Hierarchical Generative Agents for Simulating Sequential Human Behavior

Personas, identity, and agents2026arXivApproved editorial review

Authors: Maria G. Mendoza, Lucas Waldburger, Jin Lee, Shankar Sastry

Keywords: LLM agents, Evacuation simulation, Cognitive hierarchy, Persona conditioning, GPT-4.1-mini, Route-choice priors, Human validity, Reproducibility

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

4
Authors
10
Findings
19
Limitations
4
Evidence

Editorial summary

English

This preprint proposes a prototype for simulating sequential evacuation decisions with persona-conditioned agents. Each profile describes age, gender, risk perception, trust in authorities, threat response, routines, dependents, relationships, and spatial knowledge. The architecture combines a partially observable two-dimensional urban environment, perception and memory, an urgency score, GPT-4.1-mini, and three decision levels: evacuation intent, local route choice, and A* navigation. Its aim is to move beyond homogeneous or perfectly rational evacuation models and let decisions change in response to alerts, smoke, fire, traffic, and social obligations. The defensible contribution is architectural: the repository makes the links among profiles, cognitive rules, prompts, empirical priors, and navigation inspectable. It does not, however, show that simulated behavior predicts human behavior or that personality differences emerge naturally. The claimed empirical calibration consists of converting a controlled corridor-choice dataset into descriptive age-by-gender probabilities and passing those values to the route-choice prompt. Repository history contains 208 participants, 4,783 responses, and 4,617 unique corridor choices across 66 tasks. The script computes means for corridor width, transition cues, and cue conflicts. It does not fit simulator parameters, minimize a calibration objective, report intervals, split training and evaluation, or compare simulated trajectories with human trajectories. Some subgroups are very small, only three participants use the x gender key, and some conflict cells contain three or ten decisions, yet the JSON omits sample sizes and uncertainty. Age is also hard-coded as 2020 minus birth year; at least one record yields -1 and is silently excluded from the groups. The paper's main behavioral evidence is a minimal two-agent scenario with a static fire. The PDF calls the red agent risk-averse and the blue agent risk-prone and says red remains farther from the fire. The executable configuration assigns exactly the opposite labels: red Isabello is highly risk prone and blue Klaus is highly risk averse. If the published visual came from that configuration, its apparent pattern contradicts the figure's interpretation. The artifact cannot reconstruct it either: only three logs explicitly mention both profiles; their trajectory files contain headers but no positions, the historical validation notebook merely loads an author-specific local path and runs main.py, and no traceable script generates the figure. In all three runs both agents ignore the mild alert and evacuate when a textual event literally says evacuation is required. LLM rationales are semantically unstable: some use risk prone to mean underestimating danger, while another uses it as a reason to evacuate promptly. The discussion calls the profiles statistically distinguishable, but neither paper nor code reports n, a test, statistic, p value, interval, or effect size. Behavior is also prestructured by a deterministic urgency score, persona-text modifiers, and explicit safety and dependent-care rules. Without ablations of rules, persona, priors, and LLM, the source of each effect is unknown. The scalability study runs three episodes with one to five agents and shows token growth and spikes at stimulus steps. This documents a local prototype cost, not general scalability. The paper itself acknowledges that sequential human ground truth is absent and leaves distributional comparisons and matched human T-intersection tasks to future work; that admission correctly bounds the present evidence. The repository offers code, data, configurations, logs, and notebooks, but it is not a reproduction package: default main is stale relative to prompt-cleaning, no paper branch or commit is specified, and a top-level README, license, release, locked environment, tests, and CI are absent. The configured seed is not passed to env.reset, and the remote model alias is not pinned. More seriously, the public default branch contains a probable plaintext OpenAI credential; this review deliberately omits its value and recommends immediate revocation and history purging. Participant-level CSVs also lack an adjacent license, dictionary, data-use statement, and consent or de-identification documentation. The prototype is an interesting basis for research on hybrid cognitive architectures, but emergency-planning use would require corrected labels, a reproducible pipeline, baselines and ablations, construct validation, held-out human evaluation, uncertainty quantification, and resolved security and data-governance issues.

