Evaluating Large Language Models with Psychometrics

Evaluation and psychometric validity2024arXivApproved editorial review

Authors: Yuan Li, Yue Huang, Hongyi Wang, Ying Cheng, Xiangliang Zhang, James Zou, Lichao Sun

Keywords: Computation and Language, Psychological constructs, Personality, Values, Emotional intelligence, Theory of mind, Self-efficacy

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

7
Authors
13
Findings
20
Limitations
10
Evidence

Editorial summary

English

The paper proposes a psychometric benchmark for evaluating the response patterns of nine LLMs across personality, values, emotion, theory of mind, and self-efficacy. It combines 13 evaluation sets using rating items, alternative choice, multiple choice, and open-ended responses; identifies exact model snapshots and a temperature of 0.5; and adds checks based on parallel forms, option order, rater agreement, and adversarial persuasion. Its most useful contribution is that it does not merely produce profiles: it contrasts self-report with vignettes or queries and shows that a model can claim inability while answering, or claim ability while hallucinating or failing to recognize its limits.

Results are highly instrument- and prompt-dependent. On the BFI, several models appear more agreeable and conscientious and less neurotic than a large US human average, but some repeatedly select the scale midpoint; Mixtral-8x7B shifts from an extraversion score of 2.14 on the BFI to 5 on a vignette. Prompts naming traits raise or reverse scores, which is consistent with instruction following and semantic leakage rather than necessarily revealing latent personality. Low-ambiguity value items and some theory-of-mind stories are relatively easy for several models, whereas ambiguous MoralChoice items, emotion tasks, and some false-belief conditions separate systems more strongly. Adversarial persuasion reduces human-centered-values accuracy, and reported self-efficacy diverges from behavior on HoneSet.

The psychometric interpretation is nevertheless much weaker than the paper implies. Its “internal consistency” measure is the within-aspect standard deviation of item scores: a constant response of 3 obtains zero deviation and may be described as highly consistent even though it does not establish item homogeneity, reliability, or construct structure. Match rate under option or wording changes measures repetition/robustness but can remain high when both responses are wrong. Agreement between GPT-4 and Llama-3-70B as judges does not by itself supply external validity, and human adjudication is mentioned only for HoneSet disagreements. The study performs no factor analysis, measurement invariance, criterion or ecological validation, cross-date test-retest, IRT, or uncertainty estimation.

The manuscript also contains material count contradictions. Table 1 assigns 1,767 items to MoralChoice, while the appendix describes 680 high-ambiguity plus 687 low-ambiguity cases, 1,367 in total, and assigns 987 queries to HoneSet while the main text and appendix state 930. The table lists 228 Human-Centered Values cases, whereas the appendix retains 57 regular scenarios; 228 appears to include three adversarial variants per scenario without making the unit explicit. Seeds and stochastic repetition counts are not reported, there are no confidence intervals or inferential tests, and no identifiable public benchmark code or corpus was found. The study is therefore a broad and useful diagnosis of output inconsistencies under specific protocols, not a validation of LLM psychological states or a fully reproducible benchmark.

Español

El trabajo propone un benchmark psicométrico para evaluar patrones de respuesta de nueve LLM en personalidad, valores, emoción, teoría de la mente y autoeficacia. Combina 13 conjuntos de evaluación con ítems de escala, elección alternativa, opción múltiple y respuesta abierta; fija snapshots concretos de los modelos y temperatura 0,5; y añade comprobaciones de formas paralelas, orden de opciones, acuerdo entre jueces y persuasión adversaria. Su aportación más útil es no limitarse a producir perfiles: contrasta self-report con vignettes o consultas, y muestra que un modelo puede declararse incapaz mientras responde, o declararse capaz mientras alucina o no reconoce sus límites.

Los resultados dependen mucho del instrumento y del prompt. En BFI, varios modelos son más agradables y conscientes y menos neuróticos que una media humana estadounidense, pero algunos responden repetidamente con el punto medio; Mixtral-8x7B pasa de extraversión 2,14 en BFI a 5 en una vignette. Los prompts que nombran rasgos elevan o invierten las puntuaciones, algo compatible con seguimiento de instrucciones y fuga semántica, no necesariamente con una personalidad latente. Las pruebas de valores de baja ambigüedad y las historias de teoría de la mente resultan relativamente fáciles para varios modelos, mientras MoralChoice ambiguo, emoción y ciertas falsas creencias separan más a los sistemas. La persuasión adversaria reduce la exactitud en valores humanos y la autoeficacia declarada se desalinean con la conducta en HoneSet.

