acceptodds
Under review as a conference paper at ICLR 2027

Imagine a Little, Then Coordinate and Rethink: Modality-Adaptive Asynchronous Denoising for Unified Diffusion Vision-Language-Action Models

Abstract

Unified discrete-diffusion vision-language-action (VLA) models bridge world modeling and robot control by jointly generating future visual observations and actions within a single multimodal Transformer, providing explicit visual foresight for robotic manipulation. However, these models treat multimodal outputs as a homogeneous generation process, denoising them synchronously under a single shared schedule and overlooking the inherent heterogeneity across modalities: future-visual predictions are not only generation targets but also evolving conditions for action generation. Our analysis reveals distinct recovery dynamics across modalities and sample-dependent action responses to visual recovery. Motivated by these findings, we introduce MAAD-VLA, a modality-adaptive asynchronous denoising framework that preserves modality-specific progression while adaptively coordinating vision and action. During training, MAAD-VLA independently samples visual and action noise levels, exposing the model to diverse combinations of cross modal corruption so that it learns to denoise even when the two modalities are at different recovery stages. At inference, a response-guided two-stage denoising process first establishes partial visual foresight and then adaptively initiates action denoising alongside continued visual recovery under modality-specific schedules. We further introduce Action-centric Reconsideration, which enables self-correction by selectively remasking committed action tokens whose confidence drops sharply as the visual context evolves. Extensive experiments in both simulation and real-world environments demonstrate the effectiveness of MAAD-VLA, notably achieving an average task length of 4.82 on the CALVIN ABCDD benchmark.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision Language Action models, Embodied Manipulation

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_Mby3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

In this paper, the authors propose a new method for using asynchronous denoising of WM and action tokens in WM-added VLA models. During training, they use modality-decoupled noising. And during inference, they first denoise the visual tokens a little bit for priming, then begin action denoising when a inter-step KL threshold is crossed. They also add a remasking method to revisit action tokens which have lower confidence.

Critical strengths

The paper aims to solve a problem that is important in world-modeling VLAs, and their method gives novel analyses of how exactly decoupled denoising schedules should be done. Their method is quite intuitive.

Critical weaknesses

(1) Formatting and writing : The writing in this paper is quite hard to follow and does not lead naturally from concept to concept. I am genuinely interested in the problem this paper aims to fix, yet I had to spend a long time to parse what exactly the authors intended to do. The introduction gives a few analyses for motivation, but its hard to understand what exactly the analyses are supposed to be showing, and how that leads to the method the authors are proposing. The intro is too wordy on things that don't matter without actually giving the reader guidance on what problem the authors are solving, why its important, and how they solved it. I suggest a near rewrite for actual publication.

  • The formatting also needs a overhaul. For example, the real robot experiment results in Fig. 3(c) can't be seen unless you zoom in quite a bit. Also, why no parantheses around citations? Parts like "using Emu3 language tokenizer Wang et al. (2024), a VQ tokenizer Zheng et al. (2022), and FAST Pertsch et al. (2025), respectively" in the method section are hard to read.

(2) Novelty seems limited. Diffusion forcing-style decoupled timesteps have already been used a lot, with asynchronous schedules in papers like DUST. The paper's contribution is limited to the response based transition and the remasking part, which doesn't seem enough a contribution for an ICLR paper.

Questions for authors

The baselines UD-VLA seems to acheieve 4.64 in Table 1 but 4.45 in Table 3? Why is this so?

How much does the vision-first ordering help compared to just simple independent noise sampling?

3 comments

We appreciate the reviewer’s recognition of the importance of the problem we address in world-model-augmented VLAs, as well as the value of our analyses on decoupled denoising schedules. We have also carefully considered the reviewer’s constructive feedback on the presentation and substantially revised the Introduction in the updated PDF to improve the clarity of our problem formulation and motivation. We address the comments point by point below.

W1 : Formatting and writing

We thank the reviewer for this insightful comment. In response to your feedback, we have further refined the logical flow of the Introduction. For clarity, we summarize the key logic below:

  • Problem: Modality Heterogeneity. In unified discrete-diffusion VLAs, future-visual tokens serve both as generation targets and as evolving conditions for action prediction. However, visual and action tokens are denoised under a shared schedule despite their distinct functional roles and recovery characteristics.

