acceptodds
Under review as a conference paper at ICLR 2027

AsyncVLA: Asynchronous Flow Matching for Vision-Language-Action Models

13 views

Abstract

Vision-language-action (VLA) models have recently emerged as a powerful paradigm for building generalist robots. However, traditional VLA models that generate actions through flow matching (FM) typically rely on rigid and uniform time schedules, i.e., synchronous FM (SFM). Without action context awareness and asynchronous self-correction, SFM becomes unstable in long-horizon tasks, where a single action error can cascade into failure. In this work, we propose asynchronous flow matching VLA (AsyncVLA), a novel framework that introduces temporal flexibility in asynchronous FM (AFM) and enables self-correction in action generation. AsyncVLA breaks from the vanilla SFM in VLA models by generating the action tokens in a non-uniform time schedule with action context awareness. Besides, our method introduces the confidence rater to extract confidence of the initially generated actions, enabling the model to selectively refine inaccurate action tokens before execution. Moreover, we propose a unified training procedure for SFM and AFM that endows a single model with both modes, improving KV-cache utilization. Extensive experiments on robotic manipulation benchmarks demonstrate that AsyncVLA is data-efficient and exhibits self-correction ability. AsyncVLA outperforms existing methods across both simulation and real-world evaluations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
asynchronous flow matching, vision-language-action (VLA)

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ySqrOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes AsyncVLA, a two-stage framework for confidence-guided action generation in vision-language-action models. A shared flow-matching policy first generates a continuous action chunk through synchronous flow matching (SFM). A separately trained confidence rater then identifies low-confidence actions, which are re-noised and selectively regenerated through asynchronous flow matching (AFM), while retaining higher-confidence predictions as context. Unified training uses randomly sampled masks and perturbed context actions to support both generation modes, with visual-language KV-cache reuse during refinement. Experiments on LIBERO, SimplerEnv, and real-world robotic manipulation demonstrate improved success rates over the evaluated baselines, including 87.0% versus 77.0% for π0.5 on four real-world tasks.

Critical strengths
  1. Convincing evidence for selective correction.

On WidowX, AsyncVLA achieves 70.8% success versus 47.9% for SFM-only and 61.5% for cardinality-matched random masking. Paired evaluations further support the improvement, helping distinguish confidence-guided refinement from simply increasing denoising steps or randomly selecting actions.

  1. Strong real-world evaluation.

Experiments on four PiPER manipulation tasks show an average success rate of 87.0%, compared with 77.0% for π0.5. The evaluation uses matched initial configurations, counterbalanced execution, and paired statistical analysis, providing meaningful evidence beyond simulation.

Critical weaknesses
  1. Limited independent validation of confidence-guided selection.

The cardinality-matched comparison uses one action-noise realization, while the five-noise-seed experiment examines full remasking rather than confidence-guided AFM. The confidence-rater diagnostics also reuse the checkpoint-selection split, and training-seed variability is not evaluated.

  1. Inference efficiency and caching need clarification.

The paper reports 95.9 ms total inference latency versus 83.2 ms for SFM alone. However, the exact VL/action attention mask underlying KV-cache reuse is insufficiently specified, and a latency comparison against similarly cached 20-step SFM is missing. These details would help establish the practical efficiency trade-off.

  1. The data-efficiency claim is not fully isolated.

The quarter-data experiment changes both the training objective and inference procedure. AFM also conditions on retained actions, making its training loss not directly comparable to SFM loss. Consequently, the results demonstrate an advantage of the complete framework but do not isolate the contribution of unified training to data efficiency.

Questions for authors

Can you repeat the confidence-guided and cardinality-matched random-mask comparisons across multiple action-noise seeds to assess selection robustness?

What is the exact VL/action attention mask? Does cached inference match uncached computation, and how does its latency compare with similarly cached 20-step SFM?

Can you evaluate confidence-rater error discrimination on an untouched test split, particularly when most actions in a chunk are inaccurate?

9 comments

We thank the reviewer for recognizing the effectiveness of confidence-guided selective correction, the controlled WidowX evaluations, and the strong real-world results. We particularly appreciate the constructive suggestions regarding selection robustness, caching efficiency, and the isolation of training-data efficiency. We address all three weaknesses and questions with additional controlled experiments and implementation details.

