Open Models, Closed Minds? On Agents Capabilities in Mimicking Human Personalities through Open Large Language Models

Personas, identity, and agents2025AAAIApproved editorial review

Authors: Lucio La Cava, Andrea Tagarelli

Keywords: Open-weight language models, Questionnaire response conditioning, MBTI, 16Personalities NERIS, Big Five Inventory, Role conditioning, Construct validity, Reproducibility

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

2
Authors
14
Findings
26
Limitations
12
Evidence

Editorial summary

English

Open Models, Closed Minds? compares how twelve weight-accessible language models answer personality questionnaires with no conditioning, an explicit target personality, and a target personality plus profession. It is a nine-page AAAI 2025 paper with a 39-page arXiv v3 extended version. Its observable is not an internal psychological personality: it is a set of generated questionnaire choices after self-referential instructions. For RQ0, each model describes 16 MBTI-labelled types and five BFI factors at temperature 0.01; descriptions are compared with references using word overlap and `all-mpnet-base-v2` cosine similarity. Means of 0.67 and 0.66 are interpreted as awareness, but there is no validated threshold, null control, human comparison, or contamination test. The task may simply reproduce public definitions. For RQ1, 60 items copied from 16Personalities and the 44 BFI items are presented separately in randomized order. Thirty repetitions are run at temperatures 0.01 and 0.7. At low temperature, most models collapse to one result, with ENFJ and Judging dominant. Raising temperature diversifies some models but does not show a latent personality emerging. On BFI, nearly all factors are high except Neuroticism and temperature effects are heterogeneous. For RQ2, the system prompt describes one of 16 types or requests a maximal level of one BFI factor. In four-letter classification, SOLAR has mean accuracy 0.785/0.744 at temperature 0.01/0.7; Dolphin 0.654/0.633; NeuralChat 0.577/0.498; and Llama-3-8B 0.517/0.479. Mixtral and Mistral remain near 0.06, while Vicuna, Llama-2-7B, Falcon, Gemma, and Phi-3 also show limited adaptation. This is the strongest result: compliance with this steering procedure varies sharply across checkpoints and usually worsens at higher temperature. For BFI, the paper reports percentage increase in the target factor over baseline; Neuroticism reaches +233.3% for Llama-3-8B. However, only the high pole is conditioned, off-target factor changes are not reported, and a large percentage can originate from a low baseline. Selective control of a Big Five profile is therefore not established. RQ3 adds three professions considered representative of each target. Matching sometimes improves, especially for already responsive models; ENFJ paired with teacher is easiest. The design has no role-only, mismatched-role, random-role, or factorial incremental-effect condition. The profession repeats stereotypical target information, so an independent role mechanism cannot be isolated. The mapping provenance is also not auditable: the text says a group of psychologists selected roles but supplies no count, credentials, agreement, or adjudication, while the appendix displays role categorization in prompt-like form. A major construct error concerns the categorical instrument. The 60 items and web scoring come from 16Personalities, whose official documentation identifies its assessment as NERIS Type Explorer, with five spectrums and explicit differences from Myers-Briggs. The paper calls it MBTI, omits Assertive/Turbulent, and generalizes results as though they came from the genuine instrument. The response-to-type transformation is opaque and unreleased. Reproducibility also falls short of the paper's promise. The appendices disclose prompts, items, BFI key, model labels, and detailed tables, which is a genuine strength, but raw responses, per-run scores, seeds, orders, errors, code, environment, and immutable revisions are absent. AAAI, arXiv, and the author's current publication page do not link a repository, and exact title/ID searches on GitHub find bibliographic lists only. Model names omit organizations and commits; Dolphin is 2.1 in Table 2 but Appendix I links 2.2.1. Generation parameters are internally wrong or ambiguous: `top_p=50` and `top_k=1` are reported, reversing standard defaults of `top_k=50` and `top_p=1.0`. RQ3 adds three roles, yet the stated totals of 11,520 MBTI and 3,600 BFI tests do not multiply by three; Table 7 uses tenths consistent with ten repetitions per role while the method says N=30. The exact denominator cannot be recovered. Finally, the analysis is predominantly descriptive: no intervals, model contrasts, multiplicity strategy, preregistration, power analysis, or equivalence tests are supplied. The faithful conclusion is narrower than the title: these checkpoints show different questionnaire response styles and different compliance with explicit personality steering. Many resist this prompt and four adapt better. The study does not establish intrinsic personality, psychological closed-mindedness, or a footprint identity. Open scoring, raw outputs, fixed versions, seeds, convergent and behavioral psychometrics, spillover controls, retest, role ablations, and human/anti-stereotype validation are needed.

