GenPT: Beyond Self-Report for Reliable LLM Psychometrics via Generative Projective Testing

Evaluation and psychometric validity2026ACL AnthologyApproved editorial review

Authors: Ming Wang, Shuang Wu, Bixuan Wang, Lu Lin, Yuxin Chen, Xiaocui Yang, Daling Wang, Shi Feng, Yifei Zhang, Yufan Sun

Keywords: Persona conditioning, Psychometrics, Safety and bias

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

10
Authors
9
Findings
15
Limitations
5
Evidence

Editorial summary

English

GenPT proposes evaluating persona-conditioned LLM agents through newly generated projective tasks rather than relying only on direct questionnaires. Examinees are multimodal agents, not people. Each receives a fictional-character or mental-health profile and answers eight TAT-like scenes, ten Rorschach-like cards, and twenty sentence stems. The repository contains the full reported pool: twenty-eight TAT scenes in a 13/10/5 split, thirteen images representing ten cards, and ninety-seven stems. An LLM Interpreter maps responses to eight SCORS-G dimensions, four Simplified Rorschach Analysis System scores, and five sentence-completion domains; a Diagnostician using the same backbone predicts Big Five, MBTI, depression, or suicide ideation. Qwen3-8B, Phi-4-mini-reasoning, and Intern-S1-mini are compared as Interpreter and Diagnostician while Examinee outputs are fixed. Personality uses fifteen CharacterRAG characters with community labels from Personality Database; mental health uses fifteen AnnaAgent profiles sampled from 1,338 D4-derived profiles. Baselines have the role-playing agent answer BFI, 16Personalities, BDI-II, and BSS. Criterion validity is exact label match except for four-axis MBTI Hamming distance. Questionnaires outperform GenPT on personality: Big Five is .373 versus .333/.240/.293 and MBTI error is .733 versus 1.200/2.200/1.667. On depression, questionnaire accuracy is .133 and GenPT is .200/.400/.400; on suicide, .200 versus .400/.267/.067. These are samples of fifteen: .400 is six exact matches, and the advantage is not uniform because Intern-S1 is worse than the questionnaire on suicide. A trivial predictor choosing the released modal value for each Big Five trait reaches .400 across seventy-five labels, above every reported method, but the paper does not report it. AnnaAgent labels and class distributions are withheld, preventing the same check for risk tasks. Social-desirability experiments compare neutral prompts with job-interview and confidential-counselling framing using weighted kappa and Directional Consistency Ratio, which conditions only on changed items. Questionnaire suicide labels show a stable downward shift: kappa .67/.79 and downward DCR .71/.88. GenPT does not reproduce that joint signature, but its risk kappas range from -.42 to .05 and some DCRs reach .82 in varying directions. This supports absence of the same fake-good pattern, not reliability or general lack of bias; DCR near .5 can also be symmetric noise. Questionnaire trait kappas are .71-.85 versus .20-.63 for GenPT. With ten prior counselling turns, Qwen3 shifts depression .80 and suicide .20 versus questionnaire .08/.10; Phi-4 shifts .00/.07 and Intern-S1 .07/.13. The order-of-magnitude effect is Qwen3-specific, as the paper acknowledges. Further methodological limits matter. PDB consists of community votes about fictional characters, not psychometric ground truth for the assessed agent; profiles include descriptions related to labels, so the task measures recovery of an assigned persona. There are no intervals, repeated runs, significance tests, class analysis, or persona-selection sensitivity. The validity and reliability AnnaAgent subsets only partially overlap, and the latter was selected from cases with pre-generated behaviors available. The same model interprets and diagnoses. Nine human experts review stimuli and annotate responses, but labels, assignments, agreement, and quantitative comparison are absent even though the appendix calls them a gold standard. The contamination-free claim rests on an illustrative Gemini response recognizing one traditional TAT image. New stimuli reduce literal item memorization but do not prove that images, stems, rubrics, or concepts were absent from training. Adapting TAT and Rorschach to generated language also does not validate clinical equivalence. The checklist answers No on documenting identifying/offensive-content and anonymization procedures and N/A on recruitment, consent, and ethics review despite human annotators, in tension with the paper's generic anonymization claim. The paper explicitly says the method masks assessment intent and can bypass alignment filters, creating an unevaluated covert-profiling risk. The repository is inspectable but not a reproduction. It compiles, contains uv.lock, builds sdist and wheel, and publishes all stimuli and CharacterRAG profiles. Yet AnnaAgent data, outputs, results, human annotations, and kappa/DCR/longitudinal aggregators are absent. The CLI accepts model, API, and seed controls but the main runner ignores them, uses /home/aiscuser/models paths, and hard-codes seed 42. Its depression prompt asks for 0-4 although the paper uses 0-3. Reliability evaluation expects lists where the runner emits dictionaries and fails with AttributeError; validity code also expects incompatible fields. The wheel omits stimuli, characters, questionnaires, and scripts, so a clean installation reports zero loaded items. Tests, CI, license, container, and model manifest are missing. The defensible contribution is a promising research design and open stimulus set that separates framing from context, with preliminary evidence that the Qwen3 configuration can avoid one directional signature and respond to counselling. It does not establish a reliable psychometric instrument, clinical diagnosis, freedom from contamination or bias, psychological ground truth, or end-to-end reproducibility.

