acceptodds
Under review as a conference paper at ICLR 2027

ASSIST-VLA: ACCELERATING VISION-LANGUAGE- ACTION INFERENCE VIA MULTI-LEVEL REDUN- DANCY REDUCTION

Abstract

Vision-language-action (VLA) models enable robots to translate visual observa- tions and language instructions into actions, but repeated multimodal inference in- curs substantial latency during closed-loop control. We introduce ASSIST-VLA, a training-free framework that accelerates VLA inference by reducing redundancy at three levels: control cycles, visual inputs, and model structure. At the control- cycle level, consistency across successive action plans guides the reuse of cached cognition representations, allowing selected steps to bypass visual encoding and language-backbone inference. At the visual-input level, instruction-conditioned semantic relevance and local spatiotemporal cues guide token pruning to retain task-relevant information. At the model-structure level, we draw on the J-space perspective, which analyzes internal representations through their influence on model outputs, and extend it to VLA action outputs. By combining Jacobian mappings with channel activation statistics, we assess the influence of MLP chan- nels on predicted actions and structurally prune low-influence channels. Together, these components reduce the frequency and cost of full multimodal inference. On four manipulation tasks in SimplerEnv with CogACT-Base, ASSIST-VLA reduces average end-to-end inference latency from 229.1 ms to 158.5 ms per control step, achieving a 1.45× speedup with an average task success rate of 67.8%, compared with 67.1% for the original model. These results show that reducing redundancy across multiple levels can accelerate closed-loop VLA inference while maintain- ing task performance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, inference acceleration, token pruning, structural pruning, robotic manipulation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_WzqWOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ASSIST-VLA, a framework for accelerating VLA inference without additional training. It combines cognition caching based on action consistency, visual token selection using semantic and temporal cues, and structural MLP pruning guided by action influence. Experiments on CogACT in SimplerEnv report a 1.45× speedup with average success close to the original policy.

Critical strengths
  1. The paper reports measured end-to-end acceleration. Latency decreases from 229.1 to 158.5 ms, including online module overhead, providing practical evidence beyond theoretical FLOPs reductions.

  2. The ablations reveal the performance and efficiency trade-offs. Removing individual modules quantifies their marginal effects, and the appendix provides detailed cache rules, token selection procedures, and pruning configurations.

Critical weaknesses
  1. The distinctive mechanisms are insufficiently isolated from existing acceleration techniques. EfficientVLA already combines multiple acceleration strategies. The proposed action influence score is not compared with simpler pruning criteria under the same budget, and action guided caching lacks a comparison with periodic reuse at a matched refresh rate.

  2. The evaluation is narrow and does not establish robust performance preservation. Results cover one backbone and four task categories without reported uncertainty. Average success hides a 5.7 point decline on Pick Coke Can, while removing cognition reuse improves average success from 67.8% to 72.7%. Since reuse bypasses current visual observations, robustness to unexpected changes remains untested.

Questions for authors
  1. What benefits distinguish the proposed framework from EfficientVLA beyond combining different implementations of caching, token selection, and model compression?

  2. At the same pruning budget, how does the proposed action influence score compare with weight magnitude, activation statistics, and simpler action sensitivity criteria?

  3. Does action consistency outperform periodic or random caching at the same refresh rate? How does it respond to unexpected object motion or failed grasps during reuse steps?

  4. Could the authors provide latency versus success curves and repeated-run confidence intervals, and validate the method on another VLA backbone?

4 comments
AuthorsOpenReview ↗

Response to Your Questions (Part 1/2): Method Differences and Pruning Comparisons

We thank you for recognizing the measured acceleration, ablations, and implementation details. Due to the character limit, we divide our response into two parts. This comment addresses Q1–Q2; Part 2 addresses Q3–Q4.

Q1. What distinguishes ASSIST-VLA from EfficientVLA?

ASSIST-VLA introduces action consistency, semantic–spatiotemporal similarity, and J-space action influence to determine when computation is needed, which visual information to retain, and which model channels to preserve.

