acceptodds
Under review as a conference paper at ICLR 2027

Grounding Attention in Motion: Proprioception-Guided Visual Token Pruning with Adaptive Layer Selection for Vision-Language-Action Models

Abstract

Vision-Language-Action (VLA) models benefit from the perceptual generalization of pretrained vision-language backbones, but repeatedly processing hundreds of visual tokens makes closed-loop inference costly. Visual token pruning can reduce this cost, yet existing methods typically rank tokens using only vision-language interactions and prune at a fixed backbone layer, overlooking both the robot's current state and the input-dependent stability of token rankings. We introduce PropPrune, a training-free framework that adapts both which visual tokens are retained and where they are pruned. A proprio-guided importance score uses each instruction token's attention to the proprioceptive state to weight its attention to visual tokens, grounding visual relevance in the current embodied context. An adaptive layer selector then measures the Jaccard similarity between top-k visual-token sets from adjacent shallow layers and defers pruning to a later layer when the shallow ranking is unstable. Both components use attention maps already produced by the pretrained VLA, requiring no fine-tuning, auxiliary model, or additional forward pass. On LIBERO, PropPrune achieves the highest average success rate among the evaluated pruning baselines at both 50% and 25% visual-token retention, remaining within 0.7 and 1.3 points of the full-token model while reducing latency by 16.0% and 34.9% and FLOPs by 38.1% and 53.6%, respectively. On LIBERO-Plus, it improves over FastV under seven controlled distribution shifts and remains within 0.26 points of the full-token reference. On two real-world tasks, 50% retention matches full-token success while reducing latency by 19.1%.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Vision Language Action Models, Token Pruning, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hV79OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a method named PropPrune, which is a training-free visual token pruning method for VLA models. The proposed method uses proprioceptive information to score the importance of visual tokens. It also uses Jaccard similarity between adjacent layers to select the pruning layer for each input. Experiments are conducted on LIBERO, LIBERO-Plus, and two real-world tasks to evaluate the performance of PropPrune compared to baselines. The results show higher success rates than several pruning baselines while reducing inference costs.

Critical strengths
  • The proposed method is simple and does not require additional training. It achieves a 96.3% average success rate on LIBERO at 25% token retention, compared to 93.7% for FastV.

  • The experiments include distribution shifts and real-world manipulation tasks. These results help show the performance of the method beyond the standard LIBERO setting.

Critical weaknesses
  • The contribution of adaptive layer selection is not fully clear. Table 3 shows that always pruning at the later layer achieves a 97.1% average success rate, compared to 96.9% for PropPrune. However, this also increases FLOPs from 2.483T to 2.582T. The paper does not compare with randomly delaying pruning or delaying it at a fixed frequency under the same computation budget. Therefore, it is not clear if the gains come from using Jaccard similarity to select the layer. The improvement may mainly come from keeping all visual tokens for more layers on some inputs.

  • Appendix B shows that the average success rate only improves by 0.50 and 0.25 percentage points over language-only scoring at 50% and 25% retention, respectively. The improvement mainly comes from LIBERO-Long. The other suites show little improvement or worse performance. Additionally, there is no direct evidence to show if the selected visual regions change in a meaningful way when the robot state changes.

  • The novelty compared to existing action-aware VLA pruning methods is not fully clear. Action-aware Dynamic Pruning (ICLR 2026) already uses robot motion information to decide how many visual tokens to retain. SpecPrune-VLA (ICML 2026) also uses action-related signals for dynamic pruning across layers. PropPrune uses different attention weighting and layer-selection rules, but the paper does not compare with these methods. All main experiments use OpenVLA-OFT, whose Transformer includes visual, language, and proprioceptive tokens. Therefore, it is not clear if the results will hold with other VLA architectures or fusion mechanisms.

  • The real-world evaluation only includes two tabletop tasks with 20 trials for each task. At 50% retention, PropPrune and the full model both succeed in 38 of 40 trials. However, this number of trials is too small to show that the two models have the same success rate. The success rates on LIBERO are also close to 100%. It would be helpful to provide uncertainty estimates to show if the small performance differences are reliable.

Questions for authors
  • Can the authors compare Jaccard-based deferral with random deferral under the same late-pruning frequency and inference cost? This would help show the contribution of using Jaccard similarity to select the pruning layer.

  • Can the authors show if proprioceptive guidance changes token selection in a meaningful way when only the robot state changes? It would be helpful to keep the visual observation and instruction the same to show the effect of robot state.

  • Can the authors provide results on another VLA architecture to show if the performance gains will hold? This could include a model with a different proprioceptive representation or action decoder.

  • Can the authors provide confidence intervals or compare the failure cases of the methods on the same trials? This would help make the success-rate differences more clear, especially on LIBERO-Long and the real-world tasks.

6 comments

Response to Weakness 3 & Question 3 (1/3)

We thank the reviewer for the valuable suggestions on comparing our method with existing action-aware pruning methods and evaluating its performance on other VLA architectures.


