Synthetic Personality Review

Parametric Social Identity Injection and Diversification in Public Opinion Simulation

Society, culture, and collective behavior2026arXivApproved editorial review

Authors: Hexi Wang, Yujia Zhou, Bangde Du, Qingyao Ai, Yiqun Liu

Keywords: Persona conditioning, Human simulation

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

5
Authors
8
Findings
13
Limitations
6
Evidence

Editorial summary

English

The paper introduces Parametric Social Identity Injection (PSII), a technique for simulating public-opinion responses by intervening directly in an LLM's hidden states. The system combines a demographic profile in the prompt, demographic vectors built from GPT-4o-generated questions and persona instructions, embeddings associated with five languages, called value vectors, Gaussian noise, and an attribute-specific layer assignment. It is evaluated on a random sample of 100 World Values Survey Wave 7 respondents: Q1-Q259 are targets and Q260-Q290 provide identity information. Four open models answer each item independently, and their aggregate response distributions are compared with human distributions using KL divergence and Entropy Deviation (ED).

In Table 1, PSII obtains the best overall KL and ED among the compared methods for all four models: 0.4843/0.0319 for Qwen2.5-7B, 0.5814/0.2123 for Qwen2.5-14B, 0.4017/0.0040 for Llama-3.1-8B, and 0.5607/0.0774 for Mistral-24B. This supports better matching of response marginals and total entropy in this particular sample. PSII does not win every model-category cell; for example, SimVBG has lower Beliefs & Life KL on Qwen2.5-14B and Mistral-24B. Ablations attribute the largest loss to removing demographic vectors and show that noise is especially important for matching entropy.

The full-text and artifact audit narrows the central interpretation. The metrics neither compare demographic groups nor calculate within-group diversity: they aggregate all 100 people for each question. The evidence therefore supports aggregate marginal matching, not accurate recovery of between-group differences, minorities, or intersections. Hidden-state spread is illustrated with KPCA and a kNN radius for one question and is not validated against a human notion of diversity. Layer assignments are also selected by minimizing KL on the same WVS task used for final evaluation, without a separate selection set.

The code exposes a methodological mismatch: prompt language and language embedding are chosen independently at random from English, Chinese, Spanish, Arabic, and Russian, so they match only by chance and are not based on the respondent's actual language. The so-called value vectors are trained on CulturaX as language embeddings rather than against WVS values. The released artifact also does not document the analytical treatment of 1,057 negative nonresponse codes or Q223, where 70 valid human responses use country-specific party codes that the prompt cannot emit. No intervals, tests, or repeated inference runs support the paper's use of 'significantly.'

The study provides a substantial repository, a v1.0.0 tag, and a Zenodo archive with profiles, translations, and 400 vectors, but it releases neither outputs nor code for KL, ED, JS, MAE, tables, figures, SimVBG, or Persona Vectors; the reported results therefore cannot be regenerated end to end. Public CSVs contain 615 columns of WVS microdata, coordinates, identifiers, and sensitive answers. Current official WVS conditions prohibit redistribution of the data files, making their public inclusion an unresolved licensing and privacy risk. PSII is a promising technical contribution for studying representation steering, but it does not validate faithful social twins or safe population simulation beyond this controlled experiment.

Español

El artículo presenta Parametric Social Identity Injection (PSII), una técnica para simular respuestas de opinión pública mediante la intervención directa de estados ocultos de un LLM. El sistema combina un perfil demográfico en el prompt, vectores demográficos construidos a partir de preguntas e instrucciones generadas con GPT-4o, embeddings asociados a cinco idiomas, denominados «value vectors», ruido gaussiano y una asignación de atributos a capas específicas. Se evalúa con una muestra aleatoria de 100 personas de World Values Survey Wave 7: Q1-Q259 sirven como preguntas objetivo y Q260-Q290 como información de identidad. Cuatro modelos abiertos responden cada pregunta por separado y sus distribuciones agregadas se comparan con las humanas mediante divergencia KL y desviación de entropía (ED).

En la Tabla 1, PSII obtiene el mejor KL y ED global de los métodos comparados en los cuatro modelos: 0,4843/0,0319 para Qwen2.5-7B, 0,5814/0,2123 para Qwen2.5-14B, 0,4017/0,0040 para Llama-3.1-8B y 0,5607/0,0774 para Mistral-24B. Esto es evidencia de un mejor ajuste de los márgenes de respuesta y de la entropía total en esa muestra concreta. No gana cada combinación de categoría y modelo: SimVBG, por ejemplo, logra menor KL en «Beliefs & Life» con Qwen2.5-14B y Mistral-24B. Las ablaciones atribuyen la mayor parte de la mejora a los vectores demográficos y muestran que el ruido contribuye especialmente a acercar la entropía.