ComponentEfficientVLAASSIST-VLA
CachingAction-head features during diffusionCognition across control steps, gated by action consistency
Visual selectionAttention and diversityInstruction relevance and local spatiotemporal similarity
CompressionTransformer layersAction-influence-ranked MLP channels, physically compacted
ExecutionReduces computation within modulesGates full multimodal refreshes and reduces their cost

Action consistency connects reuse decisions to control outputs. During continuous manipulation, observations may change while required actions remain stable. We check whole-plan consistency, short-horizon prefixes, shifted-plan alignment, and gripper commands to decide whether cognition needs recomputation. Each hit avoids visual encoding, projection, and the entire multimodal language backbone, while retaining action-head execution. This exploits continuity across control steps rather than repetition within diffusion generation.

Semantic–spatiotemporal similarity preserves static targets and dynamic interaction regions. Motion-only selection can miss stationary targets, while single-frame importance cannot exploit observation changes. We combine instruction relevance with local temporal correlations and neighborhood displacement responses, allowing the visual budget to cover targets, the end effector, and interaction regions. This shortens the language backbone’s input on refresh steps.

J-space evaluates channels through their influence on actions. We compose layer-to-cognition and cognition-to-action mappings and combine the resulting influence with activation information. Ranking therefore considers both channel usage and its effect on control outputs. Physically reconstructing narrower gate/up/down projections converts this analysis into dense-computation and storage savings.

Together, these mechanisms reduce both refresh frequency and per-refresh cost. In the reported CogACT adaptation comparison, ASSIST-VLA achieves 158.5 ms latency and 67.8% success, versus EfficientVLA’s 168.5 ms and 66.3%: 5.9% lower latency and 1.5 points higher success. Relative to CogACT’s 229.1 ms, our speedup is 1.45×.

Removing cognition reuse, visual selection, or MLP pruning increases latency to 188.7, 171.1, or 178.4 ms, respectively. These ablations connect the mechanisms to measured end-to-end savings.

Q2. How do pruning criteria compare at the same budget?

We added closed-loop evaluations with MLP pruning only at the same overall pruning ratio. Success rates (%) and unweighted four-task averages are:

CriterionPickCanDrawerDrawerAppleCarrotAverage
Unpruned93.6773.6141.6750.0064.74
Ours93.0069.4437.9650.0062.60
Activation93.6773.1525.9350.0060.69
Magnitude90.0074.0718.5254.1759.19
Random 10.009.260.000.002.31
Random 21.008.800.0012.505.57
J-space only 190.3357.9760.1950.0064.62
J-space only 2——58.33——

Activation and magnitude pruning largely preserve PickCan and Drawer performance, but reduce DrawerApple success to 25.93% and 18.52%, versus the unpruned baseline’s 41.67%. Their scores alone do not fully capture channel contributions across tasks.

The simple action-sensitivity controls use only our J-space action-influence score, without activation information. They achieve 60.19% and 58.33% on DrawerApple under two settings, both above baseline. However, the first configuration reduces PickCan and Drawer to 90.33% and 57.97%. J-space thus identifies useful action-related channels, while using it alone changes the task profile.

Our complete criterion combines action influence and activation information. Among controls evaluated on all four tasks, its per-task success profile is closest to the unpruned baseline. Its average exceeds activation and magnitude pruning by 1.91 and 3.41 points, respectively. These closed-loop measurements evaluate task performance independently of the channel-selection score.

AuthorsOpenReview ↗

Response to Your Questions (Part 2/2): Cognition Reuse and Evaluation Uncertainty

This is Part 2 of our response, addressing Q3–Q4. Part 1 discusses the differences from EfficientVLA and the pruning comparisons.

Q3. Does action consistency outperform periodic/random caching at matched refresh rates? How does it handle disturbances?