Español

Este preprint propone un prototipo para simular decisiones secuenciales durante una evacuación mediante agentes condicionados por personas sintéticas. Cada perfil describe edad, género, percepción del riesgo, confianza en autoridades, respuesta a amenazas, rutinas, dependientes, relaciones y conocimiento espacial. La arquitectura combina un entorno urbano bidimensional parcialmente observable, memoria y percepción, una puntuación de urgencia, GPT-4.1-mini y tres niveles de decisión: intención de evacuación, elección local de ruta y navegación A*. El objetivo es superar modelos de evacuación homogéneos o perfectamente racionales y permitir que las decisiones cambien ante alertas, humo, fuego, tráfico y obligaciones sociales. La contribución defendible es arquitectónica: el repositorio permite inspeccionar cómo se conectan perfiles, reglas cognitivas, prompts, priors empíricos y planificación. No obstante, el estudio no demuestra que el comportamiento simulado prediga el comportamiento humano ni que las diferencias de personalidad emerjan de forma natural. La llamada calibración empírica consiste en convertir un conjunto controlado de elecciones de corredor en probabilidades descriptivas por edad y género y pasar esos valores al prompt de decisión de ruta. El historial contiene 208 participantes, 4.783 respuestas y 4.617 elecciones de corredor únicas en 66 tareas. El script calcula medias sobre preferencia por anchura, elementos de transición y conflictos entre ambas señales. No estima parámetros del simulador, no minimiza una función de ajuste, no informa intervalos, no separa entrenamiento y prueba y no compara trayectorias simuladas con trayectorias humanas. Algunos subgrupos son diminutos, solo tres participantes tienen la clave de género x y ciertas celdas de conflicto reúnen tres o diez decisiones, pero el JSON omite n e incertidumbre. Además, la edad se calcula como 2020 menos año de nacimiento; al menos un registro da -1 y se excluye silenciosamente de los grupos. La principal evidencia conductual del artículo es un escenario mínimo con dos agentes y un fuego estático. El PDF llama al agente rojo risk-averse y al azul risk-prone y afirma que el rojo mantiene más distancia del fuego. La configuración ejecutable contiene exactamente las etiquetas opuestas: el agente rojo Isabello es highly risk prone y el azul Klaus highly risk averse. Por tanto, el patrón visual publicado, si procede de esa configuración, contradice la interpretación de la figura. El artefacto tampoco permite reconstruirla: solo tres logs mencionan explícitamente ambos perfiles; sus archivos de trayectoria tienen cabeceras pero ninguna posición, el notebook histórico de validación se limita a cargar una ruta local y ejecutar main.py, y no existe un script trazable que genere la figura. En las tres corridas ambos agentes ignoran la alerta leve y evacuan cuando un evento textual dice literalmente que la evacuación es obligatoria. Las justificaciones del LLM son inestables: en unos casos risk prone significa subestimar el peligro y en otro se usa como razón para evacuar pronto. El discussion llama a los perfiles estadísticamente distinguibles, pero ni el PDF ni el código proporcionan n, test, estadístico, p, intervalo o tamaño de efecto. Además, el comportamiento está preestructurado por una puntuación determinista de urgencia, modificadores activados por texto de persona y reglas explícitas de seguridad y cuidado de dependientes. Sin ablaciones de reglas, persona, priors y LLM no se sabe qué componente produce cada resultado. El estudio de escalabilidad ejecuta tres episodios con uno a cinco agentes y muestra que el uso de tokens crece y tiene picos en pasos con estímulos. Esto documenta un coste del prototipo, no escalabilidad general. El propio artículo reconoce que faltan trazas humanas secuenciales y deja para trabajo futuro comparar distribuciones y repetir con agentes las tareas T-intersection realizadas por participantes; ese reconocimiento delimita correctamente el estado actual. El repositorio aporta código, datos, configuraciones, logs y notebooks, pero no es un paquete reproducible: la rama por defecto está desactualizada frente a prompt-cleaning, el paper no fija rama o commit, faltan README superior, licencia, release, entorno bloqueado, tests y CI, la semilla de configuración no se pasa a env.reset y el alias remoto del modelo no está fijado. Más grave, la rama pública por defecto contiene una probable credencial OpenAI en texto plano; este resumen omite deliberadamente su valor y exige revocarla y purgarla del historial. Los CSV con datos individuales también carecen, dentro del repositorio, de licencia, diccionario, declaración de uso y documentación de consentimiento o desidentificación. El prototipo es una base interesante para investigar arquitecturas cognitivas híbridas, pero antes de usarse en planificación de emergencias necesita corregir las etiquetas, reconstruir una canalización reproducible, incorporar baselines y ablaciones, validar constructos, evaluar contra personas fuera de muestra, cuantificar incertidumbre y resolver seguridad y gobernanza de datos.

Research question

Can a hierarchical architecture that combines textual profiles, route choice priors, cognitive rules, GPT-4.1-mini, and A* navigation produce heterogeneous and plausible sequential decisions during a dynamic evacuation?

Method

Partially observable urban environment with fire, smoke, and traffic; synthetic persons with demographic, cognitive, social, and spatial attributes; urgency scoring and memory; GPT-4.1-mini for high-level intention and local decisions; A* for navigation. Data from 208 participants are aggregated into descriptive priors by age and gender. The published evaluation consists of qualitative examples, a minimal two-agent scenario with static fire, and three scalability episodes for each size from one to five agents.

Sample: Empirical history data: 208 participants and 4,617 corridor decisions across 66 tasks. Behavioral validation: two synthetic persons, one per risk profile, with number of episodes not reported in the paper; only three logs from the repository mention both profiles. Scalability: three episodes per condition with one to five agents.

