BlossomPsy: A User-Centric AI System for Adaptive and Engaging MBTI Personality Assessments

Evaluation and psychometric validity2026arXivApproved editorial review

Authors: Bingjia Huang

Keywords: MBTI-M, Adaptive personality assessment, Conversational interface, Photo-based questions, Multi-Head Classifier, Modified UCB, PID tuning, Mixed human-LLM evaluation, Psychometric validity, Metric reproducibility, User experience

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

1
Authors
18
Findings
48
Limitations
8
Evidence

Editorial summary

English

BlossomPsy is an MBTI assessment prototype that replaces part of a questionnaire with open conversation and, when its estimator reports low confidence, image pairs. Two Doubao agents ask questions in Chinese. A shared RoBERTa-base feeds four binary heads and one 16-type classifier; modified UCB chooses which dimension to probe and PID tunes two confidence-transformation parameters during synthetic training. Thirteen MBTI-M items are converted into visual pairs by three agents plus human supervision, and nine are retained. This is an adaptive-interface concept, not a validated psychometric instrument or clinical tool.

The current arXiv v2 comparison pools 12 Chinese university students with nine Qwen/Doubao-simulated participants. All 21 complete BlossomPsy and Chinese MBTI-M. Seven match on four dimensions, nine on three, three on two, and two on one; thus 16/21 match at least three. Results are not separated for the 12 humans and nine LLMs. The latter are 43% of the sample and receive prompts instructing them to adopt an MBTI type; their consistency can stress-test the flow but does not add human psychometric evidence. MBTI-M is a reference, not personality ground truth.

The dimensional table is not arithmetically reproducible. It reports accuracy/F1/kappa of 0.76/0.79/0.63 for E/I, 0.67/0.72/0.50 for S/N, 0.95/0.92/0.90 for T/F, and 0.76/0.78/0.62 for J/P. The paper defines accuracy as correct/(correct+incorrect). In binary classification, accuracy 0.76 permits a maximum kappa of about 0.546, so 0.63 and 0.62 are impossible; accuracy 0.67 permits at most about 0.405, not 0.50. Exhaustive enumeration likewise finds no 21-case 2x2 tables reproducing those metric triplets. Only T/F is feasible. If rounded accuracies correspond to 16/21, 14/21, 20/21, and 16/21, Wilson 95% intervals are approximately [0.549,0.894], [0.454,0.828], [0.773,0.992], and [0.549,0.894]. Without labels, confusion matrices, or code, the kappas should not be cited as reliable evidence.

There is also substantive version drift. Version 1 claimed 45 participants, 12 humans and 33 LLMs, and displayed a 16-type distribution summing to 45. One day later, v2 claims 21, the same 12 humans but only nine LLMs, and replaces that distribution with a source table. Match rates, accuracy, F1, kappa, the experience radar, and conclusions remain unchanged. Match percentages correspond exactly to 7, 9, 3, and 2 of 21, but no revision explanation or data reveals which figures use 21, 45, or other subsets.

The higher-satisfaction claim rests only on a radar chart without numeric values. It does not say whether 12 humans, all 21 mixed cases, or another sample responded; items, anchors, means, spread, individual responses, order, counterbalancing, and paired tests are missing. The chart suggests a preliminary interface preference but does not establish a statistical increase in satisfaction or engagement. Completion time, dropout, errors, accessibility, and cognitive load are also absent.

Visual validation is simulation-dominated. Each of 13 pairs receives ten human responses and 40 Qwen simulations, so 80% of 50 votes are synthetic. Pairs are adapted and re-tested until exceeding 66%, and the nine retained items receive no independent validation. Matching the text item measures choice agreement, not whether the image measures an MBTI trait. No item analysis, factor structure, differential-item functioning, or psychologist validation is reported. The text recognizes gender-related visual bias and changes figures based on user gender, but does not validate the rule, address nonbinary identities, or report fairness analysis.

The classifier trains on Personality Cafe, 8,675 users with 50 posts and self-reported labels, then fine-tunes on CPME posts filtered through purported student words and translated into English. Post-filter sizes, selection vocabulary, translator, splits, indices, seeds, and author separation are unavailable. Personality Cafe explicitly discusses MBTI and can leak type names; the paper does not say whether it removes them or segments before splitting users. CPME labels users; absent an author-level split, fragments from one user may cross train and test. Duplicate controls and training-only over/undersampling are also unspecified. The strong RoBERTa row, 0.838/0.920/0.797/0.759, lacks an unambiguous test population and metric. Another table reports 84/87/86/81% in a different unspecified “contextual setting.”

Ablations use only 16 LLM simulations, one per type, without repetitions or uncertainty. The full system shows more matches than some variants, but removing PID, photos, or mUCB also changes information and conversation paths. PID converges near alpha=15.57 and beta=5.51 in a corrupted CPME environment; it is not compared with grid or random search, and the paper concedes simple alternatives may match it. Two-scalar convergence does not demonstrate calibration, better adaptive testing, or a need for PID control.

