Synthetic Personality Review

LLM-Driven Personalities for Decision Making in Emergency Simulations

Applications, bias, and safety2026arXivApproved editorial review

Authors: Stefano Calzolari, Rubens Montanha, Gabriel Schneider, Gustavo Wide, Paulo Knob, Francesco Strada, Andrea Bottino, Soraia Raupp Musse

Keywords: OCEAN personality prompts, LLM-driven virtual agents, Emergency evacuation simulation, Crowd simulation, Prompt-conditioned decision making, Rescue behavior, gpt-oss-120b

Source: Open primary source (opens in a new tab)

8
Authors
6
Findings
12
Limitations
6
Evidence

Editorial summary

English

This paper tests whether OCEAN-derived textual descriptions change the decisions of LLM-controlled virtual agents in a fire-evacuation simulation. It uses one model, gpt-oss:120b-cloud through Ollama, with temperature 0.1, top-k 10, and top-p 0.1. Six prompts are constructed: five extreme profiles and one labeled neutral. An extreme profile does not manipulate one trait in isolation. It sets the named trait high and all four other traits low, then turns the full vector into long adjective lists. One hundred agents per profile receive up to five increasingly severe warnings and choose Evacuate, Continue, or Panic. Outcomes are nearly deterministic: all 100 conscientious and neutral agents eventually evacuate, all 100 neurotic agents end in panic, about 80% of open agents evacuate while roughly 20% continue, and only about 20% of extraverted agents evacuate. A second exercise runs one 30-agent simulation for each of four helper profiles, always paired with 15 neurotic agents. Every observed rescue call for neutral (9) and agreeable (12) returns help, whereas every call for conscientious (52) and open (31) refuses help. This shows that these prompt wordings can produce different action policies in this model and scenario. It does not show that the agents reproduce human personality or that the simulation is more realistic. Profiles change all five traits simultaneously and include adjectives that directly cue the expected response: anxious and impulsive favor Panic, altruistic and cooperative favor Help, and adventurous and daring favor risk-taking. The neutral control is not personality-free either; it includes all positive and negative markers together, many of them contradictory. There is no no-persona control, human comparison, quantitative psychological model, second LLM, rule-based baseline, or validation against real evacuations. The paper uses significantly narratively but reports no hypothesis test, confidence interval, or dependence analysis. The rescue study has one simulation per condition and counts encounters nested within each trajectory rather than independent replications; it also reports no evacuation time, casualty, or collective-safety outcome. Without public code, complete user prompts, outputs, data, seeds, or a pinned model version, the figures are not reproducible. The faithful conclusion is limited but useful: an LLM can map highly explicit personality descriptors to different decisions in an evacuation prototype; psychological fidelity, realism, and improvement over traditional methods remain unvalidated.

Español

Este trabajo prueba si descripciones textuales basadas en OCEAN cambian las decisiones de agentes virtuales controlados por un LLM en una evacuación de incendio. Usa un único modelo, gpt-oss:120b-cloud mediante Ollama, con temperatura 0,1, top-k 10 y top-p 0,1. Construye seis prompts: cinco perfiles extremos y uno denominado neutral. Cada perfil extremo no modifica un rasgo de forma aislada: fija el rasgo que le da nombre en alto y los otros cuatro en bajo, convirtiendo el vector completo en largas listas de adjetivos. Cien agentes por perfil reciben hasta cinco avisos de peligro creciente y eligen Evacuate, Continue o Panic. Los resultados son casi deterministas: los 100 agentes conscientious y neutral terminan evacuando, los 100 neurotic terminan en pánico, cerca del 80% de los open evacúa y aproximadamente el 20% continúa, mientras solo alrededor del 20% de los extraverted evacúa. En un segundo ejercicio se ejecuta una sola simulación de 30 agentes por cada uno de cuatro perfiles ayudantes, siempre junto a 15 agentes neurotic. Todas las llamadas de rescate observadas para neutral (9) y agreeable (12) respondieron ayudar; todas las de conscientious (52) y open (31) respondieron no ayudar. Esto demuestra que la redacción de estos prompts puede producir políticas de acción distintas en este modelo y escenario. No demuestra que los agentes reproduzcan personalidad humana ni que la simulación sea más realista. Los perfiles cambian simultáneamente los cinco rasgos y contienen adjetivos que anticipan directamente las respuestas: anxious e impulsive favorecen Panic; altruistic y cooperative favorecen Help; adventurous y daring favorecen ignorar el riesgo. El control neutral tampoco está libre de personalidad: incluye a la vez todos los marcadores positivos y negativos, muchos contradictorios. No existe control sin persona, comparación con humanos, modelo psicológico cuantitativo, segundo LLM, baseline basado en reglas ni validación con evacuaciones reales. El artículo usa significativamente en sentido narrativo, pero no presenta pruebas de hipótesis, intervalos de confianza ni análisis de independencia. Las cuatro simulaciones de rescate aportan una ejecución por condición y cuentan encuentros dentro de cada trayectoria, no réplicas independientes; además no informan tiempos de evacuación, víctimas o una métrica colectiva de seguridad. Sin código, prompts completos de usuario, salidas, datos, semillas o versiones fijadas del modelo, las cifras no son reproducibles. La conclusión fiel es limitada pero útil: un LLM puede traducir descriptores de personalidad muy explícitos en decisiones distintas dentro de un prototipo de evacuación; la fidelidad psicológica, el realismo y la mejora frente a métodos tradicionales quedan pendientes de validación.

Research question

Can textual prompts derived from OCEAN profiles make an LLM choose different actions for virtual agents during an evacuation, both in response to fire alerts and to opportunities to help agents in panic?