In the original paper, we evaluated fixed configurations for the probe layer, deferred pruning layer, and Jaccard similarity threshold only on OpenVLA-OFT. We acknowledge that these experiments provide limited evidence for generalization across VLA architectures and raises concerns about the use of heuristic configurations, as also noted by reviewers WxwJ and 2Dvs.

To address these concerns, we conducted additional experiments on RynnVLA-002 and introduced a calibration rule to select pruning configurations for each model using training data, rather than searching for hyperparameters based on evaluation success rates.


The calibration procedure consists of the following steps:

  1. For each evaluation suite, we randomly select 10 training episodes from the corresponding tasks and sample frames every 8 frames.
  2. We forward the sampled observations through the VLA model and compute visual token importance at each Transformer layer using our proposed importance scoring method.
  3. For each pair of adjacent Transformer layers, we compute the Jaccard similarity between their selected top-k visual token sets and average the similarities across the sampled observations.
  4. We select the layer with the largest increase in average inter-layer Jaccard similarity as the probe layer and the next layer as the deferred pruning layer. The Jaccard similarity threshold is set to the average similarity at the probe layer.

This calibration rule reduces the need for manually selecting model-specific pruning configurations without changing the core components of our method, including proprioception-guided importance scoring and adaptive pruning. We also re-evaluated OpenVLA-OFT using the calibrated configuration. Although the success rates and computational costs differ slightly from the original results, the overall trends in task success rate and computational efficiency remain consistent. We will incorporate these additional experiments and the calibration procedure into the revised paper as soon as possible.

The updated results for both OpenVLA-OFT and RynnVLA-002 are presented in the following comment, together with comparisons against action-aware pruning methods.

Response to Weakness 3 & Question 3 (2/3)

Using the calibration rule described in our previous comment, we re-evaluated PROPPRUNE on OpenVLA-OFT and extended the evaluation to RynnVLA-002. The same procedure selected P4/L5 with Jaccard thresholds of 0.523 and 0.445 for OpenVLA-OFT at 50% and 25% retention, respectively, and P7/L8 with Jaccard thresholds of 0.605 for RynnVLA-002 at 50% retention.

All experiments were conducted using the corresponding model checkpoints and the same evaluation protocol, with 500 episodes per LIBERO suite.

The updated OpenVLA-OFT results are as follows.

MethodRetentionSpatialObjectGoalLongAvg. SRFLOPs (T)LLM latency (ms)
Full-token100%99.298.297.295.697.64.01377.5
PROPPRUNE (calibrated)50%99.097.697.494.297.12.54562.1
PROPPRUNE (calibrated)25%98.696.096.893.296.21.84248.6

With the calibrated configurations, PROPPRUNE achieved average success rates of 97.1% and 96.2% at 50% and 25% retention, respectively, compared with 97.6% for the full-token baseline. The corresponding LLM FLOPs decreased from 4.013T to 2.545T and 1.842T, respectively. Although these results differ slightly from the original submission due to the updated pruning configurations, the overall accuracy-efficiency trends remain consistent.


We also re-evaluated LIBERO-Plus using the calibrated pruning configuration. Independently of the configuration changes, the original LIBERO-Plus evaluation environment was lost, requiring us to reconstruct the evaluation setup. We therefore re-ran all methods, including the full-token baseline and FastV, under the same reconstructed environment to ensure a consistent comparison.

MethodCameraRobotLanguageLightBackgroundNoiseLayoutOverall
Full-token95.928.186.596.097.293.376.380.7
PROPPRUNE (50%)94.026.785.096.896.394.175.980.0
FastV (50%)92.025.582.894.094.190.475.178.0
ADP (50%)90.524.880.192.992.387.172.275.9

Under the reconstructed evaluation setup, PROPPRUNE achieved an overall success rate of 80.0%, close to the full-token baseline (80.7%) and higher than FastV (78.0%) and ADP (75.9%). Despite the updated evaluation environment and pruning configuration, the overall performance trend remains consistent with the original submission.


We then evaluated PROPPRUNE and other pruning baselines on RynnVLA-002 at 50% retention.

MethodSpatialObjectGoalLongAvg. SRFLOPs (T)LLM latency (ms)
Full-token98.698.897.692.496.97.498108.0
FastV53.80.054.022.832.74.30384.2
EfficientVLA61.643.642.458.451.54.29989.0
DivPrune12.00.00.21.23.44.09886.9
ADP77.00.844.846.242.25.78894.9
PROPPRUNE94.684.889.083.087.94.86391.5

PROPPRUNE achieved an average success rate of 87.9%, outperforming all evaluated pruning baselines, whose success rates ranged from 3.4% to 51.5%. Although PROPPRUNE showed a 9%p decrease relative to the full-token baseline, it preserved substantially higher task performance than the other pruning methods.

These results suggest that the core mechanisms of PROPPRUNE can be applied beyond OpenVLA-OFT.


To further examine the effectiveness of the calibration rule, we compared the original OpenVLA-OFT pruning configuration with the automatically calibrated configuration on RynnVLA-002.