W1: Independent validation and robustness of confidence-guided selection

We strengthen the evidence for confidence-guided selection by evaluating multiple action-noise seeds, independent training seeds, and confidence discrimination on an untouched test split.

1. Distinguishing learned selection from additional refinement. The original WidowX comparison reports 70.8% success for AsyncVLA, compared with 47.9% for 10-step SFM, 51.1% for 20-step SFM, and 61.5% for cardinality-matched random AFM. These results indicate that the improvement cannot be explained solely by additional denoising steps or the number of regenerated actions.

Importantly, cardinality matching ensures that the random baseline regenerates exactly the same number of positions as confidence-guided AFM for each action chunk. Thus, the comparison evaluates which actions are selected, rather than how many actions are regenerated.

2. Robustness to action-noise realizations. We agree that a single action-noise realization provides limited evidence for selection robustness. Different SFM noise samples produce different initial predictions, which may affect both the confidence ranking and subsequent correction. Therefore, selection robustness should be evaluated across multiple initial action-noise realizations while keeping model parameters, evaluation episodes, and refinement budgets fixed.

We also distinguish action-noise variability from training-seed variability. The former evaluates the robustness of inference-time selection under a fixed checkpoint; the latter evaluates whether the improvement persists across independently trained models.

3. Independent confidence validation. We additionally separate the confidence-rater evaluation split from the split used for checkpoint selection. Because the confidence targets are normalized within each action chunk, they represent relative imitation accuracy rather than calibrated absolute reliability. The appropriate evaluation therefore measures whether the learned scores discriminate action errors on unseen examples, including chunks containing many inaccurate actions.

Additional experiments: five action-noise seeds and independent training seeds.

We repeat the WidowX evaluation using five action-noise seeds. Each method uses the same initial SFM noise realization, paired evaluation episodes, and a matched refinement-noise tensor. The random baseline regenerates the same number of positions as confidence-guided AFM for each chunk.

Action-noise seedRandom AFM (%)Confidence AFM (%)Improvement (pp)
061.570.8+9.3
162.569.8+7.3
260.471.9+11.5
361.570.8+9.3
462.572.9+10.4
Mean ± SD61.7 ± 0.971.2 ± 1.2+9.6

The original 70.8% and 61.5% results are retained as seed 0. Confidence-guided AFM consistently outperforms cardinality-matched random refinement across all five noise realizations, with improvements ranging from 7.3 to 11.5 percentage points.

We further repeat training with three independent seeds, varying model initialization, data ordering, and training-time mask sampling while keeping the backbone initialization, training data, and evaluation protocol fixed.

Training seedRandom AFM (%)Confidence AFM (%)
061.570.8
160.468.8
262.572.9
Mean ± SD61.5 ± 1.170.8 ± 2.1

The improvement persists across independent training runs, with confidence-guided AFM consistently outperforming random selection.

Together, these experiments establish that the selection advantage is robust to both inference-time stochasticity and training variability, rather than depending on a favorable individual noise realization or model checkpoint.

W2: Inference efficiency, attention masks, and KV-cache correctness

We clarify the exact attention structure enabling VL KV-cache reuse, verify cached and uncached numerical consistency, and compare AsyncVLA against SFM with an equivalent caching implementation and denoising-step budget.

1. Exact VL/action attention structure.

Let the input sequence consist of a VL prefix followed by action tokens. The attention mask has the following block structure, where rows denote queries and columns denote keys:

Query / KeyVL tokensAction tokens
VL tokensNative VL prefix maskBlocked
Action tokensAllowedAllowed

The VL prefix retains the backbone's native causal and multimodal attention structure. Action tokens attend to the complete VL prefix and all action positions, including both retained and regenerated actions.

Crucially, VL queries do not attend to action-token keys. Therefore, the VL hidden states and their per-layer keys and values are independent of the evolving action predictions.

This block structure is what permits exact VL prefix reuse. The reference to full attention in the methodology concerns the action-query attention rows, not unrestricted bidirectional attention from VL tokens back into action tokens.

2. Cache reuse during SFM and AFM.

The VL prefix is processed once for each observation and instruction. Its per-layer KV tensors are retained throughout the SFM and AFM denoising steps.

At each FM step, action-token representations and their corresponding keys and values are recomputed because the noisy action inputs change with the integration time. Only the invariant VL prefix KV tensors are reused.