Findings

  • The prototype implements an inspectable hierarchy of intention, local route, and A* navigation conditioned by synthetic persons.
  • Qualitative responses change in the presence of alerts, smoke, fire, traffic, and family obligations, although the effect of rules, prompts, and the LLM is not separated.
  • Empirical priors are derived from 4,617 corridor choices, but they are descriptive aggregates and not a calibration of the simulator.
  • The validation figure labels red as risk-averse and blue as risk-prone, while the executable configuration assigns red as risk prone and blue as risk averse.
  • The three explicit logs make both agents evacuate after a mandatory textual order and contain contradictory justifications for the risk prone trait.
  • There is no traceable trajectory data or script to reproduce the validation figure.
  • No traceable statistical analysis exists to support statistically distinguishable proximity profiles.
  • Token consumption grows with one to five agents and has spikes at replanning steps, showing limited scalability of the prototype.
  • The paper itself acknowledges the lack of sequential human ground truth and proposes future quantitative validation.
  • The default public branch exposes a probable OpenAI credential and must be considered compromised.

Limitations

  • Simulated sequential decisions or trajectories are not compared with out-of-sample persons.
  • Priors are means by age and gender without intervals, cell size in the JSON, or a hierarchical model for repeated decisions.
  • Very small subgroups produce apparently precise probabilities.
  • Age calculation uses 2020 as a constant and generates at least one impossible age excluded silently.
  • Generalization from stairs to urban transition elements is not validated.
  • Validation has a single person per profile and confounds risk, identity, initial position, goal, and description.
  • Color and risk labels are inverted between paper and configuration.
  • The positions of the validation trajectories and the pipeline that generates the figure are missing.
  • No test, n, p, interval, or effect size is reported for the statistical claim.
  • Time steps are autocorrelated and would not be independent replicates.
  • The textual evacuation order and explicit rules introduce strong demand characteristics.
  • There are no rule-only, LLM-only, no-person, permuted-person, or no-priors baselines.
  • There is no validation of a psychological risk perception scale or manipulation check.
  • The configured seed is not passed to env.reset and the remote model is not fixed by snapshot.
  • Three episodes per size do not characterize uncertainty or general scalability.
  • The relevant branch of the paper is not identified and the default branch is outdated.
  • Top-level README, license, release, lockfile, tests, and CI are missing.
  • A probable OpenAI credential is published in the main branch and history.
  • CSV files with individual data lack repository license and governance documentation.

What the study does not establish

  • It does not demonstrate predictive accuracy over real human evacuations.
  • It does not demonstrate that the simulator is quantitatively calibrated with human data.
  • It does not validate risk prone and risk averse as psychological constructs in the agents.
  • It does not demonstrate that personality differences emerge naturally from the LLM.
  • It does not demonstrate that the published red/blue pattern corresponds to the described labels.
  • It does not demonstrate that proximity profiles are statistically distinguishable.
  • It does not identify which part of the behavior comes from the LLM, the rules, the prompt, or the priors.
  • It does not generalize to other emergencies, cities, cultures, models, or populations.
  • It does not justify operational use for resource allocation, alarm design, or urban planning.
  • It does not yet offer a safe and reproducible end-to-end artifact.

Traceability

Scope: Full text

Version: arXiv:2606.14989v1

Consulted source: https://arxiv.org/abs/2606.14989v1

Review: Codex twenty-page full-text visual, TeX, empirical-data, public-code-history, human-validity, security and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4.1-mini via OpenAI API
  • Deterministic urgency and reflection module
  • A* low-level route planner
  • Persona-conditioned hierarchical agent

Instruments and metrics

  • Dynamic MiniGrid-style urban environment
  • Hand-authored persona YAML profiles
  • Perception and associative-memory modules
  • PADM-inspired urgency representation
  • Age-by-gender route-choice priors
  • High-level intent and mid-level direction prompts
  • Trajectory, intent, urgency and token logs

Data used

  • Controlled evacuation corridor-choice data: 208 participants, 4,783 response rows and 4,617 corridor decisions
  • Hand-authored synthetic persona profiles
  • Synthetic urban maps and simulation logs
  • Aggregate mobility and PADM studies used as conceptual background, not model-evaluation data

Evidence and location

  • Metadata and version: Official arXiv record 2606.14989v1, checked 2026-07-16
  • Architecture, results, appendices, and declared limits: arXiv v1, all twenty PDF pages and TeX source
  • Code, data, labels, logs, branches, and security: Public hierarchical_LLM_agents repository; prompt-cleaning commit dda54ab9e53def1543e13ac0b0092a5a0e18749c, main commit ccccd0496e7fbd6ee27fcbeb3c6f745368321410 and validation-history commit a54ae50
  • Recalculations, human validity, reproducibility, and claim limits: reports/verification/article-299-hierarchical-evacuation-agents-human-validity-label-reversal-security-and-reproducibility-audit.json