Humanizing LLMs: A Survey of Psychological Measurements with Tools, Datasets, and Human-Agent Applications

Reviews, theory, and governance2025arXivApproved editorial review

Authors: Wenhan Dong, Yuemeng Zhao, Zhen Sun, Yule Liu, Zifan Peng, Jingyi Zheng, Zongmin Zhang, Ziyi Zhang, Jun Wu, Ruiming Wang, Shengmin Xu, Xinyi Huang, Xinlei He

Keywords: Computers and Society, Computation and Language, Human-Computer Interaction, Machine Learning, Psychological traits, Trustworthy AI

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

13
Authors
20
Findings
69
Limitations
16
Evidence

Editorial summary

English

This preprint provides a broad narrative review of how psychological theories and instruments are transferred to the study of LLMs. It organizes 154 references across six areas: assessment tools, LLM-specific datasets, consistency and stability, findings attributed to model families, methods for inducing or editing personality, and simulations of human behavior. It describes personality questionnaires such as MBTI, BFI, IPIP-NEO, SD3, HEXACO, and SSCS; affect and aggression measures such as PANAS and BPAQ; and theory-of-mind tasks including false-belief tests, Strange Stories, Imposing Memory, and Faux Pas. It also collects model-oriented or repurposed resources including MPI, TRAIT, SECEU, EQ-Bench, EmoBench, ToMi, a multidomain psychometrics benchmark, and SocialIQA.

Its most useful contribution is taxonomic. It distinguishes response consistency under changes in option order, prompts, temperature, wording, reverse-scored items, or repeated administration; statistical consistency through standard deviation, alpha, omega, ICC, and test-retest; and agreement between self-report and behavior. It separates these notions from stability after fine-tuning, parameter changes, or personality induction and editing, while acknowledging that the literature uses “stability” inconsistently. For trait simulation it summarizes explicit prompting, implicit scenarios, fine-tuning, and model editing; for applications it distinguishes social experiments, games, and interactive negotiation. This vocabulary helps avoid treating a questionnaire, a social-capability benchmark, and agent behavior as equivalent.

The profiles of GPT, LLaMA, Mistral, Qwen, Claude, and Gemini, however, do not come from a shared experiment or meta-analysis. They combine different generations, sizes, versions, languages, prompts, instruments, and dates. The paper itself reports contradictory GPT-4 SD3 results and task-dependent variability, yet the opening figure compresses each family into one “Psychological ID,” and the conclusion says GPT-4 performs best across all dimensions. No common model-by-test-by-condition matrix supports that comparison. Serving as an EQ-Bench judge, succeeding on an emotion task, producing empathic language, and answering a self-report questionnaire also do not establish the same psychological construct. Several claims in the Claude, Gemini, and Qwen sections lack a direct citation or turn how a model was used into evidence of ability.

The central limitation is that the paper calls itself systematic without reporting a systematic-review method. It gives no bibliographic databases, search string, search period, cut-off date, inclusion or exclusion criteria, screening process, flow diagram, included-study corpus, extraction protocol, reviewers, disagreement procedure, or quality and risk-of-bias assessment. The 154 references mix primary LLM studies, human psychometric validation studies, reviews, model web pages, and general resources; they do not form a defined evidence set another team could reconstruct. Reliability or validity in human respondents also does not automatically transfer to a model: saying that MBTI has a seventh-grade reading requirement only indicates that an LLM can process the text, not that its scores are valid, invariant, or interpretable as personality.

The paper appropriately warns about sensitivity to prompts, option order, temperature, limited diversity, and mismatches between human instruments and models. It also acknowledges that LLMs cannot fully replace human participants. Its proposals to improve safety by “optimizing” traits and to use models as human samples nevertheless remain hypotheses, not findings established by the review. There is no new experiment, meta-analysis, extraction table, code, or data; the official source archive contains TeX, bibliography, and seven figures. The defensible contribution is a broad, visual entry point to tools, datasets, and applications through April 2025. It is not a systematic estimate of the evidence, does not establish stable family-level psychological profiles, and does not validate clinical or population-level uses of these measurements.