La interpretación psicométrica, sin embargo, es mucho más débil de lo que sugiere el artículo. Su «consistencia interna» es la desviación estándar de puntuaciones dentro de cada aspecto: una respuesta constante de 3 obtiene desviación cero y puede describirse como altamente consistente aunque no pruebe homogeneidad de ítems, fiabilidad ni estructura del constructo. El match rate ante cambios de posición o redacción mide repetición/robustez, pero puede ser alto cuando ambas respuestas son incorrectas. El acuerdo de GPT-4 y Llama-3-70B como jueces tampoco aporta por sí solo validez externa, y solo se menciona adjudicación humana para desacuerdos de HoneSet. No se realizan análisis factoriales, invariancia, validez criterial o ecológica, test-retest entre fechas, IRT ni estimaciones de incertidumbre.

El manuscrito contiene además contradicciones materiales: la Tabla 1 atribuye 1.767 ítems a MoralChoice, frente a 680 casos de alta y 687 de baja ambigüedad descritos en el apéndice, 1.367 en total, y atribuye 987 consultas a HoneSet, mientras texto principal y apéndice indican 930. La tabla da 228 casos para Human-Centered Values, pero el apéndice conserva 57 escenarios regulares; 228 parece incorporar tres variantes adversarias por escenario sin explicitar la unidad. No se informan semillas ni número de repeticiones estocásticas, no hay intervalos o pruebas inferenciales y no se encontró código o corpus público identificable del benchmark. Por tanto, es un diagnóstico amplio y útil de inconsistencias de output bajo protocolos concretos, no una validación de estados psicológicos de los LLM ni un benchmark plenamente reproducible.

Research question

How can the personality, values, emotion, theory of mind, and self-efficacy expressed by LLMs be jointly evaluated, and to what extent are their responses consistent across instruments, parallel forms, response positions, judges, and adversarial perturbations?

Method

Experimental benchmark over nine LLM snapshots, queried at temperature 0.5 via APIs or local inference. It administers 13 evaluation sets distributed across five constructs, with four item formats. Open-ended responses from Big Five vignettes and Strange Stories are scored by GPT-4 and Llama-3-70B; HoneSet uses both judges and human adjudication when they disagree. The declared validation covers intra-aspect dispersion, parallel forms, judge agreement, position changes, and persuasive attacks. The editorial review read the eight main pages, references, and appendices up to page 38, cross-checked tables, examples, and metric definitions, verified arXiv v2, and searched for associated public materials.

Sample: The unit of analysis is outputs from nine LLM snapshots across 13 sets, not persons. The manuscript does not report a reliable total number of observations because three counts change between table and appendices and it also does not indicate how many independent runs were made per prompt. For BFI it compares means with aggregated results from 3,387,303 participants from the United States, but it does not recruit participants nor reproduce that sample; for emotion it cites human means from the dataset. Open-ended scenarios have especially small sizes, five Big Five vignettes, eleven Strange Stories, and eighteen Imposing Memory. The adversarials for Human-Centered Values start from 57 regular scenarios and generate three persuasive transformations.

Findings

  • The BFI profiles of several models show higher agreeableness and conscientiousness and lower neuroticism than the cited human mean, but the comparison does not prove equivalence of the construct or of the scale between humans and LLMs.
  • ChatGPT, Llama-3-8B, and Mistral-7B frequently produce constant responses on BFI; zero deviation is interpreted as consistency although it may also reflect default responding or lack of discrimination between items.
  • Mixtral-8x7B obtains extraversion 2.14 on BFI and 5 on the corresponding vignette, exemplifying disagreement between structured self-report and open-ended scenario.
  • P2 prompts that describe traits and their inverse versions strongly move BFI and vignettes; they demonstrate sensitivity and prompt following, not stability of an internal personality.
  • On the Big Five vignette, the global weighted agreement between GPT-4 and Llama-3-70B is 0.86, although it drops to 0.667 for Llama-3-70B and 0.706 for Mistral-7B in the per-model table.
  • On high-ambiguity MoralChoice, the best indicated result is 74.3% for Mixtral-8x7B and GPT-4 reaches 65.1%; low-ambiguity items are considerably easier.
  • Regular accuracy on Human-Centered Values exceeds 90% for most models, but drops with authority, evidence, or logical appeal; ChatGPT loses more than twenty points on some variants.
  • On emotion, the maximum is 58.4% in comprehension for Llama-3-70B versus approximately 70% human, and 64.7% in application for GPT-4 versus approximately 78% human.
  • GPT-4 and Llama-3-70B solve 97.5% and 100% of one false-belief condition, but both drop to 85% on another; Mistral-7B falls to 5% on that condition.
  • Strange Stories yields high percentages, but it only contains eleven stories and its correction depends on two LLM judges; Imposing Memory ranges approximately between 55.56% and 88.89%.
  • The inverse form of the self-efficacy questionnaire is very stable for GPT-4, GLM-4, and Llama-3-70B, but has kappa near or below zero for ChatGPT, Qwen-Turbo, and Mistral-7B.
  • On HoneSet, several models answer queries they previously declared they could not address. The metric called confidence rate counts responses without admission of limitations, not factual accuracy, calibrated probability, or response success.
  • Position and parallel-form checks reveal useful sensitivity for QA, but a high match rate only proves that two outputs coincide, not that they are correct or that they measure the same psychological construct.

