Human Psychometric Questionnaires Mischaracterize LLM Behavior

Evaluation and psychometric validity2025arXivApproved editorial review

Original title: Established Psychometric vs. Ecologically Valid Questionnaires: Rethinking Psychological Assessments in Large Language Models

Authors: Woojung Song, Dongmin Choi, Yoonah Park, Jongwook Han, Eun-Ju Lee, Yohan Jo

Keywords: Large Language Models, Personality, Psychometrics, Persona, Personality Control

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

6
Authors
15
Findings
38
Limitations
10
Evidence

Editorial summary

English

This preprint compares two ways of profiling eight open-weight LLMs on ten Schwartz values and the five Big Five traits. The first administers PVQ-40, PVQ-21, BFI-44, and BFI-10 as Likert self-reports with two counterbalanced option orders. The second scores fixed candidate responses from the Value Portrait dataset by summed token log-probability: 520 query-response pairs from 104 real-world ShareGPT, LMSYS, Reddit, and Dear Abby queries, previously validated with 681 participants; 286 pairs are value-tagged and 228 trait-tagged when the human correlation reaches r ≥ 0.3. Models are Gemma 3 4B/27B, Qwen 2.5 7B/72B, Qwen 3 30B-A3B/235B-A22B, and GPT-OSS 20B/120B. Long and short versions of the same questionnaire type agree strongly in rank, mean Spearman 0.74 for PVQ and 0.77 for BFI, while agreement with generation-probability profiles is lower: 0.31/0.28 for values and 0.26/0.11 for traits. Established instruments also show much stronger item structure. In a construct-recognition task, seven models obtain mean F1 of 0.69–0.83 on established instruments and 0.09 on VP; five sentence encoders assign the correct construct to 77–81% of established items versus 11–26% of VP items. This is the strongest result: questionnaire wording makes the measured dimension highly transparent and can support construct-consistent answers without establishing a stable disposition. Across eight demographic personas, gender, age, political orientation, and education, model-averaged PVQ shifts resemble marginal ESS subgroup differences: mean cosine 0.60 for PVQ-40 and 0.47 for PVQ-21, with 62/80 and 55/80 matching signs. VP has mean cosine −0.03, aggregate cosine 0.007 with 95% bootstrap CI [−0.221, 0.236], and 40/80 matching signs, indistinguishable from 50%. Normalized shift magnitudes are 0.67/0.71 for the questionnaires, 0.20 for humans, and 0.37 for VP; this is a within-profile relative magnitude, not Cohen’s d. Interpretation must be narrower than the title. Value Portrait is not unconstrained generation: it scores five prewritten responses per query, sums token log-probabilities without length normalization, then averages by scenario and construct. The coverage check only ranks the highest VP candidate against ten free samples; its global median rank is four but model-specific medians range from one to eleven, and this does not validate all five candidates or every construct average. Compared profiles use different instruments, items, scales, and estimators. With only five traits Spearman is coarse; the eight models come from four related families; NDCG resolves Likert ties using NumPy argsort order; and the aggregate RQ2 row is the median of eight p-values rather than a combined test. In RQ4, 80 signs are not independent trials because persona pairs oppose each other, profiles are centered, and Schwartz values are correlated. Marginal ESS subgroup contrasts also do not isolate causal effects of age, education, politics, or country. The paper acknowledges that RQ1–RQ3 lack a matched human baseline and that its probability method is a controlled measurement between questionnaire response and open generation. Code and data are promised upon publication; as of 15 July 2026 the work is listed as under review and no study artifacts were located. The defensible conclusion is that human questionnaires can strongly measure a model’s ability to recognize item meaning and desirability and are insufficient evidence of stable traits or future behavior. The study does not yet establish that VP is a more accurate measure of real behavior or that its profiles predict unconstrained generation.

