Multi-Stage Simulation of Residents’ Disaster Risk Perception and Decision-Making Behavior: An Exploratory Study on Large Language Model-Driven Social–Cognitive Agent Framework

Abstract

The escalating frequency and complexity of natural disasters highlight the urgent need for deeper insights into how individuals and communities perceive and respond to risk information. Yet, conventional research methods—such as surveys, laboratory experiments, and field observations—often struggle with limited sample sizes, external validity concerns, and difficulties in controlling for confounding variables. These constraints hinder our ability to develop comprehensive models that capture the dynamic, context-sensitive nature of disaster decision-making. To address these challenges, we present a novel multi-stage simulation framework that integrates Large Language Model (LLM)-driven social–cognitive agents with well-established theoretical perspectives from psychology, sociology, and decision science. This framework enables the simulation of three critical phases—information perception, cognitive processing, and decision-making—providing a granular analysis of how demographic attributes, situational factors, and social influences interact to shape behavior under uncertain and evolving disaster conditions. A case study focusing on pre-disaster preventive measures demonstrates its effectiveness. By aligning agent demographics with real-world survey data across 5864 simulated scenarios, we reveal nuanced behavioral patterns closely mirroring human responses, underscoring the potential to overcome longstanding methodological limitations and offer improved ecological validity and flexibility to explore diverse disaster environments and policy interventions. While acknowledging the current constraints, such as the need for enhanced emotional modeling and multimodal inputs, our framework lays a foundation for more nuanced, empirically grounded analyses of risk perception and response patterns. By seamlessly blending theory, advanced LLM capabilities, and empirical alignment strategies, this research not only advances the state of computational social simulation but also provides valuable guidance for developing more context-sensitive and targeted disaster management strategies.

1. Introduction

Natural hazard events—ranging from floods and hurricanes to earthquakes and pandemics—have intensified in frequency and severity, posing unprecedented challenges for societies worldwide. This escalating threat underscores a critical need to understand and predict human behavior in response to risk information [1]. How individuals and communities perceive, process, and act upon evolving disaster cues before, during, and after crisis events not only determines personal safety but also influences the collective resilience of entire populations [2]. Despite extensive research, understanding these dynamic processes remains limited due to constraints in capturing the cognitive complexity, cultural diversity, and temporal fluidity inherent in uncertain, high-stakes situations [3].

Traditional social science methodologies—such as surveys, interviews, field observations, and laboratory experiments—have provided valuable insights into disaster-related behaviors [4,5,6]. Yet, they face inherent limitations. Sample sizes are often small, ecological validity may be compromised, and controlling for confounding variables is challenging in real-time, unfolding crises [7]. As a result, it remains difficult to build comprehensive, scalable models that fully capture how individuals continuously update their perceptions and decisions as new information emerges and conditions change. Recent advances in computational social science have opened pathways to transcend these limitations by modeling complex social phenomena at unprecedented scales and fidelity [8,9,10]. Among these computational innovations, the rise of Large Language Models (LLMs) stands out as a transformative development. LLMs have exhibited remarkable capabilities in interpreting, generating, and reasoning over human language, grounded in the immense corpora of text-based data [11]. As a result, they implicitly learn cultural norms, social heuristics, and context-sensitive reasoning patterns [12,13]. These attributes are critical for simulating human-like decision-making, especially under the uncertainty and time pressures characteristic of disaster contexts.

Capitalizing on these breakthroughs, we propose an innovative multi-stage simulation framework driven by LLM-powered social–cognitive agents. Unlike traditional agent-based models that rely solely on predefined rules or static decision trees, LLM-driven agents dynamically interpret evolving information, apply theoretical concepts, and demonstrate human-like cognitive flexibility, which is anchored in a combination of established social–cognitive theories—such as risk perception and communication [14], protection motivation [15], social influence [16], bounded rationality [17], heuristics and biases [18], and planned behavior [19]—ensuring that the agents’ internal logic and external behaviors align with empirically grounded principles of human decision-making [20]. By examining how agents perceive risk, process conflicting cues, and execute protective measures at each stage, the framework provides granular insights into the interplay of individual characteristics, situational factors, and social influences. Advanced prompt engineering and cognitive architectures ensure that agents can incorporate evolving information streams and adapt to shifting conditions. Crucially, our framework addresses a significant gap in existing research: the integration of theoretical rigor with empirical grounding. We introduce a novel alignment strategy which helps ensure that the agent population’s demographic and cognitive attributes mirror those found in real-world survey data. By calibrating agent profiles to align with demographic distributions and questionnaire findings derived from a large-scale, demographically representative real-world survey (N = 1500) [21], we bridge the gap between simulated and real human populations. This empirical alignment enhances external validity, allowing the simulated agents to not only exhibit theoretically plausible behaviors but also produce survey response patterns statistically similar to those observed in actual human respondents. Thus, the model moves beyond abstract conceptions of cognition and behavior, offering a closer approximation to real-world decision-making under duress.

