Rethinking psychometrics through LLMs: how item semantics shape measurement and prediction in psychological questionnaires

Evaluation and psychometric validity2025Scientific ReportsApproved editorial review

Authors: Federico Ravenda, Antonio Preti, Michele Poletti, Antonietta Mira, Andrea Raballo

Keywords: Psychometrics, Questionnaire semantics, Semantic similarity, Big Five, DASS-42, GAD-7, PHQ-9, PsychoLLM, Construct validity, Reproducibility

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

5
Authors
15
Findings
27
Limitations
11
Evidence

Editorial summary

English

Rethinking psychometrics through LLMs asks whether linguistic similarity among questionnaire items resembles the correlation structure of human responses and whether that relationship can predict responses to new items. It does not evaluate synthetic personality or an internal LLM trait: pretrained text representations are used to analyze human psychometric data. The matrix analyses use 1,015,341 web responses for a 50-item Big Five questionnaire and 39,775 for DASS-42; 1,000 observations from each source are sampled for prediction. Ten representations are compared: BERT, RoBERTa, DistilBERT, three Sentence-BERT models, T5 and three OpenAI embedding APIs. Pairwise cosine similarity is compared with the magnitude of empirical item correlation. With text-embedding-3-large, the most correlated item is among the three nearest semantic neighbors for 95.2% of the 42 DASS items and 82.0% of the 50 Big Five items. DASS results are Pearson r=0.77 and Spearman rho=0.73; Big Five results are r=0.67 and rho=0.64. This is useful evidence of correspondence between wording and covariance, but it does not identify causality: questionnaire items are deliberately written to express related facets and share construct content. The paper also forces three factors on the DASS semantic matrix, applies varimax rotation and displays loadings grouped by theoretical subscale. Without an independent factor-number criterion, congruence statistic, fit measure, null matrix or external validation, this illustrates alignment but does not show that semantic analysis independently discovers the structure or that semantics predetermines the construct. PsychoLLM is a lightweight network. It selects the target's three most similar items, transforms six similarities through a Dense(10, GELU) layer and a three-weight softmax, combines the three observed responses, and applies a custom ordinal layer. The supplement states learning rate 0.01, batch size 32 and three epochs. In Leave-One-Item-Out evaluation, PsychoLLM reaches MAE/MAPE 0.54/0.28 on DASS, nearly tied with gradient boosting at 0.55/0.28 and slightly worse than random forest at 0.53/0.26. On Big Five it reaches 0.78/0.35, worse than random forest at 0.70/0.30 and gradient boosting at 0.72/0.30. An additional LOSO analysis predicts each DASS subscale from the other two with MAE 0.50-0.60. On C19PRC Wave 6, training on PHQ-9 to predict GAD-7 yields MAE 0.34 and 71% exact match; the reverse direction yields MAE 0.36 and 70%. These instruments assess related symptoms, use the same four-category scale and come from the same survey wave. The analyzed N and any majority-class, item-mode, nearest-item or balanced-accuracy baseline are absent. Consequently, the headline 70-71% does not demonstrate improvement over a trivial rule and may be driven by frequent zero responses. A decisive conceptual limitation is that the paper calls PsychoLLM unsupervised and says it does not require labeled training data, while the supplement explicitly states that it optimizes cross-entropy between predicted and actual responses. For every non-target item, real human scores serve as training labels; only the target-item labels are unseen. This is self-supervised transfer across items, not label-free learning or prediction without collected responses. In addition, supervised baselines use five respondent folds, whereas PsychoLLM is trained on other items from the same 1,000 respondents whose target item it evaluates; no clear person-level holdout is described. The comparison therefore does not isolate generalization to new respondents and is not protocol-matched. The official repository audit adds further caution. Useful public materials exist at main commit 9c7e501: DASS data, 3,072-dimensional embedding matrices, a PsychoLLM implementation and several scripts. But this is not an executable reproduction. It lacks a requirements/lockfile, environment, license, instructions, raw Big Five responses, C19PRC responses, PHQ/GAD experiment code, baselines, factor analysis, LOSO, outputs and weights. sentence_embeddings.py fails syntax compilation because of an unclosed parenthesis and contains placeholders or undefined names. Other scripts use TensorFlow before importing it, call compile/fit on a wrapper that exposes neither method, and expect absent files. The Big Five script declares four classes although the questionnaire uses five, and hard-codes embedding_dim=768 against committed 3,072-column CSVs. It does not implement the promised reverse scoring. Top-K code uses signed rather than stated absolute correlations; the supplement's ordinal formula has the opposite sign from the published layer; and the repository adds one before MAPE although the supplement prints the standard formula, which is undefined at y=0. These discrepancies can change results and block recomputation. The faithful conclusion is that pretrained embeddings contain useful information about item redundancy and covariance in these English questionnaires, and score transfer from semantic neighbors deserves further study. The paper does not show that language creates or predetermines psychological constructs, that PsychoLLM is label-free, that 70% beats simple baselines, that the model generalizes to new people, or that the released artifact reproduces the tables. Stronger claims require causal wording interventions, matrix-dependence-aware inference, respondent holdout, prevalence and baselines, zero-safe metrics, diverse instruments/languages/scales, and a versioned executable release.

