Faster-WAMDo World Action Models Need Deep Action Modules?

Liheng Ma*,1,2Rui Heng Yang*,1Amin Soleimani Abyaneh†,2George Z. Xue†,3Behnam Rahmati1Mateo Clémente1Ziwen Hu4Anlin Chen5Tongtong Cao5Zhanguang Zhang1,‡Yingxue Zhang1

1 Huawei Noah’s Ark Lab Montreal2 McGill University & Mila-Quebec AI Institute3 University of Toronto4 Huawei Celia Team5 Department of Foundation Model, Huawei 2012 Labs

* Equal contribution, listed alphabetically. † Work done as interns at Huawei Noah’s Ark Lab Montreal. ‡ Correspondence: Zhanguang Zhang.

Paper Code WeightsComing soon

Abstract

World Action Models (WAMs) benefit from representations learned by pretrained video models, but deep action modules incur high inference latency that limits their practicality in real-world robot control. Faster-WAM concentrates capacity in the video backbone and uses a single-layer action expert, connected to all video layers through Lite KV-Fusion, with world-model-only conditioning and retracted 1D-RoPE. Without additional embodied pretraining, it achieves competitive control performance on LIBERO and RoboTwin 2.0 and 78.3% success on LIBERO-Plus. It provides a 3.7–3.8× inference speedup over Fast-WAM in controlled profiling and reduces online inference latency from 502 to 129 ms on the real-robot deployment workstation. Across four Franka Panda tasks, it retains comparable success with shorter mean completion times than Fast-WAM. By reducing the delay between observation and action prediction, low-latency inference makes world-model-based control more responsive and practical for real-world deployment.

Real-World Rollouts

The recordings below compare execution of the same instruction: “Lift the lid, put it aside, and pick the black plum.” Play at 1× or 2× speed, or drag the slider to compare the policies at the same elapsed wall-clock time.

Faster-WAM21.2 s to finish
Faster-WAM pick-fruit rollout at 0 secondsIn progress

Recorded at 0 s

π0.521.6 s to finish
Pi 0.5 pick-fruit rollout at 0 secondsIn progress

Recorded at 0 s

Fast-WAM38.5 s to finish
Fast-WAM pick-fruit rollout at 0 secondsIn progress

Recorded at 0 s

LingBot-VA139.9 s to finish
LingBot-VA pick-fruit rollout at 0 secondsIn progress

Recorded at 0 s

Illustrative rollouts from Appendix G, not mean completion times. The timeline uses a uniform seconds scale, including pauses between movements. Still images are sampled every 5 s with an additional sample at 21 s. The original frame export supplies the extra snapshots omitted from the paper’s figure; each remains visible until the next recorded snapshot. Playback speed applies equally to all four policies. Completion is marked at the reported times (21.2, 21.6, 38.5, and 139.9 s), after which the final recorded frame is held. Across the four tasks, π0.5 has shorter mean completion times than Faster-WAM.

Real-Robot Evaluation

Four tasks, 100 demonstrations per task. Faster-WAM achieves 81.3% average success, compared with 80.0% for Fast-WAM.

129 ms

Faster-WAM inference

502 ms

Fast-WAM inference

3.9×

Speedup on the deployment workstation

Measured over successful rollouts on the deployment workstation; these absolute latencies should not be compared directly with the controlled benchmark below. Faster-WAM has shorter mean completion times than the evaluated WAM baselines on all four tasks. π0.5 has the highest measured average success (86.3%) and the shortest mean completion times. See the complete real-robot results ↗

These results suggest that better use of pretrained world-model representations can support efficient, generalizable control with little action-specific computation.

Scope and limitations

The results show that competitive control does not necessarily require a deep action expert; they do not establish that one layer is optimal. Physical evaluation covers one robot platform and four manipulation tasks. Additional embodied pretraining and broader evaluations across embodiments remain open directions.

