pith. sign in

arxiv: 2606.05737 · v1 · pith:NZBYWS42new · submitted 2026-06-04 · 💻 cs.CV · cs.AI· cs.LG· cs.RO

Let It Be Simple: One-Step Action Generation for Vision-Language-Action Models

Yitong Chen , Shiduo Zhang , Jingjing Gong , Xipeng Qiu This is my paper

Pith reviewed 2026-06-28 03:08 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.LGcs.RO
keywords vision-language-actiondiffusion modelsone-step generationaction chunkingLIBERO benchmarkrobot policies
0
0 comments X

The pith

Simple bias toward high-noise timesteps in diffusion training enables one-step VLA policies to match or exceed ten-step decoding on standard benchmarks.

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

The paper establishes that diffusion-based vision-language-action models do not require advanced one-step techniques from image synthesis because their conditioning is rich while their targets are compact action chunks. By keeping standard velocity prediction and only biasing the training time distribution toward high-noise states, one-step policies reach performance parity with ten-step decoding under identical recipes. On the LIBERO suite this one-step approach even surpasses ten-step policies trained with uniform time sampling, and a 1.4B-scale model attains 95.6 percent on the long-horizon split. The result follows directly from the input-output asymmetry of VLA tasks rather than from added distillation or auxiliary losses.

Core claim

Under the condition-target asymmetry of VLA models, where the policy receives rich multimodal observations yet predicts only a compact low-dimensional action chunk, biasing the diffusion training time distribution toward high-noise timesteps produces one-step generators whose performance matches or exceeds that of ten-step decoding without teacher models, distillation, or extra objectives.

What carries the argument

High-noise biased training time distribution applied to standard velocity-prediction diffusion for action-chunk generation.

If this is right

  • One-step policies match ten-step decoding across LIBERO, LIBERO-Plus, and LIBERO-Pro under the same training recipe.
  • On standard LIBERO the one-step model exceeds the ten-step model trained with uniform time distribution.
  • A 1.4B VLM with 30M action head reaches 95.6 percent success on LIBERO-Long using one-step decoding.
  • Real-robot bimanual evaluation reproduces the same sampler trend across architectures.

Where Pith is reading between the lines

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

  • The same bias schedule could be tested on other conditional generation problems that share rich context but compact output structure.
  • Optimal bias strength may vary with action dimensionality, suggesting a tunable hyperparameter rather than a fixed recipe.
  • If the asymmetry assumption holds, VLA research can focus on conditioning quality rather than on importing full image-generation few-step machinery.

Load-bearing premise

The action output remains a compact low-dimensional chunk even when the conditioning inputs are rich and multimodal.

What would settle it

A controlled experiment in which one-step policies trained with the high-noise bias fall short of ten-step decoding on a task whose action space is higher-dimensional or less structured than the LIBERO chunks.

Figures

Figures reproduced from arXiv: 2606.05737 by Jingjing Gong, Shiduo Zhang, Xipeng Qiu, Yitong Chen.

Figure 1
Figure 1. Figure 1: MNIST grid-to-sequence isolates a rich-condition, compact-target regime. The metric [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Toy diagnostics behind the condition-target view. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: VLA architecture. Image and language tokens are encoded by a vision-language backbone; [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Velocity-field diagnostics along noise-data interpolations, plotted with the common con [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: LIBERO-Plus full-condition sweep. Left: one-step versus ten-step success for comparable [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: LIBERO-Long replanning sensitivity. The tables below report the exact values. Very short [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Velocity-field diagnostics for action-horizon controls. All curves use standard-LIBERO [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Velocity-field diagnostics for condition weakening. All rows use H10, [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: LIBERO-Plus mean success over four suites for one-step and ten-step inference. [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: LIBERO-Plus suite-level success for one-step and ten-step inference. [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
read the original abstract

