acceptodds
Under review as a conference paper at ICLR 2027

Accelerating World Action Models via Action-Aware Asynchronous Caching

Abstract

Multi-view world action models enable robots to predict future interactions from multiple camera views, while introduce substantial inference cost by multi-view processing. Although training-free caching has shown promise in accelerating diffusion and video world models, existing methods largely treat multi-view representations as a unified sequence and fail to exploit the uneven dynamics across different camera views, limiting their efficiency under aggressive acceleration. In this work, we propose A3Cache: Action-Aware Asynchronous Caching, a training-free acceleration framework for multi-view embodied world action models, which transforms multi-view caching into a asynchronous computation paradigm by adapting computation according to their update demands. Specifically, we introduce view asynchronous computing paradigm, which independently evaluates cache validity for different camera views and computes only the active views while predicting the stable ones by cache. We further introduce world-action contribution-aware token delaying paradigm within active views, which jointly considers token cache deviation and changes in visual-to-action contribution to prioritize urgent local computation while delaying less urgent tokens. Extensive experiments on multi-view robotic action demonstrate that A3Cache achieves a better efficiency-performance trade-off than existing methods, delivering substantially higher acceleration ratio with minimal performance degradation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Model, Inference Acceleration, Feature Caching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_8RAdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes A3Cache, a training-free acceleration framework for multi-view world action models. The method introduces view-level asynchronous caching and action-aware token selection to reduce redundant computation. Experiments on RoboTwin 2.0 demonstrate substantial inference acceleration while largely preserving robotic manipulation performance.

Critical strengths
  1. The paper is well-written and easy to follow, with clear motivation and technical explanations.

  2. The proposed framework provides practical technical insights into efficient multi-view computation. The experimental results are promising, achieving substantial acceleration with minimal degradation in task success rates.

Critical weaknesses
  1. The proposed view cache validation and action-aware token selection rely on heuristic criteria, which may limit their generalizability across complex video dynamics. In particular, it is unclear how robust these criteria are under dynamic object interactions or inconsistent cross-view representations. Although the ablations support their effectiveness, further analysis of these limitations would strengthen the contribution.

  2. The evaluation is limited to a single backbone (Motus) in the RoboTwin simulation environment. It is unclear whether the proposed method can generalize to other world action models or real-world robotic settings with more complex dynamics.

Questions for authors
  1. Could the authors clarify whether the proposed acceleration and task performance can be maintained on physical robots?

Overall, the reviewer slightly leans toward a positive recommendation, while this topic is not entirely within my core expertise. I would like to consider other reviewers' assessments during the discussion before making a final decision.

4 comments
AuthorsOpenReview ↗

Response Part 1

Dear Reviewer 8RAd,

Thank you for highlighting the importance of robustness under complex interactions and validation on physical robots. Your comments help us examine how our caching criteria behave in more demanding settings. Below, we discuss the bimanual-task results and physical-robot evaluation in response to your concerns.

Q1: Evaluation on physical robots

Thank you for asking for evidence from physical deployment and a backbone beyond Motus.

Our test benchmark. A3Cache accelerates the inference process of WAMs independently of the simulation environment, allowing its computational benefits to transfer to physical deployment. RoboTwin 2.0 employs digital twins and domain randomization to improve sim-to-real transfer [1], providing a relevant benchmark for robotic manipulation. We further conduct real-robot experiments to verify whether the acceleration and task performance can be maintained in physical environments.

Real-robot evaluation. As shown in Table, we evaluate A3Cache using OpenWAM on the physical Stack Bowls task from RoboChallenge Table30, reporting task success rate (SR), progress score, and computation cost. Across 10 real-robot trials, A3Cache achieves 80% SR (8/10) and 82/100 PS, while reducing computation cost and maintaining a performance-acceleration trade-off comparable to the simulation results.

MethodSR (%)progress scoreTFLOPsRatio of computation
Original80%884.161.00×
A3Cache80%821.024.08×

[1] Chen et al. RoboTwin 2.0: A Scalable Data Generator and Benchmark with Strong Domain Randomization for Robust Bimanual Robotic Manipulation. ICML 2026.

AuthorsOpenReview ↗

Response Part 2

W1: Performance on bimanual tasks

Thank you for drawing attention to dynamic interactions and the limits of the cache criteria.