Español

Rethinking psychometrics through LLMs estudia si la similitud lingüística entre ítems de cuestionarios se parece a la estructura de correlaciones de respuestas humanas y si esa relación permite predecir respuestas a ítems nuevos. No evalúa personalidad sintética ni rasgos internos de un LLM: usa representaciones textuales preentrenadas para analizar datos psicométricos humanos. En Big Five se parte de 1.015.341 respuestas web y en DASS-42 de 39.775; para la predicción se toman 1.000 observaciones de cada fuente. Se comparan diez representaciones: BERT, RoBERTa, DistilBERT, tres Sentence-BERT, T5 y tres APIs de embeddings de OpenAI. Para cada par de ítems se calcula coseno semántico y se contrasta con la magnitud de su correlación empírica. Con text-embedding-3-large, el ítem más correlacionado aparece entre los tres vecinos semánticos en 95,2 % de los 42 ítems DASS y 82,0 % de los 50 Big Five. En DASS se reportan Pearson r=0,77 y Spearman rho=0,73; en Big Five, r=0,67 y rho=0,64. Es evidencia útil de correspondencia entre redacción y covariación, pero no identifica causalidad: los ítems se diseñan deliberadamente para expresar facetas relacionadas y comparten contenido de constructo. El artículo también fuerza tres factores sobre la matriz semántica DASS, rota con varimax y muestra loadings agrupados por las subescalas teóricas. Sin criterio independiente para elegir tres factores, congruencia, fit, null o validación externa, ilustra alineamiento pero no prueba que el análisis semántico descubra por sí solo la estructura ni que la semántica predetermine el constructo. PsychoLLM es una red ligera: selecciona los tres ítems más similares al target, transforma seis similitudes mediante Dense(10, GELU) y softmax de tres pesos, combina las tres respuestas y aplica una capa ordinal. El suplemento declara learning rate 0,01, batch 32 y tres epochs. En Leave-One-Item-Out, PsychoLLM obtiene MAE/MAPE 0,54/0,28 en DASS, prácticamente empatado con gradient boosting 0,55/0,28 y algo peor que random forest 0,53/0,26. En Big Five obtiene 0,78/0,35, peor que random forest 0,70/0,30 y gradient boosting 0,72/0,30. El LOSO adicional predice cada subescala DASS desde las otras dos con MAE 0,50-0,60. En C19PRC Wave 6, entrenar con PHQ-9 para predecir GAD-7 produce MAE 0,34 y 71 % de exact match; en sentido inverso, MAE 0,36 y 70 %. Esos dos instrumentos miden síntomas relacionados, comparten escala de cuatro categorías y proceden de la misma ola. No se informa el N analizado ni un baseline de clase mayoritaria, moda por ítem, nearest item o balanced accuracy. Por tanto, el 70-71 % headline no demuestra mejora sobre una regla trivial y puede estar impulsado por la abundancia de respuestas cero. Hay una limitación conceptual decisiva: el paper llama a PsychoLLM unsupervised y afirma que no necesita labeled training data, pero el suplemento dice explícitamente que optimiza cross-entropy entre respuestas predichas y reales. Para cada ítem no target usa scores humanos reales como etiquetas; solo permanece sin ver la etiqueta del ítem objetivo. Es self-supervised transfer entre ítems, no aprendizaje sin labels ni sin recogida de respuestas. Además, los baselines supervisados usan cinco folds por participantes, mientras PsychoLLM se entrena con otros ítems de los mismos 1.000 participantes sobre los que evalúa el target; no hay holdout de personas claramente descrito. La comparación