The confidence rater operates on the completed SFM action chunk and the corresponding VL representations. It does not modify the cached VL prefix.

For the same observation, language instruction, attention mask, positional encoding, and model parameters, cached and uncached inference therefore implement the same attention computation, up to floating-point numerical differences.

3. Fair comparison with additional SFM steps.

AsyncVLA uses 10 SFM steps followed by 10 AFM steps. Although 20-step SFM has the same number of FM integration steps, the two methods do not have exactly identical computational costs because AsyncVLA additionally evaluates the confidence rater.

We therefore compare wall-clock latency directly rather than assuming that equal integration-step counts imply identical runtime.

We also distinguish the total inference latency from the marginal refinement overhead. The originally reported latency increases from 83.2 ms for 10-step SFM to 95.9 ms for AsyncVLA, corresponding to 12.7 ms of additional latency, or approximately 15.3%.

Additional experiments: cache equivalence and matched latency.

We evaluate cached and uncached inference on 1,000 held-out action chunks using identical inputs, initial noise, masks, solver steps, and numerical precision.

Consistency metricResult
Mean absolute action difference1.2 × 10⁻⁵
Maximum absolute action difference3.2 × 10⁻⁴
Confidence-mask agreement100%
Cached vs. uncached task success70.8% / 70.8%

The small numerical differences arise from floating-point execution and do not change the measured task success. This supports the numerical consistency of cached and uncached inference.

We further profile inference under the same GPU, precision, batch size, attention implementation, and timing protocol:

Inference configurationUncached (ms)Cached (ms)WidowX success (%)
SFM, 10 steps111.783.247.9
SFM, 20 steps139.594.651.1
AsyncVLA, 10+10 steps140.895.970.8

All configurations use the same VL prefix-caching strategy, and the reported latency includes model execution, confidence evaluation where applicable, and synchronization overhead.

Compared with similarly cached 20-step SFM, AsyncVLA increases latency by only 1.3 ms while improving success from 51.1% to 70.8%, a gain of 19.7 percentage points.

Thus, the improvement is not obtained at zero additional cost, but confidence-guided refinement provides a substantially better success–latency trade-off than allocating the same number of FM steps to synchronous denoising.

W3: Isolating data efficiency from the second inference stage

We agree that lower training loss alone does not demonstrate data efficiency and that the complete two-stage framework does not isolate the contribution of unified training. We therefore evaluate unified training with AFM disabled at inference.

1. Interpretation of the training loss.

The AFM objective supervises selected action positions conditioned on retained action context, whereas SFM predicts the entire action chunk.

Even when the losses are normalized by the number of supervised positions, the underlying prediction tasks differ. Therefore, the lower reported training loss of AsyncVLA (0.0042 vs. 0.0076) should not be treated as direct evidence of superior data efficiency.

We instead assess data efficiency using held-out task success under an identical demonstration budget and matched backbone training schedules.

2. Separating training from inference.

Unified training samples different masks and perturbed action contexts from the same demonstrations. This exposes the model to both complete generation and conditional refinement, providing additional training configurations without collecting new demonstrations.

However, a complete AsyncVLA versus SFM comparison also includes the effect of test-time refinement. To isolate training, we evaluate a unified-trained model using SFM-only inference.

Both training variants use the same pretrained initialization, the same 25% LIBERO-Spatial demonstrations, 200 training epochs, and matched backbone optimization schedules.

The confidence rater is trained separately, and its additional training computation is accounted for rather than treated as equivalent to the SFM training budget.

3. Distinguishing training, refinement, and selection.

The comparison includes four settings:

  • Standard SFM training with 10-step SFM inference.
  • Unified training with 10-step SFM inference.
  • Unified training with 10+10-step random AFM.
  • Unified training with 10+10-step confidence-guided AFM.

The first two settings isolate the effect associated with the training objective without changing the inference procedure.

The third setting measures the benefit of adding conditional refinement without learned confidence selection.

The final setting evaluates the additional contribution of confidence-guided action selection.

Additional experiments: controlled quarter-data evaluation.

TrainingInferenceLIBERO-Spatial success (%)
SFMSFM, 10 steps86.2
UnifiedSFM, 10 steps90.4
UnifiedRandom AFM, 10+10 steps92.1
UnifiedConfidence AFM, 10+10 steps95.8

Unified training increases SFM-only success from 86.2% to 90.4%, establishing a 4.2-point improvement without additional inference.

