Seeing Through the Illusion: Test-Time Scaling for World Action Models

Cheikh Ahmed; Mahdi Mostajabdaveh; Milad Yazdani; Zirui Zhou

cs.RONov 2, 2026

Seeing Through the Illusion: Test-Time Scaling for World Action Models

Authors: Cheikh Ahmed, Mahdi Mostajabdaveh, Milad Yazdani, Zirui Zhou

Organizations: Huawei Technologies Canada · University of British Columbia

Abstract

World action models (WAMs) jointly generate actions and imagined futures, but a plausible prediction need not be realized by its co-generated action. We introduce WISE (World Action Inference Scaling through Evaluation), a pre-execution selector that scores each sampled rollout on action validity and temporal regularity, inverse-dynamics consistency between the proposed action and the imagined motion, and instruction-conditioned task progress, and executes the highest-scoring candidate while keeping the WAM frozen. With four candidates per decision, WISE raises closed-loop success from 6.0% to 24.0% on RoboLab with Cosmos~3 Edge and from 75.1% to 78.3% mean success on LIBERO with DreamZero, and exceeds the best existing selector in aggregate by 4.0 and 0.9 points. Component ablations and agreement analyses show that the three scores rank candidates differently and that task progress carries the largest weight in the fused decision.

Figures & tables

Panel of Figure 1 — Figure 1: Dream-execution divergence for Cosmos 3 Edge on RoboLab. Each pair shows the imagined future (Dream, left) and the realized execution (right) at the indicated step. The banana moves away from the predicted grasp, while the yogurt shows less motion than imagined.

Criterion	Compared quantities	Work
Future consensus	One predicted future vs. alternative futures	[ 7 ]
Cross-view geometry	Multiple views of a predicted future	[ 8 ]
Action distribution	Candidate action vs. conditional action distribution	[ 9 ]
Conditioning fidelity	Input action vs. action recovered from its rollout	[ 19 ]
Cross-output coherence	Co-generated action vs. motion in co-generated future	WISE

Table 1: Decision-time consistency criteria compare different quantities.

Figure 2: WISE at each control decision. A frozen WAM samples

B

paired action–future candidates (

B=4

in our evaluations). Each candidate receives three scores in

[0,1]

. The IDM maps imagined frames back to actions for comparison with the candidate’s proposed controls; the task evaluator scores progress at the final imagined frame. WISE sums the scores, executes the selected action chunk, and observes the resulting state before replanning.

Table 4

Table 5

Table 6

World Models for Robotics

World models for robotics predict how scenes and robot states change under actions to support planning, control, and learning.

See all 399 papers

Test-Time Scaling for World Action Models via Zero-Shot Geometric Evaluation
Jul 20, 2026Zesen Zhao, Minkyoung Cho, Hui shen +4World Models for Robotics World Model-Based Planning
University of Michigan · 2NVIDIA
Is the Future Compatible? Diagnosing Dynamic Consistency in World Action Models
May 8, 2026Bo-Kai Ruan, Teng-Fang Hsiao, Ling Lo +1World Models for Robotics World Model-Based Planning
National Yang Ming Chiao Tung University
Rethink Before You Execute: Adaptive Execution for World Action Models
Aug 10, 2026Feng Ye, Yiming Zhao, Yong Yu +5World Models for Robotics LLM Planning
Wangxuan Institute of Computer Technology, Peking University · Simplexity Robotics

Jul 20, 2026J/K move · Enter open · S save

Test-Time Scaling for World Action Models via Zero-Shot Geometric Evaluation

Zesen Zhao, Minkyoung Cho, Hui shen, Boyuan Zheng+3

University of Michigan · 2NVIDIA

World Models for Robotics World Model-Based Planning Action-Conditioned World Models Test-Time Scaling World Action Models

Open paper AskPDF arXiv

Test-time scaling improves foundation-model inference by spending additional computation, but robot control requires deciding whether extra compute is useful before executing an action. World Action Models (WAMs) make this decision natural: each rollout exposes both an action chunk and predicted future observations. We propose \methodgated, a training-free selective test-time scaling framework for WAMs. We first instantiate \method, a fixed-budget Best-of-

N

selector that ranks sampled rollouts by cross-view depth reprojection consistency of their predicted futures, computed with a frozen geometry foundation model. \methodgated\ adds a lightweight action--future consistency gate that invokes \method\ only when the initial rollout appears internally inconsistent. Across five benchmark--backbone settings on RoboCasa, LIBERO Long, and RoboTwin~2.0, fixed-budget \method\ improves

N{=}8

task success in every setting, e.g., raising the RoboCasa group average from

66.3\%

68.4\%

with Cosmos Policy and from

80.8\%

82.5\%

with X-WAM. With gating enabled, \methodgated\ recovers on average

74.8\%

of the always-on success gain while triggering additional sampling on only

26.2\%

of decision points. Offline diagnostics show that cross-view reprojection is a strong task-label-free selector, and we identify false low-score selections as a failure mode that helps explain why performance can saturate or degrade as

N

increases.

Abstract

World Models for Robotics

Explore similar work

Test-Time Scaling for World Action Models via Zero-Shot Geometric Evaluation

Is the Future Compatible? Diagnosing Dynamic Consistency in World Action Models

Rethink Before You Execute: Adaptive Execution for World Action Models