Pith. sign in

REVIEW 4 major objections 5 minor 9 cited by

A training-free combination of 4-bit quantization and token pruning speeds up a vision-language-action model 1.93x while improving average success rate by up to 4.5% over the full-precision baseline.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

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.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Useful engineering result on training-free W4A4 quantization plus token pruning for VLA models, but the paper's own ablation undercuts its 'quantization-aware' narrative, and the headline numbers are weakened by test-set selection and missing error bars. the 4 major comments →

arxiv 2509.09090 v1 pith:PBQCFJAB submitted 2025-09-11 cs.CV cs.AI

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

Hengyu Fang , Yijiang Liu , Yuan Du , Li Du , Huanrui Yang This is my paper
classification cs.CV cs.AI
keywords Vision-Language-Action modelsquantizationtoken pruningW4A4Hadamard transformfarthest point samplingtraining-free compressionrobot manipulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper tries to establish that aggressive W4A4 (4-bit weight, 4-bit activation) quantization and token pruning are not inherently incompatible in vision-language-action (VLA) models: the conflict can be resolved by co-designing the pruning criteria to survive quantization noise and shaping the quantizer to support pruning. It proposes a training-free framework that keeps the top-attention tokens whose identities are claimed to be quantization-stable, protects tokens corresponding to the robot arm via camera projection, and uses farthest-point sampling for spatial coverage, alongside a Hadamard transform to suppress activation outliers. If the claim holds, edge robots could run compressed VLA policies at roughly double speed and lower memory with no loss — and sometimes a gain — in task success. The paper reports a 1.93x speedup, a memory drop from 14.3 GB to 7.6 GB, and success-rate improvements of 4.5% and 3.1% over the full-precision baseline in its two evaluated scenarios.

Core claim

The central discovery is that poor results from combining quantization and token pruning stem from quantization distorting the attention maps that pruning uses, and that this distortion can be counteracted without retraining. The authors show that after W4A4 quantization, attention becomes scattered and shifted, but a small set of top-attention tokens keeps its identity; that a robot's known 3D position can be projected to pixel and token coordinates to anchor protected tokens; and that farthest-point sampling keeps remaining tokens spatially spread. They further show that rotating the query/key activations with a Hadamard matrix redistributes activation energy and removes outlier channels,

What carries the argument

The load-bearing mechanism is a set of three quantization-aware token selection criteria plus one quantizer enhancement. 'Quantization-insensitive preservation' selects the top-k attention tokens, relying on the empirical claim that their indices change little under quantization. 'Robot-aware token protection' projects the robot arm's known 3D world coordinates into token coordinates and keeps a Chebyshev ring of tokens around that anchor. 'Spatially-aware sampling' applies farthest-point sampling to the remaining tokens to preserve geographic spread. On the quantization side, a per-tensor Hadamard transform is applied to query and key projections (using the identity H^T H = I) to suppress a

Load-bearing premise

The framework's success depends on the claim that the identities of top-k attention tokens are largely unchanged after 4-bit quantization, but the paper offers no quantitative stability measurement and its own ablation shows that adding this top-k preservation to a quantized model decreased the average success rate (71.78% to 70.90% in visual matching).

What would settle it

Measure the Jaccard overlap between the set of top-k visual-token indices selected by the full-precision model and those selected by the W4A4 quantized model across a large sample of images from the robotics manipulation benchmark and training dataset used in the paper. If the overlap is substantially below the level needed for the pruning ratio of 0.4 (e.g., less than 50% for the k used), the claimed quantization-insensitivity of attention-based top-k selection would be refuted. A second falsifier: run the framework on tasks where the robot arm is out of view or the camera calibration is inte

