MARS: Systematic Multi-Level Acceleration for World Action Models

Zengyuan Zuo; Yongchang Zhang; Guangyu Zhao; Tianze Zhang; Biao Wu; Meng Fang; Yiwu Zhong

cs.CVOct 12, 2026

MARS: Systematic Multi-Level Acceleration for World Action Models

Authors: Zengyuan Zuo, Yongchang Zhang, Guangyu Zhao, Tianze Zhang, Biao Wu, Meng Fang, Yiwu Zhong

Organizations: School of Intelligence Science and Technology, Peking University · State Key Laboratory of General Artificial Intelligence, Peking University · University of Technology Sydney · University of Liverpool

Abstract

World Action Models (WAMs) have emerged as a promising alternative for embodied intelligence by jointly predicting future world states and robot actions. However, their heavy visual encoding and iterative generative inference incur substantial latency, limiting their applicability to high-frequency closed-loop control. Existing training-free acceleration methods typically optimize a single stage, yielding limited speedups or degraded control under aggressive reuse. In this work, we systematically profile the inference overhead of WAMs and observe abundant redundancy across Transformer computation, diffusion denoising, and consecutive policy queries. However, the safe removal of such redundancy depends strongly on its impact on action prediction and control sensitivity. Based on these observations, we introduce MARS, a training-free and plug-and-play framework for multi-level WAM acceleration. MARS selectively removes low-impact attention computation, adaptively accelerates diffusion inference according to motion granularity, and speculatively reuses predicted future visual representations across consecutive queries with online verification. Together, these complementary mechanisms provide remarkable acceleration while limiting approximation error. Extensive experiments across LIBERO, LIBERO-PRO, and RoboCasa show that MARS achieves up to $4.00\times$ , $3.94\times$ , and $2.10\times$ end-to-end speedups, respectively, while maintaining task success rate close to the original policy. Moreover, MARS generalizes across different model backbones and across simulation and real-robot experiments, substantially improving the efficiency--performance trade-off of WAM inference. Our code is available at https://github.com/PKU-VaLuE-Lab/MARS.

Figures & tables

Figure 1: WAM inference and acceleration overview. (a) The overall pipeline of a World Action Model. (b) Efficiency-performance comparison in LIBERO-PRO ( Zhou et al., 2025 ) , where our method substantially reduces latency and FLOPs with a better efficiency–performance trade-off.

Figure 2: Redundancy across multiple levels of WAM inference. (a) Analysis 1 : Importance scores are measured by the effects after removing certain components. We find that cross-attention branches at early-to-middle layers exhibit low scores (e.g., blocks 2–11) and only few attention heads show high scores (e.g., yellow entries). (b) Analysis 2 : Relative L1 differences between denoising steps and the first denoising step. The denoising input varies substantially across steps, whereas the denoising output and DiT residual remain relatively stable. (c) Analysis 3 : Comparison of observed and predicted images. The prediction closely matches the observation with minor variations.

Figure 3: The framework of MARS. MARS is a training-free and plug-and-play framework that accelerates WAM inference at three complementary levels. At the Transformer level, it skips low-impact cross-attention branches and prunes low-importance self-attention heads. At the diffusion level, it uses the provisional action to estimate the motion granularity type and selects one-step generation or the multi-step solver with residual reuse. At the cross-query level, it reuses the predicted future visual tokens when they are sufficiently consistent with the observed visual changes.

LIBERO	Success Rate (%, $\uparrow$ ) / Speedup ( $\uparrow$ )				Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )	FLOPs ( $\downarrow$ )
Method	Goal	Object	Spatial	Long	Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )	FLOPs ( $\downarrow$ )
Cosmos Policy	98.20 / 1.00	99.87 / 1.00	97.47 / 1.00	96.67 / 1.00	98.05%	1.00 $\times$	1.4001 s	100.00%
+ DiCache	98.07 / 1.70	99.87 / 1.66	97.80 / 1.67	97.27 / 1.72	98.25%	1.69 $\times$	0.8273 s	48.12%
+ C 3 ache ∗	97.80 / 2.19	99.00 / 2.16	97.53 / 2.25	94.47 / 2.30	97.20%	2.24 $\times$	0.6256 s	33.18%
+ X-Cache ∗	97.87 / 1.11	99.67 / 1.17	97.47 / 1.12	97.00 / 1.21	98.00%	1.17 $\times$	1.2017 s	80.20%
+ Ours (Speed)	96.73 / 3.89	99.07 / 4.08	97.60 / 3.87	96.27 / 4.07	97.42%	4.00 $\times$	0.3498 s	18.39%

Table 1: Comparison with representative training-free acceleration methods on the LIBERO ( Liu et al., 2023 ) and LIBERO-PRO ( Zhou et al., 2025 ) benchmarks. * denotes our re-implementation due to unavailable official code. All acceleration methods are adapted to the Cosmos Policy ( Kim et al., 2026 ) architecture for a fair comparison.

