acceptodds
Under review as a conference paper at ICLR 2027

Systematic Multi-Level Acceleration for World Action Models

Abstract

World Action Models (WAMs) are emerging as a promising alternative to Vision-Language-Action (VLA) models. Existing VLA acceleration methods primarily reduce VLM computation and are not directly applicable to diffusion-based WAMs. Moreover, WAM acceleration methods typically target redundancy along only a single computational axis, overlooking substantial redundancy across the full pipeline, resulting in either limited speedups or drops in success rate. To address these limitations, we propose MARS, a training-free multi-level acceleration framework that systematically exploits distinct sources of redundancy at three computational levels: (1) across Transformer blocks, we selectively skip redundant attention components; (2) across diffusion denoising steps, we perform action-aware routing to adaptively skip diffusion steps; (3) across consecutive action chunks, we reuse future visual tokens, thereby skipping the VAE encoding. These mechanisms are synergistically integrated in a plug-and-play manner. Extensive experiments across multiple benchmarks and base models demonstrate that our method achieves a superior efficiency-performance trade-off, delivering up to a 4.00/3.94 speedup in LIBERO/LIBERO-PRO, 2.10 in RoboCasa, and consistent speedups on real-world tasks, while maintaining comparable success rates.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Model, Acceleration, Training-Free

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.60 over 5 reviews, above 70% of the others.
Reviewer_s7vcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

MARS proposes training-free acceleration at three levels: Transformer components, diffusion denoising steps, and consecutive action chunks. The method has strong practical motivation and is evaluated on LIBERO, LIBERO-Pro, RoboCasa, several model backbones, multiple GPUs, and a real Franka robot. It achieves up to approximately 4× acceleration while largely preserving task success rates.

Critical strengths

MARS is a training-free acceleration framework for diffusion-based World Action Models. It combines calibrated attention skipping, action-dependent selection between a direct one-step output and residual-reuse sampling, and reuse of predicted visual latents after a shallow VAE consistency check. The method is mainly a coordinated combination of existing heuristic acceleration techniques, but the system-level integration and empirical results are valuable

Critical weaknesses

The main issues are that the redundancy calibration is primarily conducted on LIBERO, and it is unclear how the skip sets transfer to other benchmarks, backbones, and real-world environments. The visual-token reuse verifier lacks sufficient analysis of false positives, false negatives, reuse rates, and the action errors caused by incorrect reuse. Stability of DiT residuals does not necessarily imply stability of the final control actions, so trajectory-level and long-horizon failure analyses are still needed.

The visual verifier is not separated from reuse frequency and scheduling. Tables 4-5 do not include periodic or random reuse at a matched rate, or unconditional reuse under the same consecutive-reuse cap. The overall pipeline is empirically promising, but its results do not establish whether this verifier excludes harmful predictions better than simple scheduling.

Questions for authors

The authors should clarify calibration transfer, verifier overhead, and the relationship between feature stability and control safety.

3 comments

Re W1 & Q1: Calibration Transfer

We thank you for highlighting the need to clarify the transferability of calibration.

Transfer across benchmarks and backbones. To test configuration generality, we conducted additional experiments using shared configurations. These shared-configuration experiments were conducted on RTX 4090 GPUs, with results averaged over three seeds.

Across benchmarks, we use one fixed configuration on LIBERO, LIBERO-PRO, and RoboCasa without benchmark-specific tuning.

BenchmarkBaseline Success (%)Ours Success (%)Baseline Latency (s)Ours Latency (s)Speedup
LIBERO98.0598.370.80310.24913.22×
LIBERO-PRO45.6345.570.78000.25763.03×
RoboCasa66.6966.221.30470.61372.13×

Across backbones, we also evaluate a shared configuration on Fast-WAM, LingBot-VA, and π₀.₅ without backbone-specific tuning.

BackboneBaseline Success (%)Ours Success (%)Baseline Latency (s)Ours Latency (s)Speedup
Fast-WAM97.5097.500.52950.13453.94×
LingBot-VA95.3396.677.65551.47685.18×
π₀.₅96.0096.500.08290.06711.24×

These results show that shared configurations can reduce latency while maintaining comparable success rates across the evaluated benchmarks and backbones. We have added a new section, Appendix J, to describe these experiments and present the results in Tables 12 and 13.

The real-robot task. For real-world deployment, we directly reuse the exact hyperparameter configuration from simulation without any modification, demonstrating seamless sim-to-real transferability with zero additional tuning effort.

Re W2 & Q2: Verifier Reliability and Overhead

Benefits beyond refresh frequency. To test whether verification improves reuse decisions, we compared it with random reuse, periodic refresh, and unconditional reuse. At nearly identical skip rates (56.25% vs. 56.44%), our method improves success over random reuse from 44.87% to 45.53%, while achieving a 3.94× speedup. Periodic refresh and unconditional reuse skip more often but achieve lower success rates. The comparison suggests that choosing when to reuse cached computation matters. We have added these results and the corresponding analysis to Section 4.3 in the main paper.