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

If this is right

  • If correct, VLA models can be deployed on edge hardware at 4-bit precision with 40% token pruning and no retraining, reducing peak GPU memory from 14.3 GB to 7.6 GB.
  • The 1.93x end-to-end speedup decomposes into 2.09x from quantization and 1.21x from pruning inside the LLM backbone, showing the two techniques compound when designed together.
  • The method outperforms full-precision token-pruning baselines despite also operating under W4A4 quantization, suggesting co-design matters more than either compression alone.
  • The optimal pruning ratio of 0.4 was found through an ablation sweep; ratios from 0.3 to 0.6 all beat the unpruned W4A4 model, indicating robustness to the pruning-rate choice.
  • The framework is training-free, so it can be applied to pretrained checkpoints directly; if confirmed on more architectures and tasks, it would make quantization and token pruning a standard paired deployment step.

Where Pith is reading between the lines

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

  • The paper's own ablation shows that the 'quantization-insensitive top-k' component alone lowered the visual-matching average from 71.78 to 70.90 when added to W4A4; this suggests the headline gains are carried mainly by robot-aware projection and farthest-point sampling, not by the claimed stability of top-k. A reader should treat the top-k stability mechanism as the part needing independent verif
  • The reliance on camera intrinsics and extrinsics plus known robot coordinates means the robot-aware protection is only as good as the calibration; in settings with moving cameras or imprecise pose, the protected ring might miss the arm and the framework would lose its main performance source.
  • Since the framework is training-free and task-agnostic across the four benchmark tasks, a natural extension would be testing it on dynamic, multi-object scenes where the 'top-k stable' assumption is most likely to break; the authors' own future-work suggestion of task-adaptive pruning points in a similar direction.
  • A system-level implication left implicit: the measured 1.93x speedup assumes hardware that actually supports efficient W4A4 tensor-wise operations; on GPUs without such kernels, the quantization speedup may not materialize, leaving only the pruning gain.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes SQAP-VLA, a training-free framework that combines W4A4 post-training quantization with token pruning for Vision-Language-Action (VLA) models. The method consists of three pruning strategies (quantization-insensitive top-k attention preservation, robot-aware token protection, and farthest-point spatial sampling) plus a Hadamard transform applied to query/key projections to make attention distributions more pruning-friendly. Experiments on two CogACT variants across four manipulation tasks report a 1.93x speedup, up to a 4.5% average success-rate improvement over the FP16 baseline, and reduced GPU memory. Ablations vary the pruning ratio and the contribution of each strategy.

Significance. If the central co-design claim is sustained, this is a practically valuable result: it would show that aggressive W4A4 quantization and token pruning can be combined without retraining, while preserving or even improving task success. The paper has several genuine strengths: Eq. (6) is a correct Hadamard identity; the baselines are external FP16 pruning methods; the evaluation covers multiple tasks and two model variants; and code is released. However, the evidence as presented does not establish the claimed quantization-aware interaction, because the paper's own ablation (Table 4) shows the explicitly quantization-insensitive top-k criterion degrading W4A4 performance, and no FP16 ablation of the same pruning pipeline is provided. The contribution is therefore promising but needs re-analysis and additional experiments.

