Pith. sign in

REVIEW 3 major objections 4 minor 27 references

WorldDiT shows that a 399-million-parameter diffusion transformer, trained to both predict action chunks and future camera patches, reaches 94.9% mean LIBERO success without a large pretrained vision-language action backbone.

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 · deepseek-v4-flash

2026-08-03 01:44 UTC pith:AKWEC674

load-bearing objection A genuinely new small architecture for joint action+world modeling, but its headline 94.9% LIBERO mean is not a held-out number — 300 of 500 episodes per suite were used for checkpoint selection — so the Pareto-frontier claim rests on sand. the 3 major comments →

arxiv 2607.23909 v2 pith:AKWEC674 submitted 2026-07-27 cs.LG cs.RO

WorldDiT: A Unified Diffusion Architecture for World and Action Modeling

Sen Wang , R. Gnana Praveen , Bidhan Roy , Marcos Villagra This is my paper
classification cs.LG cs.RO
keywords diffusion transformerflow matchingworld modelrobot manipulationLIBEROvision-language-action modelparameter efficiencyPareto frontier
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.

WorldDiT sets out to show that strong robot control does not require a large pretrained vision-language model as the action backbone. The paper builds a single diffusion transformer that, during training, jointly flow-matches a seven-step continuous action chunk and 128 normalized RGB patch targets sampled from future primary- and wrist-camera frames; at deployment only the action path runs. After separate fine-tuning on each of the four LIBERO suites, the 399-million-parameter model reports 98.0/97.0/92.8/91.8% success (mean 94.9%) and lies on the reported Pareto frontier of parameter count versus mean success among methods reporting all four suites. The authors are careful to frame the result as a baseline and not as evidence of scaling behavior, and they note the aggregate includes checkpoint-selection episodes and that compared methods were not re-run under a shared protocol. If the claim holds, sub-billion-parameter non-VLM policies remain a competitive design point for robot manipulation.

Core claim

The paper claims that a single diffusion transformer can learn continuous action generation together with an auxiliary future-visual-prediction objective, and that this coupling is sufficient to reach strong LIBERO performance without a large pretrained VLM action backbone. Concretely, WorldDiT fine-tuned independently on each of the four LIBERO suites reports 98.0% on Spatial, 97.0% on Object, 92.8% on Goal, and 91.8% on Long, a mean of 94.9%, using 399.084M total parameters (135.107M trainable). The paper positions this point on the reported Pareto frontier for total instantiated parameters versus mean success among methods reporting all four suites, and explicitly frames the result as a b

What carries the argument

The central mechanism is a single flow-matching DiT backbone with two velocity heads operating over a ten-step window: three clean context slots (34 visual tokens from a frozen MAE plus Perceiver resampler, 1 CLIP language token, 1 state token), seven noised action tokens, and 128 normalized RGB patch tokens (64 per camera) from the frame at temporal offset H=7. Block-causal, action-safe attention prevents action queries from reading noised RGB targets. During training, the model regresses straight-path flow velocities for both action and RGB patch targets; at inference the RGB head is dropped and only the action velocity field is integrated over 20 Euler steps, executing the first three of

Load-bearing premise

The Pareto-frontier claim rests on treating the 23 published LIBERO scores as comparable to WorldDiT's own, even though the methods were not re-run under a shared protocol and the 94.9% aggregate includes 300 of 500 episodes per suite used for staged checkpoint selection.

What would settle it

Run WorldDiT and the top comparators from the table under one protocol with identical episode counts, seeds, held-out test sets, and checkpoint selection, then recompute the Pareto frontier. If WorldDiT's mean success falls below that frontier, the paper's central placement is an artifact of pooling incomparable published numbers.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the reported numbers hold, a 399M-parameter policy without a large pretrained VLM action backbone can achieve mean success comparable to much larger VLA policies on LIBERO.
  • The auxiliary RGB patch prediction runs only during training, so the world-modeling objective adds no inference cost or latency to the action path.
  • Because action and visual targets share one flow-matching formulation, the architecture can be decomposed into independently trained experts, including configurations suited to compute-constrained hardware.
  • The result gives the community a sub-billion-parameter baseline that separates the contributions of model scale from the choice of action backbone, which the paper identifies as a key confound in current VLA research.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the RGB patch loss is weighted at 0.001, its main benefit may be as a regularizer or representation shaper rather than as a literal world model; a head-to-head action-only ablation would be the most direct experiment to separate these effects.
  • Editorial inference: the shared flow-matching design suggests the auxiliary target could be swapped from raw normalized RGB patches to latent features or future states, potentially improving sample efficiency without changing the deployment-time action path.
  • Editorial inference: if the Pareto placement survives a controlled re-run under an identical protocol, it would strengthen the case that parameter-efficient policies can be built by coupling small backbones with auxiliary world-model objectives rather than by scaling pretrained VLMs.

