Pith. sign in

REVIEW 3 major objections 6 minor 6 cited by

EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read EfficientVLA claims a training-free, three-part compression recipe that accelerates the CogACT vision-language-action model 1.93x and cuts FLOPs to 28.9% while losing only 0.6% average task success on SIMPLER.

desk verdict Solid engineering integration with plausible speedups, but the 1.93x headline rests on an undocumented MLP sparsity step and the ablations show layer pruning hurts more than the paper admits. read the letter →

arxiv 2506.10100 v1 pith:SPMSLMFB submitted 2025-06-11 cs.CV

Yantai Yang , Yuhao Wang , Zichen Wen , Luo Zhongwei , Chang Zou , Zhipeng Zhang , Chuan Wen , Linfeng Zhang This is my paper
classification cs.CV
keywords Vision-Language-Actionmodelsinferenceaccelerationlayerpruningvisualtokendiffusioncachingtraining-freecompressionrobotmanipulationCogACT
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

This paper tries to show that a diffusion-based vision-language-action (VLA) model can be made dramatically cheaper at inference time without any retraining. It proposes EfficientVLA, a three-part recipe applied to the CogACT model: prune language-model layers whose input and output hidden states are nearly identical, keep only a task-relevant and diverse subset of visual tokens, and cache intermediate attention and MLP features across diffusion denoising steps. On the SIMPLER robot-manipulation benchmark the combined method reports a 1.93x speedup and reduces FLOPs to 28.9% of the original, with average success rate falling by 0.6 percentage points. If true, this matters because the main obstacle to real-time robot deployment of VLA models is exactly this kind of compute and memory load.

What carries the argument

The load-bearing object is the layer importance score $I^{(\ell)} = 1 - $ average cosine similarity between a layer's input and output hidden states, computed over a representative dataset; layers with the lowest scores are removed regardless of position. For vision, task relevance is scored by accumulated cross-attention from each visual token to language tokens, and a diversity term (cosine distance to the selected key tokens) fills the remaining quota. For action, a static cache interval $N=5$ recomputes attention and MLP features only every fifth denoising step and reuses cached values in between. Together these three mechanisms convert observed redundancy into a single training-free pipeline.

What would settle it

Compare EfficientVLA's layer pruning against random pruning of the same number of non-contiguous layers under identical visual-token and cache settings in SIMPLER. If random pruning matches the reported success rates, then the cosine-similarity importance metric is not responsible for the accuracy retention; if a held-out evaluation set gives a success drop larger than 0.6%, the calibration of the metric to the representative dataset is suspect.

Watch

Extended reading notes

Core claim

EfficientVLA's central claim is that the three dominant sources of redundancy in a diffusion-based VLA--depth-wise redundancy in the LLM, superfluous and overlapping visual tokens, and temporally coherent features across diffusion denoising steps--can be removed in a coordinated, training-free way. The paper measures each layer's importance as one minus the average cosine similarity between its input and output hidden states, prunes the least important non-contiguous layers, and couples this with a visual token selection that anchors on tokens with high cross-attention to the language instruction and then adds diverse tokens. In the diffusion action head it caches attention and MLP outputs and reuses them for four out of every five timesteps. Applied to CogACT in the SIMPLER environment, the framework reports 1.93x faster inference, FLOPs reduced to 28.9%, and only a 0.6% drop in average success rate, with some individual task success rates rising after pruning.

Load-bearing premise

The load-bearing premise is that a language-model layer whose input and output hidden states are highly cosine-similar performs no task-critical transformation, so it can be removed without retraining; if small residual updates carry the information the robot needs, the accuracy claims weaken.

Editorial extensions

If this is right

  • A 22-layer, 56-token configuration reaches 1.93x speedup at 28.9% FLOPs with a 0.6% average success drop, showing that roughly 71% of computation can be removed before task performance degrades noticeably.
  • On the pick coke can task, pruning 36% of parameters raises success rate from 91.3% to 94.0%, indicating that the VLA has substantial parameter slack.
  • Token-only acceleration methods such as FastV and VLA-Cache yield at most 1.38x speedup, while the combined framework yields 1.93x, so coordinated multi-module pruning is necessary to escape memory-bound bottlenecks.
  • The cache interval and token budget provide a tunable accuracy-speed trade-off, letting deployers choose an operating point suited to their latency and compute constraints.

