Do LLMs Have Distinct and Consistent Personality? TRAIT: Personality Testset designed for LLMs with Psychometrics

Evaluation and psychometric validity2024arXivApproved editorial review

Original title: Do LLMs Have Distinct and Consistent Personality? TRAIT: Personality Testset Designed for LLMs with Psychometrics

Authors: Seungbeen Lee, Seungwon Lim, Seungju Han, Giyeong Oh, Hyungjoo Chae, Jiwan Chung, Minju Kim, Beong-woo Kwak, Yeonsoo Lee, Dongha Lee, Jinyoung Yeo, Youngjae Yu

Keywords: Computation and Language, Artificial Intelligence, Large Language Models, Personality Assessment, Psychometrics

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

12
Authors
9
Findings
11
Limitations
10
Evidence

Editorial summary

English

This preprint introduces TRAIT, an English-language benchmark of 8,000 multiple-choice questions for studying response patterns associated with the Big Five and Dark Triad. It starts from 44 BFI and 27 SD-3 items: GPT-4 expands them into 1,600 descriptions, selects five of twenty ATOMIC10X situations for each description, and constructs a situation, question, and four recommendations, two labeled high and two low. The result is about 112 times the 71 seeds; the 225× claim in the pipeline figure does not match the item count. Two graduate students with psychology training label only 200 items and reach 97.5% high/low accuracy. This supports the intended polarity in a sample, but does not validate all 8,000 items, trait specificity, factor structure, invariance, or behavioral relevance. The paper calls GPT-4-assigned facet diversity and evenness “content validity,” score change after high/low instructions “internal validity,” and sensitivity to three prompts, option order, and paraphrases “reliability.” These are useful audits of coverage, instruction following, and robustness, but are not by themselves equivalent to those psychometric constructs. Averaged across eight models, TRAIT scores 71.9 for diversity and 77.5 for score difference on the Big Five, and 51.0/83.3 on the Dark Triad; mean sensitivity is 29.8% and 24.4%, respectively. It improves on the baselines, yet changing roughly one quarter to one third of answers shows relative rather than strong consistency. Refusal rates are near zero in multiple choice and 3.1–3.3% in generation, much lower than self-report tests; however, the detector only checks predefined opening phrases and may conflate response format with refusal. TRAIT scoring averages A–D token probabilities under two option layouts and reports the proportion of high choices. It is not a factor score, and two permutations do not exhaust the 24 possible orders. Across nine models, GPT-4 and Claude-Sonnet score 86–87 on Agreeableness and 0–11 on Dark Triad traits. Grouping aligned and base models yields higher Agreeableness (78.3 vs 66.7) and Conscientiousness (91.0 vs 81.7), and lower Openness (56.3 vs 67.8), Extraversion (32.8 vs 46.9), and Dark Triad scores (9.3 vs 27.0). Because this comparison mixes model families and sizes, it does not identify a causal alignment effect. The more controlled Llama2-7B→Tulu2-SFT→Tulu2-DPO sequence suggests that SFT strongly shifts choices, for example +22.9 Agreeableness, −22.9 Extraversion, and −49.8 Psychopathy, while DPO changes little. Yet Table 7's caption describes the subtraction opposite to the interpretation, and T-EVALUATOR, used to label the training data, is itself trained on synthetic TRAIT. The reported .7893 correlation between corpus balance and trait change excludes Openness post hoc, uses only seven traits, and does not establish causality. Three persona prompts yield an average directional score of 85.2, but fine-grained results vary sharply by template and model: GPT-4 high-Psychopathy scores, for example, range from 37.3 to 99.7. Induced correlations among Dark Triad traits reach .90–.97, far above human values; they reflect at least partly shared prompt and generator semantics rather than a natural psychological structure. Five scenarios derived from the same description also alter choices, which is compatible with context sensitivity but also with scenario differences or measurement noise. Finally, correlations between TRAIT and general benchmarks use eight models without controlling model family, size, or alignment and do not establish predictive power of personality. The defensible contribution is a large scenario bank and multifactor response audit, not evidence of internal personality, human-equivalent psychometrics, or behavioral validity.

