Pith. sign in

REVIEW 4 major objections 6 minor 3 cited by

Think Twice, Act Once: Token-Aware Compression and Action Reuse for Efficient Inference in Vision-Language-Action Models

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read FlashVLA claims that training-free acceleration of vision-language-action models cuts FLOPs by 55.7% and latency by 36.0% with only a 0.7% drop in task success rate.

desk verdict Promising training-free VLA acceleration with solid experiments, but the action-reuse trigger as written is inverted—a load-bearing flaw that must be fixed before the reported numbers can be trusted. read the letter →

arxiv 2505.21200 v1 pith:6AF7F2KU submitted 2025-05-27 cs.CV

classification cs.CV
keywords Vision-Language-ActionmodelsinferenceaccelerationtokenpruningactionreuseinformationcontributionscoreFlashAttentionLIBERObenchmarktraining-free
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

The paper claims that a large share of VLA inference cost is wasted on two redundancies: consecutive action outputs barely change during stable phases, and many visual tokens carry little information. It introduces FlashVLA, a training-free wrapper that decides at each step whether to reuse the previous action or run a pruned inference on a selected subset of visual tokens. On the LIBERO benchmark, FlashVLA reports 55.7% lower visual-token FLOPs and 36.0% lower latency at 160 of the original 256 visual tokens, with average task success falling from 74.4% to 73.7%. The payoff, if true, is that generalist robot policies can be made substantially cheaper to run without retraining, architectural change, or specialized kernels. The method works on FlashAttention-based VLA backbones, which standard attention-score pruning cannot touch.

What carries the argument

The load-bearing object is the information contribution score (ICS), $C(x)=\sum_{i=1}^{r}|u_{xi}\sigma_i|$, computed from the singular value decomposition of the visual-token attention output matrix; tokens with the largest ICS are kept because they carry the most energy in the dominant singular directions. The second mechanism is FlashTrigger, which combines action memory and token memory: it computes the angle $\alpha(s)$ between the previous two action vectors and the intersection ratio $\phi(s)$ of the previous two selected token sets, and reuses the previous action when $\alpha(s)>\varepsilon_1$ and $\phi(s)>\varepsilon_2$. Together the two mechanisms route each step either to direct reuse or to pruned inference, so compute is saved both across steps and inside a step, all without modifying the underlying VLA weights.

What would settle it

Run FlashVLA with 160 visual tokens on a LIBERO-Spatial episode in which the object to be grasped is small and located in a low-ICS image patch, meaning it has low $|u_{xi}\sigma_i|$ energy in the attention-output SVD, and compare success rate against the full 256-token baseline over many seeds. If the policy fails specifically on such episodes while total retained Frobenius energy remains high, then the information contribution score is not preserving the control-relevant features.

Watch

Extended reading notes

Core claim

The central claim is that VLA inference contains two independent, exploitable redundancies: temporal redundancy, because consecutive action outputs are often nearly identical during stable manipulation phases, and token redundancy, because many visual tokens contribute little to the representation once attention becomes sparse in deeper layers. FlashVLA attacks both with a dual-path framework. A FlashTrigger module keeps a two-frame memory of action vectors and selected token sets, and reuses the previous action when the angle between the two previous actions is small and the two previous token selections largely overlap. When reuse is not triggered, a pruned inference path ranks visual tokens by an information contribution score defined through the SVD of the attention output matrix and keeps only the top-K tokens. At the 160-token setting on LIBERO, the paper reports 55.7% lower visual-token FLOPs and 36.0% lower wall-clock latency while average success rate moves from 74.4% to 73.7%, with slight gains on two task suites and a larger drop on the long-horizon suite.

Load-bearing premise

The claim rests on the assumption that ranking visual tokens by their SVD-based information contribution score keeps the tokens the controller actually needs, not just the tokens that carry the most linear-representation energy.