ConfigurationAvg. SR (%)FLOPs
OpenVLA-OFT Configuration (P4/L5, θ=0.35)76.64.497
Calibrated RynnVLA-002 (P7/L8, θ=0.605)87.94.863

Directly transferring the original OpenVLA-OFT configuration to RynnVLA-002 resulted in 76.6% average success. In contrast, the calibrated configuration achieved 87.9%, an improvement of 11.3 percentage points. These results indicate that pruning configurations may not transfer directly across VLA architectures and that the proposed calibration rule can select a competitive configuration without evaluation-success-based tuning.

Response to Weakness 3 & Question 3 (3/3)

We also conducted additional comparisons with existing action-aware VLA pruning methods to clarify the methodological differences and empirical advantages of PROPPRUNE. ADP (ICLR 2026) combines text-driven visual-token selection with motion-based gating using end-effector trajectories. SpecPrune-VLA (ICML 2026) combines historical attention reuse, layer-wise token importance updates, and velocity-based control of pruning aggressiveness.

In contrast, PROPPRUNE directly incorporates proprioception-related attention into visual-token importance scoring. Specifically, we combine language-to-proprioception attention with language-to-visual attention to determine which visual tokens to retain. We then use inter-layer Jaccard similarity to select the pruning depth.

Thus, our main distinction lies in proprioception-guided visual-token ranking and Jaccard-based pruning-depth selection, rather than motion-based control of pruning aggressiveness or layer-wise token importance updates.


We evaluated these methods using their publicly released implementations. ADP produced results comparable to those reported in its original paper. We therefore compared ADP with PROPPRUNE under the same OpenVLA-OFT evaluation protocol, using 500 episodes per LIBERO suite.

MethodRetentionSpatialObjectGoalLongAvg. SRTFLOPs
ADP50%98.897.496.491.696.13.236
PROPPRUNE50%99.097.697.494.297.12.545
ADP25%98.097.294.682.293.02.844
PROPPRUNE25%98.696.096.893.296.21.842

At 50% retention, PROPPRUNE achieved a 1.0%p higher average success rate than ADP. At 25% retention, the improvement increased to 3.2%p. In particular, PROPPRUNE achieved 93.2% on LIBERO-Long at 25% retention, compared with 82.2% for ADP.

These results show that PROPPRUNE outperforms ADP at both retention budgets.


We also attempted to reproduce SpecPrune-VLA using its official public implementation with α=0.8\alpha=0.8. The reproduced success rates on LIBERO-Spatial, Object, Goal, and Long were 92.0%, 96.8%, 94.4%, and 41.4%, respectively, with an average success rate of 81.2%.

However, these results did not reproduce the performance reported in the original paper, particularly on LIBERO-Long. A related reproducibility issue was also reported in Issue #1 of the official GitHub repository.

Given the limited rebuttal period, we could not fully resolve this discrepancy with the original authors. Therefore, we focus our main quantitative comparison on ADP, for which the reported results were reproducible. We will also cite both ADP and SpecPrune-VLA in Section 2.3 of the revised manuscript and clarify their methodological differences from PROPPRUNE.

Reviewer_2DvsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a training-free visual-token pruning framework, PROPPRUNE , for vision language action models. PROPPRUNE incorporates proprioceptive state information into token importance scoring, weighting instruction-to-vision attention according to how strongly each instruction token attends to the robot's current state. More importantly, it adaptively selects where to prune: by measuring the Jaccard similarity between top-K visual-token sets from adjacent shallow layers, it prunes early when token rankings are stable and defers pruning when they are not. The method's gains are particularly notable on long-horizon tasks.

Critical strengths

S1 - action aware token selection: Incorporating proprioceptive information into visual-token ranking is a meaningful improvement. In VLA, visual relevance is state-dependent. The ablation results support the motivation, with proprioceptive guidance improving LIBERO-Long success by 1.8 and 1.2 percentage points at 50% and 25% retention, respectively.

S2 - adaptive pruning-layer selection is the key contribution: Rather than assuming that a fixed shallow layer always provides reliable importance scores, the paper uses the stability of adjacent-layer token rankings to determine where to prune. This is particularly relevant to closed-loop control, where prematurely removing task-critical visual information can cause errors that accumulate over subsequent actions. The results on LIBERO-Long are compelling.

S3 - good accuracy–efficiency trade-off: The results suggest that selecting the right tokens and selecting them at the right layer are complementary aspects of efficient VLA inference. At 50% retention, real-world success matches the full-token reference across 40 trials, with a 19.1% latency reduction. The training-free design also makes the method relatively easy to deploy.

Critical weaknesses

W1 - limited architectural generality: It remains unclear whether the method transfers to architectures with different multimodal fusion strategies, action experts, or token representations. This limits the strength of the broader claims about VLA pruning.

W2 - the proprioceptive contribution is useful but modest overall: Removing proprioceptive guidance reduces average success by only 0.50 and 0.25 percentage points at the two retention budgets. Although the improvement is larger on long-horizon tasks, the paper would benefit from stronger evidence that state-conditioned scoring is consistently valuable beyond this setting.