Action-consistency gating allocates reuse to segments where historical action plans remain stable. Periodic caching schedules reuse by timestep; random caching samples reuse steps. Our gate instead checks whole-plan, prefix, shifted-plan, first-action, and gripper consistency. Even at the same average refresh rate, these strategies can allocate computation to different steps.

Existing logs covering 300 PickCan, 216 Drawer, 108 DrawerApple, and 24 Carrot episodes show:

TaskCache hitsControl stepsHit rate
PickCan5,63024,00023.46%
Drawer5,50224,40822.54%
DrawerApple4,71021,60021.81%
Carrot5182,88017.99%
Total16,36072,88822.45%

Although at most one consecutive reuse step is permitted, actual hit rates are only 17.99%–23.46%; remaining steps perform full refreshes. Reuse frequency therefore depends on both consistency conditions and control trajectories.

After one reuse step, the next step must refresh, allowing new visual information to condition action generation. A failed consistency check exits the stable state; two consecutive qualifying evaluations are required to re-enter it. Reuse itself does not read new images, so unexpected object motion or failed grasps can first affect visual conditioning at the next refresh. This limits consecutive observation skipping to one step.

Across three seeds, complete ASSIST-VLA and the variant without reuse achieve episode-weighted success of 78.19% and 78.14%, and four-task averages of 69.88% and 68.86%. The original single-run decrease was not reproduced. Enabling reuse reduces measured latency from 188.7 to 158.5 ms.

Time constraints prevented matched-refresh-rate periodic/random comparisons and controlled disturbance experiments.

Q4. Confidence intervals, latency–success curves, and another backbone

We evaluated the baseline, complete ASSIST-VLA, and the variant without reuse across three seeds. Each seed includes 300/216/108/24 PickCan/Drawer/DrawerApple/Carrot episodes: 648 per seed and 1,944 per configuration.

Success rates (%) are:

Method/seedPickCanDrawerDrawerAppleCarrotOverall
Baseline/192.3373.6142.5945.8376.08
Baseline/295.0068.0644.4441.6775.62
Baseline/392.3373.1554.6333.3377.47
ASSIST/187.0079.6369.4450.0080.25
ASSIST/285.0077.3153.7054.1776.08
ASSIST/383.6774.0778.7045.8378.24
No reuse/187.6779.6362.0445.8379.17
No reuse/285.0077.7859.2654.1777.16
No reuse/383.3374.5479.6337.5078.09

Across-seed means and 95% Student’s t-confidence intervals (%) are:

MetricBaselineASSISTNo reuse
PickCan93.22 [89.40,97.05]85.22 [81.05,89.39]85.33 [79.90,90.76]
Drawer71.60 [63.95,79.26]77.01 [70.07,83.94]77.31 [70.91,83.72]
DrawerApple47.22 [31.12,63.32]67.28 [35.89,98.68]66.98 [39.53,94.42]
Carrot40.28 [24.47,56.09]50.00 [39.65,60.35]45.83 [25.13,66.53]
Average63.08 [61.36,64.81]69.88 [64.73,75.03]68.86 [68.46,69.27]
Overall76.39 [73.99,78.78]78.19 [73.01,83.37]78.14 [75.64,80.63]

Average weights tasks equally; Overall weights episodes. ASSIST exceeds baseline on DrawerApple in every seed, with pooled success of 67.28% versus 47.22%. PickCan remains lower at 85.22% versus 93.22%. These results show task-specific trade-offs and repeated-run uncertainty.

Latency–success curves require closed-loop evaluations and latency measurements across cache thresholds, token budgets, and pruning ratios. Time constraints prevented this systematic sweep.

Second-backbone validation requires adapting the reusable representation, action interface, and channel-influence mappings, followed by closed-loop evaluation. This validation has not yet been completed. We prioritized the direct pruning comparisons and repeated evaluations to examine action-influence scoring and cognition reuse.

Thank you for the additional experiments and clarifications. Your response addresses some of my concerns, but I still have three questions.

  1. How did you get 57.97% on Drawer for “J-space only 1” in Q2? Did you use the same 216 episodes as in Q4? If so, how many were successful?