MethodGoal SR (%) / SpeedupObject SR (%) / SpeedupSpatial SR (%) / SpeedupLong SR (%) / SpeedupSuccess Rate ↑Speedup ↑Latency ↓Skip Rate
Cosmos Policy43.67 / 1.00×46.13 / 1.00×47.40 / 1.00×45.33 / 1.00×45.63%1.00×1.4044 s0.00%
Random Reuse42.87 / 4.09×45.80 / 4.11×46.80 / 4.09×44.00 / 4.20×44.87%4.14×0.3395 s56.44%
Periodic Refresh43.13 / 4.25×45.07 / 4.32×46.53 / 4.29×44.60 / 4.38×44.83%4.32×0.3249 s64.99%
Unconditional Reuse40.67 / 5.61×40.07 / 5.63×38.53 / 5.57×32.87 / 5.88×38.03%5.72×0.2457 s94.05%
Ours (Speed)44.20 / 3.91×45.67 / 3.86×47.33 / 3.96×44.93 / 3.99×45.53%3.94×0.3564 s56.25%

Verification overhead. Verification requires shallow VAE encoding and change-map comparison. Its average cost is 0.89 ms per policy query, accounting for 0.25% of the total query time. Verification adds little overhead while enabling substantial speedups through visual token reuse.

Verification ComponentTime per Query (ms)Share of Total Query Time (%)
Shallow VAE encoding0.590.17
Change-map comparison0.300.08
Total verification0.890.25

Re W3 & Q3: Feature Stability and Control Safety

We thank the reviewer for emphasizing the connection between feature stability and control performance.

Feature Stability and Control Safety. We directly compared residual reuse with full multi-step inference on 40 LIBERO-PRO tasks. We collected queries from complete rollouts of both policies. For each query, both methods used the same observation, initial noise, and sampling schedule. For each query, we compute the L2 differences between the accelerated and original actions. The table shows that residual reuse closely preserves the original action outputs, with small translational and rotational deviations and 99.14% gripper agreement.

MethodTranslation L2 ↓Rotation L2 ↓Gripper agreement (%) ↑
Original Cosmos Policy0.00000.0000100.00
Residual Reuse0.02960.005199.14

Trajectory-level evidence. As analyzed across 12,000 paired episodes in Appendix G, our method yields a gain in overall success rate, frequently recovering trajectories involving direct object transport (Figure 5). Conversely, it is slightly more sensitive to precision-critical interactions (Figure 6).

Long-horizon performance. Table 1 also shows that Ours (Acc.) maintains success rates close to the base model on LIBERO-Long and LIBERO-PRO-Long, with 3.19× and 3.11× speedups, respectively.

BenchmarkCosmos Policy SR (%)Ours (Acc.) SR (%)Speedup
LIBERO-Long96.6796.603.19×
LIBERO-PRO-Long45.3344.603.11×

We hope these results address the reviewer's concerns, and we would be grateful if the reviewer would consider raising the score.

Dear Reviewer,

As the discussion period is nearing its end, we would greatly appreciate your feedback on whether our rebuttal has addressed your concerns. We hope our clarifications can be considered in your final assessment.

Thank you for your time and consideration.

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

This paper proposes MARS, a training-free framework for accelerating WAMs through attention pruning, action-adaptive sampling, and reuse of predicted visual latents to reduce observation encoding costs. MARS achieves substantial inference speedups while largely preserving task success.

Critical strengths

Substantial inference acceleration with broadly preserved success across 3 benchmarks.

Profiling clarifies how bottlenecks shift from DiT computation to visual encoding.

Hardware and backbone evaluations, component ablations.

Critical weaknesses

1. The effect of first-step Transformer residual reuse on action accuracy remains unclear. A comparison with full multi-step inference under identical observations, noise, and step counts would help isolate its impact.

2. High overlap between change regions does not necessarily imply an accurate prediction. The authors needs to show how gate scores relate to action errors, especially after consecutive reuse.

3. Small motions may reflect slow demonstrations or hesitation, while large motions can still demand precision. Comparisons with fixed or random routing at the same budget would better justify this heuristic design.

Questions for authors

Refer to my comments in the Weaknesses.

5 comments

Re W1: Effect of Residual Reuse

We sincerely thank the reviewer for this suggestion. We directly compared residual reuse with full multi-step inference on 40 LIBERO-PRO tasks. We collected queries from complete rollouts of both policies. For each query, both methods used the same observation, initial noise, and sampling schedule. For each query, we compute the L2 differences between the accelerated and original actions. The table shows that residual reuse closely preserves the original action outputs, with small translational and rotational deviations and 99.14% gripper agreement.

MethodTranslation L2 ↓Rotation L2 ↓Gripper agreement (%) ↑
Original Cosmos Policy0.00000.0000100.00
Residual Reuse0.02960.005199.14

We further execute both policies from 120 matched initial states. These results indicate that residual reuse has minimal impact on action accuracy while preserving closed-loop task success in our evaluation.