Reading between the lines

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

  • A direct stress test the paper does not run is random layer pruning: comparing random removal of the same number of layers against importance-ranked removal would isolate how much of the accuracy retention the cosine metric actually buys.
  • If the layer importance scores are stable across the four SIMPLER tasks, the same pruning map could be reused at deployment without per-scene calibration, turning the calibration pass into a one-time cost.
  • The diversity term in Eq. (4) is a greedy max-distance heuristic; clustering-based or sampling-based subset selection might reach equal accuracy at even smaller token budgets, which is an empirical question the paper leaves open.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes EfficientVLA, a training-free inference acceleration framework for diffusion-based Vision-Language-Action models. It combines (1) pruning of LLM layers whose input/output hidden states have high cosine similarity, (2) visual token pruning that selects task-relevant and diverse tokens using attention scores and embedding dissimilarity, and (3) periodic caching of attention/MLP features in the diffusion action head. On the CogACT model in the SIMPLER benchmark, the authors report a 1.93x wall-clock speedup, a reduction of FLOPs to 28.9% of baseline, and only a 0.6% drop in average success rate (74.8% to 74.2% on Visual Matching; 61.3% to 61.2% on Variant Aggregation), along with parameter compression from 7.63B to 4.86B. The evaluation includes baselines (Random Dropping, FastV, VLA-Cache), ablations on the 'pick coke can' task, and a three-scale scalability study.

Significance. If the results are confirmed, EfficientVLA is a meaningful contribution: it is one of the few methods that jointly targets the vision, language, and action modules of a VLA without retraining, and the comparison with VLA-Cache and FastV makes the case that module-wise speedups saturate unless the LLM memory bottleneck and action-head computation are addressed. The paper also provides a useful module-wise breakdown of parameters, FLOPs, and latency. The main empirical claims are plausible and internally consistent across Tables 2 and 3, and the ablation in Table 5 shows that each pillar adds to the overall speedup. However, the strength of the evidence is limited by the undocumented MLP sparsity component, the absence of variance reporting, and the lack of a random-layer control. These are addressable in revision.