  • Analysis 1 (Fig. 1(a)) → Modality-specific Recovery Dynamics. Under the shared schedule, visual tokens recover earlier and dominate cross-modal confidence ranking, while action tokens remain largely masked until later stages and are then committed rapidly. This heterogeneous recovery limits progressive action refinement, motivating modality-decoupled training and asynchronous inference.

  • Analysis 2 (Fig. 1(b)) → Sample-dependent Cross-modal Response. The onset of action-distribution changes varies across samples as visual tokens are progressively recovered, revealing sample-dependent sensitivity to evolving visual context. This motivates a sample-wise adaptive strategy that adjusts vision–action coordination by determining when to initiate action denoising based on each sample's action-response signal, rather than following a fixed schedule.

  • Incomplete Visual Context → Action-centric Reconsideration. During Stage II, visual and action tokens interact throughout asynchronous joint denoising, while visual tokens continue to be progressively recovered. Consequently, early action predictions may be formed under incomplete visual context, potentially leading to misleading action-token commitments. This motivates Action-centric Reconsideration, which revisits previous action predictions as richer visual information becomes available.

We have refined the Introduction to better establish the connections between our motivating analyses, the central research gap, and our methodological design. These revisions are highlighted in blue in the updated PDF.

Regarding figure readability, we have enlarged the quantitative results in Fig. 3(c) to improve their legibility. Regarding the citation style, we acknowledge this oversight and have corrected the formatting throughout the revised manuscript, ensuring that all in-text citations consistently follow the parenthetical style.

W2: Novelty.

We thank the reviewer for raising this concern. We would like to clarify that our contribution is not to introduce asynchronous denoising itself, but to address a further underexplored challenge: adaptive cross-modal coordination under sample-dependent and evolving visual conditions. While existing approaches (e.g., DUST and X-WAM) establish modality-specific denoising schedules, they lack a sample-wise adaptive mechanism for coordinating vision and action. Specifically, it remains unclear when action denoising should begin in response to evolving visual context of each sample, and how previously committed action predictions should be revised as this context continues to evolve.

To better address the your concern, we summarize our contributions as follows:

  • Contribution 1: Bridging Asynchronous Denoising and Discrete-Diffusion VLAs. We address the underexplored challenge of dynamic cross-modal coordination in discrete-diffusion VLAs, where visual tokens serve as evolving conditions for action prediction and discrete token commitments require careful coordination.
  • Contribution 2: Sample-adaptive Cross-modal Coordination. MAAD-VLA dynamically determines the onset of action denoising for each sample based on its online action-response signal, rather than relying on predefined schedules, achieving better performance than fixed asynchronous strategies.
  • Contribution 3: Action-centric Reconsideration. To address potentially misleading action commitments formed under incomplete visual context, we revisit previously committed action tokens as richer visual information becomes available.

The above contributions form a coherent framework centered on our core research question of adaptive cross-modal coordination, rather than merely combining existing techniques.

Q1: Baseline discrepancy

Thank you for this question. We clarify that 4.64 is the performance reported in the original UD-VLA paper. When evaluating the officially released checkpoint, we obtain 4.57. However, our independent SFT reproduction yields 4.43–4.45 across both 4×A100 and 4×A800 GPUs, indicating that the discrepancy is not attributable to GPU hardware.

We further examined the training dynamics and found that the original UD-VLA implementation computes cross-entropy over visual and action tokens jointly. Since visual tokens constitute the majority of output tokens, they contribute disproportionately to the overall training loss (approximately 70–80%), potentially weakening the learning signal for action prediction. To address this imbalance, we introduce modality-wise loss normalization to balance the contributions of the two modalities, helping mitigate the performance gap observed in our SFT reproduction, as evidenced by the ablation results in Table 3.

Q2: Effect of Vision-first Ordering

We evaluate vision-first ordering at both training and inference time, with controlled comparisons to isolate its effect at each stage. All comparisons are conducted without action remasking, as reported in Table VIII .

