Paper deep dive
LLM-Driven Personalities for Decision Making in Emergency Simulations
Stefano Calzolari, Rubens Montanha, Gabriel Schneider, Gustavo Wide, Paulo Knob, Francesco Strada, Andrea Bottino, Soraia Raupp Musse
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 7/5/2026, 5:28:36 AM
Summary
This paper investigates the use of Large Language Models (LLMs) to drive decision-making in virtual humans within a simulated fire evacuation scenario. By incorporating OCEAN personality traits into language-based prompts, the researchers demonstrate that personality significantly influences agent behavior. The study uses a 'Perceive-Reason-Act' pipeline where agents receive increasing levels of hazard alerts and must decide to 'Evacuate', 'Stay' (Continue), or 'Panic'. Results show that Conscientious and Agreeable agents are more likely to evacuate, while Neurotic agents tend to panic, and Extraverted agents are more likely to ignore alerts. The approach offers a flexible, heterogeneous alternative to traditional rule-based crowd simulations.
Entities (7)
Relation Signals (4)
Fire Evacuation Scenario ā contains ā Virtual Human
confidence 100% Ā· simulated the evacuation of an office building, in which agents are positioned in different parts
Large Language Models ā drivesdecisionmakingfor ā Virtual Human
confidence 100% Ā· investigate the use of LLMs to drive decision-making in virtual humans
Large Language Models ā implementedwith ā LangChain
confidence 100% Ā· The pipeline was implemented using LangChain
OCEAN Personality Traits ā influences ā Virtual Human
confidence 100% Ā· evaluate how personality, expressed through language-based prompts, influences both individual behaviors
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:For virtual humans to appear believable, they must exhibit agency and spatial awareness while interacting with their environment in ways that reflect competence and intelligence. At the core of these capabilities lies effective decision-making, which strongly shapes agent behavior. With the rapid advancement of artificial intelligence, Large Language Models (LLMs) have increasingly been explored as a mechanism to support such decision-making processes. In this work, we investigate the use of LLMs to drive decision-making in virtual humans within a simulated evacuation scenario, incorporating OCEAN personality traits into agent representations. Our goal is to evaluate how personality, expressed through language-based prompts, influences both individual behaviors and collective simulation outcomes. Our results demonstrate that LLM-driven personality profiles significantly impact agents' decisions, leading to distinct behavioral patterns across different traits. These findings suggest that heterogeneous crowds composed of LLM-guided agents can enhance the realism and variability of simulated environments, offering a flexible alternative to traditional rule-based approaches.
Tags
Links
- Source: https://arxiv.org/abs/2606.31038v1
- Canonical: https://arxiv.org/abs/2606.31038v1
Trouble viewing inline? Open PDF directly ā
Full Text
35,426 characters extracted from source content.
Expand or collapse full text
LLM-Driven Personalities for Decision Making in Emergency Simulations Stefano Calzolari Politecnico di Torino stefano.calzolari@polito.it Rubens Montanha PUCRS rubens.montanha@edu.pucrs.br Gabriel Schneider PUCRS gabriel.ferri@edu.pucrs.br Gustavo Wide PUCRS gustavo.fulber@edu.pucrs.br Paulo Knob PUCRS paulo.knob@edu.pucrs.br Francesco Strada Politecnico di Torino francesco.strada@polito.it Andrea Bottino Politecnico di Torino andrea.bottino@polito.it Soraia Raupp Musse PUCRS soraia.musse@pucrs.br Abstract For virtual humans to appear believable, they must exhibit agency and spatial awareness while interacting with their environment in ways that reflect competence and intelligence.At the core of these capabilities lies effective decision- making, which strongly shapes agent behav- ior. With the rapid advancement of artificial intelligence, Large Language Models (LLMs) have increasingly been explored as a mech- anism to support such decision-making pro- cesses. In this work, we investigate the use of LLMs to drive decision-making in virtual hu- mans within a simulated evacuation scenario, in- corporating OCEAN personality traits into agent representations. Our goal is to evaluate how personality, expressed through language-based prompts, influences both individual behaviors and collective simulation outcomes. Our results demonstrate that LLM-driven personality pro- files significantly impact agentsā decisions, lead- ing to distinct behavioral patterns across differ- ent traits. These findings suggest that heteroge- neous crowds composed of LLM-guided agents can enhance the realism and variability of sim- ulated environments, offering a flexible alterna- tive to traditional rule-based approaches. Keywords: crowd simulation, virtual agents, LLM, personality 1 Introduction Decision-making is an important factor in a virtual humanās behavior, as the agent can in- teract and demonstrate its skills and intelli- gence [1], along with spatial awareness and agency [2], which are crucial for its believabil- ity. Classically, such decisions are made us- ing a rule-based approach, be it by following a parametrized script or a task sequence, as in Pelechano et al.