How Mobile World Model Guides GUI Agents?

arXiv 2026 54.6 method, benchmark, application

TLDR

Investigates how mobile world models guide GUI agents by comparing four modalities and evaluating downstream utility on benchmarks.

Reasoning

The paper provides a systematic comparison of world model modalities and demonstrates strong empirical results on multiple benchmarks, which is a strength. However, the abstract does not detail limitations or failure cases, and the focus on mobile GUI agents may limit generalizability.

Read-first score

Read-first score 54.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 42.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

78.1

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.78057447

Topical relevance 29%

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=result

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridge

Rank sensitivity

Stability: volatile; rank range: 284.

Keyword Scores

world model

generative world model

world dynamics prediction

interactive world model

world simulator

model-based reinforcement learning world model

video world model

Deep Analysis

Innovations

Training world models across four modalities: delta text, full text, diffusion-based images, and renderable code.
Achieving state-of-the-art performance on MobileWorldBench and Code2WorldBench.
Identifying that renderable code reconstruction provides high in-distribution fidelity and effective multimodal supervision, while text-based feedback is more robust for online out-of-distribution execution.
Demonstrating that world-model-generated trajectories can improve agent end-to-end task performance despite not preserving the original distribution.
Finding that world models are more effective as prior perception or training supervision than as universal post-hoc verifiers for overconfident agents.

Methodology

The authors filter and annotate mobile world-model data, then train world models across four modalities: delta text, full text, diffusion-based images, and renderable code. They evaluate these models on MobileWorldBench and Code2WorldBench for state-of-the-art comparison, and assess downstream utility on AITZ, AndroidControl, and AndroidWorld to derive findings about representation utility, trajectory transfer, and the role of world models.

Key Results

The proposed world models achieve state-of-the-art results on both MobileWorldBench and Code2WorldBench. Three key findings emerge: renderable code excels in-distribution but text is more robust for OOD; world-model trajectories improve agent training despite distribution shift; and world models are better suited as prior perception or training supervision than as post-hoc verifiers for overconfident agents.

Limitations

World-model-generated trajectories do not preserve the original data distribution.
Posterior self-reflection provides limited gains for overconfident mobile agents with low action entropy.
Renderable code reconstruction has limited out-of-distribution robustness compared to text-based feedback.

Resources

Paper