acceptodds
Desk rejected at ICLR 2027

VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

15 views

Abstract

Flow-matching vision-language-action (VLA) models generate an action chunk by iteratively denoising Gaussian noise, but typically execute only a short prefix before replanning with updated observations. Under this receding-horizon paradigm, the unexecuted remainder—the action tail—is discarded, and the next chunk is regenerated from fresh noise. We show that this discarded tail retains transferable denoising progress: action-tail initialization can match or outperform a native 10-step Gaussian baseline using fewer denoising steps. However, the optimal denoising budget varies substantially across samples, indicating that tail reuse should be adapted to the updated environment. Based on this observation, we introduce VLA-Rolling, a lightweight inference framework that reuses action tails through three stages: action-tail inheritance, per-action confidence estimation, and confidence-guided refinement through renoising and denoising budget allocation, while the underlying VLA remains frozen. Across , SmolVLA, X-VLA, and MolmoAct2 on LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin 2.0, VLA-Rolling reduces action-chunk inference cost in all 16 model–benchmark pairs, achieving up to a speedup, while improving success rate in 14 pairs by up to percentage points. These results show that action tails are not merely discarded predictions, but reusable intermediate states that enable more efficient and accurate VLA inference. These results demonstrate that the action tail can support the generation of better actions with fewer denoising steps.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Flow Matching, Efficient Inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_1gQKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VLA-Rolling, an inference framework that reuses the unexecuted action tail from the previous prediction round to accelerate flow-matching VLA models. A lightweight confidence estimator evaluates the inherited actions using updated observations and proprioceptive states. The predicted confidence controls the mixture between inherited actions and Gaussian noise and the number of subsequent denoising steps, while the underlying VLA remains frozen. Experiments across four VLA backbones and four simulation benchmarks report reduced action-chunk inference latency and success-rate improvements in most combinations. The paper also includes detailed ablations, analyses, and real-world manipulation experiments.

Critical strengths
  1. Broad experimental coverage. The method is evaluated across four VLA backbones and four simulation benchmarks, with additional real-world experiments. This breadth provides useful evidence that the proposed inference framework is applicable across different models and environments.
  2. Promising accuracy–efficiency trade-off. The reported results show reduced action-chunk inference latency across all evaluated model–benchmark pairs, while maintaining or improving task success in most settings. The diagnostic experiments also suggest that inherited action tails can achieve comparable action accuracy with fewer denoising steps.
Critical weaknesses
  1. Missing comparisons with closely related warm-start methods. The evaluation primarily compares VLA-Rolling against the unmodified backbones and its own ablations. However, closely related methods such as STEP (ICML 2026)(https://arxiv.org/abs/2602.08245) also construct informative warm-start initializations from historical actions and current observations to reduce denoising iterations. Without a direct comparison or a detailed explanation of the methodological differences, it is difficult to establish the incremental benefit of confidence-guided tail reuse. An appropriately adapted comparison under a shared backbone, evaluation protocol, and computational budget would strengthen the contribution.
  2. The confidence-guided refinement rules are insufficiently justified. The proposed confidence is derived from quantile-normalized log demonstration error, but it is subsequently used both as the action–noise interpolation coefficient and as the basis for allocating denoising steps. In a standard linear flow-matching path, Xτ=τA∗+(1−τ)ϵ,X_\tau = \tau A^* + (1-\tau)\epsilon, the coefficients are determined by the flow time τ\tau. The paper does not establish why its learned confidence should play an equivalent role. The specific confidence-to-noise and confidence-to-budget mappings therefore require stronger justification or targeted empirical validation.
Questions for authors
  1. Comparison under identical denoising budgets. One-step and few-step distillation methods demonstrate that diffusion/flow-based policies can operate with substantially fewer iterations, although they involve additional training. This raises the question of how much of the reported advantage depends on using a 10-step baseline. Could the authors compare Gaussian initialization and VLA-Rolling under identical small, fixed denoising budgets—for example, N∈1,2,4N \in 1,2,4—while holding other parameters constant?
  2. Justification of confidence-guided refinement. Could the authors justify using action confidence, supervised by quantile-normalized log action error, to control both confidence-guided renoising (Equation 9) and adaptive-step denoising (Equation 10)? How does this confidence relate to the flow time used by the frozen VLA? Please also clarify how per-action renoising coefficients are reconciled with a shared starting flow time, and provide theoretical justification or targeted ablations supporting these mappings.
  3. Relationship to STEP. STEP already explores warm-start action generation using historical actions and current observations, followed by a small number of denoising steps. Could the authors provide a detailed distinction between STEP and VLA-Rolling, particularly regarding initialization construction, auxiliary supervision, noise injection, and iteration allocation? A direct comparison, with appropriate adaptation to the same policy backbone, would help establish the benefit of the proposed confidence-guided mechanism.
5 comments
AuthorsOpenReview ↗