No code, checkpoint, processed data, raw evaluations, or statistical scripts were found; the arXiv archive contains TeX and figures. Consent, ethics review, compensation, recruitment, and governance of personal conversations are not reported. The defensible contribution is an exploratory UX and orchestration design for making a test more conversational. It does not establish MBTI validity, real personality, Big Five portability, causal component superiority, or suitability for clinical, educational, employment, or other high-stakes decisions.

Español

BlossomPsy es un prototipo de evaluación MBTI que sustituye parte del cuestionario por conversación abierta y, cuando su estimador considera baja la confianza, muestra pares de imágenes. Dos agentes Doubao formulan preguntas en chino. Un RoBERTa-base compartido alimenta cuatro clasificadores binarios y uno de 16 tipos; una variante de UCB decide qué dimensión explorar y un PID ajusta dos parámetros de la transformación de confianza durante entrenamiento sintético. Trece ítems MBTI-M se convierten en pares visuales mediante tres agentes y supervisión humana; nueve se retienen. Es una idea de interfaz adaptativa, no un instrumento psicométrico validado ni una herramienta clínica.

La comparación principal de arXiv v2 mezcla 12 estudiantes universitarios chinos con nueve participantes simulados mediante Qwen y Doubao. Los 21 completan BlossomPsy y MBTI-M chino. Siete coinciden en cuatro dimensiones, nueve en tres, tres en dos y dos en una; es decir, 16/21 coinciden en al menos tres. No se publican resultados separados para las 12 personas y los nueve LLM. Estos últimos representan el 43 % de la muestra y reciben prompts que les ordenan adoptar un tipo MBTI; su consistencia sirve como stress test del flujo, pero no añade evidencia psicométrica humana. MBTI-M se usa como referencia, no como verdad de personalidad.

La tabla dimensional no es aritméticamente reproducible. Reporta accuracy/F1/κ de 0,76/0,79/0,63 para E/I; 0,67/0,72/0,50 para S/N; 0,95/0,92/0,90 para T/F; y 0,76/0,78/0,62 para J/P. El propio artículo define accuracy como aciertos/(aciertos+errores). En clasificación binaria, accuracy 0,76 permite como máximo κ≈0,546, por lo que 0,63 y 0,62 son imposibles; accuracy 0,67 permite como máximo κ≈0,405, no 0,50. Una enumeración exhaustiva tampoco encuentra matrices 2×2 de 21 casos que reproduzcan esos tripletes. Solo T/F es factible. Si los accuracy redondeados fueran 16/21, 14/21, 20/21 y 16/21, sus intervalos Wilson 95 % serían aproximadamente [0,549;0,894], [0,454;0,828], [0,773;0,992] y [0,549;0,894]. Sin etiquetas, matrices o código, no deben citarse los κ como evidencia fiable.

Existe además deriva sustantiva entre versiones. v1 declaraba 45 participantes, 12 humanos y 33 LLM, y mostraba una distribución por 16 tipos que sumaba 45. Un día después, v2 declara 21, los mismos 12 humanos y solo nueve LLM, y reemplaza esa distribución por una tabla de procedencia. Las cifras de coincidencia, accuracy, F1, κ, el radar de experiencia y las conclusiones permanecen iguales. Los porcentajes de coincidencia encajan exactamente con 7, 9, 3 y 2 de 21, pero no hay explicación de la revisión ni datos para saber qué figuras usan 21, 45 u otros subconjuntos.

La afirmación de mayor satisfacción se apoya solo en un radar sin valores numéricos. No se indica si respondieron 12 humanos, los 21 casos mezclados u otra muestra; faltan ítems, anclajes, medias, dispersión, respuestas individuales, orden, contrabalanceo y tests pareados. El gráfico sugiere una preferencia preliminar de interfaz, pero no establece un aumento estadístico de satisfacción o engagement. Tampoco se reportan tiempo, abandono, errores, accesibilidad o carga cognitiva.

La validación visual está dominada por simulación. Cada uno de los 13 pares recibe diez respuestas humanas y 40 de Qwen; por tanto, 80 % de los 50 votos son sintéticos. Los pares se modifican y vuelven a probar hasta superar 66 %, y los nueve retenidos no se validan en una muestra independiente. Coincidir con el ítem textual mide acuerdo de elección, no que la imagen mida el rasgo MBTI. No hay análisis de ítems, estructura factorial, funcionamiento diferencial o validación de psicólogos. El texto reconoce posibles sesgos visuales por género y cambia figuras según el género del usuario, pero no valida esa regla, no contempla identidades no binarias y no presenta análisis de equidad.

