Pith. sign in

REVIEW 3 major objections 45 references

One ~21M JEPA checkpoint serves planning, behaviour cloning, and inverse dynamics without retraining.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-11 10:36 UTC pith:DYMMO2GB

load-bearing objection Solid first sub-billion multi-paradigm JEPA controller; the multi-paradigm headline is real but scoped to near-expert data and goal-visible or tolerant envs. the 3 major comments →

arxiv 2607.04978 v1 pith:DYMMO2GB submitted 2026-07-06 cs.LG cs.CVcs.RO

Qantara: Bridge-Flow Training for Multi-Paradigm JEPA Control

Ruslan Rakhimov , George Bredis , Yuriy Maksyuta , Daniil Gavrilov This is my paper
classification cs.LG cs.CVcs.RO
keywords JEPAworld modelsflow matchingBrownian bridgelatent planningbehaviour cloninginverse dynamicsvisuomotor control
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Existing latent world models for control from pixels lock in one use at training: either plan with a dynamics model or clone expert behaviour. Qantara trains a single small joint-embedding predictor so that choice can be made later, when latency, search budget, or goal access is known. It pairs a Brownian bridge between consecutive clean latents with flow matching on actions, and samples noise levels only on the edges of the action-state noise square that inference actually queries. From the same weights the model can plan toward a goal latent, sample actions like a policy, or first predict the next latent then recover the bridging action. On a four-environment control suite the planning path averages 91.2 success rate and sets a new mark on a 3D pick-and-place task; the two goal-blind paths still reach high success on environments where the observation or success tolerance carries enough information. The result is a multi-paradigm controller rather than a single-use planner.

Core claim

A single end-to-end JEPA of roughly 21 million parameters, trained by pairing a Brownian-bridge interpolant on consecutive clean latents with noise-to-data flow matching on actions under edge-aligned (action-time, state-time) sampling, serves three inference paradigms without retraining—latent planning, behaviour-cloning sampling, and video–inverse composition—reaching a 91.2 success-rate average on the LeWM suite and 93.7 on OGBench-Cube.

What carries the argument

Bridge-flow training with edge-aligned sampling: a Brownian bridge between consecutive clean latents on the state axis, flow matching on the action axis, and mass concentrated on the four edges (plus joint diagonal) of the (action-time, state-time) noise square that the three inference dispatches actually query.

Load-bearing premise

The multi-paradigm claim depends on near-expert offline trajectories that already span the goals, so that the goal-blind paths can succeed by replaying expert-like continuations.

What would settle it

Train and evaluate the same recipe on purely random or heavily suboptimal trajectories for a sparse-reward goal-reaching task; if the behaviour-cloning and video–inverse paths collapse while planning remains usable, the multi-paradigm claim is limited to near-expert data regimes.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 0 minor

Summary. Qantara is a ~21M-parameter end-to-end pixel JEPA world model trained offline without rewards. Its joint objective pairs a Brownian-bridge interpolant between consecutive clean latents (state axis) with noise-to-data flow matching (action axis), under edge-aligned sampling of (τa, τz) on the four edges of the noise square plus a joint diagonal. From one checkpoint the model supports three inference dispatches without retraining: CEM latent planning on the forward edge, behaviour-cloning Euler sampling on the policy edge, and a video–inverse composition (action-free next-latent prediction then inverse-dynamics recovery). On the four-env LeWM suite the planning path averages 91.2 SR (3 train seeds) and reaches 93.7 SR on OGBench-Cube; BC and video–inverse from the same weights reach 82–83 SR on Push-T and ~71–73 SR on Cube at 15–65× lower per-step cost. Ablations show uniform interior sampling and several mode leave-one-outs are load-bearing, especially on Push-T.

Significance. The work fills the previously empty (JEPA-latent, multi-paradigm, sub-billion) cell: prior JEPA world models commit to one inference paradigm at training time, while multi-paradigm joint denoisers operate in pixel space at 150M–7B scale. The bridge-flow design and edge-aligned sampling are a clean, inference-matched training rule with strong ablations (uniform sampling drops Push-T CEM 90.1→53.3; residual+zero-init and γ=1 are required; mode LOO reveals unexpected trunk co-regularisation). Multi-seed evaluation (3 train × 5 eval × 50 episodes), a same-pipeline LeWM reproduction, and transparent reporting of goal-blind failures (Reacher) are strengths. If the multi-paradigm framing is scoped carefully, this is a solid step from single-paradigm JEPA planners toward deploy-time flexible controllers.