Editorial extensions

If this is right

  • An OpenVLA-style policy can be deployed at 36% lower latency with roughly one percentage point of average success-rate change, which is the practical difference between usable and unusable for high-frequency control loops.
  • Because selection operates on the attention output matrix rather than attention scores, the pruning path stays compatible with FlashAttention, so the speedup transfers to modern optimized VLA backbones without recompiling attention kernels.
  • The savings compound with architectural optimizations: action reuse skips full decoding on stable steps and token pruning shrinks the prefill stage, so the framework remains orthogonal to quantization, low-rank adaptation, and parallel decoding.
  • The $δ$ threshold in the token-stability criterion makes the accuracy-efficiency trade-off explicit and tunable: larger $δ$ permits more scene change in the token set and reuses more aggressively, trading success rate for compute.
  • Moderate pruning can slightly improve success on some LIBERO suites, suggesting that removing low-information tokens may stabilize the policy as well as speed it up.

Reading between the lines

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

  • A natural extension the paper does not test is to weight the information contribution score by the gradient of the action loss with respect to each visual token; tokens with low representation energy but high action-sensitivity would then survive pruning, addressing the weakest assumption directly.
  • The binary reuse decision could be relaxed to a predictive or interpolated action, such as a small linear model over the action memory, cutting latency further in smooth segments while still reacting within a step when the trigger fires.
  • Diffusion-policy VLA models, which generate actions by iterative denoising rather than autoregressive decoding, are a plausible target: token pruning would apply to visual inputs across denoising steps, and the reuse trigger could decide whether to warm-start the next denoising trajectory from the previous one.
  • Real-robot deployment would likely need task-adaptive thresholds, since $δ$ sensitivity in the paper suggests that scene-change dynamics differ across task suites; a policy that monitors task progress could set $ε_2$ dynamically.
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

4 major / 6 minor

Summary. FlashVLA is a training-free, plug-and-play inference acceleration framework for vision-language-action (VLA) models. It combines two mechanisms: an information-contribution-score-based visual token pruning strategy that selects a subset of visual tokens via SVD of the attention output matrix, and a token-aware action reuse mechanism that decides whether to skip decoding and reuse the previous action based on the angle between recent action vectors and the overlap of recent selected token sets. Experiments on OpenVLA over LIBERO task suites report a 55.7% reduction in visual-token FLOPs and 36.0% latency reduction with a 0.7% average success-rate drop at the 160-token configuration, alongside ablations and a comparison to FastV.

Significance. If the reported results hold, FlashVLA would be a practically valuable contribution: it is training-free, compatible with FlashAttention-based VLA backbones, and demonstrates substantial efficiency gains on a standard benchmark with limited success-rate degradation. The paper also provides ablations that isolate the contributions of token pruning and action reuse. However, the core action-reuse trigger as written is internally inconsistent with its own motivation, and the theoretical justification for the token-selection criterion is incomplete. These issues must be resolved before the empirical claims can be taken as evidence for the proposed mechanism.