no aísla generalización a nuevos respondents y no está emparejada en protocolo. La auditoría del repositorio oficial refuerza la cautela. Existe código y material público valioso en main commit 9c7e501: datos DASS, matrices de embeddings de 3.072 dimensiones, PsychoLLM y varios scripts. Pero no es una reproducción ejecutable. Falta requirements/lockfile, entorno, licencia, instrucciones, raw Big Five, respuestas C19PRC, código PHQ/GAD, baselines, factor analysis, LOSO, outputs y weights. sentence_embeddings.py ni siquiera compila por un paréntesis sin cerrar y contiene placeholders/nombres sin definir. Otros scripts usan TensorFlow antes de importarlo, llaman compile/fit sobre un wrapper que no expone esos métodos y esperan archivos ausentes. El script Big Five declara cuatro clases aunque el cuestionario usa cinco y fija embedding_dim=768 frente a los CSV de 3.072 columnas. Tampoco implementa el reverse scoring prometido. El código Top-K usa correlación con signo, no el valor absoluto declarado; la fórmula ordinal del suplemento tiene signo opuesto a la capa publicada; y el código suma uno antes de MAPE aunque el suplemento imprime la fórmula estándar, indefinida cuando y=0. Estas discrepancias pueden cambiar resultados y bloquean su recomputación. La conclusión fiel es que embeddings preentrenados capturan información útil sobre redundancia y covariación de ítems en estos cuestionarios ingleses, y que transferir scores de vecinos semánticos merece estudio. No se demuestra que el lenguaje cree o predetermine constructos psicológicos, que PsychoLLM sea label-free, que el 70 % supere baselines simples, que generalice a nuevas personas o que el artefacto publicado reproduzca las tablas. Para afirmaciones más fuertes hacen falta intervención causal sobre wording, inferencia consciente de dependencia matricial, holdout de participantes, baselines y prevalencias, métricas válidas con ceros, instrumentos/idiomas/escalas diversos y un release ejecutable versionado.

Research question

To what extent does the semantic similarity between questionnaire items reproduce their empirical correlation, and can an architecture that combines embeddings and responses from neighboring items predict scores of unseen items or questionnaires?

Method

Cosine matrices from ten representation models are compared with absolute correlation matrices of Big Five-50 and DASS-42 using Top-K, Pearson and Spearman. A three-factor DASS semantic solution is illustrated. PsychoLLM receives six similarities between target and three neighbors, learns three weights with Dense(10)-GELU/softmax and applies an ordinal layer to the weighted responses. Leave-One-Item-Out is evaluated on 1,000 cases per questionnaire, LOSO on DASS, and PHQ-9/GAD-7 transfer on C19PRC Wave 6. The audit read and rendered the 12 pages of the article and the 10 of the supplement, verified metadata, formulas, tables, protocol and claims, and inventoried, compiled and contrasted the official repository at the published commit of main.

Sample: For matrices, 1,015,341 Big Five responses and 39,775 DASS responses from voluntary web participation are declared. For Leave-One-Item-Out, 1,000 cases are randomly sampled from each dataset. For PHQ-9/GAD-7, C19PRC Wave 6 is used, but the analyzed N is not reported. No independent participant holdout is published for PsychoLLM, nor demographic, representativeness, or response quality analysis.