SuiteInitial statesFull inference: successesResidual reuse: successes
Spatial3012/3012/30
Object3012/3012/30
Goal3010/3010/30
Long3016/3016/30
Overall12050/12050/120

Re W2: Gate Scores and Action Deviations

To test whether verification filters out harmful reuse, we evaluated 40 LIBERO-PRO tasks. For each query, we compared actions with and without visual token reuse in one view, holding all other settings fixed. Accepted queries have smaller mean L2 deviations in both translation and rotation than rejected queries.

ViewAccepted queriesAccepted translation L2 ↓Accepted rotation L2 ↓Rejected queriesRejected translation L2 ↓Rejected rotation L2 ↓
Primary11390.10740.01201740.15260.0194
Wrist3420.11510.01655630.15380.0227

Verification helps reject risky reuse, but mean action deviations still increase with consecutive reuse, so we limit repeated reuse to help prevent errors from accumulating.

Primary-view reuseTranslation L2 ↓Rotation L2 ↓Actual policy decision
First reuse0.09790.0113Reuse
Second reuse0.11890.0129Reuse
Third reuse0.13620.0148Refresh

To test whether verification helps beyond forced refreshes, we conducted additional experiments. At comparable reuse rates, our method improves success over random reuse and periodic refresh by 0.66 and 0.70 percentage points, respectively. Unconditional reuse achieves the highest skip rate but substantially lowers success. These results suggest that our verification provides benefits beyond refresh frequency alone. We have added these results and the corresponding analysis to Section 4.3 in the main paper.

MethodGoalObjectSpatialLongSuccess RateSpeedupLatencySkip Rate
Cosmos Policy43.67% / 1.00×46.13% / 1.00×47.40% / 1.00×45.33% / 1.00×45.63%1.00×1.4044 s0.00%
Random Reuse42.87% / 4.09×45.80% / 4.11×46.80% / 4.09×44.00% / 4.20×44.87%4.14×0.3395 s56.44%
Periodic Refresh43.13% / 4.25×45.07% / 4.32×46.53% / 4.29×44.60% / 4.38×44.83%4.32×0.3249 s64.99%
Unconditional Reuse40.67% / 5.61×40.07% / 5.63×38.53% / 5.57×32.87% / 5.88×38.03%5.72×0.2457 s94.05%
Ours (Speed)44.20% / 3.91×45.67% / 3.86×47.33% / 3.96×44.93% / 3.99×45.53%3.94×0.3564 s56.25%

Re W3: Routing Comparisons

We sincerely thank the reviewer for suggesting this controlled comparison. To directly evaluate the routing signal, We conducted additional experiments to evaluate the benefits of action-aware routing over random routing and fixed sampling.

At the same 3.94× speedup, action-aware routing achieves 45.53% success, compared with 45.22% for random routing, with gains on the Goal and Long suites. It also achieves higher average success than always-one-step sampling (45.13%) and always-residual-reuse sampling (45.15%). These results support translation magnitude as a useful routing signal.

Sampling configurationGoalObjectSpatialLongSuccess rateLatencySpeedup
Cosmos Policy43.67 / 1.00×46.13 / 1.00×47.40 / 1.00×45.33 / 1.00×45.63%1.4044s1.00×
Always one-step (τ_act=0.0)44.00 / 4.01×45.47 / 3.95×46.80 / 4.02×44.27 / 4.10×45.13%0.3483s4.03×
Always residual reuse (τ_act=+∞)43.60% / 3.89×45.53% / 3.82×47.87% / 3.89×43.60% / 3.98×45.15%0.3655s3.84×
Random Routing43.53 / 3.91×45.53 / 3.82×47.47 / 3.90×44.33 / 4.03×45.22%0.3568s3.94×
Action-aware routing (τ_act=0.4)44.20 / 3.91×45.67 / 3.86×47.33 / 3.96×44.93 / 3.99×45.53%0.3564s3.94×

We hope these clarifications resolve the concerns about our clinical goal and technical design. We respectfully ask the reviewer to reconsider the rating.

Dear Reviewer,

As the discussion period is nearing its end, we would greatly appreciate your feedback on whether our rebuttal has addressed your concerns. We hope our clarifications can be considered in your final assessment.

Thank you for your time and consideration.

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

This paper presents MARS, an acceleration framework that speeds up the inference of WAM. At three levels, the transformer level, diffusion step level, and action generation level, the paper proposes an integrated solution to accelerate WAM. The paper evaluates MARS on multiple benchmarks, including LIBERO, LIBERO-PRO, and RoboCasa, as well as several WAM and VLA backbones.

Critical strengths
  1. MARS can be a useful integration of existing approaches for accelerating WAM.
Critical weaknesses
  1. I would treat the paper more as a ‘systems paper’. Instead of proposing a new acceleration technique, it integrates existing approaches to form an integrated system to accelerate WAM. This is very meaningful in practice if done well; however, the paper fails to justify its contribution from a systems perspective. In particular, the manuscript claims MARS is a training-free and plug-and-play “framework”. The weakness is twofold:
    1. Ease of use: It is unclear how much model-specific tuning is needed. For an end user who would like to use MARS to accelerate their own model, how much human effort/involvement is required to achieve the same level of performance reported in the manuscript? One good example would be TensorRT.
    2. Generality: Following the previous point, although the authors evaluate several models and benchmarks, it is unclear whether the same configuration works across these settings or whether each model and dataset requires its own tuning.