Thanks & Initial Response to Reviewer 1gQK (W1)

[W1 - Missing comparisons with closely related warm-start methods.]

We thank the reviewer for highlighting additional relevant work, which helps us better clarify the scope and positioning of our method. STEP directly predicts a warm-start initialization for flow matching, reducing the subsequent denoising process to two steps. This shares a similar warm-start motivation with VLA-Rolling.

To provide a direct comparison, we adapted the official STEP implementation to π0.5\pi_{0.5} and evaluated both methods on LIBERO, with 10 episodes per task. For a fair comparison, we limited VLA-Rolling to a maximum denoising budget of 2 steps, matching STEP's denoising budget. The results are summarized below.

Denoising BudgetMethodSuccess Rate ↑Avg. Denoising Steps ↓Total Inference Time (ms/chunk) ↓Denoising Time (ms) ↓
2 stepsπ0.5\pi_{0.5} + STEP90.75% (363/400)2.000119.2920.14
2 stepsπ0.5\pi_{0.5} + VLA-Rolling98.00% (392/400)1.694118.8516.21

For training, STEP was adapted to π0.5\pi_{0.5} and trained for 100K steps. Under the same GPU training-time budget, we retrained VLA-Rolling for approximately 48K steps. Despite requiring fewer training steps, VLA-Rolling achieves a higher success rate and lower inference latency under the evaluated setting.

AuthorsOpenReview ↗

Thanks & Initial Response to Reviewer 1gQK (W2)

[W2 - Insufficient justification for confidence-guided renoising and adapting denoising step rules.]

We thank the reviewer for raising this insightful question. We address the rationale behind from two perspectives.

1. Why use quantile-normalized log demonstration error?

Our confidence estimator is designed to predict the reuse reliability of inherited actions, rather than directly regress the optimal number of denoising steps. We adopt a linear mapping from predicted confidence to the refinement budget, motivated by an early experiment we have conducted during the method design stage.

Denoising Stepsπ0.5\pi_{0.5} + LIBEROπ0.5\pi_{0.5} + MetaWorldSmolVLA + LIBEROSmolVLA + MetaWorld
03.532.403.532.11
13.182.173.211.91
22.831.932.891.72
32.481.692.581.53
42.131.462.261.34
51.791.221.951.16
61.440.991.650.98
71.110.751.360.80
80.780.521.090.63
90.470.300.860.48
100.250.100.740.41

Across the evaluated models and benchmarks, action error decreases approximately linearly as the number of denoising steps increases, as shown in Figure 7. This observation provides empirical motivation for our linear confidence-to-budget mapping in Equantion (7) and Equation (10).

To further examine how closely the predicted refinement budget matches the optimal budget N∗N^* defined in Equation (2), we sampled 4,000 demonstrated actions from the LIBERO offline dataset and used demonstration replay to approximate the closed-loop execution context.

Step N12345678910Avg.
Ground Truth2.46%9.17%15.66%14.09%12.53%11.19%7.16%4.47%4.03%19.24%5.685
Predicted Step0.45%9.62%12.30%14.54%13.65%8.28%8.05%4.92%7.38%20.81%6.029

The distributions of predicted and oracle budgets exhibit closely matched average step counts and a strong positive spearman correlation (ρ=0.7858\rho=0.7858). These results provide additional empirical support for the effectiveness of our budget allocation mechanism. We will incorporate these analyses into the Appendix.

2. How are per-action renoising coefficients coordinated with a shared flow time?

In Equation (9), per-action confidence controls the proportion of inherited action information retained during initialization, rather than assigning an independent flow time to each action position. Equation (10) selects the refinement budget using the minimum confidence among inherited actions.

This choice is supported by our existing ablation in Appendix D.2, Table 6, where minimum-confidence-based allocation achieves lower cumulative trajectory errors than using the mean or maximum confidence on both LIBERO and MetaWorld.