major comments (3)
  1. [§4.1, Table 5] §4.1 states that EfficientVLA 'adopts the PruneNet [39] configuration for LLM compression,' applying 25% sparsity to the MLP layers, but the paper does not specify how the masks are computed, whether this step is truly training-free, or how 25% unstructured sparsity is converted into wall-clock speedup on the A40. Reference [39] describes a learned pruning policy, so if those masks are used, the 'training-free' characterization in the title and abstract is misleading. Table 5 never isolates the MLP-sparsity component: every condition that includes it also includes layer pruning, so its independent effect on latency and success rate is unknown. Without these details, the headline 1.93x speedup and 28.9% FLOPs are not reproducible from the method as described.
  2. [§4.2, Tables 2-5] All success rates are reported as single numbers without variance, standard deviation, or episode counts. The central accuracy claim is that the method incurs only a 0.6% average success-rate drop (Table 2, Visual Matching: 74.2% vs. 74.8%; Variant Aggregation: 61.2% vs. 61.3%). Because SIMPLER rollouts are stochastic (object poses, camera noise, diffusion sampling), a 0.6% gap may lie within run-to-run noise. Please report at least three seeds with mean and standard deviation, as well as the number of evaluation episodes per task, for the main configurations.
  3. [§3.2.2, Table 5] The layer-pruning mechanism is justified by the importance metric in Eq. (1), but no control removes the same number of layers chosen at random. The existing ablations do not isolate layer selection: Ex3 combines layer pruning with visual-token pruning and drops to 85.7%, while Ex4 combines layer pruning with MLP sparsity and reaches 92.3%, so neither condition establishes that the cosine-similarity metric, rather than the removal of any ten layers, preserves accuracy. A random-layer-pruning condition matched on the retained layer count and the other components would be needed to support the redundancy-analysis claim in §3.2.1-3.2.2.
minor comments (6)
  1. [§3.3.3, Eq. (4)] The diversity-driven augmentation computes dissimilarity to the fixed key set Vkey rather than to the growing set of already selected tokens, so as written the Kdiv tokens can be mutually similar; if the selection is sequential greedy, please say so and update the reference set after each pick.
  2. [Table 1 vs. Table 3] The module-wise EfficientVLA inference times in Table 1 sum to 110.0 ms, whereas Table 3 reports 121.3 ms for the corresponding CogACT-Base configuration; please reconcile this 11.3 ms discrepancy or state what additional overhead is included.
  3. [Table 4 (right)] The accuracy values for cache intervals 1 through 5 are non-monotonic (91.3%, 94.0%, 93.7%, 90.3%, 93.7%); the text emphasizes progressive acceleration but should also note the accuracy dip at N=4 and the absence of a clear accuracy trend.
  4. [§4.1, Eq. (1)] The size and source of the representative dataset D used to compute the layer importance scores are not specified; please report how many samples were used and from which split they were drawn.
  5. [Throughout] There are several typos and formatting issues, including 'demonstrates shows' in §1, 'Comparision' in §4.2, 'V ariant Aggregations' in §4.1, and the duplicate CogACT references [7] and [37]; these should be corrected.
  6. [Abstract, §4.2] The phrase 'reduces FLOPs to 28.9%' is ambiguous; it should read 'reduces FLOPs to 28.9% of the baseline'.

Circularity Check

0 steps flagged · score 1.0 of 10

No load-bearing circularity: the three acceleration mechanisms are empirical heuristics validated on the external SIMPLER benchmark, and the self-citations are motivational only.

full rationale

I walked the derivation chain for each of the three claimed mechanisms. The layer-pruning importance score (Eq. 1) defines importance as 1 minus input/output cosine similarity, but the claim that pruning low-scoring layers preserves task success is an empirical hypothesis tested in Tables 2 and 5, not an analytic consequence of the definition; the paper does not define 'functionally inconsequential' to mean 'prunable with no drop.' The visual-token selection (Eqs. 2-5) and the diffusion-feature cache (Eqs. 6-11) are similarly heuristic procedures whose benefit is established by SIMPLER evaluation against CogACT, FastV, and VLA-Cache, not entailed by their construction. The 25% MLP sparsity is imported from the external PruneNet reference [39], not from the authors' own prior work, and no equation in this paper reduces to its own output. The self-citations [31,32] appear only in related-work motivation about VLM computational complexity and are not load-bearing. The main methodological concern is that hyperparameters (L, T, N, alpha, Kkey) are selected using sweeps on the same SIMPLER benchmark that reports the headline 1.93x speedup and 0.6% drop; that is a selection-bias risk, not a circular reduction, because the reported success rates are not equal by construction to the fitted hyperparameters. Appendix C honestly notes the limitation that demonstrations are primarily on CogACT. I therefore find no specific circular step supported by quotation and reduction.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central numerical claims depend on several hand-chosen hyperparameters and on three domain heuristics: layers with high input-output similarity are removable, cross-attention scores measure task relevance, and diffusion features are stable across adjacent timesteps. None of these is derived; all are validated only by the reported SIMPLER experiments. There are no new physical or model entities introduced.