Español

Open Models, Closed Minds? compara cómo doce modelos de pesos accesibles contestan cuestionarios de personalidad sin condicionamiento, con una personalidad explícita y con personalidad más profesión. El trabajo es un paper de AAAI 2025 de nueve páginas y dispone de una versión arXiv v3 extendida a 39 páginas. Su objeto observable no es una personalidad psicológica interna: son elecciones de respuesta producidas tras instrucciones autorreferenciales. Para RQ0, cada modelo describe 16 tipos etiquetados como MBTI y cinco factores BFI a temperatura 0,01; las descripciones se comparan con referencias mediante solapamiento léxico y coseno de `all-mpnet-base-v2`. Los promedios 0,67 y 0,66 se interpretan como awareness, pero no hay umbral validado, control nulo, comparación humana ni prueba de contaminación: la tarea puede resolverse reproduciendo definiciones públicas. Para RQ1, se presentan por separado y en orden aleatorio 60 ítems copiados de 16Personalities y los 44 ítems BFI. Se ejecutan 30 repeticiones a temperaturas 0,01 y 0,7. A baja temperatura, la mayoría de modelos colapsa en un único resultado, con ENFJ y la preferencia J dominantes; subir temperatura diversifica algunos modelos, pero no demuestra que aparezca una personalidad latente. En BFI, casi todos obtienen factores altos salvo Neuroticismo y el efecto térmico es heterogéneo. Para RQ2, el system prompt describe uno de 16 tipos o pide exhibir al máximo un factor BFI. En la clasificación de cuatro letras, SOLAR obtiene accuracy media 0,785/0,744 a temperatura 0,01/0,7; Dolphin 0,654/0,633; NeuralChat 0,577/0,498; Llama-3-8B 0,517/0,479. Mixtral y Mistral quedan alrededor de 0,06; Vicuna, Llama-2-7B, Falcon, Gemma y Phi-3 también muestran adaptación limitada. Este es el resultado más sólido: la obediencia a este procedimiento de steering varía mucho entre checkpoints y suele empeorar al aumentar temperatura. En BFI se informa el porcentaje de subida del factor objetivo respecto al baseline; Neuroticismo llega a +233,3 % para Llama-3-8B. Sin embargo, solo se condiciona el polo alto, no se publican cambios en factores no objetivo y una gran subida porcentual puede partir de una base baja. Por eso no se demuestra control selectivo de un perfil Big Five. Para RQ3 se añaden tres profesiones consideradas representativas de cada target. A veces aumenta el match, sobre todo en modelos ya sensibles; ENFJ asociado a teacher es el caso más fácil. El diseño carece de condición role-only, profesiones incompatibles o aleatorias y análisis factorial del incremento. La profesión repite información estereotípica del target, así que no puede aislarse un mecanismo propio del rol. Además, la procedencia de esas asignaciones no es verificable: el texto dice que las eligió un grupo de psicólogos, pero no informa cuántos, credenciales, acuerdo o adjudicación, y el anexo presenta la categorización con formato de prompt. Hay un error constructivo mayor: los 60 ítems y el scoring proceden de 16Personalities. La documentación oficial de ese sitio identifica su instrumento como NERIS Type Explorer, con cinco espectros y diferencias explícitas respecto a Myers-Briggs. El paper lo llama MBTI, omite la dimensión Assertive/Turbulent y generaliza sus resultados como si fueran del instrumento genuino. La transformación de respuestas a tipo es opaca y no se publica. La reproducibilidad tampoco alcanza la promesa del paper. Los anexos revelan prompts, ítems, claves BFI, modelos y tablas, una fortaleza real, pero faltan respuestas crudas, scores por ejecución, seeds, órdenes, errores, código, environment y revisiones inmutables. AAAI, arXiv y la web actual del autor no enlazan repositorio, y las búsquedas exactas por título/ID en GitHub solo localizaron bibliografías. Los nombres de modelo carecen de organización y commit; Dolphin aparece como 2.1 en Table 2 pero el anexo enlaza 2.2.1. Los parámetros también son internamente erróneos o ambiguos: se declara `top_p=50` y `top_k=1`, al revés de los defaults estándar `top_k=50` y `top_p=1.0`. RQ3 añade tres roles, pero el total 11.520 MBTI y 3.600 BFI no multiplica por tres; Table 7 usa décimas compatibles con diez repeticiones por rol mientras el método afirma N=30. No se puede reconstruir el denominador exacto. Finalmente, casi todo el análisis es descriptivo: no hay intervalos, contrastes entre modelos, multiplicidad, preregistro, power analysis ni equivalencia. La conclusión fiel es más limitada que el título: estos checkpoints muestran estilos de respuesta y grados de obediencia distintos ante un steering explícito basado en cuestionarios. Muchos resisten este prompt y cuatro se adaptan mejor. El estudio no establece personalidad intrínseca, closed-mindedness psicológica ni footprint identity. Para sostenerlo harían falta scoring abierto, outputs crudos, versiones fijas, seeds, psicometría convergente y conductual, controles de spillover, retest, ablation de roles y validación humana contra estereotipos.