The evidence for such a claim could be a software suite that does all the tuning itself or a demonstration that the same configuration generalizes to different models, tasks, or evaluation metrics.

Questions for authors

As a more systems-oriented work:

  1. How easy is it for an end user to accelerate their own model using MARS? How much human expertise/involvement is needed?
  2. How general is the configuration once a human has fixed it? Do I have to tune the configuration and hyperparameters used in MARS for each model, dataset, and evaluation setting to achieve something similar to what is reported in the paper?
3 comments

Re W1–W2 & Q1–Q2: Ease of Use and Configuration Generality

We thank the reviewer for the questions about deployment effort and configuration generality.

General design. MARS targets standard Transformer computation, diffusion sampling, and future visual prediction. Its acceleration rules can be applied to compatible models without retraining. Our implementation in the paper supports Cosmos Policy, Fast-WAM, LingBot-VA, and π₀.₅.

Configuration choices. In the paper experiments, we use the same configuration across task suites and all three seeds within each benchmark. No separate tuning is performed for individual tasks. Some settings, such as attention-skipping budgets and verification thresholds, differ across models or benchmarks to improve the efficiency–performance trade-off.

Additional experiments with shared configurations. To test configuration generality, we conducted additional experiments using shared configurations. These shared-configuration experiments were conducted on RTX 4090 GPUs, with results averaged over three seeds.

Across benchmarks, we use one fixed configuration on LIBERO, LIBERO-PRO, and RoboCasa without benchmark-specific tuning.

BenchmarkBaseline Success (%)Ours Success (%)Baseline Latency (s)Ours Latency (s)Speedup
LIBERO98.0598.370.80310.24913.22×
LIBERO-PRO45.6345.570.78000.25763.03×
RoboCasa66.6966.221.30470.61372.13×

Across backbones, we also evaluate a shared configuration on Fast-WAM, LingBot-VA, and π₀.₅ without backbone-specific tuning.

BackboneBaseline Success (%)Ours Success (%)Baseline Latency (s)Ours Latency (s)Speedup
Fast-WAM97.5097.500.52950.13453.94×
LingBot-VA95.3396.677.65551.47685.18×
π₀.₅96.0096.500.08290.06711.24×

These results show that shared configurations can reduce latency while maintaining comparable success rates across the evaluated benchmarks and backbones. We have added a new section, Appendix J, to describe these experiments and present the results in Tables 12 and 13.

Sensitivity to parameter choices. Our existing ablations also show similar performance across the tested settings. On LIBERO-PRO, Table 6 reports success rates of 45.22%–45.53% and speedups of 3.86×–3.94× across VAE probe depths. Appendix Table 10 reports success rates of 44.98%–45.53% and speedups of 3.84×–4.03× across routing thresholds from 0.0 to +∞.

LIBERO-PROGoalObjectSpatialLongSuccess Rate ↑Speedup ↑Latency ↓VAE Skip Rate
Cosmos Policy43.67 / 1.0046.13 / 1.0047.40 / 1.0045.33 / 1.0045.63%1.00×1.4044 s–
Layer-044.13 / 3.9145.07 / 3.8646.93 / 3.9144.73 / 3.9845.22%3.93×0.3574 s56.20%
Layer-144.20 / 3.9145.67 / 3.8647.33 / 3.9644.93 / 3.9945.53%3.94×0.3564 s56.25%
Layer-243.67 / 3.8545.27 / 3.7946.87 / 3.9245.33 / 3.9345.28%3.88×0.3621 s56.41%
Layer-343.73 / 3.8345.40 / 3.7847.20 / 3.9045.33 / 3.9145.42%3.86×0.3638 s56.44%
LIBERO-PROGoalObjectSpatialLongSuccess Rate ↑Speedup ↑Latency ↓
Cosmos Policy43.67 / 1.0046.13 / 1.0047.40 / 1.0045.33 / 1.0045.63%1.00×1.4044 s
Ours-1 (0.0)44.00 / 4.0145.47 / 3.9546.80 / 4.0244.27 / 4.1045.13%4.03×0.3483 s
Ours-2 (0.2)43.73 / 3.9345.60 / 3.8446.80 / 3.9543.80 / 4.0444.98%3.96×0.3550 s
Ours-3 (0.4)44.20 / 3.9145.67 / 3.8647.33 / 3.9644.93 / 3.9945.53%3.94×0.3564 s
Ours-4 (0.6)43.93 / 3.9045.47 / 3.8647.27 / 3.9544.40 / 3.9945.27%3.94×0.3567 s
Ours-5 (0.8)43.00 / 3.8745.60 / 3.8247.47 / 3.9144.27 / 3.9645.08%3.90×0.3601 s
Ours-6 (1.0)43.73 / 3.8545.53 / 3.7847.73 / 3.8843.67 / 3.9445.17%3.87×0.3625 s
Ours-7 (+∞)43.60 / 3.8945.53 / 3.8247.87 / 3.8943.60 / 3.9845.15%3.84×0.3655 s

Real-robot support. In our real-robot experiment, we directly reuse the simulation hyperparameters without further tuning. For supported models, users can start from our supplied configurations. We will release the code, supported configurations, and usage instructions upon publication.

