Large Language Model-Enhanced General Transportation Agent Framework for Human Mobility Forecasting and Synthetic Travel Survey Data Generation

This work presents the LLM-enhanced general transportation agent, a novel framework that leverages large language models (LLMs) to simulate individual-level human mobility behavior. Unlike prior approaches that treat LLMs as generic predictors or planners, this framework employs role-play prompting and synthetic sociodemographic profiles to position LLMs as simulated individuals responding to household travel surveys. The system integrates population synthesis, persona-rich prompting, structured response tools, a dynamic survey engine, and an LLM-as-a-judge evaluation method to generate and vet context-aware, realistic behavioral data. Case studies in Chicago, USA and Lyon, France demonstrate cross-linguistic adaptability and behavioral fidelity, while highlighting challenges in non-English contexts. Quantitative evaluation shows that midsized and smaller models most accurately reproduce empirical travel survey distributions, whereas larger instruction-tuned models generate more coherent and naturalistic responses. Smaller models exhibit greater susceptibility to logical inconsistencies and stereotyped language, emphasizing potential bias propagation in synthetic datasets. The framework enables scalable, model-agnostic synthetic mobility data generation and supports applications such as scenario prototyping, discretionary activity simulation, travel diary construction, and integration into agent-based and microsimulation transportation models.