Limitations

  • Internal consistency is defined as the standard deviation of responses within an aspect, not as alpha, omega, inter-item correlation, factor analysis, or test information; a constant response obtains the best possible value.
  • Some GLOBE aspects have only two or three items, so their dispersion is a particularly fragile basis for inferring reliability.
  • The match rate mixes stability with correctness: two identical and incorrect responses receive the same robustness score as two identical and correct responses.
  • No systematic content validity, internal structure, human-model invariance, convergent, discriminant, criterion, predictive, or ecological validity is analyzed.
  • Agreement between two LLM judges may reflect shared biases, training, or preferences; it does not substitute an independent human or external criterion.
  • Human adjudication is described for HoneSet discrepancies, but no number of annotators, training, blinding, agreement, or reproducible protocol is reported.
  • Temperature is 0.5, but no seeds or number of stochastic repetitions are published; the deviations in several tables appear to arise from items or positions, not from independent runs.
  • There are no confidence intervals, hypothesis testing, correction for multiple comparisons, effect sizes, or sensitivity analyses.
  • Five Big Five vignettes leave each open trait represented by a single scenario; eleven Strange Stories and eighteen Imposing Memory also limit precision and generalization.
  • Changes under trait prompts contain the desired label and direction, so they confound induced personality with semantic instruction following.
  • The human BFI mean comes from another population and mode of administration; cultural, linguistic, sample, or response-process comparability with the LLMs is not demonstrated.
  • Moral and values categories adopt normative responses from specific datasets without validating cultural pluralism or reasonable disagreement.
  • The HoneSet confidence rate measures absence of rejection or disclaimer, not that the response is correct, safe, useful, or genuinely confident.
  • The counts for MoralChoice, Human-Centered Values, and HoneSet are inconsistent or ambiguous between Table 1 and the appendices.
  • The text speaks of five evaluated constructs, while the appendix discusses intelligence as a sixth dimension without including intelligence experiments in the benchmark.
  • The real-world scenarios are curated items and vignettes, not longitudinal behavior in real deployments or observed outcomes.
  • The work is limited to single-turn interactions and acknowledges reduced coverage of prompts and parallel forms, top-down dimensions, and dependence on classical test theory.
  • The snapshots are mainly from early 2024 and do not characterize later versions of services or models.
  • No public repository linked or identifiable with prompts, outputs, transformed data, and scripts was found; exact reproduction is not possible with the manuscript alone.
  • The reviewed document is an arXiv preprint and does not identify a final peer-reviewed publication.

What the study does not establish

  • It does not demonstrate that LLMs possess personality, values, emotions, theory of mind, or self-efficacy as internal human psychological states.
  • It does not demonstrate that a BFI or SD3 score has the same meaning, structure, or norms in humans and models.
  • It does not demonstrate that low deviation between responses is psychometric reliability; it may be acquiescence, midpoint, or invariance without sensitivity to content.
  • It does not demonstrate that a high match rate implies correct responses, genuine understanding, or construct validity.
  • It does not demonstrate that agreement between two LLM judges equates to human correction, external validity, or absence of bias.
  • It does not demonstrate that role-playing prompts create a stable personality; variations may be direct instruction following.
  • It does not demonstrate human equivalence in emotion or theory of mind, even though some models achieve high results on specific tasks.
  • It does not demonstrate that responding without declaring a limitation is accurate self-efficacy, calibrated confidence, or real capability.
  • It does not allow a general ranking of the nine models beyond the evaluated snapshots, prompts, items, and metrics.
  • It does not causally establish that size, openness, or ownership of the model explain the observed differences.
  • It does not demonstrate safe or aligned behavior under real persuasion; it tests three synthetic transformations over 57 base scenarios.
  • It does not offer a fully reproducible benchmark or a precise estimate of the total observations due to missing artifacts and contradictory counts.