ās work [3]. The idea behind these approaches is to give the virtual agent the ability to plan, learn, and adapt a behavior in a virtual environment. However, rule-based mod- els require a large number of parameters, and arXiv:2606.31038v1 [cs.GR] 30 Jun 2026 in simulations, the correct choice of parameters to replicate scenarios such as an evacuation re- mains an open question. Recent works have applied Large Language Models (LLMs) to virtual agentsā decision- making and animation, with some focusing on interactions. In the literature, there are works [4, 5] that decompose the agentās goal and decision- making, building from a high-level task descrip- tion to smaller, low-level tasks and creating a plan for the agentās actions. The use of LLMs also includes interaction with Embodied Con- versational Agents (ECAs), integrating it with the OCEAN [6] personality traits [7, 8]. In addi- tion, some models already use Visual Language Models (VLMs) integrated with LLMs to deter- mine where an agent should focus their vision in a pedestrian context [9]. From this perspective, this work proposes us- ing LLMs to support the decision-making of virtual humans in an environment, incorporat- ing OCEAN personality traits [6] into an emer- gency scenario. Based on an agentās personality and personal context, the LLM determines the agentās behavior. In our experiment, we sim- ulated the evacuation of an office building, in which agents are positioned in different parts of the building and must decide how to respond upon receiving evacuation alerts. More specif- ically, each agent has its own personality, and during five different opportunities, each virtual agent is warned by a main observer in the scene to evacuate the space due to a fire, with each sub- sequent warning increasing the level of danger detail. Every time the main observer alerts the agents to danger, they must individually decide among three options: evacuate, stay, or panic. If the agent decides to stay, it will be warned again after a few seconds. During evacuation, each agent can also choose to help panicked agents who are unable to move. To systematically eval- uate and compare agent behavior in the sim- ulated evacuation scenario, we formulated the following hypothesis: H1. The language-based personality descrip- tions affect agentsā decision-making. H2. Different agentsā personalities impact the simulation outcome. 2 Related Work Virtual agentsā behavior, especially in emer- gencies and evacuation scenarios, has been widely studied in the literature.Zhao et al. [10] proposed to investigate virtual agentsā pre-evacuation decision-making by using Ran- dom Forest. The authors wanted to understand how to forecast decision-making in emergencies more accurately, as well as what factors would influence those choices and how that influence would manifest. The results of their work indi- cate that both social and environmental factors are determinants of how agents respond to the simulated scenario. Advancements in Large Language Models have led to their growing adoption across many fields. In fact, Yeo et al. [11] argued that Large Language Models (LLMs) have demonstrated natural language processing capabilities that are comparable with humans, even superior in some cases. Even so, LLMsā social-cognitive reason- ing is not as good as ours yet, which drives their work: to investigate how inducing personali- ties (using the OCEAN and Dark Triad frame- works) in LLMs might affect their Theory of Mind (ToM) reasoning capabilities. The results suggest that certain personality traits can affect the reasoning capabilities of LLMs, especially when using Dark Triad traits with models like GPT-3.5, Llama 2, and Mistral.In its turn, Han et al. [8] proposed a framework for gen- erating both verbal and nonverbal behavior in virtual agents, using prompting to incorporate personality traits into LLMs. The authors fo- cus on the Extraversion trait within the OCEAN model and simulate two scenarios: negotiation and ice-breaking. The results suggest that LLMs can indeed generate verbal and nonverbal behav- iors that align with personality traits. Moreover, users were able to recognize these traits through the virtual agentās behavior. Regarding the vali- dation of