MobileWorldBench: Towards Semantic World Modeling For Mobile Agents

Shufan Li , Konstantinos Kallidromitis , Akash Gokul , Yusuke Kato , Kazuki Kozuka , Aditya Grover

🏛 Institutions: UCLA , Panasonic AI Research , Salesforce AI Research
📅 Date: December 16, 2025
📑 Publisher: arXiv
💻 Env: Mobile
🔑 Keywords: dataset benchmark semantic world model world model planning MobileWorldBench

TLDR

MobileWorldBench studies mobile world modeling through language-described state transitions instead of pixel prediction. It benchmarks vision-language models as mobile world models, releases a 1.4M-sample MobileWorld training set, and shows that these semantic world models can directly improve downstream mobile-agent planning.

Open paper arXiv Report issue