Español

Este preprint ofrece una revisión narrativa extensa de cómo se trasladan teorías e instrumentos psicológicos al estudio de los LLM. Organiza 154 referencias en seis frentes: herramientas de evaluación, conjuntos de datos específicos para LLM, consistencia y estabilidad, resultados atribuidos a distintas familias de modelos, métodos para inducir o editar personalidad y simulaciones de conducta humana. Describe cuestionarios de personalidad como MBTI, BFI, IPIP-NEO, SD3, HEXACO y SSCS; medidas de afecto y agresión como PANAS y BPAQ; y tareas de teoría de la mente como falsa creencia, Strange Stories, Imposing Memory y Faux Pas. También reúne recursos diseñados o reutilizados para modelos, entre ellos MPI, TRAIT, SECEU, EQ-Bench, EmoBench, ToMi, un benchmark psicométrico multidominio y SocialIQA.

Su aportación más útil es taxonómica. Distingue la consistencia de respuestas bajo cambios de orden, prompt, temperatura, formulación, ítems invertidos o repeticiones; la consistencia estadística mediante desviación estándar, alfa, omega, ICC y test-retest; y la coherencia entre autoinforme y conducta. Separa esto de la estabilidad después de fine-tuning, cambios de parámetros o inducción y edición de personalidad, aunque reconoce que la literatura usa «estabilidad» de manera poco uniforme. Para simulación de rasgos resume prompting explícito, escenarios implícitos, fine-tuning y edición de modelos; para aplicaciones distingue experimentos sociales, juegos y negociación interactiva. Este vocabulario ayuda a no tratar como equivalentes un cuestionario, un benchmark de capacidad social y una conducta de agente.

Los perfiles de GPT, LLaMA, Mistral, Qwen, Claude y Gemini, sin embargo, no proceden de un experimento común ni de un metaanálisis. Combinan generaciones, tamaños, versiones, idiomas, prompts, instrumentos y fechas diferentes. El propio texto presenta resultados contradictorios para GPT-4 en SD3 y variabilidad por tarea, pero la figura inicial condensa cada familia en un único «Psychological ID» y la conclusión afirma que GPT-4 rinde mejor en todas las dimensiones. Esa comparación no está respaldada por una matriz común de modelos, pruebas y condiciones. Ser juez de EQ-Bench, acertar una tarea emocional, producir lenguaje empático o responder un autoinforme tampoco demuestra el mismo constructo psicológico. Varias frases de las secciones de Claude, Gemini y Qwen carecen de una cita directa o convierten decisiones de uso en evidencia de capacidad.

La principal limitación es que el artículo se denomina sistemático sin proporcionar un método de revisión sistemática. No declara bases bibliográficas, consulta, periodo de búsqueda, fecha de corte, criterios de inclusión o exclusión, proceso de cribado, diagrama de flujo, corpus de estudios incluidos, protocolo de extracción, revisores, resolución de desacuerdos ni evaluación de calidad o riesgo de sesgo. Las 154 referencias mezclan estudios primarios de LLM, validaciones psicométricas humanas, revisiones, páginas de modelos y recursos generales; no forman un conjunto de evidencia definido que otro equipo pueda reconstruir. Además, la fiabilidad o validez de un cuestionario en personas no se transfiere automáticamente a un modelo: que MBTI tenga una exigencia lectora de séptimo curso solo indica que un LLM puede procesar el texto, no que sus puntuaciones sean válidas, invariantes o interpretable como personalidad.

El trabajo acierta al advertir de sensibilidad al prompt, orden de opciones, temperatura, falta de diversidad y desajuste entre herramientas humanas y modelos. También reconoce que los LLM no sustituyen de forma completa a participantes humanos. Aun así, sus propuestas de mejorar seguridad «optimizando» rasgos y de emplear modelos como muestras humanas son hipótesis, no resultados demostrados por la revisión. No hay experimento nuevo, metaanálisis, tabla de extracción, código ni datos; el archivo fuente oficial contiene TeX, bibliografía y siete figuras. La contribución defendible es una puerta de entrada amplia y visual al espacio de herramientas, conjuntos de datos y aplicaciones hasta abril de 2025. No es una estimación sistemática de la evidencia, no establece perfiles psicológicos estables por familia y no valida el uso clínico o poblacional de estas mediciones.