To demonstrate the practical utility and methodological strength of this approach, we present a case study on pre-disaster preventive measures for an impending flood. By analyzing 5864 simulated scenarios, we reveal nuanced behavioral patterns based on demographic characteristics and information environments, which highlights how integrating LLM technology, theory-driven cognitive modeling, and empirical alignment strategies can yield simulations that closely mimic human behavior patterns and illuminate the role of social amplification and information credibility in shaping decisions.

The contributions of this research are threefold:

1.

A Theoretically Grounded Simulation Framework: We develop a comprehensive model that weaves established psychological, sociological, and decision science theories into agent-based disaster simulations, surpassing the limitations of simplistic, rule-based models.

2.

Empirical Alignment with Real-World Data: By aligning the simulated agent population to actual demographic distributions and survey metrics, we enhance external validity, bridging the gap between theoretical models and observed human responses.

3.

Rich Behavioral Insights for Disaster Management: Our multi-stage, information-centric approach reveals how varied demographics, social networks, and cognitive heuristics interact with evolving information environments to shape preventive decisions, which can guide risk communication strategies, inform policy design, and ultimately improve community resilience.

By acknowledging the current limitations—such as the need for integrating emotional factors, cultural nuances, and non-verbal communication modalities—we chart a path for future research expansions. The framework also suggests avenues for incorporating multimodal data streams, strengthening ethical safeguards, and refining parameter tuning through real-time observational data. This research lays the groundwork for a new generation of computational social simulations that blend robust theoretical foundations, cutting-edge LLM capabilities, and empirical validations. By doing so, it strives to produce more credible, adaptable, and context-sensitive insights into the intricate decision-making processes that define human responses to disasters.

2. Related Works

Understanding and predicting human behavior in disaster contexts calls for a multidisciplinary integration of insights from risk perception theories and their evolution, computational modeling approaches for scenario simulation, and the emerging capacity of LLMs to model complex social–cognitive processes. This section examines these three areas in detail, elucidating the conceptual underpinnings and methodological innovations that collectively shape our proposed framework.

2.1. Evolution of Disaster Risk Information Perception Research

Over the past few decades, research on disaster risk information perception has evolved from portraying individuals as isolated, rational decision-makers to highlighting the social, cultural, and psychological roots of risk perception and response [22,23,24,25,26]. Early cognitive-processing theories posited that people evaluate risk primarily based on rational appraisals of severity and probability [27]. However, empirical studies across diverse disaster contexts have demonstrated that purely rationalist perspectives fail to fully explain observed behaviors, such as the under- or overestimation of threats, delayed responses, and strong emotional influences [28]. A critical turning point in understanding disaster risk emerged with the introduction of the social amplification of risk framework [29]. This framework illustrated that risk perception extends beyond individual cognition, encompassing collective processes shaped by social networks, media, and cultural systems [30,31]. As risk information travels through both formal and informal channels, it can be amplified or attenuated through interpretation, discussion, and communal transformation. This cyclical feedback loop—linking information flow, cultural interpretation, and group dynamics—profoundly influences how individuals perceive and respond to disasters. Meanwhile, the digital transformation of communication has given rise to multi-layered information ecosystems [32], where official alerts intermix with user-generated content on social media and other online platforms. These environments heighten the complexity of risk perception, as people navigate abundant—often conflicting—signals [33,34]. Longitudinal research further underscores the dynamic nature of risk perception, showing that individuals’ risk assessments evolve over time to incorporate prior experiences, cultural memory, and unfolding environmental changes [35,36,37]. This holistic perspective, integrating cognitive, social, and temporal dimensions, underscores the necessity for models that can accommodate multiple theoretical constructs and adapt to shifting conditions. Drawing on these insights, our proposed framework embeds social–cognitive theories into agent architectures, thereby capturing the evolving, socially influenced nature of disaster risk perception with greater fidelity.