Coordinated manipulation. We evaluate the subset of RoboTwin tasks requiring both arms, where coordinated actions can change observations in several views at once. Using the uncached model on this same subset as the reference, A3Cache retains 97.4% of average success, compared with 94.1% for SeaCache and 95.7% for CA-WorldCache. Its retention is slightly lower than the 98.1% on the full benchmark, indicating that coordinated interactions are more demanding, while A3Cache still achieves the best success and efficiency among the compared caching methods.

MethodClean SR (%)Random SR (%)Avg. SR (%)ΔSR (pp, retention)Avg. time (s)TFLOPsRatio of computation
Motus without caching85.3869.7177.550.00 (100.0%)1.313.001.00×
SeaCache81.3364.5872.96-4.59 (94.1%)0.891.272.36×
CA-WorldCache82.9265.4674.19-3.36 (95.7%)0.591.102.73×
A3Cache83.2967.7975.54-2.01 (97.4%)0.380.714.23×

Handling changes across views. Each view has its own accumulated deviation estimate, allowing a changing view to trigger an update without forcing all other views to update simultaneously. Within an active view, the combined visual and action-contribution signals prioritize tokens whose cached states are becoming less suitable for action prediction. Cached representations remain available for deferred tokens, and the final denoising step uses full computation. These mechanisms help limit the effect of delayed updates, but they do not guarantee robustness to inconsistent observations across cameras. The bimanual results provide evidence for coordinated interactions rather than a direct test of all forms of cross-view inconsistency.

*The list of tasks with coordinated actions we select: Blocks Ranking Size, Dump Bin Bigbin, Grab Roller, Handover Block, Handover Mic, Hanging Mug, Lift Pot, Pick Diverse Bottles, Pick Dual Bottles, Place Bread Basket, Place Burger Fries, Place Can Basket, Place Cans Plasticbox, Place Dual Shoes, Place Object Basket, Put Bottles Dustbin, Put Object Cabinet, Scan Object, Stack Blocks Three, Stack Blocks Two, Blocks Ranking RGB, Place Bread Skillet, Stack Bowls Three, Stack Bowls Two.

Thank you for the detailed reponses and additional experiments. Most of my concerns have been enoughly addressed. I will keep my current score. But as this topic is not entirely within my expertise, I will keep discuss during the AC-reviewer discussion.

AuthorsOpenReview ↗

Thanks for Your Response

Dear Reviewer 8RAd,

Thank you very much for your thoughtful response and for acknowledging that most of your concerns have been addressed. We greatly appreciate your suggestions regarding robustness under complex interactions, generalization beyond Motus, and physical-robot evaluation. Your feedback has been very helpful in strengthening our evaluation and understanding of A3Cache under more challenging settings. Thank you again for your time and valuable input!

Best regards,
Authors

Reviewer_v56gOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes A3Cache, a training-free inference acceleration method for multi-view world-action models. Its core is two-level selection: 1. determine which camera views need updating based on the cumulative deviation between each view's features and the cached prediction. 2. combining changes in visual features and changes in the attention contribution to the action branch, select the tokens within a view that need re-computation, while reusing the cache for the rest. In experiments, based on Motus and tested on RoboTwin 2.0, the average task success rate dropped from 81.90% to 80.35%, and the reported average inference time dropped from 1.31 s to 0.38 s, reflecting a good speed-performance trade-off.

Critical strengths
  1. The design is reasonable and practically valuable. Allocating computation separately by view and by local token can exploit the uneven change of views. Incorporating action-related signals is also closer to the control task than relying only on visual change.
  2. The reported actual speedup gain is substantial. Measured by latency, it is about a 3.45× speedup, while the average success rate drops by only 1.55 percentage points.
  3. Ablations were provided for view-level selection, token-level selection, and the two scoring signals, rather than only showing the quality of generated frames.
Critical weaknesses
  1. Novelty not sufficiently demonstrated. SeaCache already uses feature differences to decide whether to reuse the cache, ToCa has proposed token-level feature caching. The new design in this paper mainly consists of organizing the cache by camera view and introducing action-related signals to select tokens, so overall it looks like a combined improvement over existing techniques. The experiments also do not sufficiently distinguish how much of the gain comes from the new design.
  2. This paper reports the Motus baseline Clean/Random SR as 87.56%/76.24%. Motus paper reports 88.66%/87.02%. This difference needs to be clarified。
  3. The experiments are validated only on Motus and RoboTwin 2.0. In Table 3, adding action-related signals improves over using only visual change by 0.83 pp, which may be experimental fluctuation.
Questions for authors
  1. Please compare on the same caching backbone: no selection added, view selection only, token selection only, and both added together, and to report the success rate–actual latency curve.
  2. What does "Original (Aligned)" refer to? Why are the SR on RoboTwin2.0 in this paper different from those in Motus paper?
  3. For the action-related scoring, is the 0.83 pp improvement over using only visual change stable across repeated experiments?
  4. The 4.23× should be the ratio of computation, right? The actual latency ratio is only 3.45×.
5 comments
AuthorsOpenReview ↗