Conditional refinement further increases success to 92.1%, while confidence-guided selection reaches 95.8%, providing an additional 3.7-point improvement over cardinality-matched random correction.

These comparisons distinguish the improvement associated with unified training from the separate benefits of conditional refinement and confidence-guided selection.

We therefore characterize data efficiency as improved downstream task performance under the same demonstration budget, rather than lower masked training loss or lower total training computation.

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

The paper proposes AsyncVLA for flow-matching, aiming to mitigate the FM error propagation during long-horizon tasks. Specifically, AsycnVLA first generates an action chunk using synchronous flow matching, then regenerates low-confidence action tokens while retaining the remaining predictions as context. A learned confidence rater selects tokens for correction, and unified masked training enables one model to support both initial generation and refinement.Experiments on both simulation and real-world evaluations against multiple baselines are provided. AsyncVLA exhibits strong performance.

Critical strengths
  • the problem is very relevant in robotics

  • the paper is very well written

  • superior numerical performance of the proposed methods

  • the controlled WidowX comparisons show improvements over additional synchronous denoising steps and cardinality-matched random masking,

  • real-world evaluation supports the practical potential

Critical weaknesses

1.Confidence-guided token-level correction is already present in discrete-diffusion VLA methods, including secondary remasking of uncertain action tokens. The contribution appears to be adapting this principle to continuous-action flow matching through Gaussian re-noising, a learned confidence rater, and unified masked training. The paper should more clearly distinguish these contributions from existing correction mechanisms and explain why the adaptation constitutes a substantial methodological advance. Comparisons between complete systems do not isolate the benefit of the continuous-action formulation. 2. the notion of asynchrony and its isolated contribution need to be clarified. The implemented time conditioning assigns time zero to retained tokens and a shared evolving time to all regenerated tokens. Thus, the method distinguishes fixed context from tokens undergoing regeneration, rather than providing independently adaptive schedules for individual tokens. This distinction should be stated more precisely. The current ablations support the overall selective-regeneration pipeline, but do not separately establish the contribution of asynchronous time conditioning relative to masking, retained action context, and confidence-based selection. 3. The confidence targets are derived from demonstration-relative action errors normalized within each chunk. These scores therefore measure relative imitation accuracy rather than absolute action reliability. 4. The comparisons provide useful system-level evidence, but do not fully isolate the proposed mechanism. The evaluated methods differ in pretrained checkpoints, adaptation strategies, and computational budgets, while AsyncVLA additionally receives a confidence rater, extra training updates, and a second generation stage.  5. LIBERO-Long provides relevant long-horizon evaluation. However, the comparison against different VLA systems does not isolate whether selective correction improves long-horizon performance or mitigates accumulated errors.

Questions for authors

n/a. I am open to adjust my rating if my concerns are clarified.

8 comments

We thank the reviewer for recognizing the relevance of the problem, the clarity of the presentation, the strong experimental performance, and the practical potential demonstrated by real-world evaluations. We appreciate the constructive concerns regarding methodological novelty, asynchronous conditioning, confidence estimation, experimental controls, and long-horizon error propagation. We address each point below with additional clarification and controlled experiments.

W1: Methodological novelty beyond discrete-diffusion correction

AsyncVLA does not claim that selective token regeneration is itself new. Its methodological contribution is a unified continuous-action FM framework that combines learned error localization, selective Gaussian re-noising, asynchronous time conditioning, and action-context-aware refinement.

1. Continuous-action correction requires a different formulation. Discrete-Diffusion VLA models discretized actions through categorical prediction and uses token probabilities to guide decoding and secondary remasking. In contrast, AsyncVLA directly models continuous robot actions through FM velocity prediction. Continuous FM does not provide categorical token probabilities, so confidence cannot be obtained directly from decoder logits. AsyncVLA instead learns a confidence rater from demonstration-relative action errors and uses its predictions to select continuous action positions for Gaussian re-noising and regeneration.

This distinction changes both the confidence-estimation mechanism and the correction dynamics. Discrete remasking replaces selected categorical tokens with masked states, whereas AFM regenerates continuous-valued actions through a learned velocity field while conditioning on retained action values.

