Pith. sign in

REVIEW 3 major objections 6 minor 62 references

WorldCycle: Self-Verifiable Reinforcement Learning for Long-Horizon Video World Models

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Reversible action cycles turn long-horizon world-model drift into annotation-free supervision, cutting drift by up to 44 percent and nearly quadrupling composite-action accuracy.

desk verdict A genuinely useful RL post-training recipe for cycle-consistent world models, but the out-of-distribution composite-action claim is not actually established because training already includes combined actions. read the letter →

arxiv 2608.04964 v1 pith:BIUKUU7Z submitted 2026-08-05 cs.AI cs.LG

Bohai Gu , Yueyang Yuan , Taiyi Wu , Dazhao Du , Jian Liu , Xiaoyi Pang , Jie Zhang , Xiaocheng Lu
show 4 more authors
Haobin Zhong Xiaotong Zhao Alan Zhao Song Guo
This is my paper
classification cs.AIcs.LG
keywords videoworldmodelsreinforcementlearningself-verifiablesupervisionreversibleactioncyclesstate-returningdriftlong-horizongenerationcycleconsistencyBench
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Interactive video world models that roll out future frames autoregressively drift as the horizon grows, and for arbitrary action sequences there is no ground-truth future state to measure that drift against. WorldCycle's central claim is that reversible action cycles break this bottleneck: a program followed by its exact inverse must analytically return to the starting state, so the identity of the cycle is a label-free target. The paper turns that identity into two dense rewards, one comparing mirrored forward and reverse frames within a cycle and one comparing phase-aligned frames across repeated cycles, and optimizes both with reinforcement learning. On the released CycleBench it reports up to 44 percent lower state-returning drift and nearly 4x higher accuracy on out-of-distribution composite actions, without degrading local action following or visual quality.

What carries the argument

The central object is the reversible action cycle: an action program followed by its exact inverse, whose composed transformation is the identity $T_{\gamma^{-1}\circ\gamma}=I$. This identity is content-independent and annotation-free, giving a verifier for accumulated transition error. Around it the paper builds two rewards: a spatial closure reward that compares every mirrored forward-reverse frame pair using dense point tracking, and a temporal state consistency reward that anchors repeated cycle executions to the first cycle and penalizes phase-aligned drift. The optimization uses a DiffusionNFT-style reward-weighted velocity objective with a spatial-only warm-up followed by joint training, which the ablations show is necessary because temporal comparisons are uninterpretable until individual cycles roughly close.

What would settle it

Train WorldCycle on actions with irreversible dynamics, such as pushing a deformable object or leaving a persistent mark, and check whether the cycle reward improves pixel closure while the true physical state fails to return; if the model is rewarded for reaching a state that physics says is unreachable, the identity target is invalid and training should degrade rollout accuracy. A cleaner version would run the same cycles in a rigid-body simulator with known reversible kinematics and verify $\hat{T}_{a^{-1}}\circ\hat{T}_a \approx I$ in state space rather than only in pixels.

Watch

Extended reading notes

Core claim

WorldCycle claims that the verification bottleneck in post-training interactive video world models can be broken by restricting attention to reversible dynamics. If the true transformations induced by actions form a group, then any action program $\gamma$ composed with its inverse $\gamma^{-1}$ obeys $T_{\gamma^{-1}\circ\gamma}=I$, so the rollout must return to the starting observation. The paper converts this algebraic identity into dense supervision: for a cycle $(a_1,\dots,a_m,a_m^{-1},\dots,a_1^{-1})$, every intermediate forward state should match its mirrored reverse-pass state, and repeated executions of the same cycle should be phase-stable. These trajectory-level rewards, optimized with a reward-weighted diffusion RL objective, push the model to act as a consistent state operator rather than memorize temporal patterns, and they apply to composite actions because any composite program composed with its inverse is still a closed cycle. The reported result is up to 44 percent reduction in state-returning drift and composite-action accuracy rising from 0.136 to 0.553 over the base model.

Load-bearing premise

The method assumes that the actions it trains on are exactly reversible, so an action program followed by its inverse always returns the world to the same state; when that fails for deformation, contact-rich interaction, or persistent state changes, the reward's identity target is wrong.