Method

Faster-WAM uses a single-layer, 28.6M-parameter action expert with access to all 30 layers of the video backbone. Three design choices make this possible: Dock of Transformers (DoT) with Lite KV-Fusion, world-model-only conditioning, and retracted 1D-RoPE for positional alignment between video keys and action queries.

A large video world model supplies key–value representations through Lite KV-Fusion to a lightweight action expert, which controls a robot.
Camera observations, language, and proprioception enter a video hub. Native keys and values from all layers are fused with shared head-specific weights. Video keys are realigned from 3D to 1D RoPE before joint attention with a single-layer action expert.
Dock of Transformers (DoT) connects the action expert to the pretrained video backbone. Click to enlarge ↗

Lite KV-Fusion.

Lite KV-Fusion learns head-specific mixtures of native keys and values across all video layers. It adds just 720 trainable parameters and avoids dense KV reprojection.

World-model-only conditioning.

Language and proprioception enter the video backbone. The action expert receives this context only through the fused video representations.

RoPE alignment.

Retracted 1D-RoPE removes the video keys’ original 3D rotation and re-encodes them on the action queries’ temporal axis and frequency grid.

The video backbone is evaluated once per observation. Its fused KV cache is then reused across the action expert’s ten denoising steps.

Benchmark Results

Faster-WAM retains strong control performance without additional embodied pretraining, while reducing the computation required to predict an action chunk.

Inference latency & task success

RoboTwin 2.0 · model computation per action chunk

Faster-WAM
53.2 ms
π0.5
71.2 ms
Fast-WAM
196.8 ms
0← Lower is better196.8 ms
3.7×

faster than Fast-WAM

90.5 ± 0.2%

average success on RoboTwin 2.0

What is included in the latency comparison?

Measured on the same 24 GiB consumer GPU, at batch size one. Totals include text and vision encoding, multimodal conditioning, and action generation; they exclude external input preparation, transfers, and robot execution. Fast-WAM totals combine separately measured components because the full pipeline exceeds GPU memory.

Fast-WAM and Faster-WAM both predict 32-action chunks with 10 action denoising steps. π0.5 uses its native configuration (50 actions on RoboTwin; 10 on LIBERO). Thus, the WAM comparison matches action-generation settings; the VLA comparison does not match chunk length. Full profiling protocol ↗

Selected baselines · task success (%)
Model Additional embodied
pretraining		 LIBERO RoboTwin 2.0 LIBERO-Plus
π0.5 Yes 96.9 79.8 85.7
LingBot-VA Yes 98.5 92.2 69.5
Fast-WAM No 97.6 91.8 51.5
Faster-WAM No 98.3 ± 0.1 90.5 ± 0.2 78.3 ± 0.1

Faster-WAM: mean ± standard deviation over three evaluation runs. Baseline success rates are taken from the sources cited in Tables 1–3 of the paper; selected comparisons are not an exhaustive ranking.

Generalization to Distribution Shifts

The LIBERO-trained policy transfers to LIBERO-Plus without additional training or data augmentation.

+26.8 points

overall success vs. Fast-WAM

Faster-WAMFast-WAMSuccess rate (%)
Background
60.7
53.7
Camera
76.5
16.4
Language
90.9
68.9
Lighting
94.2
78.2
Object
81.7
60.7
Robot
58.3
44.5
Sensor
84.6
37.7

Faster-WAM improves over Fast-WAM across all seven perturbation categories. Both use the Wan2.2-5B video backbone without additional embodied pretraining. Values from Table 3.

Citation

@misc{ma2026fasterwam,
  title   = {Faster-WAM: Do World Action Models Need Deep Action Modules?},
  author  = {Liheng Ma and Rui Heng Yang and Amin Abyaneh and
             George Z. Xue and Behnam Rahmati and Mateo Clemente and
             Ziwen Hu and Anlin Chen and Tongtong Cao and
             Zhanguang Zhang and Yingxue Zhang},
  year    = {2026},
  eprint  = {2608.02365},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url     = {https://arxiv.org/abs/2608.02365}
}