🚀 Introducing WCM: A World Critic Model for Vision-Language-Action Reinforcement Learning</p>\n<p>We are excited to introduce the World Critic Model (WCM) , a new approach that targets a critical yet often overlooked bottleneck in VLA reinforcement learning: value estimation under partial observability. While existing critic models rely on single-frame observations or weakly supervised history, robotic manipulation is inherently a POMDP, and a single snapshot simply cannot capture motion, contact progress, or future evolution. WCM is built to fix that.</p>\n","updatedAt":"2026-08-04T02:55:42.805Z","author":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","fullname":"FSY(SII)","name":"Sylvest","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8709138035774231},"editors":["Sylvest"],"editorAvatarUrls":["/avatars/c18d2c7dbefd27ec8184622b0834531f.svg"],"reactions":[],"isReport":false},"replies":[{"id":"6a7154d0d108a943948171ab","author":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","fullname":"FSY(SII)","name":"Sylvest","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false},"createdAt":"2026-08-04T02:56:16.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"🔍 The Missing Ingredient: Temporal Dynamics, Not Just History\n\nNaively stacking more frames does not solve the problem. Scalar return regression provides too sparse a signal for learning cross-temporal dynamics, so the critic ends up treating history as a larger static feature vector without understanding how the environment evolves. Our key insight is that a critic's representation must be explicitly trained to predict the future. WCM jointly learns latent state prediction and value estimation through a lightweight LeJEPA architecture, unifying world modeling with critic learning. This produces a predictive state representation that captures what truly matters for decision making.","html":"<p>🔍 The Missing Ingredient: Temporal Dynamics, Not Just History</p>\n<p>Naively stacking more frames does not solve the problem. Scalar return regression provides too sparse a signal for learning cross-temporal dynamics, so the critic ends up treating history as a larger static feature vector without understanding how the environment evolves. Our key insight is that a critic's representation must be explicitly trained to predict the future. WCM jointly learns latent state prediction and value estimation through a lightweight LeJEPA architecture, unifying world modeling with critic learning. This produces a predictive state representation that captures what truly matters for decision making.</p>\n","updatedAt":"2026-08-04T02:56:16.377Z","author":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","fullname":"FSY(SII)","name":"Sylvest","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8709927201271057},"editors":["Sylvest"],"editorAvatarUrls":["/avatars/c18d2c7dbefd27ec8184622b0834531f.svg"],"reactions":[],"isReport":false,"parentCommentId":"6a7154ae63d384da7d8a3cd5"}},{"id":"6a7154f60d7e3bf145a6142e","author":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","fullname":"FSY(SII)","name":"Sylvest","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false},"createdAt":"2026-08-04T02:56:54.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"⚙️ Seamless Integration, Broad Compatibility\n\nWCM integrates smoothly into both on-policy and off-policy pipelines, and works out of the box with state-of-the-art VLA backbones including Pi0, Pi0.5, and OpenVLA-OFT. No heavyweight modifications, no fragile engineering. Just a principled, end-to-end trainable critic that finally makes observation history work for RL.","html":"<p>⚙️ Seamless Integration, Broad Compatibility</p>\n<p>WCM integrates smoothly into both on-policy and off-policy pipelines, and works out of the box with state-of-the-art VLA backbones including Pi0, Pi0.5, and OpenVLA-OFT. No heavyweight modifications, no fragile engineering. Just a principled, end-to-end trainable critic that finally makes observation history work for RL.</p>\n","updatedAt":"2026-08-04T02:56:54.226Z","author":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","fullname":"FSY(SII)","name":"Sylvest","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8920284509658813},"editors":["Sylvest"],"editorAvatarUrls":["/avatars/c18d2c7dbefd27ec8184622b0834531f.svg"],"reactions":[],"isReport":false,"parentCommentId":"6a7154ae63d384da7d8a3cd5"}},{"id":"6a715501dc8bda86bc042974","author":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","fullname":"FSY(SII)","name":"Sylvest","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false},"createdAt":"2026-08-04T02:57:05.