2. Asynchronous time conditioning enables conditional continuous-action generation. AsyncVLA distinguishes action positions undergoing regeneration from positions providing fixed context. The retained actions remain at FM time zero, while regenerated actions follow a shared evolving FM time. This allows the model to condition noisy action predictions on already resolved continuous actions within the same chunk.

Importantly, the retained actions are not merely excluded from the output loss. They participate in the model's attention computation and provide temporal action context for generating corrected trajectories. Unified training further exposes the model to different masks and perturbed retained actions, allowing the same backbone to support both complete generation and conditional refinement.

3. The contribution is supported by mechanism-level evidence rather than complete-system comparisons alone. The original LIBERO evaluation reports 97.4% average success for AsyncVLA versus 96.3% for Discrete-Diffusion VLA. We agree that this comparison alone cannot isolate the continuous-action formulation because the systems differ in architecture and training.

The relevant methodological question is whether the proposed continuous-action correction mechanism provides benefits under controlled backbone, data, and inference settings. We therefore distinguish the effects of action representation, confidence-guided selection, and asynchronous conditioning instead of attributing the entire improvement to continuous actions alone.

Additional experiments: controlled continuous–discrete comparison.

We implement a discrete-diffusion counterpart using the same Qwen2.5-VL-3B backbone initialization, robot demonstrations, action horizon, training-update budget, and WidowX evaluation protocol. The discrete counterpart uses discretized action prediction and confidence-guided secondary remasking, while AsyncVLA retains continuous velocity prediction and Gaussian re-noising. Both use 20 decoding or refinement steps; their actual computational costs are accounted for separately.

MethodAction representationCorrectionWidowX success (%)
Discrete-diffusion counterpartDiscreteConfidence remasking65.6
AsyncVLAContinuousConfidence AFM70.8

With the pretrained backbone, training data, and evaluation protocol controlled, AsyncVLA exceeds the discrete counterpart by 5.2 percentage points. This comparison reduces the confounding factors present in complete-system benchmarks, although it does not attribute the entire difference exclusively to action representation.

Together with the mechanism ablations below, the results support the contribution of continuous-action confidence-guided refinement rather than a claim of inventing selective regeneration.

W2: Definition and isolated contribution of asynchrony

We clarify that the implemented asynchrony consists of two time groups: retained actions fixed at time zero and regenerated actions sharing an evolving FM time. AsyncVLA does not assign independently adaptive schedules to every regenerated token.

1. Precise definition of asynchronous time conditioning.

Let mi∈{0,1}m_i\in\{0,1\} denote whether action position ii is selected for regeneration. The asynchronous time assignment is

ti(τ)=miτ.t_i(\tau)=m_i\tau.

Consequently, ti(τ)=0t_i(\tau)=0 for retained positions and ti(τ)=τt_i(\tau)=\tau for regenerated positions. All regenerated positions share the same evolving time, while retained positions remain fixed.

The AFM update is

a^τ−δ=a^τ−δVθ(o,ℓ,a^τ,τm)⊙m.\hat{\mathbf a}^{\tau-\delta}=\hat{\mathbf a}^{\tau}-\delta V_\theta(\mathbf o,\ell,\hat{\mathbf a}^{\tau},\tau\mathbf m)\odot\mathbf m.

Thus, only selected positions are updated, while the model conditions on the complete mixed-time action chunk.

We clarify that the term asynchronous refers to nonuniform time conditioning between retained and regenerated positions, not independently adaptive per-token schedules.

2. Distinguishing asynchronous conditioning from masking.

Masking determines which actions are regenerated, but does not by itself determine how the model represents the denoising states of retained and regenerated actions.

The asynchronous time embedding explicitly informs the model that retained actions correspond to resolved predictions while regenerated actions are still undergoing denoising. This provides the FM predictor with information about the different roles of the action positions.

To isolate this effect, we compare against uniform-time conditioning that preserves the same selected positions, retained action values, refinement steps, and model architecture, but assigns the same evolving time embedding to all positions. The comparison isolates the contribution of the time-conditioning scheme.

3. Distinguishing retained context from asynchronous conditioning.

Retaining reliable actions provides more than a fixed output constraint. These actions serve as continuous temporal context for predicting corrected actions.

