Simulating Public Administration Crisis: A Novel Generative Agent-Based Simulation System to Lower Technology Barriers in Social Science Research

TL;DR

This work introduces GABSS, a novel generative agent-based simulation framework built on GPT-3.5 to model public administration crises within a small-town setting. By assigning independent cognitive cycles to agents and grounding interactions in natural language, the approach enhances interpretability and flexibility beyond traditional ABMs while enabling controlled public-event insertion. Experimental results show memory-driven differences in decision-making, social networks, and rumor propagation under crisis conditions, supporting the potential of GABSS for pre-surveys and policy evaluation in computational social science. The authors also discuss limitations, including reasoning degradation, pre-training effects, and evaluation challenges, and outline avenues for deeper integration with social science research.

Abstract

This article proposes a social simulation paradigm based on the GPT-3.5 large language model. It involves constructing Generative Agents that emulate human cognition, memory, and decision-making frameworks, along with establishing a virtual social system capable of stable operation and an insertion mechanism for standardized public events. The project focuses on simulating a township water pollution incident, enabling the comprehensive examination of a virtual government's response to a specific public administration event. Controlled variable experiments demonstrate that the stored memory in generative agents significantly influences both individual decision-making and social networks. The Generative Agent-Based Simulation System introduces a novel approach to social science and public administration research. Agents exhibit personalized customization, and public events are seamlessly incorporated through natural language processing. Its high flexibility and extensive social interaction render it highly applicable in social science investigations. The system effectively reduces the complexity associated with building intricate social simulations while enhancing its interpretability.

Figures (11)

• Figure 1: Architecture of Single Generative Agent.
• Figure 2: Decision Making Structure.
• Figure 3: Hierarchical Structure of Social Environment.
• Figure 4: Township Demographics.
• Figure 5: Township Demographics.