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"📊 SOTA Results Across 149 Tasks on 4 benchmarks and 7 Real-World Tasks\n\nOn ManiSkill, WCM pushes OpenVLA-OFT from a 28% SFT baseline to 99% in-distribution performance and delivers +72.7% OOD gains, even from a near-zero 0.78% starting point. On LIBERO-Plus, WCM trained from one demonstration per task outperforms full-shot SFT trained on 20k trajectories. In the real world, across seven manipulation tasks including deformable cloth folding and dynamic conveyor-belt sushi picking, WCM consistently delivers stable improvements with just hundreds of trajectories and under one hour of training.","html":"<p>📊 SOTA Results Across 149 Tasks on 4 benchmarks and 7 Real-World Tasks</p>\n<p>On ManiSkill, WCM pushes OpenVLA-OFT from a 28% SFT baseline to 99% in-distribution performance and delivers +72.7% OOD gains, even from a near-zero 0.78% starting point. On LIBERO-Plus, WCM trained from one demonstration per task outperforms full-shot SFT trained on 20k trajectories. In the real world, across seven manipulation tasks including deformable cloth folding and dynamic conveyor-belt sushi picking, WCM consistently delivers stable improvements with just hundreds of trajectories and under one hour of training.</p>\n","updatedAt":"2026-08-04T02:57:05.427Z","author":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","fullname":"FSY(SII)","name":"Sylvest","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8840469717979431},"editors":["Sylvest"],"editorAvatarUrls":["/avatars/c18d2c7dbefd27ec8184622b0834531f.svg"],"reactions":[],"isReport":false,"parentCommentId":"6a7154ae63d384da7d8a3cd5"}},{"id":"6a7155880547549d21268aaf","author":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","fullname":"FSY(SII)","name":"Sylvest","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false},"createdAt":"2026-08-04T02:59:20.000Z","type":"comment","data":{"edited":false,"hidden":false,"latest":{"raw":"🌍 Generalization That Holds Up Under Shift\n\nOOD performance is where WCM truly shines. The world prediction objective acts as a dense, structured regularizer, preventing the critic from overfitting to spurious correlations. Unlike conventional critics that degrade under distribution shift, WCM maintains discriminative value estimates and avoids the \"dropping phenomenon\" observed in standard RL. The message is clear: predicting the future makes the present more robust.","html":"<p>🌍 Generalization That Holds Up Under Shift</p>\n<p>OOD performance is where WCM truly shines. The world prediction objective acts as a dense, structured regularizer, preventing the critic from overfitting to spurious correlations. Unlike conventional critics that degrade under distribution shift, WCM maintains discriminative value estimates and avoids the \"dropping phenomenon\" observed in standard RL. The message is clear: predicting the future makes the present more robust.</p>\n","updatedAt":"2026-08-04T02:59:20.673Z","author":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","fullname":"FSY(SII)","name":"Sylvest","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.848678708076477},"editors":["Sylvest"],"editorAvatarUrls":["/avatars/c18d2c7dbefd27ec8184622b0834531f.svg"],"reactions":[],"isReport":false,"parentCommentId":"6a7154ae63d384da7d8a3cd5"}}]},{"id":"6a715518a0845a55e21a080f","author":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","fullname":"FSY(SII)","name":"Sylvest","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false},"createdAt":"2026-08-04T02:57:28.000Z","type":"comment","data":{"edited":true,"hidden":false,"latest":{"raw":"For a deeper dive, check out our paper and project page. WCM rethinks what a critic should learn in the age of VLA-RL.","html":"<p>For a deeper dive, check out our paper and project page. WCM rethinks what a critic should learn in the age of VLA-RL.</p>\n","updatedAt":"2026-08-04T02:59:15.377Z","author":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","fullname":"FSY(SII)","name":"Sylvest","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":10,"isUserFollowing":false}},"numEdits":1,"identifiedLanguage":{"language":"en","probability":0.8448117971420288},"editors":["Sylvest"],"editorAvatarUrls":["/avatars/c18d2c7dbefd27ec8184622b0834531f.