major comments (4)
  1. [§3.3, Eq. (10) and Algorithm 1] The action-reuse trigger condition appears to be inverted relative to the paper's own motivation. Equation (7) defines α(s) as the angle between the previous two action vectors, with small α indicating similar consecutive actions; the text and Figure 1 state that stable actions (small α) should be reused. However, Eq. (10) triggers "Reuse Action" when α(s) > ε1 and ϕ(s) > ε2, i.e., reuse is applied only when consecutive actions are dissimilar. With the reported ε1 = 2 (radians), the mechanism would skip reuse during stable phases and only reuse during transitions, which cannot produce the reported FLOPs and latency savings attributed to action reuse. Algorithm 1 inherits this condition verbatim. This is a load-bearing inconsistency: either the implementation used the opposite inequality (α < ε1) and the text needs correction, or the reported efficiency gains are unexplained. Please correct Eq. (10) and Algorithm 1, or clarify the relation between the trigger condition and the stability motivation.
  2. [§3.2, Eqs. (3)–(6)] The theoretical justification for ICS-based token selection is not sufficient. The score C(x) is defined with absolute values, while the retained-information measure I(S) uses squares. The Cauchy–Schwarz inequality in Eq. (4) gives a lower bound on the squared energy from C(x)^2, but ranking by C(x) does not necessarily maximize the sum of squares in Eq. (5); a token with moderately high absolute projections on many directions can have a larger C(x) than a token with a very large projection on one dominant direction. Moreover, Eq. (5) sums over i=1..k while Eq. (3) uses i=1..r, and the claim that top-K tokens by C(x) "dominate the global sum" does not follow from the definitions. The statement I(S_C) ≥ E[I(S_rand)] is therefore not established. Please provide a correct bound or present the energy-retention comparison as an empirical observation rather than a theorem.
  3. [§4.1 and §4.2, Tables 1 and 2] The headline efficiency claim conflates visual-token FLOPs with total model FLOPs. The Abstract and Conclusion state "reduces FLOPs by 55.7%," but Table 1 and the contribution list specify "visual-token FLOPs" (e.g., the 55.7% reduction is from 1.31 to 0.58 ×10^12). Since the method does not prune language tokens or reduce LLM decoding cost on reused steps' compute for pruned inference, the total-FLOPs reduction is smaller. Please state the metric consistently as visual-token FLOPs or report total FLOPs.
  4. [§4.1 and §4.3] The choice of δ per token budget, e.g., (192,3), (160,4.5), (128,5), (96,5.5), and the subsequent promotion of the 160-token setting as "the best trade-off" introduce a selection-bias risk. These values are hand-assigned after observing performance on the same benchmark, so the reported 0.7% drop may not hold for a pre-specified configuration. Please report results with a fixed selection rule or provide a validation-based procedure for choosing δ, and discuss sensitivity to this choice beyond the two curves in Figure 6.
minor comments (6)
  1. [§3.2, Eq. (5)] The summation upper limit in Eq. (5) is written as k, but the retained information in Eq. (3) uses r. Please unify the notation.
  2. [§3.3 and §4.3] The threshold ε1 is set to 2 without specifying units; since α is an angle in radians, please clarify the unit and the meaning of the "lower limit" phrase. Also, Figure 6(a) shows that ε1 has negligible effect, which is surprising if reuse is triggered by α > ε1; please comment on why this is consistent with the corrected or intended trigger.
  3. [References] References [7] and [8] are the same paper (FastV, Chen et al., ECCV 2024) and appear twice; please deduplicate.
  4. [§4.5, Table 4] The VLAbench generalization result is based on a single task with a 7.0% baseline success rate, which is too low to support strong claims about generalization. Please either add more tasks or soften the conclusion regarding VLAbench.
  5. [Appendix A, Algorithm 1] The condition "last reuse is False" in line 9 prevents consecutive reuse, but the rationale for forbidding consecutive reuse is not explained. If consecutive reuse were allowed, the efficiency gains might be larger but stability might suffer; please clarify the design choice.
  6. [Appendix B, Eq. (11)] The text says that during decoding FlashVLA sets Lp = 0, but the FLOPs formula and the reported numbers use Lp = 2 for both prefill and decoding. Please clarify which Lp value was used to produce Table 1 and how the claim that actual FLOPs are lower than reported is consistent with the table.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: the reported FLOPs/latency/accuracy gains are measured outcomes on LIBERO/VLAbench, not derived from fitted constants; the only author-overlapping citations are related-work context, and the remaining concerns (Eq. 10 sign, ICS proof gap) are correctness issues rather than circular reductions.

full rationale

The paper's central claims are empirical measurements of a training-free acceleration scheme, not predictions recovered from fitted parameters. The 55.7% FLOPs reduction and 36.0% latency reduction are direct consequences of skipping computation (action reuse) and reducing visual tokens, with success rates reported from the benchmark; no equation in Section 3 is used to manufacture these numbers. The two references with author overlap ([9] LL3DA and [17] MotionGPT, both involving T. Chen) appear only in the related-work enumeration and are not load-bearing for the proposed mechanism. The Section 3.2 'theoretical justification' is a self-contained (though mathematically non-rigorous) comparison between the ICS score C(x) and the retention measure I(S) defined on the same SVD; it does not fit a parameter and rename it a prediction, so its weakness is a validity/correctness concern rather than circularity. The Eq. 10 reuse condition (alpha > epsilon_1) appears inconsistent with the 'stable actions' motivation in Fig. 1, but that is an internal sign/correctness inconsistency, not a derivation that reduces to its inputs. No self-citation chain, uniqueness theorem, or ansatz-imported-via-citation is used to force the central claim. Accordingly, no significant circularity is present; the score reflects only the minor, non-load-bearing author-overlapping citations.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central efficiency claims rest on two empirical observations (action stability and visual token redundancy) and two design choices (SVD-based scoring, threshold-based trigger). The scoring heuristic is supported only by an energy-retention inequality vs random selection, not by a task-level guarantee. The reuse trigger depends on two hand-set hyperparameters, and the paper does not report the resulting reuse rate. No code or data is provided.