Editorial extensions

If this is right

  • State-returning consistency can be improved without ground-truth video, because the cycle identity supplies the long-horizon target that arbitrary open trajectories cannot provide.
  • The gains grow with horizon: on 381-frame rollouts the repeated-cycle stability error drops by 34 percent relative to the strongest prior post-trained baseline.
  • Out-of-distribution composite actions become trainable: a composite sequence composed with its inverse is still a closed cycle, so the model can be optimized on motions its pretraining data rarely shows, raising accuracy from 0.136 to 0.553.
  • Cycle training does not require sacrificing quality: the full model reports the highest visual-quality score among all WorldCycle variants and comparable short-term action accuracy to the prior post-trained baseline.
  • Scaling alone does not solve state-returning consistency, since a larger 14B foundation model still lags all RL post-trained baselines on the CycleBench consistency metrics.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same self-verifiable principle should extend beyond reversibility to other analytically known constraints, such as conservation laws or fixed terminal states, which the paper flags as future work; if such constraints exist for irreversible dynamics, the identical RL recipe could apply.
  • Because the rewards are computed from pixel correspondences, the method inherits the sensitivity of the point-tracking proxy; replacing it with a geometry-aware estimator such as camera pose or depth would test whether the state-returning signal is tracking true state or only appearance.
  • The composite-action gains suggest the model learns an action-composition algebra rather than trajectory memory; a direct test would measure transfer to unseen cycles built from the same primitives, something CycleBench does not currently isolate.
  • If real scenes are only approximately reversible, the identity target becomes biased; running the same cycle in a physics simulator with known reversible kinematics could separate pixel-space closure from true state closure and quantify that bias.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. WorldCycle proposes to post-train interactive video world models with reinforcement learning using self-verifiable supervision from reversible action cycles. Because a forward action sequence followed by its inverse must analytically return to the initial state, the identity of the composed transformation (Eq. 2) provides a label-free long-horizon target. The method defines a spatial closure reward over mirrored forward/reverse frames and a temporal consistency reward over repeated cycle executions, combined with action-following and visual-quality rewards in a DiffusionNFT-style objective. The paper also introduces CycleBench, a benchmark of 47 action trajectories evaluated from 380 initial frames across four tasks (reversible, closed, repeated, and cascaded cycles) and four horizon/complexity settings, with evaluation based on RoMa to avoid overlap with the CoTracker-based training reward. On CycleBench, WorldCycle reports drift reductions up to 44% relative to WorldCompass and composite-action accuracy of 0.553 versus 0.136 for the base WorldPlay model.

Significance. The core idea is genuinely useful: the cycle-return identity is analytic and not fitted, converting an otherwise unavailable long-horizon target into dense, annotation-free supervision for reversible dynamics. The benchmark is thoughtfully designed with four complementary diagnostics of error accumulation, and the use of RoMa for evaluation provides independence from the training reward. The ablations in Table 2 give credible evidence that both spatial closure and temporal consistency contribute, and that the warm-up-and-combine schedule matters. If the out-of-distribution composite-action claim is supported with a disjoint action split, and if the headline numbers come with uncertainty quantification, this would be an important post-training method for interactive world models. At present, the empirical case is plausible but incomplete.

