PTEI: Integrating Personality Traits to Enhance Emotional Intelligence in Large Language Models

Personas, identity, and agents2026arXivApproved editorial review

Authors: Amir Reza Jafari, Praboda Rajapaksha, Reza Farahbakhsh, Noel Crespi

Keywords: Emotional understanding, Personality-aware prompting, MBTI, OCEAN, Contrastive retrieval, EmoBench, Synthetic annotations, Psychometric validity

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

4
Authors
13
Findings
24
Limitations
6
Evidence

Editorial summary

English

PTEI is an arXiv v1 preprint asking whether inferred personality traits and retrieved examples help answer emotional-understanding questions. It does not measure human personality or show that an LLM has genuine emotions or emotional intelligence. The study evaluates all 200 English Emotional Understanding scenarios in EmoBench. Each scenario has one emotion question and one cause question; Overall counts an item only when both are correct. The system uses an unspecified GPT-4 version to generate a 500-scenario synthetic memory bank matched to the test taxonomy and category proportions. GPT-4o-mini infers one MBTI type and five discrete OCEAN levels, low, medium or high, from every text. all-mpnet-base-v2 is contrastively fine-tuned to retrieve examples sharing emotion and similar profiles; PTEI-Base adds these examples and labels to the prompt, while PTEI-CoT also requests step-by-step reasoning.

Personality here is a generated interpretation of the same emotional scenario that the final model must answer. The detection prompts even receive the EmoBench category name and explanation. There is no questionnaire, self-report, longitudinal behavior or independent psychometric ground truth. Improvements may therefore come from GPT-4o-mini encoding emotional clues, extra text and compute, or labeled demonstrations rather than personality knowledge specifically. Missing controls include shuffled or wrong traits, a neutral summary matched for length and cost, inference without category metadata and validation against independent human labels. OCEAN is also mapped to 0/0.5/1 vectors and compared by cosine: an all-Low profile is a zero vector with undefined cosine, and the paper does not describe its handling.

Results are mixed rather than consistently positive. Without CoT, PTEI raises Overall for all four models: Qwen-7B 22.50→23.25 (+0.75 points), Llama-3.1-8B 16.62→17.63 (+1.01), Qwen-14B 35.50→36.12 (+0.62), and GPT-4o 60.25→62.12 (+1.87). With CoT, Llama rises 12.25→14.13 (+1.88), Qwen-14B 30.12→34.38 (+4.26), and GPT-4o 58.88→63.62 (+4.74), but Qwen-7B falls 21.38→20.88 (-0.50) and declines in three of four categories plus both marginal accuracies. Even GPT-4o PTEI-Base loses 5.10 points on Complex Emotions and 2.23 on Perspective-Taking. The abstract's “additional 4%” is best read as roughly the 4.74 percentage-point gain in one GPT-4o comparison, not a universal effect.

The word “significantly” is unsupported: no item-level outputs, intervals, paired tests, seeds or run-to-run variation are published. Several Qwen Base/CoT rows are imported from EmoBench, which uses five generations for each of four option orders, whereas PTEI reports three; the table therefore mixes inference budgets. The ablation prose is also contradicted by its cells: RAG-only is below Base for Qwen-14B without CoT, while RAG-only and full PTEI are below CoT for Qwen-7B. No interaction or synergy test separates combined MBTI/OCEAN or traits/retrieval effects.

Memory-bank quality has another internal contradiction. Gemma2 and Claude filter scenarios and only 10% proceed to blinded human evaluation, yet the later equation requires a human score and unanimous human acceptance for every retained item, and the prose says all were human-validated. It does not explain how the unreviewed 90% meet that rule. The reported κ=0.92 omits the number of annotators and items, contingency table, mapping from four ordinal ratings to categorical decisions and uncertainty; it cannot by itself validate automated judging at scale. The human-level comparison is not direct either: PTEI has no human row, while the original EmoBench human study used 30-item subsets rather than the full 200-item results reported here.

Finally, no PTEI code or data is linked. The arXiv source package contains only TeX and figures; the 500-item memory bank, trait labels, ratings, checkpoint, outputs and scripts were not located on GitHub or Hugging Face. Retrieval k, exact model snapshots, full demonstrations, emotion list, validation split and failure rules are also omitted. Overall, the table supports small gains in the Base condition and larger CoT gains for two bigger models, alongside a Qwen-7B failure. It is evidence for an EmoBench-specific combination of prompting, retrieval and synthetic annotation, not validated psychological personality, statistical significance, causal attribution, generalization or reproducibility.

Español

PTEI es un preprint arXiv v1 que estudia si añadir rasgos de personalidad inferidos y ejemplos recuperados ayuda a responder preguntas de comprensión emocional. No mide personalidad humana ni demuestra que un LLM tenga emociones o inteligencia emocional genuina. Evalúa los 200 escenarios en inglés de Emotional Understanding de EmoBench. Cada escenario contiene una pregunta sobre la emoción y otra sobre su causa; la métrica Overall solo cuenta un caso cuando ambas respuestas son correctas. El sistema genera con una versión no identificada de GPT-4 una memoria de 500 escenarios sintéticos ajustada a la taxonomía y proporciones del test. GPT-4o-mini infiere de cada texto un tipo MBTI y cinco valores OCEAN discretos, bajo, medio o alto. all-mpnet-base-v2 se afina por contraste para recuperar ejemplos que compartan emoción y perfiles similares; PTEI-Base añade esos ejemplos y etiquetas al prompt y PTEI-CoT solicita además razonamiento paso a paso.