major comments (4)
  1. The first proposed pruning strategy, quantization-insensitive top-k preservation (Eq. 1), is the paper's explicitly quantization-aware criterion. Yet adding it to W4A4 quantization lowers the Visual Matching average from 71.78 to 70.90 and the Variant Aggregation average from 58.18 to 57.23. The text in Section 4.4 states this integration 'affirms the basic compatibility' of quantization and pruning, which is contradicted by these numbers. The large improvements in Table 4 come from robot-aware protection (+3.08 Visual Matching) and spatially-aware FPS (+5.32), both of which are geometric/sampling techniques that are not specific to quantized models. To support the 'synergistic quantization-aware pruning' narrative, the authors need to show that their pruning pipeline interacts positively with quantization, e.g., by ablating the same pruning components on the FP16 model and demonstrating
  2. The pruning ratio r=0.4 is selected as the best value on the same four tasks (and two variants) that are later used for the headline results in Tables 1 and 2. This is post-hoc selection on the evaluation set, which inflates the reported success rates. Additionally, all success rates appear to be single-run numbers with no trial counts, seeds, or confidence intervals. The main visual-matching gain of +4.5% is heavily driven by the Place Apple task (Table 4: 49.7 -> 64.8), and with no variance estimate it is impossible to rule out seed noise. The authors should report multiple independent runs or at least standard errors, and ideally validate the ratio choice on held-out tasks or a separate split.
  3. There are quantitative inconsistencies in the efficiency claims. The text and Figure 4b report GPU memory decreasing from 14.3 GB to 7.6 GB, which is a 46.9% reduction, not 'over 73%' as stated in Section 5. In addition, the claim of a '36% improvement over EfficientVLA' does not match the reported speedups: 1.93x vs. 1.59x is a 21.4% relative improvement. These are load-bearing numbers in the abstract and conclusion and must be corrected.
  4. The quantization recipe is underspecified. The text says 'channel-wise quantization for activations' while the contributions and Section 1 describe a 'per-tensor Hadamard transformation'. The paper does not specify the weight quantization granularity, the calibration data or procedure, how scales and zero-points are computed, or the exact hyperparameters k (top-k count), R_t (robot ring radius), and m (FPS sample count) used in the reported results. Since SQAP-VLA is claimed to be training-free and post-hoc, a complete and reproducible recipe is essential to validate the method.
minor comments (5)
  1. The evaluation simulator is described inconsistently: Section 4.3 refers to the 'Simpler simulator', while Section 5 says 'ManiSkill2 benchmark'. These are different benchmarks; please clarify which one was used.
  2. The table header is hard to read: 'Pruning Ratio' and 'Baseline (FP16)' appear as one row, with N/A (W4A4) and 0.3-0.6 under 'Pruning Ratio'. A clearer multi-row header would avoid confusion.
  3. The model name is written inconsistently as both 'CogAct' and 'CogACT'. Please unify.
  4. The metric 'BOPs' is stated to stand for 'Basic OPerations' but no formula or hardware model is given for how it is computed. Define it precisely, since it is used to compare efficiency across methods.
  5. The phrases 'the first structured, training-free VLA inference acceleration framework' and 'state-of-the-art quantization' are stronger than the evidence presented; the comparison set does not include recent W4A4 VLA quantization methods such as QAIL. Please temper or expand the comparison.

Circularity Check

0 steps flagged

No circularity: core derivation is an orthogonal Hadamard identity plus measured benchmark comparisons; no fitted parameter is renamed as a prediction.

full rationale

Walking the claimed derivation chain: (1) The pruning criteria in Eqs. 1, 4, and 5 are defined from attention weights, robot world-coordinate projection, and geometric farthest-point sampling; none of these definitions encodes the success-rate targets or the FP16 baseline numbers. (2) The quantizer enhancement in Eq. 6 is the exact Hadamard orthogonality identity (W^T H^T)(H X)=W^T X; it is a true rotation and does not assume the claimed outcome. (3) The headline success rates in Tables 1-2 are measured against external FP16 baselines and FP16 pruning methods on the ManiSkill2 simulator, not derived from fitted constants. (4) The only tunable input is the pruning ratio (0.4), selected from the Table 3 ablation grid on the same benchmark; this is test-set selection and can inflate the reported 4.5% gain, but it is not a self-referential derivation or a renamed prediction because the reported success rate is an empirical measurement, not an algebraic consequence of the ratio. (5) Overlapping-author citations (FBQuant, Mole-VLA) are related-work references and are not used to justify the central co-design claim. (6) Table 4 shows the quant-insensitive top-k step alone degrades W4A4 performance, which weakens the paper's narrative, but that is a support/evidence issue, not circularity. No circular step found.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The method introduces no new physical entities. Its free parameters are compression hyperparameters and quantization scales, several of which are unreported. The axiomatic burden is concentrated in empirical claims about attention-score stability and the availability of robot-camera calibration, both of which are load-bearing for the proposed pruning criteria.

