pith. sign in

arxiv: 2606.20246 · v1 · pith:TULVADXYnew · submitted 2026-06-18 · 💻 cs.RO · cs.AI

Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think

Gia-Binh Nguyen , Trong-Bao Ho , Thien-Loc Ha , Khoa Vo , Philip Lund M{\o}ller , Quang T. Nguyen , Long Dinh , Tuan Dam
show 12 more authors
Vu Duong Tung M. Luu Trung Le Tran Nguyen Le Minh Vu An Thai Le Ngan Le Daniel Sonntag James Zou Jan Peters Duy M. H. Nguyen Ngo Anh Vien
This is my paper

Pith reviewed 2026-06-26 16:57 UTC · model grok-4.3

classification 💻 cs.RO cs.AI
keywords Vision-Language-Action modelsmodel compressionlayer redundancyCentered Kernel Alignmentrobotic manipulationfine-tuning efficiencyinference acceleration
0
0 comments X

The pith

Vision-language-action models contain enough layer redundancy to allow permanent 50% depth compression while preserving or improving fine-tuning performance.

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

The paper shows that large VLA models trained on diverse robot trajectories still contain many nearly identical layers. A single forward pass with Centered Kernel Alignment identifies these twin layers so they can be removed without any retraining of the compression step. The resulting half-depth models are then fine-tuned on downstream tasks and achieve the same or better success rates while cutting training time by 40-50% and inference time by up to 30%. The compression applies to both the vision-language backbone and the continuous control head and is validated on three simulation suites plus ten real-world tasks across four robot embodiments.

Core claim

Despite training on massive and varied physical trajectories, current VLA policies exhibit severe layer-wise representational redundancy. Measuring pairwise layer similarity with Centered Kernel Alignment on one forward pass reveals pairs of nearly duplicate layers; permanently excising one member of each pair compresses model depth by up to 50% in both the VLM backbone and the policy head. The resulting architecture, when fine-tuned, matches or exceeds the task performance of the original full-depth model while delivering 40-50% faster training and up to 30% faster real-time inference.

What carries the argument

Centered Kernel Alignment (CKA) applied to a single forward pass to detect and remove twin redundant layers across the VLM backbone and continuous control policy head.

If this is right

  • Model depth can be halved permanently before any fine-tuning begins.
  • Downstream training time drops 40-50% while task success stays the same or rises.
  • Real-time inference runs up to 30% faster on the same hardware.
  • The same compression works on both the vision-language backbone and the action head.
  • The benefit holds across simulation benchmarks and real robots with four different embodiments.

Where Pith is reading between the lines

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

  • Pre-training on video-robot data may routinely produce over-parameterized layers whose features are not task-specific.
  • A similar one-pass redundancy scan could be tested on other large control or multimodal models outside robotics.
  • If the redundancy pattern is stable, the method could let researchers train and deploy capable VLAs on hardware with half the memory or compute budget.

Load-bearing premise

The layer pairs found to be redundant by one forward pass of CKA stay redundant for any later task, robot body, or fine-tuning run.

What would settle it

Fine-tune the 50%-compressed model on a previously unseen task or embodiment and measure whether success rate falls more than a few percent below the full-depth baseline.

Figures

Figures reproduced from arXiv: 2606.20246 by An Thai Le, Daniel Sonntag, Duy M. H. Nguyen, Gia-Binh Nguyen, James Zou, Jan Peters, Khoa Vo, Long Dinh, Minh Vu, Ngan Le, Ngo Anh Vien, Philip Lund M{\o}ller, Quang T. Nguyen, Thien-Loc Ha, Tran Nguyen Le, Trong-Bao Ho, Trung Le, Tuan Dam, Tung M. Luu, Vu Duong.