Table VIII Effect of Vision-First Ordering on CALVIN ABCD→D

StageConfigurationVision-first OrderingAvg. Len. ↑
TrainingShared sampling—4.45
Trainingmodality-independent sampling & shared inference schedule✗4.50
Trainingmodality-independent sampling & shared inference schedule✓4.52
InferenceAsynchronous denoising✗4.68
InferenceAsynchronous denoising✓4.73

During training, vision-first ordering yields a modest improvement from 4.50 to 4.52, as it primarily shapes the distribution of cross-modal noise configurations encountered during training, while the shared inference schedule does not explicitly exploit this visual-first progression. In contrast, applying vision-first ordering during asynchronous inference improves performance from 4.68 to 4.73 (+0.05), as it aligns the denoising progression with our Visual Priming strategy.

I thank the authors for the detailed response and additional experiments. However, with W1, I appreciate the format changes, but the revised manuscript does not really fix my concerns with the writing, especially in the introduction. There is little deviation from the original structure, leaving it longer rather than clearer for the reader, while figures' legibility is low across the paper. In addition, the novelty concerns I have still stand, since the points the authors gave in their rebuttal are still incremental at best.

Overall, the paper has some merits, but given the novelty issue and weakness in writing, I would require some surprising performance gains across benchmarks for this paper to merit a higher score, but the gains in this paper are marginal in my opinion. Hence, I keep my score as 2.

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

This paper presents MAAD-VLA, which coordinates future visual prediction and action generation through modality-specific corruption during training, KL-guided two-stage asynchronous denoising, and action-token remasking. The evaluation covers CALVIN, LIBERO, and real-robot manipulation tasks. The method achieves an average completed task sequence length of 4.82 on CALVIN ABCD-to-D.

Critical strengths
  1. The research problem is well defined and clearly motivated. The paper examines the different recovery dynamics of vision and action in unified VLAs and explains the coordination requirements arising from the role of predicted future vision as a condition for action generation. The analysis, training design, inference strategy, and evaluation form a coherent and complete study.
  2. The experimental study covers both simulation benchmarks and real-robot tasks, with extensive ablations. Experiments on CALVIN, LIBERO, and a physical robot are complemented by ablations of the training components, action-denoising start strategies, remasking criteria, and related parameters, helping clarify the contributions of the individual design choices.
Critical weaknesses
  1. The rationale for starting action denoising at the first KL threshold crossing needs further clarification. Appendix D explains the trade-off between visual information and the limited action-denoising budget, and Figure 4 provides empirical support. However, these results do not directly explain why denoising should begin at the first threshold crossing rather than after a substantial response has occurred and action predictions have begun to stabilize. In addition, Figure 1(b) does not identify the tasks or trajectories represented by the colored curves or specify whether each curve represents an individual sample or an aggregate.
  2. Simulation evaluation relies on older benchmarks. The experiments use CALVIN and LIBERO but do not include newer benchmarks such as LIBERO-Plus, RoboTwin, RoboDojo, or RoboCasa365. This limits the evidence for effectiveness under more diverse evaluation conditions. Could the authors select any one of these four benchmarks and provide additional quantitative experiments?
  3. The real-robot tasks are highly homogeneous. All three tasks involve placing differently colored blocks into a bowl and therefore share essentially the same pick-and-place behavior. These tasks provide limited evidence of the method's ability to handle more complex manipulation skills and task structures.
Questions for authors
  1. Figure clarification: Which task or trajectory does each colored curve in Figure 1(b) represent? Is each curve an individual sample or an aggregate over multiple samples? Please provide a legend and clarify how the curves are computed.
  2. Start-rule clarification: Appendix D provides a reasonable explanation of the trade-off between visual information and the action-denoising budget. Why is action denoising initiated when the action KL first exceeds the threshold, rather than after a substantial response has occurred and action predictions have begun to stabilize? Is there analysis or experimental evidence supporting this specific choice?
  3. Implementation clarification: If the KL never exceeds the threshold, or the transition is triggered too late, how does the algorithm ensure that a valid action sequence is generated? Please specify the fallback strategy in Algorithm 1.
  4. Additional real-robot evaluation: Could the authors supplement the evaluation with more complex real-robot tasks beyond simple pick-and-place, such as multistage manipulation or tasks requiring more precise object interactions, and report quantitative results to assess the method's effectiveness in these settings?
3 comments

We thank the reviewer for the thoughtful assessment, particularly for recognizing our clear research motivation and the coherence of the overall study, from problem analysis to training, inference, and evaluation. We also appreciate the constructive questions, which help further clarify our contributions. We address them in order.