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 / 4 minor

Summary. WorldDiT proposes a unified diffusion transformer that jointly flow-matches a 7-step continuous action chunk and 128 normalized RGB patch targets from future camera frames, using frozen MAE/CLIP encoders and a 399.084M-parameter backbone (135.107M trainable). After pretraining on LIBERO-90 and fine-tuning on each of four LIBERO suites, the paper reports 98.0/97.0/92.8/91.8% success (mean 94.9%) and claims this places WorldDiT on the reported parameter-vs-success Pareto frontier among 24 methods, without a large pretrained VLM action backbone. The paper is transparent about a key limitation: Section 3.1.3 states that 300 of the 500 episodes per suite were used for staged checkpoint selection, so the 94.9% aggregate is not a fully held-out estimate.

Significance. If the empirical claim survives re-evaluation, WorldDiT would be a valuable sub-billion-parameter baseline: it directly addresses whether strong LIBERO performance requires a large pretrained VLM action backbone, uses a clean flow-matching formulation, and reports parameter counts in a reproducible way. The paper's explicit limitation disclosure is a strength. However, the central empirical claim and the Pareto-frontier placement currently rest on a number that was partly selected rather than purely evaluated, and on self-reported scores from heterogeneous protocols. The contributions are therefore significant in potential but not yet established.

major comments (3)
  1. [§3.1.3, Table 1, Abstract] The headline 94.9% mean is not an unbiased test estimate. Section 3.1.3 states that the aggregate includes 300 of 500 episodes per suite used during staged checkpoint selection, yet no separate result on the remaining 200 held-out episodes is reported. Because the Pareto-frontier claim in Figure 1 and the Abstract depends on this number, the reported mean is an optimized, selection-inflated estimate. Please re-evaluate the final checkpoints on episodes never used for any selection decision, report the held-out numbers separately, and describe the checkpoint-selection procedure (how many checkpoints, which criterion, how often). Without this, the frontier placement is not supported.
  2. [Table 1, Figure 1, §3.1.3] The comparison against 23 cited methods relies on self-reported success numbers obtained under different protocols, including differing episode counts, seeds, ensembling, and selection rules. The paper acknowledges this in §3.1.3 but still presents the frontier as a headline result without quantifying uncertainty. At minimum, label the comparison as 'reported under native protocols, not directly comparable,' provide a confidence interval for WorldDiT's held-out mean, and state which baselines, if any, were re-run under a shared protocol. This is load-bearing because even a small change in WorldDiT's true held-out mean (or in a baseline's reported number) could move it off the frontier.
  3. [§2.2, §3.1.2, Table 1] There is no ablation that removes the auxiliary RGB-patch loss (i.e., w_rgb = 0). The paper's central conceptual claim is that coupling action generation with future normalized RGB prediction helps or at least does not hurt action learning. Without an action-only baseline, the reported success cannot be attributed to the world-modeling objective; the architecture's contribution is therefore underdetermined. Please add an action-only ablation under the same protocol and report the resulting per-suite and mean success.
minor comments (4)
  1. [§3.1.3] The phrase 'staged checkpoint selection' is undefined. How many checkpoints were considered, at what training epochs, and what criterion selected the final one? This information is needed to judge how much selection could inflate the reported number.
  2. [Figure 3 and §2.1] The caption mentions 'Strict block-causal, action-safe attention,' but the method section does not describe the attention mask. Please specify how action/visual target tokens attend to context and to each other.
  3. [Throughout] Formatting issues: '224by224', '399.084million', '135 .107million', and 'bageldotcom/worlddit' should be cleaned up (e.g., '224×224', '399.084 million', and a full URL or repository identifier).
  4. [Figure 1] Seer is plotted at its Long-only score but excluded from the Pareto calculation. Consider marking excluded points with a distinct symbol so the frontier line is not visually confused with methods that have a full four-suite mean.

