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

“The development of large language models (LLMs) such as Transformer [38], XLNet [39], and GPT-3 [9] has enabled powerful natural language processing capabilities like text generation, translation, and question answering.”

“This framework leverages large language models’ semantic capabilities [51] to mimic human cognition.”

“The development of large language models (LLMs) such as Transformer [38], XLNet [39], and GPT-3 [9] has enabled powerful natural language processing capabilities like text generation, translation, and question answering.”

“• Used text-based Json files to input different character settings and public events, enabling the simulation of any time-based event and changed the homogeneity of the traditional ABM mathematical model [22, 23].”

“Recent advances like chain-of-thought prompting have improved LLM reasoning abilities [46], enabling applications like GABM which embeds GAs into epidemic models [47].”

“However, concerns remain about potential biases, reproducibility, interpretability, and ethics of both LLMs and GAs [48, 49].”

“Other challenges include system and simulation homogeneity [32, 33], uncertainty, and high computational cost [34].”

“Limitations include memory degradation and excessive rigidity when fine-tuned [16, 45].”

“Social information transmission greatly influences human social cognition [52–54].”

“This has sparked interest in applying LLMs to computational social science (CSS) tasks [12].”

“While Generative Agent society simulations cannot fully supplant social research towards real-world [41, 42], this work holds utility for pre-surveys or preliminary policy evaluations.”

“Social information transmission greatly influences human social cognition [52–54].”

“Other challenges include system and simulation homogeneity [32, 33], uncertainty, and high computational cost [34].”

“However, node clustering coefficients decrease [72, 73] because, in scale-free networks, average internode closeness negatively correlates with clustering coefficient.”

“This potential has attracted attention in computational social science [10].”

“S OCIAL science research methods often simplify and abstract social phenomena to establish simplified representations [1].”

“Our generative agent model can exhibit thinking patterns analogous to humans [15].”

“While Generative Agent society simulations cannot fully supplant social research towards real-world [41, 42], this work holds utility for pre-surveys or preliminary policy evaluations.”

“In the ClusterRank model [74], as node clustering coefficient declines and average internode closeness rises, final rumor propagation scope increases.”

“In this article, we introduce a novel generative agent-based social simulation (GABSS) paradigm to simulate public crisis events in a small town [17].”

“Additionally, while the LLM used targeted removal of harmful speech during training [10, 58, 59], to ensure diverse generative agent populations resembling small town residents, we de-“harmless” the LLM [60].”

“LLMs also generate synthetic social media content that is difficult for humans to distinguish from real content [43].”

“Additionally, while the LLM used targeted removal of harmful speech during training [10, 58, 59], to ensure diverse generative agent populations resembling small town residents, we de-“harmless” the LLM [60].”

“In the original incident, governmental inaction precipitated a crisis of confidence among resident agents [70].”

“Thus, water pollution has become a prevalent crisis for current township administrations due to its rapid onset and regional stability, as noted in related literature [61–66].”

“We design a town based on the administrative structure of American townships [18], set up its infrastructure [19–21], and place more than a hundred generative agents in it.”

“The concept of ‘social simulacra’ using large language models (LLMs) to generate human-like conversational agents was proposed by Park et al. (2022) as a way to design computational social systems [11].”

“ABMs have been influential in the development of computational social science [18] and can reveal and explain unpredictable social phenomena [25].”

“Additionally, while the LLM used targeted removal of harmful speech during training [10, 58, 59], to ensure diverse generative agent populations resembling small town residents, we de-“harmless” the LLM [60].”

“Despite their potential for informing policy-making [29, 30], ABMs can be difficult to interpret in natural language [1, 31].”

“With the rise of computer simulations in both qualitative and quantitative social science research [2, 3], computational models have become essential components of social simulation [4].”

“Thus, water pollution has become a prevalent crisis for current township administrations due to its rapid onset and regional stability, as noted in related literature [61–66].”

“To address these challenges, researchers have proposed using agents with diverse personalities and heterogeneous system construction [35, 36], even retaining individual radical views.”