free parameters (5)
  • token pruning ratio r = 0.4
    Selected from ablation Table 3 as the value with the highest average success rate on the same four evaluation tasks used for the final headline results.
  • top-k count k for attention preservation
    Controls the size of K_attn in Eq. 1; the paper says it is adjusted with the pruning ratio but does not report the actual value.
  • robot ring radius R_t
    Defines the Chebyshev-distance ring around the robot end-effector token in Eq. 4; no value or sensitivity analysis is reported.
  • FPS sample count m
    The number of tokens selected by farthest-point sampling in Eq. 5; determined by the remaining token quota but not specified explicitly.
  • W4A4 quantization scales and zero-points
    The per-tensor activation and weight quantization parameters are fitted to calibration data but the calibration set and procedure are not described. The quality of these scales determines whether the quantized model retains usable attention structure.
axioms (5)
  • domain assumption Top-k attention indices are largely invariant to W4A4 quantization noise
    Assumed in Section 3.2 Strategy 1 and Eq. 1 from 'empirical analysis' that is not quantified. The paper's own Table 4 shows adding this strategy alone reduced visual matching success from 71.78 to 70.90, weakening the assumption.
  • domain assumption Robot pose and camera intrinsics/extrinsics are available and accurate
    Eq. 2 projects 3D robot coordinates into pixel coordinates; deployment scenarios without calibrated cameras or robot state would break this strategy.
  • domain assumption Farthest-point sampling on token coordinates preserves task-relevant information
    Eq. 5 assumes spatial coverage is a good proxy for information content after top-k and robot-ring tokens are protected; no proof is offered, and the large ablation effect suggests task-specific behavior.
  • standard math Hadamard rotation is orthogonal and does not change attention outputs
    Eq. 6 relies on H^T H = I, which is true for a properly normalized Hadamard matrix; this is standard linear algebra.
  • domain assumption High attention scores before quantization correspond to task-critical visual objects
    The entire quantization-insensitive framing assumes the full-precision attention map is a reliable saliency signal, as illustrated qualitatively in Figure 2a but not quantitatively validated.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SQAP-VLA: A Synergistic Quantization-Aware Pruning Framework for High-Performance Vision-Language-Action Models." pith.science (2026). https://pith.science/paper/PBQCFJAB

@misc{pith2026250909090,
  author       = {Pith},
  title        = {Pith review of: SQAP-VLA: A Synergistic Quantization-Aware Pruning Framework for High-Performance Vision-Language-Action Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PBQCFJAB}},
  note         = {Machine review of arXiv:2509.09090}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Vision-Language-Action (VLA) models exhibit unprecedented capabilities for embodied intelligence. However, their extensive computational and memory costs hinder their practical deployment. Existing VLA compression and acceleration approaches conduct quantization or token pruning in an ad-hoc manner but fail to enable both for a holistic efficiency improvement due to an observed incompatibility. This work introduces SQAP-VLA, the first structured, training-free VLA inference acceleration framework that simultaneously enables state-of-the-art quantization and token pruning. We overcome the incompatibility by co-designing the quantization and token pruning pipeline, where we propose new quantization-aware token pruning criteria that work on an aggressively quantized model while improving the quantizer design to enhance pruning effectiveness. When applied to standard VLA models, SQAP-VLA yields significant gains in computational efficiency and inference speed while successfully preserving core model performance, achieving a $\times$1.93 speedup and up to a 4.5\% average success rate enhancement compared to the original model.

Figures

Figures reproduced from arXiv: 2509.09090 by Hengyu Fang, Huanrui Yang, Li Du, Yijiang Liu, Yuan Du.