Research question

What psychological tools, datasets, consistency and stability metrics, empirical results, personality simulation methods, and agent applications have been used to study psychological characteristics attributed to LLMs, and what methodological problems remain open?

Method

Narrative review organized into six dimensions and supported by 154 references. It summarizes human psychological instruments, benchmarks specific to or adapted for LLMs, work on consistency and stability, published results for six model families, prompting, fine-tuning, and editing techniques, and three types of human role simulation. It presents no search, selection, extraction, or quality evaluation protocol, so "systematic" describes the thematic organization and not a reproducible procedure. The audit read and rendered the 26 pages, inspected the seven figures and the 154 references, reviewed the TeX and the bibliography of the official source file, and performed a targeted search for artifacts.

Sample: The review cites 154 references, but does not state how many were candidates, screened, excluded, or finally considered included studies. The bibliography combines publications on LLMs, scale validations in humans, reviews, model web pages, and general works. There is no corpus table with model, version, date, language, prompt, sample, instrument, metric, and quality that would allow calculating coverage or comparing results in a homogeneous manner.

Findings

  • The review organizes the field into tools, datasets, evaluation, model analysis, personality simulation, and human applications.
  • Traditional tools are grouped into personality, emotion and mental health, and theory of mind.
  • The use in LLMs of MBTI, BFI, IPIP-NEO, SD3, HEXACO, and SSCS is documented.
  • PANAS and BPAQ are presented as measures of affect and aggression, although their described validation comes mainly from humans.
  • False belief tasks, Strange Stories, Imposing Memory, and Faux Pas cover different levels of social reasoning.
  • MPI and TRAIT attempt to adapt personality assessment to the LLM domain.
  • SECEU, EQ-Bench, and EmoBench operationalize distinct components of emotional understanding.
  • ToMi and composite benchmarks attempt to evaluate theory of mind and multiple psychological dimensions.
  • Consistency is decomposed into condition sensitivity, logical and semantic coherence, statistical consistency, and congruence with human evaluation.
  • Stability is used for persistence after fine-tuning, parameter changes, and personality induction or editing.
  • The reviewed literature reports sensitivity to option order, prompt formulation, and temperature.
  • Self-report scores do not always agree with generated behaviors or indirect tasks.
  • Results attributed to the same family change between versions, instruments, and conditions.
  • The text collects contradictory SD3 results for GPT-4 between two works.
  • Explicit and implicit prompting, fine-tuning, and editing can direct styles or responses compatible with target traits.
  • The consistency of induced personality remains a problem, especially in prolonged dialogue and different contexts.
  • Applications are divided into simulation of social experiments, games, and interactive negotiation.
  • The review recognizes reduced diversity of thought and overly uniform responses when using LLMs as simulated participants.
  • The work identifies mismatches between human tools and model capabilities as a central problem.
  • The official source file contains no executable experiments, extraction data, or reproducible review corpus.