major comments (3)
  1. [Analysis on Composite Action Generalization; Table 1; Benchmark Details] The paper's central generalization claim is that cycle supervision 'extend[s] naturally to out-of-distribution composite action cycles' (Abstract, Introduction, Analysis on Composite Action Generalization), supported by the composite-action row of Table 1. The evidence is incomplete: training 'samples a base or combined action sequence and constructs closed reversible cycles... on the fly,' while CycleBench's composite setting is described only as 'controls that combine multiple motion components.' No statement or analysis establishes that the composite programs (or the combination rules generating them) used in CycleBench are disjoint from those sampled during WorldCycle training. The base model's 0.136 accuracy shows the composites are OOD for pretraining, but not for post-training; if benchmark composites overlap training cycles, the 0.553 result is in-domain optimization rather than generalization. The residual-factorization argument in Eq. (10) is a motivation, not evidence of held-out transfer. Please add a disjoint held-out action split: enumerate composite programs, exclude them from training cycle construction, and report performance on composites with held-out component combinations or composition depths.
  2. [Table 1; Evaluation; Experiments] All empirical results are point estimates over 380 videos, with no error bars, confidence intervals, or multiple-seed results. The headline numbers (44% drift reduction, 4x composite accuracy) are small- to moderate-sized effects, and several comparisons are very close (e.g., short-term Acc 0.833 vs 0.829, Qual 11.24 vs 11.06). Because the 380 videos are generated from 47 action trajectories, the effective sample size may be much smaller than 380, and no clustered variance is reported. Please report standard errors or confidence intervals clustered by action trajectory, and run at least three training seeds for the main comparison, or explicitly justify why this is infeasible.
  3. [Discussion and Limitations; Preliminaries, Eq. (2)] The identity target T_a ∘ T_{a^{-1}} = I assumes exact reversibility. The Limitations section correctly concedes that deformation, contact-rich interaction, and persistent state changes violate this assumption, but the benchmark includes real-world photorealistic images from 4KLSDB and synthetic GPT Image scenes, where lighting, dynamic content, and non-rigid elements are present. In such scenes, a camera-ego-motion trajectory is not exactly reversible at the pixel level, so the cycle rewards can penalize physically correct dynamics. Please either restrict the claims to controlled reversible scene content or explain how the rewards are intended to treat non-reversible elements within otherwise reversible trajectories. A useful concrete test is to evaluate on synthetic scenes with known reversible dynamics and compare the result with real-scene performance.
minor comments (6)
  1. [Eqs. (5)-(6)] The reward temperature alpha is used in both spatial and temporal rewards but its value is never reported. Please give the value and, if possible, a small sensitivity check.
  2. [Eq. (4) vs Eq. (11)] The training discrepancy in Eq. (4) is normalized by (H+W), while the benchmark distance in Eq. (11) is an unnormalized mean pixel displacement. This is a reasonable design choice, but the difference should be stated explicitly so readers do not compare reward-scale distances with evaluation-scale distances.
  3. [CycleBench Overview] The benchmark is described as '47 action trajectories evaluated from 380 initial frames.' Please clarify whether the 380 videos are 380 initial frames each rolled out once, or a smaller set of trajectories repeated across initial frames, since this affects the effective sample size and correlation structure.
  4. [Experiments; General Reproducibility] No code, model checkpoints, CycleBench action definitions, inverse-mapping tables, or composite-action construction rules are linked. The action vocabulary and inverse operation are load-bearing for the method, so releasing these artifacts is important for reproducibility.
  5. [Table 1 caption; Abstract] The Table 1 caption says the reported gains are relative to WorldCompass, while the Abstract and Conclusion compare the 4x composite gain to the base model. Please state both references consistently in every place where percentage gains appear.
  6. [Introduction] The claim that WorldPlay exhibits a '5x accuracy collapse' on composite actions is based on 0.635 vs 0.136, which is approximately 4.7x. Please use the exact ratio or round consistently with the other reported ratios.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the identity target is an analytic group axiom, evaluation is decoupled from the training reward via RoMa, and the central comparisons are against external baselines; the unverified composite-action split is an experimental-design gap, not a definitional reduction.

full rationale

The load-bearing target is Eq. (2): the inverse action program cancels by construction of the group action, giving an analytic identity target rather than a fitted parameter or a quantity defined in terms of model outputs. The spatial closure and temporal consistency rewards (Eqs. 5 and 6) are built from that identity and use CoTracker only as an observable proxy for state discrepancy. CycleBench explicitly adopts RoMa as an independent correspondence estimator (Eq. 11 and the paragraph starting 'To decouple evaluation'), so evaluation is not the same function as the training reward. The central empirical claims are measured against external baselines (WorldPlay, WorldCompass, Lingbot World v2) with ablations that isolate the cycle rewards, the schedule, and the reward composition. The main caveat is that generalization to 'unseen action compositions' is not backed by a demonstrated disjoint split between training-sampled combined action sequences and CycleBench's composite-action programs; this is a missing experimental control that affects the strength of the out-of-distribution claim, but it is not a case where the paper's equations make the predicted quantity equal to its own inputs. The stated limitation about irreversible dynamics narrows the scope of the method rather than revealing a circular derivation. No load-bearing self-citation or imported uniqueness theorem appears in the derivation chain.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The paper's central improvement is empirical. It rests on the algebraic identity of reversible action cycles and on pixel-space correspondence as a state proxy, both acknowledged as limiting assumptions. Reward and schedule hyperparameters are free but not fitted to target metrics. No new physical entities are postulated; CycleBench is an evaluation artifact rather than an invented entity.