Figure 1
Figure 1. Figure 1: Overview of SQAP-VLA framework. SQAP-VLA resolves the incapability of token pruning on quantized VLA models via a quantization-aware pruning criteria. We propose insensitive preservation, robot-aware projection, and spatial-aware sampling to counter the scattered and shifted attention score of the quantized VLA model, enabling high performance, improved speed, and reduced model size with sparse tokens on a… view at source ↗
Figure 2
Figure 2. Figure 2: The attention heatmap before and after quantization. (a) and (c) Before quantization, attention is sharply [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of activation distributions. (a) Original activations are dominated by large-magnitude outliers in [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Latency and memory experiments. (a) Ablation study on pruning and quantization of SQAP-VLA. (b) GPU memory comparison of the baseline, EfficientVLA, FastV, VLA-Cache, and SQAP-VLA. We conducted ablation studies on token pruning ratios varying from 0.3 to 0.6 and compression strategies including quantization, quantization-insensitive preservation, robot-aware protection, and spatially-aware sampling. Prunin… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 9 Pith papers

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

  1. Revisiting Parameter Redundancy in Vision-Language-Action Models: Insights from VLM-to-VLA Adaptation

    cs.RO 2026-06 unverdicted novelty 7.0

    VLA models from VLM adaptation can be pruned 12-30% via multi-module joint scheme based on divergence signals while keeping ~90% performance on LIBERO without post-pruning recovery, unlike standard criteria that collapse.

  2. Latent Bridge: Feature Delta Prediction for Efficient Dual-System Vision-Language-Action Model Inference

    cs.RO 2026-05 unverdicted novelty 7.0

    Latent Bridge predicts VLM feature deltas to reduce VLM calls by 50-75% in dual-system VLA models while retaining 95-100% performance and achieving 1.65-1.73x speedup across LIBERO, RoboCasa, and ALOHA benchmarks.

  3. Towards Joint Quantization and Token Pruning of Vision-Language Models

    cs.CV 2026-04 unverdicted novelty 6.0

    QUOTA jointly optimizes low-bit quantization and visual token pruning for VLMs by deriving pruning decisions from quantized operators, achieving 95.65% average performance retention with only 30% of visual tokens vers...

  4. A1: A Fully Transparent Open-Source, Adaptive and Efficient Truncated Vision-Language-Action Model

    cs.RO 2026-04 unverdicted novelty 6.0

    A1 is a transparent VLA framework achieving state-of-the-art robot manipulation success with up to 72% lower latency via adaptive layer truncation and inter-layer flow matching.

  5. FASTER: Rethinking Real-Time Flow VLAs

    cs.RO 2026-03 unverdicted novelty 6.0

    FASTER adds a Horizon-Aware Schedule to flow VLAs that compresses immediate-action denoising to one step while keeping long-horizon trajectory quality, lowering real-robot reaction latency.

  6. FASTER: Rethinking Real-Time Flow VLAs

    cs.RO 2026-03 conditional novelty 6.0

    FASTER uses a horizon-aware flow sampling schedule to compress immediate-action denoising to one step, slashing effective reaction latency in real-robot VLA deployments.

  7. OxyGen: Unified KV Cache Management for VLA Inference under Multi-Task Parallelism

    cs.RO 2026-03 unverdicted novelty 6.0

    OxyGen unifies KV cache management in MoT VLAs to enable cross-task KV sharing and cross-frame continuous batching, delivering up to 3.7x speedup with 200+ tokens/s language and 70 Hz action on on-device platforms.

  8. DA-PTQ: Drift-Aware Post-Training Quantization for Efficient Vision-Language-Action Models

    cs.RO 2026-04 unverdicted novelty 5.0

    DA-PTQ quantizes VLAs by compensating cross-space distortions and allocating mixed precision to minimize motion errors and kinematic drift in trajectories.

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

    cs.RO 2025-10 conditional novelty 4.0

    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

27 extracted references · 24 linked inside Pith · cited by 8 Pith papers

  1. [2]

    URLhttps://arxiv.org/abs/2411.19650. Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil J...