Español

Este preprint presenta TRAIT, un benchmark en inglés de 8.000 preguntas de elección múltiple para estudiar patrones de respuesta asociados a cinco rasgos Big Five y tres rasgos de la Tríada Oscura. Parte de 44 ítems BFI y 27 SD-3: GPT-4 los expande a 1.600 descripciones, para cada descripción selecciona cinco de veinte situaciones muestreadas de ATOMIC10X y construye una situación, pregunta y cuatro recomendaciones, dos etiquetadas high y dos low. El resultado es unas 112 veces mayor que las 71 semillas; la cifra de 225× que aparece en el diagrama no coincide con el cómputo por ítems. Dos estudiantes de posgrado con formación en psicología etiquetan solo 200 ítems y alcanzan 97,5 % al distinguir high/low. Esta comprobación respalda que una muestra de opciones expresa la polaridad pretendida, pero no valida los 8.000 ítems, su especificidad de rasgo, estructura factorial, invariancia o relación con conducta. El paper denomina «validez de contenido» a la diversidad y uniformidad de facetas asignadas por GPT-4; «validez interna» a la diferencia de puntuación al instruir niveles altos o bajos; y «fiabilidad» a la sensibilidad a tres prompts, al orden de opciones y a paráfrasis. Son operaciones útiles para auditar cobertura, seguimiento de instrucciones y robustez, pero no equivalen por sí solas a esos constructos psicométricos. Promediado sobre ocho modelos, TRAIT obtiene diversidad 71,9 y diferencia 77,5 en Big Five, y 51,0/83,3 en Tríada Oscura; su sensibilidad media es 29,8 % y 24,4 %, respectivamente. Aunque mejora los baselines, que entre una cuarta y una tercera parte de las respuestas cambie muestra consistencia relativa, no estabilidad fuerte. La tasa de rechazo es casi cero en elección múltiple y 3,1–3,3 % en generación, frente a valores mucho mayores en tests de autoinforme; el detector se limita a frases iniciales predefinidas y puede confundir formato con rechazo. La puntuación TRAIT promedia probabilidades de tokens A–D bajo dos disposiciones de opciones y calcula el porcentaje de elecciones high. No es una escala factorial y dos permutaciones no agotan los 24 órdenes posibles. En nueve modelos, GPT-4 y Claude-Sonnet puntúan 86–87 en amabilidad y 0–11 en Tríada Oscura. Al agrupar modelos alineados y base, los primeros promedian más amabilidad (78,3 frente a 66,7) y responsabilidad (91,0 frente a 81,7), menos apertura (56,3 frente a 67,8), extraversión (32,8 frente a 46,9) y Tríada Oscura (9,3 frente a 27,0). Esta comparación mezcla familias y tamaños, de modo que no identifica causalmente el efecto del alineamiento. La secuencia controlada Llama2-7B→Tulu2-SFT→Tulu2-DPO sí sugiere que SFT desplaza con fuerza las elecciones, por ejemplo +22,9 en amabilidad, −22,9 en extraversión y −49,8 en psicopatía, mientras DPO cambia poco; sin embargo, el pie de la Tabla 7 describe la resta en sentido contrario a la interpretación y el clasificador usado para etiquetar los datos de entrenamiento, T-EVALUATOR, se entrena con TRAIT sintético. La correlación 0,7893 entre balance del corpus y cambio de rasgo excluye apertura post hoc, usa solo siete rasgos y no demuestra causalidad. Tres prompts de persona producen una puntuación direccional media de 85,2, pero los resultados finos varían mucho por plantilla y modelo: en GPT-4, por ejemplo, psicopatía alta va de 37,3 a 99,7. Las correlaciones inducidas entre rasgos oscuros llegan a 0,90–0,97, muy por encima de las humanas; reflejan al menos en parte el contenido compartido del prompt y del generador, no una estructura psicológica natural. Las cinco situaciones derivadas de una misma descripción también cambian la elección, algo compatible tanto con sensibilidad contextual como con ruido o diferencias entre escenarios. Finalmente, correlaciones entre TRAIT y benchmarks generales se calculan con ocho modelos y sin ajustar familia, tamaño o alineamiento; no establecen poder predictivo de la personalidad. El aporte defendible es un banco de escenarios amplio y una auditoría multifactor de respuestas, no una prueba de personalidad interna, psicometría humana equivalente ni validez conductual.

