Synthetic Personality Review

Can LLMs Generate Behaviors for Embodied Virtual Agents Based on Personality Traits?

Personas, identity, and agents2025arXivApproved editorial review

Authors: Bin Han, Deuksin Kwon, Spencer Lin, Kaleen Shrestha, Jonathan Gratch

Keywords: Human-Computer Interaction, Large Language Models, Virtual Agents, Personality Traits, Extraversion

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

5
Authors
15
Findings
45
Limitations
16
Evidence

Editorial summary

English

The work builds an embodied virtual agent that uses GPT-4o-mini-2024-07-18 to generate dialogue and select nonverbal behaviors associated with extraversion or introversion. The model receives a personality definition, a negotiation or ice-breaking context, and a closed list of facial, bodily, and vocal actions. It does not generate motion from scratch: it selects labels and descriptions executed in Unity with a male Reallusion character, SALSA expressions, Mixamo/Reallusion animations, and ElevenLabs voices. A GPT also describes the available clips offline. Eric and Brian voices are assigned to the extraverted and introverted agents after a 15-person pilot, so voice identity, utterance length, gesture, and personality change together.

Experiment 1 runs ten conversations per scenario, with up to ten negotiation turns and eight ice-breaking turns. LIWC 2015 yields 15 differing categories in negotiation and 11 in ice-breaking after the authors retain only differences with p<.05 and Cohen's d>.5. Extraverted agents produce more words: 90.58 versus 56.30 in ice-breaking and 56.32 versus 41.46 in negotiation, with p<.001. A BERT-based personality classifier provides mixed validation. In negotiation, it classifies 68% of utterances from both agents as extraverted and the condition difference is not significant (χ²=2.65, p=.10). In ice-breaking, it classifies 97.5% of extraverted-agent and 50% of introverted-agent utterances as extraverted (χ²=20.92, p<.001). Nonverbal plots largely follow the stereotypes encoded in the action list: more gaze, smiling, wide gestures, loud volume, and fast pace for extraversion; more averted gaze, narrow gestures, low volume, and slow pace for introversion. No inferential tests are reported for those probabilities.

Experiment 2 shows recorded videos to 30 Prolific participants, 17 women and 13 men, assigned to negotiation or ice-breaking. Each participant compares extraverted and introverted agents and reports perceived extraversion and influential cues; their own extraversion is estimated using the two relevant BFI-10 items. The displayed means show a clear agent-type effect: extraverted agents receive 4.8–6.0 and introverted agents 3.4–3.9 on a seven-point scale; the paper reports F=44.57 and p<.001. Scenario does not reach significance (p=.086). Participants attribute more influence to verbal than nonverbal behavior, 60% for the extraverted agent and 63.3% for the introverted agent, and most judge behavior consistent: 66.7–93.3% answer yes for nonverbal and 73.3% for verbal behavior; nobody answers no.

This evidence shows recognizability of two deliberately contrasted presentations, not that an LLM discovered or maintained a general psychological personality. Prompts, response length, voice, animations, and action labels are designed to maximize stereotypical cues and are not isolated through ablations. There is no verbal-only, nonverbal-only, no-personality, human-script, or disembodied baseline. Seeing both conditions may also disclose the hypothesis. Experiment 1 utterances are nested within only ten dialogues per scenario but are analyzed with t-tests that do not model dependence; significant LIWC categories are filtered without multiple-comparison correction. The human study omits power analysis, randomization and order, degrees of freedom, effect sizes, and exclusion details. The reported F=2.99 with p=.622 for video version is internally implausible and cannot be checked without data.

No code, prompts, videos, full questionnaire, data, derived animation assets, analysis plan, or ethics record is linked, and targeted search did not locate an official repository. The paper also reports no ethics approval or consent despite the human study and does not analyze stereotyping risks or the proposed future use in therapy. The defensible contribution is a multimodal prototype and initial evidence that observers distinguish two strongly contrasted audiovisual configurations, especially in an informal social context. It does not establish generalization to other traits, agents, models, voices, bodies, cultures, tasks, or live interaction.

Español

El trabajo construye un agente virtual encarnado que utiliza GPT-4o-mini-2024-07-18 para generar diálogo y seleccionar comportamientos no verbales asociados a extraversión o introversión. El modelo recibe una definición de personalidad, el contexto de negociación o conversación para romper el hielo y una lista cerrada de acciones faciales, corporales y vocales. No genera movimiento desde cero: escoge etiquetas y descripciones que se ejecutan en Unity mediante un personaje masculino de Reallusion, expresiones SALSA, animaciones de Mixamo/Reallusion y voces de ElevenLabs. Un GPT también describe previamente los clips disponibles. Las voces Eric y Brian se asignan respectivamente a los agentes extrovertido e introvertido después de un piloto con 15 personas, de modo que voz, longitud verbal, gesto y personalidad cambian conjuntamente.