Español

GenPT propone evaluar agentes LLM condicionados por una persona mediante tareas proyectivas generadas, en vez de depender solo de cuestionarios directos. Los Examinees son agentes multimodales, no personas. Cada agente recibe un perfil de un personaje ficticio o de salud mental y responde a ocho escenas tipo TAT, diez láminas simétricas tipo Rorschach y veinte inicios de frase. El repositorio publica el conjunto completo declarado: 28 escenas TAT en una proporción 13/10/5, 13 imágenes para diez tarjetas y 97 frases. Un Interpreter LLM convierte las respuestas en ocho dimensiones SCORS-G, cuatro puntuaciones de un Simplified Rorschach Analysis System y cinco dominios de sentence completion; un Diagnostician del mismo backbone predice Big Five, MBTI, depresión o ideación suicida. El estudio compara Qwen3-8B, Phi-4-mini-reasoning e Intern-S1-mini como Interpreter y Diagnostician, manteniendo fijas las respuestas del Examinee. La personalidad se evalúa con 15 personajes CharacterRAG y etiquetas comunitarias de Personality Database; salud mental usa 15 perfiles AnnaAgent tomados de 1.338 perfiles derivados de D4. Como baseline, el agente responde BFI, 16Personalities, BDI-II y BSS. La validez se reduce a coincidencia exacta de etiquetas, salvo MBTI, que usa distancia Hamming de cuatro ejes. El cuestionario supera a GenPT en personalidad: Big Five 0,373 frente a 0,333/0,240/0,293 y error MBTI 0,733 frente a 1,200/2,200/1,667. En depresión, questionnaire obtiene 0,133 y GenPT 0,200/0,400/0,400; en suicidio, 0,200 frente a 0,400/0,267/0,067. Son muestras de 15: 0,400 equivale a seis aciertos y la ventaja no es uniforme, pues Intern-S1 queda por debajo del cuestionario en suicidio. Además, un predictor que siempre elige la moda de cada rasgo Big Five alcanza 0,400 sobre las 75 etiquetas publicadas, por encima de todos los métodos; el paper no informa este baseline. Las etiquetas AnnaAgent y sus distribuciones no se publican, de modo que la comparación equivalente no puede comprobarse. Para resistencia a deseabilidad social, el trabajo compara neutral con entrevista laboral y counselling confidencial mediante kappa ponderada y Directional Consistency Ratio, que solo mira la dirección entre los ítems que cambian. El cuestionario muestra en suicidio una deriva estable hacia menor riesgo: kappa 0,67/0,79 y DCR descendente 0,71/0,88. GenPT no reproduce esa combinación, pero sus kappas de riesgo están entre -0,42 y 0,05 y algunas DCR llegan a 0,82 en direcciones variables. Eso respalda ausencia de la misma firma fake-good, no fiabilidad ni ausencia general de sesgo: DCR cercana a 0,5 también puede ser ruido simétrico. En personalidad, el cuestionario conserva kappa 0,71–0,85 y GenPT solo 0,20–0,63. Con diez turnos previos de counselling, Qwen3 cambia depresión 0,80 y suicidio 0,20 frente a 0,08/0,10 del cuestionario; Phi-4 cambia 0,00/0,07 e Intern-S1 0,07/0,13. El efecto de orden de magnitud es específico de Qwen3, como reconoce el texto. La auditoría metodológica encuentra más límites. PDB son votos comunitarios sobre ficción, no ground truth psicométrico del agente; los perfiles contienen descripciones relacionadas con las etiquetas y la prueba mide recuperación de una persona asignada. No hay intervalos, repeticiones, significación, análisis por clase o sensibilidad a la selección de personas. Los dos conjuntos AnnaAgent de validez y fiabilidad solo se solapan parcialmente y el segundo se eligió entre casos con conductas ya generadas. El mismo modelo interpreta y diagnostica. Nueve expertos humanos revisan estímulos y anotan respuestas, pero no se publican etiquetas, acuerdo, asignaciones o comparación cuantitativa, aunque el apéndice las llama gold standard. La supuesta ausencia de contaminación se apoya en un ejemplo donde Gemini reconoce una imagen TAT clásica; crear estímulos nuevos reduce memorización literal, pero no demuestra que imágenes, frases, rúbricas o conceptos estén libres de entrenamiento. Adaptar TAT y Rorschach a texto generado tampoco valida equivalencia clínica. El checklist marca No en documentación de identificación, contenido ofensivo y anonimización, y N/A en reclutamiento, consentimiento y revisión ética pese a usar anotadores; esto contrasta con la afirmación genérica de anonimización del paper. El propio trabajo dice que el método oculta la intención y puede evitar filtros de alineamiento, un riesgo de perfilado encubierto no evaluado. El repositorio es un prototipo inspeccionable, no una reproducción. Compila, incluye uv.lock, construye sdist y wheel y contiene todos los estímulos y perfiles CharacterRAG. Sin embargo, faltan AnnaAgent, outputs, resultados, anotaciones humanas y agregadores de kappa, DCR y longitudinalidad. El CLI acepta modelo, API y seed pero la ejecución principal ignora esos valores, usa rutas /home/aiscuser/models y semilla 42. El prompt de depresión pide 0–4 aunque el paper usa 0–3. El evaluador de fiabilidad espera listas donde el runner escribe diccionarios y falla con AttributeError; el de validez también espera campos incompatibles. El wheel omite estímulos, personajes, cuestionarios y scripts, por lo que una instalación limpia informa cero elementos. No hay tests, CI, licencia, contenedor ni manifiesto de modelos. La contribución defendible es un diseño de investigación prometedor y un conjunto de estímulos abierto que separa framing de contexto, con evidencia preliminar de que la configuración Qwen3 puede evitar una firma direccional concreta y responder a counselling. No demuestra un instrumento psicométrico fiable, diagnóstico clínico, ausencia de contaminación o sesgo, ground truth psicológico ni reproducibilidad end-to-end.

