{"paper":{"title":"CreFlow: Corrective Reflow for Sparse-Reward Embodied Video Diffusion RL","license":"http://creativecommons.org/licenses/by/4.0/","headline":"CreFlow uses automatically generated Linear Temporal Logic rewards plus corrective reflow to align video diffusion rollouts with embodied task rules and lift downstream success 23.8 points.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Minshuo Chen, Philip Torr, Qi Zhu, Ruochen Jiao, Simon Sinong Zhan, Sipeng Chen, Yijiang Li, Zhaoran Wang, Zhenfei Yin, Zhenyang Ni","submitted_at":"2026-05-14T02:18:58Z","abstract_excerpt":"Video generation models trained on heterogeneous data with likelihood-surrogate objectives can produce visually plausible rollouts that violate physical constraints in embodied manipulation. Although reinforcement-learning post-training offers a natural route to adapting VGMs, existing video-RL rewards often reduce each rollout to a low-level visual metric, whereas manipulation video evaluation requires logic-based verification of whether the rollout satisfies a compositional task specification. To fill this gap, we introduce a compositional constraint-based reward model for post-training embo"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"CreFlow yields reward judgments better aligned with human and simulator success labels than existing methods and improves downstream execution success by 23.8 percentage points across eight bimanual manipulation tasks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the automatically formulated LTL constraints provide faithful, localized rewards without significant manual engineering or domain-specific tuning, and that the corrective reflow loss reliably stabilizes high-dimensional video diffusion updates in practice.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CreFlow combines LTL compositional rewards with credit-aware NFT and corrective reflow losses in online RL to improve embodied video diffusion models, raising downstream task success by 23.8 percentage points on eight bimanual manipulation tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"CreFlow uses automatically generated Linear Temporal Logic rewards plus corrective reflow to align video diffusion rollouts with embodied task rules and lift downstream success 23.8 points.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"5405d3fb8d76c379c503dbd93c2b52aef0705e2d5dd43454a5346bafcef8f514"},"source":{"id":"2605.14274","kind":"arxiv","version":1},"verdict":{"id":"aef9cafb-2caf-4f66-bc8b-d1974185da61","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T02:39:41.086820Z","strongest_claim":"CreFlow yields reward judgments better aligned with human and simulator success labels than existing methods and improves downstream execution success by 23.8 percentage points across eight bimanual manipulation tasks.","one_line_summary":"CreFlow combines LTL compositional rewards with credit-aware NFT and corrective reflow losses in online RL to improve embodied video diffusion models, raising downstream task success by 23.8 percentage points on eight bimanual manipulation tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the automatically formulated LTL constraints provide faithful, localized rewards without significant manual engineering or domain-specific tuning, and that the corrective reflow loss reliably stabilizes high-dimensional video diffusion updates in practice.","pith_extraction_headline":"CreFlow uses automatically generated Linear Temporal Logic rewards plus corrective reflow to align video diffusion rollouts with embodied task rules and lift downstream success 23.8 points."},"references":{"count":45,"sample":[{"doi":"","year":2025,"title":"Cosmos World Foundation Model Platform for Physical AI","work_id":"a2dba24c-318d-476a-8b21-4289c265810c","ref_index":1,"cited_arxiv_id":"2501.03575","is_internal_anchor":true},{"doi":"","year":2024,"title":"Gen2Act: Human Video Generation in Novel Scenarios enables Generalizable Robot Manipulation","work_id":"a3bde288-aace-40db-8067-3ae6656f9509","ref_index":2,"cited_arxiv_id":"2409.16283","is_internal_anchor":true},{"doi":"","year":2025,"title":"Motus: A Unified Latent Action World Model","work_id":"d0b2d257-524d-4d67-9daf-5fb43e5e977a","ref_index":3,"cited_arxiv_id":"2512.13030","is_internal_anchor":true},{"doi":"","year":2025,"title":"$\\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization","work_id":"d1ad7304-d09a-49bc-809e-846439f6aff9","ref_index":4,"cited_arxiv_id":"2504.16054","is_internal_anchor":true},{"doi":"","year":2025,"title":"SAM 3: Segment Anything with Concepts","work_id":"4a72a006-2592-4554-aad0-a9c41a9f952d","ref_index":5,"cited_arxiv_id":"2511.16719","is_internal_anchor":true}],"resolved_work":45,"snapshot_sha256":"be34ede198d7cb0d7c3623048847cb2d8608395687bd45f9d6bafc3afd597210","internal_anchors":22},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}