acceptodds
Under review as a conference paper at ICLR 2027

VADP-VLA: View-Adaptive Dynamic Visual Token Pruning for Efficient VLA Inference

Abstract

Vision-language-action (VLA) models generate robot actions from visual observations, language instructions, and robot states, but the large number of visual tokens produced by multiple cameras increases inference cost. Existing methods reduce visual computation by exploiting visual saliency, temporal redundancy, or action relevance. Because camera views differ in both content and temporal dynamics, however, the validity of historical information and the appropriate visual token budget must be assessed separately for each view. We propose VADP-VLA, a training-free framework for view-adaptive dynamic visual token pruning. It estimates language-conditioned visual token saliency within each view and transfers historical saliency through a per-view keyframe cache. Agreement between current and historical saliency controls cache refresh and token retention independently for each view. Tokens that are not retained are removed before later language-model layers, shortening the sequence processed by subsequent computation. Across the four LIBERO suites, VADP-VLA achieves an average success rate of 95.25%, compared with 94.55% for the unpruned OpenVLA-OFT baseline, while reducing normalized language-model FLOPs to 41% of the baseline and model-internal latency from 73.12 ms to 48.89 ms, a 1.50× speedup. In LIBERO experiments with , average success changes from 96.30% with the unpruned baseline to 94.65%, normalized language-model FLOPs fall to 45%, and model-internal latency decreases from 37.38 ms to 25.43 ms. Across four real-robot tasks, each method is evaluated in 80 trials: VADP-VLA succeeds in 63/80 trials versus 67/80 for the unpruned baseline, while VLA model-call latency falls from 59.93 ms to 42.58 ms, approximately a 1.41× speedup. These results show reduced VLA inference cost in the evaluated settings, alongside a modest decrease in real-robot task success relative to the unpruned baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, visual token pruning, efficient inference, temporal saliency transfer, robotic manipulation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_cZVXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes VADP-VLA, a training-free visual token pruning framework for multi-camera vision-language-action policies. It computes language-conditioned visual saliency at a shallow layer and transfers historical saliency through feature matching within each camera view. Agreement between current and historical saliency controls cache refresh and token retention independently for each view. Unselected visual tokens are removed before subsequent language-model layers. Experiments evaluate OpenVLA-OFT and π0.5 on LIBERO, together with π0.5 on four real-robot tasks.

Critical strengths

The method demonstrates measurable inference savings across simulation and real-robot evaluations. With OpenVLA-OFT on LIBERO, average success changes from 94.55% to 95.25% while model-internal latency decreases from 73.12 ms to 48.89 ms. The π0.5 experiments also show measured latency reductions, including 59.93 ms to 42.58 ms for real-robot VLA calls, with success decreasing from 67/80 to 63/80. These results provide concrete evidence of practical inference savings, and the associated performance trade-offs are explicitly reported.

Critical weaknesses
  1. The methodological advance over the closest prior work is limited and insufficiently demonstrated. SAFE-Pruner already transfers saliency from cached keyframes and refreshes the reference using saliency agreement. The main distinguishing contribution of VADP-VLA is independent per-view control, but the paper does not directly compare against SAFE-Pruner. Consequently, the additional value of the proposed design over this closely related framework is not convincingly established.

  2. The experiments do not sufficiently isolate the central per-view contribution. Table 3 reports 89/100 versus 93/100 successes, but the shared-refresh rule, task composition, and uncertainty are insufficiently specified. The component ablations are limited to LIBERO-Spatial and do not establish the benefit of adaptive control over static per-view budgets at matched compute. This leaves the gains attributable specifically to the proposed controller unclear.

  3. The comparative benefit beyond OpenVLA-OFT remains unclear. The π0.5 and real-robot experiments compare only against unpruned policies. Although measured latency decreases, reported success also drops by 1.65 and 5.00 percentage points, respectively. Without comparisons to alternative pruning methods at comparable latency or success, these results do not establish a clear comparative advantage in these settings.

  4. The adaptation to π0.5 is insufficiently specified. The detailed algorithm and complexity analysis mainly follow the OpenVLA-OFT computation pipeline. The paper does not clearly explain how the saliency probe and pruning procedure are instantiated within π0.5 and interact with its flow-matching action generation. This limits reproducibility of the reported cross-architecture results.

Questions for authors
  1. What benefit comes specifically from independent per-view refresh and budgeting beyond SAFE-Pruner’s existing saliency-transfer framework? How do the methods compare under the same backbone and evaluation protocol?

  2. How is shared refresh defined in Table 3, and which tasks comprise its 100 trials? Do the improvements persist against static per-view budgets at matched average compute, with uncertainty estimates?

3.How does VADP-VLA compare with existing pruning methods at comparable latency or success in these settings? The current comparisons are only against the unpruned policy.

Reviewer_B4stOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes VADP-VLA, a training-free visual token pruning method for multi-view VLA models. It combines current text–vision relevance with historical saliency and independently adjusts cache refresh and token budgets for global and wrist-camera views. Experiments on OpenVLA-OFT and π₀.₅ show reduced computation and latency, with additional evaluation on real-robot tasks.