major comments (3)
  1. The multi-paradigm headline (abstract, §1, §5) is stronger than what Tab. 2 and §3.3 support under goal-blind dispatch. BC and video–inverse succeed when the observation already exposes the goal (Push-T) or when near-expert demos plus success tolerance absorb imitation drift (Two-Room/Cube ~69–73 SR); on Reacher they fall to ~5–6 SR (near noise). §3.1 and §3.3 already state that purely random data rarely reaches sparse goals and that goal-blind paths carry the claim only under those conditions. The central technical result—one checkpoint queryable three ways—is sound, but the abstract and conclusion should state the data/observation conditions under which non-planning dispatches are useful, and avoid implying a general three-way controller on arbitrary offline data.
  2. Table 1 mixes pipelines: PLDM/DINO-WM/LeWM-paper numbers come from Maes et al., while LeWM (reproduction) and Qantara share the authors’ pipeline. The reproduction trails the paper on Push-T (−3.7) and especially Reacher (−21.7), so head-to-head gains vs published LeWM on those envs are not cleanly attributable to method. Cube (+17.0 vs reproduction, +7.7 vs DINO-WM paper) remains the strongest cell. Please foreground same-pipeline comparisons for claims of improvement over LeWM, and treat cross-pipeline SOTA language more cautiously outside Cube.
  3. §3.1 axiom that near-expert, goal-spanning offline trajectories suffice for all three modes is load-bearing for the multi-paradigm claim and is not stress-tested. There is no ablation on suboptimal, random, or goal-sparse data (the paper notes random exploration rarely reaches success regions). Without that, it remains unclear whether edge-aligned bridge-flow training still yields usable BC/video–inverse paths, or only a strong planner, when the data distribution is weaker. A limited data-quality ablation on at least one env (e.g., Push-T or Cube) would substantially strengthen the claim.

Circularity Check

0 steps flagged

No significant circularity: empirical multi-paradigm JEPA design with external SR metrics, not a self-defining derivation.

full rationale

Qantara is an empirical methods paper. Its central claim—that one ~21M JEPA checkpoint trained with a Brownian-bridge state interpolant, action flow matching, and edge-aligned (τa, τz) sampling can serve planning, BC, and video–inverse without retraining—is tested by measuring success rates against external environment criteria (task completion on Push-T, Two-Room, Cube, Reacher), not by reconstructing quantities from fitted constants. The training–inference alignment (sample the noise-square edges that dispatches query) is an intentional design choice, falsified by the uniform-sampler ablation (Push-T CEM 90.1→53.3). Loss weights, γ, residual+zero-init, and SIGReg are hyperparameters with ablations; none define the reported SR. Citations to LeWM, SIGReg, flow matching, and Brownian bridges supply building blocks and baselines, not uniqueness theorems that force the multi-paradigm result. No equation reduces a claimed prediction to its inputs by construction. Score 0 is the correct honest finding.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The central multi-paradigm claim rests on standard flow-matching/Brownian-bridge math, JEPA training practice (SIGReg, residual heads), and several hand-chosen loss and sampling hyperparameters. The main invented construct is the edge-aligned five-mode training design itself; its necessity is argued empirically via ablations rather than derived. No new physical entities; free parameters are recipe knobs that affect stability and SR.