free parameters (3)
  • ε1 (action angle threshold) = 2 (units unclear, likely degrees in Eq. 7)
    Threshold for the action angle α in the reuse trigger (Eq. 10). The paper fixes it 'unless otherwise specified' without a search procedure; the sensitivity analysis (Fig. 6a) shows it has 'negligible effect'.
  • δ (max token-set changes) = 3, 4.5, 5, 5.5 for token budgets 192, 160, 128, 96
    Maximum allowed number of token-set changes in Eq. 9; hand-assigned per token budget in Section 4.1 to balance accuracy and efficiency. This is a fitted hyperparameter that directly controls reuse rate.
  • effective rank r (in SVD) = Not specified
    The ICS definition sums over i=1..r (Eq. 2), but the paper never states how r is chosen (e.g., energy threshold or full rank). This affects the score and is a missing free parameter.
assumptions (4)
  • domain assumption Visual token redundancy exists in VLA models, with attention becoming sparse after layer 2.
    Motivates token pruning; supported by Appendix C figures but only for OpenVLA on representative inputs.
  • domain assumption Consecutive action steps are often similar enough that skipping the model and reusing the previous action preserves task success.
    Motivates action reuse; based on Fig. 1 for OpenVLA on LIBERO.
  • ad hoc to paper The attention output matrix retains enough information about token relevance that its SVD-based score C(x) is a valid pruning criterion.
    Introduced specifically for this method; only argument is the energy-retention inequality vs random selection (Section 3.2).
  • domain assumption Flash Attention makes attention-score-based token selection infeasible, requiring a post-attention approach.
    Stated in Section 3.2, borrowing from FastV [7]; not formally proven but widely accepted.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Think Twice, Act Once: Token-Aware Compression and Action Reuse for Efficient Inference in Vision-Language-Action Models." pith.science (2026). https://pith.science/paper/6AF7F2KU

@misc{pith2026250521200,
  author       = {Pith},
  title        = {Pith review of: Think Twice, Act Once: Token-Aware Compression and Action Reuse for Efficient Inference in Vision-Language-Action Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6AF7F2KU}},
  note         = {Machine review of arXiv:2505.21200}
}
read the original abstract

Vision-Language-Action (VLA) models have emerged as a powerful paradigm for general-purpose robot control through natural language instructions. However, their high inference cost-stemming from large-scale token computation and autoregressive decoding-poses significant challenges for real-time deployment and edge applications. While prior work has primarily focused on architectural optimization, we take a different perspective by identifying a dual form of redundancy in VLA models: (i) high similarity across consecutive action steps, and (ii) substantial redundancy in visual tokens. Motivated by these observations, we propose FlashVLA, the first training-free and plug-and-play acceleration framework that enables action reuse in VLA models. FlashVLA improves inference efficiency through a token-aware action reuse mechanism that avoids redundant decoding across stable action steps, and an information-guided visual token selection strategy that prunes low-contribution tokens. Extensive experiments on the LIBERO benchmark show that FlashVLA reduces FLOPs by 55.7% and latency by 36.0%, with only a 0.7% drop in task success rate. These results demonstrate the effectiveness of FlashVLA in enabling lightweight, low-latency VLA inference without retraining.

Figures

Figures reproduced from arXiv: 2505.21200 by the authors.