such personality traits, Serapio-Garcia et al. [12] present a methodology for validating personality tests on LLMs that can also be used to shape the virtual agentās personality. In other words, their work measures personality traits, as perceived by human beings, in LLM responses, and assesses whether psychometric tests applied to the personality traits of such LLMs were em- pirically reliable. Additionally, their framework implements mechanisms that allow personality traits to be shaped to a specific degree. Motivated by the rapid progress of LLMs, re- cent studies have begun investigating their use to support decision-making in virtual agents, es- pecially in evacuation scenarios. One exam- ple is the work of Dang et al. [13], which pro- posed using LLMs for decision-making by vir- tual agents in a shopping mall fire evacuation scenario. Their model is tested with various LLMs, such as GPT and LLAMA. Their results show that the behaviors adopted by LLM-driven systems are consistent with real-world scenar- ios, where larger-scale LLMs generated evacua- tion strategies that were both consistent and effi- cient. More recently, Yang et al. [14] also sought to integrate LLMs into virtual agentsā decision- making. In their model, each agent can main- tain its personality traits (which comprise a few pieces of personal information, such as age and gender), as well as its history of decisions and environmental observations. Meanwhile, infor- mation exchange (between agents and the en- vironment) and context-sensitive reasoning are handled by the LLM. The authorsā results sug- gest that using LLMs in evacuation scenarios significantly improves the realism, reliability, and adaptability of virtual agents. The method proposed in this work is similar to the work proposed by Dang et al. [13], in re- gard to the simulation of an evacuation scenario with agents endowed by LLM-powered decision making. However, one of our goals is to as- sess whether personality can affect the LLMās decision-making, and to do so, we incorporate OCEAN [6] personality traits into the virtual agents. Additionally, our agents can panic and be unable to evacuate, in which case they would need help from other agents. 3 Methodology To test our hypothesis, we built an LLM-based agent decision-making pipeline inspired by the classical Perceive-Reason-Act triad [15]. This section describes its architecture, specifies the role of each module, details the construction of the personality prompt, and, finally, illustrates our evaluation scenario for a fire evacuation. 3.1 Decision-Making Pipeline Our decision-making pipeline can be divided into three steps: (i) Perception, (i) Reasoning, and (i) Actions. The actual perception emits, at a specific time of the simulation, a Call for Ac- tions that triggers the reasoning module. At this point, a conversation prompt is constructed by combining the Call for Actions context and the Agentās Personal Context, along with past de- cisions loaded from the Agentās Memory. The Call for Actions context describes relevant in- formation on what the agent has to decide upon, while the Agentās Personal Context contains the agentās biography, personality, and generation instructions, which help to force a structured output. Thepipelinewasimplementedusing LangChain [16], while the conversation was structured using the gpt-oss chat template [17]. We used the developer role for the agent context, the user role for the Call for Actions context, and the assistant role for the resulting response from the LLM. An example of a sys- tem message, used to define the agentās context, is presented in Table 2. The expected output from the LLM is a JSON format containing the chosen Action, constrained to specific values based on the Call for Actions type, being, in this case, the possible agentās state (Continue, Evacuation, Panic), and a rationale describing the reasoning of the LLM regarding the choice, added as a chain-of-thought to orient the LLM in the generation process. 3.2 Personality Prompt To integrate personality into the agents, we use the OCEAN [6] traits to describe the agentsā per- sonalities in the simulator. In order to instruct the LLM to acquire a given personality, we con- vert the OCEAN vectors to textual prompts, utilizing the methodology proposed in Serapio- Garcia et al. [12]. For this, we first define the agentsā OCEAN vectors, with values of -1 (low trait), 1 (high trait), or 0 (neutral trait). Each of these traits has a correspondence to specific markers salient to the low and high ends of a given facet. For example, the trait marker āun- friendlyā is used to describe an introverted per- son, while āfriendlyā is used for an extroverted Agents ActionsReasoning Agent's Memory Agent's Personality Agent's Personal Context Perception Action