Figure 1
Figure 1. Figure 1: Overview of the proposed CLP framework. CLP prunes representationally redundant transformer layers via CKA, reducing network depth by up to 66% and training/inference cost by up to 50%. Fine-tuning restores the latent geometry of the compressed model, enabling competitive performance across three simulation benchmarks, 10 real-world tasks, and four robotic embodiments. SmolVLA [12]) generally share a decou… view at source ↗
Figure 2
Figure 2. Figure 2: CKA similarity profiles across π0 and GR00T-N1.5 sub-modules. The heatmaps illustrate pairwise representation alignment among transformer layers inside the VLM backbones, action heads, and DiT blocks. The extensive, contiguous plateaus of high similarity (dark red) across both model families signify minimal representational changes between successive layers, pinpointing candidate zones for structured pruni… view at source ↗
Figure 3
Figure 3. Figure 3: Analysis of CLP evaluation across benchmarks and real-world tasks. (a) Success rate of π0 on LIBERO with different layer pruning ratio; (b) Success rate of GR00T N1.5 on RoboCasa across pruning ratios; (c) Comparison with dynamic layer skipping method (MoLe-VLA [15]) vs ours; (d) Comparison of different pruning strategies on GR00T N1.5 across LIBERO benchmark; (e) Training time and success rate on real-wor… view at source ↗
Figure 4
Figure 4. Figure 4: Robot experiments with folding shorts. RQ3. State-of-the-art Compression Compar￾ison. To evaluate the practical utility of CLP, we (i) compare with training-free pruning across multiple VLA settings. On LIBERO ( [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Real-world data samples [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Simulation data samples A.4 Simulation experiments details LIBERO: We use LIBERO as a benchmark to evaluate our method against other baselines. A 10% subset is randomly sampled from the full LIBERO dataset. We fine-tune all models - π0, GR00T-N1.5, and SmolVLA - on this subset for 100k steps with a global batch size of 64. We evaluate each task over 50 episodes with an execution length of 10 steps.Results … view at source ↗
read the original abstract

Vision-Language-Action (VLA) models pre-trained on massive video-robot datasets have revolutionized robotic manipulation, yet their multi-billion parameter architectures impose prohibitive computational burdens during downstream fine-tuning and real-time inference. In this work, we reveal a highly non-trivial architectural characteristic of these continuous control foundation policies (e.g., pi_0, GR00T-N1.5): despite being trained on diverse physical trajectories, they exhibit severe layer-wise representational redundancy. To exploit this, we introduce a structural compression pipeline that is entirely training-free, bypassing the need of existing methods to load full-scale models to learn optimized token reductions or dynamic layer selectors. Instead, using only a single forward pass via Centered Kernel Alignment to identify redundant layer features, we remove twin layers to permanently compress the model depth by up to 50% across both the VLM backbone and the continuous control policy head. Downstream fine-tuning of this streamlined architecture yields a dual acceleration benefit: a 40-50% reduction in training time and up to 30% faster real-time inference, while matching or exceeding full-scale base model performance. We comprehensively validate our method across three simulation benchmarks (LIBERO, RoboCasa, SimplerEnv) and 10 diverse real-world manipulation tasks across 4 unique robotic embodiments. These results prove that advanced VLAs require significantly fewer layers than previously assumed, offering a highly compute-efficient paradigm for scalable robot learning.

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

3 major / 2 minor

Summary. The paper claims that pre-trained VLA models (e.g., pi_0, GR00T-N1.5) exhibit severe layer-wise representational redundancy despite training on diverse trajectories; a single forward pass with Centered Kernel Alignment (CKA) identifies 'twin' layers that can be removed in a training-free manner to permanently compress both the VLM backbone and policy head by up to 50%, after which fine-tuning yields 40-50% faster training, up to 30% faster inference, and performance matching or exceeding the original model. Validation is asserted across LIBERO, RoboCasa, SimplerEnv and 10 real-world tasks on 4 embodiments.

Significance. If the central claim holds, the work would demonstrate that advanced VLAs can be structurally compressed without task-specific retraining for the compression step itself, offering substantial practical gains in fine-tuning cost and deployment speed for robotic manipulation. The training-free identification step is a clear methodological strength relative to learned token or layer selectors.

major comments (3)
  1. [§4] §4 (Experimental validation): the abstract asserts 'comprehensive validation' and 'matching or exceeding' performance across three simulation benchmarks and 10 real-world tasks, yet no details are supplied on number of random seeds, error bars, statistical tests, or how embodiment/task selection was performed; without these the claim that compression preserves capacity cannot be assessed as load-bearing evidence.
  2. [Method] Method section (CKA identification procedure): the load-bearing assumption that layer pairs identified as redundant by CKA on a single forward pass remain redundant under the distribution shift of downstream fine-tuning trajectories is stated but not tested via cross-distribution CKA comparisons or ablation on held-out embodiments; if similarity matrices change with new physical trajectories the removed layers could reduce capacity, undermining the 'permanent' and 'general' compression pipeline.
  3. [Abstract and §3] Abstract and §3 (single forward pass): the input data distribution used for the CKA forward pass is not specified (e.g., whether it is a subset of pre-training data, a generic robot trajectory set, or task-specific); this choice directly determines which layers are declared 'twins' and is therefore central to the generalizability claim across the reported 4 embodiments.
minor comments (2)
  1. Notation for the policy head compression is introduced without an explicit equation or diagram showing which layers are paired and removed; a small schematic would clarify the 50% depth reduction claim.
  2. The title uses 'Fewer Layers Than You Think' which is informal; a more precise phrasing would better match the technical contribution.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below, providing clarifications and committing to revisions that strengthen the manuscript without altering its core claims.

read point-by-point responses

  1. Referee: [§4] §4 (Experimental validation): the abstract asserts 'comprehensive validation' and 'matching or exceeding' performance across three simulation benchmarks and 10 real-world tasks, yet no details are supplied on number of random seeds, error bars, statistical tests, or how embodiment/task selection was performed; without these the claim that compression preserves capacity cannot be assessed as load-bearing evidence.

