acceptodds
Under review as a conference paper at ICLR 2027

Flash-WAM: Modality-Aware Distillation for World Action Models

Abstract

World-action models (WAMs) jointly generate future video and robot actions through iterative diffusion, achieving strong performance on manipulation benchmarks but requiring tens of sequential denoising steps per chunk, a cost that precludes real-time control. Step distillation has emerged as the natural remedy, but off-the-shelf methods break down in the joint video-action setting. Video latents are high-dimensional and redundant, so they tolerate small errors, whereas actions are low-dimensional and precision-critical. Existing distillation methods are designed for a single modality and distill both streams in the same way, which causes task success to collapse. We introduce Flash-WAM, a modality-aware step-distillation framework inspired by consistency distillation. We show that the standard consistency function provides almost no learning signal at low noise levels, where actions receive their final refinement. Flash-WAM therefore derives a separate consistency function for each stream. The action stream uses a function whose learning signal scales linearly with the noise level, and video keeps the standard function. We apply Flash-WAM to two WAMs with different architectures, LingBot-VA and Motus, reducing inference to a single denoising step per modality. On LingBot-VA, this cuts per-chunk latency from  s to  ms on an NVIDIA L40S () while retaining success on RoboTwin 2.0, versus for naive consistency distillation, and on LIBERO. In real-world experiments on a Unitree G1 humanoid robot, Flash-WAM outperforms video-only distillation by points on Lingbot-VA. On Motus, a single step retains 77.0% success, compared with 23.0% for naive distillation. Code is available at https://anonymous.4open.science/r/Flash-WAM/.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Efficient Inference, Robotic AI

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_v2QJOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

Flash-WAM accelerates world-action models through modality-aware consistency distillation. It retains the LCM parameterization for video while using a clean-sample parameterization for actions to reduce low-noise gradient attenuation. Experiments on LingBot-VA and Motus, simulation benchmarks, and a real robot demonstrate substantial inference acceleration while retaining much of the teacher’s control performance.

Critical strengths

1,The method addresses a practical bottleneck with a simple modification that requires no architectural changes. 2,The reported speed–accuracy trade-off is promising, reaching 23.3× acceleration on LingBot-VA. 3,Evaluation across two WAM architectures, two simulation benchmarks, and real hardware supports practical relevance.

Critical weaknesses

1,The mechanism is not fully isolated. Proposition 1 characterizes a scalar coefficient, rather than the complete optimization dynamics. Shared linear parameterization and noise-dependent loss reweighting are important missing controls.

2,Baseline fairness needs clarification. DMD2 uses four-step training rollouts and updates its student every five iterations, raising questions about convergence and alignment with one-/two-step evaluation.

3,The revision would benefit from discussing the following complementary approaches:From Articulated Kinematics to Routed Visual Control for Action-Conditioned Surgical Video Generation. Its modality routing and budgeted execution offer a relevant perspective on efficient controllable generation. Please discuss how reducing computation within a denoising step relates to Flash-WAM’s reduction in denoising steps, and whether the approaches could be combined. ORV: 4D Occupancy-centric Robot Video Generation. This provides useful context for evaluating whether Flash-WAM preserves control alignment and useful visual predictions after distillation.

4, Qualitative examples alone do not establish that accelerated generation retains temporal consistency and agreement with predicted actions.

5,Confidence intervals are missing, despite only 20 real-world trials per task and relatively small gains over some stronger baselines, such as video-only LCM with regularization at 1v/2a.

Questions for authors

1,Can the authors compare the proposed design against using the linear parameterization for both modalities and against noise-reweighted LCM? Gradient norms and prediction errors across noise levels would help substantiate the proposed explanation.

2,Does the DMD2 training budget of 2,000 steps count critic iterations or student updates? Please clarify the effective optimization budget and evaluate training rollouts matched to the target inference steps.

3,Please report quantitative video fidelity and action–video consistency, together with end-to-end robot latency, including observation processing and action-chunk scheduling.

4,Please reconcile the teacher results in Tables 1 and 4: 91.64/90.86 versus 92.93/91.55 on Clean/Randomized splits. The 1v/2a naive joint LCM Randomized result also differs between these tables: 22.07 versus 20.08.

7 comments

We thank the reviewer for the detailed review and the constructive suggestions, which we address below.

W1 + Q1: isolating the mechanism

We agree that the shared linear parameterization and the noise-dependent loss reweighting are informative controls, and we thank the reviewer for this suggestion. We therefore added both controls, together with direct measurements of the gradient each stream receives and of its prediction error across noise levels. In short, Flash-WAM outperforms both controls, and reweighting the loss recovers almost none of the gap.