Research question

Can a benchmark of scenarios built specifically for LLMs measure differentiated and relatively consistent response patterns, and what do those patterns show about prompting, alignment, and training data composition?

Method

Synthetic construction of 8,000 MCQs from 71 BFI/SD-3 items, 1,600 descriptions generated by GPT-4, and ATOMIC10X situations. 200 items are audited with two annotators, facet coverage with GPT-4, response to high/low prompts, rejections, and sensitivity to three templates, two orders, and paraphrases. The main evaluation uses token probabilities to score nine LLMs; base/aligned comparisons, the Llama2→Tulu2-SFT→Tulu2-DPO sequence, persona prompts, correlations between traits, data classification using T-EVALUATOR, and exploratory correlations with benchmarks are added.

Sample: The benchmark contains 1,000 questions for each of eight traits. Human validation covers 200/8,000 items with two annotators. The main comparison presents nine models and the robustness tables average eight; several appendices expand the set to 19. The prompting experiments use GPT-4, GPT-3.5, Llama2-7B-chat, and Mistral-7B-Instruct; the correlation with benchmarks uses eight models. There are no human participants who take the test or matched human behavioral data.

Findings

  • Two annotators achieve 97.5% at classifying high/low in a sample of 200 items, without evaluating the validity of the trait or of the full scale.
  • TRAIT surpasses baselines on the metrics defined by the authors, but maintains 29.8% average sensitivity in Big Five and 24.4% in Dark Triad.
  • GPT-4 and Claude-Sonnet choose many agreeableness options and few Dark Triad options; profiles vary between models.
  • The aligned group scores higher in agreeableness/conscientiousness and lower in extraversion, openness, and Dark Triad than the base group, although the compared families are not controlled.
  • In the Llama2→Tulu2 sequence, SFT produces large changes and DPO small ones; the labeled composition of the corpus is associated with the direction of seven traits.
  • High/low prompts achieve an average directional score of 85.2, with considerable variability between models, traits, and templates.
  • The induced correlations between Dark Triad are 0.90–0.97 and more extreme than human ones, which contradicts a simple structural equivalence.
  • Changing the scenario derived from the same description frequently changes the choice, showing contextual dependence but not distinguishing valid sensitivity from measurement noise.
  • Exploratory correlations with benchmarks are high for some traits, but are based on eight model-points and are confounded by capability, family, and alignment.

Limitations

  • The generation of descriptions, scenarios, options, and facets depends on GPT models, so it shares semantic and cultural biases of the generator and evaluator.
  • Only 200 of 8,000 items receive human validation; two annotators evaluate binary polarity, not inter-judge agreement, trait specificity, difficulty, ambiguity, or behavioral validity.
  • Validity and reliability labels do not fully follow their psychometric meaning: factorial structure, internal consistency, test–retest, invariance, convergence, discrimination, and external criterion are missing.
  • Scoring is the proportion of high options based on token probabilities; it is not a latent measure and may depend on tokenization, logprob access, and each provider's format.
  • Only two option arrangements are tested; residual sensitivity remains high and paraphrases are also synthetic.
  • Rejection detection uses a list of sentence beginnings and does not validate recall, precision, or semantically equivalent rejections.
  • The aligned/base comparison mixes families and scales; even the paired comparisons are few and heterogeneous.
  • T-EVALUATOR is trained on TRAIT and is used to explain the data that trains the models; this circularity does not demonstrate that its labels represent real traits of the corpus.
  • The 0.7893 correlation excludes openness post hoc and uses seven trait observations; correlations with benchmarks use only eight models and do not control confounders.
  • The evidence is in English, uses Western frameworks, MCQs, and isolated turns; it does not cover open conversation, persistence, other languages/cultures, or real behavior.
  • The article itself alternates figures 112×/225× and contains table footnotes whose direction of subtraction does not match the narrative.