  2. At the same pruning budget, J-space alone gets 64.62% average success, while the combined score gets 62.60%. What led you to choose the combined score?

  3. If you reuse the cache at regular intervals, match the hit rate for each task, and avoid reuse on consecutive steps, how do success rate and latency compare with your action-consistency gate?

Thank you for examining the experimental statistics further and raising these three specific questions. We rechecked the episode counts, completed the missing Drawer evaluations, and reviewed our existing action-consistency gating experiments.

Question 1: How was the 57.97% Drawer result for J-space only calculated? Did it use the same 216 episodes as Q4?

Like Q4, this experiment was planned for 216 Drawer episodes. Nine episodes did not finish because of GPU memory errors during model loading. Of the 207 completed episodes, 120 succeeded, giving 120/207 = 57.97%. We apologize for not stating the number of completed episodes, which made the denominator ambiguous.

We have now rerun the nine missing episodes, obtaining three successes and six failures. The complete result is 123/216 = 56.94%, and the four-task average for J-space only changes from 64.62% to 64.37%.

Pruning criterionPickCanDrawerDrawerAppleCarrotFour-task average
J-space only90.33%56.94%60.19%50.00%64.37%
Combined action influence and activation93.00%69.44%37.96%50.00%62.60%

We will correct these values in the revised manuscript and report both successful and evaluated episode counts.

Question 2: Why choose the combined score when J-space only has a higher four-task average?

The combined score considers both a channel’s influence on the final action and its activation on actual inputs. Our additional closed-loop experiments show that these factors correspond to different task outcomes. Activation-based and weight-magnitude pruning largely preserve performance on PickCan and Drawer, but achieve only 25.93% and 18.52% on DrawerApple, respectively, below the unpruned baseline’s 41.67%. The simpler action-sensitivity criterion, which uses only our J-space score, achieves 60.19% on DrawerApple but drops to 56.94% on Drawer, compared with 73.61% for the baseline.

The combined score achieves 93.00%, 69.44%, 37.96%, and 50.00% on PickCan, Drawer, DrawerApple, and Carrot, respectively. Among these pruning comparisons, its per-task success-rate profile is closest to the unpruned baseline. This motivates our choice: to compress the MLPs while preserving the original model’s performance across tasks as closely as possible.

We also acknowledge that J-space only has a higher four-task average: 64.37% versus 62.60%. It performs better on DrawerApple and is a competitive, simpler alternative. The combined score performs better on PickCan and Drawer and more closely follows the baseline’s per-task results. We will present this trade-off in the revised manuscript and narrow our claim about the combined score’s advantage accordingly.

With only the combined-score MLP pruning enabled, DrawerApple success is 37.96%. Complete ASSIST-VLA reaches 69.44%, 53.70%, and 78.70% across three random seeds, exceeding the corresponding unpruned baseline in each seed. The closed-loop results with visual selection and MLP pruning together also show why the pruning criterion should be evaluated within the complete inference pipeline.

Question 3: How does action-consistency gating compare with fixed-interval reuse at a matched hit rate, with consecutive reuse prohibited?

This comparison asks which control steps should reuse cached information under a similar computation budget. We reviewed our existing exploratory DrawerApple experiments with cognition reuse only, using 20 episodes per configuration:

ConfigurationSuccess rateCache hit rateMean latency
Full refresh at every step20/20 (100%)0%219.9 ms
Maximum refresh interval of 30/20 (0%)38.5%158.7 ms
Maximum refresh interval of 220/20 (100%)22.0%184.1 ms
Maximum interval of 3; require three consecutive stability checks0/20 (0%)24.0%182.1 ms
Maximum interval of 3; raise the stability threshold to 0.920/20 (0%)25.0%178.7 ms

The configurations with 22.0% and 24.0% hit rates have similar mean latencies (184.1 ms and 182.1 ms) but success rates of 100% and 0%, respectively. Lowering the average hit rate alone did not restore task performance; the timing of reuse and the length of consecutive observation skips also affect the result.