“ABMs developed rapidly in the 1990s with the advent of large-scale computing [27] and are now widely used in sociology and economics research [28].”

“The agent settings also reference the demographic structure and characteristics of similar societies [56].”

“Despite their potential for informing policy-making [29, 30], ABMs can be difficult to interpret in natural language [1, 31].”

“3) Rumor Propagation Coefficient: R denotes the rumor propagation coefficient at time t , N ( t ) signifies the number of individuals who heard the rumor at time t , and N represents the total network population (including rumor originators) [75].”

“• Used text-based Json files to input different character settings and public events, enabling the simulation of any time-based event and changed the homogeneity of the traditional ABM mathematical model [22, 23].”

“ABMs developed rapidly in the 1990s with the advent of large-scale computing [27] and are now widely used in sociology and economics research [28].”

“Limitations include memory degradation and excessive rigidity when fine-tuned [16, 45].”

“Subsequent work has built on this to develop ’Generative Agents’ (GAs) that can simulate realistic social behavior [16, 44, 45].”

“However, node clustering coefficients decrease [72, 73] because, in scale-free networks, average internode closeness negatively correlates with clustering coefficient.”

“We design a town based on the administrative structure of American townships [18], set up its infrastructure [19–21], and place more than a hundred generative agents in it.”

“Recent research has evaluated LLMs on diverse CSS tasks and found they can expand CSS capabilities despite lacking professional vocabulary knowledge [40].”

“Future work could employ LLM cascade models [76] to further reduce computational costs.”

“But they may help address limitations of traditional agent-based modeling like agent homogeneity [12, 13].”

“Multi-agent systems, a breakthrough in computational modeling, represent distributed artificial intelligence systems consisting of interacting agents in virtual environments [5–7].”

“The experiment references the 2014 Flint water crisis in Michigan [68, 69].”

“We utilize the open-source Omnievent toolkit for event extraction from large corpora [67].”

“However, concerns remain about potential biases, reproducibility, interpretability, and ethics of both LLMs and GAs [48, 49].”

“We establish an independent cognitive system for each agent, simulating human memory and decision-making processes [14, 24, 50].”

“Recent advances like chain-of-thought prompting have improved LLM reasoning abilities [46], enabling applications like GABM which embeds GAs into epidemic models [47].”

“Rumor graphs typically exhibit higher average degrees than normal networks [71], indicating individuals have more connections on average.”

“To address these challenges, researchers have proposed using agents with diverse personalities and heterogeneous system construction [35, 36], even retaining individual radical views.”

“Despite their potential for informing policy-making [29, 30], ABMs can be difficult to interpret in natural language [1, 31].”

“Traditional social simulation models constructed from mathematical models such as graph theory, differential equations, and statistics are difficult to express in natural language [1], and their interpretability and practicality vary [8].”

“The population configuration of this virtual society references the demographic ratios of archetypal island towns [55]: young adults aged 18-35 comprise 34.9%, middle-aged residents aged 35-60 constitute 43.1%, and the elderly populace is relatively small.”

“Per the typical township administrative model [57], there are only two full-time civil servants - the sheriff and township head.”

“With the rise of computer simulations in both qualitative and quantitative social science research [2, 3], computational models have become essential components of social simulation [4].”

“We design a town based on the administrative structure of American townships [18], set up its infrastructure [19–21], and place more than a hundred generative agents in it.”

“Finally, Section VI concludes the paper.”

“By June 2023 , the emphasis changed to multi-stage scenarios, incorporating multi-step tasks that enabled agents to engage in sequential decision-making and adaptive responses across varied contexts to achieve the more complex goal [165, 182, 190, 192].”

“…as motion planning (Mandi et al., 2024), simulations of human behavior (Park et al., 2023; Hua et al., 2024), the creation of gaming enviroments (Hu et al., 2024), recommender systems (Zhang et al., 2024a), simulations of financial trading (Li et al., 2023), and policy making (Xiao et al., 2023).”

“This often results in independently constructed task-specific simulation pipelines performing repetitive work, which reduces their generalizability and scalability [26, 58].”