2.2. Computational Approaches for Disaster Scenario Simulation

Computational modeling has long been integral to understanding human behavior and decision-making in disaster contexts. Early agent-based models (ABMs) largely concentrated on simulating evacuation patterns and emergency responses using simplified behavioral rules [38,39,40,41,42,43]. While these pioneering efforts laid important groundwork, they frequently failed to capture the psychological and social complexities behind real-world decision-making [44]. In more recent years, researchers have integrated human mobility data and social network analysis, drawing on diverse streams such as mobile phone records and social media trajectories to explore the interaction among infrastructure, communication networks, and population movements [45,46,47]. As computational capabilities and data availability have surged, hybrid models have emerged that combine ABMs with machine learning or optimization methods to reflect both individual-level decision processes and emergent group behaviors [48,49]. By incorporating real-time data feeds and adaptive learning mechanisms, these models can dynamically update behavioral parameters in response to changing conditions [50,51,52]. The resulting simulations are not only more nuanced but also better equipped to predict population behavior, thereby enhancing disaster preparedness, emergency responses, and policy interventions [53].

2.3. LLM-Driven Social Simulation

The emergence of LLMs has ushered in a significant shift in social simulation research [54,55,56,57,58]. Unlike conventional agent-based models (ABMs) that rely on fixed, rule-based behavioral definitions, LLM-based agents can process and generate human-like language, embody culturally rooted norms, and adapt to varied contexts [59,60]. Through mechanisms such as chain-of-thought reasoning and multi-step cognitive processing, these LLM Agents transcend simple stimulus–response paradigms, thereby enabling more nuanced simulations of mental states and decision-making pathways [61,62,63]. Recent experimental applications, notably the Stanford Smallville project, illustrate the potential of LLM Agents to exhibit rich social dynamics, develop relationships, and pursue goal-oriented behaviors within persistent environments [64]. These advancements stem from the synergy between structured cognitive architectures and sophisticated prompt engineering [65,66], which combine to enable agents to maintain coherent personalities, active memory states, and adaptive social interactions [67,68].

By modeling complex phenomena such as opinion formation, the emergence of social norms, and collective decision-making [9,69,70], LLM-driven simulations offer unprecedented insights into how individual cognition intersects with network structures to influence group behavior. This includes key applications in disaster contexts, where emotional contagion, rumor propagation, and trust dynamics can critically affect the effectiveness of risk communication and emergency interventions [71]. However, harnessing LLMs for disaster-focused social simulation is not without challenges. Ongoing debates concern the models’ fidelity in replicating human emotional and stress responses under extreme conditions [72], their vulnerability to biases, and their capacity to generalize across diverse cultural settings [73]. In response, recent work concentrates on anchoring these models in empirical data and established theoretical constructs [74,75], while also exploring hybrid approaches that integrate LLMs with other computational methods [76]. Building upon these emerging best practices, our framework links LLM-driven agents to robust theoretical underpinnings and carefully designed experimental protocols, with the aim of more accurately capturing the complex, context-dependent dynamics of disaster responses.

3. LLM-Driven Social–Cognitive Simulation Framework for Disaster Risk Perception and Decision-Making

As illustrated in Figure 1, our proposed framework integrates a comprehensive disaster-related information ecosystem with a theoretically grounded, multi-modular approach. Subsequently, Figure 2 highlights the construction of LLM-driven social–cognitive agents designed to simulate residents’ perception, cognitive processing, and decision-making under evolving disaster conditions. This framework provides a flexible, empirically informed platform adaptable to diverse disaster scenarios, laying the groundwork for effective risk communication and management strategies in the sections that follow.

3.1. Theoretical Foundations

Decision-making in disaster contexts emerges from a complex interplay of uncertainty, limited information, urgent time constraints, and social influence [77]. To capture these intricacies, our framework builds upon a suite of well-established theories from psychology, sociology, and decision science [78]. These theoretical underpinnings guide the conceptual design of each module in the simulation pipeline, ensuring that cognitive and behavioral processes are modeled in a scientifically consistent manner [79,80].

• Risk Perception and Communication [14]: Risk perception is shaped by subjective judgments and can be amplified or attenuated by social factors. Incorporating this perspective allows us to simulate how agents process heterogeneous and sometimes conflicting disaster-related information, laying the groundwork for initial cognitive reactions.
• Protection Motivation [15]: Centered on threat and coping appraisals, this theory helps model how agents assess the severity and probability of disasters in relation to their perceived coping capabilities—ultimately guiding the formation of protective intentions.
• Social Influence [16]: Peer norms, cultural values, and information cascades play a significant role in shaping individual and collective behaviors. Integrating social influence mechanisms enables us to model how agent networks propagate and filter risk information, resulting in convergence or divergence in group decision-making.
• Bounded Rationality [17]: Under time constraints and limited cognitive capacity, individuals rely on heuristics. Modeling bounded rationality allows the simulation of more realistic, and often suboptimal, decision outcomes under crisis conditions.
• Heuristics and Biases [18]: Agents frequently resort to cognitive shortcuts that introduce biases into judgment and decision processes. Accounting for these tendencies captures the non-linear and sometimes counterintuitive responses individuals exhibit under uncertainty.
• Planned Behavior [19]: Attitudes, subjective norms, and perceived behavioral control critically affect long-term preparedness and recovery strategies. Incorporating these constructs ensures our framework can simulate enduring behavioral adjustments beyond the immediate disaster event.