Diffusion-based vision-language-action (VLA) models often inherit the image-generation view: actions are generated by iterative denoising. We argue that VLA action generation has a different condition-target structure: the policy is conditioned on rich observations, language, and state, but predicts only a compact, low-dimensional action chunk. Under this asymmetry, strong one-step action generation should not necessarily require the advanced one-step methods developed for image synthesis. We keep standard velocity prediction and add no teacher model, distillation stage, or auxiliary objective; in our main recipe, we simply bias the training time distribution toward high-noise states. We first isolate the effect in a controlled MNIST grid-to-sequence task, then test it with extensive robot-policy experiments. Across standard LIBERO, LIBERO-Plus, and LIBERO-Pro, one-step policies trained with high-noise biased schedules generally match ten-step decoding under the same recipe, and on standard LIBERO can exceed ten-step policies trained with a uniform time distribution. A real-robot bimanual YAM RSS evaluation gives a small-sample cross-architecture check of the same sampler trend. On a 1.4B VLM model with a 30M action head, one-step decoding reaches 95.6\% on LIBERO-Long. These results show that strong one-step VLA action generation can emerge from standard diffusion training, without importing the full few-step diffusion machinery developed for image generation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that due to the asymmetry in VLA models (rich conditioning on observations/language/state versus compact low-dimensional action chunks), standard velocity-prediction diffusion training with a high-noise biased timestep distribution enables effective one-step action generation without advanced image-synthesis techniques, teacher models, or distillation. This is isolated on an MNIST grid-to-sequence task and validated empirically on LIBERO, LIBERO-Plus, LIBERO-Pro (one-step matching or exceeding ten-step, e.g. 95.6% on LIBERO-Long), plus a real-robot bimanual YAM RSS check.

Significance. If the results hold, the work shows that VLA policies can achieve strong one-step performance from minimal modifications to standard diffusion training. Credit is due for the controlled isolation experiment, direct comparisons on standard benchmarks without new parameters beyond bias strength, absence of auxiliary objectives, and the real-robot cross-check providing an external validation point.

major comments (2)
  1. [Abstract] Abstract: the claim that one-step policies 'can exceed ten-step policies trained with a uniform time distribution' on standard LIBERO rests on point estimates (e.g., 95.6% on LIBERO-Long) without reported variance, number of runs, or standard deviations; this weakens assessment of whether the exceedance is reliable.
  2. [Experiments] Experiments section: the repeated assertion of results 'under the same recipe' requires explicit confirmation that all hyperparameters except the time-distribution bias (learning rate, optimizer, total steps, batch size, etc.) are identical between one-step and ten-step conditions; without this, the comparison is not fully controlled.
minor comments (2)
  1. [Abstract] Abstract: specify the numerical value or schedule parameters of the 'high-noise bias strength' used in the primary LIBERO experiments to aid reproducibility.
  2. [MNIST task] The MNIST isolation task description should include exact grid dimensions, sequence lengths, and noise schedule details so the controlled effect can be replicated independently.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the positive assessment and for recognizing the value of the controlled MNIST isolation experiment, direct benchmark comparisons, and real-robot validation. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that one-step policies 'can exceed ten-step policies trained with a uniform time distribution' on standard LIBERO rests on point estimates (e.g., 95.6% on LIBERO-Long) without reported variance, number of runs, or standard deviations; this weakens assessment of whether the exceedance is reliable.

    Authors: We agree that variance estimates or multiple runs would allow a stronger statistical assessment of the exceedance. The reported numbers (including 95.6% on LIBERO-Long) are single-run point estimates, which is common in large-scale robotics experiments given the compute cost. We will revise the abstract to qualify the claim by noting that results are from single training runs per condition and will not perform additional runs for this revision. revision: partial

  2. Referee: [Experiments] Experiments section: the repeated assertion of results 'under the same recipe' requires explicit confirmation that all hyperparameters except the time-distribution bias (learning rate, optimizer, total steps, batch size, etc.) are identical between one-step and ten-step conditions; without this, the comparison is not fully controlled.

    Authors: All other hyperparameters are identical by design; the only change is the timestep sampling distribution. We will add an explicit statement in the Experiments section confirming that learning rate, optimizer, total training steps, batch size, model architecture, and all other settings remain unchanged between the one-step and ten-step conditions. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper advances an empirical claim that biasing the training-time distribution toward high-noise timesteps enables one-step velocity-prediction policies to match or exceed multi-step decoding on LIBERO benchmarks. No derivation chain, equations, or fitted-parameter predictions are present; the argument rests on controlled experiments (MNIST grid-to-sequence followed by robot-policy suites) and direct performance comparisons under identical recipes. No self-citations, ansatzes, or uniqueness theorems are invoked as load-bearing premises. The result is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The work relies on standard diffusion assumptions and empirical validation on public benchmarks; the only added choice is the training-time distribution bias.