free parameters (6)
  • Number of retained LLM layers = 22 or 28 of 32 layers (prune 10 or 4)
    Chosen by hand as an operating point; directly sets the parameter and FLOP reduction from layer pruning (Table 2).
  • Number of retained visual tokens Kfinal = 56 or 112 of 256 tokens
    Chosen by hand; sets the visual sequence length and most of the vision pathway FLOP reduction (Table 4).
  • Cache interval N in the action head = 5
    Selected from the sweep in Table 4; trades action fidelity against diffusion-head speedup.
  • Task-to-diversity augmentation ratio alpha = 0.5
    Chosen by hand in implementation details (§4.1); controls the mix of relevance-driven and diversity-driven visual tokens.
  • Number of core key tokens Kkey = 4
    Chosen by hand in implementation details (§4.1); anchors the visual token selection with instruction-relevant tokens.
  • MLP weight sparsity ratio = 25%
    Adopted from the PruneNet configuration (§4.1); contributes to both FLOPs and the claimed speedup, but its application and hardware support are unspecified.
assumptions (6)
  • domain assumption Layers with high input-output cosine similarity perform minimal functional transformation and can be removed without significant task degradation.
    Eq. (1) in §3.2.1-3.2.2 defines importance as 1 minus cosine similarity; the paper assumes low importance implies low criticality, with no random-layer control.
  • domain assumption Cross-attention scores from visual tokens to instruction tokens measure task relevance of visual tokens.
    §3.3.1, Eq. (2); the basis for key and task token selection, inherited from FastV-style methods.
  • domain assumption Diffusion action-head features are temporally coherent enough that cached attention and MLP outputs remain valid for N-1 skipped timesteps.
    §3.4.2, Eqs. (8)-(11); static caching reuses features even though the input z_t changes each timestep.
  • domain assumption Applying 25% MLP weight sparsity via the PruneNet configuration preserves accuracy without retraining.
    §4.1 'Implementation Details'; the sparsity mechanism and its hardware support are not described.
  • domain assumption The representative dataset D used to compute layer importance represents the evaluation distribution.
    Eq. (1) uses D of 'VLA training samples'; the composition and size are unspecified.
  • domain assumption SIMPLER simulation success rate approximates real-world VLA performance.
    §4.1 relies on SIMPLER's design; no real robot evaluation is performed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models." pith.science (2026). https://pith.science/paper/SPMSLMFB

@misc{pith2026250610100,
  author       = {Pith},
  title        = {Pith review of: EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SPMSLMFB}},
  note         = {Machine review of arXiv:2506.10100}
}
read the original abstract

Vision-Language-Action (VLA) models, particularly diffusion-based architectures, demonstrate transformative potential for embodied intelligence but are severely hampered by high computational and memory demands stemming from extensive inherent and inference-time redundancies. While existing acceleration efforts often target isolated inefficiencies, such piecemeal solutions typically fail to holistically address the varied computational and memory bottlenecks across the entire VLA pipeline, thereby limiting practical deployability. We introduce EfficientVLA, a structured and training-free inference acceleration framework that systematically eliminates these barriers by cohesively exploiting multifaceted redundancies. EfficientVLA synergistically integrates three targeted strategies: (1) pruning of functionally inconsequential layers from the language module, guided by an analysis of inter-layer redundancies; (2) optimizing the visual processing pathway through a task-aware strategy that selects a compact, diverse set of visual tokens, balancing task-criticality with informational coverage; and (3) alleviating temporal computational redundancy within the iterative diffusion-based action head by strategically caching and reusing key intermediate features. We apply our method to a standard VLA model CogACT, yielding a 1.93X inference speedup and reduces FLOPs to 28.9%, with only a 0.6% success rate drop in the SIMPLER benchmark.

Figures

Figures reproduced from arXiv: 2506.10100 by the authors.

Figure 1
Figure 1. VLA inference bottleneck and redundancy analysis: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Efficiency analysis in simulation, comparing FLOPs and inference time of our Effi￾cientVLA variants against the original model backbone. EfficientVLA-22 and EfficientVLA-28 denote configurations retaining 22 and 28 LLM layers, respectively [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Representative robotic manipulation tasks for the Google robot in the SIMPLER environ [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 6 Pith papers

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

  1. TempoVLA: Learning Speed-Controllable Vision-Language-Action Policies

    cs.RO 2026-06 unverdicted novelty 6.0 of 10

    TempoVLA learns a single VLA policy with controllable execution speed via variable-speed trajectory augmentation and explicit speed conditioning.