free parameters (4)
  • λz, λa, λSIG loss weights = 3, 1, 0.09
    Fixed at 3, 1, 0.09 without a derivation; balance between bridge, flow-matching, and anti-collapse terms is recipe-level and affects training stability.
  • Brownian-bridge noise scale γ = 1
    Chosen as γ=1 (stochastic-interpolant canonical); off-canonical values collapse seeds. Not fitted to SR but selected by ablation.
  • Inference Euler/CEM step counts K = CEM K=4; BC/VI K=2
    Defaults K=4 (CEM) and K=2 (BC/video–inverse) chosen for latency/quality tradeoff; suite average is fairly flat but still a free deployment knob.
  • Five-mode edge sampling mixture = 5 equal modes
    Equal replication of forward, policy, inverse, video, and joint modes per batch is a design choice; dropping modes destabilizes Push-T even off-corner.
axioms (5)
  • domain assumption Noise-to-data flow matching is a valid generative model for continuous actions conditioned on latent context.
    Invoked for the action axis (§2.1, Eq. 3) from Lipman et al. / rectified-flow literature; not re-proved.
  • domain assumption A Brownian bridge between consecutive clean latents is the right state interpolant when the previous latent is available at τz=0.
    Core modeling choice (§2.1, Eq. 4); justified by matching planner inputs, supported by residual/bridge ablations, not derived from first principles.
  • domain assumption SIGReg (sketched isotropic Gaussian regularizer) prevents encoder collapse without stop-grad or EMA targets.
    Adopted from LeJEPA/LeWM recipe (§2.1); sole anti-collapse mechanism in the paper.
  • ad hoc to paper Near-expert offline trajectories suffice for all three inference modes on the evaluated envs.
    Stated in §3.1; BC cannot exceed the demo policy and video–inverse needs goal-spanning coverage—assumption fails on Reacher goal-blind paths.
  • standard math Standard transformer block-causal + within-block bidirectional attention can jointly denoise action and state tokens.
    Architectural premise of §2.1; standard sequence modeling assumption.
invented entities (3)
  • Qantara bridge-flow multi-paradigm JEPA predictor no independent evidence
    purpose: Single shared trunk that parameterizes planning, BC, and inverse dynamics via different (τa, τz) edge queries.
    The model and joint objective are the paper’s main construct; evidence is empirical SR and ablations, not independent external measurement.
  • Edge-aligned five-mode (τa, τz) sampling (four edges + joint diagonal) no independent evidence
    purpose: Concentrate training mass on inference-queried corners and co-regularize the shared trunk.
    Introduced in §2.3; necessity shown by uniform-sampler and mode-LOO ablations, not by external theory.
  • Video–inverse composition dispatch no independent evidence
    purpose: Factor BC through action-free next-latent prediction then inverse dynamics without a pixel decoder.
    Composition of two edges (§2.2); related to UniPi-style video-as-policy but in JEPA latent space.

pith-pipeline@v1.1.0-grok45 · 23487 in / 3892 out tokens · 34598 ms · 2026-07-11T10:36:11.237549+00:00 · methodology

0 comments
read the original abstract

Joint-Embedding Predictive Architectures (JEPAs) underpin a growing family of latent world models for control from raw pixels, but every existing JEPA world model commits at training time to a single inference paradigm: either trajectory optimisation in a learned dynamics model, or direct behaviour cloning. A single checkpoint that serves both would defer this choice to inference, when deployment constraints (rollout cost, observation accessibility) determine which path wins. We present Qantara, an end-to-end JEPA whose joint training objective pairs a Brownian-bridge interpolant between consecutive clean latents on the state axis with noise-to-data flow matching on the action axis. The same checkpoint serves three inference paradigms without retraining: latent planning, behaviour-cloning action sampling, and inverse dynamics, which we query through a video-inverse composition that first predicts the next latent without action conditioning, then extracts the action. Training concentrates mass on the edges of the (action-time, state-time) noise square, where inference queries the predictor: replacing it with uniform interior sampling drops Push-T planning from 90.1 to 53.3 SR at matched compute. On the LeWM control suite, Qantara reaches a 91.2 SR three-train-seed average and sets new SOTA on OGBench-Cube (+7.7 SR over DINO-WM, +19.7 over LeWM). From the same weights, the behaviour-cloning and video-inverse paths reach 82-83 SR on Push-T and 71-73 SR on Cube. These results move JEPA world models from single-paradigm planners to multi-paradigm controllers.

Figures

Figures reproduced from arXiv: 2607.04978 by Daniil Gavrilov, George Bredis, Ruslan Rakhimov, Yuriy Maksyuta.

Figure 1
Figure 1. Figure 1: Qantara training and edge-aligned (τ a , τ z ) sampling. Left: block-causal predictor on (a, z) tokens; a clean prefix conditions the noisy continuation and per block the action head emits an action velocity v a while the state head emits a latent residual that yields zˆ. Right: per-token (τ a , τ z ) are drawn under five 1D modes: the four edges of [0, 1]2 (forward, policy, inverse, video) and the joint d… view at source ↗
Figure 2
Figure 2. Figure 2: Same checkpoint, three action sources, three coherent decoded rollouts (Push-T top, Cube bottom; Two-Room and Reacher continued in [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 2
Figure 2. Figure 2: (Continued.) Two-Room (top) and Reacher (bottom) decoded rollouts under the same three [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 1 canonical work pages

  1. [1]

    A. Ajay, Y . Du, A. Gupta, J. Tenenbaum, T. Jaakkola, and P. Agrawal. Is conditional generative modeling all you need for decision-making?International Conference on Learning Represen- tations, 2022. doi: 10.48550/arxiv.2211.15657. URL https://doi.org/10.48550/arxiv. 2211.15657. Decision Diffuser; arXiv:2211.15657