El clasificador se entrena en Personality Cafe, 8.675 usuarios con 50 posts y etiquetas auto-reportadas, y se afina en posts CPME filtrados por palabras supuestamente propias de estudiantes, traducidos al inglés. No se publican tamaños después de filtros, vocabulario de selección, modelo de traducción, splits, índices, seeds o separación por autor. Personality Cafe trata explícitamente MBTI y puede filtrar nombres de tipos; el paper no dice si los elimina ni si segmenta antes de dividir usuarios. CPME tiene etiquetas por usuario; sin split por autor, fragmentos del mismo usuario podrían cruzar entrenamiento y test. También faltan controles de duplicados y asegurar que over/undersampling ocurre solo en training. La fila fuerte de RoBERTa, 0,838/0,920/0,797/0,759, carece de una definición inequívoca de población de test y métrica. Otra tabla publica 84/87/86/81 % en un «contextual setting» diferente que no se especifica.

Las ablaciones usan solo 16 LLM simulados, uno por tipo, sin repeticiones o incertidumbre. El sistema completo muestra más coincidencias que algunas variantes, pero eliminar PID, fotos o mUCB cambia también la información y la ruta conversacional. El PID converge cerca de α=15,57 y β=5,51 en un entorno CPME corrompido; no se compara con grid search o random search, y el propio artículo admite que alternativas simples podrían igualarlo. La convergencia de dos escalares no demuestra calibración, mejor testing adaptativo o necesidad de control PID.

No se encontró código, checkpoint, datos procesados, evaluaciones brutas o scripts estadísticos; el tar de arXiv contiene TeX y figuras. Tampoco se informa consentimiento, revisión ética, compensación, reclutamiento o gobernanza de conversaciones personales. La contribución defendible es un diseño exploratorio de UX y orquestación para hacer un test más conversacional. No demuestra validez MBTI, personalidad real, transferibilidad a Big Five, superioridad causal de sus componentes ni aptitud para clínica, educación, empleo o decisiones de alto impacto.

Research question

Can an MBTI conversational interface with adaptive selection of dimensions and visual fallback questions increase perceived experience while maintaining preliminary agreement with MBTI-M?

Method

Prototype with two Doubao agents, multi-head RoBERTa classifier, mUCB and PID tuning. Compares BlossomPsy with MBTI-M on a combined sample of 12 humans and nine LLMs, uses a Likert UX radar, validates 13 visual pairs with 10 humans and 40 Qwen simulations, and performs ablations with 16 LLM persons.

Sample: v2 declares 12 Chinese university students and nine simulated LLMs for consistency; v1 declared 12 and 33. Ten humans and 40 Qwen simulations evaluate each visual pair. The ablations use 16 LLMs, one per MBTI type.

Findings

  • In v2, 7/21 match on four dimensions and 9/21 on three.
  • In total, 16/21 match with MBTI-M on at least three dimensions.
  • Human and LLM results do not separate.
  • Three of four accuracy-kappa pairs are mathematically incompatible.
  • Accuracy 0.76 cannot produce kappa 0.63 or 0.62 in binary classification.
  • Accuracy 0.67 cannot produce kappa 0.50 in binary classification.
  • The 0.95/0.90 pair for T/F is feasible.
  • v1 declared 45 participants and v2 21 without changing results.
  • The match percentages of v2 correspond exactly to 7, 9, 3 and 2 cases.
  • The UX radar offers no values or statistical inference.
  • 80% of visual validation votes come from Qwen.
  • Nine of 13 visual pairs exceed the threshold after adaptation and retest.
  • There is no out-of-sample visual validation.
  • The highest row of the MHC reports 0.838/0.920/0.797/0.759 without a reproducible protocol.
  • The contextual table of the MHC publishes another set 84/87/86/81%.
  • The ablations are LLM-only and without repetitions.
  • PID converges in simulation, but is not compared with simple search.
  • The defensible evidence is of interaction concept, not psychometric validity.

