Beyond Demographics: Enhancing Cultural Value Survey Simulation with Multi-Stage Personality-Driven Cognitive Reasoning

Society, culture, and collective behavior2025ACL AnthologyApproved editorial review

Authors: Haijiang Liu, Qiyuan Li, Chao Gao, Yong Cao, Xiangyu Xu, Xun Wu, Daniel Hershcovich, Jinguang Gu

Keywords: Large Language Models, Survey Simulation, Cultural Values, World Values Survey, MBTI Type Dynamics, Synthetic Respondents, Demographic Stereotyping, Reproducibility

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

8
Authors
13
Findings
24
Limitations
12
Evidence

Editorial summary

English

This EMNLP 2025 paper proposes MARK, a multi-call prompting pipeline for predicting individual World Values Survey answers and aggregate response distributions. It does not measure participant personality or stress. An LLM first assigns each demographic attribute a 0-100 stress score and summarizes the profile; it then infers dominant and auxiliary MBTI cognitive functions, completes tertiary and inferior functions by rule, generates one answer and weight under each function, and uses another call to review, reweight, and synthesize the final option. Stress, personality, cognitive trajectory, and explanation are therefore model-generated text rather than psychological observations. The paper uses WVS Wave 7 U.S. and China case studies, claims thirty demographic features, twenty clusters, 400 subjects and 102,800 question-answer pairs in the U.S., and twenty subjects/4,340 pairs in China. It compares GLM-4-air, GPT-4o-2024-08-06, and Doubao-1.5-pro at temperature .9 without published repetitions or uncertainty. On the U.S. sampled distribution, MARK reports 33.69%, 38.11%, and 46.98% accuracy. Against each model's strongest accuracy baseline, the actual gains are 6.71, 1.15, and 15.53 points, not a uniform 10%. Section 5.2 also conflicts with Table 2: against Demo.+Ideo.+Opiniontop3 the table yields +8.63, +5.04, and +15.53 points, not +4.3, +8.9, and +15.5. In China, MARK is not the most accurate GLM method: 31.58% versus 35.21% with Chinese prompts and 31.37% versus 35.60% with English prompts, despite the table caption's broad superiority claim. Artifact audit further changes the methodological interpretation. The extraction notebook does not cluster the thirty described demographics; it uses only SACSECVAL and RESEMAVAL, WVS secular-value and emancipative-value indices derived from survey responses. Group and representative selection are thus informed by the target domain. The same notebook hard-codes five clusters rather than twenty, selects at most twenty people within each, has no seed, calls drop_duplicates without keeping the result, and labels an SSE curve as silhouette analysis. Oracle personality is also not measured personality: personal_assign.py iterates Q1-Q258, uses each subject's actual answers, including target items, and accumulates LLM-generated personality scores. Table 17 compares a demographic predictor against this response-derived synthetic label; 70% role accuracy and 15/31% type accuracy do not validate MBTI. The robustness analysis in fact shows nearly overlapping curves for random, predicted, and oracle assignments, and the paper says personality similarity has minimal influence. This weakens causal attribution to personality. The released metrics also lack valid implementations. The notebook computes one minus SciPy Jensen-Shannon distance, not one minus the divergence it names. For EMD it passes probability-mass vectors as sample locations to wasserstein_distance; reversed binary distributions [.7,.3] and [.3,.7] therefore receive zero instead of .4 on ordered supports. The kappa script passes per-option count vectors to cohen_kappa_score as if they were two raters' labels, while the paper alternates Cohen and Fleiss. Isolated p-values are printed, including .6, but no test, sampling unit, code, predictions, repetitions, intervals, bootstrap, or multiplicity correction is released. The repository omits WVS data, samples, outputs, results, and requirements.txt; all three notebooks have zero executed cells and zero outputs. Paper, README, and code respectively name Gemini-2-pro, Gemini-2.5, and gemini-1.5-flash-exp-0827; augmentation code fails at print(a), appends each result twice in memory, and the 21 released JSON files contain 1,254 empty probability objects. The main runner also exposes a non-placeholder API credential that must be treated as compromised. The defensible conclusion is narrow: more structured calls and context can change and sometimes improve synthetic-survey predictions, especially with Doubao. The study does not establish personality, stress, cognition, faithful interpretability, significance, unseen-demographic generalization, or replacement of human surveys. A credible reference requires response-independent sampling, respondent-level metric recalculation, measured constructs or construct-neutral labels, held-out people/questions/groups, compute-matched ablations, repeated runs, stereotype auditing, and complete immutable artifacts.