El Experimento 1 ejecuta diez conversaciones por escenario. En negociación se permiten hasta diez turnos y en ice-breaking hasta ocho. LIWC 2015 muestra 15 categorías diferentes en negociación y 11 en ice-breaking después de que los autores retengan únicamente diferencias con p<0,05 y d de Cohen>0,5. Los agentes extrovertidos generan más palabras: 90,58 frente a 56,30 en ice-breaking y 56,32 frente a 41,46 en negociación, con p<0,001. Un clasificador de personalidad basado en BERT ofrece una validación desigual: en negociación clasifica como extrovertidas el 68 % de las intervenciones de ambos agentes y la diferencia entre condiciones no es significativa (χ²=2,65; p=0,10); en ice-breaking clasifica así el 97,5 % de las intervenciones extrovertidas y el 50 % de las introvertidas (χ²=20,92; p<0,001). Las figuras no verbales siguen en general los estereotipos incorporados a la lista: más mirada, sonrisa, gesto amplio, volumen alto y velocidad rápida para extraversión; más evitación de mirada, gesto estrecho, volumen bajo y ritmo lento para introversión. No se aplican pruebas inferenciales a esas probabilidades.

El Experimento 2 presenta vídeos pregrabados a 30 participantes de Prolific, 17 mujeres y 13 hombres, asignados a negociación o ice-breaking. Cada participante compara agentes extrovertido e introvertido y responde sobre extraversión percibida y señales influyentes; su propia extraversión se estima con los dos ítems correspondientes del BFI-10. El efecto principal del tipo de agente es claro en las medias mostradas: los agentes extrovertidos reciben entre 4,8 y 6,0 puntos y los introvertidos entre 3,4 y 3,9 en una escala de siete puntos; el artículo informa F=44,57 y p<0,001. El contexto no alcanza significación (p=0,086). Los participantes atribuyen mayor influencia a lo verbal que a lo no verbal, 60 % para el agente extrovertido y 63,3 % para el introvertido, y la mayoría considera los comportamientos coherentes: 66,7–93,3 % responde «sí» para lo no verbal y 73,3 % para lo verbal; nadie responde «no».

Esta evidencia demuestra reconocibilidad de dos presentaciones deliberadamente contrastadas, no que el LLM haya descubierto o mantenido una personalidad psicológica general. Los prompts, la longitud de respuesta, la voz, las animaciones y las etiquetas de acción están diseñados para maximizar señales estereotípicas y no se aíslan mediante ablaciones. El estudio no incluye condición solo verbal, solo no verbal, sin personalidad, guion humano o agente no encarnado. Ver dos condiciones puede además revelar la hipótesis a los participantes. Las unidades del Experimento 1 son intervenciones anidadas en apenas diez diálogos por escenario, pero se analizan con t-tests sin modelar dependencia; se filtran categorías LIWC significativas sin corrección por comparaciones múltiples. En el estudio humano faltan potencia, randomización y orden, grados de libertad, tamaños de efecto y detalles de exclusión. El dato F=2,99 con p=0,622 para versión de vídeo es internamente inverosímil y no puede verificarse con los datos ausentes.

No se enlazan código, prompts, vídeos, cuestionario completo, datos, animaciones derivadas, plan analítico o registro ético, y una búsqueda dirigida no localizó un repositorio oficial. El artículo tampoco informa aprobación ética o consentimiento, pese al estudio humano, ni analiza riesgos de estereotipar introversión/extraversión o de aplicar estos agentes en terapia, mencionada solo como futuro posible. La contribución defendible es un prototipo multimodal y evidencia inicial de que observadores distinguen dos configuraciones audiovisuales muy contrastadas, especialmente en un contexto social informal. No prueba generalización a otros rasgos, agentes, modelos, voces, cuerpos, culturas, tareas o interacción en tiempo real.

Research question

Can an LLM generate for virtual agents verbal and non-verbal behaviors that reproduce signals attributed to extraversion and introversion, can human observers distinguish those conditions, and which modality most influences their judgment?

Method

Two experiments with a Unity prototype. GPT-4o-mini-2024-07-18 generates dialogue and selects actions from a closed list for introverted, extraverted, and generic agents in negotiation and ice-breaking. Ten simulations per scenario are analyzed with length, LIWC 2015, a Big Five classifier based on BERT, and action distributions. Afterwards, 30 Prolific participants watch prerecorded videos, rate extraversion, and indicate verbal and non-verbal signals; their extraversion is estimated with two BFI-10 items. The editorial audit read and rendered the ten pages, checked tables, figures, statistics, metadata, and license, and searched for public artifacts.

