- 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
- 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 produces a different mask than (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.
- 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.