Research question

Can a generative adaptation of projective tests assess the persona expressed by LLM agents more robustly against contamination, social framing, and longitudinal context than direct questionnaires?

Method

Comparative protocol on agents conditioned by profiles. The Examinee responds to new TAT, Rorschach, and sentence completion stimuli; an Interpreter extracts SCORS-G, SRAS, and SCT indicators; a Diagnostician predicts four targets. Three backbones of 3.84-8B are compared with fixed Stage 1 responses, direct questionnaires, two framings, and a counselling context. The independent audit reviews the 17 ACL v2 pages, 2 checklist pages, TeX, stimuli, published labels, code, evaluators, clean build, and wheel.

Sample: The unit is an agent conditioned by a persona. Validity uses 15 CharacterRAG and 15 AnnaAgent; reliability uses the same 15 CharacterRAG and another fixed set of 15 AnnaAgent partially overlapping. Each session has 8 TAT, 10 Rorschach cards, and 20 SCT. There are no human participants as Examinees; there are nine human annotators.

Findings

  • The questionnaire outperforms GenPT on Big Five and MBTI for the three backbones.
  • All Big Five results fall below the independent modal baseline of 0.400 calculated over the published labels.
  • GenPT reaches 0.400 on depression with Phi-4 and Intern-S1 and on suicide with Qwen3, but each 0.400 is six correct out of fifteen.
  • Intern-S1 obtains 0.067 on suicide, below the 0.200 of the questionnaire; there is no uniform advantage in risk.
  • The questionnaire shows a stable downward drift in suicide, DCR 0.71/0.88 with kappa 0.67/0.79.
  • GenPT avoids that specific combination, but its risk kappas are low or negative and do not certify reliability.
  • Qwen3 responds to counselling with a 0.80 change in depression; Phi-4 and Intern-S1 do not reproduce that effect size.
  • The release contains exactly the 28/13/97 declared stimuli and 15 CharacterRAG profiles.
  • The public artifact cannot reconstruct any table and its evaluators do not accept the format produced by its runner.