Sample: Experiment 1: ten trials per scenario; conversations of up to ten turns in negotiation and eight in ice-breaking, with no published count of utterances analyzed per condition. Voice pilot: 15 participants, with no detailed characteristics or results. Experiment 2: 30 Prolific participants, 17 women and 13 men, distributed across two scenarios; age, country, language, inclusion criteria, exclusions, power, or exact cell sizes are not reported.

Findings

  • GPT-4o-mini produces dialogue and selects distinct non-verbal actions under extraversion and introversion prompts.
  • Extraverted agents generate more words than introverted agents in both scenarios.
  • In ice-breaking, the mean number of words is 90.58 versus 56.30 and the mean number of sentences is 5.80 versus 4.92.
  • In negotiation, the mean number of words is 56.32 versus 41.46.
  • Fifteen filtered LIWC categories differ in negotiation and eleven in ice-breaking.
  • In negotiation, the classifier labels 68% of utterances from both conditions as extraverted.
  • The classifier difference between agents is not significant in negotiation, with p=0.10.
  • In ice-breaking, the classifier marks 97.5% of extraverted utterances and 50% of introverted utterances as extraverted.
  • The selected actions follow the predefined associations for gaze, expressivity, gesture amplitude, volume, and speed.
  • The 30 observers rate extraverted agents between 4.8 and 6.0 and introverted agents between 3.4 and 3.9.
  • The article reports a main effect of agent type on perceived extraversion, F=44.57 and p<0.001.
  • Scenario does not show a significant main effect, although a trend appears with p=0.086.
  • The interaction between agent type and participant extraversion is reported with p=0.009.
  • Verbal signals are reported as more influential than non-verbal signals in both conditions.
  • Most participants consider the behaviors consistent with the perceived personality and nobody chooses the no answer.

Limitations

  • Only extraversion versus introversion is manipulated; the other four Big Five traits are not evaluated.
  • The system uses a single GPT-4o-mini checkpoint and does not test replication in other models.
  • Temperature is left at the default value, which may change or may not remain frozen reproducibly.
  • Complete prompts, system messages, structured outputs, or validation rules are not published.
  • The model chooses from a list of behaviors designed to encode stereotypical personality differences.
  • Movements are not generated by the LLM: preexisting animations and expressions are selected.
  • Clip descriptions produced offline by GPT are not published or evaluated for accuracy.
  • Behaviors are synchronized at the utterance level, not at the word or fine-grained temporal level.
  • A single male character is used and alternative bodies, appearances, genders, or styles are not tested.
  • The Eric and Brian voices are confounded with personality because they are not counterbalanced across conditions.
  • The 15-person pilot does not report procedure, questions, scores, statistics, or sample characteristics.
  • Verbal length, content, voice, gesture, posture, and face change simultaneously across conditions.
  • There are no ablations for verbal-only, non-verbal-only, no personality prompt, human script, or non-embodied agent.
  • The generic agent participates in generation but does not appear as a control condition in the human study.
  • Only ten conversations per scenario are run and are not justified by power or saturation.
  • The total number of utterances used in t-tests, LIWC, and the classifier is not reported.
  • Utterances are nested within conversations, but the analysis does not model that dependence.
  • Many LIWC categories are examined and only those exceeding p<0.05 and d>0.5 are reported.
  • No correction for multiple comparisons is applied nor is the full family of tests published.
  • The Cohen d values used for filtering do not appear in the tables.
  • The star notation in the tables is not explicitly defined.
  • LIWC 2015 is used while the main methodological reference cited describes LIWC-22.
  • The binary personality classifier is not validated for synthetic dialogue or for these lengths.
  • That 68% of both negotiation conditions are classified as extraverted contradicts a robust separation there.
  • The 50% extraverted classification for the introverted agent in ice-breaking is equivalent to chance in a binary output.
  • Non-verbal action probabilities are described without tests, uncertainty, or variability across conversations.
  • The human study has only 30 participants and approximately 15 per scenario.
  • No power analysis, preregistration, directional hypotheses, or prior analytic plan is provided.
  • Randomization, order counterbalancing, attention checks, exclusions, or missing data are not reported.
  • Viewing contrasted agents within the same scenario may reveal the manipulation and favor experimental demand.
  • The textual description of videos and versions is ambiguous: two videos are mentioned, but the ANOVA includes two versions per type.
  • The videos are not published, so it cannot be audited whether they differ in content, quality, or duration.
  • The ANOVA does not report degrees of freedom, effect sizes, intervals, or assumption checks.
  • F=2.99 with p=0.622 for video version is an internally implausible combination and suggests a reporting error.
  • For scenario, p=0.086 is given without the corresponding F statistic.
  • The interaction regression does not report coefficients, errors, intervals, R-squared, coding, or group split.
  • Two BFI-10 items provide a very brief measure and inversion or scoring is not explained.
  • Influence categories allow multiple selection, but normalized percentages and chi-square are not explained with counts.
  • The coherence option asks about alignment with the already perceived personality, which may be tautological.
  • Naturalness, likability, credibility, presence, task effectiveness, or behavioral consequences are not evaluated.
  • Interaction is through prerecorded videos, not real-time conversation.
  • Ethical approval, consent, exact compensation, or data protection are not reported.
  • Risks of stereotyping introversion as withdrawal, sadness, low voice, or gaze avoidance are not discussed.
  • Code, data, prompts, stimuli, questionnaire, or supplementary materials are not linked.
  • The possible future application in therapy is mentioned without clinical evaluation, safety, or governance.