Figure 1
Figure 1. Motivation behind our proposed FLASHVLA. The figure shows the change in the VLA model’s output vector at each time step relative to the previous one. The vertical axis indicates the directional difference between consecutive actions. Most actions remain highly consistent with the previous step and appear in the stable area of the figure, while only a few exhibit significant changes. this, recent work focuses on arch… view at source ↗
Figure 2
Figure 2. Framework of our FLASHVLA. We give the way our method works as the action step changes. Before each inference, FlashTrigger will think about whether it can reuse the output of the previous action based on action memory and token memory (as shown in blue block). If the trigger condition is met, this inference is skipped. If the trigger condition is not met, proceed to the pruned inference step. In pruned inference st… view at source ↗
Figure 3
Figure 3. Comparison of visual token selection strategies on a sample image. Left: patches selected using the proposed ICS. Right: patches selected uniformly at random. Patches selected by ICS tend to focus on semantically meaningful and information-dense regions. Theoretical Justification. To theoretically justify the superiority of ICS-based token selection over random sampling, we analyze the information retention in the t… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: FLOPs breakdown of FlashVLA across four LIBERO tasks under different visual token [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Trajectory of action. We visualize the trajectory of action in 3-dimensional space. The location of red dashed box illustrates the smoother trajectory of FLASHVLA for the same task. a b [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 7
Figure 7. Figure 7: Attention Map: Layer-wise attention map visualizations in VLA and VLM models. Both models exhibit uniform attention distribution in the first layer, while attention becomes increasingly sparse from the second layer onward. This pattern suggests growing redundancy in to…
Figure 8
Figure 8. Figure 8: Attention Score: Attention score distributions across transformer layers in a VLA model. The scores are computed by averaging attention weights over heads and selecting the attention received by each token from the final query position. The results show increasing spar…
Figure 9
Figure 9. Figure 9: Sample Frame of Four Main Task Suites. down complex goals into manageable sub - goals and accomplish them step by step. The evaluation aims to assess the robot’s ability to transfer and apply goal - oriented knowledge effectively. LIBERO-Long It has 10 tasks primarily …

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. TIDAL: Temporally Interleaved Diffusion and Action Loop for High-Frequency VLA Control

    cs.RO 2026-01 conditional novelty 6.0 of 10

    TIDAL raises VLA control feedback from ~2.4 Hz to ~9 Hz by caching semantic intent and interleaving one-step flow generation with execution, doubling dynamic interception success in simulation.

  2. Efficient Vision-Language-Action Models for Embodied Manipulation: A Systematic Survey

    cs.RO 2025-10 conditional novelty 4.0 of 10

    A survey that groups VLA efficiency techniques into four categories: model architecture, perception features, action generation, and training/inference strategies.

  3. RLRC: Reinforcement Learning-based Recovery for Compressed Vision-Language-Action Models

    cs.RO 2025-06

Reference graph

Works this paper leans on

47 extracted references · 10 canonical work pages · cited by 3 Pith papers

  1. [1]

    M. Ahn, A. Brohan, N. Brown, Y . Chebotar, O. Cortes, B. David, C. Finn, C. Fu, K. Gopalakrishnan, K. Hausman, et al. Do as i can, not as i say: Grounding language in robotic affordances.arXiv preprint arXiv:2204.01691, 2022

  2. [2]

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities.arXiv preprint arXiv:2308.12966, 1(2):3, 2023

  3. [3]

    Black, N

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichter, et al. π0: A vision-language-action flow model for general robot control, 2024.URL https://arxiv. org/abs/2410.24164

  4. [4]

    Brohan, N

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, J. Dabis, C. Finn, K. Gopalakrishnan, K. Hausman, A. Herzog, J. Hsu, et al. Rt-1: Robotics transformer for real-world control at scale.arXiv preprint arXiv:2212.06817, 2022

  5. [5]

    Cheang, G

    C.-L. Cheang, G. Chen, Y . Jing, T. Kong, H. Li, Y . Li, Y . Liu, H. Wu, J. Xu, Y . Yang, et al. Gr-2: A generative video-language-action model with web-scale knowledge for robot manipulation.arXiv preprint arXiv:2410.06158, 2024

  6. [6]

    J. Chen, L. Ye, J. He, Z.-Y . Wang, D. Khashabi, and A. Yuille. Efficient large multi-modal models via visual context compression. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

  7. [8]

    L. Chen, H. Zhao, T. Liu, S. Bai, J. Lin, C. Zhou, and B. Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. InEuropean Conference on Computer Vision, pages 19–35. Springer, 2024

  8. [9]