Method	S.R. ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )	FLOPs ( $\downarrow$ )
Cosmos Policy	66.69%	1.00 $\times$	2.2958 s	100.00%
+ DiCache	66.47%	1.50 $\times$	1.5305 s	49.99%
+ C 3 ache ∗	65.00%	1.81 $\times$	1.2653 s	33.64%
+ X-Cache ∗	66.37%	1.20 $\times$	1.9202 s	72.86%
+ Falcon	66.47%	1.21 $\times$	1.8968 s	74.56%
+ Ours	66.94%	2.10 $\times$	1.0911 s	22.50%

Table 2: Evaluation on the RoboCasa benchmark ( Nasiriany et al., 2024 ) .

LIBERO	Success Rate (%, $\uparrow$ ) / Speedup ( $\uparrow$ )				Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )
Method	Goal	Object	Spatial	Long	Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )
Fast-WAM	95.87 / 1.00	99.40 / 1.00	97.13 / 1.00	93.93 / 1.00	96.58%	1.00 $\times$	0.5365 s
+ Ours	96.07 / 3.43	99.00 / 3.25	96.93 / 3.43	93.20 / 2.93	96.30%	3.17 $\times$	0.1693 s
LingBot-VA	–	–	–	97.40 / 1.00	97.40%	1.00 $\times$	13.4105 s
+ Ours	–	–	–	98.13 / 2.88	98.13%	2.88 $\times$	4.6558 s
$\pi_{0.5}$	97.80 / 1.00	98.67 / 1.00	98.40 / 1.00	92.93 / 1.00	96.95%	1.00 $\times$	0.1462 s

Table 3: Our method with additional WAM and VLA base models on LIBERO ( Liu et al., 2023 ) . LingBot-VA can only be evaluated in Long, as only the Long checkpoint is publicly available.

LIBERO-PRO	Success Rate (%, $\uparrow$ ) / Speedup ( $\uparrow$ )				Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )	FLOPs ( $\downarrow$ )
Method	Goal	Object	Spatial	Long	Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )	FLOPs ( $\downarrow$ )
Cosmos Policy	43.67 / 1.00	46.13 / 1.00	47.40 / 1.00	45.33 / 1.00	45.63%	1.00 $\times$	1.4044 s	100.00%
Ours (Speed)	44.20 / 3.91	45.67 / 3.86	47.33 / 3.96	44.93 / 3.99	45.53%	3.94 $\times$	0.3564 s	18.80%
w/o Action-Aware Step Skipping	43.67 / 1.18	44.67 / 1.18	47.00 / 1.18	44.47 / 1.21	44.95%	1.19 $\times$	1.1803 s	92.81%
w/o Attention Skipping	43.80 / 3.64	45.67 / 3.59	47.80 / 3.69	43.93 / 3.69	45.30%	3.66 $\times$	0.3841 s	20.19%
w/o Visual Token Reuse	43.60 / 3.07	46.87 / 3.01	47.53 / 3.08	44.67 / 3.07	45.67%	3.06 $\times$	0.4595 s	19.05%

Table 4: Ablation study on the LIBERO-PRO benchmark ( Zhou et al., 2025 ) . We evaluate three acceleration components of our method.

LIBERO-PRO	Success Rate (%, $\uparrow$ ) / Speedup ( $\uparrow$ )				Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )	VAE Skip Rate
Method	Goal	Object	Spatial	Long	Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )	VAE Skip Rate
Cosmos Policy	43.67 / 1.00	46.13 / 1.00	47.40 / 1.00	45.33 / 1.00	45.63%	1.00 $\times$	1.4043 s	–
Layer-0	44.13 / 3.91	45.07 / 3.86	46.93 / 3.91	44.73 / 3.98	45.22%	3.93 $\times$	0.3574 s	56.20%
Layer-1	44.20 / 3.91	45.67 / 3.86	47.33 / 3.96	44.93 / 3.99	45.53%	3.94 $\times$	0.3564 s	56.25%
Layer-2	43.67 / 3.85	45.27 / 3.79	46.87 / 3.92	45.33 / 3.93	45.28%	3.88 $\times$	0.3621 s	56.41%
Layer-3	43.73 / 3.83	45.40 / 3.78	47.20 / 3.90	45.33 / 3.91	45.42%	3.86 $\times$	0.3638 s	56.44%

Table 5: Sensitivity Analysis of the VAE Feature Layer in LIBERO-PRO. Layer-1 achieves the best task performance and inference efficiency among the evaluated feature layers.