The 22.0% hit-rate configuration permits at most one reuse step before a forced refresh. It retained the baseline’s 20/20 successes while reducing mean latency from 219.9 ms to 184.1 ms, a reduction of approximately 16.3%. These results provide empirical motivation for action-consistency scheduling. Because the configurations also change gating parameters, they do not isolate the contribution of the consistency test itself. We will add the proposed direct comparison with fixed-interval reuse at a matched actual hit rate and with consecutive reuse prohibited, reporting per-task success, actual hit rate, and end-to-end latency.

Reviewer_ng24OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ASSIST-VLA, a training-free method for accelerating VLA inference. The method consists of three components, including cached representation reuse based on action consistency, visual token pruning using semantic and motion information, and MLP channel pruning based on estimated action influence. Experiments are conducted on four SimplerEnv manipulation tasks using CogACT-Base. The results show that the average inference latency decreases from 229.1 ms to 158.5 ms per step, achieving a 1.45× speedup. The average success rate remains comparable to the original model.

Critical strengths
  1. The paper addresses a practical issue in VLA deployment. It reports actual inference latency and task success rates rather than only theoretical FLOPs reduction. The reported 1.45× speedup is encouraging, especially since the average success rate remains close to that of the original CogACT-Base model.

  2. The paper provides detailed implementation settings in Appendix A, including cache thresholds, token selection, pruning ratios, and calibration parameters. The ablation experiments also show how each component affects inference speed and task performance. These details are helpful for understanding and reproducing the method.

Critical weaknesses
  1. The technical novelty of the individual components is limited. Cross-step caching, visual token pruning, and MLP pruning have already been widely studied. The proposed action-Jacobian pruning is potentially interesting, but its advantage over existing pruning criteria is not sufficiently demonstrated. The paper lacks comparisons with activation-based or magnitude-based pruning under the same model size. Moreover, Figure 3(c) evaluates retained influence using the same score used for channel selection. This does not directly show whether the pruned model can preserve action prediction accuracy.

  2. The three components seem to operate independently rather than form a unified method. There is no clear shared objective or information exchange between them. The framework mainly applies these techniques sequentially.

  3. The experiments are limited to four SimplerEnv tasks using CogACT-Base. Although these tasks cover different manipulation behaviors, the paper does not include evaluations on other common benchmarks such as LIBERO or RoboTwin 2.0. Real robot experiments are also missing. This is particularly important because the proposed caching mechanism skips visual processing based on previous action consistency. Its behavior under unexpected object movements, contact failures, or other disturbances remains unclear.

Questions for authors
  1. See weaknesses.

  2. What explains the 4.9-point average success-rate decrease when cognition reuse is enabled (Table 2)? Are per-task cache-hit rates or failure analyses available from existing evaluations?

  3. What is the one-time wall-clock/GPU cost of Jacobian calibration? Are matched-parameter comparisons against simpler structural-pruning scores available to justify the action-influence criterion?

2 comments
AuthorsOpenReview ↗

Response to Your Questions: Cognition Reuse, Calibration Cost, and Pruning Comparisons

We thank you for recognizing the measured latency, task success, and implementation details. We added three-seed evaluations, cache-hit statistics, pruning comparisons, and calibration timings.

Q1. Success-rate decrease and cache-hit statistics

Table 2 reports 72.72% success without cognition reuse versus 67.83% for the complete method, a 4.89-point difference. To examine reproducibility, we evaluated both across three seeds, each with 300 PickCan, 216 Drawer, 108 DrawerApple, and 24 Carrot episodes. Success rates (%) are:

Method/seedPickCanDrawerDrawerAppleCarrotOverall
ASSIST/187.0079.6369.4450.0080.25
ASSIST/285.0077.3153.7054.1776.08
ASSIST/383.6774.0778.7045.8378.24
ASSIST/pooled85.2277.0167.2850.0078.19
No reuse/187.6779.6362.0445.8379.17
No reuse/285.0077.7859.2654.1777.16
No reuse/383.3374.5479.6337.5078.09
No reuse/pooled85.3377.3166.9845.8378.14

Overall is episode-weighted success; pooled rows combine all seeds. With/without reuse, overall success is 78.19%/78.14%, and four-task averages are 69.88%/68.86%. The original decrease was not reproduced, so the 4.9-point difference is not established as a consistent reuse penalty.