Response Part 1

Dear Reviewer v56g,

Thank you for carefully examining our method and evaluation. We appreciate your suggestions on isolating the contributions of view and token selection, checking the stability of action-aware scoring, and clarifying the baseline and efficiency metrics. Below, we present the corresponding experiments and clarify these aspects of our approach.

Q1: View and token selection on the same caching backbone

Thank you for suggesting a controlled ablation to distinguish the contributions of view and token selection.

Comparison under a matched computation budget. We evaluate no selection, view selection only, token selection only, and both together on the same caching backbone. For each configuration, we adjust the caching parameters to obtain approximately the same computation budget, with costs ranging from 0.71 to 0.77 TFLOPs. This comparison tests how effectively each configuration uses a comparable budget to preserve task performance. The table also reports the measured latency of each configuration, providing a practical efficiency reference for these operating points.

Complementary contributions of view and token selection. At this comparable budget, view selection improves average success from 75.51% to 78.82%, and token selection alone reaches 77.06%. Combining both achieves 80.35%, exceeding either component alone on both Clean and Random. View selection allocates updates across camera views according to their different dynamics, while token selection prioritizes updates within active views using visual and action-related signals. The gains therefore support the complementary value of these two ways of allocating computation, rather than a progressive reduction in the computation budget across configurations.

MethodClean SR (%)Random SR (%)Avg. SR (%)ΔSR (pp, retention)Avg. time (s)TFLOPsRatio of computation
Original(Aligned)87.5676.2481.900.00 (100.0%)1.313.001.00×
No selection81.0869.9475.51-6.39 (92.2%)0.450.763.95×
View selection only84.2673.3878.82-3.08 (96.2%)0.450.773.90×
Token selection only82.9271.2077.06-4.84 (94.1%)0.410.754.00×
Full A3Cache85.9474.7680.35-1.55 (98.1%)0.380.714.23×
AuthorsOpenReview ↗

Response Part 2

Q2: Original (Aligned) and the evaluation setup

Thank you for asking us to clarify the baseline and the source of the reported success rates.

The Original (Aligned) refers to the unaccelerated Motus model evaluated locally using the official open-source code and pretrained weights. To ensure a fair comparison, we evaluate the original Motus baseline and all caching methods under the same experimental environment and settings. All results in the comparison tables are obtained from our unified local evaluation, and all changes in success rate are calculated relative to this locally reproduced Motus baseline.

Q3: Stability of the gain from action-aware scoring

Thank you for highlighting the need to check the stability of a relatively small improvement. We repeat the visual-only scoring experiment under the same evaluation protocol and compare it with Full A3Cache. The repeated results with different random seeds consistently show the combined visual and action-related score, supporting that the benefit is not specific to the original visual-only run. The original comparison improves average success from 79.52% to 80.35% at the same 0.71 TFLOPs.

Visual deviation alone cannot distinguish visual changes that affect action prediction from those with limited impact. By measuring changes in visual-to-action contributions, A3Cache prioritizes action-relevant updates while delaying less urgent tokens. Meanwhile, accumulated visual deviation increases the priority of delayed tokens in subsequent denoising steps, preventing excessive error accumulation. The two complementary signals thus enable more effective token selection under the same computation budget, better preserving action prediction accuracy.