Controls. All variants use the same recipe and training budget (2,000 steps) and differ only in the consistency functions. We evaluated them on a 10-task RoboTwin subset.

Action function / video function1v/1a1v/2a
standard / standard (naive joint distillation)56.042.6
standard with noise-reweighted loss / standard58.845.2
linear / linear (shared linear parameterization)75.687.2
linear / standard (Flash-WAM)81.490.0

Reweighting the standard function's action loss so that its gradient matches ours at every noise level does not recover performance. Reweighting changes how strongly each noise level is fit, but not what the network is fit to. Under the standard function, the learned velocity does not yield the clean-sample estimate that the few-step sampler uses at inference, as the prediction errors below show. The action function must therefore both scale linearly near zero noise and coincide with the clean-sample estimate. The shared linear parameterization also satisfies both conditions on the action stream and recovers most of the gap. Keeping the standard function for the video stream, as Flash-WAM does, performs better still at both budgets.

Gradient norms across noise levels. We set every frame of one stream to noise level σ\sigma, build the consistency target exactly as in training, and measure the norm of the gradient of that stream's loss with respect to the model parameters. The table below reports the action stream at the start of distillation, where both functions share the same weights.

σ\sigma0.010.050.10.20.51.0
standard function1.4e-33.1e-29.0e-20.230.660.85
ours 7.9e-20.370.540.741.553.11

The gradient of the standard function vanishes quadratically as σ→0\sigma \to 0 and is 58× smaller than ours at σ=0.01\sigma = 0.01, while ours decreases only linearly. The video stream is trained mostly at high noise, where the standard function already provides a large gradient.

Prediction errors across noise levels. Per the reviewer's request, we also measure the mean distance between the model's clean action estimate and the distillation target at each noise level, before and after distillation.

σ\sigma0.010.050.10.20.5
before distillation0.00070.00080.00090.00110.0026
standard function, after distillation0.00690.02090.03450.05660.0805
ours, after distillation0.00080.00120.00140.00180.0036

Under the standard function, the action estimate at low noise drifts 10–40× away from the target during distillation, while ours stays at its initial accuracy.

W2 + Q2: DMD2 training

We thank the reviewer for raising these points and address each of them below.

Rollout steps in training versus inference. To align the training rollout exactly with one-step inference, we implemented DMD2 with a one-step rollout, so that the student is trained at exactly the noise level used for one-step sampling. All other settings are unchanged, including the flow-matching loss on the action stream, and we report it after 2,000, 4,000 and 6,000 steps (400, 800 and 1,200 student updates). Even with this exact alignment, DMD2 does not close the gap and stays below Flash-WAM on both splits at every checkpoint. Success rates on the 10-task subset (50 trials per task) are below.

MethodTraining stepsClean 1v/1aRandomized 1v/1a
DMD2 baseline with one-step rollout2,00071.060.8
DMD2 baseline with one-step rollout4,00070.261.4
DMD2 baseline with one-step rollout6,00065.659.0
Flash-WAM2,00081.470.8

Longer training does not narrow the gap either, since the baseline gains at most 0.6 points from 2,000 to 4,000 steps and falls below its 2,000-step result at 6,000 steps, which shows the gap comes from the distillation objective, not from the rollout length or the training budget.

Training budget and convergence. Each training step updates the critic, and the student is updated at every fifth step, so 2,000 steps correspond to 2,000 critic updates and 400 student updates. The number of student updates is not directly comparable between DMD and consistency distillation, since the two methods use entirely different training signals and update rules. To check whether DMD2 had converged, we continued training the DMD2 baseline of the paper to 6,000 steps. As shown in the above table, further training did not improve the result.

Updating the student every five iterations. DMD training is generally unstable, because the critic must keep track of the student's distribution for the distribution-matching gradient to remain informative. In our initial runs with a 1:1 update ratio, the student moved ahead of the critic within the first iterations and the DMD loss signal collapsed. We therefore followed the DMD hyperparameters of Helios [1], which update the student once every five critic updates. With this ratio, the student and the critic learned in step and training converged stably.

[1] Helios: Real Real-Time Long Video Generation Model, Yuan et. al.

