Panasonic HD and UCLA researchers have developed a world model that anticipates and supports work performed on mobile devices using AI. They created approximately 1.4 million training records and a benchmark, improving the operating success rate of an AI agent. The work is scheduled to be presented at EMNLP 2026.
Panasonic Holdings, Panasonic R&D Company of America, and researchers from the University of California, Los Angeles, have jointly developed a world model that predicts and supports work performed on mobile devices such as smartphones.
Unlike conventional methods that reproduce the screen after an operation as an image, the technology predicts operation-related changes in natural language, evaluates multiple possible operations, and selects the operation most likely to achieve the objective. The MobileWorld dataset contains operation details, screens before and after operations, descriptions of changes, and question-and-answer data about the next state. It includes approximately 942,000 screen-change records and approximately 543,000 question-and-answer records.
When Qwen3-VL-8B-Instruct, additionally trained using MobileWorld, was evaluated with MobileWorldBench, it demonstrated higher predictive performance than existing methods. In evaluations using AndroidWorld, the task success rate improved from the baseline of 46.9% to 54.3%.
The technology is intended to support work across multiple applications, including facilities management, parts inventory, and work records. Its potential application to websites and PC applications, as well as smartphones, will also be examined. The paper information includes the title of MobileWorldBench and an arXiv number. The authors include Konstaninos Kallidromitis, Kato Yusuke (加藤祐介), Kozuka Kazuki (小塚 和紀), and Shufan Li.