La lectura completa y la auditoría del artefacto acotan la interpretación central. Las métricas no comparan grupos demográficos ni calculan diversidad dentro de cada grupo: agregan a las 100 personas por pregunta. Por ello, demuestran ajuste marginal agregado, no que PSII recupere correctamente diferencias entre grupos, minorías o intersecciones. La dispersión de estados ocultos se muestra con KPCA y radio kNN para una sola pregunta y no se valida contra una noción humana de diversidad. Además, las capas se eligen minimizando KL en la misma tarea WVS usada para evaluar, sin conjunto de selección separado.

El código revela una desalineación metodológica: el idioma del prompt y el embedding de idioma se eligen de forma aleatoria e independiente entre inglés, chino, español, árabe y ruso, de modo que coinciden solo por azar y no se basan en el idioma real de la persona. Los llamados value vectors se entrenan en CulturaX como embeddings lingüísticos, no contra valores WVS. Tampoco se publica el tratamiento analítico de 1.057 códigos negativos de no respuesta ni de Q223, donde 70 respuestas humanas válidas usan códigos de partido que el prompt no puede emitir. No hay intervalos, pruebas ni repeticiones de inferencia que respalden la palabra «significativamente».

La publicación ofrece un repositorio considerable, una etiqueta v1.0.0 y un archivo Zenodo con perfiles, traducciones y 400 vectores, pero no incluye outputs ni código para KL, ED, JS, MAE, tablas, figuras o los baselines SimVBG y Persona Vectors; por tanto, los resultados no se pueden regenerar de extremo a extremo. Los CSV públicos contienen 615 columnas de microdatos WVS, coordenadas, identificadores y respuestas sensibles. Las condiciones oficiales actuales de WVS prohíben redistribuir los archivos de datos, por lo que su inclusión pública plantea un riesgo de licencia y privacidad no resuelto. El trabajo es una contribución técnica prometedora para estudiar steering representacional, pero no valida gemelos sociales fieles ni simulación poblacional segura fuera de este experimento controlado.

Research question

Can the parametric injection of profiles, demographic vectors, linguistic embeddings, and noise into internal layers of an LLM bring the distributions of synthetic responses closer to the World Values Survey and avoid the homogenization attributed to the «Diversity Collapse»?

Method

The authors sample 100 of 97,220 responses from WVS Wave 7, convert Q260-Q290 into profiles, and use Q1-Q259 as targets. Each agent answers each item separately with Qwen2.5-7B/14B, Llama-3.1-8B, or Mistral-24B. PSII adds a structured profile to the prompt, injects demographic vectors obtained by subtracting means of hidden states for synthetic instructions into selected layers, incorporates an embedding trained per language on 20,000 examples from CulturaX over three epochs, and applies Gaussian noise calibrated per model. It compares Direct, High-Temp, Multilingual, Diversity Request, Prompt Engineering, SimVBG, and Persona Vectors using KL and normalized entropy absolute difference; JS and MAE appear in the appendix. The layers of Qwen2.5-7B are selected by sweeping 1-28 and choosing the one that minimizes KL on the task itself.

Sample: Main experiment with 100 individuals randomly sampled from 97,220 WVS records, coming from 48 countries in the published CSV; 259 target questions per individual. The representation visualizations use 500 agents. The sampling robustness repeats only PSII with Qwen2.5-7B in five groups of 100.

Findings

  • PSII obtains the best global KL and ED of Table 1 for the four models, although it does not dominate all categories.
  • The global PSII results are KL/ED 0.4843/0.0319, 0.5814/0.2123, 0.4017/0.0040, and 0.5607/0.0774 for Qwen2.5-7B, Qwen2.5-14B, Llama-3.1-8B, and Mistral-24B, respectively.
  • The ablation of demographic vectors produces the greatest overall degradation; removing noise markedly worsens the entropy fit.
  • Five samples of 100 with PSII and Qwen2.5-7B yield mean KL 0.4732 and mean ED 0.0290, but do not repeat baselines or paired contrasts.
  • The main metrics measure aggregate marginal distributions and total entropy, not intergroup or intragroup diversity.
  • The prompt language and the linguistic embedding are assigned independently in the code, creating an expected match of only approximately 20%.
  • The repository provides 400 vectors, samples, and a Zenodo file, but lacks outputs and the analytical code necessary to regenerate tables and figures.
  • The publication of WVS microdata apparently contradicts the current official non-redistribution clause and adds privacy and linkage risk.