“…et al. , 2023; Qian et al. , 2023 ] , multi-robot systems [ Mandi et al. , 2023; Zhang et al. , 2023c ] , society simulation [ Park et al. , 2023; Park et al. , 2022 ] , policy simulation [ Xiao et al. , 2023; Hua et al. , 2023 ] , and game simulation [ Xu et al. , 2023c; Wang et al. , 2023c ] .”

“…– achieving a shared goal - and adverserial – debate and competition – settings [2], with applications including debating [16], policy making [17, 18], economy [19, 20], epidemic modeling [21], (online) society simulation [14, 22, 23], psychology [24, 25], gaming [26, 27], software…”

“For instance, leveraging LLMs for role-playing can be used to accomplish tasks such as software development[14, 36], societal simulation[34, 35], policy simulation[51, 15], game simulation[52, 49] and video generation[53].”

“Such interfaces have been successfully applied across a range of domains, including software development (Qian et al., 2023), games (Park et al., 2023), robotics (Zhang et al., 2023), economics (Li et al., 2023), political science (Xiao et al., 2023), and public health (Williams et al., 2023).”

“Sepa-rately, they’ve been used to model developmental psychology (25; 61), game theory (32), macroeco-nomics (29; 63), social policies (41; 54; 19), and community dynamics (40; 39; 10).”

“Moreover, LLMs are increasingly being used in multi-agent settings (de Zarzà et al., 2023; Park et al., 2023; 2022; Xiao et al., 2023; Hua et al., 2023; Vezhnevets et al., 2023; Chuang et al., 2023) and are already interacting with one another as chatbots on social media 1 .”

“Beyond text analysis, these models are leveraged for annotating images (Zhang & Pan, 2019), and conducting social simulations (Bail, 2023; Xiao et al., 2023).”

“Xiao et al. [57] introduced a GABSS framework to simulate public administration crises, demonstrating how LLM-driven agents can adapt and respond to events like water pollution incidents based on memory and interactions.”

“Recent advances have demonstrated the ability to deploy LLMs as agents [2] in specific scenarios to provide more specialized and insightful advice for these specific domains, which is di ff erent from traditional agent architecture [3], [4].”

“LLM-MA have shown great promise in software engineering [10, 33], robotics [34], policy and society simulation [35, 36], and game environments [37].”

“Additionally, some research explores the social aspects of collaborative multi-agent systems, including the use of “social simulacra” to simulate realistic interactions within a virtual community of agents [19, 20].”

“The second category investigates specific social phenomena, such as public administration crises (Xiao et al. , 2023), health policy (Hou et al. , 2025), political manipulations (Touzel et al. , 2024), financial market (Gao et al. , 2024b), and deduction games (Xu et al. , 2025).”

“…in across diverse fields: we encounter them in robotics [23], [24], [25], [26], [27]; in game-playing for adaptive decision-making, [28], [29], [30], in software development [31], [32], in simulations of human behavior [33], [34], financial trading [35] policy-making [36] and negotiations [37].”

“In the fields of social simulation and policy simulation, works by Park et al. [8], Xiao et al. [13], and Hua et al. [3] demonstrated how multi-agent systems use the expertise and interaction mechanisms of various agents to simulate complex social behaviors and policy decisions.”

“In such systems, multiple LLM-based agents exchange information, coordinate strategies, and collectively generate solutions, often outperforming single agents across a variety of tasks and fields, such as building robots, creating public policies, and software engineering [2]–[6].”

“…from thousands of LLM-driven agents, are increasingly used to investigate emergent social phenomena such as polarization, cooperation, and cultural diffusion (Piao et al. 2025a, b; Tang et al. 2025; Wang et al. 2024; Xiao et al. 2023; Yan et al. 2024; Yang et al. 2025; Zhang et al. 2025).”

“Xiao et al. (2023) proposed the Generative Agent-Based Simulation System (GABSS) specifically designed to emulate human cognition, memory, and decision-making frameworks within virtual social systems, highlighting capabilities for personalized agent customization and natural language processing…”