Using the minimum confidence allows the refinement budget to account for the least reliable inherited action, reducing the risk of prematurely terminating denoising based on the average reliability of the chunk. This constitutes a conservative chunk-level scheduling strategy.

Thus, per-action confidence governs the initialization, while the minimum confidence determines the shared refinement schedule. This is a practical approximation design.

AuthorsOpenReview ↗

Thanks & Initial Response to Reviewer 1gQK (Q1, Q2, Q3)

[Q1 - Comparison under identical denoising budgets.]

To address the reviewer's concern, we conducted additional comparisons using π0.5\pi_{0.5} with smaller denoising budgets of 2 and 4 steps, under identical evaluation settings. We did not include the 1-step setting, as VLA-Rolling cannot further reduce the denoising budget in this case. For these experiments, we retained the VLA-Rolling estimator trained under the original 10-step configuration and adjusted its maximum denoising budget accordingly.

Denoising BudgetMethodSuccess Rate ↑Avg. Denoising Steps ↓Inference Time (ms/chunk) ↓
2 stepsGaussian Initialization97.75% (391/400)2.000121.08
2 stepsVLA-Rolling98.00% (392/400)1.694118.85
4 stepsGaussian Initialization96.75% (387/400)4.000140.43
4 stepsVLA-Rolling97.50% (390/400)3.078132.17

The results show that VLA-Rolling maintains success-rate while further reducing denoising steps and inference latency, even under substantially smaller denoising budgets. This suggests that our method's benefits are not limited to the original 10-step baseline, as it continues to reduce redundant computation without sacrificing task performance.

We appreciate the reviewer's suggestion and would be happy to evaluate our method on additional one-step or few-step distilled baselines of interest.

[Q2 - Justification of confidence-guided refinement.]

As detailed in our response to W2, we clarify that confidence serves as an empirical proxy for denoising progress rather than an exact equivalent of flow time. We provide additional evidence supporting the confidence-to-budget mapping, including Spearman correlation coefficients. We also explain that the confidence level for each action renoising during initialization, while the minimum confidence level determines the shared optimization plan, a point supported by the ablation results in Table 6 of Appendix D.2.

[Q3 - Methodological differences between STEP and VLA-Rolling.]

We thank the reviewer for encouraging a clearer comparison with STEP, which have already been included in our related work. Their mechanisms differ in four aspects:

  1. Initialization Construction: STEP learns to predict a new action sequence from historical actions and current observations, whereas VLA-Rolling directly reuses the unexecuted action tail from the previous inference round.
  2. Auxiliary Supervision: STEP trains its predictor using action prediction error, while VLA-Rolling learns action-wise reuse confidence from quantile-normalized log demonstration error.
  3. Noise Injection: STEP employs velocity-aware perturbation to address execution stagnation, whereas VLA-Rolling uses confidence-guided renoising to adjust how much historical action information is retained at each position.
  4. Iteration Allocation: STEP uses a predefined few-step refinement budget, while VLA-Rolling adaptively determines the remaining denoising steps based on the predicted confidence.

For direct empirical evidence, please refer to our response to W1, where we adapt STEP to the same π0.5\pi_{0.5} backbone and compare both methods under a two-step denoising budget.

Reviewer_FfUBOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a technic, VLA-Rolling. This method reuses the unexecuted tail of a previous action chunk to initialize subsequent action generation. A learned confidence estimator controls per-action renoising and the chunk’s denoising budget while keeping the underlying VLA frozen. Experiments across four models and four benchmarks report reduced inference latency in all 16 combinations and improved success in 14. The speedup with up to 1.48x.

Critical strengths
  1. The method addresses a practical source of redundant computation. Its latency measurements include the additional confidence estimator, making the reported inference savings meaningful and practical.

  2. Comprehensive results across multiple backbones and benchmarks provides encouraging evidence of effectiveness of the proposed method. Component ablations also shows the necessity of different design choices of the paper.

Critical weaknesses
  1. The incremental contribution over existing action-reuse methods is insufficiently established. Real Time Chunking (RTC) already uses overlapping actions from the previous chunk to guide subsequent generation while incorporating new observations, and has been used in many existing work. This paper provides no direct RTC comparison. Moreover, the ablations lack fixed-budget warm-start baselines combining fixed renoising with fewer denoising steps. Without matched latency–success comparisons, the benefit of the learned adaptive mechanism remains unclear.