Limitations

  • The document is an arXiv v1 preprint and does not declare peer review or editorial acceptance.
  • It calls itself a systematic review without including a review method section.
  • It does not identify the bibliographic databases consulted.
  • It does not publish search strings or queries.
  • It does not declare search dates or an explicit cutoff date.
  • It does not define inclusion criteria.
  • It does not define exclusion criteria.
  • It does not report eligible languages or document types.
  • It does not describe deduplication of results.
  • It does not report the number of records found.
  • It does not describe screening of titles, abstracts, or full texts.
  • It does not present a PRISMA diagram or equivalent flow.
  • It does not list excluded studies or reasons for exclusion.
  • It does not provide a separate list of included studies.
  • It does not publish an extraction form or protocol.
  • It does not declare how many reviewers selected or extracted evidence.
  • It does not describe disagreement resolution or inter-reviewer reliability.
  • It does not assess methodological quality or risk of bias of the cited studies.
  • It does not register a prior protocol.
  • It does not provide a corpus characteristics table.
  • It does not allow distinguishing the exhaustiveness of narrative selection.
  • The 154 references mix primary evidence, human validation, reviews, and product pages.
  • Evidence is not weighted by design, sample size, replication, or peer review.
  • No meta-analysis or quantitative synthesis is performed.
  • Metrics with different scales and directions are not harmonized.
  • Model names, sizes, and snapshots are not harmonized.
  • Access dates of changing APIs are not harmonized.
  • Languages, prompts, temperatures, or response formats are not harmonized.
  • Comparisons between families come from different studies and conditions.
  • The claim that GPT-4 is better in all dimensions does not come from a common benchmark.
  • Figure 1 compresses multiple generations of each family into a single psychological profile.
  • Family profiles hide contradictions and internal variability.
  • The figure uses trait labels as if they were stable attributes of the model.
  • Self-report scores are conceptually combined with agent capabilities and behaviors.
  • Measured trait, output style, and competence in a task are not consistently separated.
  • Human reliability of a scale does not demonstrate reliability in LLMs.
  • Human construct validity does not demonstrate the same construct in a text generator.
  • The applicability of MBTI is justified by reader demand, not by psychometric validation in models.
  • The presentation of MBTI does not sufficiently develop its psychometric limitations despite citing a critique.
  • Measurement invariance across models, versions, languages, or prompts is not required.
  • Recoverable factorial structure is not required to interpret questionnaire dimensions.
  • Repeatability of a contextual performance and persistence of a trait are not always distinguished.
  • The work itself recognizes that stability has heterogeneous definitions.
  • The categories of emotion, mental health, aggression, and personality encompass non-equivalent constructs.
  • The mental health label rests mainly on affect and aggression, not on comprehensive clinical evaluation.
  • Theory of mind results may depend on contamination, wording, format, and explanation demand.
  • Analogies with child ages do not necessarily share protocol, modality, or cognitive process.
  • The expression human-like is used for response similarity without establishing a human mechanism.
  • Being used as a judge for EQ-Bench does not by itself validate emotional intelligence.
  • Being used as a role model in an evaluation does not demonstrate emotional experience.
  • Several claims about Claude lack complete data or direct citation.
  • Several claims about intercultural advantages of Gemini do not provide a direct reference in the paragraph.
  • Several claims about multimodal Qwen and industrial level do not provide a direct measure or source.
  • The text presents as evidence some non-verifiable engineering inferences.
  • There is no formal analysis of citation errors or certainty of each claim.
  • There is no dedicated section on limitations of the review itself.
  • There is no dedicated section on ethics or conflicts of interest visible in the manuscript.
  • Benchmark contamination and questionnaire knowledge are not sufficiently discussed.
  • The dependence of responses on alignment training and provider policies is not analyzed.
  • The speed of the field means that profiles from April 2025 age with each snapshot.
  • The recommendation to optimize personality to improve safety is based on associations, not on general causal validation.
  • Safety effects in deployments, attacks, or harm outcomes are not tested.
  • The substitution of humans is discussed without a complete framework of population validity.
  • Similarity in specific tasks does not reproduce demographic, cultural, or clinical heterogeneity.
  • No review code or bibliography encoded as a dataset is published.
  • No extraction notes or inclusion decisions are published.
  • The official source file contains TeX, bibliography, and figures, but no additional reproducible evidence.
  • The targeted search did not locate an official repository associated with the article.
  • The seven figures are legible, but their visual clarity does not resolve the lack of methodological traceability.

What the study does not establish

  • It does not establish that LLMs possess mind, emotion, or human personality.
  • It does not demonstrate stable psychological profiles for the GPT, LLaMA, Mistral, Qwen, Claude, or Gemini families.
  • It does not allow comparing which family is better in all dimensions under equivalent conditions.
  • It does not prove that questionnaires validated in humans are valid for LLMs.
  • It does not demonstrate psychometric invariance across models, prompts, languages, or versions.
  • It does not convert response consistency into evidence of an internal state.
  • It does not correctly equate self-report, emotional capacity, theory of mind, and social behavior.
  • It does not demonstrate that prompting or editing creates a persistent personality outside the context.
  • It does not establish that optimizing traits causally improves general safety.
  • It does not validate the use of LLMs as representative substitutes for human participants.
  • It does not validate clinical, diagnostic, or therapeutic applications.
  • It does not quantify prevalence, effect size, or certainty of the field's findings.
  • It does not demonstrate exhaustiveness of the reviewed literature.
  • It does not allow reproducing the selection and synthesis of the 154 works.
  • It does not by itself constitute a systematic review according to reproducible reporting standards.