    Pith/arXiv arXiv

  2. [3]

    9 SQAP-VLAA PREPRINT Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W

    URL https://arxiv.org/abs/2410.24164. 9 SQAP-VLAA PREPRINT Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W. Mahoney, and Kurt Keutzer. A survey of quantization methods for efficient neural network inference,

    Pith/arXiv arXiv

  3. [5]

    Hanrui Wang, Zhekai Zhang, and Song Han

    URLhttps://arxiv.org/abs/2106.08295. Hanrui Wang, Zhekai Zhang, and Song Han. Spatten: Efficient sparse attention architecture with cascade token and head pruning. In2021 IEEE International Symposium on High-Performance Computer Architecture (HPCA). IEEE, February

    Pith/arXiv arXiv

  4. [6]

    URL http://dx.doi.org/10.1109/HPCA51647.2021

    doi:10.1109/hpca51647.2021.00018. URL http://dx.doi.org/10.1109/HPCA51647.2021. 00018. Yujun Lin, Haotian Tang, Shang Yang, Zhekai Zhang, Guangxuan Xiao, Chuang Gan, and Song Han. Qserve: W4a8kv4 quantization and system co-design for efficient llm serving,

    arXiv 2021

  5. [8]

    Wenqi Shao, Mengzhao Chen, Zhaoyang Zhang, Peng Xu, Lirui Zhao, Zhiqian Li, Kaipeng Zhang, Peng Gao, Yu Qiao, and Ping Luo

    URL https://arxiv.org/abs/2310.19102. Wenqi Shao, Mengzhao Chen, Zhaoyang Zhang, Peng Xu, Lirui Zhao, Zhiqian Li, Kaipeng Zhang, Peng Gao, Yu Qiao, and Ping Luo. Omniquant: Omnidirectionally calibrated quantization for large language models,

    Pith/arXiv arXiv

  6. [9]

    Yifei Liu, Mathias Gehrig, Nico Messikommer, Marco Cannici, and Davide Scaramuzza

    URL https://arxiv.org/abs/2308.13137. Yifei Liu, Mathias Gehrig, Nico Messikommer, Marco Cannici, and Davide Scaramuzza. Revisiting token pruning for object detection and instance segmentation, 2023a. URLhttps://arxiv.org/abs/2306.07050. Andrey Kuzmin, Markus Nagel, Mart van Baalen, Arash Behboodi, and Tijmen Blankevoort. Pruning vs quantization: Which is...

    Pith/arXiv arXiv

  7. [10]

    Tailin Liang, John Glossner, Lei Wang, Shaobo Shi, and Xiaotong Zhang

    URLhttps://arxiv.org/abs/2307.02973. Tailin Liang, John Glossner, Lei Wang, Shaobo Shi, and Xiaotong Zhang. Pruning and quantization for deep neural network acceleration: A survey,

    Pith/arXiv arXiv

  8. [11]

    Benjamin Hawks, Javier Duarte, Nicholas J

    URLhttps://arxiv.org/abs/2101.09671. Benjamin Hawks, Javier Duarte, Nicholas J. Fraser, Alessandro Pappalardo, Nhan Tran, and Yaman Umuroglu. Ps and qs: Quantization-aware pruning for efficient low latency neural network inference.Frontiers in Artificial Intelligence, 4, July

    Pith/arXiv arXiv

  9. [12]

    doi:10.3389/frai.2021.676564

    ISSN 2624-8212. doi:10.3389/frai.2021.676564. URL http://dx.doi.org/10.3389/frai.2021. 676564. Ye Li, Yuan Meng, Zewen Sun, Kangye Ji, Chen Tang, Jiajun Fan, Xinzhu Ma, Shutao Xia, Zhi Wang, and Wenwu Zhu. Sp-vla: A joint model scheduling and token pruning approach for vla model acceleration, 2025b. URL https://arxiv.org/abs/2506.12723. Weihao Ye, Qiong W...

    arXiv 2021

  10. [13]

    Kilian Kleeberger and Marco F

    URLhttps://arxiv.org/abs/2409.10197. Kilian Kleeberger and Marco F. Huber. Single shot 6d object pose estimation,

    Pith/arXiv arXiv

  11. [16]

    URLhttps://arxiv.org/abs/2308.07633. Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakr- ishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Le...

