Cognitive Alignment in Personality Reasoning: Leveraging Prototype Theory for MBTI Inference

Evaluation and psychometric validity2025arXivApproved editorial review

Authors: Haoyuan Li, Yuanbo Tong, Yuchen Li, Zirui Wang, Chunhou Liu, Jiamou Liu

Keywords: Computation and Language, Artificial Intelligence

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

6
Authors
7
Findings
14
Limitations
9
Evidence

Editorial summary

English

The paper proposes ProtoMBTI to infer one of 16 MBTI types from user posts. It augments Kaggle and Pandora with GPT-4o and GPT-4o-mini, filters synthetic examples through a four-dichotomy classifier, LoRA-fine-tunes encoders of at most 2B parameters, and stores each text, embedding, and label as a prototype. At inference, GPT-4o-mini, Qwen2-72B, or Llama-3.1-70B receives a query and its three nearest prototypes for reasoning and voting. In the main results, the best single-model variant reaches 85.14% average dichotomy accuracy and 71.42% direct 16-type accuracy on Kaggle; on Pandora, GPT-4o-mini reaches 71.41% and 60.22%. Mixed-training experiments report as much as 96.41% average dichotomy accuracy and 92.13% 16-type accuracy for within-Pandora evaluation, and roughly 81% 16-type accuracy in cross-domain settings. The baseline comparison is not equivalent for the 16-type task: baseline figures are calculated by multiplying four dichotomy accuracies under an independence assumption, whereas ProtoMBTI is evaluated using a direct multiclass prediction. More importantly, Algorithm 3 compares every test prediction with the example’s ground-truth label and, only when correct, inserts that now-labeled test example into the prototype bank for subsequent predictions. This is sequential supervised test-time adaptation, not label-free inference; it makes results order-dependent and may inflate the mixed-setting results in particular. No ablation separately removes retention. Although the paper says test sets remain raw, every test input also undergoes LLM-generated explanation/augmentation before classification. The appendices do not provide a coherent reconstruction of evaluation: post-augmentation totals differ between Tables 7 and 8, Pandora test counts change across dichotomies, and several confusion matrices contain about 4,800 balanced observations rather than the stated 835 raw test users. Figure 10–12 plot titles also conflict with their captions. ProtoMBTI is therefore an interesting retrieval and augmentation architecture with useful component ablations, but the reported gain magnitude and claimed human-like cognitive alignment are not established reliably. Labels are self-reported MBTI types from online communities, class prompts encode type stereotypes, and interpretability rests on model-generated explanations without human faithfulness validation.

Español

El artículo propone ProtoMBTI para inferir uno de los 16 tipos MBTI a partir de publicaciones de usuarios. El sistema aumenta los datos de Kaggle y Pandora con GPT-4o y GPT-4o-mini, filtra ejemplos sintéticos con un clasificador de cuatro dicotomías, ajusta mediante LoRA codificadores de hasta 2B parámetros y almacena cada texto, su representación y su etiqueta como prototipo. En inferencia, GPT-4o-mini, Qwen2-72B o Llama-3.1-70B recibe el texto consultado y los tres prototipos más próximos para razonar y votar una clase. En los resultados principales, la mejor variante individual alcanza en Kaggle un promedio de 85,14 % en las cuatro dicotomías y 71,42 % de exactitud directa en los 16 tipos; en Pandora, GPT-4o-mini obtiene 71,41 % y 60,22 %, respectivamente. Los experimentos con entrenamiento mixto informan hasta 96,41 % por dicotomía y 92,13 % en 16 tipos dentro de Pandora, y alrededor de 81 % en la clasificación de 16 tipos al transferir entre dominios. Los baselines, sin embargo, no se comparan de forma equivalente en la tarea de 16 tipos: su cifra se calcula multiplicando las cuatro exactitudes bajo un supuesto de independencia, mientras ProtoMBTI se evalúa con una predicción multiclase directa. La amenaza principal afecta al propio protocolo de test. El Algoritmo 3 compara cada predicción con la etiqueta real del ejemplo de prueba y, solo si acierta, incorpora ese ejemplo ya etiquetado al banco para las predicciones posteriores. Esto es adaptación supervisada y secuencial sobre el test, no inferencia sin etiquetas; hace que el resultado dependa del orden y puede inflar especialmente los resultados mixtos. No se ofrece una ablación que elimine por separado esa retención. Además, aunque el texto afirma que el test permanece crudo, cada entrada de test pasa por una explicación/aumento generado por LLM antes de clasificarse. Los anexos tampoco permiten reconstruir con seguridad la evaluación: los totales posteriores al aumento difieren entre las Tablas 7 y 8, los recuentos de test de Pandora cambian según la dicotomía y varias matrices contienen unas 4.800 observaciones equilibradas, no las 835 muestras crudas declaradas. Los rótulos de las Figuras 10–12 se contradicen con sus títulos y pies. Por ello, el trabajo aporta una arquitectura interesante y ablations útiles sobre recuperación y aumento, pero no demuestra de forma fiable la magnitud de las mejoras ni una alineación cognitiva humana. Las etiquetas son autodeclaraciones MBTI de foros, los prompts incorporan estereotipos de cada tipo y la interpretabilidad se apoya en explicaciones generadas por el mismo tipo de modelo, sin validación humana de fidelidad.