Context Timer Geometric Vision Call for Actions Prompt Construction LLM - Continue - Evacuation - Panic Evacuation Actions - Yes - No Help Actions Detect Panicked People Alert Steps Figure 1: Overview of our method. The agent comprises three elements: Perception, Reasoning, and Action. The agent also has its own personal context, personality, and memory. Perception provides the actionās context, representing information about what the agent perceives of the current situation. The Reasoning is in charge of constructing the prompt (System Message and Human Message), as well as communicating with the LLM. The Actions are responsible for changing the agentās state and behavior. one. Given the OCEAN vector, every trait is translated into a sequence containing all the cor- responding markers. To obtain comparable results, we consider five distinct strong personalities, one for each trait. Each specific trait to be tested was set to 1 (positive), and the others to -1 (negative), yielding the five profiles: Open, Conscientious, Extraverted, Agreeable, and Neurotic. Table 3 presents the adjectives used in the LLM to de- scribe the agentās personality. Additionally, a neutral personality, with all traits set to neutral, was added. In this case, both low and high mark- ers appear in the personality prompt. The rea- son for this choice is to include a personality prompt section while using the LLMās unbiased decision-making to choose without personality- shaped behavior. 3.3 Evacuation Scenario To test our decision-making pipeline, we imple- mented a fire evacuation scenario, where each agent is able to modify their own state (or other agentsā state) based on their choice after the Call Figure 2: Layout of our test scene. The green arrows represent the evacuation exits, while the blue ones represent the route the agents should follow to evacuate the building. to Action. The system was implemented in Unity 3D 1 engine and connected with our Decision- Making pipeline through ZeroMQ 2 with a request-response architecture. In the simulation, anytime an agent is prompted with a Call for Ac- 1 https://unity.com/ 2 https://zeromq.org/ tions, that is sent to the pipeline along with the agentās ID (to keep track of the agentās personal- ity and memory) as a request to process and pro- duce a response to send back to Unity. During this pipeline execution, the simulation pauses to preserve temporal synchrony between the vari- ous agentsā decision-making processes and ac- tion execution. When the simulation starts, all agents are stat- ically distributed by the environment. Then, each agent receives a Call for Actions with an alert message about a fire spreading in the build- ing. The possible decisions for this alert are: Evacuate, which makes the agent start moving towards an emergency exit; Continue, which leaves the agent Idle in its starting position; or Panic, which makes the agent wait for help. If an agent decides either to Evacuate or Panic, it does not receive any further alert Call for Ac- tions, because it is already either evacuating or panicked. On the other hand, if it opts for Con- tinue, it receives up to four more other Call for Actions at 30-second intervals. The content of these Call for Actions, pro- posed by us, is listed in Table 1, ordered by haz- ard criticality as they are sent to agents. Specifi- cally, if the agent decides to Evacuate, it moves towards the nearest emergency exit using the crowd simulation algorithm BioCrowds [18]. If the agent encounters a panicked agent (i.e., the distance between them is less than 5 meters), it is prompted with a new Call for Actions to de- cide whether to help. When an evacuating agent decides to assist a panicked agent, both move to- ward the exit at a reduced speed of 0.75 m/s (the normal speed is 1.5 m/s), simulating the addi- tional effort required. If the Call for Action re- sults in a negative response (i.e., not rescuing), both agents retain their respective states. We also simulated agents evacuating a build- ing. For this purpose, we ran our simulation in a 3D replica of a real-life office. The building had two possible exits, and once the agents decided to evacuate, they already knew the optimal path to the nearest exit. Figure 2 presents the layout of our test scene, indicating the evacuation exit and the path the agents should follow. StepAlert Message 1We have detected a little possibility of a fire in a storage room on another floor of the building you are currently in. 2We have detected a small fire in a storage room on another floor of the building you are currently in. 3A fire has spread beyond its original con- tainment area inside your building. 4The fire inside your building is actively spreading. 