W3 - the adaptive gate relies on a heuristic stability measure: Adjacent-layer top-K similarity is a plausible indicator of ranking stability, but stable rankings do not necessarily imply that the selected tokens contain all information needed for accurate action prediction. Conversely, a changing ranking does not necessarily mean early pruning would be harmful. The threshold and probe/deferred layers also introduce hyperparameters, and the paper does not establish how robust these choices are across architectures and tasks.

W4 - real-world validation remains limited: The physical evaluation covers only two tabletop tasks, with 20 trials per task and setting, which is insufficient to establish robust gains on genuinely long-horizon, contact-rich tasks.

Questions for authors

Q1 - can the authors validate the adaptive layer-selection mechanism across additional VLA architectures, particularly models with different action-generation heads or multimodal fusion designs?

Q2 - how sensitive are the results to the Jaccard threshold, probe layer, and deferred pruning layer? Can these parameters be selected automatically without task-specific tuning?

Q3 - since proprioceptive guidance yields modest average gains but larger improvements on LIBERO-Long, can the authors characterize when state-conditioned scoring is most beneficial and whether it complements adaptive layer selection independently?

7 comments

Response to Weakness 1 & Question 1-2 (1/2)

To address these concerns, we extended our evaluation to RynnVLA-002 and introduced a calibration procedure to select pruning configurations for each VLA model using training observations, rather than tuning hyperparameters based on evaluation success rates.

Importantly, the core mechanisms of PROPPRUNE remain unchanged, including proprioception-guided visual-token importance scoring and Jaccard-based adaptive layer deferral. The calibration procedure only determines the probe layer, deferred pruning layer, and Jaccard similarity threshold for each model.


The calibration procedure consists of the following steps:

  1. For each evaluation suite, we randomly select 10 training episodes from the corresponding tasks and sample frames every 8 frames.
  2. We forward the sampled observations through the VLA model and compute visual-token importance at each Transformer layer using our proposed importance scoring method.
  3. For each pair of adjacent Transformer layers, we compute the Jaccard similarity between their selected top-K visual-token sets and average the similarities across the sampled observations.
  4. We select the layer with the largest increase in average inter-layer Jaccard similarity as the probe layer and the next layer as the deferred pruning layer. The Jaccard similarity threshold is set to the average similarity at the probe layer.

Using this procedure, we obtained the following configurations:

ModelRetentionProbe / Deferred LayerJaccard Threshold
OpenVLA-OFT50%P4/L50.523
OpenVLA-OFT25%P4/L50.445
RynnVLA-00250%P7/L80.605

While the calibration uses training observations from the corresponding tasks, it does not require task-specific hyperparameter search based on evaluation performance. The same calibration rule is applied across models and retention budgets.


We first re-evaluated PROPPRUNE on OpenVLA-OFT using the calibrated configurations, following the same evaluation protocol with 500 episodes per LIBERO suite. We also extended our baseline comparisons to include Action-aware Dynamic Pruning (ADP), an existing action-aware VLA pruning method.

MethodRetentionSpatialObjectGoalLongAvg. SRFLOPs (T)
Full-token100%99.298.297.295.697.64.013
ADP50%98.897.496.491.696.13.236
PROPPRUNE (calibrated)50%99.097.697.494.297.12.545
ADP25%98.097.294.682.293.02.844
PROPPRUNE (calibrated)25%98.696.096.893.296.21.842

With the calibrated configurations, PROPPRUNE achieved average success rates of 97.1% and 96.2% at 50% and 25% retention, respectively, compared with 97.6% for the full-token baseline.

Compared with ADP, PROPPRUNE achieved 1.0 and 3.2 %p higher average success rates at 50% and 25% retention, respectively, while requiring fewer FLOPs at both retention budgets.

Although these results differ slightly from the original submission due to the updated pruning configurations, the overall accuracy-efficiency trends remain consistent. These results indicate that the calibrated configurations maintain the effectiveness of PROPPRUNE without modifying its underlying proprioception-guided scoring or adaptive layer-deferral mechanisms.


To further examine sensitivity to pruning-layer configurations, we evaluated alternative probe and deferred pruning layers on OpenVLA-OFT at 50% retention. For each configuration, the Jaccard threshold was determined using the same calibration procedure.

ConfigurationθSpatialObjectGoalLongAvg. SRFLOPs (T)
P3/L40.73897.692.696.884.092.752.491
P4/L5 (calibrated)0.52399.097.697.494.297.052.545
P5/L60.60898.896.497.694.496.802.599
P4/L80.52398.897.697.893.496.902.626

The results show that the earlier P3/L4 configuration leads to a noticeable performance drop, particularly on LIBERO-Long. In contrast, P4/L5, P5/L6, and P4/L8 achieve comparable average success rates, ranging from 96.80% to 97.05%, with modest differences in computational cost.

These results suggest that several pruning configurations can achieve comparable performance, while overly early pruning may lead to performance degradation. The calibration procedure provides a systematic way to select a suitable configuration from these candidates without manually tuning pruning layers and thresholds based on evaluation success rates. This is particularly useful when applying PROPPRUNE to different VLA architectures, where suitable pruning configurations may differ across models.