Limitations

  • PDB labels are community votes about characters, not psychometric ground truth of the agent.
  • AnnaAgent profiles, labels, class distributions, and selections are not public.
  • There are only 15 personas per family, without replications, intervals, significance, or sensitivity analysis.
  • DCR ignores stable items and a symmetric DCR may reflect noise; it should be read with very low kappas.
  • Job interview and honest counselling are not symmetric manipulations and the expected clinical direction is presupposed.
  • Interpreter and Diagnostician share backbone, without independent validation of the inferences.
  • The nine expert annotations do not include labels, agreement, adjudication, or quantitative results.
  • The contamination test is a Gemini example on a classic image, not a systematic study.
  • The protocol does not demonstrate equivalence with human clinical projective tests.
  • The checklist omits concrete privacy, consent, and ethical review procedures for annotators.
  • Hiding the intention to avoid filters creates a risk of covert profiling that is not evaluated.
  • Results, AnnaAgent, aggregators, exact Examinee model, and reproducible configuration are missing.
  • CLI and runner ignore flags, use private paths, and inconsistent depression scale.
  • The evaluators fail with the native format and the wheel is distributed without essential data.
  • There are no tests, CI, license, container, or version and hardware manifest.

What the study does not establish

  • It does not demonstrate that GenPT is a reliable or valid psychometric test in a clinical sense.
  • It does not demonstrate intrinsic psychological states of the model; it evaluates recovery of conditioned profiles.
  • It does not demonstrate advantage over trivial baselines on Big Five nor does it allow testing them in mental health.
  • It does not demonstrate absence of bias from DCR close to 0.5 with low stability.
  • It does not demonstrate zero contamination of stimuli, rubrics, or concepts.
  • It does not demonstrate that adapted TAT or Rorschach are equivalent to human instruments.
  • It does not demonstrate generality of the longitudinal effect beyond Qwen3.
  • It does not authorize human diagnosis, clinical use, or profiling without consent.
  • It does not allow reproducing tables or conclusions end-to-end with the current repository.

Traceability

Scope: Full text

Version: ACL Anthology 2026.acl-long.1901v2; DOI 10.18653/v1/2026.acl-long.1901

Consulted source: https://aclanthology.org/2026.acl-long.1901/

Review: Codex seventeen-page ACL-v2 visual, two-page checklist, TeX, stimulus, label-distribution, psychometric-claim, human-annotation, ethics, repository, evaluator-schema, build and clean-wheel audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Qwen3-8B como Interpreter y Diagnostician
  • Phi-4-mini-reasoning 3.84B como Interpreter y Diagnostician
  • Intern-S1-mini aproximadamente 8B como Interpreter y Diagnostician
  • Qwen3-VL-8B-Instruct como Examinee por defecto en el release; el paper no especifica por completo el modelo Stage 1
  • FLUX.1-dev y Stable Diffusion para generar estímulos
  • Gemini en la prueba ilustrativa de contaminación

Instruments and metrics

  • SCORS-G, ocho dimensiones TAT de 1 a 7
  • Simplified Rorschach Analysis System, 22 variables y cuatro dominios
  • Sentence Completion Test, ítems 0-6 y cinco dominios
  • Big Five Inventory y 16Personalities
  • Beck Depression Inventory II y Beck Scale for Suicide Ideation
  • Exact-match accuracy y distancia Hamming MBTI
  • Cohen kappa linealmente ponderada
  • Directional Consistency Ratio
  • Cambio absoluto medio bajo contexto longitudinal

Data used

  • CharacterRAG: 15 personajes de anime y etiquetas PDB
  • AnnaAgent ampliado desde D4: 1.338 perfiles reportados, 15 por experimento
  • Pool GenPT: 28 TAT, 13 Rorschach para 10 tarjetas y 97 SCT
  • Release público: 15 perfiles CharacterRAG, sin perfiles AnnaAgent ni resultados
  • Nueve anotadores expertos descritos, anotaciones no publicadas

Evidence and location

  • Metadata, published version, DOI, and pages: ACL Anthology 2026.acl-long.1901v2, checked 2026-07-17
  • Method, results, tables, limitations, ethics, prompts, sampling, and annotation: ACL v2, all seventeen PDF pages and complete TeX source
  • Privacy, humans, consent, ethical review, and AI assistance: Responsible NLP Checklist, both pages visually inspected
  • Stimuli, labels, implementation, build, wheel, and reproducibility: sci-m-wang/GenPT commit 1baeb5513fcd8efbe6acbd387ef1633304188ae7, checked 2026-07-17
  • Validity audit, baselines, uncertainty, ethics, and code: reports/verification/article-309-genpt-projective-psychometrics-validity-uncertainty-human-annotation-ethics-and-repository-audit.json