acceptodds
Withdrawn from ICLR 2027

SparseVLA: Towards Smaller And Faster Vision-Language-Action Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-language-action (VLA) models offer a promising approach to general-purpose robotic control, but their large parameter counts contribute to substantial storage requirements and inference latency, hindering efficient deployment. Weight pruning is a representative approach to addressing this problem. However, when applied to VLA models, existing weight pruning methods typically use information from each control step in isolation, overlooking the sequential nature of VLA control. In practice, two challenges arise: 1) Methods may reconstruct individual outputs well but fail to preserve how they change across steps; 2) Sparse execution can slow down inference when runtime overheads outweigh computational savings. To solve these challenges, we introduce SparseVLA, a training-free framework for hardware-efficient VLA weight pruning and deployment. We find that activation changes between steps reflect task progression, suggesting their potential as a complementary signal for weight pruning. Since preserving inter-step activation changes can favor different pruning patterns than reconstructing individual outputs, SparseVLA balances the two criteria when selecting and reconstructing weights. To turn this sparsity into practical inference gains, SparseVLA combines shape-aware sparse execution with operator fusion, reducing memory traffic and kernel launch overhead. Across LIBERO, SimplerEnv, and real-world manipulation tasks, SparseVLA maintains or improves aggregate task success relative to dense baselines on two VLA backbones, and GR00T, while reducing model footprint and inference latency.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Post-Training Weight Pruning, Model Compression, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_X8TrOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents SparseVLA, a training-free method for weight pruning of vision-language action models. The key idea is to add an objective that preserves changes in layer activations between consecutive control steps along calibration trajectories, in addition to the standard per-step reconstruction objective used by prior methods. This objective is then used both to select a pruning mask and to reconstruct the retained weights. Additionally, the paper proposes an inference optimization strategy that uses sparse kernel only where they help and merges adjacent MLP operations into one kernel, which helps reduce inference latency. The paper reports that SparseVLA reduces model footprint and inference latency while maintaining or improving performance relative to the dense models, including pi-0.5 on LIBERO benchmark, GR00T N1.7 on SimplerEnv, and GR00T N1.6 on real-world manipulation tasks.

Critical strengths
  • The paper tries to address a timely problem by reducing the deployment cost (memory and latency) of VLA models, which is important for robotics.
  • The proposed method is training-free and does not require fine-tuning the VLA backbone, which makes it simple and practically useful.
  • The evaluation covers multiple VLA architectures (including Gr00t N1.6/N1.7 and pi-0.5) across two simulation benchmarks and real-world manipulation tasks.
Critical weaknesses
  1. The problem statement is not well motivated: The paper argues that per-step reconstruction lets action errors compound across control steps, so pruning should preserve inter-step activation changes. This claim relies on weak and inconsistent evidence:
  • Finding 1 (Figure 2a) consists of one example rollout plus a mean over three rollouts of a single task, with no variance shown. Moreover, the dense policy is itself stochastic (flow-matching action sampling), so part of the 5.26 cm deviation is sampling noise rather than pruning error. The paper does not report a dense-vs-dense control under different sampling noise, so it is unclear how much of the deviation is caused by pruning at all.
  • Finding 2 shows that HJH_J produces a different mask than H0H_0 (3.5–14.6% disagreement). However, this only shows that changing the curvature changes the mask, which any reweighting would do. It does not show the new mask is better.
  1. The efficiency gains claimed in the paper are not attributable to its contributions: The title and abstract claim that the proposed method is smaller and faster, but the results do not support this:
  • On inference latency, under selective sparse execution (Table 2a), the speedup on pi-0.5 is 1.018$\times$ (1.8% faster), which is negligible and indistinguishable from the other baselines (Magnitude, Wanda, SparseGPT). Under full-sparse execution, inference is slower than the dense baseline for both pi-0.5 and GR00T N1.7.
  • On memory, the 20.95% saving on pi-0.5 is identical for every 2:4 baselines in Table 1, so it is a property of 2:4 sparsity rather than of the proposed method. On GR00T, the saving is 5.2% in simulation and 4.8% on the real robot. For GR00T, the full framework therefore delivers ~5% less memory (Table 1) and ~1% lower latency (0.9% and 0.4% faster than dense in eager mode, Table 2a), which does not justify the “smaller and faster” claim.