Research question

Can a retrieval and reasoning architecture inspired by prototype theory improve inference of the four dichotomies and the 16 MBTI types from text, explain its decisions, and transfer between the Kaggle and Pandora datasets?

Method

Kaggle and Pandora are split into training, validation, and test with a declared ratio of 8:1:1. Only training and validation are balanced by generating examples per class and semantic, linguistic, and sentiment transformations with GPT-4o/GPT-4o-mini, filtered by a BERT, RoBERTa, or DeBERTa classifier of four dichotomies. LoRA fine-tuning of DeepSeek-1B, Qwen2.5-1.5B, and Llama-3-1B is used to represent the examples and build a text–embedding–type triplet bank. At test, the text is transformed with an LLM, the three nearest neighbors are retrieved, and GPT-4o-mini, Qwen2-72B, or Llama-3.1-70B reasons over them. The algorithm checks the prediction against the real label and adds the correct ones to the bank. Accuracy is reported per dichotomy and full type, plus F1, recall, confusion matrices, and ROC in appendices; no repetitions, intervals, statistical tests, or specific seeds are reported.

Sample: Kaggle contains 8,675 users from PersonalityCafe, each with extracts from their 50 most recent posts, and Pandora contains 9,067 users from Reddit. Table 6 confirms these gross totals. After augmentation, Table 7 declares 34,050 examples for Kaggle and 33,994 for Pandora, while Table 8 declares 34,068 and 34,079; the difference is not explained. Table 9 attributes 853 test cases to Kaggle in each dichotomy, except for an error of 91 examples in the I/E training set. In Pandora the test counts sum to 835, 843, 855, and 845 depending on the dichotomy, although each user should contribute the four letters. The mixed Pandora and transfer matrices show approximately 300 observations for each of the 16 types, around 4,800, incompatible with the 835 declared raw test cases. The split identifiers and per-example results are not published.

Findings

  • On Kaggle, ProtoMBTI with Qwen obtains the best individual average per dichotomy, 85.14%, and 71.42% direct accuracy on the 16 types; the variant with GPT-4o-mini reaches 82.66% and 68.39%.
  • On Pandora, GPT-4o-mini obtains the best individual result: 71.41% average per dichotomy and 60.22% on 16 types; Qwen achieves 70.55% and 41.86%, showing that the dichotomous average does not determine multiclass performance.
  • With mixed training, the article reports 93.50%/85.54% for Kaggle test and 96.41%/92.13% for Pandora test in same-domain validation; in cross-transfer it reports around 91% per dichotomy and 81% on 16 types.
  • The Kaggle ablations reduce 16-type accuracy from 71.42% to 50.77% with random prototypes, 54.15% without prototypes, 45.37% with raw text, and 60.62% with only the encoder; there is no ablation of supervised retention during test.
  • Retrieval with k=3 yields 59.38% in the shown sensitivity experiment; k=1 achieves 56.92% and k=9, 55.38%.
  • The DeBERTa control classifier obtains 88.63% average and 71.08% on 16 types in Kaggle validation, a result very close to the 71.42% main result of ProtoMBTI with Qwen.
  • The published improvements cannot be cleanly attributed to prototype reasoning because the test uses its labels to retain examples and the number/origin of evaluated cases does not match across tables and figures.

Limitations

  • Algorithm 3 queries the real label of each test case and adds the correct ones to the bank to predict subsequent cases. It is a supervised and transductive adaptation over the test set, dependent on order, that would not be available in deployment.
  • There is no ablation comparing the same pipeline with and without retention based on the real label; therefore it is not possible to separate how much performance comes from initial retrieval and how much from progressively contaminating the bank with the test set.
  • The 16-type comparison is not equivalent: for baselines the accuracy of four classifiers is multiplied assuming independence, but ProtoMBTI receives credit for a directly observed multiclass prediction.
  • The text says the test remains raw, but Algorithm 3 applies transformation A2 to each post before inferring. The evaluated input thus includes semantic, affective, and stylistic explanations generated by an LLM.
  • The data totals are internally inconsistent: Tables 7 and 8 differ by 18 examples for Kaggle and 85 for Pandora; Table 9 changes the Pandora test total between dichotomies; and several matrices show around 4,800 balanced observations instead of 835 raw cases.
  • Figures 10–12 contain contradictions between the chart title and its caption: Qwen-Pandora is described as Llama in Kaggle, and charts labeled Qwen are presented as Llama. This prevents safely assigning those matrices to a model and dataset.
  • Significance and robustness are claimed without statistical tests, confidence intervals, standard deviations, repeated runs, seed values, or sensitivity analysis to test order.
  • The labels are self-reported MBTI types by forum users, not diagnoses or verified psychometric measurements; label reliability or human agreement is not measured.
  • The augmentation starts from prompts that encode stereotyped descriptions of the 16 types and uses a classifier trained with the same labels as a filter, which can amplify linguistic artifacts and circular separability.
  • The explanations and the qualitative case come from LLMs; there is no human evaluation, causal fidelity test, or verification that the textual reasoning reflects the mechanism that produced the class.
  • t-SNE and embedding separation do not validate the psychological prototype theory or the claim of cognitive alignment. The study includes no cognitive experiment with humans nor comparison with graded human judgments.
  • Group biases, residual privacy, consent, misuse, or harm from inferring personality from posts are not evaluated. PII cleaning is asserted without describing rules or audit.
  • Version v1 does not link a reproducible repository. The manuscript declares code partially generated with GPT-4.1 and validated by the authors, but does not allow inspection of implementation, splits, executed prompts, raw results, or exact API snapshots.
  • The layout retains an unrelated footer of “NeurIPS 2023 Generative AI and Biology Workshop” in a 2025 preprint and figure/table errors, additional signals of insufficient editorial control.

