Paper Title: Generative Agents for Realistic Mobile Computing and User Behavior Case
Conference Name: The 4th International Conference on Mobile • Military • Maritime IT Convergence
Abstract: Realistic mobile user behavior is essential for evaluating mobile systems, but traditional methods often lack context and adaptability. We explore using Large Language Models
(LLMs) to simulate human-like mobile interactions by developing generative agents with defined personas and goals. These agents query Google’s Gemini model within a discrete-time simulation environment to determine actions based on dynamic context. We compared LLM-powered agents against a random-action baseline across key metrics like location, activity, application usage, and charging patterns, using simulated 24-hour histories
for student and office worker personas. Results show that LLM-agents exhibit more plausible, goal-oriented behavior that better aligns with persona expectations and real-world benchmarks. Despite some limitations, such as overestimated application usage, this work demonstrates the potential of LLM-powered agents for creating rich, context-aware mobile behavior scenarios.