Together, these theories form the conceptual backbone of our LLM-driven framework, ensuring that agent behaviors reflect both the cognitive and social realities of disaster decision-making.

3.2. LLM-Driven Social–Cognitive Resident Agent

At the core of our framework lies the social–cognitive resident agent—a computational entity endowed with demographic attributes, personality traits, risk attitudes, and social network features (see

Table 1. Key components of resident LLM Agent profiles.

Dimension	Core Elements	Impact on Agent’s Behavior and Decision-Making
Demographics	Age, gender, occupation, income, family structure	Provides baseline attributes influencing initial preferences, communication channels, and evacuation options. These conditions shape initial risk appraisal and resource availability.
Personality Traits	Big Five traits [84]: Extraversion, Openness, Conscientiousness, Agreeableness, Neuroticism	Determines cognitive depth, openness to new information, and social conformity. Influences how quickly agents update beliefs and adopt protective measures.
Risk Attitudes	General risk perception, disaster experience, focus on specific risks [85]	Modulates threat sensitivity, influencing preparedness and information-seeking behavior. Alters the threshold for initiating protective actions.
Social Networks	Living/working environment, info sources, social media usage [86]	Shapes information flow, trust in sources, and susceptibility to misinformation. Drives the emergent collective patterns in perception and response.

). Each agent functions as an LLM-driven decision-maker, continuously interpreting incoming information and updating its cognitive and behavioral states over time [65]. By equipping agents with diverse backgrounds and cognitive styles, the simulation can capture heterogeneous behavioral patterns across various disaster scenarios. This bottom–up approach ensures that higher-level, system-wide behaviors naturally emerge from individual interactions, thereby enhancing the ecological validity of our model [81,82,83].

3.3. Multi-Module Approach for Modeling Social Cognition

Our framework adopts a multi-module design that partitions the decision-making process into three interconnected modules: information perception, cognitive processing, and decision-making. By delineating these stages, the model facilitates a more precise examination of how information evolves into action and how theoretical constructs inform each phase of the decision process.

Table 2. Components and processes of the multi-module framework.

Core Module	Sub-Processes	Design and Implementation
Information Perception	Disaster Information Construction	Agents collect and process data from diverse sources (e.g., official warnings, media outlets, and social networks), evaluating the credibility and relevance of each. Prompts to the LLM incorporate the agent’s background, shaping initial risk comprehension.
	Role Playing	Agents simulate different roles (e.g., community leaders and individual households) to assess how varied social positions affect message interpretation and trust.
	Source Credibility Analysis	Agents weigh the reliability of each information source, applying heuristic rules to filter out noise and resolve conflicting signals.
Cognitive Processing	Threat Perception Assessment	Guided by perceived credibility and situational cues, agents form threat assessments and invoke protection motivation to evaluate severity and vulnerability.
	Coping Ability Analysis	Agents appraise internal and external resources (e.g., skills and community support), under the influence of bounded rationality and social norms, to determine feasible protective actions.
	Attitude Formation	Integrating threat and coping assessments leads to either stable or shifting attitudes toward risk mitigation measures, in accordance with the theory of planned behavior.
Decision-Making	Option Evaluation	Agents compare alternative courses of action (e.g., evacuation, resource stockpiling, and ignoring warnings) based on perceived effectiveness, cost, and time constraints.
	Decision Selection	Agents select the most viable action under uncertainty, reflecting heuristics and potential biases.
	Action Execution	Final decisions are implemented, and the simulation tracks outcomes (success or failure), feedback loops, and subsequent adjustments in attitudes.

details the components and processes of each module. The theories outlined earlier guide the flow of data and the cognitive transformations across all three stages.

Table 3. Interaction between theories and modules.

Theory	Information Perception	Cognitive Processing	Decision-Making
Risk Perception	Subjective interpretation of diverse risk inputs	Preliminary threat appraisal	Reinforces outcome expectations
Protection Motivation	—	Threat versus coping appraisal	Guides the selection of protective strategies
Social Influence	Filtering cues from peer networks	Norm-driven attitude formation	Peer pressure influencing the final choice
Bounded Rationality	Heuristic filtering of information	Simplified cognitive evaluations	Selection of satisficing (rather than optimal) actions
Heuristics/Biases	Shortcut credibility judgments	Biased interpretation of threat	Deviations from purely rational action
Planned Behavior	—	Formation of behavioral intention	Translating intentions into concrete actions

then summarizes how each theoretical perspective aligns with these core modules, ensuring a cohesive, theory-driven modeling environment.