What the study does not establish

  • It does not demonstrate that the LLM possesses its own personality.
  • It does not demonstrate induction of a stable psychometric construct beyond designed signals.
  • It does not test consistency in other Big Five traits.
  • It does not test generalization to other models, prompts, voices, bodies, languages, or cultures.
  • It does not demonstrate reliable separation between introversion and extraversion in negotiation.
  • It does not establish that the LLM is necessary versus equivalent scripts or rules.
  • It does not causally identify whether perception stems from language, voice, face, gesture, or their combination.
  • It does not demonstrate generation of non-verbal movement from scratch.
  • It does not demonstrate effective real-time interaction.
  • It does not test improvement in negotiation, collaboration, satisfaction, or task outcomes.
  • It does not demonstrate that observers perceive personality without comparing contrasted conditions.
  • It does not establish that participant personality improves accuracy rather than introducing bias.
  • It does not validate use in therapy, education, health, or sensitive decisions.
  • It does not evaluate safety, fairness, or harm from stereotypes.
  • It does not offer a reproducible experiment with the available public artifacts.

Traceability

Scope: Full text

Version: arXiv:2508.21087v1 (27 Aug 2025); arXiv non-exclusive distribution license 1.0

Consulted source: https://arxiv.org/pdf/2508.21087v1

Review: Codex full-text, visual, statistical, human-study and artifact audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o-mini-2024-07-18 with default temperature
  • Generic, introverted and extraverted prompted agents
  • Bagged SVM over BERT embeddings personality classifier from Kazameini et al. (2020)
  • Unity male Reallusion embodied character
  • ElevenLabs Eric and Brian English voices

Instruments and metrics

  • Closed nonverbal action list covering face, body and voice
  • LIWC 2015 lexical categories
  • Binary pretrained Big Five personality detector
  • Seven-point perceived-extraversion rating
  • Two extraversion items from BFI-10
  • Cue-influence and verbal/nonverbal consistency questions
  • Mixed ANOVA, chi-square tests, linear regression, t-tests and Cohen's d filtering

Data used

  • Ten agent-to-agent trials per scenario for negotiation and ice-breaking
  • Recorded virtual-agent video stimuli generated from those conversations
  • Prolific user study with 30 participants
  • Pilot voice-selection study with 15 participants

Evidence and location

  • Objectives and research questions: arXiv v1, abstract and section 1, pp. 1-2
  • Closed list of actions and behavioral rationale: arXiv v1, section 3.1 and Table 1, pp. 3-4
  • Unity, GPT, and animation pipeline: arXiv v1, section 3.2 and Figure 2, p. 4
  • Voices and 15-participant pilot: arXiv v1, section 3.2, p. 4
  • Checkpoint, scenarios, turns, and ten trials: arXiv v1, section 4.1 and footnote 1, p. 5
  • LIWC and personality classifier: arXiv v1, section 4.2, p. 5
  • Verbal length and filtered LIWC categories: arXiv v1, section 4.3.1 and Tables 2-3, pp. 5-6
  • Classifier results by scenario: arXiv v1, Figure 3 and section 4.3.1, pp. 5-6
  • Non-verbal action distributions: arXiv v1, Figures 4-5 and section 4.3.2, p. 6
  • Design, sample, and BFI-10 measure: arXiv v1, section 5.1, pp. 6-7
  • Means and ANOVA of perceived personality: arXiv v1, Figure 6 and section 5.2.2, p. 7
  • Interaction with observer personality: arXiv v1, Figure 7 and section 5.2.3, p. 7
  • Signal influence and perceived coherence: arXiv v1, Table 4 and sections 5.2.4-5.2.5, pp. 7-8
  • Scope, interpretation, and future work: arXiv v1, sections 6-7, p. 8
  • Version and license: arXiv:2508.21087v1 metadata; submitted 27 Aug 2025; non-exclusive distribution license 1.0
  • Absence of linked public artifacts: arXiv v1 paper and metadata plus targeted repository search; audited 15 Jul 2026