    Pith/arXiv arXiv doi:10.48550/arxiv.2211.15657 2022

  2. [2]

    Albergo, N

    M. Albergo, N. Boffi, and E. Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.Journal of machine learning research, 2023. doi: 10.48550/arxiv.2303. 08797. URLhttps://doi.org/10.48550/arxiv.2303.08797

    Pith/arXiv arXiv doi:10.48550/arxiv.2303 2023

  3. [3]

    Assran, A

    M. Assran, A. Bardes, D. Fan, Q. Garrido, R. Howes, M. Muckley, A. Rizvi, C. Roberts, K. Sinha, A. Zholus, et al. V-JEPA 2: Self-supervised video models enable understanding, prediction and planning.arXiv preprint arXiv:2506.09985, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Balestriero and Y

    R. Balestriero and Y . LeCun. LeJEPA: Provable and scalable self-supervised learning without the heuristics.arXiv preprint arXiv:2511.08544, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    A. Bar, G. Zhou, D. Tran, T. Darrell, and Y . LeCun. Navigation world models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15791–15801, 2025

    2025

  6. [6]

    J. Cen, S. Huang, Y . Yuan, K. Li, H. Yuan, C. Yu, Y . Jiang, J. Guo, X. Li, H. Luo, F. Wang, D. Zhao, and H. Chen. RynnVLA-002: A unified vision-language-action and world model. arXiv preprint arXiv:2511.17502, 2025

    Pith/arXiv arXiv 2025

  7. [7]

    J. Cen, C. Yu, H. Yuan, Y . Jiang, S. Huang, J. Guo, X. Li, Y . Song, H. Luo, F. Wang, D. Zhao, and H. Chen. WorldVLA: Towards autoregressive action world model.arXiv preprint arXiv:2506.21539, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    B. Chen, D. M. Monso, Y . Du, M. Simchowitz, R. Tedrake, and V . Sitz- mann. Diffusion forcing: Next-token prediction meets full-sequence diffusion. In NeurIPS, 2024. URL http://papers.nips.cc/paper_files/paper/2024/hash/ 2aee1c4159e48407d68fe16ae8e6e49e-Abstract-Conference.html

    2024

  9. [9]

    C. Chi, S. Feng, Y . Du, Z. Xu, E. Cousineau, B. Burchfiel, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion. InRobotics: Science and Systems (RSS), 2023

    2023

  10. [10]

    Destrade, O

    M. Destrade, O. Bounou, Q. L. Lidec, J. Ponce, and Y . LeCun. Value-guided action planning with JEPA world models.arXiv preprint arXiv:2601.00844, 2026

    arXiv 2026

  11. [11]

    Y . Du, M. Yang, B. Dai, H. Dai, O. Nachum, J. B. Tenenbaum, D. Schuurmans, and P. Abbeel. Learning universal policies via text-guided video generation. InAdvances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  12. [12]

    D. Gao, B. Zhao, A. Lee, I. Chuang, H. Zhou, H. Wang, Z. Zhao, J. Zhang, and I. Soltani. VITA: Vision-to-action flow matching policy. InICLR, 2026. URL https://openreview. net/forum?id=BTe5VLBjPg

    2026

  13. [13]

    Ghosh, A

    D. Ghosh, A. Gupta, A. Reddy, J. Fu, C. M. Devin, B. Eysenbach, and S. Levine. Learning to reach goals via iterated supervised learning. InICLR, 2021. URL https://openreview. net/forum?id=rALA0Xo6yNJ

    2021

  14. [14]

    R. G. Goswami, A. Bar, D. Fan, T.-Y . Yang, G. Zhou, P. Krishnamurthy, M. Rabbat, F. Khorrami, and Y . LeCun. World models for learning dexterous hand-object interactions from human videos. arXiv preprint arXiv:2512.13644, 2025

    arXiv 2025

  15. [15]

    R. G. Goswami, P. Krishnamurthy, Y . LeCun, and F. Khorrami. OSVI-WM: One-shot visual imitation for unseen tasks using world-model-guided trajectory generation.arXiv preprint arXiv:2505.20425, 2025

    arXiv 2025

  16. [16]

    Y . Guo, Y . Hu, J. Zhang, Y .-J. Wang, X. Chen, C. Lu, and J. Chen. Prediction with action: Visual policy learning via joint denoising process. InAdvances in Neural Information Processing Systems (NeurIPS), 2024. 10

    2024

  17. [17]

