Generative agents in crowd simulation: a cognitive approach with Large Language Models

Abstract

Crowd simulation is a powerful tool used in fields such as urban planning, emergency response, and entertainment to model and predict human movement and behavior in various scenarios. As society becomes increasingly complex and interconnected, the need for simulations that accurately capture human behavior at both the individual and group level grows. Understanding these interactions can help institutions and experts develop more effective mitigation strategies in dynamic social environments.
This research explores the potential of the power of Large Language Models (LLMs) in crowd simulations, leveraging their capabilities to model individual behavior and enable the emergence of realistic crowd dynamics through agent-level interactions.
We propose a novel architecture that integrates key cognitive components – such as perception, planning, memory, reflection, and action – on an algorithmic level. This approach allows generative agents to process environmental and social contexts in a human-like manner. Our findings show that these agents exhibit diverse and contextually appropriate behaviors, closely resembling human decision-making, particularly in crisis situations.