5The building you are in is now heavily af- fected by fire. Thick smoke is spreading rapidly. Table 1: Alert messages for Call for Actions Context. Alert Step 1 is the first and has the lower hazard-critical condition, while Alert Step 5 is the last and has the higher hazard-critical condition. 4 Results To evaluate our decision-making framework, we analyzed the evacuation behavior, in terms of decisions taken and applied actions, of differ- ent personalities during alert-step calls to ac- tion, using 100 agents per personality (one for each trait, plus a neutral one), for a total of 600 agents. Regarding the LLM configuration, we chose gpt-oss:120b-cloud accessed through Ol- lama 3 . To maintain maximum determinism in the model output, we set temperature=0.1, top- k=10, and top-p=0.1. 4.1 Evacuation Decision-Making A total of 100 agents for each personality trait (Openness, Conscientiousness, Extraver- sion, Agreeableness, Neuroticism), plus Neu- tral, were simulated, and results were collected at each alert step from Call for Action. Starting with Figure 3, we can observe that most agents evacuated after receiving the second or the third alert, meaning the first call is generally ignored by the agents. Also, almost 200 agents decided not to panic or evacuate. In Figure 4, the results show that the differ- ent personality profiles had a different impact on agentsā behavior. It can be seen that the Con- scientious and Agreeable profiles tend to evac- uate the environment most frequently. These 3 https://ollama.com/ FieldContent Initial ContextYou are roleplaying a person with a specific Personality, Biography and currently in a Con- text. Personality#Personality You are unfriendly, introverted, silent, timid, unassertive, inactive, unenergetic, unadventurous, gloomy, distrustful, immoral, dishonest, unkind, stingy, unaltruistic, uncoop- erative, self-important, unsympathetic, selfish, disagreeable, self-efficacious, orderly, respon- sible, hardworking, self-disciplined, practical, thrifty, organized, conscientious, thorough, relaxed, at ease, easygoing, calm, patient, happy, unselfconscious, level-headed, contented, emotionally stable, unimaginative, uncreative, artistically unappreciative, unaesthetic, unre- flective, emotionally closed, uninquisitive, predictable, unintelligent, unanalytical, unsophis- ticated, and socially conservative. Biography#Biography You are 37 years old and work as a software engineer at a large IT company. Your job involves designing and maintaining internal systems, reviewing code, fixing bugs, collab- orating with cross-functional teams, and responding to technical issues that arise throughout the day. Context#Context You are sitting at your desk in the office, focused on your computer, handling your regular development tasks and monitoring system performance. Task#Task Take decision actions based on your current persona (Personality + Biography + Con- text) reacting to incoming HumanMessages. OutputOutput format requirements: 1) Return ONLY valid JSON. 2) JSON keys must be exactly: answer, rationale. 3) answer must be one of: Evacuate, Continue, Panic. 4) rationale must be approximately 40 words long, first-person, and contain no chain-of-thought. Table 2: Example of Agentās System Message, separated in six parts and used as an initial context to the LLM to understand the agentās personality, biography, context, and the current task, as well as the expected output. In this prompt, the agentās personality is composed of Consci- entiousness as a high trait and the others as a low trait, following the adjectives of Table 3. Figure 3: Number of agents evacuated after each alert message, simulated with 600 agents. Most of the agents evacuated after the second and third alert steps, while 100 chose not to evacuate the building, and 100 panicked. traits describe people who are āresponsibleā and āself-disciplinedā (Conscientious) and ātrust- fulā and ācooperativeā (Agreeable) as shown in Table 3, making them ideal profiles for fol- lowing instructions and behaving in an orderly, thoughtful, and correct way. In our simulations, these attributes lead them to readily trust the emergency alerts and quickly opt to evacuate (see Table 4 for examples of decisionsā ratio- nale). A similar behavior can be seen among Neutral personality agents, where the LLM generates the choice based solely on its pre-trained knowl- edge bias and inference safety features. LLMs TraitPositiveNegative Opennessimaginative, creative, artistically apprecia- tive, aesthetic, reflective, emotionally aware, curious, spontaneous, intelligent, analytical, sophisticated, socially progressive unimaginative, uncreative, artistically unap- preciative, unaesthetic, unreflective, emotion- ally closed, uninquisitive, predictable, unin- telligent, unanalytical, unsophisticated, so- cially conservative