Method	S.R. ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )
Cosmos Policy	90.00%	1.00 $\times$	0.6949 s
+ Ours	95.00%	2.65 $\times$	0.2625 s
$\pi_{0.5}$	25.00%	1.00 $\times$	0.1463 s
+ Ours	25.00%	1.21 $\times$	0.1209 s

Table 6: Evaluation on the real-robot task on a Franka robot.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

A40
LIBERO	Success Rate (%, $\uparrow$ ) / Speedup ( $\uparrow$ )				Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )
Method	Goal	Object	Spatial	Long	Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )
Cosmos Policy	98.20 / 1.00	99.87 / 1.00	97.47 / 1.00	96.67 / 1.00	98.05%	1.00 $\times$	1.4001 s
+ Ours (Speed)	96.73 / 3.89	99.07 / 4.08	97.60 / 3.87	96.27 / 4.07	97.42%	4.00 $\times$	0.3498 s
RTX 3090
Cosmos Policy	97.67 / 1.00	100.00 / 1.00	97.53 / 1.00	97.40 / 1.00	98.15%	1.00 $\times$	1.4181 s

Appendix

Table 7: Comparison of our method across different GPUs in LIBERO.

LIBERO	Success Rate (%, $\uparrow$ ) / Speedup ( $\uparrow$ )				Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )
Method	Goal	Object	Spatial	Long	Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )
Seed 195
Cosmos Policy	98.20 / 1.00	99.80 / 1.00	98.00 / 1.00	96.00 / 1.00	98.00%	1.00 $\times$	1.4194 s
+ DiCache	98.20 / 1.69	100.00 / 1.59	98.60 / 1.65	96.80 / 1.60	98.40%	1.62 $\times$	0.8737 s
+ C 3 ache $*$	98.40 / 2.20	99.00 / 2.18	97.00 / 2.26	95.20 / 2.39	97.40%	2.28 $\times$	0.6215 s
+ X-Cache ∗	97.80 / 1.10	99.40 / 1.17	97.80 / 1.12	95.80 / 1.24	97.70%	1.18 $\times$	1.2036 s

Appendix

Table 8: Results in LIBERO with different random seeds.

Figure 4: Analysis of the latency. Component-wise latency breakdown and the resulting shift in computational bottlenecks as different acceleration components are progressively introduced.

LIBERO-PRO	Success Rate (%, $\uparrow$ ) / Speedup ( $\uparrow$ )				Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )
Method	Goal	Object	Spatial	Long	Success Rate ( $\uparrow$ )	Speedup ( $\uparrow$ )	Latency ( $\downarrow$ )
Cosmos Policy	43.67 / 1.00	46.13 / 1.00	47.40 / 1.00	45.33 / 1.00	45.63%	1.00 $\times$	1.4043 s
Ours-1 (0.0)	44.00 / 4.01	45.47 / 3.95	46.80 / 4.02	44.27 / 4.10	45.13%	4.03 $\times$	0.3483 s
Ours-2 (0.2)	43.73 / 3.93	45.60 / 3.84	46.80 / 3.95	43.80 / 4.04	44.98%	3.96 $\times$	0.3550 s
Ours-3 (0.4)	44.20 / 3.91	45.67 / 3.86	47.33 / 3.96	44.93 / 3.99	45.53%	3.94 $\times$	0.3564 s
Ours-4 (0.6)	43.93 / 3.90	45.47 / 3.86	47.27 / 3.95	44.40 / 3.99	45.27%	3.94 $\times$	0.3567 s

Appendix

Table 9: Sensitivity Analysis of the Action Routing Threshold

\tau_{\mathrm{act}}

in LIBERO-PRO. Our method maintains comparable task performance and substantial acceleration across a broad range of action threshold values.

Figure 5: Qualitative examples where our method improves task completion. Representative comparisons in LIBERO and LIBERO-PRO where Ours (Acc.) succeeds while the original Cosmos Policy fails.

Figure 6: Qualitative examples of failure cases under acceleration. Representative comparisons in LIBERO and LIBERO-PRO where the original Cosmos Policy succeeds while Ours (Acc.) fails.

Efficient Inference for World Action Models

Efficient inference for world action models cuts the cost of predicting action-conditioned futures for responsive control and planning.

