Hugging Face Daily Papers · September 4, 2026 · 5 min read

VeriPhy: Agentic Physical Reasoning for World Model Evaluation and Refinement

Mirrored from Hugging Face Daily Papers for archival readability. Support the source by reading on the original site.

Like Read original ↗

This is an automated message from the <a href=\"https://huggingface.co/librarian-bots\">Librarian Bot</a>. I found the following papers similar to this paper. \nThe following papers were recommended by the Semantic Scholar API \n<ul>\n<li><a href=\"https://huggingface.co/papers/2608.25622\">Plans You Can Check: Verifier-Grounded Learning of an Open-Weight Planner for Executable Video-Editing</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2608.16859\">HarnessEval-W: Agentifying the Evaluation of Visual Worlds</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2608.27334\">BTS-AgentBench: A Deterministic, Replayable Pipeline from Read-Only Telemetry Logs to Agent Benchmarks</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2608.02680\">TraceCompiler: Skill-Guided Mining and Compilation of LLM Agent Traces into Mostly Deterministic Workflows</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2607.16387\">Fantastic Adaptive Taxonomies and How to Use Them</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2607.19790\">Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning</a> (2026)</li>\n<li><a href=\"https://huggingface.co/papers/2607.06256\">Diagnosing Semantic Handoff Failures in Agent-Orchestrated Vision-Language-Action Skill Composition</a> (2026)</li>\n</ul>\n Please give a thumbs up to this comment if you found it helpful!\n If you want recommendations for any Paper on Hugging Face checkout <a href=\"https://huggingface.co/spaces/librarian-bots/recommend_similar_papers\">this</a> Space\n You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: <code>@librarian-bot recommend</code>\n","updatedAt":"2026-09-05T01:21:00.209Z","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":379,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.709033727645874},"editors":["librarian-bot"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/1674830754237-63d3e0e8ff1384ce6c5dd17d.jpeg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2609.03153","authors":[{"_id":"6a9b306eec9341424a87f5e8","user":{"_id":"63c19dbea0ffa3857eae1b88","avatarUrl":"/avatars/fc22a2ea47d3cd765635ef7a77e5f525.svg","isPro":false,"fullname":"Wenzhuo Xu","user":"littleshark2000","type":"user","name":"littleshark2000"},"name":"Wenzhuo Xu","status":"claimed_verified","statusLastChangedAt":"2026-09-05T08:45:03.857Z","hidden":false},{"_id":"6a9b306eec9341424a87f5e9","name":"Yuchen Zhu","hidden":false},{"_id":"6a9b306eec9341424a87f5ea","name":"Chongjian Ge","hidden":false},{"_id":"6a9b306eec9341424a87f5eb","name":"Xuan Shen","hidden":false},{"_id":"6a9b306eec9341424a87f5ec","name":"Jing Shi","hidden":false},{"_id":"6a9b306eec9341424a87f5ed","name":"Jason Kuen","hidden":false},{"_id":"6a9b306eec9341424a87f5ee","name":"Yongxin Chen","hidden":false},{"_id":"6a9b306eec9341424a87f5ef","name":"Molei Tao","hidden":false},{"_id":"6a9b306eec9341424a87f5f0","name":"Christopher McComb","hidden":false},{"_id":"6a9b306eec9341424a87f5f1","name":"Noelia Grande Gutiérrez","hidden":false},{"_id":"6a9b306eec9341424a87f5f2","name":"Jiuxiang Gu","hidden":false}],"publishedAt":"2026-09-02T00:00:00.000Z","submittedOnDailyAt":"2026-09-04T00:00:00.000Z","title":"VeriPhy: Agentic Physical Reasoning for World Model Evaluation and Refinement","submittedOnDailyBy":{"_id":"642467708d97ce93878e8124","avatarUrl":"/avatars/f0e464ddb4bd790f470fc0f10275fa26.svg","isPro":false,"fullname":"Jiuxiang Gu","user":"gujiuxiang","type":"user","name":"gujiuxiang"},"summary":"Visual fluency in generated video does not imply physical reliability, and a scalar quality score alone is incapable of indicating the obligation a clip violates or the moment it fails. We present VeriPhy, an auditable physical-verification system in which a text-only planner compiles the prompt into typed physical obligations and a statically validated execution plan before any frame is observed. During execution, observations gate and scope only declared calls to frozen low-level experts (e.g., segmentation and tracking, counting, eleven typed physical measurements over the resulting tracks, depth, OCR, and audio-event detection). Each action returns a provenance-carrying evidence record whose payload, when usable, is either a typed measurement or an explicitly tagged learned state. Typed resolvers and fixed composition map usable records to a three-valued state (supported, contradicted, or unknown, surfaced as plausible, implausible, or abstain) with full provenance, so that every verdict is traceable to the evidence that produced it. We anchor evaluation in a 1,500-clip corpus of human-annotated flaw records that localize real generation failures in prompt reference, space, and time. On a 149-clip core carrying 304 such records, VeriPhy accounts for 228, against 164 for a published question-decomposition evaluator given the same clips and the same claims. Recall alone does not separate it from prompting the same backbone monolithically, which reaches 222; what separates them is that each decision retains its evidence record and provenance, making the traces auditable one verdict at a time and usable as the interface through which a critic verdict could be written back into generation.","upvotes":12,"discussionId":"6a9b306fec9341424a87f5f3","projectPage":"https://veriphy-ai.github.io/","ai_summary":"VeriPhy verifies generated video by compiling prompts into typed physical obligations, executing frozen expert analyses with provenance tracking, and mapping evidence to auditable three-valued verdicts.","ai_keywords":["physical verification","typed physical obligations","frozen low-level experts","segmentation and tracking","depth estimation","OCR","audio-event detection","provenance-carrying evidence","typed measurements","three-valued logic","auditable verdicts"],"ai_summary_model":"thinkingmachines/Inkling-Small"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"642467708d97ce93878e8124","avatarUrl":"/avatars/f0e464ddb4bd790f470fc0f10275fa26.svg","isPro":false,"fullname":"Jiuxiang Gu","user":"gujiuxiang","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"},{"_id":"6a6a9b6cd3e6cc76822373e9","avatarUrl":"/avatars/7a93065f42475ca37c70031b6031f694.svg","isPro":false,"fullname":"James White","user":"wildreed","type":"user"},{"_id":"6a6c8cac42585612a497ec3b","avatarUrl":"/avatars/6a54f4673501dbd51c7b5a88a2753c20.svg","isPro":false,"fullname":"Sarah Moore","user":"Sarah-Moore","type":"user"},{"_id":"6a6aa77d342f6961a3ac5feb","avatarUrl":"/avatars/5b391a7dd212e505964da69ce2f00d24.svg","isPro":false,"fullname":"David Lopez","user":"Cobalt-Glade","type":"user"},{"_id":"6a81166e570086915074c6b2","avatarUrl":"/avatars/1cd0c1f5fcbe2da4ca59cefa510e86ed.svg","isPro":false,"fullname":"zeyu xu","user":"atlasridge","type":"user"},{"_id":"6a82621b6f3e370df0aeb3ce","avatarUrl":"/avatars/a9547a8c8818a0cfaaa06eba18bdad72.svg","isPro":false,"fullname":"Xinyi Li","user":"ZenithRidge67","type":"user"},{"_id":"6a9b5952fca080bdb0437909","avatarUrl":"/avatars/d2bab5a98161cd5581b1307021ba923d.svg","isPro":false,"fullname":"한서연","user":"GraniteEvan","type":"user"},{"_id":"620f126891e167b068fa76f8","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/620f126891e167b068fa76f8/NaPyS5lFjgZYJZrWaf0OI.jpeg","isPro":false,"fullname":"ChongjianGE","user":"RhettGee","type":"user"},{"_id":"66cd1bb86aee4c865fb8fec5","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66cd1bb86aee4c865fb8fec5/E34rNm28KXVgPkUH7U__I.jpeg","isPro":false,"fullname":"Yuchen Zhu","user":"yuchen-zhu-zyc","type":"user"},{"_id":"63c19dbea0ffa3857eae1b88","avatarUrl":"/avatars/fc22a2ea47d3cd765635ef7a77e5f525.svg","isPro":false,"fullname":"Wenzhuo Xu","user":"littleshark2000","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":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2609/2609.03153.md","query":{}}">