To isolate this contribution, we remove retained-action information from the conditioning inputs of the AFM stage while preserving the original retained values in the final output. The model therefore regenerates the same selected positions but cannot use the retained predictions as informative action context.

This separates the contribution of conditional action generation from that of selective masking or asynchronous time embeddings.

Additional experiments: factorized mechanism ablations.

We conduct controlled ablations on WidowX using the same backbone initialization, training data, evaluation identities, and 10+10-step refinement schedule. Each modified conditioning scheme is trained consistently with its inference configuration.

VariantSelectionRetained contextAsync timeSuccess (%)
A: Random AFMRandomYesYes61.5
B: No retained contextConfidenceNoYes63.5
C: Uniform-time AFMConfidenceYesNo66.7
D: AsyncVLAConfidenceYesYes70.8

All random-selection comparisons use cardinality-matched masks. Variant B removes retained-action information only from AFM conditioning, while preserving retained values in the final action chunk. Variant C preserves the confidence-selected mask and context but replaces asynchronous time conditioning with uniform-time conditioning.

The results establish three complementary effects. Confidence-guided selection improves over random regeneration, retaining action context improves conditional correction, and asynchronous time conditioning provides a further 4.1-point gain over its uniform-time counterpart.

The performance of the complete method is therefore not explained by masking alone. It arises from the interaction of confidence-guided selection, retained action context, and nonuniform time conditioning.

W3: Relative confidence versus absolute action reliability

We agree that the confidence targets represent relative imitation accuracy within each action chunk, rather than calibrated probabilities of absolute action reliability. We clarify this distinction and evaluate the rater accordingly.

1. Interpretation of the confidence targets.

For an action chunk, let eie_i denote the demonstration-relative prediction error at position ii. The confidence target is defined by

qi=1−α−βei−min⁡jejmax⁡jej−min⁡jej+ϵ.q_i=1-\alpha-\beta\frac{e_i-\min_j e_j}{\max_j e_j-\min_j e_j+\epsilon}.

The normalization is performed within each chunk. Consequently, higher confidence indicates a smaller error relative to other positions in the same chunk.

This definition is intentional: AFM uses confidence primarily to identify which positions should be reconsidered and which predictions provide relatively more reliable context.

However, the resulting scores should not be interpreted as calibrated probabilities that an action is correct or that a task will succeed. We make this distinction explicit in the revised description.

2. Why relative confidence is useful for selective refinement.

The correction mechanism requires a ranking of candidate action positions rather than an absolute estimate of task success. Relative confidence supports this objective by identifying comparatively inaccurate positions even when action-error scales differ across tasks or action dimensions.

The model then regenerates selected actions while retaining comparatively accurate predictions as context.

This formulation is particularly useful when errors are concentrated in a subset of the action chunk, because selective correction can target problematic positions without unnecessarily modifying the remaining predictions.

3. Behavior when most actions are inaccurate.

We also acknowledge an important limitation of within-chunk normalization. When all action predictions have large absolute errors, the normalization can still assign relatively high confidence to the least inaccurate positions.

Therefore, relative confidence does not guarantee that retained actions are absolutely correct. In such cases, the available action context may itself be unreliable, limiting the effectiveness of selective refinement.

We distinguish this limitation from the intended role of the confidence rater and avoid interpreting its output as an absolute safety or reliability guarantee.

Additional experiments: held-out error discrimination and high-error analysis.

We evaluate the confidence rater on an untouched test split not used for confidence-rater training or checkpoint selection. We measure the Spearman correlation between confidence scores and demonstration-relative action errors, together with recall of the highest-error 25% of positions when selecting the same number of positions for correction.

We additionally evaluate the subset of chunks belonging to the highest quartile of chunk-mean action error.

Evaluation subsetSpearman correlationTop-25% error recall (%)
All held-out chunks-0.6267.8
High-error chunks-0.4957.3
Random selection—25.0

The negative correlations confirm that higher predicted confidence is associated with lower demonstration-relative error. The rater identifies high-error positions substantially more effectively than random selection, including on the difficult high-error subset.

The lower discrimination performance on high-error chunks also clarifies the method's limitation: relative confidence remains useful for ranking actions but does not provide an absolute guarantee of reliable retained context.

These results support the intended interpretation of the confidence rater as an error-localization mechanism rather than an absolute reliability estimator.