    S. Chen, X. Chen, C. Zhang, M. Li, G. Yu, H. Fei, H. Zhu, J. Fan, and T. Chen. Ll3da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26428–26438, 2024

Show all 47 references
  1. [10]

    X. Chen, B. Jiang, W. Liu, Z. Huang, B. Fu, T. Chen, and G. Yu. Executing your commands via motion diffusion in latent space. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18000–18010, 2023

  2. [11]

    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, page 02783649241273668, 2023

  3. [12]

    T. Dao, D. Fu, S. Ermon, A. Rudra, and C. Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness.Advances in neural information processing systems, 35:16344–16359, 2022

  4. [13]

    J. Duan, W. Yuan, W. Pumacay, Y . R. Wang, K. Ehsani, D. Fox, and R. Krishna. Manipulate-anything: Automating real-world robots using vision-language models.arXiv preprint arXiv:2406.18915, 2024

  5. [14]

    H.-S. Fang, H. Fang, Z. Tang, J. Liu, C. Wang, J. Wang, H. Zhu, and C. Lu. Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot.arXiv preprint arXiv:2307.00595, 2023. 10

  6. [15]

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

  7. [16]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, W. Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

  8. [17]

    Jiang, X

    B. Jiang, X. Chen, W. Liu, J. Yu, G. Yu, and T. Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067–20079, 2023

  9. [18]

    Khazatsky, K

    A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, S. Karamcheti, S. Nasiriany, M. K. Srirama, L. Y . Chen, K. Ellis, et al. Droid: A large-scale in-the-wild robot manipulation dataset.arXiv preprint arXiv:2403.12945, 2024

  10. [19]

    M. J. Kim, C. Finn, and P. Liang. Fine-tuning vision-language-action models: Optimizing speed and success.arXiv preprint arXiv:2502.19645, 2025

  11. [20]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

  12. [21]

    K. Y . Li, S. Goyal, J. D. Semedo, and J. Z. Kolter. Inference optimal vlms need only one visual token but larger models.arXiv preprint arXiv:2411.03312, 2024

  13. [22]

    Q. Li, Y . Liang, Z. Wang, L. Luo, X. Chen, M. Liao, F. Wei, Y . Deng, S. Xu, Y . Zhang, et al. Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation. arXiv preprint arXiv:2411.19650, 2024

  14. [23]

    X. Li, K. Hsu, J. Gu, K. Pertsch, O. Mees, H. R. Walke, C. Fu, I. Lunawat, I. Sieh, S. Kirmani, et al. Evaluating real-world robot manipulation policies in simulation.arXiv preprint arXiv:2405.05941, 2024

  15. [24]

    X. Li, M. Liu, H. Zhang, C. Yu, J. Xu, H. Wu, C. Cheang, Y . Jing, W. Zhang, H. Liu, et al. Vision-language foundation models as effective robot imitators.arXiv preprint arXiv:2311.01378, 2023

  16. [25]

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

  17. [26]

    J. Liu, M. Liu, Z. Wang, L. Lee, K. Zhou, P. An, S. Yang, R. Zhang, Y . Guo, and S. Zhang. Robo- mamba: Multimodal state space model for efficient robot reasoning and manipulation.arXiv preprint arXiv:2406.04339, 2024

  18. [27]

    S. Liu, L. Wu, B. Li, H. Tan, H. Chen, Z. Wang, K. Xu, H. Su, and J. Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation.arXiv preprint arXiv:2410.07864, 2024

  19. [28]

    T. Liu, L. Shi, R. Hong, Y . Hu, Q. Yin, and L. Zhang. Multi-stage vision token dropping: Towards efficient multimodal large language model.arXiv preprint arXiv:2411.10803, 2024

  20. [29]

    Y . Liu, J. I. Hamid, A. Xie, Y . Lee, M. Du, and C. Finn. Bidirectional decoding: Improving action chunking via closed-loop resampling.arXiv preprint arXiv:2408.17355, 2024

  21. [30]

    S. Nair, A. Rajeswaran, V . Kumar, C. Finn, and A. Gupta. R3m: A universal visual representation for robot manipulation.arXiv preprint arXiv:2203.12601, 2022

  22. [31]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, R. Howes, P.-Y . Huang, H. Xu, V . Sharma, S.-W. Li, W. Galuba, M. Rabbat, M. Assran, N. Ballas, G. Synnaeve, I. Misra, H. Jegou, J. Mairal, P. Labatut,...