  2. SAFE-Pruner: Semantic Attention-Guided Future-Aware Token Pruning for Efficient Vision-Language-Action Manipulation

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    SAFE-Pruner forecasts deep-layer visual-token saliency from historical attention maps and refreshes at subtask boundaries, enabling up to 1.89x faster VLA inference with minimal success-rate drop.

  3. 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.

  4. VLSA: Vision-Language-Action Models with Plug-and-Play Safety Constraint Layer

    cs.RO 2025-12 conditional novelty 6.0 of 10

    AEGIS wraps VLA robot policies in a CBF-based safety layer that uses VLM obstacle identification, raising collision avoidance from 18.69% to 77.85% and task success by 17.25 points on the new SafeLIBERO benchmark.

  5. SQAP-VLA: A Synergistic Quantization-Aware Pruning Framework for High-Performance Vision-Language-Action Models

    cs.CV 2025-09 conditional novelty 6.0 of 10

    A training-free framework jointly quantizes a VLA model to 4 bits and prunes visual tokens, recovering or exceeding full-precision success rates at 1.93x speedup.

  6. 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.

Reference graph

Works this paper leans on

38 extracted references · 8 canonical work pages · cited by 6 Pith papers

  1. [39]

    You only prune once: Designing calibration-free model compression with policy learning,

    A. Sengupta, S. Chaudhary, and T. Chakraborty, “You only prune once: Designing calibration-free model compression with policy learning,”arXiv preprint arXiv:2501.15296, 2025. 13 A Experimental Settings A.1 SIMPLER Environment The SIMPLER simulation environment serves as our primary benchmark for evaluating VLA models. It is specifically designed to closel...

    arXiv 2025

  2. [2]

    Blip: Bootstrapping language-image pre-training for unified vision- language understanding and generation,

    J. Li, D. Li, C. Xiong, and S. Hoi, “Blip: Bootstrapping language-image pre-training for unified vision- language understanding and generation,” inInternational conference on machine learning. PMLR, 2022, pp. 12 888–12 900

    work page 2022

  3. [3]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PMLR, 2021, pp. 8748–8763

    2021

  4. [4]

    Mc-llava: Multi-concept personalized vision-language model,

    R. An, S. Yang, M. Lu, K. Zeng, Y . Luo, Y . Chen, J. Cao, H. Liang, Q. She, S. Zhanget al., “Mc-llava: Multi-concept personalized vision-language model,”arXiv preprint arXiv:2411.11706, 2024

    arXiv 2024

  5. [5]

    Llm as dataset analyst: Subpopulation structure discovery with large language model,

    Y . Luo, R. An, B. Zou, Y . Tang, J. Liu, and S. Zhang, “Llm as dataset analyst: Subpopulation structure discovery with large language model,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 235–252

    work page 2024

  6. [6]

    Openvla: An open-source vision-language-action model,

    M. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. 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,”arXiv preprint arXiv:2406.09246, 2024

    arXiv 2024

  7. [7]

    Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation,

    Q. Li, Y . Liang, Z. Wang, L. Luo, X. Chen, M. Liao, F. Wei, Y . Deng, S. Xu, Y . Zhang, X. Wang, B. Liu, J. Fu, J. Bao, D. Chen, Y . Shi, J. Yang, and B. Guo, “Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation,” 2024

    work page 2024

  8. [8]

    π0: A vision-language-action flow model for general robot control,

    K. Black, N. Brown, D. Driess, A. Esmail, M. Equi, C. Finn, N. Fusai, L. Groom, K. Hausman, B. Ichteret al., “ π0: A vision-language-action flow model for general robot control,”arXiv preprint arXiv:2410.24164, 2024

    arXiv 2024

Show all 38 references
  1. [9]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control,

    A. Brohan, N. Brown, J. Carbajal, Y . Chebotar, X. Chen, K. Choromanski, T. Ding, D. Driess, A. Dubey, C. Finn, P. Florence, C. Fu, M. G. Arenas, K. Gopalakrishnan, K. Han, K. Hausman, A. Herzog, J. Hsu, B. Ichter, A. Irpan, N. Joshi, R. Julian, D. Kalashnikov, Y . Kuang, I. L...