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2607.29613","authors":[{"_id":"6a71535bec5082b9f872cdd1","name":"Senyu Fei","hidden":false},{"_id":"6a71535bec5082b9f872cdd2","name":"Xiaopeng Yu","hidden":false},{"_id":"6a71535bec5082b9f872cdd3","name":"Siyin Wang","hidden":false},{"_id":"6a71535bec5082b9f872cdd4","name":"Xianzhong Zhao","hidden":false},{"_id":"6a71535bec5082b9f872cdd5","name":"Jingjing Gong","hidden":false},{"_id":"6a71535bec5082b9f872cdd6","name":"Xipeng Qiu","hidden":false}],"publishedAt":"2026-07-31T00:00:00.000Z","submittedOnDailyAt":"2026-08-04T00:00:00.000Z","title":"WCM: A World Critic Model for Vision-Language-Action Reinforcement Learning","submittedOnDailyBy":{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","isPro":false,"fullname":"FSY(SII)","user":"Sylvest","type":"user","name":"Sylvest"},"summary":"Reinforcement learning (RL) post-training of Vision-Language-Action (VLA) models has shown strong promise for robotic manipulation. Among RL methods, critic-based approaches rely on a value estimator that predominantly operates on single-frame observations or single-frame VLM backbone latents, which is a fundamental mismatch with the partially observable nature of robot control. A naive approach to incorporate observation history into the critic incurs exponential complexity with high-dimensional visual space, and still fails because pure scalar-return regression provides insufficient supervision for learning cross-temporal dynamics. We identify the root cause as a state approximation problem: without an explicit world modeling objective, the critic's representation cannot capture the temporal structure needed for accurate value estimation. To address this, we propose the World Critic Model (WCM), built on a lightweight LeJEPA architecture; WCM jointly predicts future latent state and estimates values, such that the critic's representation is explicitly trained to capture temporal dynamics rather than merely regress scalar returns. WCM integrates seamlessly into both on-policy and off-policy training pipelines and is compatible with state-of-the-art VLA backbones including Pi0, Pi0.5, and OpenVLA-OFT. Extensive experiments on 149 tasks across four benchmarks demonstrate that WCM consistently achieves state-of-the-art performance in both in-distribution and out-of-distribution settings, with particularly strong generalization gains. We further validate WCM on seven real-world manipulation tasks using OpenVLA-OFT and Pi0.5 with off-policy RL, confirming stable deployment across diverse settings.","upvotes":14,"discussionId":"6a71535cec5082b9f872cdd7","projectPage":"https://sylvestf.github.io/wcm-homepage/","githubRepo":"https://github.com/sylvestf/WCM","githubRepoAddedBy":"user","githubStars":11,"organization":{"_id":"613b0dee83ec35d460684607","name":"OpenMOSS-Team","fullname":"OpenMOSS","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/61457b8deff2c9fdb4de4988/N5b9663zQ4uq5_OTNlnmw.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6755c067119ff854c890c2c1","avatarUrl":"/avatars/c18d2c7dbefd27ec8184622b0834531f.svg","isPro":false,"fullname":"FSY(SII)","user":"Sylvest","type":"user"},{"_id":"68721f6af8a466c2cb0e5987","avatarUrl":"/avatars/fd1ed49c18b61b4c80160506c4c3e83d.svg","isPro":false,"fullname":"akanflew","user":"TMSSmith","type":"user"},{"_id":"6a70a02b87a687524a44d852","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6a70a02b87a687524a44d852/buOoE8_Nr51VTtwbovVIm.jpeg","isPro":false,"fullname":"ethan.j","user":"ethan-0230","type":"user"},{"_id":"6a706ceaf5ce69f2faa213d5","avatarUrl":"/avatars/b9bf40d9ddd773969ae2c262d24a2bff.svg","isPro":false,"fullname":"xue","user":"vin0204","type":"user"},{"_id":"6a7157948fad438210d3c7c3","avatarUrl":"/avatars/77311ad4dbe2b8819aa3dceb0b00d3ec.svg","isPro":false,"fullname":"satokenta","user":"satokenta","type":"user"},{"_id":"6a7158e01d662199cb6e48b4","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/sGJjNMNw7cJ9yp7GRdiX7.png","isPro":false,"fullname":"opp","user":"moonquakes024","type":"user"},{"_id":"6a71598c16b81721d7865952","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/Rmy5YD8MJBm4uFguamcKV.png","isPro":false,"fullname":"JianxiangLi","user":"FrayNexus71","type":"user"},{"_id":"6463554dd2044cd1d7c6e0bf","avatarUrl":"/avatars/d7653623117268c545a7063fec69664b.svg","isPro":false,"fullname":"Bingzheng Wei","user":"Bingzheng","type":"user"},{"_id":"6a715c2ea2c5513911301e79","avatarUrl":"/avatars/3d81a0a7f212cf2ffa52924b2faf65d3.svg","isPro":false,"fullname":"E. Carter","user":"Leyuwn","type":"user"},{"_id":"63c1699e40a26dd2db32400d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63c1699e40a26dd2db32400d/3N0-Zp8igv8-52mXAdiiq.jpeg","isPro":false,"fullname":"Chroma","user":"Chroma111","type":"user"},{"_id":"6a7161971d1ec44efec31484","avatarUrl":"/avatars/2f6cdbc8342bee4cb21e90cc77a2f5d2.svg","isPro":false,"fullname":"Yunchen Xu","user":"Shewynn","type":"user"},{"_id":"6900b7a8ecfa14b7b16368fb","avatarUrl":"/avatars/183456c0fa0d5e21c27813ce48d3bad2.svg","isPro":false,"fullname":"Sentinel","user":"Sentinel7","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"613b0dee83ec35d460684607","name":"OpenMOSS-Team","fullname":"OpenMOSS","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/61457b8deff2c9fdb4de4988/N5b9663zQ4uq5_OTNlnmw.png"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2607/2607.29613.md","query":{}}">