reasoning. Trace factorizes task construction into a scene grammar and an executable task program, separating visual realization from answer computation. A shared semantic state determines the rendered image, prompt, typed answer, verifier state, and replayable instance trace. The resulting environment comprises 1,000 tasks over 277 scene grammars and 11 visual domains, with controlled semantic and visual variation. RLVR on 64,000 Trace instances improves the macro-average across 24 external benchmarks by 3.51 percentage points for Qwen2.5-VL-3B and 4.06 points for Qwen2.5-VL-7B, providing evidence that broad procedural training can transfer beyond the generated task distributions. Project page: https://maveryn.github.io/trace/.","upvotes":2,"discussionId":"6a619685c3792c34f5a04166","projectPage":"https://maveryn.github.io/trace/","githubRepo":"https://github.com/maveryn/trace","githubRepoAddedBy":"user","githubStars":2},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"69f9d640f8ea92de979d9b72","avatarUrl":"/avatars/d813ac5b90456cb265a5fa0b912a7c56.svg","isPro":false,"fullname":"Anonymous User","user":"hfanon13","type":"user"},{"_id":"63ac5701c21e60a3e9b58aa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63ac5701c21e60a3e9b58aa7/g6EX7diOpuA94R2ab-rZC.png","isPro":true,"fullname":"Dipankar Sarkar","user":"dipankarsarkar","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.19790.md","query":{}}">
Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning
Abstract
Reinforcement learning with verifiable rewards (RLVR) has substantially improved language-model reasoning, yet its extension to vision-language models remains constrained by the lack of training data that are simultaneously broad, exactly verifiable, and reproducible. We introduce Trace, a taxonomy-guided environment for multidomain visual reasoning. Trace factorizes task construction into a scene grammar and an executable task program, separating visual realization from answer computation. A shared semantic state determines the rendered image, prompt, typed answer, verifier state, and replayable instance trace. The resulting environment comprises 1,000 tasks over 277 scene grammars and 11 visual domains, with controlled semantic and visual variation. RLVR on 64,000 Trace instances improves the macro-average across 24 external benchmarks by 3.51 percentage points for Qwen2.5-VL-3B and 4.06 points for Qwen2.5-VL-7B, providing evidence that broad procedural training can transfer beyond the generated task distributions. Project page: https://maveryn.github.io/trace/.
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/2607.19790 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.