MethodClean SR (%)Random SR (%)Avg. SR (%)ΔSR (pp, retention)Avg. time (s)TFLOPsRatio of computation
Original(Aligned)87.5676.2481.900.00 (100.0%)1.313.001.00×
Full A3Cache85.9474.7680.35-1.55 (98.1%)0.380.714.23×
w/o action contribution deviation (original run)85.0673.9879.52-2.38 (97.1%)0.370.714.23×
w/o action contribution deviation (repeat 2)85.1273.9079.51-2.39 (97.1%)0.370.714.23×
w/o action contribution deviation (repeat 3)85.0473.8679.45-2.45 (97.0%)0.370.714.23×
AuthorsOpenReview ↗

Response Part 3

Q4: Ratio of computation and actual latency ratio

Thank you for pointing out this distinction. 4.23× is the ratio of computation, calculated as 3.00/0.71 TFLOPs. To minimize the influence of hardware-dependent factors, we report FLOPs-based computation ratios, following existing caching methods such as TeaCache [1], SeaCache [2], and DPCache [3]. We also report inference latency measured on our hardware platform as a practical reference, showing a consistent acceleration trend with the FLOPs-based computation ratios.

References

[1] Liu et al. Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model. CVPR 2025.

[2] Chung et al. SeaCache: Spectral-Evolution-Aware Cache for Accelerating Diffusion Models. CVPR 2026.

[3] Cui et al. Denoising as Path Planning: Training-Free Acceleration of Diffusion Models with DPCache. CVPR 2026.

W1: View scheduling and action-aware token selection

Thank you for asking us to clarify the contribution relative to existing feature and token caching methods.

Multi-view WAMs exhibit heterogeneous dynamics across views and varying visual contributions to action prediction. Existing methods, such as SeaCache, use global feature differences without exploiting view-specific redundancy. While token-level methods, such as ToCa, prioritize visual generation fidelity, A3Cache exploits the action-oriented objective of WAMs to defer visually changing tokens with stable contributions to action prediction to subsequent denoising steps. Since WAM inference targets action generation, visual updates with stable action contributions can be deferred to subsequent denoising steps without compromising action prediction.

A3Cache therefore introduces hierarchical asynchronous computation, independently scheduling view updates and prioritizing tokens based on visual deviations and changes in action contributions. This design reduces redundant computation while preserving action-relevant information, improving the efficiency-performance trade-off. The results in ablation Table further validate the effectiveness of both components and demonstrate their complementary contributions to the overall performance improvement.

MethodClean SR (%)Random SR (%)Avg. SR (%)ΔSR (pp, retention)Avg. time (s)TFLOPsRatio of computation
Original(Aligned)87.5676.2481.900.00 (100.0%)1.313.001.00×
No selection81.0869.9475.51-6.39 (92.2%)0.450.763.95×
View selection only84.2673.3878.82-3.08 (96.2%)0.450.773.90×
Token selection only82.9271.2077.06-4.84 (94.1%)0.410.754.00×
Full A3Cache85.9474.7680.35-1.55 (98.1%)0.380.714.23×
Reviewer_Si3cOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a training-free caching method for multi-view world action models. Tokens can be cached and reused across denoising iterations, reducing the amount of computation required during diffusion model inference. For multi-view inputs, however, tokens corresponding to multiple modes of input (action, multiple views) evolve at different rates. To address this, the authors propose (1) view asynchronous computing: each camera view gets its own cache, and (2) action-contribution aware token delay: recomputing only the visual tokens that influence action tokens the most. Using this approach, the authors demonstrate speedups on worldaction model inference.

Critical strengths
  1. Strong speed–accuracy trade-off
  2. Well motivated approximation method
Critical weaknesses
  1. Evaluation over only one model
  2. No comparison with few-step or distilled models
  3. Limited novelty
Questions for authors

Please address the following comments:

  1. All experiments use one world action model (Motus) on one simulated benchmark (RoboTwin 2.0), with no real-robot experiments. The method is proposed for multi-view world action models as a class (Genie Envisioner, EnerVerse, Ctrl-World) but these are not evaluated. With this model alone evaluated, it is hard to say whether the properties would transfer to other models. The authors must evaluate atleast another world action model using a3cache.

  2. Method does not work on distilled variants on Motus. Caching across denoising steps relies on redundancy between consecutive steps. Distilled models requiring fewer denoising steps sees less benefit. Does A3Cache still yield a meaningful speedup on a step-distilled or few-step model?