free parameters (4)
  • reward temperature alpha (Eqs. 5 and 6) = not reported
    Controls the exponential sharpness of both cycle rewards; the exact value is needed to reproduce the reward scale and is omitted.
  • reward weights lambda_s, lambda_t, lambda_v, lambda_a = 1.0, 0.5, 1.0, 2.0
    Hand-set trade-offs between spatial closure, temporal consistency, action fidelity, and visual quality; they define the final objective but are not fitted to the benchmark.
  • spatial warm-up threshold t_w = 300 steps
    Determines when the temporal reward is activated; the schedule ablation shows this choice matters.
  • cycle repetition K and rollout group size G = K up to 8, G=16
    Rollout structure chosen by hand; affects reward variance and the horizon over which temporal consistency is measured.
assumptions (4)
  • domain assumption Action-induced transformations form a group with exact inverses: T_a composed with T_{a^{-1}} equals identity (Eq. 2).
    The entire self-verification target depends on exact reversibility; the authors note in the Limitations section that it fails for irreversible and contact-rich dynamics.
  • domain assumption Dense pixel correspondences from CoTracker or RoMa faithfully measure state identity.
    Equations 4 and 11 use image-space displacement as a proxy for state distance; occlusion, appearance drift, or camera effects can decouple visual alignment from true state.
  • domain assumption The autoregressive diffusion world model behaves as a Markovian state-transition system f_theta(s, a).
    Equation 1 and Algorithm 1 treat the video model as a deterministic transition operator, while generation is stochastic and the true state is latent.
  • domain assumption DiffusionNFT reward-weighted velocity regression is a valid RL update for long rollouts.
    Borrowed from Zheng et al. 2025; the paper does not prove convergence or unbiasedness in the autoregressive world-model setting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WorldCycle: Self-Verifiable Reinforcement Learning for Long-Horizon Video World Models." pith.science (2026). https://pith.science/paper/BIUKUU7Z

@misc{pith2026260804964,
  author       = {Pith},
  title        = {Pith review of: WorldCycle: Self-Verifiable Reinforcement Learning for Long-Horizon Video World Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BIUKUU7Z}},
  note         = {Machine review of arXiv:2608.04964}
}
read the original abstract

Interactive video world models are essential for long-horizon planning and exploration, yet they suffer from compounding errors. Post-training methods such as reinforcement learning (RL) can improve these models, but they hit a verification bottleneck: for arbitrary action sequences, no ground-truth future state exists to measure long-term drift. Our key insight is that reversible action cycles make this verification possible: a sequence composed with its inverse must analytically return to the initial state, yielding annotation-free supervision on long-horizon correctness. Building on this, we introduce WorldCycle, a self-verifiable RL framework that constructs closed action cycles and their repeated executions from ordinary action sequences, and optimizes two complementary rewards: a spatial closure reward enforcing symmetry between mirrored forward and reverse segments, and a temporal consistency reward aligning states across repeated cycle executions. These rewards force the model to learn actions as consistent state operators rather than memorized temporal patterns, and extend naturally to out-of-distribution composite action cycles that the base model handles poorly. We further release CycleBench, a diagnostic benchmark for state-returning ability under complex action structures. WorldCycle reduces state returning drift by up to 44% and lifts composite-action accuracy nearly 4x over the base model, providing a vital foundation for physically grounded world models.

Figures

Figures reproduced from arXiv: 2608.04964 by the authors.