Circularity Check

1 steps flagged

Headline 94.9% success and Pareto-frontier claim are not held out: 300 of 500 episodes per suite were used for checkpoint selection.

specific steps
  1. fitted input called prediction [Section 3.1.3 (Evaluation), second paragraph; used again in Section 3.2 and in Figure 1/Table 1 for the Pareto-frontier claim.]
    "For each suite, the reported WorldDiT score aggregates500simulator episodes. The aggregate includes300episodes per suite used during staged checkpoint selection. It is therefore not fully held out, and the reported94.9%should not be interpreted as an unbiased test estimate."

    The headline 94.9% mean and the Pareto-frontier placement are computed from the same 300 episodes per suite that were used to select the checkpoint. The reported success is therefore the selection objective evaluated on the selection data, not an independent estimate of the policy's success. Because the remaining 200 episodes are not reported separately, the abstract's claim that WorldDiT 'lies on the reported Pareto frontier' rests on a number that is, by the paper's own admission, not fully held out. This is statistically forced by construction: selecting a checkpoint to maximize success on a set and then reporting success on that same set guarantees an optimistically biased 'prediction' of generalization.

full rationale

The method-level derivation is self-contained: the flow-matching loss in Section 2.2 is a standard regression against ground-truth action and normalized-RGB-patch targets, and no fitted parameter in the equations is renamed as a prediction. The 'unified world-and-action modeling' claim is a concrete training objective, not a renamed known result. The two self-citations (Paris [9], Paris 2.0 [19]) appear only in the Discussion as a future decomposition direction and are not load-bearing for the reported performance or frontier claim. The Pareto frontier computation itself is a descriptive summary of the table and is not circular. The circularity I find is at the evaluation-to-claim step: the reported 94.9% mean and the resulting Pareto-frontier placement depend on 300 of 500 episodes per suite used for staged checkpoint selection, as the paper explicitly admits. Since the headline claim is that WorldDiT lies on the reported Pareto frontier, and that placement depends on a number that is not fully held out, the central 'prediction' is partly forced by the selection procedure. The remaining 200 episodes per suite are not reported separately, so no unbiased held-out estimate supports the abstract's strong-baseline claim. I also note that the cross-method Pareto comparison uses published numbers without a shared protocol and reconstructed parameter counts; those are external-validity risks rather than constructional circularity. The absence of an action-only ablation for the auxiliary RGB-patch loss is a missing support, not a circular step.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central empirical result rests on a stack of untested choices: the precise loss weighting, window sizes, patch subsampling, and the assumption that the auxiliary RGB objective helps. None of these are derived or ablated. No new theoretical entities are introduced.

free parameters (7)
  • action loss weight w_action = 0.1
    Hand-chosen; no sensitivity analysis reported; balances action and RGB losses in Eq. 5.
  • RGB loss weight w_rgb = 0.001
    Hand-chosen; no sensitivity analysis; central to the auxiliary world-modeling contribution.
  • context length C = 3
    Window size chosen for LIBERO; no study of effect on success.
  • action chunk length H = 7
    Chunk size and execution prefix 3 chosen without ablation.
  • retained RGB patches per camera = 64
    Subsampling of 16x16 patch grid; arbitrary and untested.
  • sampling steps / execution prefix = 20 / 3
    Euler integration steps and receding-horizon prefix chosen by hand; no analysis.
  • backbone capacity = depth 4, hidden 1024, 16 heads, 4 register tokens
    Architecture size determines the 399M parameter count; no scaling or sensitivity study.
axioms (5)
  • domain assumption LIBERO success rate measures manipulation competence relevant to real robots
    Used as the sole evaluation; no real-robot results.
  • domain assumption Published LIBERO scores for 23 methods are accurate and comparable despite differing protocols
    Paper relies on cited reports (Section 3.1.3) without shared-protocol reproduction.
  • ad hoc to paper Auxiliary future RGB-patch prediction improves or does not degrade action learning
    Loss included with weight 0.001; no action-only ablation in Section 3.
  • domain assumption Frozen MAE and CLIP encoders provide sufficient representations for LIBERO
    Encoders are fixed; no study of encoder choice.
  • standard math Flow matching objective (Lipman et al.) is a valid generative training framework
    Adopted from [13]; standard.

