{"paper":{"title":"Do-Undo Bench: Reversibility for Action Understanding in Image Generation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Requiring image generators to apply and then undo real actions tests genuine cause-and-effect understanding.","cross_cats":["cs.LG"],"primary_cat":"cs.CV","authors_text":"Apratim Bhattacharyya, Fatih Porikli, Hoang Le, Munawar Hayat, Rajeev Yasarla, Shreya Kadambi, Shweta Mahajan","submitted_at":"2025-12-15T18:03:42Z","abstract_excerpt":"We introduce the Do-Undo task and benchmark to address a critical gap in vision-language models: understanding and generating plausible scene transformations driven by real-world actions. Unlike prior work that relies on prompt-based image generation and editing to perform action-conditioned image manipulation, our training hypothesis requires models to simulate the outcome of a real-world action and then reverse it to the original state. This forward-reverse requirement tests genuine cause-and-effect understanding rather than stylistic or semantic edits. We curate a high-quality benchmark of "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"our training hypothesis requires models to simulate the outcome of a real-world action and then reverse it to the original state. This forward-reverse requirement tests genuine cause-and-effect understanding rather than stylistic or semantic edits.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the curated high-quality benchmark of reversible actions from real-world scenarios actually isolates genuine cause-and-effect understanding rather than other visual or linguistic cues.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Do-Undo Bench is a new evaluation task and dataset that forces models to simulate forward action effects and then undo them to measure genuine action understanding in image generation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Requiring image generators to apply and then undo real actions tests genuine cause-and-effect understanding.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"480eb27bf12949776cf156439f290f5f8e923735597504dedb52e3f3c03f13f6"},"source":{"id":"2512.13609","kind":"arxiv","version":2},"verdict":{"id":"85ac5aef-347b-45f3-a495-4dd4ef9829eb","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T22:07:13.024071Z","strongest_claim":"our training hypothesis requires models to simulate the outcome of a real-world action and then reverse it to the original state. This forward-reverse requirement tests genuine cause-and-effect understanding rather than stylistic or semantic edits.","one_line_summary":"Do-Undo Bench is a new evaluation task and dataset that forces models to simulate forward action effects and then undo them to measure genuine action understanding in image generation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the curated high-quality benchmark of reversible actions from real-world scenarios actually isolates genuine cause-and-effect understanding rather than other visual or linguistic cues.","pith_extraction_headline":"Requiring image generators to apply and then undo real actions tests genuine cause-and-effect understanding."},"references":{"count":36,"sample":[{"doi":"","year":2024,"title":"Unibench: Visual reasoning requires rethinking vision- language beyond scaling.Advances in Neural Information Processing Systems, 37:82411–82437, 2024","work_id":"2e38bbe8-063d-44d3-93a3-0ff530253581","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Cosmos-Reason1: From Physical Common Sense To Embodied Reasoning","work_id":"09215448-e68a-4168-89b7-9d9a83a0e51f","ref_index":2,"cited_arxiv_id":"2503.15558","is_internal_anchor":true},{"doi":"","year":2018,"title":"Long-term image boundary prediction","work_id":"c8bd781e-7050-49eb-b5f4-0ef9c394bda3","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Visual jenga: Discovering object dependencies via counterfactual inpainting","work_id":"0cbf5c2c-76ca-4283-bd7e-ab9252f14a18","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"In- structpix2pix: Learning to follow image editing instructions","work_id":"02ea0df2-52bb-4d47-b481-bdaec89376a0","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":36,"snapshot_sha256":"3e3aa11e2e71f2080b66d45639eff485d2a00ca2e04c2251ce08c679381666fd","internal_anchors":5},"formal_canon":{"evidence_count":2,"snapshot_sha256":"ad0664dae266f8553b3dd2d1f167e20b6bd8e7e3311768e922c050a849d2cfd9"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}