    2023 arXiv

  2. [10]

    Rh20t: A comprehensive robotic dataset for learning diverse skills in one-shot,

    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,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 653–660

    2024

  3. [11]

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0,

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jainet al., “Open x-embodiment: Robotic learning datasets and rt-x models: Open x-embodiment collaboration 0,” in2024 IEEE International Conference on Robotics and Automat...

    2024

  4. [12]

    Diffusion-vla: Scal- ing robot foundation models via unified diffusion and autoregression,

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

    2024 arXiv

  5. [13]

    Dexvla: Vision-language model with plug-in diffusion expert for general robot control,

    J. Wen, Y . Zhu, J. Li, Z. Tang, C. Shen, and F. Feng, “Dexvla: Vision-language model with plug-in diffusion expert for general robot control,”arXiv preprint arXiv:2502.05855, 2025

    2025 arXiv

  6. [14]

    Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation,

    J. Wen, Y . Zhu, J. Li, M. Zhu, K. Wu, Z. Xu, N. Liu, R. Cheng, C. Shen, Y . Penget al., “Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation,”arXiv preprint arXiv:2409.12514, 2024

    2024 arXiv

  7. [15]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosaleet al., “Llama 2: Open foundation and fine-tuned chat models,”arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  8. [16]

    Gpt-3: Its nature, scope, limits, and consequences,

    L. Floridi and M. Chiriatti, “Gpt-3: Its nature, scope, limits, and consequences,”Minds and Machines, vol. 30, pp. 681–694, 2020. 11

    2020

  9. [17]

    Gpt-4 technical report,

    OpenAIet al., “Gpt-4 technical report,” 2024. [Online]. Available: https://arxiv.org/abs/2303.08774

    2024 arXiv

  10. [18]

    Roberta: A robustly optimized bert pretraining approach,

    Y . Liu, M. Ott, N. Goyal, J. Du, M. Joshi, D. Chen, O. Levy, M. Lewis, L. Zettlemoyer, and V . Stoyanov, “Roberta: A robustly optimized bert pretraining approach,” 2019. [Online]. Available: https://arxiv.org/abs/1907.11692

    2019 arXiv

  11. [19]

    Data whisperer: Efficient data selection for task-specific llm fine-tuning via few-shot in-context learning,

    S. Wang, Z. Wang, X. Jin, J. Wang, J. Zhang, K. Li, Z. Wen, Z. Li, C. He, X. Huet al., “Data whisperer: Efficient data selection for task-specific llm fine-tuning via few-shot in-context learning,”arXiv preprint arXiv:2505.12212, 2025

    2025 arXiv

  12. [20]

    Deer-vla: Dynamic inference of multimodal large language models for efficient robot execution,

    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,” inThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  13. [21]

    Mole- vla: Dynamic layer-skipping vision language action model via mixture-of-layers for efficient robot manipulation,

    R. Zhang, M. Dong, Y . Zhang, L. Heng, X. Chi, G. Dai, L. Du, D. Wang, Y . Du, and S. Zhang, “Mole- vla: Dynamic layer-skipping vision language action model via mixture-of-layers for efficient robot manipulation,”arXiv preprint arXiv:2503.20384, 2025

    2025 arXiv

  14. [22]

    Vla-cache: Towards efficient vision-language-action model via adaptive token caching in robotic manipulation,

    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

    2025

  15. [23]

    Evaluating real-world robot manipulation policies in simulation,

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

    2024 arXiv

  16. [24]

    Vision-language foundation models as effective robot imitators,

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

    2023 arXiv

  17. [25]

    Rt- 2: Vision-language-action models transfer web knowledge to robotic control,

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahidet al., “Rt- 2: Vision-language-action models transfer web knowledge to robotic control,” inConference on Robot Learning. PMLR, 2023, pp. 2165–2183

    2023

  18. [26]

    Flamingo: a visual language model for few-shot learning,

    J.-B. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y . Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, R. Ring, E. Rutherford, S. Cabi, T. Han, Z. Gong, S. Samangooei, M. Monteiro, J. Menick, S. Borgeaud, A. Brock, A. Nematzadeh, S. Sharifzadeh, M. Binkowski, R. Barre...

