Leveraging Large Language Models for Agent‐Based Simulation of Human‐Water System Interactions

Abstract

Modeling human–water system interactions is essential for understanding the co‐evolution of human society and hydrological extremes. Agent‐based models (ABMs) are widely used for this purpose, yet they face persistent challenges, including manually coded rules, simplified behavioral assumptions, and limited integration of qualitative information. Recent advances in large language models (LLMs) offer promising opportunities to address these challenges by simulating agents' decisions in natural language reasoning and guided by behavioral theory frameworks. This study presents a proof‐of‐concept LLM‐ABM framework that models household flood adaptations including buying insurance, elevating homes, relocating, or doing nothing in a synthetic flood‐prone city. Agents (i.e., households) are initialized with trust levels and narrative‐style memory. Protection Motivation Theory (PMT) is structured to guide decision‐making as an example. Results show that LLM‐driven ABM generates memory‐mediated, narrative‐driven behavioral adjustments, in contrast to the threshold‐driven, utility‐mediated shifts produced by a traditional stochastic ABM, illustrating how architectural choices shape emergent behavioral dynamics. Year‐by‐year natural‐language appraisals and explicit reasoning can enhance the interpretability and diagnostic visibility of LLM‐driven ABM, which allows scientists to trace how threat perception, coping beliefs, or social cues potentially shape individual decisions. Sensitivity analyses demonstrate that LLM‐driven ABM can evaluate uncertainty across new qualitative and parameter dimensions, including prompt phrasing and memory‐window length. While limitations remain including LLM reliability and architecture choices, output versus process transparency, simplified agent population, and simulation scalability, the findings in this paper illustrate the potential of LLM‐driven ABMs as a complementary tool for socio‐hydrological research.