Critical strengths
  1. A practical motivation for multi-view inference.

Global and wrist cameras can have different temporal changes and information requirements. Handling their historical information and pruning decisions separately is a reasonable approach to reducing VLA inference cost.

  1. Useful latency reductions and substantial simulation evaluation.

The main LIBERO comparison includes 2,000 episodes per configuration. On OpenVLA-OFT, reported average success increases from 94.55% to 95.25%, while model-internal latency decreases from 73.12 to 48.89 ms. The π₀.₅ and real-robot experiments also show latency reductions, and the paper reports the accompanying success-rate losses clearly.

Critical weaknesses
  1. The comparison with fixed per-view budgets needs to control for computational cost.

    The method combines dynamic decisions with manually chosen, asymmetric budgets for the two cameras. Table 4 includes a static Raw-QK baseline, but the configurations have different computational costs. It remains unclear whether the dynamic controller outperforms a tuned fixed allocation between the two views. Comparing them at matched latency or average token counts would show whether the dynamic rules improve on a simpler allocation.

  2. The shared-refresh comparison needs more detail.

    Table 3 reports 89/100 successes with shared refresh and 93/100 with independent refresh, alongside a higher pruning ratio for the latter. This is encouraging, but the difference of four successful trials is reported without uncertainty estimates. This concern applies to the 100-trial ablation, not the larger main evaluation. The shared-refresh decision rule is also not clearly specified. Please explain how that baseline combines the two views' signals and whether the budget settings are otherwise identical. Repeated runs or uncertainty estimates would help assess how reliable the improvement is.

  3. Other acceleration baselines are missing on π₀.₅ and real hardware.

    Comparisons with other acceleration methods are concentrated on OpenVLA-OFT. The π₀.₅ and real-robot experiments compare VADP-VLA with the unpruned policy. They report lower latency and lower success rates: 96.30% to 94.65% on π₀.₅'s LIBERO evaluation and 83.75% to 78.75% on the real robot. These results show that pruning reduces latency on both, but leave open whether another method could achieve the same speed with less loss in success. Please compare with at least one compatible acceleration baseline at similar success rates or latency.

Questions for authors

Please refer to the weaknesses above.

Reviewer_1SmUOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VADP-VLA, a training-free visual token pruning framework for efficient vision-language-action (VLA) inference. The method addresses differences in temporal dynamics across camera views by maintaining independent historical saliency caches and dynamically allocating token budgets for each view. Specifically, it estimates task-conditioned visual saliency using language-model query-key interactions, transfers historical saliency through within-view token matching, and adjusts cache refresh and token retention based on saliency agreement. Experiments on LIBERO and real-world robotic tasks demonstrate reduced inference latency and computational cost, with varying effects on task success rates.

Critical strengths
  1. The paper identifies a relevant problem in multi-view VLA inference: different camera views exhibit distinct temporal dynamics and may require different token retention strategies. This provides a reasonable motivation for view-adaptive pruning.

  2. The proposed method is training-free and achieves meaningful reductions in computational cost and inference latency. Its compatibility with existing VLA architectures further supports its practical applicability.

Critical weaknesses

1. Misalignment Between Motivation and Experimental Validation.

The main concern is that the experiments do not sufficiently address the central challenges raised in the Introduction. The paper emphasizes asynchronous changes across camera views, the validity of historical information, and the necessity of view-specific token allocation. However, the evaluation primarily focuses on success rates, FLOPs, and inference latency, without adequately validating these underlying assumptions. Although Figure 2 provides some preliminary observations, the paper lacks systematic evidence that saliency agreement reliably reflects the validity of historical information or that the proposed adaptive allocation effectively responds to view-specific task dynamics. Consequently, the connection between the motivation, the proposed method, and the experimental results remains insufficiently established.

2. Insufficient Experimental Validation.

The real-world evaluation is limited to four tasks on a single robotic platform, with only 20 trials per task. Furthermore, the overall success rate decreases from 83.75% to 78.75%. The lack of statistical analysis and evaluations across camera configurations, or more complex manipulation scenarios further limits the generalizability of the conclusions.

3. Inadequate Ablation Studies.

The ablation studies do not sufficiently isolate the contributions of the proposed components. Although the paper compares shared and independent cache refresh, it lacks comprehensive comparisons between fixed token budgets, single-view adaptive budgets, and fully view-adaptive budgets under matched computational costs. While the full method provides lower latency, the necessity and specific benefits of its complete view-adaptive design remain insufficiently demonstrated.

Questions for authors
  1. How do the authors verify that the proposed saliency agreement reliably reflects the validity of historical information and changes in task-relevant visual content? Can the authors provide quantitative evidence linking saliency agreement to action prediction sensitivity?

  2. Can the authors provide computationally matched comparisons between fixed-budget pruning, single-view adaptive pruning, and the proposed fully view-adaptive strategy? This would help clarify whether the improvements arise specifically from view adaptivity rather than general token reduction.

  3. Can the authors provide quantitative analysis of view-specific token retention ratios and cache refresh decisions across different manipulation stages? How do these decisions correlate with changes in action prediction sensitivity?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.