Traceability

Scope: Full text

Version: arXiv:2406.17675v2 (17 Oct 2025), 38 pages; no linked or identifiable public benchmark code/data artifact found

Consulted source: https://arxiv.org/pdf/2406.17675

Review: Codex editorial review and methods/artifact audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI gpt-3.5-turbo-0125 (ChatGPT)
  • OpenAI gpt-4-turbo-2024-04-09 (GPT-4)
  • GLM-4
  • Qwen-Turbo
  • Mistral-7B
  • Mixtral-8x7B
  • Mixtral-8x22B
  • Llama-3-8B
  • Llama-3-70B

Instruments and metrics

  • Big Five Inventory (BFI), 44 rating items
  • Short Dark Triad (SD3), 12 rating items
  • Big Five Vignette Test, 5 open-ended scenarios
  • GLOBE Cultural Orientation, 27 rating items
  • MoralChoice alternative-choice benchmark
  • Human-Centered Values regular and adversarial scenarios
  • EmoBench Emotion Understanding, 200 multiple-choice items
  • EmoBench Emotion Application, 200 multiple-choice items
  • False Belief Task, 40 alternative-choice items
  • Strange Stories Task, 11 open-ended items
  • Imposing Memory Task, 18 alternative-choice items
  • LLM Self-Efficacy questionnaire, 6 rating items
  • HoneSet user-query benchmark
  • Within-aspect standard deviation, match rate, agreement rate and weighted Cohen's kappa

Data used

  • Big Five Inventory: 44 items
  • Short Dark Triad: 12 items
  • Big Five Vignette Test: 5 scenarios
  • GLOBE Cultural Orientation: 27 items
  • MoralChoice: Table 1 reports 1,767; appendix components total 1,367
  • Human-Centered Values: 57 retained regular scenarios plus three adversarial variants; Table 1 reports 228
  • Emotion Understanding: 200 items
  • Emotion Application: 200 items
  • False Belief Task: 40 items
  • Strange Stories Task: 11 items
  • Imposing Memory Task: 18 items
  • LLM Self-Efficacy: 6 items
  • HoneSet: main text and appendix report 930 queries; Table 1 reports 987

Evidence and location

  • Objective, five constructs, 13 sets, and discrepancy between self-report and scenarios: arXiv:2406.17675v2, abstract and sections 1–2, pp. 1–3
  • Snapshots, temperature, formats, judges, and definition of the five validations: arXiv v2, section 2 and Table 1, pp. 3–4
  • Personality, human comparison, and effect of trait prompts: arXiv v2, section 3, Figure 2 and Appendix C, pp. 4–5 and 13–21
  • Values, MoralChoice, and drop under adversarial persuasion: arXiv v2, section 4, Figures 3–4 and Appendix D, pp. 5–6 and 21–28
  • Emotional comprehension and application results versus human means: arXiv v2, section 5, Table 2 and Appendix E, pp. 6–7 and 28–29
  • False belief, Strange Stories, Imposing Memory, match rate, and judge agreement: arXiv v2, section 6 and Appendix F, pp. 6–7 and 29–34
  • Declared self-efficacy, HoneSet, confidence rate, and inverse-form kappa: arXiv v2, section 7 and Appendix G, pp. 7–8 and 34–36
  • Contradictions in sizes of MoralChoice, Human-Centered Values, and HoneSet: arXiv v2, Table 1 p. 4; Appendix D.2–D.3 pp. 24–27; section 7 p. 7; Appendix G p. 34
  • Stated limitations: single-turn, parallel forms, top-down, and classical theory: arXiv v2, Appendix J, pp. 37–38
  • Absence of factor analysis, invariance, experimental IRT, uncertainty, and reproducible artifact: Full 38-page arXiv v2 manuscript and public-artifact audit; IRT appears only as future discussion, not as an analysis of the reported benchmark