  2. The confidence estimator’s suitability for allocating computation is not adequately validated. Its supervision measures distance to demonstrated actions, rather than the refinement budget needed under the current observation. These quantities need not correspond reliably, particularly when multiple actions are valid. Moreover, several success gains are small and lack uncertainty across training seeds; the real-world improvement is only from 14/20 to 15/20 successes.

Questions for authors
  1. How does VLA-Rolling compare with RTC and fixed-budget warm-start methods under matched backbones, hardware, execution schedules, and inference delays?

  2. What is the quantitative relationship between predicted confidence and the optimal refinement budget, particularly on states encountered during closed-loop execution?

6 comments
AuthorsOpenReview ↗

Thanks & Initial Response to Reviewer FfUB (W1)

[W1 - Insufficient distinction from RTC and lack of fixed-budget warm-start comparisons.]

We thank the reviewer for highlighting RTC and the importance of matched-backbone comparisons. RTC is also discussed in our Related Work. Our key distinction is that RTC primarily addresses temporal consistency under asynchronous execution, whereas VLA-Rolling focuses on reducing redundant action generation computation through action-tail reuse and adaptive refinement. While both methods reuse information from previous action chunks, they differ in their primary objectives and how historical actions are incorporated into inference. We clarify these distinctions and provide direct empirical comparisons below.

1. Distinction from RTC.

RTC primarily addresses temporal continuity and robustness to inference delays during asynchronous execution. It conditions newly generated actions on previously committed actions through inference-time guidance, which introduces additional gradient computations.

VLA-Rolling focuses on reducing redundant action generation computation. Rather than using previous actions only as consistency constraints, we reuse the unexecuted action tail as an informative initialization that preserves previous denoising progress. A lightweight confidence estimator then evaluates its reliability under updated observations to guide renoising and adaptive-step allocation. Thus, our approach aims to improve both action quality and inference efficiency without modifying the underlying VLA parameters.

2. VLA-Rolling achieves higher success rates and lower total inference latency than RTC and fixed-step inference.

Following the reviewer's suggestion, we conducted additional experiments using the same π0.5\pi_{0.5} backbone and evaluation device (NVIDIA RTX 4090 GPUs) on LIBERO and MetaWorld. We compare inference-time RTC, VLA-Rolling, and a fixed-step baseline whose task-wise denoising budget is set to the ceiling of VLA-Rolling's average predicted steps. We report success rate, average denoising steps, and per-chunk inference latency, including both non-flow and flow computation.

BenchmarkMethodSR (%) ↑Avg. Steps ↓Total Inference Time (ms) ↓
LIBEROVLA-Rolling97.006.291159.43
LIBEROFixed Steps95.256.980166.13
LIBERORTC94.7510.000242.21
MetaWorldVLA-Rolling67.806.025289.42
MetaWorldFixed Steps62.406.772297.43
MetaWorldRTC57.8010.000371.05

For a fair comparison, we follow the execution settings used in our main experiments (Section 4.1), with Chunk Horizon = 50, Execution Horizon = 10, and Inference Delay = 10. For RTC, we set Max Guidance Weight = 10.0. 10 episodes were tested for each task.

AuthorsOpenReview ↗

Thanks & Initial Response to Reviewer FfUB (W2)

[W2 - Insufficient validation of confidence-based computation allocation and uncertainty in performance gains.]

We thank the reviewer for highlighting directions in which we can further validate our method. We provide additional empirical evidence for confidence-guided denoising budget allocation and clarify the intended purpose of our real-world experiments.

1. Supporting evidence for confidence-guided denoising budget allocation.

Our confidence estimator is designed to predict the reuse reliability of inherited actions, rather than directly regress the optimal number of denoising steps. We adopt a linear mapping from predicted confidence to the refinement budget, motivated by an early experiment we have conducted during the method design stage.

Denoising Stepsπ0.5\pi_{0.5} + LIBEROπ0.5\pi_{0.5} + MetaWorldSmolVLA + LIBEROSmolVLA + MetaWorld
03.532.403.532.11
13.182.173.211.91
22.831.932.891.72
32.481.692.581.53
42.131.462.261.34
51.791.221.951.16
61.440.991.650.98
71.110.751.360.80
80.780.521.090.63
90.470.300.860.48
100.250.100.740.41

Across the evaluated models and benchmarks, action error decreases approximately linearly as the number of denoising steps increases, as shown in Figure 7. This observation provides empirical motivation for our linear confidence-to-budget mapping in Equantion (7) and Equation (10).