What the study does not establish

  • It does not demonstrate that ProtoMBTI can learn in production from correct predictions without receiving confirmation or an external label; the evaluation protocol does have access to the test truth.
  • It does not establish a valid and comparable improvement over baselines in the full 16-type classification, because those figures are calculated with different procedures.
  • It does not demonstrate that prototype theory describes the cognitive mechanism of the LLM or that there is alignment with human reasoning; it uses an architectural analogy and MBTI labels.
  • It does not validate MBTI as a psychometric construct or convert self-reported forum types into reliable clinical, occupational, or personal traits.
  • It does not prove that the generated explanations are faithful, that the prototypes are interpretable to humans, or that the decision causally depends on the retrieved examples.
  • It does not allow verification of the main figures or knowing which exact set feeds the confusion matrices and ROC from the materials published in the audited version.
  • It does not justify using the system for personnel selection, mental health, surveillance, sensitive personalization, or decisions about individuals.

Traceability

Scope: Full text

Version: arXiv:2511.00115v1, submitted 31 October 2025, 43 pages

Consulted source: https://arxiv.org/pdf/2511.00115v1

Review: Codex full-text, bilingual-fidelity, 43-page visual, source-integrity, test-leakage, split-accounting, baseline-comparability, construct-validity, statistical, interpretability, privacy, reproducibility and editorial-consistency audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o
  • GPT-4o-mini
  • GPT-5 for manuscript refinement
  • GPT-4.1-based coding assistant
  • Qwen2-72B
  • Llama-3.1-70B
  • DeepSeek-1B encoder
  • Qwen2.5-1.5B encoder
  • Llama-3-1B encoder
  • BERT
  • RoBERTa
  • DeBERTa

Instruments and metrics

  • LLM-guided class balancing and semantic, linguistic and sentiment augmentation
  • Four-dichotomy gatekeeper classifier
  • LoRA prototype encoder and cosine top-k retrieval
  • Retrieve–reuse–revise–retain inference cycle with prompt voting
  • Direct four-dichotomy and 16-type accuracy
  • Macro F1, recall, confusion matrices and micro/macro ROC AUC
  • t-SNE visualization
  • Author-designed MBTI stereotype prompts

Data used

  • Kaggle MBTI Personality Type Dataset / PersonalityCafe posts
  • PANDORA Reddit personality dataset
  • LLM-augmented derivatives of Kaggle and PANDORA

Evidence and location

  • Objective, architecture, and main results: Paper, pp. 1–3, Abstract and Introduction; pp. 8–10, Tables 1–2
  • Augmentation, prototype construction, and inference with test truth: Paper, pp. 4–7, Sections 3.1–3.3 and Figure 3; pp. 18–20, Algorithms 1–3
  • Models, splits, metrics, and experimental configuration: Paper, pp. 7–8 and 17–18, Experiment Setup and Implementation Details
  • Ablations of prototypes, augmentation, encoder, and k: Paper, pp. 9–10, Table 2; pp. 23–24, More Results and Figure 6
  • Internally incompatible distributions and counts: Paper, pp. 28–29, Tables 6–9
  • F1, recall, and gatekeeper comparison: Paper, p. 30, Tables 10–13
  • Matrices with contradictory labels and unreconciled sizes: Paper, pp. 31–36 and 40–43, Figures 8–13 and 17–20
  • Use of LLMs, PII, and absence of linked reproducible materials: Paper, p. 15, Appendix A, The Use of Large Language Models; arXiv:2511.00115v1 metadata
  • Comprehensive visual inspection and editorial defects: Paper, all 43 rendered pages; p. 1 unrelated NeurIPS 2023 footer; pp. 33–35 title/caption mismatches