Large Language Models for Human Mobility Generation: Methods, Applications and Open Issues

Accurate human mobility data are crucial for urban planning, transportation management, and disaster preparedness, yet traditional data sources often face limitations due to privacy issues, cost, and data sparsity. Recently, large language models (LLMs) have emerged as promising tools for generating synthetic human mobility data. This paper reviews state-of-the-art LLM-based methodologies, categorized into training-from-scratch, hybrid fine-tuning, and prompt-constrained autoregression, highlighting their respective strengths and limitations. We further compare various spatial tokenization approaches—including grid-based, road network-based, POI-based, coordinate quantization, and hierarchical encoding—and examine practical applications such as event-driven forecasting, missing-data imputation, and interactive scenario simulations. Finally, we discuss remaining challenges and future directions concerning spatial consistency, computational efficiency, evaluation standardization, multi-resolution modeling, and human-AI collaborative frameworks. We provide practical insights to advance realistic yet privacy-preserving human mobility simulations.