To further examine how closely the predicted refinement budget matches the optimal budget N∗N^* defined in Equation (2), we sampled 4,000 demonstrated actions from the LIBERO offline dataset and used demonstration replay to approximate the closed-loop execution context.

Step N12345678910Avg.
Ground Truth2.46%9.17%15.66%14.09%12.53%11.19%7.16%4.47%4.03%19.24%5.685
Predicted Step0.45%9.62%12.30%14.54%13.65%8.28%8.05%4.92%7.38%20.81%6.029

The distributions of predicted and oracle budgets exhibit closely matched average step counts and a strong positive spearman correlation (ρ=0.7858\rho=0.7858). These results provide additional empirical support for the effectiveness of our budget allocation mechanism. We will incorporate these analyses into the Appendix.

2. Clarification of the purpose of our real-world experiments.

We understand the reviewer's concern regarding the modest success-rate gains and their statistical uncertainty. However, our primary research objective is to reduce inference cost without sacrificing action accuracy, rather than to maximize success-rate improvements.

We are happy to provide further explanations below. Our real-world experiments are primarily intended to validate the confidence estimator's sensitivity to environmental changes during execution. As illustrated in Figure 6, when the target changes unexpectedly, the predicted confidence decreases and the allocated refinement budget increases. This demonstrates that, under the tested distribution shift, VLA-Rolling does not blindly reduce denoising steps or over-rely on potentially outdated actions.

AuthorsOpenReview ↗

Thanks & Initial Response to Reviewer FfUB (Q1, Q2)

[Q1 - Comparison with RTC and fixed-budget warm-start methods under matched settings.]

VLA-Rolling and RTC differ in their objectives and inference mechanisms. RTC primarily addresses temporal consistency under asynchronous execution through inference-time guidance on previously committed actions, whereas VLA-Rolling reuses unexecuted action tails as informative initializations and employs confidence-guided re-noising and adaptive denoising to reduce redundant action generation computation.

Compared with RTC and fixed-step baseline, VLA-Rolling reduces total inference latency while improves success rates by 2.25 and 10.00 percentage points, respectively. These results demonstrate that reusing historical action tails with confidence-guided adaptive refinement provides a inference efficiency–task success trade-off . We provide the full latency breakdown in our response to W1.

[Q2 - Quantitative relationship between predicted confidence and the optimal refinement budget.]

As detailed in our response to W2, we provide empirical evidence supporting the use of a linear mapping between predicted confidence and the optimal refinement budget, with reasonably accurate budget predictions. We appreciate this valuable question, as it helps us better clarify the rationale behind our method. We will include this analysis in the Appendix.

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

This method aims to accelerate inference without losing accuracy by reusing the unexecuted end of the previous chunk (action tail), which a flow-matching VLA discards during replanning, as the denoising starting point for the next chunk (the research question on p.1). A lightweight estimator predicts confidence at each tail position to determine how much noise to mix back in (renoising) and the number of denoising steps. The paper reports lower inference time per chunk and higher success rates across 4 models × 4 benchmarks.

Critical strengths
  • This is an inference method that can be attached to multiple models without modifying the underlying VLA. It freezes the VLA and separately trains and attaches only an estimator with approximately 30M parameters (Appendix B.1), and across π0.5, SmolVLA, X-VLA, and MolmoAct2 with 4 benchmarks, all 16 combinations show a reduction in inference time per chunk of 5–33% (Table 1).
Critical weaknesses
  1. The problem addressed is narrow inference acceleration, but the contribution is claimed more broadly, and even that problem is of limited importance. The authors themselves limit the problem to "generate new actions faster without sacrificing action accuracy" (p.1), but the abstract presents increased success rates (up to +15.08pp) as a contribution, and the same gain is obtained with tail initialization alone without reducing steps (MetaWorld +9.3pp, Table 2). The remaining contribution, a reduction of approximately 38ms per chunk (π0.5 LIBERO, Table 1), could largely be hidden by the chunk currently being executed during asynchronous execution on a real robot. In the asynchronous real-robot experiment, the difference is only 1 of 20 trials (Table 4), so the practical benefit relative to the cost of training an estimator for each VLA is not apparent.
  2. It has not been verified whether the acceleration comes from the tail and estimator. There is no baseline that runs the same number of steps from Gaussian noise (Fig. 3 compares only with 10 steps), so the results cannot be distinguished from the possibility that the default 10 steps were already more than necessary. Falcon (arXiv:2503.00339) and RTI-DP (arXiv:2508.05396) have already proposed tail reuse without training, so the distinctive contribution is the learned estimator, yet there is no evidence that it outperforms fixed 50% renoising or a fixed configuration with the same average number of steps (Table 2).
Questions for authors
  1. Compared with Gaussian initialization using the same average number of steps, how much does tail reuse improve success rate and inference time?
  2. Why is increased success rate claimed as a contribution when the research question is acceleration without losing accuracy (p.1)?
11 comments
AuthorsOpenReview ↗