Español

Este artículo de EMNLP 2025 propone MARK, un pipeline de prompting de varias llamadas para predecir respuestas individuales y distribuciones agregadas del World Values Survey. No mide personalidad ni estrés en los participantes. Primero, un LLM asigna a cada atributo demográfico un nivel de estrés de 0 a 100 y resume el perfil; después infiere funciones cognitivas MBTI dominante y auxiliar, completa las funciones terciaria e inferior mediante reglas; genera una respuesta y un peso bajo cada función; y otro paso revisa, repondera y sintetiza la opción final. Por tanto, estrés, personalidad, trayectoria cognitiva y explicación son texto generado por el modelo, no observaciones psicológicas. El paper usa WVS Wave 7 de Estados Unidos y China, declara 30 rasgos demográficos, 20 clusters, 400 sujetos y 102.800 pares pregunta-respuesta en EE. UU., y 20 sujetos/4.340 pares en China. Compara GLM-4-air, GPT-4o-2024-08-06 y Doubao-1.5-pro a temperatura 0,9, sin repeticiones ni incertidumbre publicada. En la distribución muestreada de EE. UU., MARK obtiene 33,69 %, 38,11 % y 46,98 % de accuracy. Frente al baseline de mayor accuracy de cada modelo, las ganancias reales son 6,71, 1,15 y 15,53 puntos, no un 10 % uniforme. La prosa de Section 5.2 tampoco coincide con Table 2: contra Demo.+Ideo.+Opiniontop3, la tabla da +8,63, +5,04 y +15,53 puntos, no +4,3, +8,9 y +15,5. En China, MARK no es el método más preciso con GLM: 31,58 % frente a 35,21 % con prompts chinos y 31,37 % frente a 35,60 % con prompts ingleses, aunque el caption afirma superioridad general. La auditoría del artefacto cambia además la interpretación metodológica. El notebook de extracción no agrupa por los 30 datos demográficos descritos: usa únicamente SACSECVAL y RESEMAVAL, índices WVS de valores seculares y emancipativos derivados de las respuestas. La selección de grupos y representantes está así informada por el dominio objetivo. El mismo notebook fija 5 clusters, no 20, selecciona como máximo 20 personas en cada uno, no define seed, llama a drop_duplicates sin conservar el resultado y etiqueta una curva SSE como silhouette analysis. La llamada personalidad oracle tampoco es personalidad medida: personal_assign.py recorre Q1-Q258, usa las respuestas reales del sujeto, incluida cada pregunta objetivo, y acumula probabilidades de personalidad generadas por otro LLM. Table 17 compara el predictor con esa etiqueta sintética filtrada por respuestas; 70 % de acierto a nivel de role y 15/31 % a nivel de type no valida MBTI. De hecho, el análisis de robustez muestra que asignaciones random, predicted y oracle producen curvas casi solapadas, y el propio texto dice que la similitud de personalidad tiene influencia mínima. Esto debilita la atribución causal a la personalidad. Las métricas publicadas tampoco cuentan con una implementación válida. El notebook calcula 1 menos la distancia Jensen-Shannon de SciPy, no 1 menos la divergencia que nombra. Para EMD pasa los vectores de masas como posiciones muestrales a wasserstein_distance; dos distribuciones binarias invertidas [0,7, 0,3] y [0,3, 0,7] reciben así distancia cero en vez de 0,4 sobre soportes ordenados. El script de kappa entrega recuentos por opción a cohen_kappa_score como si fueran etiquetas de dos jueces, y el paper alterna Cohen y Fleiss. Se muestran p-values aislados, incluido 0,6, pero no se publica test, unidad muestral, código, predicciones, repeticiones, intervalos, bootstrap ni corrección múltiple. El repositorio omite WVS, muestras, outputs, resultados y requirements.txt; los tres notebooks tienen cero celdas ejecutadas y cero outputs. Paper, README y código nombran respectivamente Gemini-2-pro, Gemini-2.5 y gemini-1.5-flash-exp-0827; el script de augmentación falla en print(a), duplica append en memoria y los 21 JSON publicados contienen 1.254 objetos de probabilidad vacíos. El runner central expone además una credencial API no placeholder, que debe considerarse comprometida. La conclusión defendible es limitada: una estructura de deliberación con más llamadas y contexto puede cambiar y a veces mejorar predicciones sintéticas de encuestas, especialmente con Doubao. No se demuestra personalidad, estrés, cognición, interpretabilidad fiel, significación, generalización a demografías nuevas ni sustitución de encuestas humanas. Antes de usar MARK como evidencia psicológica o social se necesitan muestreo independiente de respuestas, métricas recalculadas a nivel de participante, constructos medidos o labels neutrales, tests held-out de personas/preguntas/grupos, ablations igualadas en cómputo, repeticiones, auditoría de estereotipos y artefactos completos.

