{"paper":{"title":"Video models are zero-shot learners and reasoners","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Generative video models like Veo 3 perform zero-shot object segmentation, edge detection, physics understanding, affordance recognition, tool simulation, and early visual reasoning such as maze and symmetry solving.","cross_cats":["cs.AI","cs.CV","cs.RO"],"primary_cat":"cs.LG","authors_text":"Been Kim, Kevin Swersky, Nick Matarese, Paul Vicol, Priyank Jaini, Robert Geirhos, Shixiang Shane Gu, Thadd\\\"aus Wiedemer, Yuxuan Li","submitted_at":"2025-09-24T17:17:27Z","abstract_excerpt":"The remarkable zero-shot capabilities of Large Language Models (LLMs) have propelled natural language processing from task-specific models to unified, generalist foundation models. This transformation emerged from simple primitives: large, generative models trained on web-scale data. Curiously, the same primitives apply to today's generative video models. Could video models be on a trajectory towards general-purpose vision understanding, much like LLMs developed general-purpose language understanding? We demonstrate that Veo 3 can solve a broad variety of tasks it wasn't explicitly trained for"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Veo 3 can solve a broad variety of tasks it wasn't explicitly trained for: segmenting objects, detecting edges, editing images, understanding physical properties, recognizing object affordances, simulating tool use, and more. These abilities enable early forms of visual reasoning like maze and symmetry solving.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the demonstrated capabilities are genuinely zero-shot and not the result of implicit task information in the prompts, data contamination, or post-hoc selection of successful examples.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Generative video models exhibit emergent zero-shot capabilities across perception, manipulation, and basic reasoning tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Generative video models like Veo 3 perform zero-shot object segmentation, edge detection, physics understanding, affordance recognition, tool simulation, and early visual reasoning such as maze and symmetry solving.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ff0c595c84bd51c5e2400df16dae0c52e3aeab41d2f12f05e722417cba55e8d9"},"source":{"id":"2509.20328","kind":"arxiv","version":2},"verdict":{"id":"5e2c70ea-d0de-416b-a455-834dd2668077","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T02:11:09.410472Z","strongest_claim":"Veo 3 can solve a broad variety of tasks it wasn't explicitly trained for: segmenting objects, detecting edges, editing images, understanding physical properties, recognizing object affordances, simulating tool use, and more. These abilities enable early forms of visual reasoning like maze and symmetry solving.","one_line_summary":"Generative video models exhibit emergent zero-shot capabilities across perception, manipulation, and basic reasoning tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the demonstrated capabilities are genuinely zero-shot and not the result of implicit task information in the prompts, data contamination, or post-hoc selection of successful examples.","pith_extraction_headline":"Generative video models like Veo 3 perform zero-shot object segmentation, edge detection, physics understanding, affordance recognition, tool simulation, and early visual reasoning such as maze and symmetry solving."},"references":{"count":98,"sample":[{"doi":"","year":2024,"title":"A Survey on Large Language Models for Code Generation","work_id":"7d829146-3160-44c7-9ce4-4201ed9b654c","ref_index":1,"cited_arxiv_id":"2406.00515","is_internal_anchor":true},{"doi":"","year":2025,"title":"Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities","work_id":"008df105-2fdd-45d8-857a-8e35868aecb6","ref_index":2,"cited_arxiv_id":"2507.06261","is_internal_anchor":true},{"doi":"","year":2024,"title":"Weaver: Foundation models for creative writing","work_id":"1d3d18bc-fb96-451d-b2e5-aec2c15eb603","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Multilingual machine translation with large language models: Empirical results and analysis","work_id":"8aedba3a-0fa4-4db0-94b7-ed63580404df","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery","work_id":"56b6b58d-e73a-4317-896e-36ac5f84e957","ref_index":5,"cited_arxiv_id":"2408.06292","is_internal_anchor":true}],"resolved_work":98,"snapshot_sha256":"f08f5b262f8e19d6e3a421f4ce2444e499a2bf0eed658cf2405abffc07e7ee8f","internal_anchors":15},"formal_canon":{"evidence_count":1,"snapshot_sha256":"72b87a91f7f0f708c7eea993a55ddf90800983ff1e85b7072e7aead1f82d25f4"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}