Thanks & Initial Response to Reviewer ipFs (W1)

[W1 - The practical significance of inference acceleration, the claimed success-rate improvements, and the additional training cost of the estimator, particularly under asynchronous real-world execution.]

We sincerely apologize that our original presentation may have caused confusion regarding the practical objectives and contributions of VLA-Rolling. Our intended message is that VLA-Rolling improves the overall inference efficiency–task success trade-off: it reduces inference latency while maintaining or improving task success rates across most evaluated settings. We have published the presentation in our introduction, and added additional experiments as supporting evidence for improved action quality

1. VLA-Rolling improves action quality

(1) Second-order action differences. Reusing the unexecuted action tail can preserve motion continuity across replanning rounds, reducing abrupt changes between consecutive actions. We measure the second-order action differences (lower values indicate smoother motion):

MethodLIBEROMetaWorld
Baseline (π0.5\pi_{0.5})0.1011570.830971
VLA-Rolling0.0956760.709536
Relative Reduction5.42%14.61%

(2) Accumulated trajectory action error. We also evaluated predictions along demonstration trajectories and accumulated their errors against the corresponding demonstrated actions:

MethodLIBEROMetaWorld
Baseline (π0.5\pi_{0.5})36.133 ± 15.27812.892 ± 16.564
VLA-Rolling33.450 ± 14.8109.5394 ± 15.704
Relative Error Reduction7.43%26.01%

These results demonstrate smoother trajectories and lower accumulated action errors, providing further insight into the observed success-rate improvements. Together with reduced inference latency, they support the Pareto improvements reported in Table 1.

2. VLA-Rolling maintains effectiveness in larger real-world experiments

To strengthen our real-world evidence, we expanded the evaluation from 20 to 100 trials per method under the same task setting. The updated results are:

MethodSuccess RateAverage Inference Latency
Baseline (π0.5\pi_{0.5})71% (71/100)228.13 ms
VLA-Rolling77% (77/100)173.29 ms

The expanded evaluation confirms that the inference-latency reduction persists over repeated real-world trials, without an observed decrease in task success rate.

We view VLA-Rolling as an improved inference alternative for existing VLA models, rather than a competing acceleration method. The original VLA architecture and pretrained parameters remain unchanged, while allowing the it to adapt its inference computation to the current execution context, avoiding unnecessary denoising while retaining sufficient refinement when needed.

AuthorsOpenReview ↗

Thanks & Initial Response to Reviewer ipFs (W2)

[W2 - Insufficient evidence for the benefits of action-tail initialization and learned confidence-guided refinement over simpler fixed strategies.]

We sincerely thank the reviewer for highlighting Falcon and RTI-DP , which demonstrate the effectiveness of training-free historical action and denoising-state reuse. These works provide valuable references for positioning our method and motivate us to further clarify the distinctive contribution of learned confidence-guided refinement. Our key finding is that VLA-Rolling provides benefits beyond simply exploiting redundant computation in the original 10-step baseline: it improves action quality under comparable denoising budgets and further reduces computation through adaptive refinement. We provide two complementary experiments to validate these benefits.

1. VLA-Rolling maintains its advantages under matched low-step budgets, rather than merely exploiting redundancy in the 10-step baseline.