    Authors: We agree that these experimental details are necessary for rigorous assessment. In the revised manuscript we will report all results as means over 3 random seeds with standard deviations, include paired statistical tests (e.g., Wilcoxon) for performance comparisons, and describe the task/embodiment selection criteria used to ensure coverage of diverse manipulation scenarios across the four platforms. These additions will make the evidence for preserved capacity explicit. revision: yes

  2. Referee: [Method] Method section (CKA identification procedure): the load-bearing assumption that layer pairs identified as redundant by CKA on a single forward pass remain redundant under the distribution shift of downstream fine-tuning trajectories is stated but not tested via cross-distribution CKA comparisons or ablation on held-out embodiments; if similarity matrices change with new physical trajectories the removed layers could reduce capacity, undermining the 'permanent' and 'general' compression pipeline.

    Authors: This is a fair methodological concern. While the successful transfer of the compressed models to four distinct embodiments provides supporting evidence that the identified redundancies are not brittle, we did not perform explicit cross-distribution CKA ablations. In revision we will add a targeted analysis comparing CKA similarity matrices computed on pre-training-style trajectories versus held-out fine-tuning trajectories from each embodiment to directly test stability of the twin-layer identifications. revision: partial

  3. Referee: [Abstract and §3] Abstract and §3 (single forward pass): the input data distribution used for the CKA forward pass is not specified (e.g., whether it is a subset of pre-training data, a generic robot trajectory set, or task-specific); this choice directly determines which layers are declared 'twins' and is therefore central to the generalizability claim across the reported 4 embodiments.

    Authors: We will revise §3 and the method section to explicitly state that the single forward pass uses a generic collection of robot trajectories sampled from a broad distribution matching the pre-training corpus but deliberately independent of the downstream task data. This choice was made precisely to support the generalizability claim; the clarification will remove ambiguity. revision: yes

Circularity Check

0 steps flagged

No circularity; method uses external CKA metric on forward passes

full rationale

The paper identifies redundant layers via Centered Kernel Alignment (CKA) computed on a single forward pass, an independent external similarity measure unrelated to any fitted parameters or self-derived quantities within the work. Layer removal is a direct structural edit based on this measurement, followed by standard downstream fine-tuning on the compressed model. No equations or steps reduce by construction to the inputs (no self-definitional mappings, no fitted inputs renamed as predictions, no load-bearing self-citations, and no uniqueness theorems imported from prior author work). The derivation chain is empirically grounded in an off-the-shelf metric and remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides insufficient detail to identify any free parameters, axioms, or invented entities; the method relies on CKA which is a standard metric.

pith-pipeline@v0.9.1-grok · 5864 in / 1044 out tokens · 27206 ms · 2026-06-26T16:57:13.079737+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

44 extracted references · 14 canonical work pages · 10 internal anchors

  1. [1]

    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

    2023

  2. [2]

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    $\pi_0$: 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. Ichter, et al. π 0: A vision-language-action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  5. [5]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

    J. Bjorck, F. Casta˜neda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huang, et al. Gr00t n1: An open foundation model for generalist humanoid robots.arXiv preprint arXiv:2503.14734, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  6. [6]

    M. J. Kim, C. Finn, and P. Liang. Fine-tuning vision-language-action models: Optimizing speed and success.Conference on Neural Information Processing Systems (NeurIPS), 2024

    2024

  7. [7]

    Driess, J

    D. Driess, J. Springenberg, B. Ichter, L. Yu, A. Li-Bell, K. Pertsch, A. Ren, H. Walke, Q. Vuong, L. X. Shi, et al. Knowledge insulating vision-language-action models: Train fast, run fast, generalize better.Advances in Neural Information Processing Systems, 38:102867–102888, 2026

    2026

  8. [8]

    Y . Yang, Y . Wang, Z. Wen, L. Zhongwei, C. Zou, Z. Zhang, C. Wen, and L. Zhang. Efficientvla: Training-free acceleration and compression for vision-language-action models.Advances in Neural Information Processing Systems, 38:40891–40914, 2026. 9

