CityReal framework aligns LLM agents with real human urban behavior

CityReal, a new framework by Nicolas Bougie, Xiaotong Ye, and Narimasa Watanabe, tackles a core flaw in urban simulation: LLM agents often reproduce the model's behavioral priors instead of the target population's actual behavior. Previous approaches relied on few-shot prompting, which made agents act like generic ChatGPT characters rather than real city residents. CityReal replaces this with an intention-driven architecture where agents form coherent mobility and activity plans, then adapt over time by learning habits and preferences from their experiences and constraints. This makes their decisions look less like isolated step-by-step choices and more like genuine human routines.

The framework's key innovation is learning textual adapters for each behavior module. These adapters tune agent decisions to match observed population statistics, improving realism at both micro and macro levels. In experiments, CityReal outperformed baseline methods in aligning with real-world human behavior. Crucially, it scales to tens of thousands of agents, enabling researchers to simulate and analyze crowd density, place popularity, mobility flows, and even well-being across different urban scenarios. For urban planners, transport engineers, and AI researchers, CityReal provides a far more reliable testbed for forecasting how cities respond to policy changes, new infrastructure, or disruptive events—without needing real-world trials.