Existing logs with 300/216/108/24 episodes provide:

TaskHits / control stepsHit rate
PickCan5,630 / 24,00023.46%
Drawer5,502 / 24,40822.54%
DrawerApple4,710 / 21,60021.81%
Carrot518 / 2,88017.99%
Total16,360 / 72,88822.45%

Two qualifying action-consistency evaluations enter stability; one nonqualifying refresh exits it. At most one consecutive reuse step is permitted, bounding observation skipping. Unexpected disturbances cannot be detected visually during reuse.

Reuse reduces measured latency from 188.7 to 158.5 ms, without a noticeable overall success decline in the repeated evaluations. Time constraints prevented controlled disturbance experiments and failure-level attribution.

Q2. Calibration cost and pruning comparisons

Calibration cost. Calibration uses 7,000 contexts, 131,072 directional trials per language layer across 32 layers, and 32,768 cognition-to-action trials: 4,194,304 language-layer trials plus 32,768 action-mapping trials.

StageRecorded wall-clock time
Context collection4 h 9 min 4 s
Layer-24 calibration, 131,072 trials7 h 33 min 26 s

Other per-layer records typically require 6.7–8.2 hours. These are stage-level timings, not total pipeline GPU-hours. Calibration incurs a substantial one-time cost, but its mappings and scores are reusable for pruning configurations of the same model. No Jacobian computation occurs during deployed control; preparation costs are amortized across deployments and inference.

Pruning comparisons. We added MLP-pruning-only evaluations at the same overall pruning ratio. Success rates (%) and unweighted four-task averages are:

CriterionPickCanDrawerDrawerAppleCarrotAverage
Unpruned93.6773.6141.6750.0064.74
Ours93.0069.4437.9650.0062.60
Activation93.6773.1525.9350.0060.69
Magnitude90.0074.0718.5254.1759.19
Random 10.009.260.000.002.31
Random 21.008.800.0012.505.57
J-space only 190.3357.9760.1950.0064.62
J-space only 2——58.33——

Activation and magnitude pruning largely preserve PickCan and Drawer performance, but reduce DrawerApple success to 25.93% and 18.52%, versus 41.67% for the baseline. Their scores alone do not reliably preserve capabilities across tasks.

Simple action sensitivity uses only our J-space action-influence score, without activation information. It achieves 90.33% on PickCan and 57.97% on Drawer, below baseline values of 93.67% and 73.61%. Yet DrawerApple reaches 60.19% and 58.33% under two settings, both above baseline. J-space scoring thus identifies useful channels for DrawerApple while changing performance elsewhere.

Our criterion combines action influence and activation information. Among controls evaluated on all four tasks, its per-task profile is closest to the unpruned baseline. Its average exceeds activation and magnitude pruning by 1.91 and 3.41 points. These closed-loop results evaluate task performance independently of the channel-selection score.

Area_Chair_FHzyOpenReview ↗

Reminder to respond to the authors’ rebuttal

With one day left in the rebuttal period, please read the authors’ response and leave a comment. It would help the authors to know which of your concerns they have addressed and which questions remain open, so they can make the most of the time left.

Thanks for your time. Your constructive feedback helps the authors improve their work and strengthens the ICLR community.

Best regards,

Area Chair

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

ASSIST-VLA is a training-free acceleration framework combining action-consistency-based cognition reuse, semantic-spatiotemporal visual-token selection, and Jacobian-guided structural MLP pruning. Cognition reuse skips visual encoding and language-backbone inference on eligible control steps, while token selection and smaller MLPs reduce computation on refresh steps.