Overall, our method supports plug-and-play acceleration across models and benchmarks without additional hyperparameter tuning. We thank again the reviewer for the valuable feedback, which has helped us clarify MARS’s usability and generality.

We hope these results address the reviewer's concerns, and we would be grateful if the reviewer would consider raising the score.

Dear Reviewer,

As the discussion period is nearing its end, we would greatly appreciate your feedback on whether our rebuttal has addressed your concerns. We hope our clarifications can be considered in your final assessment.

Thank you for your time and consideration.

We would be grateful if you could revisit our response and the additional experiments, and consider revising your score upward if you feel that they adequately address your concerns.

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

This paper proposes an acceleration framework for world action models that: (1) dynamically skips low-impact attention components; (2) adapts the number of inference steps based on action magnitude; and (3) reuses visual tokens to bypass part of the VAE encoding process when consecutive scenes are sufficiently similar.

Critical strengths

1、The results are supported by extensive experiments. 2、Attention skipping is more akin to pruning, but it works well, and the idea is promising.

Critical weaknesses

1、Dynamic step selection delivers the greatest speedup with the smallest performance drop, but the method itself seems somewhat simplistic. Using only translation magnitude as the action signal may not reliably capture motion complexity. Although the paper includes a threshold sensitivity analysis, the evaluation is limited; more robust motion measures could be explored. 2、The modules appear to rely on similar forms of pruning based on visual or feature similarity. This raises concerns about error accumulation over long action sequences. 3、The experimental results show a trade-off between speed and task performance rather than a clear overall improvement: each acceleration component appears to introduce some performance degradation.

Questions for authors

Explain the paper’s primary contribution and the technical challenge it addresses.

3 comments

We thank the reviewer for the constructive feedback and helpful questions. Below, we clarify our choice of routing signal, safeguards against error accumulation, efficiency–performance trade-off, and main contribution.

Re W1: Translation Magnitude as a Routing Signal

Design motivation. Translation magnitude serves as a lightweight proxy for motion granularity. Larger movements frequently occur during reaching and transport, whereas smaller movements are common during approach, grasping, and placement.

Preliminary exploration. As requested by the reviewer, we perform additional experiments with richer motion signals. Specifically, we add a rotation constraint to the original translation-based routing rule. We evaluate rotation thresholds of 0.02, 0.05, and 0.10, chosen around the median rotation magnitude. As shown in the table below, translation alone achieves the highest mean success rate among the evaluated acceleration configurations, reaching 45.53% on LIBERO-PRO averaged over three random seeds. Adding rotation provides no further improvement. We conjecture that rotation is weakly related to diffusion acceleration. These results support translation as a simple and effective routing signal in this setting. We have added a new section, Appendix I, to describe these experiments and present the results.

Routing signalTranslation thresholdRotation thresholdSpatial (%)Goal (%)Long (%)Object (%)Average success (%)
Cosmos Policy (base model)——47.4043.6745.3346.1345.63
Translation only0.4—47.3344.2044.9345.6745.53
Translation + rotation0.40.0247.0744.1344.0045.4745.17
Translation + rotation0.40.0547.0044.0043.8045.5345.08
Translation + rotation0.40.1047.3343.9344.7345.6745.42

Future extensions. We agree that translation alone may not fully capture control sensitivity. Our ablation study in Appendix Table 10 shows robust performance across the translation thresholds. Future work could explore rotation magnitude and changes in gripper state as richer routing signals.

Re W2: Error Accumulation

Limiting error accumulation. Our design skips the VAE encoding to reduce computation. However, we do not allow skipping to continue indefinitely. We limit how many times encoding can be skipped to reduce error accumulation over long action sequences. Specifically, as described in Section 3.3.3, visual tokens are reused for at most Nmax⁡N_{\max} consecutive policy queries, after which full encoding is required.

Long-horizon evidence. Table 1 in the main paper already reports the results on long-horizon tasks. Our method can largely preserve success rate while achieving large speedup. Specifically, on LIBERO-Long, Ours (Acc.) achieves 96.60% success compared with 96.67% for the base model, with a 3.19× speedup. On LIBERO-PRO-Long, it achieves 44.60% success compared with 45.33% for the base model, with a 3.11× speedup.

Re W3: Efficiency–Performance Trade-off

Accuracy-preserving configuration. Table 4 evaluates Ours (Speed), which prioritizes acceleration. Ours (Acc.) provides a more conservative operating point, achieving higher average success rates than Cosmos Policy on both benchmarks across three seeds. For ease of comparison, we summarize the key results in the tables below.