Q1/W1: Clarification of curves in Fig. 1(b)

Thank you for pointing this out, and we apologize for the confusion. Each curve in Fig. 1(b) corresponds to an individual sample trajectory sampled from the CALVIN ABCD→D evaluation set, illustrating the variation in action responses across different samples. We have clarified this in the revised figure caption.

Q2: Choice of Transition Criterion

To capture how action predictions respond to progressive visual recovery, we adopt inter-step KL divergence as an intuitive measure of action distribution changes. Similar measures have also been explored in other diffusion settings to characterize temporal changes in prediction distributions [1, 2]. We adopt this straightforward signal as an online gate for action denoising.

To further address this concern, we compare First Crossing against two alternative transition rules, Post-Peak Decline and After-Stabilization, on CALVIN ABCD→D. As shown in Table IV, First Crossing achieves the highest Avg. Len., although all three strategies yield comparable performance. Moreover, First Crossing triggers action denoising earlier (26.96 steps vs. 27.91 and 30.24), leaving more of the denoising budget for action refinement. The comparable performance of the three strategies suggests that delaying the transition provides no additional benefit under our default inference budget. These results support our choice of First Crossing as the default transition strategy in MAAD-VLA.

Table IV Comparison of Action-Denoising Transition Criteria

Start ruleCriterionAvg. Len. ↑Avg. transition step
First Crossing (ours)�0��0�4.8226.96
Post-Peak DeclineFirst decline after KL peak4.8027.91
After-Stabilization�1��1�4.7930.24

Q3: Fallback when the KL never exceeds the threshold

Thank you for pointing this out. Our implementation includes a fallback for this case, which was omitted from Algorithm 1: if the KL signal has not exceeded δa\delta_a, Stage II is triggered at a fixed visual-recovery bound. The bound follows from requiring that at least half of the denoising steps be reserved for action denoising. With our default T = 72, half of the steps corresponds to about 29% visual recovery under the cosine schedule. This is also consistent with Fig. 4, where fixed transitions at around 30% visual recovery perform comparably well. We have added this fallback to Algorithm 1 in the revised version.

W2: Evaluation on a newer benchmark

Following the reviewer's recommendation, we additionally evaluate UD-VLA and MAAD-VLA on LIBERO-Plus. Due to limited computational resources and the tight rebuttal timeline, we focus on the Spatial and Object suites. We further compare our method with representative baselines, including WorldVLA[3] and UniVLA[4], using their reproduced results on the LIBERO-Plus benchmark. For UD-VLA, we use its officially released checkpoint and the same evaluation setup as MAAD-VLA to ensure a fair comparison. In Table V, we report overall success rates aggregated across all evaluation rollouts.

Table V Overall Success Rates (%) on LIBERO-Plus Spatial and Object Suites

MethodSpatialObject
WorldVLA [3]32.528.6
UniVLA [4]55.536.7
UD-VLA59.653.3
MAAD-VLA (Ours)61.755.8

These results demonstrate improved generalization performance of MAAD-VLA over UD-VLA under the evaluated environmental perturbations. We note that differences in training data and data diversity may also influence robustness under environmental perturbations. Moving forward, exploring more diverse training strategies, such as camera viewpoint augmentation, remains an important direction for further improving the generalization capability of MAAD-VLA.

[1] Stability-Weighted Decoding for Diffusion Language Models. Arxiv: 2604.17068.

[2] KLASS: KL-Guided Fast Inference in Masked Diffusion Models. NeurIPS 2025.

[3] WorldVLA: Towards Autoregressive Action World Model. NeurIPS 2026.

[4] UniVLA: Learning to Act Anywhere with Task-Centric Latent Actions. RSS 2025.

Q4/W3 : Additional Real-Robot Evaluation

Thank you for this valuable suggestion, which helps us further strengthen the empirical validation of our approach. We extend our real-world evaluation with two more challenging multi-stage manipulation tasks:

  • Task A: Plate Retrieval and Cloth Placement. The robot retrieves a plate, places it on the table, and puts a cloth inside it.
  • Task B: Three-Plate Stacking. The robot retrieves three plates and stacks them together on the table in a specified color order (purple →\rightarrow green →\rightarrow pink).