pith-pipeline@v1.3.0-alltime-deepseek · 7386 in / 14020 out tokens · 140902 ms · 2026-08-03T01:44:08.753486+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of WorldDiT: A Unified Diffusion Architecture for World and Action Modeling." pith.science (2026). https://pith.science/paper/AKWEC674

@misc{pith2026260723909,
  author       = {Pith},
  title        = {Pith review of: WorldDiT: A Unified Diffusion Architecture for World and Action Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AKWEC674}},
  note         = {Machine review of arXiv:2607.23909}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Many recent robot policies pursue stronger control by using large pretrained vision-language models (VLMs) as the action backbone. We introduce WorldDiT, a unified diffusion transformer architecture that couples action generation with visual world modeling and achieves strong performance without a large pretrained VLM action backbone. During training, a single diffusion transformer generates continuous action chunks and predicts normalized RGB patch targets from future camera frames. Across four LIBERO simulation suites, WorldDiT lies on the reported Pareto frontier for total model parameters and mean success among methods reporting all four suites. These results provide a strong sub-billion-parameter baseline for future scaling studies.

Review history (2 revisions) →

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

27 extracted references · 5 linked inside Pith

  1. [1]

    R., Finn, C., Fusai, N., Galliker, M

    Black, K., Brown, N., Darpinian, J., Dhabalia, K., Driess, D., Esmail, A., Equi, M. R., Finn, C., Fusai, N., Galliker, M. Y., Ghosh, D., Groom, L., Hausman, K., ichter, b., Jakubczak, S., Jones, T., Ke, L., LeBlanc, D., Levine, S., Li-Bell, A., Mothukuri, M., Nair, S., Pertsch, K., Ren, A. Z., Shi, L. X., Smith, L., Springenberg, J. T., Stachowicz, K., Ta...

    2025

  2. [2]

    R., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., Jakubczak, S., Jones, T., Ke, L., Levine, S., Li-Bell, A., Mothukuri, M., Nair, S., Pertsch, K., Shi, L

    Black, K., Brown, N., Driess, D., Esmail, A., Equi, M. R., Finn, C., Fusai, N., Groom, L., Hausman, K., Ichter, B., Jakubczak, S., Jones, T., Ke, L., Levine, S., Li-Bell, A., Mothukuri, M., Nair, S., Pertsch, K., Shi, L. X., Smith, L., Tanner, J., Vuong, Q., Walling, A., Wang, H., and Zhilinsky, U.𝜋0: A Vision-Language-Action Flow Model for General Robot ...

    doi:10.15607/rss.2025.xxi.010 2025

  3. [3]

    WorldVLA: Towards autoregressive action world model.arXiv preprint, 2025

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

    2025

  4. [4]

    Unified diffusion VLA: Vision-language-action model via joint discrete denoising diffusion process.International Conference on Learning Representations, 2026

    Chen, J., Song, W., Ding, P., Zhou, Z., Zhao, H., Tang, F., Wang, D., and Li, H. Unified diffusion VLA: Vision-language-action model via joint discrete denoising diffusion process.International Conference on Learning Representations, 2026

    2026

  5. [5]

    C., and Song, S

    Chi, C., Feng, S., Du, Y., Xu, Z., Cousineau, E., Burchfiel, B. C., and Song, S. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. InProceedings of Robotics: Science and Systems, Daegu, Republic of Korea, July 2023. doi: 10.15607/RSS.2023.XIX.026

    doi:10.15607/rss.2023.xix.026 2023

  6. [6]

    R., Pertsch, K., Black, K., Mees, O., Dasari, S., Hejna, J., Kreiman, T., Xu, C., Luo, J., Tan, Y