free parameters (1)
  • high-noise bias strength
    The degree to which the training schedule favors high-noise timesteps is a hyperparameter selected to produce the reported one-step performance.
axioms (1)
  • domain assumption Standard velocity-prediction diffusion process applies to compact action chunks.
    Invoked when retaining velocity prediction without modification.

pith-pipeline@v0.9.1-grok · 5803 in / 1170 out tokens · 43205 ms · 2026-06-28T03:08:17.932019+00:00 · methodology

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

50 extracted references · 16 canonical work pages · 7 internal anchors

  1. [1]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models. InNeurIPS, 2020

    2020

  2. [2]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations. InICLR, 2021

    2021

  3. [3]

    T. X. Pham, K. Zhang, J. W. Hong, and C. D. Yoo. A hidden semantic bottleneck in conditional embeddings of diffusion transformers. InICLR, 2026

    2026

  4. [4]

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever. Consistency models. InICML, 2023

    2023

  5. [5]

    T. Yin, M. Gharbi, R. Zhang, E. Shechtman, F. Durand, W. T. Freeman, and T. Park. One-step diffusion with distribution matching distillation. InCVPR, pages 6613–6623, 2024

    2024

  6. [6]

    Frans, D

    K. Frans, D. Hafner, S. Levine, and P. Abbeel. One step diffusion via shortcut models. InICLR, 2025

    2025

  7. [7]

    Z. Geng, M. Deng, X. Bai, J. Z. Kolter, and K. He. Mean flows for one-step generative modeling. InNeurIPS, 2025. URL https://papers.neurips.cc/paper_files/paper/ 2025/hash/6d13e085b79d454da5910e4ca82a3d9d-Abstract-Conference.html

    2025

  8. [8]

    N. M. Boffi, M. S. Albergo, and E. Vanden-Eijnden. Flow map matching with stochastic interpolants: A mathematical framework for consistency models.Transactions on Machine Learning Research, 2025. URLhttps://openreview.net/forum?id=cqDH0e6ak2

    2025

  9. [9]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. R. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, S. Jakubczak, T. Jones, L. Ke, S. Levine, A. Li-Bell, M. Mothukuri, S. Nair, K. Pertsch, L. X. Shi, L. Smith, J. Tanner, Q. Vuong, A. Walling, H. Wang, and U. Zhilinsky.π0: A vision- language-action flow model for general robot control. InRobotics: Sc...

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

  10. [10]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 44(10–11):1684–1704, 2025. doi:10.1177/02783649241273668

    work page doi:10.1177/02783649241273668 2025

  11. [11]

    Ghosh, H

    Octo Model Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, C. Xu, J. Luo, T. Kreiman, Y . Tan, L. Y . Chen, P. Sanketi, Q. Vuong, T. Xiao, D. Sadigh, C. Finn, and S. Levine. Octo: An open-source generalist robot policy. InRobotics: Science and Systems, Delft, Netherlands, 2024. doi:10.15607/rss.2024.xx.090

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

  12. [12]

    Y . Chen, C. Liang, H. Sui, R. Guo, C. Cheng, J. You, and G. Liu. LangFlow: Continuous diffusion rivals discrete in language modeling. arXiv preprint arXiv:2604.11748, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  13. [13]

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone. LIBERO: Benchmarking knowledge transfer for lifelong robot learning. InNeurIPS Datasets and Benchmarks, pages 44776–44791, 2023

    2023

  14. [14]

    S. Fei, S. Wang, J. Shi, Z. Dai, J. Cai, P. Qian, L. Ji, X. He, S. Zhang, Z. Fei, J. Fu, J. Gong, and X. Qiu. LIBERO-Plus: In-depth robustness analysis of vision-language-action models. arXiv preprint arXiv:2510.13626, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  15. [15]

    X. Zhou, Y . Xu, G. Tie, Y . Chen, G. Zhang, D. Chu, P. Zhou, and L. Sun. LIBERO-PRO: Towards robust and fair evaluation of vision-language-action models beyond memorization. arXiv preprint arXiv:2510.03827, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  16. [16]

    Zhang, Y

    S. Zhang, Y . Wang, H. Chang, H. Zhao, Y . Liu, V . Guizilini, A. Bobu, A. Wagenmaker, A. Dixit, C. Yu, D. Shah, and M. Simchowitz. Post-training for robotics foundation models dataset and challenge. RSS 2026 Workshop & Challenge, 2026. URL https: //posttraining-for-robotics.github.io. 9

    2026

  17. [17]

    Black, N

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

    2025

  18. [18]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Haus- man, A. Herzog, J. Hsu, J. Ibarz, B. Ichter, A. Irpan, T. Jackson, S. Jesmonth, N. J. Joshi, R. Julian, D. Kalashnikov, Y . Kuang, I. Leal, K.-H. Lee, S. Levine, Y . Lu, U. Malla, D. Manju- nath, I. Mordatch, O. Nachum, C. Parada, J. Peralta, E. Perez, K. Pertsc...