  23. [32]

    O’Neill, A

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jain, et al. Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0. In2024 IEEE International Conference on Robotics and Automati...

  24. [33]

    S. Park, H. Kim, W. Jeon, J. Yang, B. Jeon, Y . Oh, and J. Choi. Quantization-aware imitation-learning for resource-efficient robotic control.arXiv preprint arXiv:2412.01034, 2024

  25. [34]

    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.arXiv preprint arXiv:2501.09747, 2025. 11

  26. [35]

    Singh, V

    I. Singh, V . Blukis, A. Mousavian, A. Goyal, D. Xu, J. Tremblay, D. Fox, J. Thomason, and A. Garg. Progprompt: Generating situated robot task plans using large language models. In2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523–11530. IEEE, 2023

  27. [36]

    W. Song, J. Chen, P. Ding, H. Zhao, W. Zhao, Z. Zhong, Z. Ge, J. Ma, and H. Li. Accelerating vision-language-action model integrated with action chunking via parallel decoding.arXiv preprint arXiv:2503.02310, 2025

  28. [37]

    Touvron, T

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, et al. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971, 2023

  29. [38]

    J. Wen, M. Zhu, Y . Zhu, Z. Tang, J. Li, Z. Zhou, C. Li, X. Liu, Y . Peng, C. Shen, et al. Diffusion-vla: Scaling robot foundation models via unified diffusion and autoregression.arXiv preprint arXiv:2412.03293, 2024

  30. [39]

    J. Wen, Y . Zhu, J. Li, M. Zhu, Z. Tang, K. Wu, Z. Xu, N. Liu, R. Cheng, C. Shen, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation.IEEE Robotics and Automation Letters, 2025

  31. [40]

    S. Xu, Y . Wang, C. Xia, D. Zhu, T. Huang, and C. Xu. Vla-cache: Towards efficient vision-language-action model via adaptive token caching in robotic manipulation.arXiv preprint arXiv:2502.02175, 2025

  32. [41]

    Yan, Y .-H

    G. Yan, Y .-H. Wu, and X. Wang. Dnact: Diffusion guided multi-task 3d policy learning.arXiv preprint arXiv:2403.04115, 2024

  33. [42]

    S. Yang, Y . Chen, Z. Tian, C. Wang, J. Li, B. Yu, and J. Jia. Visionzip: Longer is better but not necessary in vision language models.arXiv preprint arXiv:2412.04467, 2024

  34. [43]

    Y . Yue, Y . Wang, B. Kang, Y . Han, S. Wang, S. Song, J. Feng, and G. Huang. Deer-vla: Dynamic inference of multimodal large language models for efficient robot execution.Advances in Neural Information Processing Systems, 37:56619–56643, 2024

  35. [44]

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975–11986, 2023

  36. [45]

    Zhang, Q

    S. Zhang, Q. Fang, Z. Yang, and Y . Feng. Llava-mini: Efficient image and video large multimodal models with one vision token.arXiv preprint arXiv:2501.03895, 2025

  37. [46]

    Zhang, Z

    S. Zhang, Z. Xu, P. Liu, X. Yu, Y . Li, Q. Gao, Z. Fei, Z. Yin, Z. Wu, Y .-G. Jiang, and X. Qiu. Vlabench: A large-scale benchmark for language-conditioned robotics manipulation with long-horizon reasoning tasks, 2024

  38. [47]

    Zhang, C.-K

    Y . Zhang, C.-K. Fan, J. Ma, W. Zheng, T. Huang, K. Cheng, D. Gudovskiy, T. Okuno, Y . Nakata, K. Keutzer, et al. Sparsevlm: Visual token sparsification for efficient vision-language model inference.arXiv preprint arXiv:2410.04417, 2024

  39. [48]

    Zitkovich, T

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, et al. Rt- 2: Vision-language-action models transfer web knowledge to robotic control. InConference on Robot Learning, pages 2165–2183. PMLR, 2023. 12 Appendix for FLASHVLA A Algorithm...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.