To directly address the missing Gaussian-initialization comparison, we evaluate VLA-Rolling against Gaussian initialization under identical maximum denoising budgets of 2 and 4 steps on LIBERO, using the same π0.5\pi_{0.5} backbone and evaluation settings. Importantly, we retain the confidence estimator trained under the original 10-step configuration without additional retraining.

Maximum BudgetMethodSuccess Rate ↑Avg. Steps ↓Inference Time (ms/chunk) ↓
2 StepsGaussian Initialization97.75% (391/400)2.000121.08
2 StepsVLA-Rolling98.00% (392/400)1.694118.85
4 StepsGaussian Initialization96.75% (387/400)4.000140.43
4 StepsVLA-Rolling97.50% (390/400)3.078132.17

Under both budgets, VLA-Rolling achieves comparable or higher success rates while requiring fewer average denoising steps and lower inference latency. These results demonstrate that our method remains effective even when the maximum denoising budget is substantially reduced.

2. Learned confidence-guided re-noising improves action quality beyond fixed reuse strategies at comparable computational costs.

To isolate the contribution of the learned confidence estimator, we conduct a 2×2 ablation study with four configurations: M1 (fixed re-noising + 10 steps), M2 (confidence-guided re-noising + 10 steps), M3 (fixed re-noising + adaptive steps), and M4 (confidence-guided re-noising + adaptive steps, i.e., VLA-Rolling).

MethodLIBERO Mean Error ↓Std.Avg. Steps ↓MetaWorld Mean Error ↓Std.Avg. Steps ↓
M135.145118.660310.00012.393624.813710.000
M234.383615.581810.0009.755215.716510.000
M334.768918.19396.62012.346324.73437.137
M4 (Ours)33.449714.81026.6119.539415.70417.127

M2 achieves lower trajectory errors on both benchmarks without additional denoising computation. M3 vs. M4 further isolates this benefit under adaptive computation:** despite requiring nearly identical average denoising steps, M4 achieves lower trajectory errors on both benchmarks. These comparisons demonstrate that the improvement is attributable to selective preservation and refinement of inherited actions

We have incorporated the corresponding ablation and matched-budget comparisons into Section 4.2 and the Appendix. We have also added Falcon and RTI-DP to the revised Related Work section, discussing their training-free action-reuse paradigms and clarifying how VLA-Rolling differs.

Update 11.17 (Retest few step baseline according to unified standards)

Following the reviewer's concerns regarding latency consistency, we re-evaluated the 2-step and 4-step baselines using the unified RTX 4090 hardware and inference timing protocol described in our response to C3.

BudgetMethodSR (%) ↑Avg. Steps ↓Total (ms/chunk) ↓
2 StepsGaussian Initialization97.752.00077.34
2 StepsVLA-Rolling98.001.69474.68
4 StepsGaussian Initialization96.754.00096.86
4 StepsVLA-Rolling97.503.07887.94
AuthorsOpenReview ↗

Thanks & Initial Response to Reviewer ipFs (Q1, Q2)

[Q1 - Compared with Gaussian initialization using the same average number of steps, how much does tail reuse improve success rate and inference time?]

We thank the reviewer for suggesting this comparison. To further evaluate the benefit of adaptive denoising, we conducted an additional fixed-step experiment using π0.5\pi_{0.5}. Specifically, we first ran 10 episodes per task to record the average denoising steps assigned by VLA-Rolling. We then fixed the denoising budget under Gaussian initialization for each task to the ceiling of its average, consistent with the rounding rule in Equation (10), and evaluated both methods using the same random seeds.

BenchmarkMethodSuccess Rate ↑Avg. Denoising Steps ↓Total Inference Time (ms/chunk) ↓
LIBEROVLA-Rolling97.00%6.291159.43
LIBEROFixed Steps95.25%6.980166.13
MetaWorldVLA-Rolling67.80%6.025289.42
MetaWorldFixed Steps62.40%6.772297.43

Compared with the fixed-step baseline, VLA-Rolling improves success rates while reducing inference latency respectively. These results indicate that adaptively allocating denoising steps according to the current action context provides a better accuracy–efficiency trade-off than using a task-wise fixed budget.

[Q2 - Why is increased success rate claimed as a contribution when the research question is acceleration without losing accuracy?]

Our intended message is that VLA-Rolling improves the overall inference efficiency–task success trade-off: it reduces inference latency while maintaining or improving task success rates across most evaluated settings. Rather than treating inference acceleration and task success as competing objectives, our method adaptively reuses reliable inherited actions and allocates additional refinement when necessary.