Español

Este preprint compara dos formas de perfilar ocho LLM abiertos en diez valores de Schwartz y los cinco rasgos Big Five. La primera administra PVQ-40, PVQ-21, BFI-44 y BFI-10 como autoinformes Likert, con dos órdenes de opciones contrabalanceados. La segunda calcula la suma de log-probabilidades de respuestas candidatas fijas del dataset Value Portrait: 520 pares procedentes de 104 consultas reales de ShareGPT, LMSYS, Reddit y Dear Abby, validados previamente con 681 participantes; 286 pares se etiquetan con valores y 228 con rasgos cuando la correlación humana alcanza r ≥ 0,3. Los modelos son Gemma 3 4B/27B, Qwen 2.5 7B/72B, Qwen 3 30B-A3B/235B-A22B y GPT-OSS 20B/120B. El trabajo encuentra alta concordancia de rango entre versiones largas y cortas del mismo tipo de cuestionario, Spearman medio 0,74 para PVQ y 0,77 para BFI, pero concordancia menor con el perfil de generación: 0,31/0,28 para valores y 0,26/0,11 para rasgos. Los cuestionarios también muestran mucha más estructura entre ítems que Value Portrait. En una prueba de reconocimiento de constructos, siete modelos obtienen F1 media 0,69–0,83 en instrumentos establecidos y 0,09 en VP; cinco encoders asignan el constructo correcto a 77–81% de los ítems establecidos, frente a 11–26% de VP. Esta es la evidencia más sólida: la redacción de los cuestionarios revela de forma muy explícita qué dimensión se mide y permite respuestas coherentes con el constructo sin demostrar una disposición estable. En ocho personas demográficas, género, edad, ideología y educación, los desplazamientos medios del PVQ se parecen a diferencias marginales de ESS: coseno medio 0,60 para PVQ-40 y 0,47 para PVQ-21, con 62/80 y 55/80 signos coincidentes. VP obtiene coseno medio −0,03, coseno agregado 0,007 con IC bootstrap del 95% [−0,221, 0,236] y 40/80 signos, indistinguible de 50%. Los desplazamientos normalizados del cuestionario son 0,67/0,71 frente a 0,20 en humanos y 0,37 en VP; esta medida es una magnitud relativa dentro del perfil, no Cohen’s d. La interpretación debe ser más estrecha que el título. Value Portrait no observa generación libre: puntúa cinco respuestas preescritas por consulta, suma log-probabilidades sin normalizar por longitud y luego promedia por escenario y constructo. La validación de cobertura solo comprueba el candidato VP mejor situado entre diez muestras libres; su mediana global es rango 4, pero varía de 1 a 11 según modelo y no valida los cinco candidatos ni el promedio de cada constructo. Los perfiles comparados usan instrumentos, ítems, escalas y estimadores distintos. Además, con solo cinco rasgos Spearman es discreto; los ocho modelos pertenecen a cuatro familias relacionadas; NDCG rompe empates Likert con el orden arbitrario de argsort; y la fila agregada de RQ2 es la mediana de ocho p-valores, no una prueba conjunta. En RQ4, las 80 señales no son ensayos independientes: los pares de personas son opuestos, los perfiles se centran y los valores de Schwartz están correlacionados. Las diferencias marginales de ESS tampoco aíslan efectos causales de edad, educación, política o país. El propio paper reconoce que no existe un baseline humano emparejado para RQ1–RQ3 y que el método de probabilidad es una medición controlada entre cuestionario y generación abierta. Código y datos se prometen para cuando se publique; a 15 de julio de 2026 el trabajo figura como under review y no se localizaron artefactos del estudio. La conclusión defendible es que los cuestionarios humanos pueden medir con mucha fuerza la habilidad del modelo para reconocer el significado y la deseabilidad de sus ítems, y no bastan como prueba de rasgos estables o conducta futura. El estudio no demuestra todavía que VP sea una medida más exacta de comportamiento real ni que sus perfiles predigan generación libre.

Research question

To what extent do the value and trait profiles obtained by asking Likert self-reports to LLMs coincide with profiles constructed from response probabilities to realistic queries, why do the questionnaires appear internally coherent, and do their shifts under demographic personas transfer to that controlled generative behavior?

Method

PVQ-40/PVQ-21 and BFI-44/BFI-10 are administered, in two response orders, to eight models from four families. In parallel, for 104 Value Portrait scenarios, the unnormalized by length sum of log-probabilities of five fixed candidate responses is calculated, and the labeled pairs with r >= 0.3 per scenario and construct are aggregated. Rankings are compared with Spearman and NDCG, item structure with eta squared and intra-construct variance, textual transparency with yes/no recognition and five sentence encoders, and shifts of eight demographic personas with ESS marginal differences.

Sample: Eight open models, organized into four families with small and large variants, are evaluated on 104 VP scenarios and four questionnaires. Value Portrait contains 520 candidates and its prior validation gathered 681 participants. RQ4 uses seven models (Gemma 3 4B is excluded due to high non-response under personas), eight demographic conditions, and 37,398 ESS responses; comparisons are restricted to values because ESS does not contain Big Five.