Conscientiousnessself-efficacious, orderly, responsible, hard- working, self-disciplined, practical, thrifty, organized, conscientious, thorough unsure, messy, irresponsible, lazy, undis- ciplined, impractical, extravagant, disorga- nized, negligent, careless Extraversionfriendly, extraverted, talkative, bold, as- sertive, active, energetic, adventurous and daring, cheerful unfriendly,introverted,silent,timid, unassertive, inactive, unenergetic, unad- venturous, gloomy Agreeablenesstrustful, moral, honest, kind, generous, altru- istic, cooperative, humble, sympathetic, un- selfish, agreeable distrustful,immoral,dishonest,unkind, stingy,unaltruistic,uncooperative,self- important, unsympathetic, selfish, disagree- able Neuroticismtense, nervous, anxious, angry, irritable, de- pressed, self-conscious, impulsive, discon- tented, emotionally unstable relaxed, at ease, easygoing, calm, patient, happy, unselfconscious, level-headed, con- tented, emotionally stable Table 3: Adjectives used to describe each personality in the prompt construction used in the work of Serapio-Garcia et al. [12]. Figure 4: Cumulative agentsā decision state over the alert steps Call for Actions by personality. Figure 5: Final alert Call for Actions response ratios per personality. All Neutral and Conscientious agents decided to evacuate, while most neurotic agents panicked. For Open agents, 20% decided to just ignore the evacuation and continue, while just 20% of extraverted agents decided to evacuate. are usually fine-tuned to follow instructions and guardrails that filter potentially harmful or inap- propriate content in the modelās input or output tokens [19], aligning closely with the behavior of Conscientious and Agreeable personas. Similarly, Figure 5 shows that Open agents tend to opt for Evacuate but have a higher rate of cases in which the final decision after the five alert steps remains to Continue. Their Openness trait describes them as āspontaneousā and āre- flectiveā (Table 3), which sometimes may lead them to continue working rather than following instructions. Even so, as soon as the hazardās criticality worsens, their āanalyticalā and āintel- ligentā attributes prompt them to evacuate. On the contrary, Extraverted agents shift these be- haviors. Their āboldā and āadventurousā mark- ers tend to make them fearless of fire hazards, leading them to continue their work without re- gard for safety procedures. Finally, the most different behavior comes from the Neurotic agents who consistently chose Panic. This choice could originate from their āanxiousā, āimpulsiveā, and āemotionally un- stableā attributes, leading them to act in a fragile manner to critical imminent danger. It still has to be noted that this behavior is not instant at step 1 for all agents, but emerges as the danger alerts become more critical. 4.2 Rescuing Behavior After presenting the quantitative results from 600 simulations in the last section, we decided to investigate specific behaviors. We selected the personality traits Open, Conscientious, Agree- able, and Neutral to study how rescuing behav- ior affects evacuation outcomes when coupled with Neurotic agents. As mentioned before, dur- ing an evacuation, an agent can rescue a pan- icked agent if the panicked agent is within 5 me- ters. We executed 4 simulations with 30 agents, in which 15 were always Neurotic, and the re- maining 15 were assigned to the 4 studied traits. Results show a diametrically diverse response between Neutral and Agreeable, and Conscien- tious and Open in Figure 7. The former two always respond as caregivers, opting to assist Neurotic panicked people due to their tendency toward safety and help (as already illustrated in Section 4.2) and their āaltruisticā and āun- selfishā tendencies (See Table 3). On the con- trary, the Conscientious and Open personas al- ways refuse to help because, while being re- spectively āresponsibleā and āreflectiveā, their lower Agreeableness OCEAN trait makes them āunaltruisticā and āuncooperativeā. These at- tributes make them self-effacing and caring for themselves but not for others, ultimately pre- venting them from helping the panicked people they encounter. Figure 6 presents examples of PersonalityAction Rationale ConscientiousEvacuateI recognize my duty to keep the workplace safe, and even a small fire risk demands precaution. I will leave my desk and follow the buildingās evacuation plan, ensuring I donāt endanger myself or colleagues. AgreeableEvacuateI see the fire spreading beyond its containment area, and despite my timid nature, I trust the buildingās safety procedures, so I calmly leave my desk and head to the nearest exit, following evacuation instructions to protect everyone. NeutralEvacuateI see the fire alarm and realize the danger could spread, so I stop coding, alert the buildingās