MobileWorldBench: Towards Semantic World Modeling For Mobile Agents

TL;DR

This paper introduces MobileWorldBench, a benchmark and dataset for semantic world modeling in mobile GUI agents using vision-language models, demonstrating improved task success rates through natural language-based state transition modeling.

Contribution

It presents a new benchmark and large-scale dataset for semantic world modeling in GUI agents, and integrates vision-language models into planning for better task performance.

Findings

VLMs can effectively model GUI environments using natural language.

Semantic world models improve task success rates for mobile agents.

MobileWorldBench facilitates evaluation of language-based world modeling approaches.

Abstract

World models have shown great utility in improving the task performance of embodied agents. While prior work largely focuses on pixel-space world models, these approaches face practical limitations in GUI settings, where predicting complex visual elements in future states is often difficult. In this work, we explore an alternative formulation of world modeling for GUI agents, where state transitions are described in natural language rather than predicting raw pixels. First, we introduce MobileWorldBench, a benchmark that evaluates the ability of vision-language models (VLMs) to function as world models for mobile GUI agents. Second, we release MobileWorld, a large-scale dataset consisting of 1.4M samples, that significantly improves the world modeling capabilities of VLMs. Finally, we propose a novel framework that integrates VLM world models into the planning framework of mobile agents, demonstrating that semantic world models can directly benefit mobile agents by improving task success rates. The code and dataset is available at https://github.com/jacklishufan/MobileWorld