Method

Unity 3D prototype with BioCrowds connected via ZeroMQ to a LangChain pipeline. gpt-oss:120b-cloud receives a biography, work context, an OCEAN profile converted into adjectives, decision memory, and one of five warnings of increasing danger; it returns JSON with Evacuate, Continue, or Panic and a rationale. 100 trajectories are run for each of six profiles. Then four simulations of 30 agents are performed, with 15 neurotic and 15 of a helper profile, to record binary rescue responses when encountering agents at less than five meters.

Sample: First exercise: 600 agents or trajectories, 100 for each of six profiles, with up to five sequential decisions. Second exercise: four simulations of 30 agents; each combines 15 neurotic with 15 neutral, agreeable, conscientious, or open. 9, 12, 52, and 31 rescue encounters are observed respectively, but only one collective trajectory per condition.

Findings

  • The 100 conscientious profiles and the 100 neutral profiles end up evacuating; the 100 neurotic profiles end up in panic.
  • By the end of the fifth warning, approximately 80% of open evacuates and 20% continues, while only around 20% of extraverted evacuates and the rest continues.
  • Most evacuations occur after the second and third warnings, not after the first danger message.
  • In the four rescue simulations, neutral helps in 9/9 encounters and agreeable in 12/12; conscientious refuses in 52/52 and open in 31/31.
  • The generated rationales repeat the vocabulary of the prompt, making it visible that the model associates the explicit adjectives with the available actions.
  • The system integrates structured LLM calls with a spatial simulator that pauses while each agent decides, demonstrating technical viability of the prototype.

Limitations

  • Only one model served through a mutable cloud alias is evaluated; there is no comparison between models nor an immutable version of the weight or service.
  • Each profile elevates one trait and reduces the other four, so no difference identifies the causal effect of the trait that names the profile.
  • The lists include direct semantic cues toward the available actions, such as anxious, altruistic, cooperative, adventurous, and daring.
  • The neutral assumption concatenates contradictory high and low markers; it is not a control without personality conditioning.
  • No condition without persona, adjective ablation, rule-based baseline, human comparison, or psychometric or behavioral validation is included.
  • Sampling seeds, number of retries, parsing errors, stability across runs, or sensitivity to temperature and wording are not reported.
  • The article presents no statistical tests, confidence intervals, inferential effect sizes, or correction for sequential decisions and nested encounters.
  • The rescue study has one collective simulation per condition; the 9, 12, 52, and 31 encounters are not independent replicates.
  • Evacuation times, congestion, exit usage, casualties, completed rescues, or any other variable quantifying collective safety are not measured.
  • The spatial dynamics are simplified: those who evacuate know the optimal path; those who enter panic stop moving; helping activates at less than five meters and fixes speeds manually.
  • No code, full configuration, user prompts, decisions, rationales, trajectory data, logs, or reproducible manifest are published.
  • A single building, a software engineer biography, and three predefined actions limit generalization.

What the study does not establish

  • It does not establish that the profiles correspond to real human personality or that they validly measure the five OCEAN factors.
  • It does not isolate conscientiousness, agreeableness, neuroticism, openness, or extraversion because each profile changes the five traits simultaneously.
  • It does not demonstrate that an LLM crowd is more realistic, credible, or adaptable; there are no human observers or real evacuation data as a criterion.
  • It does not demonstrate superiority or practical flexibility over rule-based systems, because no comparable baseline is implemented.
  • It does not allow interpreting the nearly deterministic patterns as a distribution of human behavior; they may be literal following of lexical cues.
  • It does not support the word significantly as a statistical inference, since no formal contrast is presented.
  • It does not demonstrate that helping improves collective safety or efficiency; it only counts LLM responses to encounters and not evacuation outcomes.
  • It does not allow reproducing the figures or evaluating robustness due to the lack of public artifacts and seeds.

Traceability

Scope: Full text

Version: arXiv:2606.31038v1; CASAXR 2026 conference contribution listed in press by Politecnico di Torino

Consulted source: https://arxiv.org/pdf/2606.31038

Review: Codex 12-page full-text visual, TeX source, publication-status, construct, prompt, experimental-design, statistical, artifact, reproducibility and claim audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • gpt-oss:120b-cloud through Ollama

Instruments and metrics

  • Five OCEAN-derived extreme adjective prompts
  • Contradictory all-markers neutral prompt
  • Five escalating fire-alert messages
  • Evacuate, Continue, or Panic structured action
  • Binary rescue Call for Actions
  • Unity 3D and BioCrowds evacuation prototype
  • Agent rationale generated by the same LLM

Data used

  • Unreleased outputs from 600 prompt-conditioned agent trajectories
  • Unreleased encounter decisions from four 30-agent rescue simulations
  • 3D replica of a real office used as a synthetic scene

Evidence and location

  • Status, authors, and contribution CASAXR 2026 in press: Official arXiv record 2606.31038v1 and Politecnico di Torino IRIS record 11583/3010356, checked 2026-07-16
  • Architecture, OCEAN conversion, prompts, simulator, and actions: arXiv v1, Sections 3.1-3.3 and Tables 1-3
  • Six hundred trajectories, model configuration, and decision patterns: arXiv v1, Sections 4 and 4.1, Figures 3-5 and Table 4
  • Four rescue simulations, encounters, and responses: arXiv v1, Section 4.2, Figure 7 and Table 5
  • Acknowledged limitations and pending real validation: arXiv v1, Section 5
  • Absence of repository or locatable data and audit of construct, statistics, and reproducibility: reports/verification/article-287-arxiv-ocean-emergency-prompt-confounding-realism-statistics-artifact-and-claim-audit.json