    2022 arXiv

  19. [27]

    Prismatic vlms: Investigating the design space of visually-conditioned language models,

    S. Karamcheti, S. Nair, A. Balakrishna, P. Liang, T. Kollar, and D. Sadigh, “Prismatic vlms: Investigating the design space of visually-conditioned language models,” 2024. [Online]. Available: https://arxiv.org/abs/2402.07865

    2024 arXiv

  20. [28]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,

    J. Li, D. Li, C. Xiong, and S. Hoi, “Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,” 2022. [Online]. Available: https://arxiv.org/abs/2201.12086

    2022 arXiv

  21. [29]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” 2023. [Online]. Available: https://arxiv.org/abs/2304.08485

    2023 arXiv

  22. [30]

    Openflamingo: An open-source framework for training large autoregressive vision-language models,

    A. Awadalla, I. Gao, J. Gardner, J. Hessel, Y . Hanafy, W. Zhu, K. Marathe, Y . Bitton, S. Gadre, S. Sagawa, J. Jitsev, S. Kornblith, P. W. Koh, G. Ilharco, M. Wortsman, and L. Schmidt, “Openflamingo: An open-source framework for training large autoregressive vision-language m...

    2023 arXiv

  23. [31]

    Shifting ai efficiency from model-centric to data-centric compression,

    X. Liu, Z. Wen, S. Wang, J. Chen, Z. Tao, Y . Wang, X. Jin, C. Zou, Y . Wang, C. Liaoet al., “Shifting ai efficiency from model-centric to data-centric compression,”arXiv preprint arXiv:2505.19147, 2025

    2025

  24. [32]

    Stop looking for important tokens in multimodal language models: Duplication matters more,

    Z. Wen, Y . Gao, S. Wang, J. Zhang, Q. Zhang, W. Li, C. He, and L. Zhang, “Stop looking for important tokens in multimodal language models: Duplication matters more,”arXiv preprint arXiv:2502.11494, 2025

    2025 arXiv

  25. [33]

    Robomamba: Efficient vision-language-action model for robotic reasoning and manipulation,

    J. Liu, M. Liu, Z. Wang, P. An, X. Li, K. Zhou, S. Yang, R. Zhang, Y . Guo, and S. Zhang, “Robomamba: Efficient vision-language-action model for robotic reasoning and manipulation,” 2024. [Online]. Available: https://arxiv.org/abs/2406.04339

    2024 arXiv

  26. [34]

    Efficientvlm: Fast and accurate vision-language models via knowledge distillation and modal-adaptive pruning,

    T. Wang, W. Zhou, Y . Zeng, and X. Zhang, “Efficientvlm: Fast and accurate vision-language models via knowledge distillation and modal-adaptive pruning,” 2022. [Online]. Available: https://arxiv.org/abs/2210.07795

    2022 arXiv

  27. [35]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Noubyet al., “Dinov2: Learning robust visual features without supervision,”arXiv preprint arXiv:2304.07193, 2023. 12

    2023 arXiv

  28. [36]

    Sigmoid loss for language image pre-training,

    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, 2023, pp. 11 975–11 986

    2023

  29. [37]

    Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation,

    Q. Li, Y . Liang, Z. Wang, L. Luo, X. Chen, M. Liao, F. Wei, Y . Deng, S. Xu, Y . Zhang, X. Wang, B. Liu, J. Fu, J. Bao, D. Chen, Y . Shi, J. Yang, and B. Guo, “Cogact: A foundational vision-language-action model for synergizing cognition and action in robotic manipulation,” 2...

    2024 arXiv

  30. [38]

    An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models,

    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. Springer, 2025, pp. 19–35

    2025

Pith tools

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