Research question

What response patterns do twelve accessible-weight LLMs produce on a 16Personalities questionnaire and on the BFI, how much do they change when congruent types/factors and professions are imposed on them, and does that obedience differ across models and temperatures?

Method

RQ0 asks each model to describe 16 types and five factors and computes word overlap and cosine all-mpnet-base-v2. RQ1 administers 60 questions from 16Personalities and 44 from BFI, one per prompt and in random order, at temperature 0.01/0.7 with 30 repetitions. RQ2 adds system messages with descriptions of each type or factor; it evaluates exact match of four letters and percentage increase of the target BFI factor. RQ3 adds three congruent professions per target. Twelve models are run locally via text-generation-webui and AutoGen is used for interviewer/interviewee roles. The audit read the complete AAAI and extended arXiv v3, visually inspected the 48 pages, contrasted NERIS/MBTI, reviewed formulas, tables, counts, parameters and versions, verified official metadata and searched for the promised code/data on official surfaces and GitHub.

Sample: Twelve checkpoints between 3.8B and 46.7B parameters, two temperatures and 30 declared repetitions. RQ2 reports 12x2x16x30=11,520 categorical tests and 12x2x5x30=3,600 BFI. The same text attributes those totals jointly to RQ2/RQ3 although RQ3 adds three roles; if each role had 30 repetitions, factors x3 would be missing. Table 7 advances in tenths, compatible with ten repetitions per role, but this is not documented. There are no evaluated persons; the supposed group of psychologists that assigns professions is not described.

Findings

  • Awareness obtains mean cosine 0.67 on MBTI-labeled types and 0.66 on BFI, without calibration or baseline.
  • At temperature 0.01, ENFJ and Judging predominate; several models produce a single type almost always.
  • SOLAR achieves mean exact accuracy 0.785/0.744 at temperature 0.01/0.7.
  • Dolphin obtains 0.654/0.633, NeuralChat 0.577/0.498 and Llama-3-8B 0.517/0.479.
  • Mixtral and Mistral remain at approximately 0.06; five other models also do not exceed 0.30 on average.
  • Higher temperature usually reduces exact match, although it diversifies some unconditioned outputs.
  • On BFI, Neuroticism is the most sensitive factor; Llama-3-8B records +233.3 % from its baseline at temperature 0.01.
  • The BFI results show no selectivity because they do not publish spillover on non-target factors.
  • Congruent roles help in some models; ENFJ/teacher is the easiest pairing, without incongruent role control.
  • The real categorical questionnaire is NERIS from 16Personalities, not the genuine MBTI.
  • top_p=50 and top_k=1 are reported, apparently swapped with respect to standard defaults.
  • The total for RQ3 and its denominator per role are not reconstructable from the method and the tables.
  • Dolphin is identified as 2.1 in Table 2 and linked as 2.2.1 in Appendix I.
  • There are no raw outputs, categorical scoring, seeds, environment, code or official data release despite the promise.

Limitations

  • Questionnaire responses do not validate psychological personality, identity or stable internal trait.
  • Each item is answered separately by repeating context; there is no conversational consistency or temporal retest.
  • 16Personalities uses NERIS, not MBTI, and its exact scoring is not public in the artifact.
  • The Assertive/Turbulent dimension of NERIS is omitted although it is part of its five-spectrum framework.
  • Awareness lacks null, threshold, human baseline and memorization/training contamination control.
  • ENFJ/J may arise from alignment, prompt, wording or scoring, not from an inherent personality.
  • Only the high pole of each BFI factor is conditioned.
  • Off-target changes and distance to the full profile are not published.
  • Percentages may be inflated by low baselines and do not include inference or intervals.
  • There is no role-only, mismatched roles, random roles or control against stereotypes.
  • Professions are chosen by congruence, so they duplicate target cues.
  • The number, credentials, agreement or adjudication of the psychologists who supposedly select roles is not reported.
  • The presentation of the appendix leaves ambiguous whether role categorization was also an LLM prompt.
  • The RQ3 counts omit or do not explain the factor of three roles.
  • There are no raw responses, per-run results or reproducible scoring.
  • Seeds, orders, failures, retries and timestamps are missing.
  • Namespaces, hashes and immutable revisions of models are missing.
  • Dolphin 2.1/2.2.1 is inconsistent.
  • top_p/top_k appear swapped, so the actual decoding is ambiguous.
  • No version of Transformers, text-generation-webui, AutoGen, loader, dtype, quantization or chat templates is reported.
  • Model-specific templates add an uncontrolled confound.
  • Models differ simultaneously in size, architecture, base, alignment and data.
  • There is no preregistration, power, intervals, model tests, equivalence or multiplicity.
  • Everything is in English and the StereoSet/profession pool may reinforce occupational stereotypes.
  • There is no behavioral, human, ecological or convergent validation of the obtained profiles.
  • The absence of the promised code prevents verifying calculations and pipeline decisions.