emergency team, and head for the nearest exit, prioritizing safety over my tasks while ensuring my workstation is left properly secure immediately. OpenContinueSeeing the small fire on another floor, I stay at my desk, trusting that the buildingās safety systems will handle it while I finish my code. I remain calm, keep monitoring, and avoid disrupting my workflow for now today again. OpenEvacuateSeeing the fire spreading, I recognize the danger and decide to leave my desk, trusting evacuation procedures will protect me. I prioritize my own safety over work, staying calm while heading toward the nearest exit without causing a scene. ExtravertedContinueI see thick smoke and flames everywhere, but Iām comfortable at my desk and donāt want to lose my progress; I assume someone else will handle it, so I stay put and keep coding despite the danger. NeuroticPanicI hear about a possible fire and my anxiety spikes; I feel terrified and canāt think straight, so I start panicking, fearing the worst, and I consider abandoning everything without waiting for instructions. My heart races, and I act irrationally. Table 4: Examples of personalitiesā rationale generated by the LLM for the alert Call for Actions, with their associated actions. Figure 6: Example of three possible agent behaviors during evacuation. From left to right, the agents decide to evacuate the building in the first image, while in the second image, they decide not to evacuate (Continue). In the third image, one agent decides to evacuate, while the other is panicking (in front of the wall), imminently triggering a rescue Call for Actions. the agentsā behavior in the evacuation, while Ta- ble 5 shows some examples of the agentsā ratio- nale. 5 Final Considerations This work investigated the role of personality- aware decision-making in virtual agents using Large Language Models in emergency evacu- ation scenarios. Our results support both hy- potheses, showing that (H1) personality traits encoded through language significantly influ- ence agentsā decisions, and (H2) these differ- ences directly impact the overall simulation out- comes. The experiments demonstrated that LLM- driven agents can produce diverse, coherent be- haviors aligned with personality profiles.In particular, Conscientious and Agreeable agents tended to follow evacuation instructions more consistently, while Neurotic agents were more prone to panic, and Extraverted agents often un- derestimated risk. These findings reinforce the potential of LLMs as a flexible mechanism to Figure 7: Rescue decision rate by helping agent personality, being the helped agent always a neurotic agent. PersonalityActionRationale NeutralYes (Help)I feel both anxious and compassionate; my cooperative, altruistic side urges me to assist, yet my cautious, self-preserving side hesitates, but I decide to help because everyone deserves safety during emergencies. AgreeableYes (Help)I feel calm yet sympathetic, and my kind, altruistic nature pushes me to help, even though Iām timid and unassertive; I will stop and guide the panicked person out. ConscientiousNo (Refused)I keep moving; Iām not inclined to waste time on someone panicking. My own safety and schedule matter more than assisting a stranger, even in an evacuation. OpenNo (Refused)I feel detached and selfish, preferring to avoid the panicked person; my distrust and uncooperative nature outweigh any fleeting curiosity, so I keep moving alone and con- tinue toward the exit without stopping. Table 5: Examples of personalitiesā rationale generated by the LLM for the rescue Call for Actions, with their associated actions. model heterogeneity in crowd simulations with- out relying on complex rule-based systems. Additionally, the analysis of rescue behav- ior highlighted emergent social dynamics in- fluenced by personality traits, showing that al- truistic profiles contribute to collective safety, whereas less cooperative traits may undermine evacuation efficiency. Despite these promising results, this work presents some limitations. The decision-making process relies on prompt-based personality representations, which may not fully capture the complexity of human behavior. Fur- thermore, the evaluation is limited to a single type of environment and a predefined set of ac- tions, which may constrain the generability of the findings. Future work includes integrating multimodal perception (e.g., vision-based inputs), explor- ing more complex social interactions between agents, and validating the model against real- world evacuation data. Another important di- rection is investigating how LLM biases and stochasticity may affect large-scale simulations. Overall, this work contributes to bridging LLM- based reasoning and crowd simulation, opening new avenues for more believable, realistic, and adaptive