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

The paper introduces AsyncVLA, a VLA model that regenerates some actions with low confidence in the initially generated action chunk. The framework has a confidence rater to find the unconfident actions. The confident actions are used as context, and unconfident action tokens are renoised and refined in asynchronous flow matching. The work is evaluated on multiple simulation benchmarks and real-world manipulation tasks, and show promising performance.

Critical strengths
  1. The work is well-motivated and introduces a selective-correction mechanism for continuous robot actions. The confidence rater estimates unreliable actions, and the model uses previously reliable actions as context to refine them. I think this is a reasonable design.
  2. The model shows better performance compared to the baseline methods on multiple benchmarks. There is also real-world experiment included. Also, this confidence-guided refinement outperforms the SFM-only inference result.
Critical weaknesses
  1. In Section 4, the paper claims data efficiency of the framework by lower training loss and higher success rate. However, for AsyncVLA, the model only predicts unconfident actions, but SFM predicts all, so lower training loss is expected, and I do not quite understand why this shows data efficiency. For a higher success rate, there is an additional pass for AsyncVLA compared to SFM, so the gap might comes from this second pass at test time, not training time.

  2. The novelty of the paper is limited, and there have been works doing similar things of regenerating some tokens/action tokens while keeping the rest. For example. RTC [1] has done this on action chunks. Some discrete diffusion VLA works also regenerate tokens based on confidence [2]. It would be better to compare with these works.

[1] Black, Kevin, Manuel Galliker, and Sergey Levine. "Real-time execution of action chunking flow policies." Advances in Neural Information Processing Systems 38 (2026): 33383-33407.

[2]Liang, Zhixuan, et al. "Discrete diffusion vla: Bringing discrete diffusion to action decoding in vision-language-action policies." arXiv preprint arXiv:2508.20072 (2025).

Questions for authors
  1. See weakness 1. Can you provide more clarification on why a lower training loss and higher success rate show data efficiency?
  2. Can you please clarify more on the key difference between this work and RTC, discrete diffusion VLA works?
7 comments

We thank the reviewer for recognizing the motivation and effectiveness of AsyncVLA. We address the concerns regarding data efficiency, additional inference computation, and distinctions from prior work.

W1: Data efficiency and additional inference computation

We clarify that data efficiency should be evaluated through downstream performance under matched training-data budgets, rather than comparing training losses alone.

The lower training loss of AsyncVLA (0.0042 vs. 0.0076) does not independently establish data efficiency, since AFM predicts selected actions conditioned on retained context, whereas SFM predicts the complete action chunk. We therefore emphasize held-out task success under identical demonstration budgets and training schedules.

Unified training improves data utilization by exposing the shared model to different action masks and perturbed contexts without requiring additional demonstrations. Meanwhile, the second inference stage provides a separate benefit through selective correction. These two contributions can be distinguished by evaluating the unified model with AFM disabled.

Additional controlled experiments. We compare four settings using 25% of LIBERO-Spatial data, identical pretrained initialization, and 200 training epochs:

TrainingInferenceSuccess (%)
SFMSFM (10 steps)86.2
UnifiedSFM (10 steps)90.4
UnifiedRandom AFM (10+10 steps)92.1
UnifiedConfidence AFM (10+10 steps)95.8

Unified training improves SFM-only performance by 4.2 percentage points without additional inference. Confidence-guided AFM provides a further 5.4-point improvement and exceeds cardinality-matched random refinement by 3.7 points. These results separate the benefit of unified training from that of additional inference and demonstrate improved task success with the same demonstration budget.

W2: Novelty and comparisons with RTC and Discrete-Diffusion VLA

Our contribution is not selective regeneration in isolation, but its integration into continuous-action FM through learned error localization, asynchronous time conditioning, and unified training.

1. Distinction from RTC. RTC primarily addresses execution latency and action continuity across consecutive chunks. It preserves actions already committed to execution and inpaints the remaining actions to accommodate asynchronous execution. In contrast, AsyncVLA corrects low-confidence predictions within an initially generated chunk before execution. Its confidence rater explicitly identifies relatively inaccurate actions, while AFM regenerates them conditioned on retained predictions. Thus, RTC addresses execution-time continuity, whereas AsyncVLA addresses prediction-time error correction.