What the study does not establish

  • It does not demonstrate that LLMs possess human personality, consciousness, emotions, or persistent internal traits.
  • It does not validate TRAIT as a psychometric instrument equivalent to BFI or SD-3 for persons or models.
  • It does not demonstrate that choice profiles predict behavior outside the items, multi-turn conversations, or real decisions.
  • It does not causally identify alignment as the origin of differences between models or the trait balance of the corpus as a mechanism of change.
  • It does not prove that prompting creates a stable trait; it shows directional instruction following with marked template dependence.
  • It does not justify interpreting correlations between trait scores and benchmarks as predictive power of personality.
  • It does not establish cross-cultural, clinical, diagnostic, criterion, or human-evaluation-substitution validity.

Traceability

Scope: Full text

Version: arXiv:2406.14703v3 (19 March 2025)

Consulted source: https://arxiv.org/pdf/2406.14703

Review: Codex editorial review, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4 Turbo (gpt-4-turbo-2024-04-09)
  • GPT-3.5 Turbo (gpt-3.5-turbo-0125)
  • Claude Sonnet
  • Llama-2-7B and Llama-2-7B-chat
  • Llama-3-8B and Llama-3-8B-Instruct
  • Mistral-7B and Mistral-7B-Instruct
  • Gemma-2B and Gemma-2B-it
  • Tulu2-7B-SFT and Tulu2-7B-DPO
  • OLMo-7B and OLMo-7B-SFT
  • Additional appendix models: Gemini, Claude Opus, Qwen, Zephyr and others
  • T-EVALUATOR: Mistral-7B-v0.1 with LoRA
  • GPT-4o-2024-08-06 (facet evaluator)
  • GPT-3.5 and Gemini-Pro (paraphrase generation)

Instruments and metrics

  • TRAIT 8,000-item multiple-choice benchmark
  • 44-item Big Five Inventory (BFI)
  • 27-item Short Dark Triad (SD-3)
  • 300-item IPIP-NEO-PI and 120-item subset
  • Anthropic personality evaluation set
  • T-EVALUATOR trait and level classifier
  • Facet diversity and evenness metrics
  • High/low prompt score difference
  • Prompt, option-order and paraphrase sensitivity
  • Keyword-based refusal detector
  • Pearson correlations among traits and with capability benchmarks

Data used

  • TRAIT: 8,000 synthetic scenarios, questions and four-option answers
  • ATOMIC10X commonsense knowledge graph
  • 71 seed items from BFI and SD-3
  • 200-item human-validation subsample
  • Tulu2Mix and UltraFeedback alignment datasets
  • HH-RLHF harmlessness and helpfulness splits
  • UltraChat
  • General capability benchmark leaderboard results for eight models

Evidence and location

  • Objective, construction of 8,000 items and comparison with seeds: arXiv v3, pp. 1–4, Abstract, sections 1–3.1 and Figures 1–2
  • Human validation, T-EVALUATOR, and variation between scenarios: arXiv v3, pp. 5–6, sections 3.2–3.4 and Tables 5–6
  • Model profiles and effect attributed to alignment: arXiv v3, pp. 6–7, sections 4.1–4.2, Figures 3–4 and Table 7
  • Prompting elicitation and intercorrelations: arXiv v3, pp. 7–8, sections 4.3–4.4 and Figures 5–6
  • Definition of diversity, difference, sensitivity, scoring, and rejection: arXiv v3, pp. 13–16 and 24–27, Appendices C–D, G.2 and Tables 15–20
  • Alignment data balance and post hoc correlation: arXiv v3, pp. 17–19 and 31, Appendix E.2, Tables 7 and 14, Figure 13
  • Fine-grained prompt results and human/LLM correlations: arXiv v3, pp. 20‐24, Appendix F–G.3, Tables 11‐13 and 21‐23
  • Correlations with benchmarks and model size: arXiv v3, pp. 25 and 28–29, Appendix H, Figure 12 and Table 22
  • Annotators, interface, and exact prompts: arXiv v3, pp. 25 and 35–38, Appendix I–K, Figure 14 and Tables 27–30
  • Cultural limits, MCQ, anthropomorphism, and misuse: arXiv v3, pp. 9–10, Limitations and Ethical Considerations