On CogACT-Base and four SimplerEnv task groups, the method reduces mean observation-to-action inference latency from 229.1 ms to 158.5 ms on an RTX A6000, achieving a 1.45x speedup. Average success is 67.8%, compared with 67.1% for CogACT, with substantial variation in the per-task effects.

Critical strengths
  1. ASSIST-VLA achieves a concrete end-to-end latency reduction of 30.8% while maintaining a similar reported average success rate to CogACT (Table 1). Reconstructing smaller dense MLP layers also reduces the parameter count from 7.63B to 6.55B, providing practical computation and storage savings.

  2. The leave-one-out ablations expose the chosen efficiency-performance trade-off: removing cognition reuse raises average success from 67.8% to 72.7% while increasing latency from 158.5 ms to 188.7 ms (Section 4.3). Appendix A documents the implementation in considerable detail, including cache transitions, token selection, and pruning configuration.

Critical weaknesses
  1. The value of the proposed channel-ranking criterion and consistency gate is insufficiently established. Table 1 compares complete systems, while Table 2 switches modules on and off. Neither tests the channel ranking against random, weight-magnitude, or activation-based alternatives under matched pruning constraints, or the gate against periodic refresh and direct plan reuse at comparable refresh rates. Figure 1(c) provides a qualitative pruning example, and Figure 3(c) reports retained influence mass, but these do not establish the ranking's quantitative advantage in action preservation or task success. These omissions leave the contribution of the specific criteria unresolved.

  2. The evaluation does not establish generality and leaves the comparative results insufficiently supported. All experiments use CogACT-Base, whose separate cognition feature and action head are central to the reuse mechanism. Transfer to other VLA backbones is not evaluated. Structural pruning is calibrated on specific manipulation trajectories; Carrot is the only evaluation task group excluded from that calibration, providing limited evidence beyond the calibration task families. The success-rate average also combines a 5.7-percentage-point loss on PickCan with a 10.8-point gain on DrawerApple, without explicit evaluation episode counts or uncertainty estimates. Method-specific settings for the reimplemented baselines are insufficiently documented. These omissions limit assessment of both transferability and the latency-success advantage over existing methods.

Questions for authors
  1. Under matched pruning budgets and layer constraints, how does the proposed channel ranking compare with random, weight-magnitude, and activation-based rankings in task success or held-out action deviation?

  2. At comparable refresh rates, how does consistency gating compare with periodic refresh or direct shifted-plan reuse in latency and task success?

  3. What method-specific token budgets, compression settings, and cache schedules were used for the reimplemented baselines? Please also provide per-task evaluation episode counts and any available uncertainty estimates across repeated evaluation runs for the main success-rate comparisons.

2 comments
AuthorsOpenReview ↗

Response to Your Questions: Pruning Comparisons, Cognition Reuse, and Evaluation Details

We thank you for recognizing the measured acceleration and implementation details. We added pruning comparisons, cache-hit statistics, and three-seed evaluations.

Q1. Comparison of pruning criteria at the same budget

We added closed-loop evaluations with MLP pruning only, using the same overall pruning ratio. Success rates (%) are:

CriterionPickCanDrawerDrawerAppleCarrotAverage
Unpruned93.6773.6141.6750.0064.74
Ours93.0069.4437.9650.0062.60
Activation93.6773.1525.9350.0060.69
Magnitude90.0074.0718.5254.1759.19
Random 10.009.260.000.002.31
Random 21.008.800.0012.505.57
J-space only90.3357.9760.1950.0064.62

Ours improves the four-task average over activation and magnitude pruning by 1.91 and 3.41 percentage points. On DrawerApple, it achieves 37.96%, versus 25.93% and 18.52%.