Table VI: Real-World Tasks and Subtask Instructions

TaskSubtaskInstruction
A: Plate Retrieval and Cloth PlacementA1Pick up the plate and place it on the table.
A2Put the cloth inside the plate.
B: Three-Plate StackingB1Pick up the purple plate and place it on the table.
B2Pick up the green plate and stack it on the purple one.
B3Pick up the pink plate and stack it on the green one.

Experiments are conducted on a single-arm RealMan RM65 robot, with 50 demonstrations collected per task. Table VI presents the decomposition of the two tasks into sequential subtasks. For Task B, we vary initial plate locations during training, placing some plates on the table rather than the rack to encourage visual object identification beyond fixed spatial priors. Detailed experimental configurations and additional results are provided in Appendix G.

We evaluate both MAAD-VLA and UD-VLA over 20 evaluation trials per method for each task, and report cumulative success rates at each manipulation stage, as summarized in Table VII.

Table VII: Real-World Cumulative Success Rates (%)

TaskMethodStage: A1/B1Stage: A2/B2Stage: B3
A: Plate Retrieval and Cloth PlacementMAAD-VLA95%80%--
UD-VLA95%75%--
B: Three-Plate StackingMAAD-VLA90%80%65%
UD-VLA85%70%45%

Overall, MAAD-VLA achieves higher cumulative success rates than UD-VLA across both tasks. Notably, MAAD-VLA outperforms UD-VLA by 20 percentage points (65% vs. 45%) in terms of final cumulative success rate on the more challenging three-plate stacking task. This superior long-horizon performance is consistent with our simulation findings, which demonstrates the effectiveness of MAAD-VLA in challenging multi-stage real-world manipulation tasks.

Thank you for the additional real-robot experiments and the comparison of KL start rules. These additions strengthen the paper and address my main concerns. I maintain my positive rating of 3.

Reviewer_qWxPOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper build on top of UD-VLA a discrete diffusion based VLA, which denoises future image and action tokens on a shared schedule. MAAD-VLA hypothesizes that visual and action tokens have different learning dynamics and hence modality gap, then recover information at different rates. Building on this, it addresses this decoupling of denoising schedules during both training and inference. During training, it samples independent noise levels for vision and action, and during inference it introduces adaptive asynchronous denoising mechanism. First denoising the images and then it uses uncertainty scores from the action predictions to decide when to start the action denoising. Earlier works explored asynchronous denoising in continuous diffusion, this work fills the gap in discrete diffusion setting.

Critical strengths
  1. Strong empirical motivation from figure 1 showing modality recovery gap
  2. Good theoretical grounding and rigorous ablations provided in the experiments
  3. Clear and intuitive paper writing
  4. Good improvement in success rate for CALVIN, libero-long especially for long horizon tasks and the numbers are supported by the analysis as well
  5. Architectural improvements on both training and inference side are complementary for the overall approach. Adaptive trigger and remasking claims are especially well supported by the experiments.
Critical weaknesses
  1. No direct comparison with DUST/X-WAM style fixed asynchronous schedule other than baseline UD-VLA
  2. Role of visual foresight is not ablatied, the appendix shows how predicted images are blurry from wrist views. It is not isolated well, if the ‘visual priming’ helps because of the image content or because the model gets extra computation
  3. Given adaptive asynchronous mechanism, there could be potential gains on latency side which has not been provided for the method.
  4. Add related works from the field such as Diffusion forcing (per token individual noise levels), UVA (Unified video action model with decouple denoising), DAM-VLA (Decoupled asynchronous Multimodal VLA)
  5. Incremental work on existing ideas but a solid an principled execution, so this is a minor point for the review
Questions for authors
  1. For weakness 1, can you provide one external baseline ?
  2. If ‘visual primining’ uses corrupted/more blurry visual tokens, how does it affect the performance?
  3. For weakness 3, can you provide wall clock latency per action chunk, time to first action , inference-control frequency and timing profiling ?
  4. What happens if the KL signal never crosses the threshold or crosses very late, is there a limit on when action denoising must start ?
4 comments

We thank the reviewer for the thoughtful and expert review, and for recognising the empirical motivation, the theoretical grounding, and the complementary contributions of our training and inference strategies. We address the reviewer's concerns point by point below.