See all 80 papers

Sep 28, 2026

Efficient World Action Model Inference with Adaptive Intermediate States

World Action Models (WAMs) enable future-aware control by jointly modeling actions and environment dynamics. However, iterative diffusion or flow inference incurs substantial denoising latency. Prior inference state offers a natural opportunity for acceleration, yet changing planning contexts, observations, and intermediate representations can quickly render retained state stale. Preserving useful computation therefore requires adapting inference state rather than reusing it as-is. To this end, we present

\mathrm{WAM}{\scriptstyle\mathrm{ACHINE}}

, a training-free framework that accelerates WAM inference by preserving and adapting inference state for efficient and accurate continuation as the control loop evolves. Across closed-loop replans, Trajectory Remapping remaps replan state from the preceding replan to initialize the next replan, reducing redundant trajectory generation. Across denoising steps, Observation Rebinding performs anticipatory inference during action execution and rebinds retained denoising state to the real observation for continuation when consistency checks pass, reducing latency exposed to the control loop. Across Transformer layers, Residual Rescaling selectively rescales retained layer state and refreshes it through full computation of the middle layers when probe checks fail, reducing repeated Transformer computation. Evaluations of three representative WAM architectures on LIBERO and RoboTwin 2.0 show that

\mathrm{WAM}{\scriptstyle\mathrm{ACHINE}}

achieves 1.47-3.05

\times

speedups in observation-to-action latency and 2.23-3.27

\times

speedups in GPU inference time per replan, while preserving 96.69-99.54% of native WAM task success.

Zhinnan Liu, Haozhi Han, Ruge Zhang +8

Xiamen University · Institute for AI Industry Research, Tsinghua University · School of Computer Science, Peking University +2

Efficient Transformer Inference World Model-Based Planning Efficient Inference for World Action Models Robotic Manipulation World Action Models

Ask PDF

Oct 7, 2026

RealtimeWAM: How Fast Can I Run My World Action Model?

World Action Models (WAMs) combine visual dynamics modeling with action generation, but their high inference latency limits responsive robot control. Recent efforts accelerate inference by removing explicit future-video generation at test time, as in FastWAM, an approach that requires a specially tailored architectural design. More general caching strategies exploit feature redundancy, but redundancy alone does not capture the changing computational demands of closed-loop control. To address these challenges, we present RealtimeWAM, a general, training-free framework that coordinates parallel execution with adaptive computation for low-latency inference across diverse WAM architectures. We exploit layerwise dependencies to overlap observation processing with prediction. However, concurrent branches still compete for GPU resources, limiting the benefit of parallel execution. We therefore adapt computation throughout the pipeline through selective reuse, caching observation features in visually stable regions and reusing Transformer residuals while reserving additional refinement for small predicted adjustments. We evaluate RealtimeWAM on FastWAM and OpenWAM across RoboTwin, LIBERO, and LIBERO-Plus. On an RTX 4090, measured mean inference latencies are 24.09 and 63.09 ms, corresponding to average speedups of 8.90

\times

and 10.67

\times

. Average success rates are 82.75% and 87.41%, respectively, within 0.02 and 0.53 percentage points of native inference. Across five real-world tasks, RealtimeWAM improves average success rates over native inference by 17.2 and 37.2 percentage points on FastWAM and OpenWAM, respectively.

Huanan Liu, Ye Li, Kangye Ji +8

Tsinghua University · YuanxingGuangnian Robotics · Nanjing University

Efficient Inference for World Action Models World Action Models Efficient Inference Robotic Manipulation

Ask PDF

Oct 24, 2026

TurboWAM: Learning What Matters for Fast World Action Models

World action models have demonstrated promising performance in robot policy learning, yet their high inference latency remains a critical bottleneck for real-world deployment. Recent approaches have sought to improve inference efficiency by reducing test-time future-frame imagination or accelerating action prediction. However, these methods often leverage representations from all video world model layers without explicitly considering their varying relevance to action prediction, leading to redundant computation and suboptimal representation utilization. In this work, we propose TurboWAM, a fast world action model that efficiently aggregates structured world representations conditioned on action predictions. Specifically, TurboWAM introduces a structured world representation aggregation module that adaptively fuses representations from different world model layers according to their relevance to the evolving action denoising state. Furthermore, TurboWAM employs a structured action supervision mechanism that adaptively modulates gradients from action supervision across different world model layers during backward propagation, facilitating world representation learning. During inference, by reducing the number of world model layers executed and caching relevant representations across action denoising steps, TurboWAM substantially reduces computational overhead. Extensive experiments on multiple robot manipulation benchmarks demonstrate that TurboWAM achieves low inference latency while maintaining competitive task success rates compared with state-of-the-art methods.

Guile Wu, Shuangzhi Li, Xingxin Chen +5

Huawei Noah’s Ark Lab · University of Alberta · Department of Foundation Model, Huawei

Efficient Inference for World Action Models World Action Models Action-Conditioned World Models World Model Learning

Ask PDF

Abstract

Efficient Inference for World Action Models

Explore similar work

Efficient World Action Model Inference with Adaptive Intermediate States

RealtimeWAM: How Fast Can I Run My World Action Model?

TurboWAM: Learning What Matters for Fast World Action Models