Driving · 2026-10-14

MARS: Systematic Multi-Level Acceleration for World Action Models

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

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

MARS is a training-free, plug-and-play framework for speeding up World Action Models by removing non-uniform redundancy at three levels: selectively skipping low-impact Transformer attention, adaptively accelerating diffusion denoising based on motion granularity (one-step generation for coarse motions; residual reuse for fine motions), and speculatively reusing predicted future visual tokens across consecutive queries with online verification. The authors report up to 4.00× end-to-end speedups while maintaining task success near the original policy.

Why it matters

WAM inference is described as too slow for high-frequency closed-loop control due to expensive visual encoding and iterative generative inference. The authors argue that single-stage acceleration is insufficient because redundancy is distributed across Transformer computation, diffusion denoising, and cross-query processing, and safe removal is stage-dependent. MARS targets this multi-level structure to improve the efficiency–performance trade-off for both simulation and real-robot experiments.

Method

Transformer-level: compute calibration importance offline to skip low-impact cross-attention branches and prune low-importance self-attention heads during DiT forward passes.
Diffusion-level: use the provisional action after the first denoising step to estimate motion granularity; run one-step generation for coarse motions, and for fine-grained stages reuse stable DiT residuals while preserving the original multi-step solver.
Cross-query level: treat predicted future visual tokens as speculative inputs for the next query; compare predicted vs observed visual changes using shallow VAE features and reuse tokens only when sufficiently consistent.

Abstract (from arXiv)

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 \textbf{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.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-12 · Indexed by Pub-AI: 2026-10-14 UTC

preprint