Findings

  • The current source is arXiv:2509.10078v4, revised on 29 May 2026, with 38 pages and a title and authorship different from the historical record.
  • The 38 pages were rendered and visually inspected.
  • The current PDF matches byte for byte with the cache and has SHA-256 21bd4cc2209ed1ece2306b146c36024317fb7979a3e88139fc887c0d82f4d127.
  • PVQ-40 versus PVQ-21 obtains a mean Spearman of 0.74 and BFI-44 versus BFI-10 obtains 0.77.
  • Generation versus PVQ-40/PVQ-21 obtains 0.31/0.28 and versus BFI-44/BFI-10 obtains 0.26/0.11.
  • The exact sign-change test over eight models reports p=0.004 for values, p=0.016 for Big Five, and p<0.001 when combining both groups.
  • The structure among items is high in questionnaires and close to permutation in VP according to eta squared and intra-construct variance.
  • Seven models recognize constructs in questionnaires with mean F1 0.69-0.83 and in VP with 0.09; GPT-OSS 20B is excluded due to frequent null responses.
  • Five encoders assign the correct construct to 77-81% of the established items and to 11-26% of the VP scenarios.
  • The best VP candidate occupies a global median of 4 among 15 responses and a mean of 5.7, but the per-model medians range from 1 to 11.
  • RQ4 obtains a mean cosine of 0.595 for PVQ-40, 0.466 for PVQ-21, and -0.033 for VP per condition.
  • The aggregated VP cosine is 0.007, 95% CI [-0.221, 0.236], with permutation p of 0.957.
  • The direction coincides in 62/80 dimensions for PVQ-40, 55/80 for PVQ-21, and 40/80 for VP.
  • The mean normalized magnitude is 0.665 for PVQ-40, 0.711 for PVQ-21, 0.373 for VP, and 0.202 for ESS.
  • No repository, derived dataset, or raw results were located; the PDF promises code and data only after publication and the work remains listed as under review.

Limitations

  • Value Portrait scores pre-written candidate responses and does not observe open generation.
  • The sum of log-probabilities is not normalized by length, so longer responses accumulate more penalty even if the content is comparable.
  • The paper does not show an ablation that rules out length as a partial explanation of the differences between candidates or constructs.
  • Checking only the best VP candidate does not validate the representativeness of the five candidates or of all labeled pairs.
  • The position of the best VP candidate varies greatly by model, with medians from 1 to 11.
  • In 43 of 832 model-scenario pairs, fewer than ten free samples remain after filtering.
  • The r >= 0.3 validation relates human judgments about candidates with human self-reports, not LLM VP profiles with observed real behavior.
  • Labeling the same candidate with multiple constructs introduces dependence between scores.
  • The counts per construct are unequal: from 11 to 64 responses and from 9 to 45 scenarios in values.
  • The macro and micro averages change some Spearman cells by up to approximately 0.31.
  • The questionnaires and VP use different items, scales, instructions, and estimators, so divergence does not identify which method is more accurate.
  • There is no external criterion of free behavior with which to measure the predictive accuracy of either of the two profiles.
  • The paper acknowledges that RQ1-RQ3 lack a matched human baseline.
  • With five traits, Spearman has low resolution and the individual tests have low power.
  • The eight models are not independent replicates: they form four families and two sizes per family.
  • The combined test of 16 observations treats value and trait groups of the same model as separate observations.
  • The percentile bootstrap with n=8 can be anti-conservative, a limitation acknowledged by the authors.
  • NDCG assigns different degrees to Likert ties through the NumPy argsort order; the result is deterministic but not psychometrically identified.
  • Using an established questionnaire as the ideal NDCG ranking favors by construction proximity to that instrument, not behavioral truth.
  • The aggregated row of RQ2 is the median of eight p-values per model, not a combined p-value of the sample.
  • The RQ2 permutation tests relabel constructs within the same set of items and do not evaluate generalization to new scenarios.
  • GPT-OSS 20B is excluded from RQ3 due to frequent null responses, but the exact rate and the effect of exclusion are not integrated into the main metric.
  • The recognition F1 depends on unequal multilabel prevalences; near chance it does not have the same baseline value for each construct.
  • Textual transparency is well supported, but the design does not causally isolate how much of the effect comes from recognition, social desirability, alignment, or Likert format.
  • RQ4 uses brief demographic prompts that invite stereotypes and do not represent complete identities or contexts.
  • Gemma 3 4B is excluded from RQ4 due to high non-response; no end-to-end result is offered that penalizes this lack of compliance.
  • The ESS differences are marginal contrasts between subgroups and may mix country, age, education, ideology, and other variables.
  • No causal estimation or multivariable adjustment of the human differences used as a reference is documented.
  • The 80 signs of the binomial test are not independent: the persona pairs are opposites, centering forces dependence, and the ten values form a correlated system.
  • The RQ4 bootstrap resamples only eight conditions, a small sample structured in four pairs.
  • Averaging deltas of seven models can hide disagreement between models; the appendix shows substantial variance, especially in politics and age.
  • The normalized magnitude uses the deviation among ten values within a profile, not a deviation between participants or models; it is not Cohen's d.
  • RQ4 only evaluates Schwartz values because ESS does not provide Big Five.
  • The comparison is limited to eight open models and requires access to token log-probabilities.
  • Closed models, languages other than English, and multi-turn conversations are not evaluated.
  • The fixed score is invariant to temperature and paraphrase only because it does not generate freely; this advantage does not prove complete ecological validity.
  • No code, derived data, per-run profiles, or analysis scripts have been published yet, which prevents reproducing the tables.
  • The article is under review and no peer-reviewed published version was located as of 15 July 2026.

