Towards Foundation Model-based Generation of Human Mobility Trajectories

Siyu Li, Toan Tran, Lingyi Zhao, Khurram Shafique, Li Xiong

Abstract: Generating realistic human mobility trajectories is critical for urban computing, transportation, and epidemiology, since privacy constraints and limited data availability often restrict access to real-world data. Recent advances apply Large Language Models (LLMs) to this task. Existing methods are limited to either fine-tuning on structured data or prompting LLMs, each with trade-offs between semantic coherence and spatial-temporal precision. In this work, we propose a taxonomy of LLM-based trajectory generation approaches based on interaction modality, knowledge injection granularity, and controllability. Building upon the insights, we propose a hierarchical LLM agent workflow under the orchestrator-worker architecture for controllable trajectory generation, where the orchestrator first produces activity chains with contextual reasoning and the workers refine them into complete trajectories with a workflow that injects fine-grained spatiotemporal knowledge via a series of prefixed tools.

External IDs:doi:10.1145/3764925.3770908