{"paper":{"title":"CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action Models","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Adding explicit visual chain-of-thought by predicting future image frames before actions improves vision-language-action model performance on complex robot tasks.","cross_cats":["cs.AI","cs.LG","cs.RO"],"primary_cat":"cs.CV","authors_text":"Ankur Handa, Chelsea Finn, Donglai Xiang, Gordon Wetzstein, Ming-Yu Liu, Moo Jin Kim, Qianli Ma, Qingqing Zhao, Song Han, Tsung-Yi Lin, Yao Lu, Yecheng Wu, Zhaoshuo Li, Zhuoyang Zhang, Zipeng Fu","submitted_at":"2025-03-27T22:23:04Z","abstract_excerpt":"Vision-language-action models (VLAs) have shown potential in leveraging pretrained vision-language models and diverse robot demonstrations for learning generalizable sensorimotor control. While this paradigm effectively utilizes large-scale data from both robotic and non-robotic sources, current VLAs primarily focus on direct input--output mappings, lacking the intermediate reasoning steps crucial for complex manipulation tasks. As a result, existing VLAs lack temporal planning or reasoning capabilities. In this paper, we introduce a method that incorporates explicit visual chain-of-thought (C"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our experimental results demonstrate that CoT-VLA achieves strong performance, outperforming the state-of-the-art VLA model by 17% in real-world manipulation tasks and 6% in simulation benchmarks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That autoregressive prediction of future image frames produces reliable visual goals that meaningfully improve downstream action generation for complex manipulation tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CoT-VLA is a 7B VLA that generates future visual frames autoregressively as planning goals before actions, outperforming prior VLAs by 17% on real-world tasks and 6% in simulation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Adding explicit visual chain-of-thought by predicting future image frames before actions improves vision-language-action model performance on complex robot tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"3b52b88d2af5ea114c21b3b378c53ee08b402920c61bbf8d8ff99ba3ab2c10d3"},"source":{"id":"2503.22020","kind":"arxiv","version":1},"verdict":{"id":"5fa2ec6f-345f-40fb-a915-12c6d595b5c6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T05:17:41.158602Z","strongest_claim":"Our experimental results demonstrate that CoT-VLA achieves strong performance, outperforming the state-of-the-art VLA model by 17% in real-world manipulation tasks and 6% in simulation benchmarks.","one_line_summary":"CoT-VLA is a 7B VLA that generates future visual frames autoregressively as planning goals before actions, outperforming prior VLAs by 17% on real-world tasks and 6% in simulation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That autoregressive prediction of future image frames produces reliable visual goals that meaningfully improve downstream action generation for complex manipulation tasks.","pith_extraction_headline":"Adding explicit visual chain-of-thought by predicting future image frames before actions improves vision-language-action model performance on complex robot tasks."},"references":{"count":85,"sample":[{"doi":"","year":2024,"title":"Gen2Act: Human Video Generation in Novel Scenarios enables Generalizable Robot Manipulation","work_id":"a3bde288-aace-40db-8067-3ae6656f9509","ref_index":1,"cited_arxiv_id":"2409.16283","is_internal_anchor":true},{"doi":"","year":2023,"title":"Zero-Shot Robotic Manipulation with Pretrained Image-Editing Diffusion Models","work_id":"954b4359-f4ed-4c73-ae5b-f75d486b6fc8","ref_index":2,"cited_arxiv_id":"2310.10639","is_internal_anchor":true},{"doi":"","year":2022,"title":"RT-1: Robotics Transformer for Real-World Control at Scale","work_id":"e11bda85-8531-46bc-a07f-d0ade3643ab1","ref_index":3,"cited_arxiv_id":"2212.06817","is_internal_anchor":true},{"doi":"","year":2021,"title":"Emerg- ing properties in self-supervised vision transformers","work_id":"d783e7b6-6d26-419d-a3e8-aae0948df097","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"GR-2: A Generative Video-Language-Action Model with Web-Scale Knowledge for Robot Manipulation","work_id":"843ab5eb-2815-4db8-b3bc-890b23fa5ffa","ref_index":5,"cited_arxiv_id":"2410.06158","is_internal_anchor":true}],"resolved_work":85,"snapshot_sha256":"ed33a648d307c2bf7df5fb3321c74515d25cf806f5865b17937f09810b96bb56","internal_anchors":28},"formal_canon":{"evidence_count":2,"snapshot_sha256":"38c0d167a3af73b21e802bd9b1a98b72fe48248e34edd20fb8a7e992da0f3826"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}