Method (LIBERO)GoalObjectSpatialLongAverage success (%) ↑Speedup ↑Latency (s) ↓FLOPs (%) ↓
Cosmos Policy98.20 / 1.0099.87 / 1.0097.47 / 1.0096.67 / 1.0098.051.00×1.4001100.00
Ours (Speed)96.73 / 3.8999.07 / 4.0897.60 / 3.8796.27 / 4.0797.424.00×0.349818.39
Ours (Acc.)98.47 / 3.0699.60 / 3.2997.80 / 3.1096.60 / 3.1998.123.17×0.441419.19
Method(LIBERO-PRO)GoalObjectSpatialLongAverage success (%) ↑Speedup ↑Latency (s) ↓FLOPs (%) ↓
Cosmos Policy43.67 / 1.0046.13 / 1.0047.40 / 1.0045.33 / 1.0045.631.00×1.4044100.00
Ours (Speed)44.20 / 3.9145.67 / 3.8647.33 / 3.9644.93 / 3.9945.533.94×0.356418.80
Ours (Acc.)43.67 / 3.0547.00 / 3.0847.60 / 3.0644.60 / 3.1145.723.08×0.455819.05

Practical objective. Our objective is to improve the efficiency–performance trade-off and provide configurable operating points for different deployment requirements. In addition, we have updated the Appendix (see Appendix H) to include the corresponding efficiency–performance rate trade-off curves.

Re Q1: Main Contribution and Technical Challenge

Core contribution.

  1. We systematically analyze WAM architectures that were proposed this year and identify substantial redundancy across Transformer computation, diffusion denoising, and visual encoding.
  2. We design simple, training-free methods to exploit these redundancies, with components that can be applied to other WAMs in a plug-and-play manner.
  3. Across multiple base models and benchmarks, we demonstrate substantial speedups with minimal loss in task performance.

Technical challenge. Identifying redundancy in emerging WAM architectures requires careful, systematic analysis. We then need to design training-free methods that exploit this redundancy and work consistently across benchmarks. Combining multiple acceleration components also risks compounding errors. Our design allows these components to work together, reducing computation while largely preserving task performance.

Dear Reviewer,

As the discussion period is nearing its end, we would greatly appreciate your feedback on whether our rebuttal has addressed your concerns. We hope our clarifications can be considered in your final assessment.

Thank you for your time and consideration.

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

This paper presents MARS, a training-free framework for accelerating diffusion-based World Action Models. Its three modules skip attention components using offline importance scores; route between one-step and residual-reuse sampling based on the first-step action prediction; and verify the previous policy query’s predicted future visual latents against shallow VAE features of the current observation, reusing them when verification passes. With Cosmos Policy as the primary backbone, MARS reduces LIBERO inference latency from 1.4001 s to 0.3498 s, while success decreases from 98.05% to 97.42%. Its conservative configuration achieves a 3.17× speedup with 98.12% success. The paper also reports 3.94× acceleration on LIBERO-PRO, 2.10× on RoboCasa, and evaluations across additional models, GPUs, and a real-world task.

Critical strengths
  1. MARS achieves about 4× acceleration on LIBERO using roughly 19% of the original FLOPs. Evaluations across benchmarks, GPUs, seeds, and a real-world task support its practical value.

  2. Figure 4 identifies VAE encoding as the next bottleneck after reducing DiT computation. Visual reuse raises speedup from 3.06× to 3.94×, supporting optimization across both stages.

Critical weaknesses
  1. Table 9 reports 45.13% success for τ_act=0 versus 45.53% for τ_act=0.4. Table 4 ablates the entire denoising module, leaving routing’s advantage over fixed sampling unclear.

  2. MARS reuses predictions after denoising acceleration, but IoU checks only change locations. Experiments neither test rejection of incorrect states nor isolate verification’s benefit beyond forced refreshes.

  3. Each baseline has only one reported operating point, without a rationale for its acceleration strength or a trade-off curve (Tables 1–2). The advantage over tuned baselines remains unclear.

Questions for authors
  1. With the other modules fixed, on which types of manipulation does action-aware routing provide measurable success-rate gains over always using one-step or residual-reuse sampling?

  2. At the same refresh frequency, can the verifier reject predictions with matching change regions but incorrect object states, and reduce subsequent control errors?

  3. After tuning the caching strength of DiCache and C³ache to match success rates, does MARS retain a speed advantage? Could the authors provide the corresponding curves of speed versus success rate?

6 comments

Re W1 & Q1: Benefit of action-aware routing

We sincerely thank you for highlighting this important issue. Following your suggestions step by step, we have conducted additional experiments to evaluate (1) the advantage of routing over fixed sampling and (2) the task-specific benefits of routing across manipulation tasks.