What the study does not establish

  • It does not demonstrate that human questionnaires lack all utility for studying LLMs.
  • It does not demonstrate that the VP profile is a more accurate measure of real behavior than questionnaires.
  • It does not demonstrate that probabilities over fixed candidates predict free generation.
  • It does not demonstrate that the divergence between methods is exclusive to LLMs or greater than in humans.
  • It does not causally demonstrate that lexical transparency or social desirability alone explain the observed structure.
  • It does not validate that each r >= 0.3 label represents a single or stable construct across all contexts.
  • It does not demonstrate statistical independence of models, constructs, demographic conditions, or signs.
  • It does not demonstrate that brief personas simulate complete human demographic identities.
  • It does not demonstrate that ESS differences are causal effects of gender, age, ideology, or education.
  • It does not allow reproducing the results without the artifacts promised after publication.
  • It does not justify using either of the profiles as a psychological diagnosis or safety criterion on its own.

Traceability

Scope: Full text

Version: arXiv:2509.10078v4, initially submitted 12 September 2025, revised 29 May 2026, 38 pages

Consulted source: https://arxiv.org/abs/2509.10078

Review: Codex complete bilingual full-text fidelity pass, current arXiv-v4 metadata reconciliation, all-page PDF visual inspection, generation-probability scoring audit, rank and tie-handling audit, item-structure and construct-recognition assessment, persona/ESS inference audit, and artifact-availability check; summaries written from the full paper and reported appendices rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Gemma 3 4B instruction-tuned
  • Gemma 3 27B instruction-tuned
  • Qwen 2.5 7B Instruct
  • Qwen 2.5 72B Instruct
  • Qwen 3 30B-A3B Instruct 2507
  • Qwen 3 235B-A22B Instruct 2507 in the official FP8 storage checkpoint
  • GPT-OSS 20B
  • GPT-OSS 120B
  • All inference with vLLM 0.16.0 and default Hugging Face chat templates; float16 weights except the stored FP8 Qwen 3 235B checkpoint, bfloat16 compute

Instruments and metrics

  • Portrait Values Questionnaire PVQ-40 and PVQ-21
  • Big Five Inventory BFI-44 and BFI-10
  • Two reversed Likert option-order prompt variants
  • Value Portrait generation-probability scoring over fixed candidate responses
  • Spearman rank correlation and full-list NDCG
  • Eta-squared and z-scored within-construct mean variance
  • LLM binary item-construct recognition with mean F1
  • Five sentence-embedding families for top-1 construct assignment, discrimination and clustering gap
  • Cosine similarity, direction agreement, bootstrap, permutation and within-profile normalized shift magnitude for persona deltas

Data used

  • Value Portrait: 104 real-world queries, five candidates each, 520 query-response pairs
  • Value Portrait validation: 681 human participants and about 46 ratings per query-response pair
  • 286 value-tagged and 228 trait-tagged item-construct pairs at Pearson r ≥ 0.3
  • Source queries from ShareGPT, LMSYS-Chat-1M, Reddit advisory data and Dear Abby
  • European Social Survey round 11 integrated file edition 4.1: 37,398 respondents across 29 European countries plus Israel for RQ4
  • No released snapshot of this study’s code, derived profiles, prompts-as-run, model outputs or analysis tables was located

Evidence and location

  • Version, title, authors, date, and artifact promise: arXiv:2509.10078v4 title page, page 1; metadata checked 15 July 2026
  • Questionnaire design, Value Portrait, models, and scoring: Sections 3.1-3.3, pages 3-4; Appendices B-C, pages 11-13
  • Rankings and RQ1 tests: Section 4.1 and Table 2, pages 4-5; Appendices F.1-F.3 and Tables 8-10, pages 17-19
  • Construct structure RQ2: Section 4.2 and Table 3, pages 5-6; Appendix G.1 and Tables 15-20, pages 21-23
  • Textual transparency and recognition RQ3: Section 4.3, Figure 1 and Table 4, pages 6-7; Appendix H, pages 24-29
  • Personas, ESS reference, cosines, direction, and magnitude: Section 4.4 and Table 5, pages 7-8; Appendix I and Tables 28-39, pages 30-38
  • Limitations acknowledged by the authors: Limitations, page 9
  • Complete visual inspection: All 38 pages of arXiv:2509.10078v4 rendered and visually inspected on 15 July 2026
  • Artifact status: Paper, exact-title, arXiv-ID, GitHub, Hugging Face and official author-page checks on 15 July 2026
  • Methodological audit and metadata reconciliation: reports/verification/article-182-method-and-metadata-audit.json