    Ghosh, D., Walke, H. R., Pertsch, K., Black, K., Mees, O., Dasari, S., Hejna, J., Kreiman, T., Xu, C., Luo, J., Tan, Y. L., Chen, L. Y., Vuong, Q., Xiao, T., Sanketi, P. R., Sadigh, D., Finn, C., and Levine, S. Octo: An Open-Source Generalist Robot Policy. InProceedings of Robotics: Science and Systems, Delft, Netherlands, July 2024. doi: 10.15607/RSS.2024.XX.090

    doi:10.15607/rss.2024.xx.090 2024

  7. [7]

    VLA-0: Building state-of-the-art vlas with zero modification.arXiv preprint, 2025

    Goyal, A., Hadfield, H., Yang, X., Blukis, V., and Ramos, F. VLA-0: Building state-of-the-art vlas with zero modification.arXiv preprint, 2025

    2025

  8. [8]

    Diffusion transformer policy.arXiv preprint arXiv:2410.15959, 2024

    Hou, Z., Zhang, T., Xiong, Y., Pu, H., Zhao, C., Tong, R., Qiao, Y., Dai, J., and Chen, Y. Diffusion transformer policy.arXiv preprint arXiv:2410.15959, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Paris: A decentralized trained open-weight diffusion model

    Jiang, Z., Seraj, R., Villagra, M., and Roy, B. Paris: A decentralized trained open-weight diffusion model. arXiv preprint arXiv:2510.03434, 2025

    Pith/arXiv arXiv 2025

  10. [10]

    J., Finn, C., and Liang, P

    Kim, M. J., Finn, C., and Liang, P. Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. InProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June 2025. doi: 10.15607/RSS.2025.XXI.017. 8 WorldDiT

    doi:10.15607/rss.2025.xxi.017 2025

  11. [11]

    J., Pertsch, K., Karamcheti, S., Xiao, T., Balakrishna, A., Nair, S., Rafailov, R., Foster, E

    Kim, M. J., Pertsch, K., Karamcheti, S., Xiao, T., Balakrishna, A., Nair, S., Rafailov, R., Foster, E. P., Sanketi, P. R., Vuong, Q., et al. OpenVLA: An open-source vision-language-action model. InConference on Robot Learning, pp. 2679–2713. PMLR, 2025

    2025

  12. [12]

    Discrete diffusion VLA: Bringing discrete diffusion to action decoding in vision-language-action policies

    Liang, Z., Li, Y., Yang, T., Wu, C., Mao, S., Pei, L., Nian, T., Zhou, S., Yang, X., Pang, J., Mu, Y., and Luo, P. Discrete diffusion VLA: Bringing discrete diffusion to action decoding in vision-language-action policies. arXiv preprint arXiv:2508.20072, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  14. [14]

    LIBERO: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776–44791, 2023

    Liu, B., Zhu, Y., Gao, C., Feng, Y., Liu, Q., Zhu, Y., and Stone, P. LIBERO: Benchmarking knowledge transfer for lifelong robot learning.Advances in Neural Information Processing Systems, 36:44776–44791, 2023

    2023

  15. [15]

    Mmada-vla: Large diffusion vision-language-action model with unified multi-modal instruction and generation.arXiv preprint arXiv:2603.25406, 2026

    Liu, Y., Ding, P., Jiang, T., Wang, X., Song, W., Lin, M., Zhao, H., Zhang, H., Zhuang, Z., Zhao, W., et al. Mmada-vla: Large diffusion vision-language-action model with unified multi-modal instruction and generation.arXiv preprint arXiv:2603.25406, 2026

    arXiv 2026

  16. [16]

    GR00T N1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    NVIDIA. GR00T N1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    Pith/arXiv arXiv 2025

  17. [17]

    FAST: Efficient Action Tokenization for Vision-Language-Action Models

    Pertsch, K., Stachowicz, K., Ichter, B., Driess, D., Nair, S., Vuong, Q., Mees, O., Finn, C., and Levine, S. FAST: Efficient Action Tokenization for Vision-Language-Action Models. InProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June 2025. doi: 10.15607/RSS.2025.XXI.012

    doi:10.15607/rss.2025.xxi.012 2025

  18. [18]

    SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Models

    Qu, D., Song, H., Chen, Q., Yao, Y., Ye, X., Gu, J., Wang, Z., Ding, Y., Zhao, B., Wang, D., and Li, X. SpatialVLA: Exploring Spatial Representations for Visual-Language-Action Models. InProceedings of Robotics: Science and Systems, LosAngeles, CA, USA, June 2025. doi: 10.15607/RSS.2025.XXI.011

    doi:10.15607/rss.2025.xxi.011 2025

  19. [19]

    Paris 2.0: A decentralized diffusion model for video generation.arXiv preprint arXiv:2605.26064, 2026

    Rouzbayani, A., Roy, B., Villagra, M., and Jiang, Z. Paris 2.0: A decentralized diffusion model for video generation.arXiv preprint arXiv:2605.26064, 2026

    Pith/arXiv arXiv 2026

  20. [20]

    MemoryVLA: Perceptual-cognitive memory in vision-language-action models for robotic manipulation.International Conference on Learning Representations, 2026

    Shi, H., Xie, B., Liu, Y., Sun, L., Liu, F., Wang, T., Zhou, E., Fan, H., Zhang, X., and Huang, G. MemoryVLA: Perceptual-cognitive memory in vision-language-action models for robotic manipulation.International Conference on Learning Representations, 2026

    2026

  21. [21]

    Predictive inverse dynamics models are scalable learners for robotic manipulation

    Tian, Y., Yang, S., Zeng, J., Wang, P., Lin, D., Dong, H., and Pang, J. Predictive inverse dynamics models are scalable learners for robotic manipulation. InInternational Conference on Learning Representations, 2025

    2025

  22. [22]

    Vla-adapter: An effective paradigm for tiny-scale vision-language-action model.Proceedings of the AAAI Conference on Artificial Intelligence, 2026

    Wang, Y., Ding, P., Li, L., Cui, C., Ge, Z., Tong, X., Song, W., Zhao, H., Zhao, W., Hou, P., et al. Vla-adapter: An effective paradigm for tiny-scale vision-language-action model.Proceedings of the AAAI Conference on Artificial Intelligence, 2026

    2026

  23. [23]

    Wu, X.-M., Fan, B., Liao, K., Jiang, J.-J., Yang, R., Luo, Y., Wu, Z., Zheng, W.-S., and Loy, C. C. VLANeXt: Recipes for building strong vla models. InForty-third International Conference on Machine Learning, 2026

    2026

  24. [24]

    Dreamvla: a vision-language-action model dreamed with comprehensive world knowledge.Advances in Neural Information Processing Systems, 38:24195–24228, 2025

    Zhang, W., Liu, H., Qi, Z., Wang, Y., Yu, X., Zhang, J., Dong, R., He, J., Wang, H., Zhang, Z., et al. Dreamvla: a vision-language-action model dreamed with comprehensive world knowledge.Advances in Neural Information Processing Systems, 38:24195–24228, 2025

    2025

  25. [25]

    J., Fu, Z., Zhang, Z., Wu, Y., Li, Z., Ma, Q., Han, S., Finn, C., Handa, A., Lin, T.-Y., Wetzstein, G., Liu, M.-Y., and Xiang, D

    Zhao, Q., Lu, Y., Kim, M. J., Fu, Z., Zhang, Z., Wu, Y., Li, Z., Ma, Q., Han, S., Finn, C., Handa, A., Lin, T.-Y., Wetzstein, G., Liu, M.-Y., and Xiang, D. Cot-vla: Visual chain-of-thought reasoning for vision-language- action models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1702–1713, June 2025

    2025

  26. [26]

    Acot-vla: Action chain-of-thought for vision- language-action models

    Zhong, L., Liu, Y., Wei, Y., Xiong, Z., Liu, S., and Ren, G. Acot-vla: Action chain-of-thought for vision- language-action models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 8152–8162, 2026

    2026

  27. [27]

    FlowVLA: Visual chain of thought-based motion reasoning for vision-language-action models.arXiv preprint arXiv:2508.18269, 2025

    Zhong, Z., Yan, H., Li, J., Liu, X., Gong, X., Zhang, T., Song, W., Chen, J., Zheng, X., Wang, H., and Li, H. FlowVLA: Visual chain of thought-based motion reasoning for vision-language-action models.arXiv preprint arXiv:2508.18269, 2025. 9

    arXiv 2025