    2026

  9. [9]

    H. Wang, J. Xu, Y . Xiang, J. Pan, Y . Zhou, Y .-L. Li, and G. Dai. Specprune-vla: Accelerat- ing vision-language-action models via action-aware self-speculative pruning.International Conference on Machine Learning, 2026

    2026

  10. [10]

    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. Advances in Neural Information Processing Systems, 37:40085–40110, 2024

    2024

  11. [11]

    Reuss, H

    M. Reuss, H. Zhou, M. R¨uhle, ¨O. E. Ya˘gmurlu, F. Otto, and R. Lioutikov. Flower: Democratizing generalist robot policies with efficient vision-language-action flow policies.Conference on Robot Learning (CoRL), 2025

    2025

  12. [12]

    SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics

    M. Shukor, D. Aubakirova, F. Capuano, P. Kooijmans, S. Palma, A. Zouitine, M. Aractingi, C. Pascal, M. Russi, A. Marafioti, et al. Smolvla: A vision-language-action model for affordable and efficient robotics.arXiv preprint arXiv:2506.01844, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    C.-Y . Hung, Q. Sun, P. Hong, A. Zadeh, C. Li, U. Tan, N. Majumder, S. Poria, et al. Nora: A small open-sourced generalist vision language action model for embodied tasks.arXiv preprint arXiv:2504.19854, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  14. [14]

    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

    2024

  15. [15]

    Zhang, M

    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. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 18764–18772, 2026

    2026

  16. [16]

    S. Xu, Y . Wang, C. Xia, D. Zhu, T. Huang, and C. Xu. Vla-cache: Efficient vision-language- action manipulation via adaptive token caching.Advances in Neural Information Processing Systems, 38:164448–164473, 2026

    2026

  17. [17]

    D. M. Nguyen, N. T. Diep, B. G. Nguyen, T.-B. Ho, D. Le, T. Nguyen, T.-L. Ha, T. Nhiem, B. Thach, N. Tran, T. A. Tran, A. Habuda, P. L. Moeller, T. N. Le, D. Sonntag, M. Niepert, K. Doan, V . Duong, H. Ngo, M. Vu, D. M. Nguyen, A. Le, and V . Ngo. Foca: Future- oriented conditioning for data-efficient vision-language-action adaptation. InProceedings of th...

    2026

  18. [18]

    M. Koo, D. Choi, T. Kim, K. Lee, C. Kim, Y . Seo, and J. Shin. Hamlet: Switch your vision- language-action model into a history-aware policy.International Conference on Learning Representations (ICLR), 2025

    2025

  19. [19]

    Kornblith, M

    S. Kornblith, M. Norouzi, H. Lee, and G. Hinton. Similarity of neural network representations revisited. InInternational conference on machine learning, pages 3519–3529. PMlR, 2019

    2019

  20. [20]

    Cortes, M

    C. Cortes, M. Mohri, and A. Rostamizadeh. Algorithms for learning kernels based on centered alignment. InInternational Conference on Artificial Intelligence and Statistics (AISTATS), pages 301–309, 2012

    2012

  21. [21]

    Nguyen, M

    T. Nguyen, M. Raghu, and S. Kornblith. Do wide and deep networks learn the same things? uncovering how neural network representations vary with width and depth. InInternational Conference on Learning Representations (ICLR), 2021

    2021

  22. [22]

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone. Libero: Benchmarking knowledge transfer for lifelong robot learning. InAdvances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, 2023. 10

    2023

  23. [23]

    Nasiriany, A

    S. Nasiriany, A. Maddukuri, L. Zhang, A. Parikh, A. Lo, A. Joshi, A. Mandlekar, and Y . Zhu. Robocasa: Large-scale simulation of everyday tasks for generalist robots. InRobotics: Science and Systems (RSS), 2024

    2024

  24. [24]

    X. Li, K. Gao, H. Zhou, A. Yu, Y . Zhu, H. Ku, S. Tao, J. Gu, S. Ha, X. Peng, and H. Su. Evaluating real-world robot manipulation policies in simulation. InConference on Robot Learning (CoRL), 2024

    2024

  25. [25]

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn. Learning fine-grained bimanual manipulation with low-cost hardware.arXiv preprint arXiv:2304.13705, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    K. P. Hawkins. Analytic inverse kinematics for the universal robots UR-5/UR-10 arms. Technical report, Georgia Institute of Technology, 2013

    2013

  27. [27]