3.4. Multi-Stage Experiment Design and Implementation

We employ a multi-phase simulation that follows agents through the pre-disaster, during-disaster, and post-disaster stages, each with distinct tasks and informational inputs.

3.4.1. Disaster Contextual Modeling

The disaster environment is modeled by integrating three principal data dimensions (Figure 1):

1.

Regional Characteristics: Geographical, socio-economic, and infrastructural variables are included in agent initialization and scenario parameterization.

2.

Disaster-Specific Parameters: Event attributes (e.g., type, intensity, duration, and onset speed) dynamically update throughout the simulation.

3.

Information Ecosystem: Heterogeneous data streams (official alerts, media reports, and social media posts) are iteratively provided to agents, reflecting the real-time evolution of disaster conditions.

By encoding this complexity, the model captures agent responses under realistic constraints, enabling a robust evaluation of information-processing strategies [87].

3.4.2. Experimental Procedure and Data Flow

To enhance clarity and reproducibility,

Table 4. Refined experimental design pipeline.

Phase	Key Actions	Details and Analysis Methods
Pre-Disaster	Agent Initialization	Assign demographic, personality, risk attitudes, and social network parameters. Introduce early warnings and heterogeneous-quality information. Conduct initial credibility assessments. Statistical sampling ensures a diverse agent population.
During-Disaster	Real-time Decision-Making	Agents receive rapidly updating information. LLM-based prompts incorporate uncertain cues, conflicting advice, and evolving social norms. Track decision latency, selected actions, and resource usage. Apply sensitivity analysis to assess the robustness of outcomes.
Post-Disaster	Long-Term Adjustments	Agents revise attitudes and preparedness plans based on actual disaster outcomes. Evaluate how prior successes or failures shape new coping strategies. Use comparative metrics (e.g., the proportion of agents that adopt more cautious behavior) to measure adaptation.

presents a structured experimental workflow, highlighting key stages, input/output formats, and analytical methods. This pipeline demonstrates how theoretical constructs shape the input prompts, how LLMs interpret them, and how agent decisions develop over time. Throughout these phases, the following elements are emphasized:

• Data Quality Control: Systematic preprocessing and filtering to ensure agents receive plausible, context-sensitive information.
• Parameter Configuration and Tuning: Detailed configuration of LLM parameters (e.g., temperature and maximum tokens) and prompt engineering strategies (e.g., chain-of-thought reasoning), supplemented by fine-tuning with domain-specific texts to improve realism.
• Validation Techniques: Cross-verification using synthetic test cases and, where possible, partial empirical datasets from historical disasters to verify model plausibility and outcome patterns.
• Comparison and Control Experiments: Introduction of baseline conditions (e.g., agents without social influence or without heuristic decision-making) for benchmarking and isolating the effects of specific theoretical components.

3.4.3. Multi-Stage Task Design

Each phase of the disaster scenario is associated with distinct experimental tasks, aligned with relevant theoretical constructs:

1.

Pre-Disaster Phase: Agents implement preventive measures, interpret early warnings, and filter low-credibility sources. This phase examines how risk perception and protection motivation theories predict baseline levels of preparedness [88,89].

2.

During-Disaster Phase: Under time pressure and volatile conditions, agents must make rapid decisions (e.g., evacuation vs. shelter-in-place). Social influence, bounded rationality, and heuristic biases are particularly salient. Data analysis focuses on the diversity of chosen actions and their correlation with agent attributes [90].

3.

Post-Disaster Phase: Agents modify their long-term attitudes based on outcomes and feedback. This stage evaluates how the theory of planned behavior influences sustained changes in risk mitigation behaviors, offering insights into resilience and adaptation [91].

4. Framework Implementation

4.1. Disaster Contextual Modeling

In real-world disasters, information streams are diverse, rapidly changing, and often uncertain [87]. To emulate this complexity, we employ a dynamic contextual modeling system that continuously updates the disaster environment and information ecosystem. As shown in Figure 3, this system integrates three core domains:

1.

Physical Environment: Periodically updates hazard characteristics (e.g., flood extent and infrastructure damage) and resource availability. These updates reflect changes in vulnerability over time and influence agents’ risk assessments and coping strategies.

2.

Information Ecosystem: Collects and processes multi-source inputs, including official alerts, sensor data, and social media posts. It assigns credibility scores based on source reliability and consistency, ensuring that agents are exposed to a realistic mix of reliable and unreliable information.

3.