Response to Weakness 1 & Question 1-2 (2/2)

We then extended the evaluation to RynnVLA-002 using the same calibration procedure and evaluation protocol, with 500 episodes per LIBERO suite. We compared PROPPRUNE against existing visual-token pruning baselines, including ADP, at 50% retention.

MethodSpatialObjectGoalLongAvg. SRFLOPs (T)LLM latency (ms)
Full-token98.698.897.692.496.97.498108.0
FastV53.80.054.022.832.74.30384.2
EfficientVLA61.643.642.458.451.54.29989.0
DivPrune12.00.00.21.23.44.09886.9
ADP77.00.844.846.242.25.78894.9
PROPPRUNE94.684.889.083.087.94.86391.5

PROPPRUNE achieved an average success rate of 87.9%, outperforming all evaluated pruning baselines, whose success rates ranged from 3.4% to 51.5%. In particular, PROPPRUNE substantially outperformed ADP (42.2%), another action-aware pruning method.

Although PROPPRUNE showed a 9%p decrease relative to the full-token baseline, it preserved substantially higher task performance than the other pruning methods while reducing FLOPs from 7.498T to 4.863T.

Importantly, the same Jaccard-based adaptive deferral mechanism was applied to both OpenVLA-OFT and RynnVLA-002, with only the probe layer, deferred layer, and threshold determined through calibration. These results therefore provide evidence that the adaptive layer-selection mechanism can be applied to another VLA backbone without modifying its underlying decision rule.


To further examine the importance of architecture-specific calibration, we compared the original OpenVLA-OFT pruning configuration with the automatically calibrated configuration on RynnVLA-002.

ConfigurationAvg. SR (%)FLOPs (T)
OpenVLA-OFT Configuration (P4/L5, θ=0.35)76.64.497
Calibrated RynnVLA-002 (P7/L8, θ=0.605)87.94.863

Directly transferring the original OpenVLA-OFT configuration to RynnVLA-002 resulted in 76.6% average success. In contrast, the calibrated configuration achieved 87.9%, an improvement of 11.3%p.

These results indicate that pruning configurations may not transfer directly across VLA architectures. However, the proposed calibration procedure can identify suitable pruning layers and thresholds for different models without tuning against evaluation success rates.


These experiments provide additional evidence that PROPPRUNE can be applied beyond OpenVLA-OFT while retaining its original proprioception-guided scoring and Jaccard-based adaptive layer-deferral mechanisms. The calibrated OpenVLA-OFT results preserve the accuracy-efficiency trends observed in the original submission, and the RynnVLA-002 results demonstrate the applicability of the same mechanisms to another VLA backbone.

The configuration sensitivity and transfer experiments further suggest that pruning-layer and threshold selection can affect performance, while the proposed calibration procedure offers a systematic way to select competitive configurations without tuning against evaluation success rates.

Response to Weakness 2 & Question 3

We thank the reviewer for highlighting the modest average improvements from proprioceptive guidance and for suggesting a more detailed analysis of when state-conditioned scoring is beneficial and how it interacts with adaptive layer selection.

To address these concerns, we re-evaluated the language-only ablation using the calibrated pruning configurations described in our response to W1/Q1/Q2.

MethodRetentionSpatialObjectGoalLongAvg. SR
PROPPRUNE50%99.097.697.494.297.1
Language-only50%99.498.695.493.296.7
PROPPRUNE25%98.696.096.893.296.2
Language-only25%98.695.896.291.495.5

The average improvements over language-only scoring are 0.4 and 0.65%p at 50% and 25% retention, respectively, showing a similar overall trend to the original submission.

We acknowledge that the average improvements from proprioceptive guidance are modest and that the improvements are not uniform across all LIBERO suites. As the reviewer noted, the benefits are more pronounced on LIBERO-Long, where PROPPRUNE improves success by 1.0 and 1.8%p at 50% and 25% retention, respectively. We also observe a 2.0%p improvement on LIBERO-Goal at 50% retention.

These results suggest that proprioceptive guidance is more beneficial in tasks involving sequential interactions, rather than consistently improving performance across all task suites. We will revise the manuscript to clarify this distinction and place greater emphasis on its observed benefits in long-horizon manipulation tasks, rather than presenting proprioceptive guidance as a source of uniform performance improvements.


To further examine how proprioceptive guidance affects visual-token selection, we conducted qualitative analyses using real LIBERO-Long demonstration observations.

We compared language-only scoring and PROPPRUNE using identical visual observations, language instructions, and robot states. Both methods retained 25% of visual tokens per camera, with pruned patches represented in white.

In the examined examples, incorporating proprioception-related attention changed the spatial distribution of retained visual tokens. Compared with language-only scoring, PROPPRUNE preserved more patches around task objects and robot interaction regions, including objects involved in different stages of the task.

These qualitative observations provide additional insight into why proprioception-guided scoring may be more beneficial on LIBERO-Long, where sequential interactions require maintaining information about multiple task-relevant regions.

We have added these visualizations to Appendix E (Figures 6 and 7) of the updated manuscript. We will also update the manuscript with the additional quantitative results and corresponding discussion as soon as possible.