    Universal Robots A/S, Odense, Denmark, 2013

    Universal Robots UR5/UR10 User Manual. Universal Robots A/S, Odense, Denmark, 2013. Available athttps://www.universal-robots.com

    2013

  28. [28]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces

    A. Gu and T. Dao. Mamba: Linear-time sequence modeling with selective state spaces.arXiv preprint arXiv:2312.00752, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  29. [29]

    B. Xiao, H. Wu, W. Xu, X. Dai, H. Hu, Y . Lu, M. Zeng, C. Liu, and L. Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4818–4829, 2024

    2024

  30. [30]

    Q. Team. Qwen2.5-VL technical report.arXiv preprint arXiv:2502.13923, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    S. Ge, Y . Zhang, L. Liu, M. Zhang, J. Han, and J. Gao. Model tells you what to discard: Adaptive kv cache compression for llms. InInternational Conference on Learning Representations, volume 2024, pages 22975–22988, 2024

    2024

  32. [32]

    X. Zhou, W. Wang, M. Zeng, J. Guo, X. Liu, L. Shen, M. Zhang, and L. Ding. Dynamickv: Task- aware adaptive kv cache compression for long context llms.arXiv preprint arXiv:2412.14838, 2024

    work page arXiv 2024

  33. [33]

    X. Pei, Y . Chen, S. Xu, Y . Wang, Y . Shi, and C. Xu. Action-aware dynamic pruning for efficient vision-language-action manipulation.The Fourteenth International Conference on Learning Representations (ICLR), 2026

    2026

  34. [34]

    W. Yu, T. Wang, F. Li, J. Li, and L. Zhu. Acˆ 2-vla: Action-context-aware adaptive com- putation in vision-language-action models for efficient robotic manipulation.arXiv preprint arXiv:2601.19634, 2026

    work page arXiv 2026

  35. [35]

    G. R. Team, A. Abdolmaleki, S. Abeyruwan, J. Ainslie, J.-B. Alayrac, M. G. Arenas, A. Balakr- ishna, N. Batchelor, A. Bewley, J. Bingham, et al. Gemini robotics 1.5: Pushing the frontier of generalist robots with advanced embodied reasoning, thinking, and motion transfer.arXiv preprint arXiv:2510.03342, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  36. [36]

    Huang, Y .-H

    C.-P. Huang, Y .-H. Wu, M.-H. Chen, F. Wang, and F.-E. Yang. Thinkact: Vision-language-action reasoning via reinforced visual latent planning.Advances in Neural Information Processing Systems, 38:82782–82802, 2026

    2026

  37. [37]

    Zhang, H

    W. Zhang, H. Liu, Z. Qi, Y . Wang, X. Yu, J. Zhang, R. Dong, J. He, H. Wang, Z. Zhang, et al. Dreamvla: a vision-language-action model dreamed with comprehensive world knowledge. Advances in Neural Information Processing Systems, 38:24195–24228, 2026

    2026

  38. [38]

    Lipman, R

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

    2023

  39. [39]

    The unreasonable ineffectiveness of the deeper layers.arXiv preprint arXiv:2403.17887,

    A. Gromov, K. Tirumala, H. Shapourian, P. Glorioso, and D. A. Roberts. The unreasonable ineffectiveness of the deeper layers.arXiv preprint arXiv:2403.17887, 2024. 11

    work page arXiv 2024

  40. [40]

    X. Men, M. Xu, Q. Zhang, Q. Yuan, B. Wang, H. Lin, Y . Lu, X. Han, and W. Chen. Shortgpt: Layers in large language models are more redundant than you expect. InFindings of the Association for Computational Linguistics: ACL 2025, pages 20192–20204, 2025

    2025

  41. [41]

    X. Chen, Y . Hu, J. Zhang, Y . Wang, C. Li, and H. Chen. Streamlining redundant layers to compress large language models.arXiv preprint arXiv:2403.19135, 2024

    work page arXiv 2024

  42. [42]

    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. In European Conference on Computer Vision, pages 19–35. Springer, 2024

    2024

  43. [43]

    S. R. Alvar, G. Singh, M. Akbari, and Y . Zhang. Divprune: Diversity-based visual token pruning for large multimodal models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 9392–9401, 2025

    2025

  44. [44]

    S. Wang, R. Yu, Z. Yuan, C. Yu, F. Gao, Y . Wang, and D. F. Wong. Spec-vla: speculative decoding for vision-language-action models with relaxed acceptance. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 26916–26928, 2025. 12 Supplement to ”Finetuning Vision-Language-Action Models Requires Fewer Layers Than ...

    2025