In agent‑based simulations of multimodal urban mobility, traditional approaches let agents (i.e., simulated residents) decide travel modes using discrete‑choice models, machine‑learning models fitted on structured travel‑survey data, or expert‑written rules. These methods fix the agents' behavior space beforehand, so factors not encoded in variables or rules cannot influence decisions. In reality, mode choice is a multidimensional decision where objective and subjective factors intertwine and evolve over time, making it hard to embed this complexity into city‑scale simulations.
Generative agents (GAs) built on large language models (LLMs) carry decision heuristics that surveys do not capture, offering a possible breakthrough. However, aligning their decisions with real household travel surveys remains an open question.
This paper builds a daily multimodal mobility model for Toulouse, France, and evaluates fifteen decision makers on the actual household travel survey, including:
- Simple baselines (random, most common mode, etc.)
- Machine‑learning models fitted on the survey
- GAs with minimal and tuned prompts
- The newly released Jev‑1.13 TypeSafe classifier
Results show that, in a simulated day without transport disruptions, tuned‑prompt GAs approximate the modal split of fitted models, yet they lag behind on a trip‑by‑trip basis. GAs also respond sensibly to unforeseen events. Surprisingly, the TypeSafe classifier achieves excellent matching without generating text, indicating that text generation is not a prerequisite for accurate decisions.
These findings are relevant to transport policymakers and service providers, suggesting that generative agents can improve urban mobility services while staying aligned with real survey data.
Review: The study highlights the promise of generative agents for urban mobility simulation, especially the closeness achieved through prompt tuning, but further work is needed to enhance trip‑level accuracy and interpretability.