Q1: Comparison with a DUST/X-WAM-style asynchronous schedule.

Following this recommendation, we further include comparisons with inference-time scheduling strategies from DUST[1] and X-WAM[2]. However, DUST and X-WAM adopt different attention-mask designs and cross-modal information flows from MAAD-VLA. Directly incorporating their complete methods would require substantial architectural modifications and retraining. We therefore adapt their inference-time scheduling strategies to our discrete-diffusion setting to investigate their effectiveness under a unified MAAD-VLA architecture. This allows us to isolate the effects of scheduling strategies without confounding architectural and training differences.

For a fair comparison, all schedules are evaluated using the same trained MAAD-VLA checkpoint, with only the inference schedule modified. We keep the total denoising steps at (T=72) with the same cosine schedule and disable action remasking for all compared strategies. We consider the following representative fixed asynchronous schedules:

  • DUST-style (fixed update-rate ratio). DUST adopts different denoising rates by updating action tokens less frequently than visual tokens, reflecting their distinct denoising requirements: high-dimensional visual generation benefits from more refinement, whereas low-dimensional actions typically converge in fewer steps. We adapt this strategy to discrete diffusion: action tokens are updated once every qq steps (Na=72/qN_a=72/q) updates, e.g., Na=18N_a=18 for q=4q=4 and Na=9N_a=9 for q=8q=8. We evaluate q∈{4,8}q \in \lbrace 4, 8 \rbrace, following DUST's test-time-scaling settings.
  • X-WAM-style (early action finish). In X-WAM, both modalities start from noise. The action branch is denoised for TaT_a steps, while the video branch undergoes TO>TaT_O > T_a denoising steps (10 and 50, respectively, in its main setting). After TaT_a steps, the clean actions are dispatched for execution, while video denoising continues for the remaining TO−TaT_O-T_a steps. Under discrete diffusion with T=72T=72, visual tokens are updated at every step, whereas action tokens are fully revealed within the first TaT_a steps and remain fixed thereafter. We set Ta=16T_a=16, approximately following X-WAM's default 1:5 action-to-video denoising-step ratio, and additionally evaluate Ta∈{24,36}T_a\in \lbrace 24,36 \rbrace​.

Table I Comparison of Inference-Time Asynchronous Scheduling Strategies

ScheduleSettingAvg. Len. ↑
Synchronous baselineOriginal shared schedule, w/o our inference strategy4.61
DUST-styleq=4q = 44.29
DUST-styleq=8q = 83.22
X-WAM-styleTa=16T_a = 162.87
X-WAM-styleTa=24T_a = 243.30
X-WAM-styleTa=36T_a = 363.79
Ours (KL, w/o remask)δa=0.005δ_a = 0.0054.73

The results are summarized in Table I. We note that directly transferring these fixed schedules to MAAD-VLA has certain limitations:

  • Different attention-mask designs: DUST adopts bidirectional cross-modal interaction, while X-WAM uses bidirectional full attention over a unified video-state-action sequence. In contrast, MAAD-VLA employs full attention within each modality and causal attention across modalities, allowing visual tokens to condition action generation without reverse action-to-vision conditioning. Thus, the original scheduling strategies cannot be directly transferred without considering these differences.
  • Different diffusion formulations: DUST and X-WAM are originally developed under continuous diffusion formulations, whereas MAAD-VLA operates in discrete token diffusion. Directly adapting their complete methods would therefore require architectural modifications and substantial retraining.

We will further clarify these architectural differences and adaptation considerations in Section 3.3 of the revised paper. A more comprehensive investigation of asynchronous scheduling strategies across different attention designs and discrete-diffusion architectures is left for future work.

Q2: Effect of Future-Visual Token Corruption

To investigate the effect of degraded future-visual representations, we conduct inference-time ablations by replacing 50% and 100% of future-visual tokens with a special blurry token (ID: 5), without modifying the training procedure.