W3: related work

We thank the reviewer for suggesting these two papers. KVLR (Li et al., 2026) reduces the computation within each denoising step by routing control signals and executing under a compute budget, whereas Flash-WAM reduces the number of denoising steps. The two are complementary and can be combined. Savings within a step also become relatively more important once inference uses only one or two steps, because these forward passes and observation processing then make up most of the latency per chunk (W4). ORV (Yang et al., 2026) generates robot videos conditioned on 4D occupancy aligned with actions. Its focus on control alignment relates directly to the action–video consistency we now report under W4 response. In the revision, we discuss both papers in the Appendix, next to the video fidelity and action–video consistency results reported under W4.

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

Flash-WAM aims to accelerate diffusion-based world-action models (WAMs) by converting their original multi-step diffusion into a single-step process. Its main contribution is the discovery that, in action generation — a setting that demands very high precision — standard consistency distillation suffers from a gradient vanishing problem at low noise levels. The paper addresses this by modifying the noise-scaling coefficient in the consistency function, changing its decay near zero noise from quadratic to linear, while the video diffusion module retains the original LCM scaling. The method is evaluated in both simulation and on a real robot, achieving strong results — far better than standard consistency distillation (on RoboTwin 2.0 at 1v/2a, 85.5% vs. only 24.0% for naive joint LCM).

Critical strengths

S1. A meaningful problem with sharp insight and solid analysis. Action latency in embodied manipulation remains an important open problem. The paper analyzes, from both theoretical and empirical perspectives, the key obstacle to distilling WAMs to a single step — the vanishing gradient as the noise level approaches zero — providing a direct theoretical explanation (the order analysis via Taylor expansion in Proposition 1) and indirect empirical evidence (the performance collapse of naive joint LCM and its recovery after modifying the coefficient).

S2. A small and precise fix and Thorough experiments. The authors propose a precise solution targeting exactly this point (taking a constant a and a linear b in the negative sigma for the action-stream consistency function), achieving large performance gains without adding redundant modules. Relevant ablations are provided (Table 4 and Table 8). The method is evaluated on multiple models (Motus and LingBot-VA, which differ in both architecture and noise schedule) as well as in simulation and on a real robot, with complete results reported that clearly demonstrate its advantages.

Critical weaknesses

W1. The investigation into the root cause of the problem is not sufficiently thorough. The paper supports the claimed mechanism only through theory (order analysis of the Taylor expansion as noise approaches zero) and indirect experiments (e.g., performance recovery after adjusting the noise-scaling coefficient), without directly observing, during training, the gradient magnitudes actually received by the video and action modules at different noise levels.

W2. The ablation of the method itself is not detailed enough. The paper proposes only one way to modify the noise coefficient in consistency distillation. A natural alternative, however, is to keep the standard LCM parametrization and instead compensate for the gradient scaling at low noise levels through loss reweighting; this is not compared against. It is also unclear whether other linear coefficients (or other coefficients satisfying the boundary condition with a nonzero derivative at zero) would perform better.

Questions for authors

Q1. Could the authors provide the gradient magnitudes actually received by the action and video streams at different noise levels during training, for both standard LCM and the proposed method (e.g., curves of the gradient norm versus the noise level, or the effective learning signal weighted by the noise coefficient)?

Q2. Could the authors run experiments with a loss-reweighting variant (standard LCM parametrization with a loss weighted toward the low-noise region) or with other linear scaling coefficients?

Q3. Some numbers in Table 1 and Table 4 appear inconsistent in their reporting conventions. For example, the teacher's Clean/Randomized success is reported as 91.64/90.86 in Table 1 but as 92.93/91.55 in Table 4; naive joint LCM at 1v/2a is reported as 25.88/22.07 in Table 1 but as 25.88/20.08 in Table 4. Could the authors explain the inconsistencies between Table 1 and Table 4 noted in W3?

5 comments

We thank the reviewer for the careful review and the constructive suggestions, which we address below.

W1 + Q1: gradient magnitudes

We agree with the reviewer that this is an informative experiment. We measured the gradient that each stream receives at different noise levels, for the standard consistency function and for ours. In short, the measurements confirm the mechanism, since under the standard function the action stream receives almost no gradient at low noise and its predictions there drift during distillation, whereas under ours the gradient scales linearly and the predictions stay accurate. Using fixed training batches, we set every frame of one stream to noise level σ\sigma, build the consistency target exactly as in training, and record the norm of the gradient of that stream's loss with respect to the model parameters. The table reports the action stream at the start of distillation, where both functions share the same weights.

