The task utility gap in human mobility modeling: evidence from environmental exposure assessment

Human mobility data are essential for applications such as environmental exposure assessment, but privacy constraints limit access to large-scale trajectory datasets. Trajectory generative models offer a promising solution, yet whether statistical-fidelity-based evaluation guarantees task utility remains unclear. Task utility is defined as the agreement between downstream results derived from generated and observed trajectories. Using trajectory data from Beijing and Shenzhen, two cities with contrasting urban structures, we identified a systematic gap between statistical fidelity and task utility. We showed that commonly used trajectory-centered metrics were often weak predictors of downstream performance, whereas metrics reflecting spatial extent and collective flow structure provided more reliable indicators of task utility. Motivated by this gap, we proposed StayMove, an interpretable mechanistic framework coupling empirical activity-state transitions with a behaviorally informed spatial choice model. Across both cities, StayMove achieved high statistical fidelity and robust task utility with transferable parameters and no task-specific tuning. These findings highlight the importance of task-oriented evaluation and provide a practical framework for developing mobility models that are both behaviorally realistic and useful for downstream applications.