Traceability

Scope: Full text

Version: arXiv:2505.00049v1 (30 Apr 2025); DOI 10.48550/arXiv.2505.00049; CC BY 4.0

Consulted source: https://arxiv.org/pdf/2505.00049v1

Review: Codex full-text, visual, systematic-review-method, psychometric, construct-validity, citation-traceability and source-integrity audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT family: GPT-3, GPT-3.5, GPT-4, GPT-4o and smaller variants reported across cited studies
  • LLaMA family: LLaMA, LLaMA 2 and LLaMA 3 variants from 7B to 70B reported across cited studies
  • Mistral family, principally Mistral-7B and Mixtral variants
  • Qwen family: Qwen, Qwen 1.5 and Qwen 2.5 variants
  • Claude family: Claude 3 Opus and Sonnet through Claude 3.7 references
  • Gemini family: Gemini 1.0, 1.5, 2.0 and 2.5 references
  • No models were newly run or compared by the survey authors

Instruments and metrics

  • Myers-Briggs Type Indicator (MBTI)
  • Big Five Inventory (BFI)
  • IPIP-NEO 300-, 120- and 60-item variants
  • Short Dark Triad (SD3)
  • HEXACO-60 and HEXACO-100
  • Short Scale for Creative Self (SSCS)
  • Positive and Negative Affect Schedule (PANAS)
  • Buss-Perry Aggression Questionnaire and short form
  • False-belief tasks including Sally-Anne and Smarties
  • Strange Stories Test
  • Imposing Memory Test
  • Faux Pas Test
  • Consistency measures: option-order symmetry, prompt sensitivity, reverse-item agreement, standard deviation, Cronbach alpha, McDonald omega, ICC and test-retest
  • Behavioral and self-report congruence measures

Data used

  • Machine Personality Inventory (MPI)
  • TRAIT personality assessment dataset
  • Situational Evaluation of Complex Emotional Understanding (SECEU)
  • EQ-Bench
  • EmoBench
  • Theory of Mind interactions dataset (ToMi)
  • Li et al. multidomain psychometrics benchmark built from thirteen datasets
  • SocialIQA
  • ToMBench and other ToM datasets discussed in model-family sections
  • Narrative bibliography of 154 mixed references rather than a declared included-study dataset

Evidence and location

  • Objective and six dimensions of the review: arXiv:2505.00049v1, abstract and introduction, pp. 1–2
  • Tools for personality, emotion, aggression, and theory of mind: Paper, section 2, pp. 2–6
  • Datasets specific to or adapted for LLMs: Paper, section 3 and Figure 5, pp. 6–8
  • Taxonomy of consistency and stability: Paper, section 4 and Figure 6, pp. 9–12
  • Narrative profiles of GPT, LLaMA, Mistral, Qwen, Claude, and Gemini: Paper, section 5, pp. 13–15; Figure 1, p. 1
  • Contradictory SD3 results for GPT-4: Paper, section 5.1 Personality Traits, p. 13
  • Prompting, fine-tuning, and personality editing: Paper, section 6, pp. 15–16
  • Social experiments, games, and negotiation: Paper, section 7 and Figure 7, pp. 16–18
  • Comparison with previous reviews and future proposals: Paper, sections 8–9, p. 18
  • Claim of GPT-4 as better in all dimensions: Paper, conclusion, p. 18
  • Absence of search, selection, and quality protocol: Full-paper structure and official main.tex audit: sections 1–10 contain no review-method section; checked 15 Jul 2026
  • Composition and number of references: Paper, references 1–154, pp. 19–26; official main.bbl count
  • Version, date, DOI, and license: arXiv:2505.00049v1 metadata; DOI 10.48550/arXiv.2505.00049; CC BY 4.0; checked 15 Jul 2026
  • Comprehensive reading and visual inspection: All 26 PDF pages rendered and inspected, including Figures 1–7 and references; checked 15 Jul 2026
  • Official source file without review code, data, or corpus: .cache/editorial-sources/article-076/supplements/audit/arxiv-2505.00049v1-source.tar; sha256 e07a1870bf07bbe6d385023f4390ea0999f385bae7a9eebb8162f5781a4ee920
  • Absence of associated official repository: Paper, arXiv page and targeted title/author/GitHub search; checked 15 Jul 2026