Papers

arxiv:2609.03153

VeriPhy: Agentic Physical Reasoning for World Model Evaluation and Refinement

Published on Sep 2

· Submitted by

Jiuxiang Gu on Sep 4

Upvote

Authors:

Wenzhuo Xu ,

Abstract

VeriPhy verifies generated video by compiling prompts into typed physical obligations, executing frozen expert analyses with provenance tracking, and mapping evidence to auditable three-valued verdicts.

Generated by thinkingmachines/Inkling-Small

Visual fluency in generated video does not imply physical reliability, and a scalar quality score alone is incapable of indicating the obligation a clip violates or the moment it fails. We present VeriPhy, an auditable physical-verification system in which a text-only planner compiles the prompt into typed physical obligations and a statically validated execution plan before any frame is observed. During execution, observations gate and scope only declared calls to frozen low-level experts (e.g., segmentation and tracking, counting, eleven typed physical measurements over the resulting tracks, depth, OCR, and audio-event detection). Each action returns a provenance-carrying evidence record whose payload, when usable, is either a typed measurement or an explicitly tagged learned state. Typed resolvers and fixed composition map usable records to a three-valued state (supported, contradicted, or unknown, surfaced as plausible, implausible, or abstain) with full provenance, so that every verdict is traceable to the evidence that produced it. We anchor evaluation in a 1,500-clip corpus of human-annotated flaw records that localize real generation failures in prompt reference, space, and time. On a 149-clip core carrying 304 such records, VeriPhy accounts for 228, against 164 for a published question-decomposition evaluator given the same clips and the same claims. Recall alone does not separate it from prompting the same backbone monolithically, which reaches 222; what separates them is that each decision retains its evidence record and provenance, making the traces auditable one verdict at a time and usable as the interface through which a critic verdict could be written back into generation.

View arXiv page View PDF Project page Add to collection