In this experiment, MAAD-VLA achieves average task lengths of 4.824, 4.25, and 3.89 with the original, 50%-corrupted, and fully corrupted future-visual tokens, respectively. Given our IDM-style formulation, action generation is conditioned on intermediate future-visual representations. Replacing these representations with a blurry token therefore introduces an inference-time distribution shift, leading to performance degradation. Nevertheless, even when all future-visual tokens are replaced with the blurry token, the model retains some long-horizon manipulation capability by relying on current visual observations and learned action patterns.

Regarding the blurry wrist-view predictions noted in Appendix Fig. 7, their reconstruction is more challenging due to substantial viewpoint variations during manipulation, compared with the relatively static global view. However, wrist-view predictions already capture task-relevant spatial cues, such as target-object locations and their approximate positions relative to the gripper, which provide useful guidance for action generation.

Q3: Latency analysis

Thank you for raising this concern. We provide additional latency profiling on a single RTX 4090, as summarized in Table II and Table III.

Table II Inference Latency and Throughput

MetricValue
Latency per chunk / time to first action (ms)2767
Policy inference frequency (Hz)0.36
Action throughput (chunk = 10, CALVIN)3.61 actions/s
Action throughput (chunk = 20, real robot)7.23 actions/s

Table III Time Profiling

ComponentLatency Breakdown (%)
Transformer Forward90.0%
Logits Projection + Softmax4.2%
Token Sampling2.7%
KV-Cache Prefill1.2%
VQ Encoding0.6%
Reveal Selection (Confidence Top-k)0.4%
KL Computation0.3%
CPU-side Input Preparation0.3%
Action Remasking0.2%
Token Update0.1%

Our profiling shows that Transformer forward accounts for approximately 90% of the total inference time, whereas the additional operations introduced by MAAD-VLA, including KL computation, action-centric remasking, and token updates, collectively account for only 0.6%. This indicates that our modality-adaptive coordination and action-centric refinement introduce negligible computational overhead, with the overall latency remaining predominantly dominated by Transformer forward passes.

MAAD-VLA adopts non-streaming parallel decoding, with inference latency primarily dominated by iterative Transformer forward passes over future-visual tokens, whose sequence length is approximately 10× that of action tokens. We leave more efficient visual-token refinement while preserving task-relevant action guidance to future work.

Q4: What if the KL signal never crosses the threshold, or crosses very late?

Thank you for raising this. Our implementation bounds Visual Priming, which we omitted from Algorithm 1. Stage II starts at the latest once half of the denoising steps have elapsed, serving as a fallback if the KL signal has not exceeded δa\delta_a by then. With a total budget of T=72T=72 steps, this bound corresponds to approximately 29% of future-visual tokens being recovered under the cosine schedule. This is also consistent with Fig. 4, where a fixed transition at 30% visual-token recovery achieves relatively strong performance. Since vision and action share the same denoising budget, this fallback ensures that action denoising is not left with too few steps. We have added this fallback to Algorithm 1 and Sec. 3.3 in the revised version, with the corresponding changes highlighted in blue.

W4: Related work.

Thank you for pointing out these works. We have added them to the Related Work section. In brief, Diffusion Forcing [3] introduces token-wise noise-level assignments for flexible denoising schedules across sequence positions, while UVA [4] learns unified video-action latent representations with decoupled diffusion heads. DAM-VLA [5] maintains modality-specific latent buffers updated at their respective sensor rates, which are accessed by the action head through gated cross-attention. In contrast, MAAD-VLA conditions action denoising on progressively recovered future-visual tokens and adaptively determines when to initiate action denoising based on the sample-wise action response to evolving visual foresight.

[1] Dual-Stream Diffusion for World-Model Augmented Vision-Language-Action Model. ICLR 26.

[2] Unified 4D World Action Modeling from Video Priors with Asynchronous Denoising. Arxiv :2604.26694.

[3] Diffusion Forcing: Next-token Prediction Meets Full-Sequence Diffusion. NeurIPS 2024.

[4] Unified Video Action Model. RSS 2025.

[5] DAM-VLA: Decoupled Asynchronous Multimodal Vision Language Action Model. CoRL 2026.

I appreciate the response and most of my concerns seem to be addressed. As I stated in my original review I find this to be a very well executed but incremental work. I already scored the submission quite positive before, but for a clear acceptance I would require stronger insights.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.