Advantages of Routing over Fixed Sampling

Sampling configurationGoalObjectSpatialLongSuccess rateLatencySpeedup
Cosmos Policy43.67% / 1.00×46.13% / 1.00×47.40% / 1.00×45.33% / 1.00×45.63%1.4044s1.00×
Always one-step (τ_act=0.0)44.00% / 4.01×45.47% / 3.95×46.80% / 4.02×44.27% / 4.10×45.13%0.3483s4.03×
Action-aware routing (τ_act=0.4)44.20% / 3.91×45.67% / 3.86×47.33% / 3.96×44.93% / 3.99×45.53%0.3564s3.94×
Mostly residual reuse (τ_act=1.0)43.73% / 3.85×45.53% / 3.78×47.73% / 3.88×43.67% / 3.94×45.17%0.3625s3.87×
Always residual reuse (τ_act=+∞)43.60% / 3.89×45.53% / 3.82×47.87% / 3.89×43.60% / 3.98×45.15%0.3655s3.84×

Table 10 in the appendix already provides initial evidence for this trade-off. We conducted additional experiments to further evaluate the benefits of routing over fixed sampling. Compared with always-one-step sampling, our routing improves success rates across all four suites. Compared with always-multi-step sampling, it achieves a higher average success rate with lower latency. These results show that our routing offers a better efficiency–performance trade-off than fixed sampling. We have added these results and the corresponding analysis to Appendix F.

Benefits of Routing across Manipulation Tasks

Action CategoryRepresentative TasksAlways One-Step SRAlways Residual Reuse SRAction-Aware Routing SR
Spatial Manipulationtransfer bowl to plate38.00%36.00%39.33%
place book in caddy31.33%30.00%32.00%
Object Interactionplace mug and close microwave27.33%23.33%28.67%
place pudding in basket43.33%42.67%44.00%
Goal-Directed Placementplace bottle atop cabinet36.00%33.33%38.67%
place bowl atop cabinet44.67%40.00%45.33%
Long-Horizon Multi-stepplace two moka pots54.00%50.00%54.67%
place mug and pudding57.33%56.67%59.33%

To show which tasks benefit from routing with the other modules fixed, we conducted additional experiments on representative manipulation tasks. Routing achieves higher success rates than both fixed sampling strategies on all eight selected tasks, covering spatial alignment, object interaction, placement, and multi-step manipulation.

Re W2 & Q2: Effectiveness of verification

We thank you for highlighting the distinction between change consistency and state accuracy.

Role as a Consistency Check. We regard the IoU checks as a consistency check for speculative reuse, rather than a certificate of prediction correctness. To prevent errors from accumulating over long horizons, we limit the maximum number of consecutive reuses, even when all checks pass.

The Verifier’s Advantage. To test whether verification helps beyond forced refreshes, we conducted additional experiments. At comparable reuse rates, our method improves success over random reuse and periodic refresh by 0.66 and 0.70 percentage points, respectively. The comparison with random reuse suggests that choosing when to reuse cached computation matters even at nearly identical skip rates, while periodic refresh incurs a larger success-rate drop from the original policy. Unconditional reuse achieves the highest skip rate but substantially lowers success. These results suggest that our verification provides benefits beyond refresh frequency alone. We have added these results and the corresponding analysis to Section 4.3 in the main paper.

MethodGoalObjectSpatialLongSuccess RateSpeedupLatencySkip Rate
Cosmos Policy43.67% / 1.00×46.13% / 1.00×47.40% / 1.00×45.33% / 1.00×45.63%1.00×1.4044 s0.00%
Random Reuse42.87% / 4.09×45.80% / 4.11×46.80% / 4.09×44.00% / 4.20×44.87%4.14×0.3395 s56.44%
Periodic Refresh43.13% / 4.25×45.07% / 4.32×46.53% / 4.29×44.60% / 4.38×44.83%4.32×0.3249 s64.99%
Unconditional Reuse40.67% / 5.61×40.07% / 5.63×38.53% / 5.57×32.87% / 5.88×38.03%5.72×0.2457 s94.05%
Ours (Speed)44.20% / 3.91×45.67% / 3.86×47.33% / 3.96×44.93% / 3.99×45.53%3.94×0.3564 s56.25%

Rejection of Incorrect States. To test whether verification filters out harmful reuse, we evaluated 40 LIBERO-PRO tasks. For each query, we compared actions with and without visual token reuse, holding all other settings fixed. Accepted queries have smaller mean L2 deviations in both translation and rotation than rejected queries.

ViewAccepted queriesAccepted translation L2 ↓Accepted rotation L2 ↓Rejected queriesRejected translation L2 ↓Rejected rotation L2 ↓
Primary11390.10740.01201740.15260.0194
Wrist3420.11510.01655630.15380.0227

Mean action deviations increase with consecutive reuse, even among queries passing the check. We therefore limit the maximum number of consecutive reuses. Verification helps reject risky reuse, while the reuse limit helps prevent errors from accumulating.