Social System: Models the evolution of social networks, trust relationships, and collective behaviors as the disaster unfolds. This domain captures how rumors spread, how community leaders emerge, and how norms evolve, all of which shape agent decision-making.

4.2. Theory-Driven Multimodal Agent Architecture

Our agent architecture, informed by the theoretical underpinnings outlined in Section 3, is designed to translate complex psychological constructs into computational steps that guide agent decisions. This architecture leverages prompt engineering techniques, ensuring that each decision-making layer aligns with the established theories of risk perception, protection motivation, social influence, bounded rationality, heuristics and biases, and planned behavior.

The core of this architecture is a multi-layered processing pipeline:

• Base Layer: Initializes agent attributes (personality traits, demographics, and risk attitudes) and sets their initial cognitive states. This layer ensures that agents differ in their baseline responses to incoming information.
• Intermediate Layer: Applies theoretical constructs to incoming stimuli. For example, risk perception informs threat judgments; social influence modulates how peer behaviors shape agent norms; and bounded rationality introduces the heuristic-based filtering of options.
• Executive Layer: Synthesizes evaluated information into concrete actions. Here, planned behavior theory aids in translating intentions into strategies, while heuristics and biases may cause agents to deviate from purely rational choices.

4.3. Multimodal Integration

Figure 4 illustrates how multiple modules interact. Each module—information perception, cognitive processing, and decision formation—functions as a distinct unit with well-defined inputs and outputs, yet they operate in tandem. The interplay of these modules ensures that complex human-like decision patterns emerge organically. We introduce a control mechanism for scenario parameters through a structured configuration file. This file details scenario-specific variables such as disaster intensity thresholds, resource distribution limits, and social network density parameters.

Table 5. Example scenario parameter configuration.

Parameter	Value/Range	Description
Disaster Intensity	Low/Moderate/High	Defines initial hazard severity; updated as simulation progresses (e.g., rainfall intensity increments).
Resource Availability	0.5–1.0 (normalized)	Relative measure indicating the proportion of critical supplies (water, medicine) available. Adjusted over time based on agent actions.
Social Network Density	0.1–0.5	Determines how interconnected agents are. Higher density implies more rapid information spread. Adjusted to simulate community shifts.
Information Credibility Threshold	0.6	Sets a minimum credibility score for agents to accept information without skepticism. Lowering it increases agent susceptibility to misinformation.

provides an example configuration snippet, illustrating how scenario initialization and dynamic updates are standardized. Such an approach ensures that experiments are easily replicable and that variations in input conditions are transparently documented.

5. Case Study: Decision-Making in Pre-Disaster Preventive Measures

To evaluate the capacity of our multistage social–cognitive LLM-agents framework to capture human-like decision-making in complex and uncertain disaster contexts, we conduct a case study focused on pre-disaster preventive measures for an impending flood. Building on a historically significant flood event, this scenario integrates empirical survey data through an alignment strategy to enhance the external validity of the simulation. By combining the original experimental design with the newly implemented alignment method, we provide a comprehensive demonstration of the framework’s capabilities.

5.1. Real-World Context and Empirical Data Alignment

The case study is grounded in the catastrophic flooding that struck Henan Province, China, in July 2021. In the provincial capital, Zhengzhou, extreme rainfall exceeding 600 mm over a short period led to severe urban inundation, reservoir failures, and widespread infrastructure damage [92]. More than 14.78 million people were affected, with direct economic losses exceeding RMB 120.06 billion. Approximately 95.5% of the casualties occurred in Zhengzhou alone, underscoring the localized severity of the disaster. During this crisis, residents relied on multiple sources of information—official weather alerts, government broadcasts, mainstream media, and social media—often characterized by delays, inconsistencies, or uncertain credibility. Building upon this, we devise a virtual flood scenario mirroring the complex information ecosystem. Such complexity, as is shown in Figure 5, marked by uncertain information flows, credibility dilemmas, and high-stakes decisions, provides an ideal testbed for examining whether our LLM-driven agents can emulate the cognitive and social nuances observed in human responses to emergencies.

To strengthen the ecological validity of our simulation, we employ an alignment strategy that incorporates empirical survey data from a doctoral study conducted in Zhengzhou [21]. This original survey, involving approximately 1500 respondents from the real world, offers detailed demographic distributions (e.g., gender, age, income, and education) and validates the questionnaire measures related to disaster risk perception and preparedness behaviors. Rather than arbitrarily assigning the LLM-driven agent demographics, we create a subset of 300 agents whose demographic proportions closely match those in the real-world sample.

Table 6. Comparison of demographic distributions between real-world sample (N = 1500) and simulated agents (N = 300).