  3. The primary novelty here seems to be two things: 1) action-contribution aware token delay: recomputing the visual tokens that influence action tokens the most, and 2) view asynchronous computing: each camera view gets its own cache. In Table 1, prior caching works are applied to the entire multi-view sequence, so A3Cache’s advantage may come from per-view scheduling. This could be applied to any existing method. Could the authors include baselines that run an existing caching method (e.g., TeaCache or CA-WorldCache) independently on each camera view? This would show A3Cache's performance improvement due to action-contribution-aware token delay alone.

5 comments
AuthorsOpenReview ↗

Response Part 1

Dear Reviewer Si3c,

Thank you for your thoughtful comments on generalization, few-step inference, and the contribution of per-view scheduling. Your suggestions help us examine the scope of A3Cache and better distinguish the benefits of its components. Below, we address each concern with additional comparisons and explanations.

Q1: Generalization to other models

Thank you for pointing out the need to test whether the benefits extend beyond Motus. We evaluate A3Cache on OpenWAM and compare it with representative existing methods D2Cache and CA-WorldCache using the same checkpoint and evaluation protocol. This additional backbone tests whether our view and token scheduling remains useful when the underlying model changes, rather than relying only on the redundancy observed in Motus.

MethodClean SR (%)Random SR (%)Avg. SR (%)ΔSR (pp, retention)Avg. time (s)TFLOPsRatio of computation
OpenWAM without caching92.7891.9492.360.00 (100.0%)3.754.161.00×
D2Cache91.7290.2090.96-1.40 (98.5%)1.341.662.50×
CA-WorldCache91.3089.1690.23-2.13 (97.7%)1.471.792.32×
A3Cache92.4891.0291.75-0.61 (99.3%)1.131.024.08×

As the Table shows, A3Cache achieves the highest Clean and Random success rates among the compared caching methods, retaining almost full of the uncached average success rate. It also reduces computation by 4.07× and inference time from 3.75 s to 1.13 s. These results support the applicability of A3Cache beyond Motus, with improvements in both task performance and efficiency over the compared caching methods.

AuthorsOpenReview ↗

Response Part 2

Q2: Performance on few-step model

Thank you for raising the question of whether caching remains useful when fewer denoising steps are available. We agree that reducing the number of steps leaves fewer opportunities for feature reuse. However, redundancy need not be uniform across camera views or tokens within the same step. Unlike existing methods that rely on synchronous cache updates, A3Cache makes separate update decisions for each view and selectively recomputes visual tokens within active views, allowing it to exploit the remaining redundancy at a finer level.

We evaluate Motus with five denoising steps. As the Table shows, A3Cache retains 97.1% of this baseline's average success rate while reducing computation by 2.08× and latency from 0.64 s to 0.26 s. It also achieves higher success rates on both Clean and Random than SeaCache and CA-WorldCache. Thus, the benefit remains meaningful in five-step inference, although the available computation reduction is smaller than in the default ten-step setting.

MethodClean SR (%)Random SR (%)Avg. SR (%)ΔSR (pp, retention)Avg. time (s)TFLOPsRatio of computation
Motus without caching (5 steps)82.7468.1675.450.00 (100.0%)0.641.501.00×
SeaCache78.8062.4870.64-4.81 (93.6%)0.431.031.46×
CA-WorldCache79.6663.9471.80-3.65 (95.2%)0.390.911.65×
A3Cache80.5465.8273.18-2.27 (97.0%)0.260.722.08×
AuthorsOpenReview ↗

Response Part 3

Q3: Effect of independent per-view caching

Thank you for suggesting this comparison to separate the benefit of per-view scheduling from token selection.

Applying per-view scheduling to existing methods. We apply each baseline's cache rule independently to each camera view. SeaCache improves from 78.56% to 79.88% average success, and CA-WorldCache improves from 79.12% to 80.00%, while both also reduce computation and latency. We therefore obtain improvements on both existing methods, showing that independent view scheduling contributes beyond their original global caching rules.