σ\sigma0.010.050.10.20.51.0
standard function1.4e-33.1e-29.0e-20.230.660.85
ours7.9e-20.370.540.741.553.11

As predicted by Proposition 1, the action gradient of the standard function vanishes quadratically as σ→0\sigma \to 0 and is 58× smaller than ours at σ=0.01\sigma = 0.01, while ours decreases only linearly. This starvation persists after 2,000 steps of distillation, and under the standard function the model's clean action estimate at low noise drifts 10–40× away from the teacher, whereas ours keeps it at its initial accuracy. The video stream is trained mostly at high noise, where the standard function already provides a large gradient, which is why only the action stream is affected.

W2 + Q2: loss reweighting and other coefficients

Per reviewer request, we ran both suggested controls with the same recipe and training budget as in the paper (2,000 steps), and evaluated them on a 10-task RoboTwin subset (Clean split, 50 trials per task). In short, loss reweighting does not recover performance, whereas another coefficient that also scales linearly performs on par with Flash-WAM, so the form of the action function, not the gradient magnitude alone, decides the result.

Action consistency function1v/1a1v/2a
standard function (naive joint distillation)56.042.6
standard function + loss reweighting58.845.2
LCM boundary scaling with timestep rescaling80.089.6
Flash-WAM81.490.0

For the reweighting control, we weight the standard function's action loss per frame by σ/∣b(σ)∣\sigma / |b(\sigma)|, so that its per-element gradient equals ours at every noise level. This does not recover performance. Reweighting changes how strongly each noise level is fit, but not what the network is fit to. Under the standard function, the network matches its targets through a(σ)xσ+b(σ)vθa(\sigma) x_\sigma + b(\sigma) v_\theta with a(σ)≠1a(\sigma) \neq 1. The velocity it learns therefore does not give the correct clean-sample estimate xσ−σvθx_\sigma - \sigma v_\theta, which is what the few-step sampler uses at inference. The drift of this estimate reported above under W1 is consistent with this. Linear gradient scaling is thus necessary but not sufficient, and the action function must also coincide with the clean-sample estimate. Our choice (a≡1a \equiv 1, b(σ)=−σb(\sigma) = -\sigma) satisfies both.

For other coefficients, we used the boundary scaling of the original LCM implementation, which rescales the noise level by a large constant before computing cskipc_\text{skip} and coutc_\text{out}. Over the whole training noise range this gives a(σ)≈1a(\sigma) \approx 1 and b(σ)≈−σb(\sigma) \approx -\sigma, so it satisfies both conditions, and it performs on par with Flash-WAM. Any function that satisfies the criterion therefore works, and ours is its simplest member, with no scaling hyperparameter. Our naive joint baseline uses the standard parametrization of Eq. 3 without this rescaling, and the revision names the baseline accordingly. Both controls and the refined statement of the criterion are added to the revision.

Q3: Table 1 vs. Table 4

We thank the reviewer for pointing this out. The two teacher rows came from different sources. Table 1 reported our own evaluation of LingBot-VA (91.64 Clean / 90.86 Randomized), obtained by following the training and evaluation recipe of the original paper. Table 4 reported the numbers published by LingBot-VA (92.93 / 91.55). To make the tables consistent, the revision moves our reproduction to the appendix and uses the published numbers in both tables. We would also be happy to adopt any other suggestions. The naive joint LCM entry at 1v/2a in Table 4 (Randomized 20.08) is a typo. The correct value is 22.07, as in Table 1, and it matches the per-horizon entries of that row. We have corrected it.

We hope these additions address the reviewer's questions.

Reviewer_fD4vOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Flash-WAM, a modality-aware consistency distillation method for world-action models. By using different consistency functions for video and action generation, the method reduces denoising steps while preserving task performance. Experiments are conducted on RoboTwin, LIBERO, Motus, and a Unitree G1 robot.

Critical strengths
  1. The paper addresses an important practical limitation of world-action models. The reported inference speedups are substantial, with relatively small performance degradation compared to the original models.
  2. The method is evaluated on two different WAM architectures, with comparisons against existing distillation approaches and several ablation studies.
Critical weaknesses
  1. The proposed action consistency function closely resembles the standard clean-sample prediction in flow matching. The main contribution is adapting existing consistency distillation techniques to different modalities, rather than introducing a fundamentally new approach.
  2. The real-world experiments cover only three relatively simple manipulation tasks, with 20 trials per task. The paper provides no complete demonstration videos, making it difficult to assess execution quality and failure cases. The limited number of trials also raises concerns about the statistical reliability of the reported improvements.
Questions for authors
  1. Can the authors clarify the methodological novelty of the proposed action consistency function compared with standard flow-matching formulations?
  2. Can the authors provide complete real-robot demonstration videos, including successful and failed trials, along with additional trials or statistical analysis to support the reported improvements?
3 comments

We thank the reviewer for the thoughtful review. We address the weaknesses and questions below.