Additionally, the presentation is hard to follow because the problem setting is not clearly established. Overall, the proposed pruning criterion has no clearly demonstrated effect on efficiency, and the efficiency gains that are demonstrated come from a prior backend. I therefore do not recommend acceptance of the current version.

Questions for authors
  • EfficientVLA [1] and module-aware pruning from [2] are cited but not compared against, even though they are training-free methods that directly address VLA pruning. Why were they excluded? Could the authors compare against these methods under the same protocol as in Table 1? A latency comparison against a token-pruning method such as ADP [3], which gives wall-clock speedups without custom kernels, would also be informative.
  • No standard deviations, seeds, or confidence intervals are reported for any success rate in the main results. Could the authors report results across multiple pruning seeds (e.g., different calibration-trajectory draws) and evaluation seeds, with standard deviations? In particular, are the improvements of SparseVLA over Dense (+1.1 on LIBERO and +2.6 on SimplerEnv, Table 1) and over SparseGPT, as well as the 0.10 / 0.90 point ablation gains on pi-0.5 (Table 4a), statistically distinguishable from evaluation noise?

[1] Yang et al., “EfficientVLA: Training-free acceleration and compression for vision-language-action models,” NeurIPS 38, 2025.
[2] Zhang, Fengnian, et al. "Revisiting Parameter Redundancy in Vision-Language-Action Models: Insights from VLM-to-VLA Adaptation." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.
[3] Pei, Xiaohuan, et al. "Action-aware dynamic pruning for efficient vision-language-action manipulation." International Conference on Learning Representations. Vol. 2026. 2026.

Reviewer_BA8ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose SparseVLA, a training-free weight pruning method for VLA models using semi-structured sparsity. Its key insight is that preserving per-step outputs alone may be insufficient for closed-loop control, where action errors can accumulate over time. The method combines transition-aware mask selection, OBS-based weight reconstruction that balances output fidelity and inter-step activation changes, and hardware-aware sparse execution with operator fusion. Experiments on pi0.5 and GR00T on LIBERO, SimplerEnv, and real-robot manipulations show that the proposed method largely preserves task performance while reducing model footprint and inference latency.

Critical strengths

S1 - good problem formulation. The paper explicitly considers temporal consistency in closed-loop VLA control, extending beyond conventional per-sample reconstruction objectives.

S2 - a technically coherent approach. The combination of correlated 2:4 mask scoring, transition-aware selection, and reconstruction is well motivated. Ablations provide evidence that temporal ordering contributes to performance.

S3 - broad empirical evaluation. Results cover two VLA families, multiple simulation benchmarks, and real-robot experiments. SparseVLA achieves 97.25% versus 96.15% on LIBERO, 65.77% versus 63.15% on SimplerEnv, and the same aggregate 59/80 successes as the dense baseline on the real robot.

S4 - practical deployment considerations. Shape-aware kernel selection and mask-preserving operator fusion address the important gap between theoretical sparsity and actual inference speed.

Critical weaknesses

W1 - architecture-dependent efficiency gains. pi0.5 achieves a 1.215× speedup, while GR00T achieves only approximately 1.03×. The practical benefit therefore varies substantially by architecture and execution path.

W2 - limited evidence for generalization. Calibration relies on a relatively small number of trajectories. The experiments do not sufficiently establish robustness to calibration distribution shifts, unseen tasks, or substantially different deployment conditions.

W3 - the closed-loop motivation is not fully established. The ablations support the usefulness of transition-aware statistics, but provide limited direct evidence that the method consistently reduces error accumulation or improves long-horizon robustness.

Questions for authors

Q1 - how sensitive is SparseVLA to calibration trajectory coverage and distribution shift? Can the authors evaluate smaller calibration sets or unseen task conditions?

Q2 - why are the matched-backend speedups so different between π0.5 and GR00T? Can the authors quantify the contributions of pruning, kernel selection, and operator fusion separately?