To further investigate the relationship between proprioception-guided scoring and adaptive layer selection, we conducted an additional ablation study comparing the two scoring methods under adaptive pruning and fixed early pruning at L4.

Scoring50% Adaptive50% Fixed L425% Adaptive25% Fixed L4
PROPPRUNE97.0584.5096.1571.35
Language-only96.6589.0595.5076.45

The adaptive gate substantially improves performance for both scoring methods. At 50% retention, adaptive pruning improves success by 12.55%p with proprioception-guided scoring and 7.60%p with language-only scoring. At 25% retention, these improvements increase to 24.80 and 19.05%p, respectively.

Interestingly, the relative benefit of proprioceptive guidance depends on the pruning strategy. Under adaptive pruning, PROPPRUNE outperforms language-only scoring by 0.40 and 0.65%p at 50% and 25% retention, respectively. However, under fixed early pruning at L4, language-only scoring performs better by 4.55 and 5.10%p.

These results indicate that proprioceptive guidance is more effective when combined with adaptive layer selection. Moreover, the performance improvement from adaptive pruning is larger with proprioception-guided scoring than with language-only scoring at both retention budgets, suggesting an interaction between the two components.

Overall, proprioception-guided scoring determines which visual tokens to retain, while adaptive layer selection determines when pruning should occur based on ranking stability. Their combination enables PROPPRUNE to leverage state-conditioned token importance while reducing the risk of premature pruning.

Reviewer_WxwJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper propose a training free token prunning method for VLA. The evaluations on conducted on LIBERO、LIBERO-Plus and two real-world cases.

Critical strengths
  • The target problem is meaningful.
  • Good overall performance, and especially on the long sequence task.
Critical weaknesses
  • The proposed is largely an engineering heuristic and lacks a clear theoretical motivation.
  • Most of the gains come from layer selection, while proprioceptive weighting provides only modest additional benefits. Always pruning at layer 6 achieves the highest average success rate with only a small increase in FLOPs, calling into question the added value of the adaptive mechanism.
  • Combining the two heuristics offers practical value, but the overall novelty is limited.
  • The method lacks validation across different model families.
Questions for authors

See the weakness.

3 comments

Response to Weaknesses (1/3)

We thank the reviewer for the constructive feedback regarding the heuristic nature of PROPPRUNE, the relative contributions of its components, and its generalizability across VLA architectures.


To address these concerns, we extended our evaluation to RynnVLA-002 and introduced a calibration procedure to select pruning configurations for each VLA model using training observations, rather than tuning hyperparameters based on evaluation success rates. We also conducted additional experiments examining Jaccard-based adaptive layer selection and the contribution of proprioceptive guidance.

Importantly, the core mechanisms of PROPPRUNE remain unchanged, including proprioception-guided visual-token importance scoring and Jaccard-based adaptive layer deferral. The calibration procedure only determines the probe layer, deferred pruning layer, and Jaccard similarity threshold for each model.


The calibration procedure consists of the following steps:

  1. For each evaluation suite, we randomly select 10 training episodes from the corresponding tasks and sample frames every 8 frames.
  2. We forward the sampled observations through the VLA model and compute visual-token importance at each Transformer layer using our proposed importance scoring method.
  3. For each pair of adjacent Transformer layers, we compute the Jaccard similarity between their selected top-K visual-token sets and average the similarities across the sampled observations.
  4. We select the layer with the largest increase in average inter-layer Jaccard similarity as the probe layer and the next layer as the deferred pruning layer. The Jaccard similarity threshold is set to the average similarity at the probe layer.

Using this procedure, we obtained the following configurations:

ModelRetentionProbe / Deferred LayerJaccard Threshold
OpenVLA-OFT50%P4/L50.523
OpenVLA-OFT25%P4/L50.445
RynnVLA-00250%P7/L80.605

While calibration uses training observations from the corresponding tasks, it does not require task-specific hyperparameter search based on evaluation performance. The same calibration rule is applied across models and retention budgets.


We first re-evaluated PROPPRUNE on OpenVLA-OFT using the calibrated configurations, following the same evaluation protocol with 500 episodes per LIBERO suite.

MethodRetentionSpatialObjectGoalLongAvg. SRFLOPs (T)
Full-token100%99.298.297.295.697.64.013
PROPPRUNE (calibrated)50%99.097.697.494.297.12.545
PROPPRUNE (calibrated)25%98.696.096.893.296.21.842

With the calibrated configurations, PROPPRUNE achieved average success rates of 97.1% and 96.2% at 50% and 25% retention, respectively, compared with 97.6% for the full-token baseline.

Although these results differ slightly from the original submission due to the updated pruning configurations, the overall accuracy-efficiency trends remain consistent. These results indicate that the calibration procedure maintains the effectiveness of PROPPRUNE without modifying its underlying proprioception-guided scoring or adaptive layer-deferral mechanisms.


We then extended the evaluation to RynnVLA-002 using the same calibration procedure and evaluation protocol, with 500 episodes per LIBERO suite. We compared PROPPRUNE against existing visual-token pruning baselines, including ADP, at 50% retention.