Limitations

  • Only 100 individuals support the main comparison; the additional robustness evaluates PSII on a single model and not the difference against baselines.
  • There are no group-conditioned metrics, worst group, minorities, intersections, joint coherence, or individual stability.
  • Layers are selected with the same human objective and the same task that is later used for evaluation, without an independent validation split.
  • Noise calibration also has no script, outputs, or published selection set.
  • The value vectors are language embeddings; their interpretation as value orientations has no construct validation.
  • The uniform and random assignment of five languages ignores the actual language and linguistic distribution of the respondent.
  • Of 25,900 target human cells, 1,057 are negative WVS codes; no exclusion, denominator, or smoothing is documented.
  • Q223 has 70 valid national party codes that cannot be generated from the four generic options published.
  • The parser accepts the first number, truncates decimals, and does not validate the range; invalid outputs are not reported.
  • No intervals, tests, complete seeds, or repeated inferences are offered for the differences between methods.
  • SimVBG and Persona Vectors appear in the tables, but not in run.sh or in published outputs or analysis scripts.
  • The environment does not fix exact model revisions and requires paths, APIs, full WVS, CulturaX, and GPU A100.
  • The public CSVs include coordinates, identifiers, and political, religious, and ethnic responses without privacy assessment or documented redistribution permission.

What the study does not establish

  • It does not demonstrate faithful diversity between groups, within groups, minorities, or intersectional identities.
  • It does not demonstrate that entropy deviation or hidden state dispersion equate to human social diversity.
  • It does not demonstrate that language embeddings represent individual value orientations.
  • It does not demonstrate individual prediction, digital twin coherence, or longitudinal stability.
  • It does not demonstrate generalization to other questionnaires, populations, languages, countries, or models.
  • It does not support statistical inference of significance between methods.
  • It does not allow end-to-end reproduction of tables and figures with the published artifact.
  • It does not demonstrate that redistribution of WVS microdata complies with its license or is safe against linkage and reidentification.
  • It does not validate real use for decisions, campaigns, public policies, or substitute representation of populations.

Traceability

Scope: Full text

Version: arXiv:2603.16142v2; KDD 2026 accepted version; repository HEAD 5bae420 and archival tag v1.0.0 at f437bf5

Consulted source: https://arxiv.org/abs/2603.16142

Review: Codex 18-page visual full-text, representation-steering, aggregate-diversity, language-vector alignment, layer-selection, WVS data-quality, statistics, code, privacy, license and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Qwen2.5-7B-Instruct
  • Qwen2.5-14B-Instruct
  • Llama-3.1-8B-Instruct
  • Mistral-24B-Instruct
  • GPT-4o for demographic-vector instruction generation
  • GPT-5-mini, DeepSeek-V3 and Claude-Haiku-4.5 in robustness comparisons

Instruments and metrics

  • KL divergence
  • Entropy Deviation (absolute normalized-entropy difference)
  • Jensen-Shannon divergence
  • Mean Absolute Error
  • Kernel PCA of final-token hidden states
  • k-nearest-neighbor radius dispersion
  • WVS Q1-Q290

Data used

  • World Values Survey Wave 7 cross-national data, version 6.0 sample files
  • CulturaX language subsets
  • GPT-4o-generated demographic questions and persona instructions

Evidence and location

  • Method, models, sample, metrics, results, ablations, layers, robustness, ethics, limitations, and appendices: arXiv:2603.16142v2, 18/18 pages rendered and individually inspected
  • Version, KDD acceptance, DOI, authors, dates, and official link to the repository: Official arXiv abstract and Atom records inspected 2026-07-17
  • Code, language-vector misalignment, parser, pipeline, vectors, samples, absence of analysis, and version drift: https://github.com/halsayxi/PSII at HEAD 5bae420 and archival tag v1.0.0 f437bf5; local full repository audit 2026-07-17
  • Frozen reproducible file, size, MD5 sum, version, and MIT license: Zenodo concept DOI 10.5281/zenodo.20465632; version DOI 10.5281/zenodo.20466113 inspected 2026-07-17
  • Official non-commercial use condition, mandatory citation, and prohibition of redistributing WVS files: World Values Survey official download conditions and documentation pages inspected 2026-07-17
  • Construct audit, sample, weights, non-response, Q223, statistics, code, privacy, license, and reproducibility: reports/verification/article-389-psii-aggregate-diversity-language-vector-mismatch-layer-selection-leakage-statistics-code-data-privacy-and-reproducibility-audit.json