Limitations

  • Preprint v2 without verified peer review.
  • Only 12 human participants in the main comparison.
  • Nine LLMs are mixed with humans in the same metrics.
  • There are no separate results by origin.
  • LLMs receive the target MBTI type in the prompt.
  • MBTI-M is a reference, not ground truth.
  • Three kappa values contradict the printed accuracies.
  • No confusion matrices or individual labels are published.
  • There are no intervals, tests, power or multiplicity correction.
  • v1 and v2 disagree on 45 versus 21 participants.
  • Tables and figures do not change with the sample reduction.
  • The revision between versions is not explained.
  • The UX radar contains no extractable figures.
  • It is not identified who responded to the radar.
  • Items, anchors, distribution and individual UX responses are missing.
  • There is no counterbalancing or random order of MBTI-M and BlossomPsy.
  • Time, dropout, errors or accessibility are not measured.
  • Visual validation uses only ten humans per pair.
  • Qwen provides 40 of 50 votes per pair.
  • Selection, adaptation and visual evaluation reuse the same loop.
  • There is no independent visual confirmation set.
  • Text-image agreement does not validate the MBTI construct.
  • There is no item analysis or factorial structure.
  • No psychologists participate in the reported validation.
  • Gender personalization is not validated.
  • Non-binary gender is not explicitly considered.
  • There is no bias or equity evaluation.
  • Personality Cafe uses self-reported labels and personality context.
  • Removal of MBTI type names is not documented.
  • It is not known whether the split is by user before segmenting posts.
  • CPME is filtered with an unpublished vocabulary.
  • The CPME translation system is not identified.
  • Dataset sizes after filters are missing.
  • Splits, indices, seeds and duplicate control are missing.
  • It is not ensured that oversampling is only on training.
  • There is no external test of students with validated labels.
  • MHC tables use unreconciled configurations.
  • Confidence is not evaluated with probabilistic calibration.
  • The visual reward is set to 1 without justifying correctness.
  • Ablations use a single simulation per type.
  • There is no human ablation or uncertainty.
  • PID is not compared with grid or random search.
  • Exact models, generation parameters or dependencies are not reported.
  • No code, checkpoint or processed data were found.
  • There is no data-availability statement.
  • Consent or ethical review is not reported.
  • Privacy, retention or deletion of conversations is not documented.
  • Transferability to Big Five is not tested.

What the study does not establish

  • That BlossomPsy is a psychometrically valid test
  • That agreement with MBTI-M measures true personality
  • That the published kappas are reproducible
  • That the combined results represent people
  • That the simulated LLMs provide human validity
  • That BlossomPsy statistically increases satisfaction or engagement
  • That the visual pairs measure the intended MBTI traits
  • That gender adaptation is fair or safe
  • That the MHC generalizes beyond the unpublished datasets and splits
  • That mUCB is calibrated as psychological confidence
  • That PID is necessary or superior to simple search
  • That the ablations identify stable causal effects
  • That the system transfers to Big Five without new validation
  • That it is suitable for diagnosis, advice, education or employment
  • That the artifact can be reproduced end to end

Traceability

Scope: Full text

Version: arXiv:2607.06149v2; 23-page preprint; PDF, all pages, v1/v2 TeX, publication status, method, metrics, figures, datasets, UX and claim boundaries audited 2026-07-16

Consulted source: https://arxiv.org/abs/2607.06149v2

Review: Codex 23-page full-text visual, arXiv v1/v2 TeX diff, metric arithmetic, dataset, psychometric, UX, ethics and artifact audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • Chinese RoBERTa-base encoder
  • BERT classifier baseline
  • XGBoost baseline
  • SVM baseline
  • Doubao-1.5-pro family dialogue agents; exact snapshot not reported
  • Qwen2.5-VL/Qwen family simulated users; exact snapshot not reported
  • Unspecified interpreter, image-generator and monitor LLMs

Instruments and metrics

  • Chinese MBTI-M reference scale
  • Four binary MBTI dimensions and one 16-class head
  • Multi-Head Classifier
  • Modified UCB and LCB intervals
  • PID tuning of alpha and beta
  • Five-dimension Likert UX radar
  • Text-photo response agreement threshold of 0.66
  • Accuracy, F1 and Cohen kappa
  • Wilson interval sensitivity audit

Data used

  • Personality Cafe Kaggle: 8,675 users with 50 posts and self-reported MBTI labels
  • CPME: 11,338 Weibo users and 566,900 posts with user-level MBTI and affect labels
  • Unreleased filtered and translated CPME college-language subset
  • 21-case mixed MBTI-M consistency comparison in v2
  • 13 generated photo pairs; nine retained
  • 16 LLM-only ablation trajectories
  • No released code, processed data, raw labels, ratings or outputs

Evidence and location

  • Method, results, figures, appendices and limitations: arXiv:2607.06149v2 PDF, 23 pages; every page rendered and visually inspected
  • Version, date, category and editorial status: Official arXiv record for 2607.06149v2
  • Drift from 45 to 21 participants: Line-by-line TeX diff between arXiv v1 source sha256:45dab150 and v2 source sha256:6884c54f
  • Accuracy-kappa incompatibility and descriptive intervals: Exhaustive binary confusion-matrix and Wilson interval audit from printed Table 2, 2026-07-16
  • Size and structure of CPME: Official arXiv:2411.08347 record
  • Structure and leakage risk of Personality Cafe: Cited Kaggle dataset description and published user-level dataset documentation
  • Absence of publication and official repo: Exact-title arXiv, OpenReview and GitHub searches checked 2026-07-16
  • Consolidated audit: reports/verification/article-280-blossompsy-mbti-mixed-human-llm-kappa-arithmetic-version-drift-psychometric-code-data-and-ux-audit.json