Findings

  • text-embedding-3-large achieves Top-3 95.2% on DASS and 82.0% on Big Five.
  • On DASS it obtains Pearson r=0.77 and Spearman rho=0.73; on Big Five r=0.67 and rho=0.64.
  • The set of models shows a positive semantic-response association, with considerable variation across representations.
  • PsychoLLM obtains MAE/MAPE 0.54/0.28 on DASS, close to the ensembles.
  • On Big Five PsychoLLM 0.78/0.35 falls behind random forest 0.70/0.30 and gradient boosting 0.72/0.30.
  • DASS LOSO produces MAE 0.50 for Depression, 0.60 for Anxiety and 0.57 for Stress.
  • PHQ-9 to GAD-7 reports 71% exact match and MAE 0.34; GAD-7 to PHQ-9, 70% and 0.36.
  • The 70-71% lacks a prevalence baseline, per-item mode, or balanced accuracy.
  • PsychoLLM uses real responses as labels for non-target items; it is not label-free.
  • The evaluation does not document a holdout of new participants for PsychoLLM.
  • The observational association does not demonstrate that semantics cause or predetermine the construct.
  • The DASS solution fixes three known factors and does not quantify independent recovery.
  • The official repository contains DASS, embeddings and partial code, but does not reproduce the tables.
  • The published code has syntax/runtime errors and material contradictions with article and supplement.
  • The work is psychometry of human responses assisted by embeddings, not evidence of personality in an LLM.

Limitations

  • Only two instruments are evaluated in the matrix phase and related scales in the transfer.
  • All textual material is in English.
  • The OpenPsychometrics samples are web-based, opt-in, and representativeness is not demonstrated.
  • Demographics are excluded and there is no subgroup, culture, or response quality analysis.
  • The semantic-response correlation shares construct content and does not identify causality.
  • There is no intervention with paraphrases, random wording, controlled negations, or matched items.
  • Matrix cells are not independent and standard p-values do not correct for this dependence.
  • Multiplicity is not reported despite multiple models, metrics, and instruments.
  • The DASS factor analysis fixes three factors a priori and groups by theoretical subscale.
  • Parallel analysis, fit, congruence, null, and factorial cross-validation are missing.
  • PsychoLLM is described as unsupervised/label-free even though it optimizes with real human scores.
  • There is no clear participant holdout for PsychoLLM.
  • The baselines use five folds and a comparable protocol is not demonstrated.
  • The PHQ/GAD exact match does not include per-class distribution or a majority baseline.
  • PHQ-9 and GAD-7 share four categories and related clinical constructs.
  • The analyzed C19PRC N is not reported.
  • MAPE with zero scores requires a convention; code and supplement do not match.
  • The ordinal thresholds of the code are not ordered and the direction of the formula differs.
  • There is no preregistration, power, intervals, or independent external validation.
  • The repository has no requirements, lockfile, environment, license, or instructions.
  • Raw Big Five, C19PRC, baselines, factor analysis, LOSO, outputs, and weights are missing.
  • sentence_embeddings.py does not compile and contains placeholders/undefined names.
  • The Big Five pipeline uses four classes and dimension 768 versus five classes and 3,072 embeddings.
  • The reverse scoring declared for Big Five does not appear in the published code.
  • Top-K, ordinal sign, and MAPE show paper-code discrepancies.
  • There is no immutable tag/release; the last audited main commit is prior to publication.
  • A mixture of encoders, sentence transformers, T5, and embedding APIs is called an LLM.