What the study does not establish

  • It does not demonstrate a personality, mind, identity or psychological footprint in the LLMs.
  • It does not demonstrate equivalence between NERIS/16Personalities and MBTI.
  • It does not demonstrate awareness beyond similarity to public definitions.
  • It does not demonstrate that ENFJ/J is an intrinsic trait.
  • It does not demonstrate that temperature reveals latent personality.
  • It does not demonstrate selective control of Big Five profiles.
  • It does not demonstrate an independent effect of role conditioning.
  • It does not demonstrate that profession-personality mappings are valid or free of stereotype.
  • It does not demonstrate causal effects of architecture, size, alignment or openness.
  • It does not demonstrate that top_p=50/top_k=1 were executed exactly under the declared semantics.
  • It does not demonstrate N=30 for each role of RQ3.
  • It does not allow reproducing the figures without outputs, scoring, code and versions.
  • It does not generalize to other languages, instruments, checkpoints or real behavior.

Traceability

Scope: Full text

Version: AAAI proceedings version, published 11 April 2025, 9 pages; audited with arXiv:2401.07115v3, revised 22 March 2025, 39-page extended version

Consulted source: https://doi.org/10.1609/aaai.v39i2.32125

Review: Codex complete bilingual fidelity pass using the full 9-page AAAI proceedings paper and 39-page arXiv v3 extended version, all-page visual inspection, official DOI/publication/version verification, construct comparison of NERIS and genuine MBTI, detailed table/repetition/parameter/model-version audit, prompt and role-control analysis, and current official/GitHub artifact searches; summaries written from full evidence rather than abstract keywords, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Mixtral-8x7B-Instruct-v0.1
  • Llama-2-13b-chat-hf
  • SOLAR-10.7B-Instruct-v1.0
  • Llama-3-8B-Instruct
  • Mistral-7B-Instruct-v0.1
  • Neural-chat-7b-v3-1
  • Dolphin-2.1-mistral-7b, with appendix link to Dolphin 2.2.1
  • Vicuna-7b-v1.5
  • Llama-2-7b-chat-hf
  • Falcon-7b-instruct
  • Gemma-1.1-7b-it
  • Phi-3-mini-4k-instruct

Instruments and metrics

  • 16Personalities NERIS Type Explorer 60-item public questionnaire, mislabelled as MBTI
  • Big Five Inventory 44-item questionnaire
  • all-mpnet-base-v2 cosine similarity
  • Lemmatized word-overlap ratio
  • Exact four-letter target classification frequency
  • Target BFI factor percentage increase
  • Personality-conditioned system prompts
  • Role-plus-personality-conditioned system prompts
  • Temperature comparison 0.01 versus 0.7

Data used

  • Sixty 16Personalities questions reproduced in arXiv Appendix C
  • Forty-four BFI items and scoring key reproduced in Appendices E-F
  • StereoSet-derived list of 120 profession labels
  • Aggregate tables and figures in the paper; raw model outputs not released

Evidence and location

  • Metadata, scope, RQs and headline results: AAAI proceedings pages 1355-1363; DOI 10.1609/aaai.v39i2.32125
  • Design, prompts, temperatures, repetitions and deployment: arXiv:2401.07115v3 pages 3-5, Sections 3.1-3.6 and Figures 1-3
  • RQ1 and RQ2 per model: arXiv v3 pages 6-8 and 19-28, Figures 4-5 and 6-23, Table 3
  • RQ3 roles and results: arXiv v3 pages 8 and 29-30, Tables 7-8
  • Limitations, ethics and code promise: arXiv v3 page 9, Limitations, Ethics Statement and Transparency and reproducibility
  • Items, BFI scoring and personality references: arXiv v3 pages 15-18 and 31-39, Appendices B-F, K-L
  • Difference between NERIS and MBTI: Official 16Personalities Our Framework and Myers-Briggs Company official MBTI assessment pages, checked 16 July 2026
  • Semantics and defaults top_k/top_p: Official Hugging Face Transformers text-generation documentation; top_k default 50 and top_p default 1.0
  • History and change of authors: Official arXiv:2401.07115 submission history v1 13 January 2024, v2 23 June 2024, v3 22 March 2025
  • Absence of locatable official artifact: Official AAAI/arXiv/current author publication links plus exact-title and arXiv-ID GitHub repository/code searches on 16 July 2026
  • Integral audit of construct, method and reproducibility: reports/verification/article-205-open-models-personality-validity-and-reproducibility-audit.json
  • Complete visual inspection: All 9 AAAI pages and all 39 arXiv v3 pages rendered and visually inspected on 16 July 2026