We therefore include the success-rate gains as an additional empirical contribution, while maintaining inference acceleration with preserved action accuracy as our central claim. We also provide further supporting analyses in our response to W1.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

For clarity, we refer to Reviewers ipFs, FfUB, and 1gQK as R1, R2, and R3, respectively.

We sincerely thank all reviewers for their thoughtful feedback and constructive suggestions. We are encouraged by their recognition of several key strengths of VLA-Rolling.

In particular, we sincerely appreciate the recognition that VLA-Rolling can be applied to multiple VLA backbones without modifying the underlying models, while consistently reducing inference latency across all 16 evaluated model–benchmark combinations (R1). We are grateful for the acknowledgment that our method addresses a practical source of redundant computation, with inference measurements already accounting for the additional estimator overhead, making the reported savings meaningful in practice (R2). Furthermore, we appreciate the reviewers' positive assessments of our extensive evaluation across four VLA backbones and four benchmarks (R1, R2, R3), especially the component ablations demonstrating the importance of different design choices (R2) and the additional real-world manipulation experiments supporting applicability across different models and environments (R3). Finally, we are encouraged by the recognition of the promising accuracy–efficiency trade-off, particularly the observation that inherited action tails can achieve comparable action accuracy with fewer denoising steps (R3).

These encouraging assessments, together with the reviewers' concerns about comparative baselines, adaptive refinement, and real-world significance, have helped us identify where stronger evidence and clearer explanations are needed. We have addressed these points individually and summarize the main contributions and rebuttal updates below.

Core Contributions of Our Work

  1. Uncovering reusable denoising progress in action tails. We show that unexecuted action tails retain useful information from previous inference rounds, enabling comparable action accuracy with fewer denoising steps. We further reveal that the optimal refinement budget varies considerably across samples, motivating adaptive rather than uniform computation allocation.

  2. Confidence-guided adaptive action reuse. VLA-Rolling introduces an action-wise confidence estimator that jointly controls how much inherited action information to preserve and how many denoising steps to allocate. The original VLA remains frozen, enabling adaptive inference while maintaining its generalization and scalability.

  3. Generalizable inference acceleration. Across four VLA models and four benchmarks, VLA-Rolling reduces action-chunk inference latency in all 16 combinations (5%–33%), achieving up to 1.48× speedup while maintaining or even improving task success rates in most settings.

Updates of experimental results during Rebuttal

  • Section 4.1: Compared Gaussian fixed-step initialization, RTC and confirmed VLA-Rolling's advantages in success rate and inference latency.

  • Section 4.2: Added ablation to separate fixed and confidence-guided renoising from fixed and adaptive denoising budgets, verifying the contribution of confidence estimator mechanism.

  • Section 4.4: Expanded our real-world evaluation trials per method, providing stronger evidence of consistent inference savings without an observed loss in task success.

  • Appendix D.1: Further demostrated the strong correlation and the reasonableness of linear mapping between predicted confidence and denoising requirements, providing empirical motivation for the confidence-to-budget mapping.

  • Appendix E: Added few step action generation method as our comparison, showing the independence of our work from the 10-step baseline.

Updates of in-depth discussions during Rebuttal

  • Abstract & Introduction: Clarified that our work focuses on faster action generation without sacrificing accuracy. Modified our expression to maintaining the success rate to avoid overcolaiming.

  • Appendix F: Explored action smoothness and accumulated trajectory prediction errors, providing possible reasons for the increase in success rate.

  • Related Work: Included related warm-start action reuse method in our discussion to distinguish our work from other works.

We believe these additions comprehensively address the reviewer concerns and enhance the overall quality of our manuscript. All revisions will be highlighted in blue\color{blue}{\text{blue}} text in our updated version.

We look forward to the reviewers' favorable consideration.

Update 11.17 (Retesting under the unified standard of baseline and STEP)

  • Appendix E: Found STEP did not understand language modalities well, and retraining and retesting were conducted for fair comparison.

  • Appendix D.3: Unified inference delay testing standards with Overall performance (Table 1) in the main text and retested.

As we are now halfway through the discussion period, I encourage reviewers to follow up on the authors’ responses and authors to address any remaining questions to help ensure a constructive discussion.

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

Related papers

Discussion (0)

Sign in to comment.