    Hafner, W

    D. Hafner, W. Yan, and T. Lillicrap. Training agents inside of scalable world models.arXiv preprint arXiv:2509.24527, 2025

    Pith/arXiv arXiv 2025

  18. [18]

    Hansen, H

    N. Hansen, H. Su, and X. Wang. Td-mpc2: Scalable, robust world models for continuous control. International Conference on Learning Representations, 2023. doi: 10.48550/arxiv.2310.16828. URLhttps://doi.org/10.48550/arxiv.2310.16828

    Pith/arXiv arXiv doi:10.48550/arxiv.2310.16828 2023

  19. [19]

    Hansen, H

    N. Hansen, H. Su, and X. Wang. Learning massively multitask world models for continuous control. InICLR, 2026. URLhttps://openreview.net/forum?id=MPabX9LEds

    2026

  20. [20]

    Janner, Y

    M. Janner, Y . Du, J. Tenenbaum, and S. Levine. Planning with diffusion for flexible behavior synthesis.International Conference on Machine Learning, 2022. doi: 10.48550/arxiv.2205. 09991. URLhttps://doi.org/10.48550/arxiv.2205.09991. arXiv:2205.09991

    Pith/arXiv arXiv doi:10.48550/arxiv.2205 2022

  21. [21]

    M. J. Kim, Y . Gao, T.-Y . Lin, Y .-C. Lin, Y . Ge, G. Lam, P. Liang, S. Song, M.-Y . Liu, C. Finn, and J. Gu. Cosmos policy: Fine-tuning video models for visuomotor control and planning. In Proceedings of the International Conference on Learning Representations (ICLR), 2026

    2026

  22. [22]

    S. Li, Y . Gao, D. Sadigh, and S. Song. Unified video action model. InProceedings of Robotics: Science and Systems (RSS), 2025

    2025

  23. [23]

    Lipman, R

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. InICLR, 2023. URLhttps://openreview.net/forum?id=PqvMRDCJT9t

    2023

  24. [24]

    G.-H. Liu, A. Vahdat, D.-A. Huang, E. A. Theodorou, W. Nie, and A. Anandkumar. I2sb: Image-to-image schrödinger bridge.International Conference on Machine Learning, 2023. doi: 10.48550/arxiv.2302.05872. URLhttps://doi.org/10.48550/arxiv.2302.05872

    Pith/arXiv arXiv doi:10.48550/arxiv.2302.05872 2023

  25. [25]

    K. Liu, W. Hu, J. Xu, Y . Shan, and S. Lu. Rolling forcing: Autoregressive long video diffusion in real time. InInternational Conference on Learning Representations (ICLR), 2026

    2026

  26. [26]

    X. Liu, C. Gong, and Q. Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.International Conference on Learning Representations, 2022. doi: 10.48550/arxiv.2209.03003. URLhttps://doi.org/10.48550/arxiv.2209.03003

    Pith/arXiv arXiv doi:10.48550/arxiv.2209.03003 2022

  27. [27]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Decoupled weight decay regularization. InICLR, 2019. URL https://openreview.net/forum?id=Bkg6RiCqY7

    2019

  28. [28]

    Z. Lyu, M. Li, J. Jiao, and C. Chen. Frame interpolation with consecutive Brownian bridge diffusion. InProceedings of the 32nd ACM International Conference on Multimedia (MM ’24),

  29. [29]

    doi: 10.1145/3664647.3680961

    work page doi:10.1145/3664647.3680961

  30. [30]

    L. Maes, Q. Le Lidec, D. Scieur, Y . LeCun, and R. Balestriero. LeWorldModel: Stable end- to-end joint-embedding predictive architecture from pixels.arXiv preprint arXiv:2603.19312, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    H. Nam, Q. L. Lidec, L. Maes, Y . LeCun, and R. Balestriero. Causal-JEPA: Learning world models through object-level latent interventions.arXiv preprint arXiv:2602.11389, 2026

    Pith/arXiv arXiv 2026

  32. [32]

    S. Park, K. Frans, B. Eysenbach, and S. Levine. Ogbench: Benchmarking offline goal- conditioned rl.International Conference on Learning Representations, 2024. doi: 10.48550/ arxiv.2410.20092. URLhttps://doi.org/10.48550/arxiv.2410.20092

    Pith/arXiv arXiv doi:10.48550/arxiv.2410.20092 2024

  33. [33]