Figure 1
Figure 1. State-returning failure in interactive video world models. Even state-of-the-art models fail a minimal sanity check: an inverse action sequence such as forward-then-backward does not recover the starting view (spatial closure failure), and executing the same action from what should be the same state at different rollout steps produces inconsistent displacements (temporal consistency failure). Both failures accumulat… view at source ↗
Figure 2
Figure 2. Overview of WorldCycle. From ordinary action sequences we construct closed reversible cycles and their repeated executions. The spatial closure reward compares every mirrored forward-reverse frame pair within a single cycle, providing dense supervision on where the trajectory begins to drift. The temporal consistency reward compares phase-aligned frames of the same action across repeated cycles, penalizing drift of … view at source ↗
Figure 3
Figure 3. Overview of CycleBench. Four complementary tasks probe different aspects of accumulated transition error: T1 measures where error appears along a reversible path; T2 measures the residual after a non-symmetric closed trajectory; T3 tracks error growth across repeated cycles; T4 measures error propagation across structurally different cycles. All tasks are further evaluated under three horizon settings (short/mid/lon… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on CycleBench (253 frames). Left: composite-action setting (forward+turn then inverse). Right: simple-action setting (forward then backward). Each method row contains eight sampled frames in total, with four keyframes from each setting: two from …
Figure 5
Figure 5. Figure 5: Evolution of cycle consistency and visual quality [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Long-horizon qualitative comparison over five repeated cycles (381 frames). Each row shows sampled frames from the same long-term rollout. The bottom row enlarges the red-boxed region in the final return frame and compares it with the corresponding region in the first …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 36 canonical work pages

  1. [1]

    arXiv preprint arXiv:1803.10122 , year=

    World models , author=. arXiv preprint arXiv:1803.10122 , year=

    arXiv

  2. [2]

    Proceedings of the 36th International Conference on Machine Learning , series =

    Learning Latent Dynamics for Planning from Pixels , author =. Proceedings of the 36th International Conference on Machine Learning , series =. 2019 , publisher =

    work page 2019

  3. [3]

    Nature , volume =

    Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model , author =. Nature , volume =. 2020 , doi =

    work page 2020

  4. [4]

    Nature , volume=

    Mastering diverse control tasks through world models , author=. Nature , volume=. 2025 , publisher=

    2025

  5. [5]

    Advances in Neural Information Processing Systems , volume =

    Scheduled Sampling for Sequence Prediction with Recurrent Neural Networks , author =. Advances in Neural Information Processing Systems , volume =

  6. [6]

    Advances in Neural Information Processing Systems , volume =

    When to Trust Your Model: Model-Based Policy Optimization , author =. Advances in Neural Information Processing Systems , volume =

    work page

  7. [7]

    Advances in Neural Information Processing Systems , volume =

    Gradient Surgery for Multi-Task Learning , author =. Advances in Neural Information Processing Systems , volume =. 2020 , publisher =

    work page 2020

  8. [8]

    Proceedings of the 36th International Conference on Machine Learning , series =

    Policy Consolidation for Continual Reinforcement Learning , author =. Proceedings of the 36th International Conference on Machine Learning , series =. 2019 , publisher =

    work page 2019

Show all 62 references
  1. [9]

    Proceedings of the National Academy of Sciences , volume =

    Overcoming Catastrophic Forgetting in Neural Networks , author =. Proceedings of the National Academy of Sciences , volume =. 2017 , doi =

    2017

  2. [10]

    Advances in Neural Information Processing Systems , year =

    Training Language Models to Follow Instructions with Human Feedback , author =. Advances in Neural Information Processing Systems , year =

  3. [11]

    arXiv preprint arXiv:2402.03300 , year =

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author =. arXiv preprint arXiv:2402.03300 , year =

    arXiv

  4. [12]

    arXiv preprint arXiv:2501.12948 , year =

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning , author =. arXiv preprint arXiv:2501.12948 , year =

    arXiv

  5. [13]

    International Conference on Learning Representations , year =

    Training Diffusion Models with Reinforcement Learning , author =. International Conference on Learning Representations , year =

  6. [14]

    arXiv preprint arXiv:2310.03739 , year =

    Aligning Text-to-Image Diffusion Models with Reward Backpropagation , author =. arXiv preprint arXiv:2310.03739 , year =

    arXiv

  7. [15]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

    Diffusion Model Alignment Using Direct Preference Optimization , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , year =

  8. [16]

    arXiv preprint arXiv:2505.05470 , year =

    Flow-GRPO: Training Flow Matching Models via Online RL , author =. arXiv preprint arXiv:2505.05470 , year =

    arXiv

  9. [17]

    arXiv preprint arXiv:2505.07818 , year =

    DanceGRPO: Unleashing GRPO on Visual Generation , author =. arXiv preprint arXiv:2505.07818 , year =

    arXiv

  10. [18]

    arXiv preprint arXiv:2509.16117 , year =

    DiffusionNFT: Online Diffusion Reinforcement with Forward Process , author =. arXiv preprint arXiv:2509.16117 , year =

    arXiv

  11. [19]