Q3 - can the authors directly connect transition-reconstruction error to action deviation and closed-loop task failure, beyond aggregate success-rate ablations?

Reviewer_hR4ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SparseVLA, a training-free framework for VLA model weight pruning and inference optimization. The authors point out that existing weight pruning methods primarily focus on output reconstruction errors at individual control steps while overlooking temporal relationships in sequential robotic control, which may lead to the accumulation of pruning errors during closed-loop execution. To address this issue, SparseVLA leverages activation changes between consecutive control steps in calibration trajectories to guide 2:4 semi-structured weight pruning while preserving per-step output reconstruction accuracy, and compensates for pruning errors in the retained weights. In addition, the authors combine shape-aware sparse execution and operator fusion to reduce model storage requirements and inference overhead. Experiments are conducted on two VLA models, π0.5\pi_{0.5} and GR00T, covering LIBERO, SimplerEnv, and real-world robotic manipulation tasks. The results show that SparseVLA can maintain or improve overall task success rates while reducing model memory usage and inference latency.

Critical strengths
  • Introducing temporal information from robotic control into weight pruning provides a reasonably motivated and somewhat novel approach. Existing weight pruning methods primarily focus on per-step output reconstruction, whereas this paper further considers activation changes between consecutive control steps to improve the performance of pruned models during sequential control. The authors incorporate temporal information into both pruning pattern selection and weight reconstruction, and validate its effectiveness through ablation studies and experiments that shuffle the temporal order of calibration trajectories. This design meaningfully extends traditional weight pruning methods by considering the closed-loop nature of VLA control.
  • The paper considers not only model compression but also the practical inference efficiency of sparse models. The authors recognize that weight sparsity does not necessarily translate into actual inference speedups. They therefore develop shape-aware sparse execution and operator fusion techniques to reduce the additional overhead associated with sparse computation. The paper also provides latency comparisons under the same optimized backend and ablation studies on operator fusion, offering some experimental evidence for the effectiveness of the system optimizations.
  • The experiments cover different VLA models, simulation benchmarks, and real-world robotic tasks. The paper evaluates two VLA models, π0.5\pi_{0.5} and GR00T, across LIBERO, SimplerEnv, and four real-world robotic manipulation tasks, and compares SparseVLA with several existing weight pruning methods. The results show that SparseVLA can maintain or improve overall task success rates while reducing model memory usage. In particular, it achieves notable performance improvements over existing pruning baselines on GR00T, providing experimental support for the effectiveness of the proposed method.
Critical weaknesses
  • The inference acceleration experiments do not sufficiently distinguish the contributions of weight pruning from other backend optimizations. The paper reports an overall inference speedup of approximately 2×, but this result includes additional inference backend optimizations and cannot be entirely attributed to SparseVLA's weight pruning. It remains unclear how much of the approximately 2× speedup in the real-robot experiments comes from sparsity itself. The authors are encouraged to provide latency comparisons for the real-robot models under the same optimized backend and further clarify the contributions of different system optimization components.
  • The inference speedup is limited. Under the same optimized backend, the speedup on π0.5\pi_{0.5} is approximately 1.2×, while the actual sparse execution speedup on GR00T is only 1.026–1.029×. This suggests that the acceleration benefits of SparseVLA itself may not generalize to models beyond π0.5\pi_{0.5}.
  • The mechanism by which transition-aware pruning improves closed-loop control performance lacks sufficient experimental analysis. The paper validates the contribution of temporal information to task success rates through ablation studies and shows in Figure 2(a) that SparseVLA reduces closed-loop trajectory deviations relative to the Dense policy. However, the figure only includes three paired rollouts from a single task, which is insufficient to establish whether this observation holds across different tasks and execution conditions. The authors are encouraged to report closed-loop trajectory deviations across more tasks and further investigate the underlying mechanism through an analysis of action prediction errors.
Questions for authors
  • SparseVLA introduces multiple hyperparameters. How sensitive is the final task success rate to these parameters? Do different VLA models require separate hyperparameter tuning?
  • The authors should clearly specify which components are included in the backend optimizations.
  • Could the authors provide comparisons with some structured weight pruning methods?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.