  19. [19]

    doi:10.15607/rss.2023.xix.025

    work page doi:10.15607/rss.2023.xix.025 2023

  20. [20]

    Zitkovich, T

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, Q. Vuong, V . Vanhoucke, H. Tran, R. Soricut, A. Singh, J. Singh, P. Sermanet, P. R. Sanketi, G. Salazar, M. S. Ryoo, K. Reymann, K. Rao, K. Pertsch, I. Mordatch, H. Michalewski, Y . Lu, S. Levine, L. Lee, T.-W. E. Lee, I. Leal, Y . Kuang, D. Kalashnikov, R. Julia...

    2023

  21. [21]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn. OpenVLA: An open-source vision-language-action model. InProceedings of The 8th Conference on Robot Learning, volume 270 ofProceedings of Machine Learni...

    2025

  22. [22]

    Pertsch, K

    K. Pertsch, K. Stachowicz, B. Ichter, D. Driess, S. Nair, Q. Vuong, O. Mees, C. Finn, and S. Levine. FAST: Efficient action tokenization for vision-language-action models. InRobotics: Science and Systems, 2025. doi:10.15607/rss.2025.xxi.012

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

  23. [23]

    K. Tian, Y . Jiang, Z. Yuan, B. Peng, and L. Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. InNeurIPS, pages 84839–84865, 2024

    2024

  24. [24]

    L. Yu, Y . Cheng, K. Sohn, J. Lezama, H. Zhang, H. Chang, A. G. Hauptmann, M.-H. Yang, Y . Hao, I. Essa, and L. Jiang. MAGVIT: Masked generative video transformer. InCVPR, pages 10459–10469, 2023

    2023

  25. [25]

    L. Yu, J. Lezama, N. B. Gundavarapu, L. Versari, K. Sohn, D. Minnen, Y . Cheng, A. Gupta, X. Gu, A. G. Hauptmann, B. Gong, M.-H. Yang, I. Essa, D. Ross, and L. Jiang. Language model beats diffusion: Tokenizer is key to visual generation. InICLR, 2024

    2024

  26. [26]

    Y . Liu, S. Zhang, Z. Dong, B. Ye, T. Yuan, X. Yu, L. Yin, C. Lu, J. Shi, L. J.-T. Yu, L. Zheng, J. Gong, T. Jiang, X. Qiu, and H. Zhao. FASTer: Toward powerful and efficient autoregressive vision–language–action models with learnable action tokenizer and block-wise decoding. In ICLR, 2026. URLhttps://openreview.net/forum?id=k6nTUFoqeT

    2026

  27. [27]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. InCVPR, pages 10674–10685, 2022. 10

    2022

  28. [28]

    Greenberg

    O. Greenberg. Demystifying Flux architecture. arXiv preprint arXiv:2507.09595, 2025

    work page arXiv 2025

  29. [29]

    Dieleman

    S. Dieleman. Generative modelling in latent space. Blog post, 2025. URL https://sander. ai/2025/04/15/latents.html

    2025

  30. [30]

    J. Yao, B. Yang, and X. Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. InCVPR, pages 15703–15712, 2025

    2025

  31. [31]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

    NVIDIA, J. Bjorck, F. Casta˜neda, N. Cherniadev, X. Da, R. Ding, L. J. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, J. Jang, Z. Jiang, J. Kautz, K. Kundalia, L. Lao, Z. Li, Z. Lin, K. Lin, G. Liu, E. Llontop, L. Magne, A. Mandlekar, A. Narayan, S. Nasiriany, S. Reed, Y . L. Tan, G. Wang, Z. Wang, J. Wang, Q. Wang, J. Xiang, Y . Xie, Y . Xu, Z. Xu, S. Ye, Z. Yu...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Y . Luo, W. Chen, T. Liang, B. Wang, and Z. Li. SimVLA: A simple VLA baseline for robotic manipulation. arXiv preprint arXiv:2602.18224, 2026

    work page arXiv 2026

  33. [33]