Research question

Can a multistage pipeline inspired by MBTI type dynamics, which infers stress and cognitive functions from demographics, improve the zero-shot prediction of responses and distributions of the World Values Survey compared to demographic and opinion prompts?

Method

MARK transforms WVS attributes into stress scores and a sociodemographic profile generated by an LLM, infers the dominant/auxiliary MBTI functions and completes the other two, decides for each question whether stress activates its positive or negative version, generates four weighted responses and synthesizes a final option. The paper declares K-means on demographics and evaluates exact match, 1-JSD, EMD and kappa on sampled/global distributions of the U.S. and sampled distributions of China. The audit contrasts the 23 published pages with the official commit, reconstructs the real sampling, profiles 25,473 augmentation rows and reviews formulas and metric code.

Sample: The paper declares 400 U.S. representatives, 20 per each of 20 clusters, and 102,800 responses, plus 20 Chinese centroids and 4,340 responses. It does not publish the workbooks used, IDs, cluster composition or outputs. The released notebook contradicts that design: it clusters with two value indices derived from responses, fixes 5 clusters and selects up to 20 representatives per cluster, so it does not reconstruct the 400 cases.

Findings

  • The authoritative source is EMNLP 2025, DOI 10.18653/v1/2025.emnlp-main.928, pages 18406-18428; the 23 published pages and the 23 arXiv v1 pages were visually inspected.
  • In sampled U.S., MARK obtains ACC 33.69/38.11/46.98% for GLM/GPT-4o/Doubao; the gains over the best baseline of each model are 6.71/1.15/15.53 points.
  • The numbers 4.3/8.9/15.5 in Section 5.2 are not derived from Table 2; against the baseline it names the difference is 8.63/5.04/15.53 points.
  • With GLM in China, MARK falls below Demo.+Ideo.+Opiniontop3 with both Chinese prompts, 31.58 versus 35.21, and English prompts, 31.37 versus 35.60.
  • Random, predicted and oracle personality show very similar performance; the accuracy of declared personality is 70% role but only 15%/31% type.
  • Oracle personality is derived from real Q1-Q258 of the same participant and is not an independent MBTI measurement.
  • The sampling notebook uses SACSECVAL/RESEMAVAL, two value indices derived from responses, not the 30 demographic variables described.
  • The notebook fixes n_clusters=5 although paper/Figure 6 declare 20 and does not implement silhouette_score.
  • The 1-JSD implementation uses 1 minus JS distance; EMD uses masses as positions; kappa uses count vectors as labels.
  • There is no significance code or predictions; at least one p-value 0.6 is shown.
  • The 21 augmentation JSON files total 25,473 rows and 1,254 empty probability objects; role has 33.39% empty.
  • The three notebooks have zero executed cells and zero outputs; WVS data, samples, results, requirements and license are missing.
  • The repository exposes a non-placeholder API credential and it should be revoked.

Limitations

  • There are no independent ground truths for stress, MBTI, cognitive functions or explanations.
  • Psychological profiles are fabricated from demographics and can turn stereotypes into plausible reasoning.
  • Sample selection uses value indices derived from the target responses.
  • Oracle personality uses all real responses, including the evaluated question.
  • Table 17 validates against synthetic response-derived labels, not against an MBTI instrument.
  • The released sampling does not reproduce 20 clusters x 20 people and lacks seeds or stability.
  • There are no held-out splits of participants, questions, cultures or demographic groups.
  • The global evaluation compares with the histogram of the same population and does not demonstrate unseen demographics.
  • JSD is misnamed and the EMD/kappa implementations are conceptually invalid.
  • Cohen and Fleiss kappa are used inconsistently.
  • No tests, unit of analysis, multiple runs, CIs, bootstrap or multiple correction are published.
  • Temperature 0.9 without repetitions prevents estimating generation variance.
  • The ablation of one function versus four confounds mechanism with additional calls, tokens and context.
  • The synthesis example uses weights that sum to 1.6 and neither prompts nor code require normalization.
  • The improvements are not consistent across models, metrics, languages and baselines.
  • GPT-4o is only run on centroids due to budget, reducing comparability.
  • The 21 augmentation files contain only 242 unique questions while the assigner iterates over 258.
  • Paper, README and code disagree on the Gemini model used.
  • The augmentation code contains a NameError and double append; dependencies and outputs are missing.
  • There is no license, model/data cards, API traces, complete configuration or privacy statement for derivatives.
  • Only the U.S. and China are studied; there is no cultural invariance or translation equivalence.
  • The analogy with PTSD is applied without trauma or mental health data.
  • Errors or stereotypes by sex, age, ethnicity, citizenship, religion, migration, occupation, income or marital status are not audited.
  • There is no prospective validation that would allow replacing representative human surveys.

