Urban-simulation study tests whether LLM agents move like real people

A new arXiv paper tests whether LLM-based generative agents in urban simulations reproduce realistic human mobility patterns. The authors compare agent behavior with real-world mobility data from the Greater Paris region and Shanghai.

The finding is that plausible movement narratives do not necessarily translate into empirically realistic mobility. The paper evaluates mobility laws, temporal rhythms, network motifs, semantic transitions, and behavioral profiles.

That matters for planners and researchers using LLM agents in simulations. If the movement looks believable but fails against real data, downstream policy or design conclusions could be misleading.