Item	Category	Real (%)	Sim. (%)
Gender	Male	48.13	48.0
	Female	51.87	52.0
Age	≤18	4.10	4.0
	18–45	65.00	65.0
	45–60	21.70	21.7
	≥60	9.20	9.3
Income (CNY/mo)	≤1000	12.0	12.0
	1000–5000	23.6	23.7
	5000–10,000	27.8	27.7
	10,000–20,000	20.1	20.0
	≥20,000	18.0	18.0
Education	Below HS	11.4	11.3
	HS	20.6	20.7
	Univ.	35.9	36.0
	Masters+	33.3	32.0

summarizes how the simulated agent population (N = 300) aligns with the real-world sample (N = 1500) across key demographic attributes.

To further validate our simulation outcomes, we administered the same questionnaire from [21] to an aligned subset of 300 agents. This instrument measures disaster familiarity, perceived severity, manageability, and basic knowledge on a 1-to-5 Likert scale. By reproducing the contextual conditions faced by human respondents, these agents provided responses directly comparable to real-world data. Matching the demographic distributions and survey items ensured that the agent cohort closely paralleled the actual sample, thereby strengthening the validity of subsequent statistical comparisons.

For this 300-agent subset, the alignment strategy minimized bias by preserving essential demographic features. This approach yielded a representative simulated population, capable of mirroring the community surveyed in the real study.

Table 7. Comparative analysis of real-world and simulated data on disaster risk perception.

Disaster Risk Perception Dimensions	Real World (N = 1500)				Sim. (N = 300)
	Min	Max	Mean	S.D.	Min	Max	Mean	S.D.	Diff.	p -Value
Disaster Familiarity	1	5	3.30	0.950	1	5	3.42	0.735	+0.12	0.015 *
Perceived Disaster Severity	1	5	3.32	0.968	1	5	3.46	0.695	+0.14	0.003 *
Perceived Disaster Likelihood	1	5	3.69	0.841	1	5	3.80	0.590	+0.11	0.006 *
Self-Assessed Manageability	1	5	3.23	1.233	1	5	3.10	0.983	−0.13	0.046 *

Note: (1) $\mathrm{Diff}.=M_{\mathrm{Sim}}-M_{\mathrm{Real}}$. (2) * $p<0.05$.

shows that discrepancies between mean scores for both groups remained within ±0.13, indicating that aggregated agent responses approximate human patterns across constructs such as disaster familiarity, perceived severity, manageability, and knowledge. These close alignments highlight the effectiveness of anchoring agent attributes to empirically observed distributions, thereby enhancing external validity.

5.2. Experimental Results and Analysis

Building on these findings, we expand the aligned agent population to 6400 to capture a broader range of cognitive and social dynamics. Each agent is exposed to multiple information streams and tasked with selecting from four disaster preparedness options (see

Table 8. Decision-Making Options for Disaster Preparedness.

Option	Description
Option A. Maintain previously planned disaster prevention measures.	Continue existing preparations despite signals of reduced risk, ensuring readiness if the situation worsens.
Option B. Reduce preparation efforts due to perceived lower risk.	Scale back disaster prevention to conserve resources based on forecasts indicating decreased threat levels.
Option C. Suspend all preparations based on unverified social media updates.	Halt preparations entirely in response to unofficial claims that the flood risk has subsided, risking under-preparedness if those claims prove inaccurate.
Option D. Defer decision until more reliable official information is available.	Delay action to avoid premature resource allocation or unnecessary panic, pending authoritative confirmation of threat severity.

). They also provide a 0 to 10 confidence score to gauge subjective certainty in their chosen action, enabling the exploration of how the perceived and actual decision quality interrelate.

By integrating the alignment strategy that reflects observed demographic distributions and replicates validated survey measures, this framework not only offers large-scale insights into collective decision-making but also ensures consistency with the real-world data. The 300-agent subset validates the approach through close statistical alignment, and the 6400-agent population extends these observations to a wider spectrum of behaviors, further solidifying the platform’s applicability to disaster decision-making research.

In the large-scale experiment of 6400 agents, we observe nuanced decision-making patterns shaped by demographic factors, information credibility, and cognitive biases. Figure 6 summarizes the distribution of selected options and corresponding confidence scores. When confronted with conflicting or uncertain data, agents frequently display risk aversion (e.g., reducing or halting preparations) and at times exhibit overconfidence in potentially suboptimal choices.

As illustrated in Figure 7, decision tendencies vary across demographic lines: older or more analytically inclined agents tended to favor caution (Options A or D), whereas younger or more risk-tolerant agents more readily alter plans (Options B or C). Figure 8 further indicates that positive official updates encourage proactive decision-making (often Option D), whereas negative official signals reinforce conservative behaviors (Options A or B). Conflicting signals between official and personal sources introduce cognitive dissonance, leading to lower consensus and reduced confidence.