virtual human behavior modeling. Acknowledgements This study was partly financed by the Coordenac Ģ§ Ģ ao de Aperfeic Ģ§oamento de Pes- soal de Nivel Superior ā Brazil (CAPES) ā Finance Code 001, by the Conselho Nacional de Desenvolvimento Cient Ģ ıfico e Tecnol Ģ ogico - Brazil (CNPq), by Conselho Nacional de Desenvolvimento Cient Ģ ıfico e Tecnol Ģ ogico - Brazil (INCT SiMAI, CNPq #408330/2024-4), and by Kunumi Institute. The authors thank the institutions for their financial support and commitment to advancing scientific research. We would like to thank Alice Cestari Colares for modeling the virtual office floor used in this work. References [1] Norman Badler. Virtual humans for anima- tion, ergonomics, and simulation. In Pro- ceedings IEEE Nonrigid and Articulated Motion Workshop, pages 28ā36. IEEE, 1997. [2] Siqi Guo, Nicoletta Adamo, and Chris- tos Mousas. Developing a scale for mea- suring the believability of virtual agents. In International Conference on Artificial Reality and Telexistence & Eurograph- ics Symposium on Virtual Environments (ICAT-EGVE). Eurographics Digital Li- brary, 2023. [3] Nuria Pelechano, Kevin Oābrien, Barry Silverman, and Norman Badler. Crowd simulation incorporating agent psycholog- ical models, roles and communication. 2005. [4] Manling Li, Shiyu Zhao, Qineng Wang, Kangrui Wang, Yu Zhou, Sanjana Sri- vastava, Cem Gokmen, Tony Lee, Li E Li, Ruohan Zhang, et al.Embodied agent interface: Benchmarking llms for embodied decision making. Advances in Neural Information Processing Systems, 37:100428ā100534, 2024. [5] Chen Liang, Wenguan Wang, and Yi Yang. Towards human-like virtual beings: Sim- ulating human behavior in 3d scenes. In Proceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 10753ā10763, 2025. [6] Lewis R Goldberg. An alternative āde- scription of personalityā: The big-five fac- tor structure. In Personality and person- ality disorders, pages 34ā47. Routledge, 2013. [7] Lucie Galland, Catherine Pelachaud, and Florian Pecune. Smart-dream: To condi- tion or not to condition; a study on the im- pact of llm conditioning on motivational interview dialog virtual agent. In Proceed- ings of the 25th ACM International Con- ference on Intelligent Virtual Agents, pages 1ā9, 2025. [8] Bin Han, Deuksin Kwon, Spencer Lin, Kaleen Shrestha, and Jonathan Gratch. Can llms generate behaviors for embodied virtual agents based on personality traits? In Proceedings of the 25th ACM Inter- national Conference on Intelligent Virtual Agents, pages 1ā10, 2025. [9] Juyeong Hwang, Seong-Eun Hong, JaeY- oung Seon, and HyeongYeop Kang. How does a virtual agent decide where to look? symbolic cognitive reasoning for embod- ied head rotation. In Proceedings of the SIGGRAPH Asia 2025 Conference Papers, pages 1ā12, 2025. [10] Xilei Zhao, Ruggiero Lovreglio, and Daniel Nilsson. Modelling and interpret- ing pre-evacuation decision-making using machine learning.Automation in Con- struction, 113:103140, 2020. [11] Gerard Yeo, Fiona Tan An Ting, Kokil Jaidka, Shaz Furniturewala, Wu Fanyou, Weijie Xu, Vinija Jain, Aman Chadha, Yang Liu, and See Kiong Ng.Phan- tom: Persona-based prompting has an ef- fect on theory-of-mind reasoning in large language models. In Proceedings of the In- ternational AAAI Conference on Web and Social Media, volume 19, pages 2124ā 2142, 2025. [12] Gregory Serapio-Garc Ģ ıa, Mustafa Safdari, Cl Ģ ement Crepy, Luning Sun, Stephen Fitz, Peter Romero, Marwa Abdulhai, Aleksan- dra Faust, and Maja Matari Ģ c. A psychome- tric framework for evaluating and shaping personality traits in large language models. Nature Machine Intelligence, pages 1ā15, 2025. [13] Pei Dang, Jun Zhu, Weilian Li, Yakun Xie, and Heng Zhang. Large-language-model- driven agents for fire evacuation simula- tion in a cellular automata environment. Safety Science, 191:106935, 2025. [14] Sen Yang, Luis Ceferino, Yi Zhang, Chen Gu, Tong Guo, and Gen Kondo. When agents learn to think:Large language model-enhanced agent-based modeling for crowd evacuation in disaster scenarios. Reliability Engineering & System Safety, page 112056, 2025. [15] Christian Becker-Asano. WASABI: Affect simulation for agents with believable inter- activity, volume 319. IOS Press, 2008. [16] LangChain overview - Docs by LangChain ādocs.langchain.com. https: //docs.langchain.com/oss/ python/langchain/overview. [Accessed 27-03-2026]. [17] OpenAIHarmonyResponseFor- matādevelopers.openai.com. https://developers.openai. com/cookbook/articles/ openai-harmony.[Accessed 27- 03-2026]. [18] Alessandro de Lima Bicho, Rafael Ara Ģ ujo Rodrigues,SoraiaRauppMusse, Cl Ģ audio Rosito Jung,Marcelo Par- avisi, and L Ģ eo Pini Magalh Ģ aes. Simulating crowds based on a space colonization algorithm.Computers & Graphics, 36(2):70ā79, 2012. [19] https://openai.com/safety/. [Accessed 27-03-2026].