Primary-view reuseTranslation L2 ↓Rotation L2 ↓Actual policy decision
First reuse0.09790.0113Reuse
Second reuse0.11890.0129Reuse
Third reuse0.13620.0148Refresh

Re W3 & Q3: Comparison with tuned baselines

We sincerely thank you for this constructive suggestion. We fully agree that evaluating baselines across multiple operating points provides a much more rigorous and comprehensive view of the efficiency–performance trade-off than a single default setting.

Caching Mechanisms. DiCache reuses features within an action chunk, using a feature-change threshold to control cache refreshes. C³ache reuses features across consecutive action chunks, using a refresh interval to control caching strength. Both controls balance computational savings and approximation error.

Multi-Point Evaluation. We further evaluated both baselines at multiple caching strengths on LIBERO and LIBERO-PRO. The table below reports success rates and end-to-end latency. We also added the corresponding efficiency–performance trade-off curves to Appendix H.

Method(LIBERO)GoalObjectSpatialLongSuccess RateSpeedupLatency
Cosmos Policy98.20% / 1.00×99.87% / 1.00×97.47% / 1.00×96.67% / 1.00×98.05%1.00×1.4001 s
+ DiCache (Setting 1)98.07% / 1.70×99.87% / 1.66×97.80% / 1.67×97.27% / 1.72×98.25%1.69×0.8273 s
+ DiCache (Setting 2)97.87% / 1.46×99.80% / 1.46×97.07% / 1.46×97.33% / 1.48×98.02%1.47×0.9544 s
+ DiCache (Setting 3)98.07% / 1.89×99.93% / 1.89×97.60% / 1.89×97.47% / 1.92×98.27%1.90×0.7359 s
+ C³Cache* (Setting 1)97.67% / 1.83×99.33% / 1.86×97.73% / 1.85×96.47% / 1.95×97.80%1.89×0.7410 s
+ C³Cache* (Setting 2)97.80% / 2.19×99.00% / 2.16×97.53% / 2.25×94.47% / 2.30×97.20%2.24×0.6256 s
+ C³Cache* (Setting 3)98.20% / 2.31×98.80% / 2.40×97.47% / 2.29×92.93% / 2.42×96.85%2.37×0.5907 s
+ X-Cache*97.87% / 1.11×99.67% / 1.17×97.47% / 1.12×97.00% / 1.21×98.00%1.17×1.2017 s
+ Ours (Speed)96.73% / 3.89×99.07% / 4.08×97.60% / 3.87×96.27% / 4.07×97.42%4.00×0.3498 s
+ Ours (Acc.)98.47% / 3.06×99.60% / 3.29×97.80% / 3.10×96.60% / 3.19×98.12%3.17×0.4414 s
Method(LIBERO-PRO)GoalObjectSpatialLongSuccess RateSpeedupLatency
Cosmos Policy43.67% / 1.00×46.13% / 1.00×47.40% / 1.00×45.33% / 1.00×45.63%1.00×1.4044 s
+ DiCache (Setting 1)44.13% / 1.67×46.27% / 1.66×47.73% / 1.63×45.07% / 1.73×45.80%1.68×0.8347 s
+ DiCache (Setting 2)43.73% / 1.45×46.33% / 1.46×47.67% / 1.46×45.07% / 1.49×45.70%1.47×0.9568 s
+ DiCache (Setting 3)44.13% / 1.89×46.47% / 1.89×47.47% / 1.89×44.87% / 1.92×45.73%1.90×0.7380 s
+ C³Cache* (Setting 1)43.13% / 1.91×46.00% / 1.90×48.27% / 1.88×43.73% / 1.96×45.28%1.92×0.7313 s
+ C³Cache* (Setting 2)42.80% / 2.22×45.53% / 2.20×47.53% / 2.28×42.60% / 2.30×44.62%2.26×0.6222 s
+ C³Cache* (Setting 3)42.40% / 2.39×45.27% / 2.41×47.00% / 2.29×41.13% / 2.50×43.95%2.42×0.5801 s
+ X-Cache*43.73% / 1.17×46.20% / 1.22×47.80% / 1.17×44.33% / 1.26×45.52%1.22×1.1549 s
+ Ours (Speed)44.20% / 3.91×45.67% / 3.86×47.33% / 3.96×44.93% / 3.99×45.53%3.94×0.3564 s
+ Ours (Acc.)43.67% / 3.05×47.00% / 3.08×47.60% / 3.06×44.60% / 3.11×45.72%3.08×0.4558 s

Our Method’s Advantage. Our method addresses multi-level costs in Transformer computation, diffusion denoising, and visual encoding. Even after tuning DiCache and C³ache, MARS retains lower end-to-end latency at comparable success rates. Compared with DiCache (Setting 3), MARS (Acc.) reduces latency by 40.0% and 38.2% on LIBERO and LIBERO-PRO, respectively, while maintaining comparable success rates. Compared with C³Cache* (Setting 3), MARS (Speed) reduces latency by 40.8% and 38.6% while improving success rates, respectively. This demonstrates that our method achieves a superior speed–performance trade-off compared with the baselines.

If these answers have addressed the reviewer's concerns, we hope that the reviewer could consider raising the score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.