The simple action-sensitivity control uses only our J-space action-influence score, without activation information. It has a higher average, but changes the task profile substantially: Drawer drops from 73.61% to 57.97%, while DrawerApple rises from 41.67% to 60.19%. Combining action influence with activation information more closely preserves the baseline’s per-task profile. These experiments measure closed-loop success directly, independently of the ranking score.

Q2. Consistency gating versus periodic refresh or shifted-plan reuse

Recorded cache hits are:

TaskHits / control stepsHit rate
PickCan5,630 / 24,00023.46%
Drawer5,502 / 24,40822.54%
DrawerApple4,710 / 21,60021.81%
Carrot518 / 2,88017.99%
Total16,360 / 72,88822.45%

These logs cover 300/216/108/24 episodes, respectively. Although at most one consecutive reuse step is permitted, the system does not simply alternate refresh and reuse. It checks whole-plan, prefix, shifted-plan, first-action, and gripper consistency. Two qualifying evaluations enter stability; one nonqualifying refresh exits it. Unlike periodic scheduling, reuse timing depends on action consistency.

Unlike direct shifted-plan execution, reuse still runs the action head with cached cognition and combines its output with the shifted historical plan at weights 0.25/0.75, followed by gripper discretization and temporal ensembling. No new visual observation is processed during reuse.

Across three seeds, complete ASSIST-VLA and the variant without reuse achieve episode-weighted success of 78.19%/78.14%, and four-task averages of 69.88%/68.86%. The original single-run decrease was not reproduced. Enabling reuse reduces measured latency from 188.7 to 158.5 ms.

Time constraints prevented matched-refresh-rate experiments against periodic refresh and direct plan reuse. The statistics explain operation and performance stability; they do not establish superiority in those matched comparisons.

Q3. Baseline settings, episode counts, and uncertainty

Confirmed local implementation settings are:

  • FastV: K=3, removal ratio 0.75; 256 visual tokens in the first three layers, then 64, selected by attention with original ordering preserved. No parameter compression or cognition reuse.
  • VLA-Cache standard path: patch cosine similarity ≥0.996; at most 150 static candidates, excluding the top-100 task-relevant positions using previous-frame layer-16 text-to-visual attention. Attention-entropy scheduling operates at zero-indexed layers 2/6/9/11. Cold starts have no reuse candidates. No parameter compression.
  • EfficientVLA local proxy: 56 visual tokens, 10 of 32 language layers skipped, action-head cache interval of five diffusion calls, reset before each action-sampling invocation.

The original evaluation uses 300/216/108/60 episodes for PickCan/Drawer/DrawerApple/Carrot. Carrot was expanded from the official 24 to 60 episodes. We will document this extension.

Repeated evaluations use the official 300/216/108/24 episodes: 648 per seed and 1,944 per configuration. Results (%) are:

Method/seedPickCanDrawerDrawerAppleCarrotOverall
Baseline/192.3373.6142.5945.8376.08
Baseline/295.0068.0644.4441.6775.62
Baseline/392.3373.1554.6333.3377.47
ASSIST/187.0079.6369.4450.0080.25
ASSIST/285.0077.3153.7054.1776.08
ASSIST/383.6774.0778.7045.8378.24
No reuse/187.6779.6362.0445.8379.17
No reuse/285.0077.7859.2654.1777.16
No reuse/383.3374.5479.6337.5078.09

Overall is episode-weighted success. Pooled ASSIST/baseline success is 78.19%/76.39%; DrawerApple is 67.28%/47.22%, with ASSIST higher in every seed. PickCan remains lower: 85.22%/93.22%. Overall success with/without reuse is 78.19%/78.14%, showing no noticeable decline in these repeated evaluations.

Area_Chair_FHzyOpenReview ↗

Reminder to respond to the authors’ rebuttal

With one day left in the rebuttal period, please read the authors’ response and leave a comment. It would help the authors to know which of your concerns they have addressed and which questions remain open, so they can make the most of the time left.

Thanks for your time. Your constructive feedback helps the authors improve their work and strengthens the ICLR community.

Best regards,

Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.