What Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling

Renping Zhou; Zanlin Ni; Zihao Fan; Guohao Fu; Zeyu Liu; Hao Shi; Jie Zhang; Chi Bene Chen

cs.CVSep 28, 2026

What Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling

Authors: Renping Zhou, Zanlin Ni, Zihao Fan, Guohao Fu, Zeyu Liu, Hao Shi, Jie Zhang, Chi Bene Chen, +3 more

Organizations: Leap Lab, Tsinghua University · University of Science and Technology of China · Beijing Institute of Technology

Abstract

World action models (WAMs) predict the future alongside actions during training. Due to the heavy computation cost of video denoising, whether the future must still be generated during inference is disputed: Explicit WAMs denoise it into clean frames along with every action chunk, whereas Latent WAMs discard it entirely for acceleration. We find that latent WAMs, despite matching explicit ones on in-distribution tasks, fail to retain the generalization benefits that originally motivated WAMs. To demonstrate this, we evaluate generalization along three axes: environmental perturbation, data efficiency, and task generalization. Controlled comparisons with a matched backbone, training data, and budget reveal consistent degradation across all three axes when the action expert no longer conditions on future representations. Further analysis shows that the gap arises almost entirely from the first denoising step: the benefit comes from preparing the future, not generating it. We therefore propose Simple-WAM, which simplifies future modeling into a single forward pass of fully noised video tokens and adapts the training-time noise schedule to this inference behavior. Across simulation and real-world tasks, Simple-WAM achieves the best of both worlds, leading explicit WAMs in generalization performance with efficiency comparable to Latent WAMs. Project Page: https://zrporz.github.io/Simple-WAM-Web/

Figures & tables

Figure 1: An empirical study of test-time future modeling for WAM generalization. Left: the three axes of the protocol, environmental perturbation, data efficiency, and task generalization. Right: the explicit and latent paradigms are within $0.9$ points on in-distribution tasks yet separate on all three axes (F1), and a single forward pass over a fully noised future recovers almost all of that separation (F2).

	In-dist.	Generalization
	In-dist.	Perturb.	Data Eff.	Task Gen.
Paradigm	LIBERO	LIBERO-Plus	10-shot	w/o vid	w/ vid
Explicit	97.75	67.72	96.95	6.25	69.90
Latent	96.85	53.75	88.50	2.10	5.90
$\Delta$	$0.90$	$\mathbf{13.97}$	$\mathbf{8.45}$	$\mathbf{4.15}$	$\mathbf{64.00}$

Table 1: Generalization Deteriorates without Inference-time Future Modeling. Success rate (%) under the matched setting of Sec. 4.1 . w/ vid: action-free video of the held-out tasks with video available.

\Delta

is Explicit minus Latent; the in-distribution column is greyed , and the generalization deltas are bold .

Figure 2: Fully Noised Future Is Sufficient for Generalization. Success rate (%) on LIBERO-Spatial for a model trained under the explicit paradigm. The video expert is run for

t_{\mathrm{denoise}}

of the

K

denoising steps; no retraining is performed.

Figure 3: The three conditioning schemes. The explicit paradigm denoises the future video tokens across the schedule; the latent paradigm drops them and keeps only the single forward pass. We propose Simple-WAM, which keeps both the future video tokens and the single forward pass: future video tokens are input as pure Gaussian noise at $\tau=1$ and are never denoised.

Environmental perturbation

Data efficiency

Task generalization

LIBERO-Plus

LIBERO

RoboTwin

LIBERO

RoboTwin

Method

Emb. P.T.

Init.

Cam.

Lang.

Noise

Bg.

Lay.

Light

Avg.

5-shot

10-shot

w/o vid

w/ vid

w/o vid

w/ vid

Vision-language-action models

\pi_{0.5}

( 2025 )

✓

73.6

78.4

80.8

94.1

84.5

96.2

84.4

87.5

91.5

33.9

9.4

9.4*

2.5

2.5*

VLA-Adapter ( 2025 )

✗

37.4

36.4

73.8

57.2

76.6

70.2

71.0

59.0

77.6

85.9

–

0.3

0.3*

–

Explicit WAMs

Table 2: Main results across the three generalization axes. Success rate (%) under the protocol of Sec. 4.1 , per factor for environmental perturbation and averaged over suites elsewhere. w/o vid. and w/ vid. denote task generalization without and with action-free video of the held-out tasks; such video provides no training signal in a VLA, so “*” repeats the w/o vid. entry. “Emb. P.T.” denotes large-scale embodied pretraining; those rows are grey , and bold indicates the best among the rest. “–”: not covered by the official release. Detailed results are in Tab. A2 .

Figure 4: Inference Latency. Per-chunk latency and mean success rate over the four generalization settings of Tab. 2 .