2. Distinction from Discrete-Diffusion VLA. Discrete-Diffusion VLA performs confidence-based remasking over discrete action tokens using categorical prediction scores. AsyncVLA operates directly on continuous actions through Gaussian re-noising and FM velocity prediction. Since continuous actions lack categorical token probabilities, AsyncVLA learns a confidence rater from relative action errors. Moreover, retained tokens remain fixed at time zero, while regenerated tokens share an evolving FM time. Unified masked training enables the same model to support both SFM and AFM.

3. Controlled mechanism comparisons. The original LIBERO evaluation reports 97.4% average success for AsyncVLA versus 96.3% for Discrete-Diffusion VLA. We further isolate the contribution of different correction mechanisms on WidowX:

VariantSuccess (%)
SFM, 10 steps47.9
SFM, 20 steps51.1
AFM, cardinality-matched random masking61.5
AFM, Bernoulli random masking62.5
AFM, uniform-time conditioning66.7
AsyncVLA, confidence-guided AFM70.8

The two random-mask controls use different selection protocols. The uniform-time control retains confidence-selected masks and action context while removing asynchronous time assignments. Increasing SFM to 20 steps yields only 51.1%, while selective refinement substantially improves success. Under matched refinement computation, confidence-guided AFM achieves 70.8%, exceeding uniform-time conditioning by 4.1 points and cardinality-matched random masking by 9.3 points. These comparisons establish the complementary contributions of confidence-guided selection and asynchronous conditioning beyond additional denoising.

Q1: Why do lower training loss and higher success demonstrate data efficiency?

We agree that lower training loss alone is insufficient and that the additional inference stage must be distinguished from the training contribution.

The AFM and SFM objectives differ in supervised positions and conditioning, so their losses are not directly comparable as measures of data efficiency. We instead define data efficiency through held-out success under identical demonstration budgets and training schedules.

The additional inference cost is also explicitly accounted for. On WidowX, 20-step SFM improves success only from 47.9% to 51.1%, compared with 70.8% for AsyncVLA. The measured inference latency is 95.9 ms for AsyncVLA versus 83.2 ms for 10-step SFM. Therefore, AsyncVLA improves control reliability with a moderate additional inference cost rather than reducing inference computation.

We additionally isolate the training contribution using SFM-only inference after unified training. With 25% LIBERO-Spatial data, unified training achieves 90.4% success using 10-step SFM, compared with 86.2% for standard SFM training. Adding random AFM increases success to 92.1%, while confidence-guided AFM reaches 95.8%. These controlled comparisons directly establish the separate contributions of training-data utilization, additional refinement, and confidence-guided selection.

Q2: What are the key differences from RTC and Discrete-Diffusion VLA?

The principal differences concern the correction objective, action representation, confidence estimation, and training mechanism.

RTC focuses on maintaining continuous execution across successive action chunks by preserving committed actions and inpainting future predictions. AsyncVLA instead identifies and corrects relatively inaccurate predictions within a generated chunk before execution.

Discrete-Diffusion VLA uses categorical confidence to select discrete action tokens for remasking. AsyncVLA performs continuous-action correction through a learned error-based confidence rater, Gaussian re-noising, and action-conditioned FM. Its unified training further supports initial generation and selective refinement within the same model.

Importantly, the learned confidence represents relative imitation accuracy within each chunk, rather than an absolute probability of successful execution. AsyncVLA also uses two time assignments—zero for retained tokens and a shared evolving time for regenerated tokens—rather than independently adaptive schedules for all positions.

We further provide controlled comparisons to substantiate these distinctions. On WidowX, AsyncVLA achieves 70.8% success, compared with 61.5% for cardinality-matched random refinement and 66.7% for uniform-time refinement. Together with the original LIBERO comparison against Discrete-Diffusion VLA (97.4% vs. 96.3%), these results demonstrate the effectiveness of continuous-action confidence-guided correction while distinguishing AsyncVLA from existing selective-regeneration methods.

Thank you once again for taking the time to review our manuscript. We have carefully addressed all the concerns raised in your review and provided detailed responses to fully resolve each of the points raised.

As the discussion period is coming to a close, we would like to kindly follow up on our previous response in case you have not yet had a chance to review it. We appreciate the many demands on your time and would be grateful for any opportunity you may have to review our responses.

Thank you once again for your time and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.