What the study does not establish

  • It does not demonstrate that semantics create or predetermine psychological constructs.
  • It does not demonstrate that factor analysis primarily recovers linguistic artifacts.
  • It does not demonstrate learning without labels or without human responses.
  • It does not demonstrate that 70-71% exceeds a majority class rule or mode.
  • It does not demonstrate an advantage over baselines under the same protocol.
  • It does not demonstrate generalization to new participants.
  • It does not demonstrate transfer to heterogeneous scales, constructs, languages, or cultures.
  • It does not demonstrate that the MAPE values follow the printed formula.
  • It does not demonstrate that the published code produces the tables.
  • It does not demonstrate that the Big Five reverse scoring was executed as stated.
  • It does not demonstrate that three DASS factors were discovered without prior knowledge.
  • It does not attribute the results specifically to generative LLM reasoning.
  • It does not demonstrate personality, mental state, or internal trait in any LLM.
  • It does not generalize the opt-in web sample to the general population.

Traceability

Scope: Full text

Version: Scientific Reports version of record, volume 15 article 37313, published 24 October 2025; audited with the official 10-page supplement and official GitHub main commit 9c7e50136b67154205eba94d12c9abfe9c9fad75

Consulted source: https://doi.org/10.1038/s41598-025-21289-8

Review: Codex complete bilingual fidelity pass using the full 12-page Scientific Reports version of record and 10-page official supplement, all-page visual inspection, official metadata and DOI verification, full method/table/formula/claim audit, and inventory, syntax check and paper-code consistency review of the official GitHub repository at immutable main commit 9c7e50136b67154205eba94d12c9abfe9c9fad75; 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

  • all-MiniLM-L6-v2
  • all-MiniLM-L12-v2
  • all-mpnet-base-v2
  • BERT
  • RoBERTa
  • DistilBERT
  • text-embedding-3-small
  • text-embedding-3-large
  • text-embedding-ada-002
  • T5
  • PsychoLLM custom dense-plus-ordinal architecture
  • Feed-forward neural network baseline
  • Random forest baseline
  • Gradient boosting baseline

Instruments and metrics

  • 50-item Big Five questionnaire from OpenPsychometrics
  • DASS-42
  • GAD-7
  • PHQ-9
  • Cosine semantic similarity
  • Top-1, Top-2 and Top-3 retrieval
  • Pearson and Spearman matrix-entry correlations
  • MAE
  • MAPE
  • Exact response match
  • Leave-One-Item-Out
  • Leave-One-Subscale-Out

Data used

  • OpenPsychometrics Big Five raw dataset: 1,015,341 respondents reported; not included in the official repository
  • OpenPsychometrics DASS raw dataset: 39,775 respondents; included in the official repository
  • C19PRC Wave 6 PHQ-9 and GAD-7 responses from OSF; analyzed N not reported and response data not included in GitHub
  • Committed 3,072-dimensional Big Five and DASS text-embedding-3-large matrices
  • Official repository at main commit 9c7e50136b67154205eba94d12c9abfe9c9fad75

Evidence and location

  • Metadata, abstract, datasets and semantic models: Scientific Reports version of record pages 1-4; DOI 10.1038/s41598-025-21289-8
  • Top-K and correlations per model: Version of record pages 4-5, Figure 2 and Table 1
  • PsychoLLM architecture and new item evaluation: Version of record pages 6-7, Figure 3 and Table 2; Supplementary pages 5-7
  • DASS LOSO: Version of record pages 7-8, Figures 4-5
  • PHQ-9/GAD-7 transfer and 70-71%: Version of record pages 8-9, Table 3; Supplementary page 10
  • Limitations and causal/epistemological scope: Version of record pages 9-10, Discussion, Implications and Conclusions
  • Data, code and official dates: Version of record pages 11-12 and official Nature article metadata checked 16 July 2026
  • Factor analysis, MAPE, training labels, hyperparameters and negation: Official Supplementary Information pages 3-7
  • Code audit and artifact completeness: Official GitHub repository main commit 9c7e50136b67154205eba94d12c9abfe9c9fad75; file inventory, Python compile check and paper-code comparison on 16 July 2026
  • Comprehensive validity and reproducibility report: reports/verification/article-207-questionnaire-semantics-psychollm-validity-and-reproducibility-audit.json
  • Complete visual inspection: All 12 version-of-record pages and all 10 supplementary pages rendered and visually inspected on 16 July 2026