MethodSpatialObjectGoalLongAvg. SRFLOPs (T)
Full-token98.698.897.692.496.97.498
FastV53.80.054.022.832.74.303
EfficientVLA61.643.642.458.451.54.299
DivPrune12.00.00.21.23.44.098
ADP77.00.844.846.242.25.788
PROPPRUNE94.684.889.083.087.94.863

PROPPRUNE achieved an average success rate of 87.9%, outperforming all evaluated pruning baselines, whose success rates ranged from 3.4% to 51.5%.

Although PROPPRUNE showed a 9.0%p decrease relative to the full-token baseline, it preserved substantially higher task performance than the other pruning methods while reducing FLOPs from 7.498T to 4.863T.


These results provide additional evidence that PROPPRUNE can be applied beyond OpenVLA-OFT while retaining its original proprioception-guided scoring and Jaccard-based adaptive layer-deferral mechanisms. Although suitable pruning configurations may differ across VLA architectures, the proposed calibration procedure provides a systematic way to select them without tuning against evaluation success rates.

Response to Weaknesses (2/3)

The motivation behind adaptive layer selection is that visual-token importance rankings may not be equally stable across inputs and Transformer layers. To further clarify the contribution of this mechanism, we conducted additional experiments examining whether Jaccard-based layer selection provides benefits beyond simply delaying pruning.

PROPPRUNE uses adjacent-layer Jaccard similarity as a lightweight indicator of ranking stability, pruning early when token rankings are stable and deferring pruning otherwise. Although this criterion is heuristic rather than theoretically guaranteed, it provides a training-free, input-dependent rule for selecting the pruning layer.

For each LIBERO suite and retention budget, random deferral was configured to match the actual late-pruning frequency observed with PROPPRUNE. Both methods used the same checkpoint, proprioception-guided token scoring, retention budget, and candidate pruning layers.

MethodRetentionSpatialObjectGoalLongAvg. SRFLOPs (T)
Full-token100%99.298.297.295.697.64.013
PROPPRUNE (Jaccard gate)50%99.097.697.494.297.12.545
Random deferral50%98.694.896.091.295.22.550
PROPPRUNE (Jaccard gate)25%98.696.096.893.296.21.842
Random deferral25%98.083.895.290.491.91.840

Under matched late-pruning frequencies and nearly identical computational costs, PROPPRUNE outperforms random deferral by 1.9 and 4.3%p at 50% and 25% retention, respectively.

The differences are particularly pronounced on LIBERO-Object and LIBERO-Long. At 25% retention, PROPPRUNE achieves 96.0% success on Object, compared with 83.8% for random deferral, a 12.2%p improvement.

These results suggest that the benefit of adaptive pruning cannot be explained solely by delaying pruning more frequently; selecting which inference steps should defer pruning is also important.


We also acknowledge that proprioceptive guidance provides modest average improvements over language-only scoring.

Using the calibrated configurations, we re-evaluated both scoring methods.

MethodRetentionSpatialObjectGoalLongAvg. SR
PROPPRUNE50%99.097.697.494.297.1
Language-only50%99.498.695.493.296.7
PROPPRUNE25%98.696.096.893.296.2
Language-only25%98.695.896.291.495.5

The average improvements over language-only scoring are 0.40 and 0.65%p at 50% and 25% retention, respectively. Although these improvements are modest and not uniform across LIBERO suites, they are more pronounced on LIBERO-Long, reaching 1.0 and 1.8%p.

To further examine the relationship between proprioceptive guidance and adaptive layer selection, we conducted an additional ablation comparing both scoring methods under adaptive pruning and fixed early pruning at L4.

Scoring50% Adaptive50% Fixed L425% Adaptive25% Fixed L4
PROPPRUNE97.0584.5096.1571.35
Language-only96.6589.0595.5076.45

Under fixed early pruning at L4, language-only scoring performs better than proprioception-guided scoring. However, under adaptive pruning, proprioception-guided scoring achieves higher success rates at both retention budgets.

The performance improvement from adaptive pruning is also larger with proprioception-guided scoring than with language-only scoring. At 50% retention, adaptive pruning improves success by 12.55%p with proprioception-guided scoring and 7.60 %p with language-only scoring. At 25% retention, these improvements increase to 24.80 and 19.05%p, respectively.

These results suggest that the two components interact in determining pruning effectiveness, rather than providing independent, uniform improvements. While proprioception-guided scoring determines which visual tokens to retain, adaptive layer selection determines when pruning should occur based on ranking stability.

Response to Weaknesses (3/3)

We would like to further clarify the methodological contributions of PROPPRUNE in relation to existing action-aware VLA pruning methods.

In VLA models, visual-token importance depends not only on language instructions but also on the robot's current state. Moreover, the reliability of token importance rankings can vary across inputs and Transformer layers. PROPPRUNE addresses these two aspects jointly by incorporating proprioception-related attention into visual-token importance scoring and using inter-layer Jaccard similarity to determine when pruning should occur.

