Psychologically Enhanced AI Agents

Personas, identity, and agents2025arXivApproved editorial review

Authors: Maciej Besta, Shriram Chandran, Robert Gerstenberger, Mathis Lindner, Marcin Chrapek, Sebastian Hermann Martschat, Taraneh Ghandi, Patrick Iff, Hubert Niewiadomski, Piotr Nyczyk, Jürgen Müller, Torsten Hoefler

Keywords: Artificial Intelligence, Computation and Language, Computers and Society, Human-Computer Interaction, Multiagent Systems

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

12
Authors
7
Findings
11
Limitations
10
Evidence

Editorial summary

English

The paper presents MBTI-in-Thoughts (MiT), a prompting framework that assigns one of 16 MBTI profiles to LLM agents and examines whether those instructions alter psychometric answers, stories, game strategies, and multi-agent collaboration. For induction validation, GPT-4o mini answers the 60-item 16Personalities assessment five times per profile at temperature 1. The prompt contains both a profile description and four answer exemplars already aligned with the requested axes. The plots clearly separate E/I, T/F, and J/P, while S/N is weaker. This supports consistent instruction following on the same construct explicitly encoded by the prompt, not an internal personality or persistence beyond the test. For narrative generation, the study samples 100 WritingPrompts items and generates stories for all 16 types plus EXPERT and NONE controls. In the displayed Qwen3-235B-A22B, temperature-0 result, automated scores assign Feeling profiles greater emotional chargedness, happier endings, and personalness. Yet the controls also improve cohesion and redundancy over human stories, and the paper acknowledges that the personality-specific effect on those quality properties is small. In repeated games, honesty is operationalized as agreement between an action announced in a message, as inferred by another LLM, and the selected action. In the aggregated GPT-4 figure, Thinking profiles defect in roughly 90% of rounds versus 50% for Feeling; Thinking switches strategy at about 0.07 versus 0.16 for Feeling; and Introversion reaches about 0.54 honesty versus 0.33 for Extraversion. These are protocol-specific patterns under instructions that explicitly permit deception, not general measures of honesty, cognition, or safety. On BIG-Bench and SOCKET, communication with private reflection outperforms plain interactive communication but is approximately comparable to independent voting, so it does not establish an improvement over the simpler control. The claimed extension to Big Five, HEXACO, Enneagram, and DISC is an illustrative vector formalization rather than an experimental evaluation. The paper explicitly says it presents representative results and omits data that does not yield relevant insights. The official repository also does not provide an executable reproduction: it contains no raw results or locked environment, three modules fail compilation because of hyphenated import names, several scripts reference missing paths, the interactive-protocol judge is not passed the concluding conversation, and the published game configuration fixes temperature 0 or invokes GPT-4o mini whereas the figure reports GPT-4 at temperature 1. Claims about healthcare, negotiation, safety, or judicial use therefore remain speculative.

Español

El artículo presenta MBTI-in-Thoughts (MiT), un marco de prompts que asigna a agentes LLM uno de los 16 perfiles MBTI y estudia si esas instrucciones cambian respuestas psicométricas, relatos, estrategias en juegos y colaboración multiagente. Para validar la inducción, GPT-4o mini responde los 60 ítems de 16Personalities cinco veces por perfil, a temperatura 1; los prompts incluyen tanto una descripción del tipo como cuatro ejemplos de respuesta ya alineados con sus ejes. Los diagramas separan con claridad E/I, T/F y J/P, mientras S/N resulta más débil. Esta prueba demuestra seguimiento coherente de instrucciones en el mismo constructo que el prompt explicita, no una personalidad interna ni persistencia fuera del test. En generación narrativa se toman 100 prompts de WritingPrompts y se generan historias para los 16 tipos y dos controles, EXPERT y NONE. En el resultado mostrado, Qwen3-235B-A22B a temperatura 0 produce diferencias de tono evaluadas automáticamente: los perfiles Feeling reciben mayores puntuaciones de carga emocional, final feliz y carácter personal. Sin embargo, los controles también mejoran cohesión y redundancia respecto de historias humanas, por lo que el propio texto reconoce que el efecto específico de personalidad sobre esas propiedades de calidad es pequeño. En juegos repetidos, el trabajo define honestidad como coincidencia entre la acción anunciada, inferida por otro LLM, y la acción elegida. En la figura agregada para GPT-4, Thinking aparece con aproximadamente 90 % de defección frente a 50 % en Feeling; Thinking cambia de estrategia alrededor de 0,07 frente a 0,16 en Feeling; e Introversion alcanza una honestidad cercana a 0,54 frente a 0,33 en Extraversion. Son patrones del protocolo, cuyo prompt permite expresamente mentir, no medidas generales de honestidad, cognición o seguridad. En BIG-Bench y SOCKET, la comunicación con reflexión privada supera a la comunicación interactiva simple, pero queda aproximadamente al nivel de la votación independiente; por tanto, no se demuestra una mejora sobre el control más sencillo. La extensión a Big Five, HEXACO, Eneagrama y DISC es una formalización vectorial ilustrativa, no una evaluación experimental de esos marcos. El artículo admite que selecciona resultados “representativos” y omite datos sin “insights relevantes”. Además, el repositorio oficial no reproduce de forma ejecutable el estudio: no contiene resultados ni una instantánea de entorno, tres módulos no compilan por imports con guion, varios scripts apuntan a rutas inexistentes, el juez de los protocolos interactivos no recibe la conversación final y la configuración publicada de teoría de juegos usa temperatura 0 o GPT-4o mini, frente a GPT-4 y temperatura 1 declarados en la figura. Por estas razones, las conclusiones aplicadas a salud, negociación, seguridad o justicia son especulativas.