    Pith/arXiv arXiv

  12. [18]

    org/abs/2412.01034

    URL https://arxiv. org/abs/2412.01034. Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. Gptq: Accurate post-training quantization for generative pre-trained transformers,

    Pith/arXiv arXiv

  13. [19]

    Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han

    URLhttps://arxiv.org/abs/2210.17323. Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, and Song Han. Awq: Activation-aware weight quantization for llm compression and acceleration,

    Pith/arXiv arXiv

  14. [20]

    Yuxuan Sun, Ruikang Liu, Haoli Bai, Han Bao, Kang Zhao, Yuening Li, Jiaxin Hu, Xianzhi Yu, Lu Hou, Chun Yuan, Xin Jiang, Wulong Liu, and Jun Yao

    URLhttps://arxiv.org/abs/2306.00978. Yuxuan Sun, Ruikang Liu, Haoli Bai, Han Bao, Kang Zhao, Yuening Li, Jiaxin Hu, Xianzhi Yu, Lu Hou, Chun Yuan, Xin Jiang, Wulong Liu, and Jun Yao. Flatquant: Flatness matters for llm quantization,

    Pith/arXiv arXiv

  15. [21]

    Zechun Liu, Changsheng Zhao, Igor Fedorov, Bilge Soran, Dhruv Choudhary, Raghuraman Krishnamoorthi, Vikas Chandra, Yuandong Tian, and Tijmen Blankevoort

    URL https: //arxiv.org/abs/2410.09426. Zechun Liu, Changsheng Zhao, Igor Fedorov, Bilge Soran, Dhruv Choudhary, Raghuraman Krishnamoorthi, Vikas Chandra, Yuandong Tian, and Tijmen Blankevoort. Spinquant: Llm quantization with learned rotations, 2025b. URL https://arxiv.org/abs/2405.16406. Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang ...

    Pith/arXiv arXiv

  16. [22]

    Qizhe Zhang, Aosong Cheng, Ming Lu, Renrui Zhang, Zhiyong Zhuo, Jiajun Cao, Shaobo Guo, Qi She, and Shanghang Zhang

    URL https://arxiv.org/abs/2403.06764. Qizhe Zhang, Aosong Cheng, Ming Lu, Renrui Zhang, Zhiyong Zhuo, Jiajun Cao, Shaobo Guo, Qi She, and Shanghang Zhang. Beyond text-visual attention: Exploiting visual cues for effective token pruning in vlms, 2025a. URL https://arxiv.org/abs/2412.01818. Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christop...

    Pith/arXiv arXiv

  17. [23]

    Yantai Yang, Yuhao Wang, Zichen Wen, Luo Zhongwei, Chang Zou, Zhipeng Zhang, Chuan Wen, and Linfeng Zhang

    URLhttps://arxiv.org/abs/2210.09461. Yantai Yang, Yuhao Wang, Zichen Wen, Luo Zhongwei, Chang Zou, Zhipeng Zhang, Chuan Wen, and Linfeng Zhang. Efficientvla: Training-free acceleration and compression for vision-language-action models,

    Pith/arXiv arXiv

  18. [24]

    Siyu Xu, Yunke Wang, Chenghao Xia, Dihao Zhu, Tao Huang, and Chang Xu

    URL https://arxiv.org/abs/2506.10100. Siyu Xu, Yunke Wang, Chenghao Xia, Dihao Zhu, Tao Huang, and Chang Xu. Vla-cache: Towards efficient vision- language-action model via adaptive token caching in robotic manipulation,

    Pith/arXiv arXiv

  19. [25]

    Rongyu Zhang, Menghang Dong, Yuan Zhang, Liang Heng, Xiaowei Chi, Gaole Dai, Li Du, Yuan Du, and Shanghang Zhang