    Advances in Neural Information Processing Systems , volume =

    Video Diffusion Models , author =. Advances in Neural Information Processing Systems , volume =

  12. [20]

    Proceedings of the 41st International Conference on Machine Learning , year =

    VideoPoet: A Large Language Model for Zero-Shot Video Generation , author =. Proceedings of the 41st International Conference on Machine Learning , year =

  13. [21]

    2024 , howpublished =

    Video Generation Models as World Simulators , author =. 2024 , howpublished =

    2024

  14. [22]

    International Conference on Learning Representations , year =

    Learning Interactive Real-World Simulators , author =. International Conference on Learning Representations , year =

  15. [23]

    Wu, Jialong and Yin, Shaofeng and Feng, Ningya and He, Xu and Li, Dong and Hao, Jianye and Long, Mingsheng , booktitle =

  16. [24]

    and Hu, Zhiting , journal =

    Xiang, Jiannan and Liu, Guangyi and Gu, Yi and Gao, Qiyue and Ning, Yuting and Zha, Yuheng and Feng, Zeyu and Tao, Tianhua and Hao, Shibo and Shi, Yemin and Liu, Zhengzhong and Xing, Eric P. and Hu, Zhiting , journal =

  17. [25]

    Proceedings of the 41st International Conference on Machine Learning , year =

    Genie: Generative Interactive Environments , author =. Proceedings of the 41st International Conference on Machine Learning , year =

  18. [26]

    2024 , howpublished =

    Genie 2: A Large-Scale Foundation World Model , author =. 2024 , howpublished =

    2024

  19. [27]

    2025 , howpublished =

    Genie 3: A New Frontier for World Models , author =. 2025 , howpublished =

    2025

  20. [28]

    arXiv preprint arXiv:2408.14837 , year =

    Diffusion Models Are Real-Time Game Engines , author =. arXiv preprint arXiv:2408.14837 , year =

    arXiv

  21. [29]

    2024 , howpublished =

    2024

  22. [30]

    Advances in Neural Information Processing Systems , year =

    Diffusion for World Modeling: Visual Details Matter in Atari , author =. Advances in Neural Information Processing Systems , year =

  23. [31]

    arXiv preprint arXiv:2506.18701 , year =

    Matrix-Game: Interactive World Foundation Model , author =. arXiv preprint arXiv:2506.18701 , year =

    arXiv

  24. [32]

    Proceedings of the 43rd International Conference on Machine Learning , note =

    WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling , author =. Proceedings of the 43rd International Conference on Machine Learning , note =

  25. [33]

    arXiv preprint arXiv:2607.07534 , year =

    Infinite Worlds with Versatile Interactions , author =. arXiv preprint arXiv:2607.07534 , year =

    arXiv

  26. [34]

    arXiv preprint arXiv:2604.14268 , year =

    HY-World 2.0: A Multi-Modal World Model for Reconstructing, Generating, and Simulating 3D Worlds , author =. arXiv preprint arXiv:2604.14268 , year =

    arXiv

  27. [35]

    arXiv preprint arXiv:2501.03575 , year =

    Cosmos World Foundation Model Platform for Physical AI , author =. arXiv preprint arXiv:2501.03575 , year =

    arXiv

  28. [36]

    arXiv preprint arXiv:2507.21809 , year =

    HunyuanWorld 1.0: Generating Immersive, Explorable, and Interactive 3D Worlds from Words or Pixels , author =. arXiv preprint arXiv:2507.21809 , year =

    arXiv

  29. [37]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Epona: Autoregressive Diffusion World Model for Autonomous Driving , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  30. [38]

    2025 , doi =

    Zhu, Fangqi and Wu, Hongtao and Guo, Song and Liu, Yuxiao and Cheang, Chilam and Kong, Tao , booktitle =. 2025 , doi =

    2025

  31. [39]

    2024 , publisher =

    Zhou, Siyuan and Du, Yilun and Chen, Jiaben and Li, Yandong and Yeung, Dit-Yan and Gan, Chuang , booktitle =. 2024 , publisher =

    2024

  32. [40]