Research question

Can prompt conditioning based on MBTI profiles produce measurable behavioral biases in LLM agents and be leveraged for narrative generation, strategic interaction, and multiagent reasoning; and is the same mechanism formally adaptable to other personality frameworks?

Method

Prompting study in four blocks. (1) GPT-4o mini responds to the 60 items of 16Personalities for each of the 16 profiles, with five responses per item and temperature 1; the responses are sent to the 16Personalities backend and the four dichotomies are represented. (2) Qwen3-235B-A22B generates stories at temperature 0 for 100 prompts from WritingPrompts under 16 profiles and the EXPERT and NONE controls; an LLM judge scores eleven narrative attributes and lexical and readability metrics are additionally calculated in the code. (3) Pairs of agents play repeated dilemmas with one message and one action per round; GPT-4o mini classifies the message intention and defection, strategy changes, and message-action match are calculated. (4) Teams of three agents solve six BIG-Bench/SOCKET tasks through voting, interactive communication, or communication with a private scratchpad. The manuscript does not completely report the number of executions, statistical units, or all the combinations included in the figures, and declares selection of representative results.

Sample: The described psychometric protocol involves 16 profiles × 60 items × 5 responses, that is, 4,800 responses from GPT-4o mini. The narrative task uses 100 prompts × 18 conditions (16 types and two controls), up to 1,800 stories per model if all cells are complete. For the games, the repository proposes seven rounds per pairing and traverses combinations of 16 types plus NONE, ALTRUISTIC, and SELFISH, but the manuscript does not document which exact executions feed each figure nor does it publish the CSVs. In the multiagent task, the script uses teams of three, six team configurations, six tasks, and one execution per configuration; the article does not report the item counts per bar or reproducible intervals.

Findings

  • Verification with 16Personalities strongly separates E/I, T/F, and J/P; S/N is less clear-cut. Since the prompt describes the profile and includes four examples already aligned with the axes, the result is compatible with instruction following and test contamination.
  • Across 100 narrative prompts, Feeling profiles receive higher automatic scores for emotional load, happy endings, and personal character, especially INFP, INFJ, and ISFP; Extraversion is associated in the figure with greater readability, humor, and happy endings.
  • Improvements in cohesion and redundancy over human stories also appear with NONE and EXPERT, so the manuscript considers the specific effect of the profile on those quality properties to be small.
  • In the game protocol, Thinking shows approximately 90% defection compared to 50% in Feeling and changes strategy less (approx. 0.07 vs. 0.16); Introversion shows higher message-action match than Extraversion (approx. 0.54 vs. 0.33).
  • Communication with private self-reflection outperforms interaction without scratchpad on the six tasks shown, but its accuracy is comparable to independent voting and shows no general advantage over it.
  • The supposed generalization to Big Five, HEXACO, Enneagram, and DISC consists of representing each framework as a vector of traits; no experiments, validity tests, or results with those frameworks are presented.
  • The official repository contains a substantial implementation, but the audited version does not compile completely nor reproduce the figures: data/results are missing, the MBTI and games paths are broken, and the interactive judge solves from the task without receiving the dialogue it is supposedly meant to judge.