    URL https://arxiv.org/abs/ 2502.02175. Rongyu Zhang, Menghang Dong, Yuan Zhang, Liang Heng, Xiaowei Chi, Gaole Dai, Li Du, Yuan Du, and Shanghang Zhang. Mole-vla: Dynamic layer-skipping vision language action model via mixture-of-layers for efficient robot manipulation, 2025b. URLhttps://arxiv.org/abs/2503.20384. Siddharth Karamcheti, Suraj Nair, Ashwin B...

    arXiv

  20. [26]

    URL https://arxiv.org/ abs/2402.07865. Embodiment Collaboration, Abby O’Neill, Abdul Rehman, Abhinav Gupta, Abhiram Maddukuri, Abhishek Gupta, Abhishek Padalkar, Abraham Lee, Acorn Pooley, Agrim Gupta, Ajay Mandlekar, Ajinkya Jain, Albert Tung, Alex Bewley, Alex Herzog, Alex Irpan, Alexander Khazatsky, Anant Rai, Anchit Gupta, Andrew Wang, Andrey Kolobov,...

    Pith/arXiv arXiv

  21. [27]

    URLhttps://arxiv.org/abs/2310.08864. 12

    Pith/arXiv arXiv

  22. [2017]

    Albert Tseng, Jerry Chee, Qingyao Sun, V olodymyr Kuleshov, and Christopher De Sa

    URLhttps://arxiv.org/abs/1706.02413. Albert Tseng, Jerry Chee, Qingyao Sun, V olodymyr Kuleshov, and Christopher De Sa. Quip#: Even better llm quantization with hadamard incoherence and lattice codebooks,

    Pith/arXiv arXiv

  23. [2020]

    Charles R

    URL https://arxiv.org/abs/ 2004.12729. Charles R. Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space,

    Pith/arXiv arXiv 2004

  24. [2021]

    Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort

    URLhttps://arxiv.org/abs/2103.13630. Markus Nagel, Marios Fournarakis, Rana Ali Amjad, Yelysei Bondarenko, Mart van Baalen, and Tijmen Blankevoort. A white paper on neural network quantization,

    Pith/arXiv arXiv

  25. [2023]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee

    URL https://arxiv.org/abs/2303.03378. Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023b. URL https: //arxiv.org/abs/2304.08485. Seongmin Park, Hyungmin Kim, Wonseok Jeon, Juyoung Yang, Byeongwook Jeon, Yoonseon Oh, and Jungwook Choi. Quantization-aware imitation-learning for resource-efficient robotic control,

    Pith/arXiv arXiv

  26. [2024]

    URL https://arxiv.org/abs/2406.09246. Qixiu Li, Yaobo Liang, Zeyu Wang, Lin Luo, Xi Chen, Mozheng Liao, Fangyun Wei, Yu Deng, Sicheng Xu, Yizhong Zhang, Xiaofan Wang, Bei Liu, Jianlong Fu, Jianmin Bao, Dong Chen, Yuanchun Shi, Jiaolong Yang, and Bain- ing Guo. Cogact: A foundational vision-language-action model for synergizing cognition and action in robo...

    Pith/arXiv arXiv

  27. [2025]

    Yijiang Liu, Hengyu Fang, Liulu He, Rongyu Zhang, Yichuan Bai, Yuan Du, and Li Du

    URLhttps://arxiv.org/abs/2405.04532. Yijiang Liu, Hengyu Fang, Liulu He, Rongyu Zhang, Yichuan Bai, Yuan Du, and Li Du. Fbquant: Feedback quantization for large language models, 2025a. URLhttps://arxiv.org/abs/2501.16385. Muyang Li, Yujun Lin, Zhekai Zhang, Tianle Cai, Xiuyu Li, Junxian Guo, Enze Xie, Chenlin Meng, Jun-Yan Zhu, and Song Han. Svdquant: Abs...

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.