Unlike existing action-aware methods such as ADP (ICLR 2026), which uses robot motion information to control pruning aggressiveness, PROPPRUNE uses proprioceptive information to determine which visual tokens to retain and ranking stability to select the pruning layer. Thus, its contribution lies in jointly addressing state-conditioned token selection and input-dependent pruning depth without additional training.

The additional experiments presented in our previous comments support the complementary roles of these mechanisms. Under matched late-pruning frequencies and computational costs, Jaccard-based deferral outperforms random deferral by 1.9 and 4.3 percentage points at 50% and 25% retention, respectively. Moreover, proprioceptive guidance achieves higher success rates than language-only scoring when combined with adaptive pruning, particularly on LIBERO-Long.

To further examine the practical effectiveness of these methodological differences, we compared PROPPRUNE with ADP using its publicly released implementation under the same OpenVLA-OFT evaluation protocol, with 500 episodes per LIBERO suite.

MethodRetentionAvg. SR (%)FLOPs (T)
Full-token100%97.64.013
ADP50%96.13.236
PROPPRUNE50%97.12.545
ADP25%93.02.844
PROPPRUNE25%96.21.842

PROPPRUNE outperforms ADP by 1.0 and 3.2 percentage points at 50% and 25% retention, respectively, while requiring fewer FLOPs at both retention budgets.

Together with the RynnVLA-002 results presented in our first comment, these findings provide evidence that PROPPRUNE's state-conditioned token selection and adaptive pruning-layer selection can achieve favorable accuracy-efficiency trade-offs across different VLA backbones.


Overall, our contribution is not simply the combination of two pruning heuristics, but a training-free framework that jointly considers which visual tokens are relevant to the robot's current state and when their importance rankings are sufficiently stable for pruning. The additional experiments provide empirical support for both decisions and demonstrate their practical effectiveness beyond the original OpenVLA-OFT setting.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions and Responses

Dear Reviewers and Area Chair,

Thank you very much for your insightful and constructive feedback. We have revised the manuscript to address the concerns raised by the reviewers and incorporated additional experiments and analyses to strengthen the empirical evaluation.

The main updates are as follows:

  • Introducing a calibration procedure for pruning-layer selection and extending the evaluation to another VLA architecture, RynnVLA-002.

  • Expanding baseline comparisons to include Action-aware Dynamic Pruning (ADP).

  • Adding matched random-deferral experiments and component ablations to further evaluate Jaccard-based adaptive pruning and its interaction with proprioceptive guidance.

  • Extending the real-world evaluation to a challenging multi-stage manipulation task under different background conditions.

We sincerely appreciate the reviewers' thoughtful comments and suggestions, which have been invaluable in improving the quality and clarity of our work.

Below, we provide a consolidated summary of the strengths and key concerns raised across multiple reviews, along with our corresponding responses. Detailed responses to individual comments and questions are provided in our reviewer-specific replies.


Strengths

1. Meaningful problem and strong empirical performance (Reviewers WxwJ, 2Dvs, and hV79)

  • Efficient VLA inference is an important problem, and the method demonstrates strong empirical performance, particularly on long-horizon tasks.

2. Training-free design and practical evaluation (Reviewers 2Dvs and hV79)

  • The training-free design offers practical benefits, supported by evaluations on standard benchmarks and real-world manipulation tasks.

Weaknesses and Responses

1. Architectural generality and justification of adaptive layer selection (Reviewers WxwJ, 2Dvs, and hV79)

  • Concerns were raised about generalization beyond OpenVLA-OFT, heuristic pruning configurations, and whether Jaccard-based deferral provides benefits beyond simply delaying pruning.

  • We introduced a calibration procedure that selects pruning layers and thresholds using training observations without tuning against evaluation success rates.

  • We extended the evaluation to RynnVLA-002, where PROPPRUNE achieves 87.9% average success rate at 50% retention, outperforming all evaluated pruning baselines.

  • Under matched late-pruning frequencies and nearly identical computational costs, PROPPRUNE outperforms random deferral by 1.9 and 4.3 percentage points at 50% and 25% retention, respectively, supporting the effectiveness of Jaccard-based adaptive layer selection.

2. Modest contribution of proprioceptive guidance and interaction between components (Reviewers WxwJ, 2Dvs, and hV79)

  • The average contribution of proprioceptive guidance was considered modest, motivating further analysis of its benefits and interaction with adaptive layer selection.

  • Additional ablations show that proprioceptive guidance is more effective when combined with adaptive layer selection, achieving higher success rates than language-only scoring, particularly on long-horizon tasks. This highlights the complementary benefits of the two components.

3. Limited real-world evaluation (Reviewers 2Dvs and hV79)

  • Concerns were raised about the limited number and complexity of real-world manipulation tasks.

  • We extended the evaluation to a multi-stage drawer manipulation task on the SO-101 robot under two background conditions, with 25 trials per condition for each method.

  • At 50% visual-token retention, PROPPRUNE maintains a success rate close to the full-token model (62.0% vs. 66.0% overall), while matching its 64.0% success rate under the additional-object background condition.


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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.