    Hu, Anthony and Russell, Lloyd and Yeo, Hudson and Murez, Zak and Fedoseev, George and Kendall, Alex and Shotton, Jamie and Corrado, Gianluca , journal =

  33. [41]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  34. [42]

    Gao, Shenyuan and Yang, Jiazhi and Chen, Li and Chitta, Kashyap and Qiu, Yihang and Geiger, Andreas and Zhang, Jun and Li, Hongyang , journal =

  35. [43]

    2026 , note =

    Huang, Junchao and Ye, Ziyang and Hu, Xinting and He, Tianyu and Zhang, Guiyu and Shi, Shaoshuai and Bian, Jiang and Jiang, Li , booktitle =. 2026 , note =

    2026

  36. [44]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Learning Correspondence from the Cycle-Consistency of Time , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  37. [45]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Temporal Cycle-Consistency Learning , author =. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  38. [46]

    Su, Zihan and Hu, Teng and Zhang, Jiangning and Wang, Ruiyan and Yi, Ran and Ma, Lizhuang and Tao, Dacheng , journal =

  39. [47]

    2602.08025 , archivePrefix =

    Ye, Yixuan and Lu, Xuanyu and Jiang, Yuxin and Gu, Yuchao and Zhao, Rui and Liang, Qiwei and Pan, Jiachun and Zhang, Fengda and Wu, Weijia and Wang, Alex Jinpeng , year =. 2602.08025 , archivePrefix =

  40. [48]

    Ying, Kaining and Hu, Hengrui and Ren, Siyu and Li, Jiamu and Chen, Fengjiao and Wang, Ziwen and Cao, Xuezhi and Cai, Xunliang and Ding, Henghui , journal =

  41. [49]

    Proceedings of the 40th International Conference on Machine Learning , series =

    Temporally Consistent Transformers for Video Generation , author =. Proceedings of the 40th International Conference on Machine Learning , series =. 2023 , publisher =

    2023

  42. [50]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Long-Context State-Space Video World Models , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  43. [51]

    arXiv preprint arXiv:2404.08471 , year =

    Revisiting Feature Prediction for Learning Visual Representations from Video , author =. arXiv preprint arXiv:2404.08471 , year =

    arXiv

  44. [52]

    arXiv preprint arXiv:2505.21996 , year =

    Learning World Models for Interactive Video Generation , author =. arXiv preprint arXiv:2505.21996 , year =

    arXiv

  45. [53]

    arXiv preprint arXiv:2506.05284 , year =

    Video World Models with Long-Term Spatial Memory , author =. arXiv preprint arXiv:2506.05284 , year =

    arXiv

  46. [54]

    Advances in Neural Information Processing Systems , year =

    RLVR-World: Training World Models with Reinforcement Learning , author =. Advances in Neural Information Processing Systems , year =

  47. [55]

    arXiv preprint arXiv:2509.23958 , year =

    Reinforcement Learning with Inverse Rewards for World Model Post-Training , author =. arXiv preprint arXiv:2509.23958 , year =

  48. [56]

    Wang, Zehan and Wang, Tengfei and Zhang, Haiyu and Zuo, Xuhui and Wu, Junta and Wang, Haoyuan and Sun, Wenqiang and Wang, Zhenwei and Cao, Chenjie and Zhao, Hengshuang and Guo, Chunchao and Zhao, Zhou , booktitle =

  49. [57]

    2024 , publisher =

    Karaev, Nikita and Rocco, Ignacio and Graham, Benjamin and Neverova, Natalia and Vedaldi, Andrea and Rupprecht, Christian , booktitle =. 2024 , publisher =

    2024

  50. [58]

    IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

    Edstedt, Johan and Sun, Qiyu and B. IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages =

  51. [59]

    2025 , eprint=

    HPSv3: Towards Wide-Spectrum Human Preference Score , author=. 2025 , eprint=

    2025

  52. [60]

    arXiv preprint arXiv:2605.24578 , year =

    World Models as Group Actions , author =. arXiv preprint arXiv:2605.24578 , year =. 2605.24578 , archivePrefix =

    arXiv

  53. [61]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , pages =

    4KLSDB: A Large-Scale Dataset for 4K Image Restoration and Text-to-Image Generation , author =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , pages =

  54. [62]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Vbench: Comprehensive benchmark suite for video generative models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.