Limitations

  • The authors state that they present representative results and omit data that does not produce relevant insights; they do not define selection criteria in advance or deliver the complete set of results, which introduces reporting bias.
  • The so-called MBTI evaluation uses the NERIS Type Explorer from 16Personalities. 16Personalities' own documentation explains that its model combines Big Five traits and defines types differently from MBTI; it is not the official MBTI instrument.
  • Induction and testing are not independent: the prompts detail stereotypes such as empathy, honesty, structure, or spontaneity and add four examples with the expected answer for each axis before administering the items.
  • The narrative evaluation depends on a single LLM judge without human validation, inter-judge agreement, sensitivity analysis, or per-story results; furthermore, the prompt requires the model to reason explicitly about its profile, which reinforces the treatment.
  • The article uses significance language but does not provide a complete table of tests, sample sizes, intervals, correction for multiple comparisons, or treatment of dependence between rounds, profiles, and opponents.
  • Honesty is a narrow proxy: another LLM infers the message intention and it is checked whether it matches the action in a game whose prompt expressly authorizes deception. It does not equate to factual truthfulness, integrity, or reliability in other contexts.
  • Length of the justification and response latency do not prove internal deliberation, self-reflection, or greater cognitive depth; they may also arise from style, prompt length, infrastructure, or service variation.
  • The multiagent comparison does not equalize tokens or cost between protocols and the most complex result only approximately matches voting; furthermore, the published code does not pass the conversation to the final judge.
  • Reproducibility is insufficient: there are no raw results, tests, or CI, no lockfile is set, three files fail compilation due to invalid imports, several paths do not exist, and the model/temperature configurations differ from the figures.
  • There is no evaluation with users, patients, healthcare professionals, negotiators, security systems, or justice; the recommendations for those domains extrapolate games and narratives to high-risk contexts.
  • It is an arXiv v1 version and working paper, not a peer-reviewed publication identified in the audit; nor does it include a systematic limitations section or risk evaluation.

What the study does not establish

  • It does not demonstrate that an LLM has human personality, affect, cognition, honesty, intention, or self-reflection; it demonstrates output changes under explicit prompts.
  • It does not prove that a Feeling or Introverted profile makes an agent safer, more empathetic, more reliable, or more suitable for health, negotiation, or justice.
  • It does not demonstrate that self-reflection improves reasoning over independent voting or that personality diversity reduces correlated errors.
  • It does not validate the experimental generalization to Big Five, HEXACO, Enneagram, DISC, multimodal modalities, or embodied agents.
  • It does not causally separate personality from content, length, stereotypes, and examples included in the prompt, nor from differences in model, temperature, cost, or protocol.
  • It does not allow reproducing the main figures from the current official repository nor verifying whether the omitted results contradict the selected conclusions.

Traceability

Scope: Full text

Version: arXiv:2509.04343v1, submitted 4 September 2025, 37 pages

Consulted source: https://arxiv.org/pdf/2509.04343

Review: Codex full-text, bilingual-fidelity, 37-page visual, official-repository, reproducibility, construct-validity, treatment-contamination, selective-reporting, statistical, automated-judge, game-theory, multi-agent, high-risk-extrapolation and evidence-level audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o mini
  • GPT-4o
  • Qwen3-235B-A22B
  • Qwen2.5-14B-Instruct

Instruments and metrics

  • 16Personalities NERIS Type Explorer, 60 items and 7-point response scale
  • MBTI profile prompts with four type-aligned few-shot exemplars
  • PersonaLLM-style automated narrative attributes
  • Lexical richness and readability metrics in the released code
  • Repeated Prisoner's Dilemma and Hawk-Dove games
  • Majority voting, interactive communication and interactive communication with self-reflection
  • Welch t-tests and two-proportion tests in the released analysis code

Data used

  • WritingPrompts
  • BIG-Bench: causal judgment, dark humor detection, simple ethical questions and disambiguation QA
  • SOCKET: complaints and Stanford politeness
  • Model-generated repeated-game transcripts

Evidence and location

  • Objective, contributions, and applied claims: Paper, pp. 1–2, Abstract and Introduction; p. 9, Conclusion
  • Prompt construction, 60-item test, and multiagent protocols: Paper, pp. 4–6, Sections 3.1–4.1; pp. 13–31, Appendix C
  • WritingPrompts design and results: Paper, pp. 6–7, Section 4.2 and Figure 3
  • Defection, strategy changes, honesty, and extrapolations: Paper, pp. 7–8, Section 4.3 and Figures 4–5; pp. 32–33, Appendix D.2
  • Multiagent communication results: Paper, pp. 32–33, Appendix D.1 and Figure 6
  • Selection of results and only formal generalization: Paper, p. 5, Evaluation & Use Cases; pp. 9–12, Sections 6 and Appendices A–B
  • Difference between NERIS Type Explorer and MBTI: 16Personalities, Our Framework, sections Historical Detour and Our Approach: https://www.16personalities.com/articles/our-theory
  • Compilation failures, paths, absence of results, and configuration: Official repository commit ba760ed73b9592798c964ad64db11a1ca4c7ce8f; pyproject.toml; src/MBTITest; src/WritingPrompt; src/MultiAgent-GameTheory
  • The interactive judge does not receive the final message: Official repository commit ba760ed73b9592798c964ad64db11a1ca4c7ce8f; src/MultiAgent-BenchmarkTasks/solve_blackboard_independent.py and solve_blackboard_independent_scratchpad.py
  • Comprehensive visual inspection and layout defects: Paper, all 37 rendered pages; p. 8 malformed text marker and p. 29 stray [H] marker