    W. S. Peebles and S. Xie. Scalable diffusion models with transformers.IEEE International Conference on Computer Vision, 2022. doi: 10.1109/iccv51070.2023.00387. URL https: //doi.org/10.1109/iccv51070.2023.00387

    doi:10.1109/iccv51070.2023.00387 2022

  34. [34]

    Perez, F

    E. Perez, F. Strub, H. D. Vries, V . Dumoulin, and A. C. Courville. Film: Visual reasoning with a general conditioning layer.AAAI Conference on Artificial Intelligence, 2017. doi: 10.1609/aaai.v32i1.11671. URLhttps://doi.org/10.1609/aaai.v32i1.11671

    doi:10.1609/aaai.v32i1.11671 2017

  35. [35]

    R. Po, Y . Nitzan, R. Zhang, B. Chen, T. Dao, E. Shechtman, G. Wetzstein, and X. Huang. Long-context state-space video world models. InICCV, 2025. URL https://openaccess.thecvf.com/content/ICCV2025/html/Po_Long-Context_ State-Space_Video_World_Models_ICCV_2025_paper.html. 11

    2025

  36. [36]

    R. Y . Rubinstein and D. P. Kroese.The Cross-Entropy Method: A Unified Approach to Combinatorial Optimization, Monte-Carlo Simulation, and Machine Learning. Information Science and Statistics. Springer, 2004

    2004

  37. [37]

    Sobal, W

    V . Sobal, W. Zhang, K. Cho, R. Balestriero, T. G. J. Rudner, and Y . LeCun. Stress-testing offline reward-free reinforcement learning: A case for planning with latent dynamics models. In7th Robot Learning Workshop: Towards Robots with Human-Level Abilities, 2025. URL https://openreview.net/forum?id=jON7H6A9UU

    2025

  38. [38]

    Tassa, Y

    Y . Tassa, Y . Doron, A. Muldal, T. Erez, Y . Li, D. de Las Casas, D. Budden, A. Abdolmaleki, J. Merel, A. Lefrancq, T. P. Lillicrap, and M. A. Riedmiller. DeepMind control suite.arXiv preprint arXiv:1801.00690, 2018

    Pith/arXiv arXiv 2018

  39. [39]

    Terver, T.-Y

    B. Terver, T.-Y . Yang, J. Ponce, A. Bardes, and Y . LeCun. What drives success in physical planning with joint-embedding predictive world models?arXiv preprint arXiv:2512.24497, 2025

    Pith/arXiv arXiv 2025

  40. [40]

    J. Won, K. Lee, H. Jang, D. Kim, and J. Shin. DUST: Dual-stream diffusion for world-model augmented vision-language-action model.arXiv preprint arXiv:2510.27607, 2025

    Pith/arXiv arXiv 2025

  41. [41]

    H. Wu, Y . Jing, C. Cheang, G. Chen, J. Xu, X. Li, M. Liu, H. Li, and T. Kong. Unleashing large-scale video generative pre-training for visual robot manipulation. InProceedings of the International Conference on Learning Representations (ICLR), 2024

    2024

  42. [42]

    D. Xie, Z. Xu, Y . Hong, H. Tan, D. Liu, F. Liu, A. E. Kaufman, and Y . Zhou. Progressive autoregressive video diffusion models.2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. doi: 10.1109/cvprw67362.2025.00628. URL https://doi.org/10.1109/cvprw67362.2025.00628

    doi:10.1109/cvprw67362.2025.00628 2025

  43. [43]

    G. Zhou, H. Pan, Y . LeCun, and L. Pinto. Dino-wm: World models on pre-trained visual features enable zero-shot planning.International Conference on Machine Learning, 2024. doi: 10.48550/arxiv.2411.04983. URLhttps://doi.org/10.48550/arxiv.2411.04983

    Pith/arXiv arXiv doi:10.48550/arxiv.2411.04983 2024

  44. [44]

    C. Zhu, R. Yu, S. Feng, B. Burchfiel, P. Shah, and A. Gupta. Unified world models: Coupling video and action diffusion for pretraining on large robotic datasets.Robotics, 2025. doi: 10.48550/arxiv.2504.02792. URLhttps://doi.org/10.48550/arxiv.2504.02792

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.02792 2025

  45. [45]

    H. Zhu, M. Zhao, G. He, H. Su, C. Li, and J. Zhu. Causal forcing: Autoregressive diffusion distillation done right for high-quality real-time interactive video generation.arXiv preprint arXiv:2602.02214, 2026. Appendix A Environments and datasets All four envs use continuous action spaces and are goal-conditioned on a future state sampled from the same tr...

    Pith/arXiv arXiv 2026