MethodClean SR (%)Random SR (%)Avg. SR (%)ΔSR (pp, retention)Avg. time (s)TFLOPsRatio of computation
Original(Aligned)87.5676.2481.900.00 (100.0%)1.313.001.00×
SeaCache (global)84.2072.9278.56-3.34 (95.9%)0.891.272.36×
SeaCache (per view)86.0873.6879.88-2.02 (97.5%)0.851.242.42×
CA-WorldCache (global)84.8873.3679.12-2.78 (96.6%)0.591.102.73×
CA-WorldCache (per view)85.9274.0880.00-1.90 (97.7%)0.581.072.80×

Contribution beyond per-view scheduling. The shared-backbone ablation provides a complementary comparison between view selection alone and the full method. Adding token selection increases average success from 78.82% to 80.35% by the same computation cost. Together, these comparisons distinguish the benefit of scheduling views independently from the additional benefit of selecting tokens within active views. The latter comparison measures the complete token-selection component, whose visual and action-related signals are separately examined in the token-scoring ablation.

MethodClean SR (%)Random SR (%)Avg. SR (%)ΔSR (pp, retention)Avg. time (s)TFLOPsRatio of computation
Original(Aligned)87.5676.2481.900.00 (100.0%)1.313.001.00×
View selection only84.2673.3878.82-3.08 (96.2%)0.450.773.90×
Token selection only82.9271.2077.06-4.84 (94.1%)0.410.754.00×
Full A3Cache85.9474.7680.35-1.55 (98.1%)0.380.714.23×

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Additional Experiments, and Clarifications

We thank the reviewers for their thoughtful comments and constructive suggestions. Their feedback has motivated us to further investigate the generalizability, component contributions, robustness, and practical efficiency of A3Cache. To address these concerns, we have extended our evaluation with additional model comparisons, controlled ablations, repeated experiments, and physical-robot validation. We first outline the central ideas of A3Cache and then highlight the new evidence and clarifications provided in our responses.

Key Contributions

Hierarchical Asynchronous Caching. A3Cache exploits view-specific redundancy in multi-view WAMs through independent view scheduling, reducing redundant visual computation across denoising steps.

Action-Aware Token Selection. Beyond visual feature differences, A3Cache incorporates changes in visual-to-action contributions to prioritize action-relevant token updates, improving the efficiency–performance trade-off.

Additional Experiments

Generalization Across WAMs. We extend our evaluation from Motus to OpenWAM, where A3Cache retains 99.3% of the uncached success rate with a 4.08× computation reduction, outperforming the compared caching methods.

Few-Step Inference. We evaluate Motus with five denoising steps, achieving 97.1% success-rate retention and a 2.08× computation reduction, demonstrating effectiveness under shorter inference schedules.

Independent Per-View Caching. We apply per-view scheduling to SeaCache and CA-WorldCache, improving their average success rates from 78.56% to 79.88% and from 79.12% to 80.00%, respectively. These comparisons isolate the benefits of view-specific scheduling.

View and Token Selection Ablations. Under comparable computation budgets, we compare four configurations: no selection, view-only, token-only, and both. Full A3Cache achieves the highest average success rate (80.35%), demonstrating the complementary benefits of the two components.

Stability of Action-Aware Scoring. We repeat the visual-only scoring experiment with different random seeds. The consistently lower results (79.45%–79.52%) compared with Full A3Cache (80.35%) provide further evidence for the benefit of action-related signals.

Robustness on Bimanual Tasks. We evaluate 24 tasks involving both robot arms, where A3Cache retains 97.4% of the uncached average success rate, outperforming the compared caching methods.

Physical-Robot Evaluation. We additionally evaluate OpenWAM on the real-world Stack Bowls task from RoboChallenge Table30. A3Cache achieves the same observed success rate as the original model with a 4.08× computation reduction, while also reporting task progress scores.

Further Clarifications

Evaluation Fairness. We clarify that Original (Aligned) refers to the unaccelerated Motus baseline, and that all caching methods are evaluated under the same settings.

Computation vs. Latency. We distinguish FLOPs-based computation reduction from actual inference speedup, clarifying that A3Cache achieves 4.23× computation reduction and 3.45× measured latency speedup on Motus.

These results provide broader empirical support for A3Cache across different models, inference settings, and manipulation scenarios. They also help distinguish the individual contributions of view scheduling and action-aware token selection, while clarifying the scope and limitations of our evaluation. We address the reviewers' specific questions in the following responses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.