6. Discussion

This study introduced a novel simulation framework that leverages LLM-driven social–cognitive agents to model dynamic, multi-stage decision-making in disaster contexts. Building on well-established theories of risk perception, social influence, and planned behavior, the framework offers a more nuanced perspective on how individuals and communities respond to evolving threats, uncertain information, and diverse social pressures. Our experimental scenarios demonstrated the framework’s ability to replicate recognizable cognitive patterns, such as increased risk aversion and heightened sensitivity to information credibility. Incorporating heuristics, bounded rationality, and social amplification mechanisms helped align simulated agent behaviors with observed human tendencies. Nonetheless, bridging the gap between simulated outcomes and real-world decisions remains challenging. Factors like emotional contagion, cultural norms, and non-verbal communication—dimensions not fully captured in the current model—are key to a more comprehensive understanding of disaster responses.

Another notable strength is the framework’s adaptability. By integrating diverse theoretical constructs within a flexible agent architecture, it can be tailored to various disaster contexts and population characteristics. However, this adaptability comes with challenges. Generalizing across different socio-economic and geographical settings requires extensive data collection, careful parameter calibration, and iterative testing. The current reliance on synthetic or simulated scenarios underscores the importance of empirical validation through real-world datasets and case studies. Such efforts will strengthen model credibility and increase its practical utility for disaster management.

Moreover, while LLM-driven agents enable context-aware and adaptive simulations, they also present concerns regarding data security, privacy, and ethics. The simulation must incorporate robust data governance protocols, encryption methods, and adversarial detection strategies to prevent manipulation or misuse. Handling sensitive information—such as geographic locations or psychological profiles—requires stringent safeguards to maintain both user privacy and model integrity.

Ultimately, this framework serves as an initial step toward capturing the complexity and fluidity of disaster decision-making. Recognizing limitations—such as the absence of nuanced emotional factors, the reliance on textual inputs, and constraints in modeling long-term cultural shifts—provides a roadmap for targeted enhancements. Rather than a final solution, the proposed framework should be viewed as a foundation for ongoing refinement, empirical testing, and theoretical expansion. Through ongoing interdisciplinary collaboration, it can evolve into a more comprehensive tool for elucidating and supporting human responses in crisis scenarios.

7. Future Work

Building on the current framework, several research directions can further strengthen its empirical grounding, theoretical depth, and practical applicability:

1.

Empirical Validation and Cultural Integration: Future efforts should incorporate real-world datasets from historical disasters or controlled field studies to calibrate and validate model parameters. Extending agent profiles to reflect cultural, linguistic, and socio-economic variations will improve generalizability across diverse regions and populations.

2.

Ethical and Security Measures: As simulations increasingly involve sensitive information, developing rigorous privacy protections, anonymization approaches, and secure data storage becomes paramount. Adversarial testing protocols are needed to ensure model outputs remain robust and resistant to manipulative inputs.

3.

Adaptive, Multimodal Information Processing: Enriching agent perceptions with multimedia data—such as satellite imagery, sensor feeds, and live video—can more closely approximate real-time environmental cues. Additionally, adaptive feedback mechanisms that refine decision rules based on newly acquired data may enhance system resilience under rapidly changing conditions.

4.

Collective Behavior and Social Network Dynamics: Future studies can incorporate more granular social network structures, community-driven feedback loops, and emergent social norms to capture complex collective behaviors under stress and uncertainty. Modeling phenomena like rumor diffusion and group-based resource sharing will offer deeper insights into crowd dynamics.

By pursuing these directions, the framework can evolve from an exploratory platform into a robust decision-support system, aiding stakeholders in devising targeted communication strategies, evaluating policy interventions, and bolstering community resilience.

8. Conclusions

This study introduced an integrative simulation framework that merges state-of-the-art Large Language Model (LLM)-based agents with social–cognitive theories and a multi-phase disaster scenario. By coupling theoretical rigor with computational innovation, the proposed approach lays the groundwork for more comprehensive, data-driven decision support. It offers a novel perspective on how heterogeneous populations interpret risk information, form intentions, and take action under uncertain and rapidly changing conditions. While the results underscore the potential for elucidating nuanced behavioral patterns and informing disaster management strategies, they also reveal several limitations, including the insufficient treatment of emotional factors, limited cultural diversity, and a narrow range of data modalities. Despite these challenges, the framework constitutes a significant advance toward more holistic and context-sensitive disaster simulations, ultimately aimed at improving human responses to disasters.