Figure 5: Real-world results. Success rate (%) over 30 evaluation runs. T1–T4 are the four training tasks of Fig. 6 , in order; task generalization is the held-out Store in Order .

Figure 6: The real-world setup. Left: the four training tasks, one per column, each shown at three moments of a single rollout. Middle: the three environmental perturbations, varying the lighting, the object layout and the background. Right: Store in Order , the held-out task.

Table 3: Ablations on LIBERO-Spatial. The shaded entries are the default setting and bold marks the best.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

	LIBERO	RoboTwin 2.0	Real world
Optimizer	AdamW, $\beta=(0.9,\,0.95)$
Learning rate	$1\times 10^{-4}$
Weight decay	$1\times 10^{-2}$
LR schedule	cosine, $5\%$ warmup
Precision	BF16
Batch size	$128$	$1{,}024$	$512$

Appendix

Table A1: Training configuration.

	LIBERO					RoboTwin 2.0
	Spatial	Object	Goal	Long	Avg.	Clean	Rand.	Avg.
Data efficiency, 5-shot
$\pi_{0.5}$	92.1	94.5	89.7	73.6	87.5	–	–	–
VLA-Adapter	85.7	95.0	75.8	54.0	77.6	–	–	–
Lingbot-VA	59.0	96.3	88.6	70.2	78.5	–	–	–
FastWAM-Joint	97.0	99.4	92.4	77.2	91.5	–	–	–

Appendix

Table A2: Success rate (%) behind the data-efficiency and task-generalization columns of Tab. 2 .

	T1	T2	T3	T4	Avg.
In-distribution
FastWAM	$83.3$ $\pm$ 19.2	$92.5$ $\pm$ 16.9	$81.1$ $\pm$ 12.9	$93.3$ $\pm$ 14.1	$87.6$
FastWAM-Joint	$80.0$ $\pm$ 21.9	$92.5$ $\pm$ 16.9	$83.3$ $\pm$ 13.1	$92.2$ $\pm$ 11.8	$87.0$
Simple-WAM	$81.7$ $\pm$ 16.6	$95.0$ $\pm$ 10.5	$80.0$ $\pm$ 15.5	$84.4$ $\pm$ 18.3	$85.3$
Environmental perturbation
FastWAM	$23.3$ $\pm$ 26.3	$17.5$ $\pm$ 23.7	$20.0$ $\pm$ 13.7	$40.0$ $\pm$ 23.5	$25.2$

Appendix

Table A3: Real-world results per task.

Explore similar work

World Action Models: A Survey
Jun 18, 2026Qiuhong Shen, Shihua Zhang, Yue Liao +5Recent World-Action Models World Models
National University of Singapore
Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination
Jun 8, 2026Jiajun Li, Tiecheng Guo, Yifan Ye +9Efficient World-Action Model Faster-Wam
1The University of Hong Kong · 2Peking University · 3Muka Robotics +2
MoWAM: Explicit Future Motion Prediction for Efficient World Action Models
Sep 17, 2026Jiayu Wang, Bin Zhu, Yue Yu +1Efficient World-Action Model World Models
College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China · Singapore Management University, Singapore · Institute of Trustworthy Embodied AI, Fudan University, Shanghai, China

Jun 18, 2026, cs.ROJ/K move · Enter open · S save

World Action Models: A Survey

Qiuhong Shen, Shihua Zhang, Yue Liao, Qi Li+4

National University of Singapore

Recent World-Action Models World Models Action Prediction World Survey

Open paper AskPDF arXiv

World Action Models (WAMs) are embodied predictive-action models that make a forecast of the future available to action. Recent WAMs repurpose large video generation models, and a parallel line relies on language or vision-language backbones without a video-generation core. This rapid expansion has blurred the boundary among broad world models, video generation models, action-grounded video world models, Vision-Language-Action policies, and WAMs. This survey gives the field a common account. It first clarifies these boundaries, then organizes existing works through two complementary views. The first view asks what each method is required to generate, spanning rendered futures, latent futures, and video-generation-free action reasoning. The second view decomposes each method by predictive substrate, backbone, action coupling, and deployment regime. This anatomy supports a unified discussion of interactability, causality, persistence, physical plausibility, and generalization, followed by data, evaluation, and open challenges. Across these axes, a consistent design pattern emerges: WAMs are not simply video generators with action heads, but predictive-action methods whose design choices trade representational richness against compute, memory, latency, and action-label cost. The field is moving toward methods that generate less of the future while preserving what control requires. The survey homepage is available at https://world-action-models.github.io/.

Abstract

Explore similar work

World Action Models: A Survey

Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination

MoWAM: Explicit Future Motion Prediction for Efficient World Action Models