W1 + Q1: main contribution

We clarify that our contribution is more than a choice of consistency functions. The motivating finding is that applying consistency distillation with the standard consistency function to both streams of a world-action model collapses task success on RoboTwin from 92% to 24%, even though the same function distills the video stream alone. Our contribution is a formal analysis that explains why this happens and derives the optimal fix directly from it. We (i) identify this failure mode and trace it to a structural cause, namely that the per-modality noise schedules place the action stream in a low-noise regime where the gradient of the standard consistency loss vanishes; (ii) formalize the achievable gradient scaling under the consistency boundary condition (Proposition 1); and (iii) use this analysis to select each modality's consistency function rather than choosing it heuristically. The two consistency functions are therefore a consequence of the analysis, not the contribution itself.

We summarize the analysis of Sections 4.1 and 4.2 below. A consistency function f=a(σ)xσ+b(σ)vθf = a(\sigma)x_\sigma + b(\sigma)v_\theta depends on the network only through vθv_\theta, so its gradient at noise level σ\sigma is scaled by ∣b(σ)∣|b(\sigma)|. Proposition 1 shows that, under the boundary condition, the best achievable scaling near zero noise is linear, ∣b(σ)∣=Θ(σ)|b(\sigma)| = \Theta(\sigma), and that it is attained if and only if b′(0)≠0b'(0) \neq 0. The standard function has b′(0)=0b'(0) = 0, so its gradient vanishes quadratically where the action stream is trained and refined. The video stream, trained mostly at high noise, is largely unaffected. The action stream therefore needs a function with b′(0)≠0b'(0) \neq 0, and our action function (a≡1a \equiv 1, b(σ)=−σb(\sigma) = -\sigma) is the minimal such choice, with no additional parameters or hyperparameters. The revision also measures this effect directly. At σ=0.01\sigma = 0.01, the gradient of the standard action loss is 58× smaller than that of ours.

W2 + Q2: real robot

We thank the reviewer for this suggestion. As requested, we have provided real-world demonstration videos of both successful and failed trials. They are in the supplementary material, in the real-world folder, for each task at 1v/2a and 1v/1a. We have also increased the number of trials from 20 to 30 per task and method. Table below reports the updated results.

MethodNvN^v/NaN^aT1T2T3Success
LingBot-VA (teacher)3 / 1016/3022/3024/3068.9%
LingBot-VA (reduced steps)1 / 28/308/3018/3037.8%
LingBot-VA + Video-only LCM1 / 210/3014/3017/3045.6%
Flash-WAM1 / 214/3019/3020/3058.9%
LingBot-VA (reduced steps)1 / 15/309/309/3025.6%
LingBot-VA + Video-only LCM1 / 16/3012/3012/3033.3%
Flash-WAM1 / 112/3015/3018/3050.0%

With the additional trials, the improvements are statistically significant (p-values from a two-sided Cochran–Mantel–Haenszel test, stratified by task and step budget). Across the three tasks and both step budgets, Flash-WAM outperforms video-only LCM (p=0.005p = 0.005) and reduced-step inference (p<10−4p < 10^{-4}), and it is better than both on every task at both budgets. At 1v/2a, Flash-WAM is the only accelerated configuration whose gap to the teacher is not statistically significant (p=0.21p = 0.21).

Task complexity. Each task tests a different capability. T1 is a multi-stage task in which the robot opens the pot's lid and then places the plastic potato inside. T2 requires picking the correct object next to a distractor. T3 requires precise placement on a marked target. These skills are comparable to those in RoboTwin, and the tasks are far from saturated, since the teacher itself reaches 68.9%. The hardware experiment tests whether a one- or two-step student keeps the teacher's performance under real execution timing, and the setting is sufficient to separate the methods.

We hope these clarifications address the reviewer's concerns and would be glad to answer any further questions.

AuthorsOpenReview ↗

Follow-up on our response

We thank the reviewer again for the thoughtful review. As the discussion ends on November 18, we would like to check whether our response addressed your two concerns.

  • Novelty of the action function. We clarified that our contribution is the formal analysis of why the standard consistency function fails for actions, from which our function follows as the optimal choice.
  • Real-robot evaluation. We increased the trials to 30 per task and method and added videos of successful and failed trials to the supplementary material.

If our response resolves these concerns, we would be grateful if you would consider revisiting your score. We are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.