What the study does not establish

  • It does not demonstrate that participants have the inferred stress, MBTI or cognitive functions.
  • It does not demonstrate that LLMs simulate human mental processes or that their explanations are faithful.
  • It does not demonstrate that correctly assigning personality is the cause of the improvements.
  • It does not establish a uniform 10% improvement over the best baseline.
  • It does not establish valid EMD or kappa with the released code.
  • It does not test statistical significance of the results.
  • It does not demonstrate generalization to new demographics, new participants, new questions or new countries.
  • It does not demonstrate safety or absence of stereotypes in demographic profiling.
  • It does not justify using synthetic persons as substitutes for human population or social surveys.

Traceability

Scope: Full text

Version: Proceedings of EMNLP 2025, Anthology ID 2025.emnlp-main.928, DOI 10.18653/v1/2025.emnlp-main.928, pages 18406-18428, 23 pages

Consulted source: https://aclanthology.org/2025.emnlp-main.928/

Review: Codex complete bilingual full-text fidelity pass using the published EMNLP version and arXiv v1, all-page visual inspection of both PDFs, official repository commit audit, response-derived sampling analysis, oracle target-leakage tracing, exact Table 2 and Table 3 arithmetic reconciliation, metric-implementation review, personality augmentation profiling, construct-validity assessment, demographic-stereotyping review, credential exposure check, and end-to-end reproducibility inventory; summaries written from paper and artifact evidence rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GLM-4-air
  • openai/gpt-4o-2024-08-06
  • Doubao-1.5-pro
  • Llama-3.1-7B and Llama-3.1-13B exploratory runs
  • Personality augmentation described inconsistently as Gemini-2-pro in paper, Gemini-2.5 in README, and gemini-1.5-flash-exp-0827 in code
  • bge-m3-v3 embedding model for the opinion baseline

Instruments and metrics

  • World Values Survey Wave 7 multiple-choice responses
  • LLM-generated 0-100 demographic stress scores
  • MBTI Type Dynamics cognitive-function prompts
  • Response-derived synthetic oracle personality assignment
  • Exact-match accuracy
  • Released 1 minus Jensen-Shannon distance implementation, mislabeled as 1-JSD
  • Released invalid Wasserstein/EMD implementation
  • Released invalid count-vector Cohen kappa implementation; paper also calls it Fleiss kappa
  • Independent construct, leakage, arithmetic, metric, stereotyping, and reproducibility audit

Data used

  • World Values Survey Wave 7, United States case study
  • World Values Survey Wave 7, China case study
  • Reported U.S. sample: 20 clusters x 20 subjects, 400 subjects and 102,800 QA pairs
  • Reported China sample: 20 cluster-centroid subjects and 4,340 QA pairs
  • Official artifact: 21 augmented personality JSON files, 25,473 rows, 242 unique question texts, 1,254 empty probability objects
  • Raw WVS, sampled workbooks, predictions, results, and execution outputs are not released

Evidence and location

  • Abstract, scope and MARK pipeline: Published pages 18406-18409, Abstract and Sections 1-3
  • Data, settings, declared models and metrics: Published pages 18409-18411, Sections 4.1-4.4 and Table 2
  • Results, arithmetic mismatch, robustness and China: Published pages 18412-18414, Sections 5.1-5.5 and Tables 2-4
  • MBTI limitations, ethics and recognized stereotypes: Published pages 18414-18415, Limitations and Ethics Statement
  • Prompts, declared clustering, oracle and complete tables: Published pages 18418-18428, Appendices B-E and Tables 5-17
  • Real sampling with value indices, n_clusters=5 and absence of seed/silhouette: Official repository commit 394e92b, Data Extract.ipynb audited 15 July 2026
  • Target leakage of oracle personality: Official repository commit 394e92b, utils/personal_assign.py audited 15 July 2026
  • 1-JSD, EMD and kappa errors: Official repository commit 394e92b, Result Evaluation.ipynb and utils/cohen_kappa.py audited 15 July 2026
  • Incomplete augmentation, Gemini identity and code failures: Official repository commit 394e92b, 21 data/augmented JSON files, README and utils/personality_augmentation.py audited 15 July 2026
  • Reproducibility, missing files and exposed credential: Official repository tree at commit 394e92b audited 15 July 2026; credential value omitted
  • Comprehensive audit of construct, metrics and artifact: reports/verification/article-193-mark-construct-and-artifact-audit.json
  • Complete visual inspection: All 23 published EMNLP pages and all 23 arXiv v1 pages rendered and visually inspected on 15 July 2026