    Y . Wang, P. Ding, L. Li, C. Cui, Z. Ge, X. Tong, W. Song, H. Zhao, W. Zhao, P. Hou, S. Huang, Y . Tang, W. Wang, R. Zhang, J. Liu, and D. Wang. VLA-Adapter: An effective paradigm for tiny-scale vision-language-action model.Proceedings of the AAAI Conference on Artificial Intelligence, 40(22):18638–18646, 2026. doi:10.1609/aaai.v40i22.38931

    work page doi:10.1609/aaai.v40i22.38931 2026

  34. [34]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever. Learning transferable visual models from natural language supervision. InICML, 2021

    2021

  35. [35]

    Raffel, N

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of Machine Learning Research, 2020

    2020

  36. [36]

    W. Kong, Q. Tian, Z. Zhang, R. Min, Z. Dai, J. Zhou, J. Xiong, X. Li, B. Wu, J. Zhang, K. Wu, Q. Lin, J. Yuan, Y . Long, A. Wang, A. Wang, C. Li, D. Huang, F. Yang, H. Tan, H. Wang, J. Song, J. Bai, J. Wu, J. Xue, J. Wang, K. Wang, M. Liu, P. Li, S. Li, W. Wang, W. Yu, X. Deng, Y . Li, Y . Chen, Y . Cui, Y . Peng, Z. Yu, Z. He, Z. Xu, Z. Zhou, Z. Xu, Y . ...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  37. [37]

    X. Liu, C. Gong, and Q. Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InICLR, 2023. URLhttps://openreview.net/forum?id=XVjTT1nw5z

    2023

  38. [38]

    W. Song, J. Chen, P. Ding, Y . Huang, H. Zhao, D. Wang, and H. Li. CEED-VLA: Consistency vision-language-action model with early-exit decoding. arXiv preprint arXiv:2506.13725, 2025

    work page arXiv 2025

  39. [39]

    W. Luan, J. Li, W. Zhao, W. Zhang, T. Wu, and R. Ma. SnapFlow: One-step action generation for flow-matching VLAs via progressive self-distillation. arXiv preprint arXiv:2604.05656, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  40. [40]

    Esser, S

    P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. M¨uller, H. Saini, Y . Levi, D. Lorenz, A. Sauer, F. Boesel, D. Podell, T. Dockhorn, Z. English, and R. Rombach. Scaling rectified flow trans- formers for high-resolution image synthesis. InICML, 2024

    2024

  41. [41]

    Li and K

    T. Li and K. He. Back to basics: Let denoising generative models denoise. InCVPR, pages 36115–36125, 2026

    2026

  42. [42]

    Zheng, N

    B. Zheng, N. Ma, S. Tong, and S. Xie. Diffusion transformers with representation autoencoders. InThe Fourteenth International Conference on Learning Representations, 2026. URLhttps: //openreview.net/forum?id=0u1LigJaab. 11

    2026

  43. [43]

    Lipman, R

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling. InICLR, 2023

    2023

  44. [44]

    M. S. Albergo, N. M. Boffi, and E. Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.Journal of Machine Learning Research, 26(209):1–80,

  45. [45]

    URLhttps://jmlr.org/papers/v26/23-1605.html

  46. [46]

    LeCun, C

    Y . LeCun, C. Cortes, and C. J. C. Burges. The MNIST database of handwritten digits. Website,

  47. [47]

    URLhttps://yann.lecun.com/exdb/mnist/

  48. [48]

    Krizhevsky

    A. Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009

    2009

  49. [49]

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer. Sigmoid loss for language image pre-training. InICCV, pages 11941–11952, 2023

    2023

  50. [50]

    PaliGemma: A versatile 3B VLM for transfer

    L. Beyer, A. Steiner, A. S. Pinto, A. Kolesnikov, X. Wang, D. Salz, M. Neumann, I. Alab- dulmohsin, M. Tschannen, E. Bugliarello, T. Unterthiner, D. Keysers, S. Koppula, F. Liu, A. Grycner, A. Gritsenko, N. Houlsby, M. Kumar, K. Rong, J. Eisenschlos, R. Kabra, M. Bauer, M. Boˇsnjak, X. Chen, M. Minderer, P. V oigtlaender, I. Bica, I. Balazevic, J. Puigcer...

    work page internal anchor Pith review Pith/arXiv arXiv 2024