Hierarchical Generative Agents create realistic disaster evacuation simulations

Accurate modeling of human behavior during disasters is critical for effective evacuation planning, but collecting real-world data is extremely difficult and existing models assume rational, homogeneous actors—leading to overly optimistic predictions. To bridge this gap, researchers Maria Mendoza, Lucas Waldburger, Jin Lee, and Shankar Sastry from UC Berkeley have developed Hierarchical Generative Agents, a new simulation framework that uses persona-conditioned large language models (LLMs) paired with a cognitive module. The framework simulates sequential decision-making in a grid-based urban environment that evolves over time, capturing dynamic hazards like fire.

The agents are structured with a three-level cognitive hierarchy: high-level evacuation goals, mid-level route reasoning, and low-level navigation. Each agent's decisions are driven by LLMs and calibrated with empirical human evacuation data, allowing the simulation to reflect real, non-rational behaviors—including panic, hesitation, and social influence. This approach addresses a key limitation in current computational evacuation models by introducing heterogeneity and realistic response to environmental stimuli. The 20-page paper (with 6 figures) demonstrates how these cognitive AI agents can generate actionable insights for disaster response teams, urban planners, and emergency managers looking to improve resource allocation and evacuation strategies.