La personalidad aquí es una interpretación generada a partir del mismo escenario emocional que después se debe resolver. Los prompts de detección reciben incluso el nombre y explicación de la categoría de EmoBench. No hay cuestionarios, autoinforme, conducta longitudinal ni etiquetas psicométricas independientes. Por ello, la mejora puede proceder de que GPT-4o-mini ya codifica pistas emocionales, del texto y cómputo adicionales o de los ejemplos etiquetados, no específicamente de conocimiento de personalidad. Faltan controles con rasgos barajados o erróneos, un resumen neutral de igual longitud y coste, inferencia sin metadatos de categoría y validación contra anotaciones humanas. Además, OCEAN se convierte en vectores 0/0,5/1 y se compara por coseno: un perfil con todos los rasgos bajos es el vector cero, cuyo coseno no está definido, y el artículo no explica cómo lo trata.

Los resultados son mixtos, no consistentemente positivos. Sin CoT, PTEI eleva Overall en los cuatro modelos: Qwen-7B 22,50→23,25 (+0,75 puntos), Llama-3.1-8B 16,62→17,63 (+1,01), Qwen-14B 35,50→36,12 (+0,62) y GPT-4o 60,25→62,12 (+1,87). Con CoT, Llama sube 12,25→14,13 (+1,88), Qwen-14B 30,12→34,38 (+4,26) y GPT-4o 58,88→63,62 (+4,74), pero Qwen-7B baja 21,38→20,88 (-0,50) y empeora en tres de cuatro categorías y en ambas exactitudes marginales. Incluso GPT-4o PTEI-Base pierde 5,10 puntos en Complex Emotions y 2,23 en Perspective-Taking. El “4% adicional” del abstract describe aproximadamente los 4,74 puntos porcentuales de una comparación GPT-4o concreta; no es un efecto universal.

Tampoco se sostiene la palabra “significativamente”: no se publican outputs por ítem, intervalos, pruebas pareadas, semillas ni variación entre ejecuciones. Varias filas Base/CoT de Qwen se importan de EmoBench, que usa cinco generaciones por cada uno de cuatro órdenes de opciones, mientras PTEI declara tres; se mezclan así presupuestos de inferencia. La ablación también contradice su prosa: RAG-only baja frente a Base para Qwen-14B sin CoT, y con CoT RAG-only y PTEI bajan para Qwen-7B. No se prueba una interacción o sinergia entre MBTI y OCEAN, ni entre rasgos y recuperación.

La calidad de la memoria presenta otra contradicción. Gemma2 y Claude filtran escenarios y solo el 10% pasa a evaluación humana ciega, pero después la ecuación exige puntuación y aceptación humana unánime para cada caso retenido y el texto afirma que todos fueron validados por humanos. No se explica cómo cumple el 90% no revisado. El κ=0,92 carece de número de anotadores e ítems, tabla de contingencia, regla para convertir cuatro escalas ordinales en categorías e intervalo; no permite validar por sí solo el juez automático. La comparación con nivel humano tampoco es directa: PTEI no incluye fila humana y el estudio humano original de EmoBench usó subconjuntos de 30 preguntas, no estos 200 resultados completos.

Finalmente, no hay código ni datos PTEI enlazados. El paquete fuente de arXiv solo contiene TeX y figuras; no se localizaron la memoria de 500 casos, etiquetas, ratings, checkpoint, outputs o scripts en GitHub o Hugging Face. Se omiten además k de recuperación, versiones exactas de modelos, demostraciones completas, lista de emociones, partición de validación y reglas de fallos. En conjunto, la tabla apoya mejoras pequeñas en la condición Base y mejoras mayores con CoT para dos modelos grandes, junto con un fallo en Qwen-7B. Es evidencia de una combinación de prompting, recuperación y anotación sintética específica de EmoBench; no valida personalidad psicológica, significación estadística, causalidad, generalización ni reproducibilidad.

Research question

Does the joint accuracy of emotion and cause improve on the 200 English items of EmoBench when adding MBTI/OCEAN traits inferred by another LLM and synthetic examples retrieved through a contrastive encoder?

Method

Experimental study on EmoBench EU. Generates 500 synthetic scenarios with GPT-4, filters them with Gemma2/Claude and humanly audits 10%; GPT-4o-mini infers MBTI and OCEAN; all-mpnet-base-v2 is fine-tuned with NT-Xent; and four LLMs are compared on Base, CoT, TraitsOnly, RAG-only, PTEI-Base and PTEI-CoT through voting and four option orders.

Sample: Two hundred English items from EmoBench, each with one emotion question and one cause question, plus 500 synthetic memory scenarios. No real people or observed psychometric profiles are studied.

Findings

  • PTEI-Base improves Overall over Base across the four models, with gains between 0.62 and 1.87 points.
  • PTEI-CoT improves over CoT in Llama-3.1-8B, Qwen-14B and GPT-4o.
  • The largest gain is GPT-4o: 58.88 to 63.62, +4.74 percentage points.
  • Qwen-7B PTEI-CoT worsens Overall from 21.38 to 20.88 and falls in most columns.
  • GPT-4o PTEI-Base improves Overall but worsens Complex Emotions and Perspective-Taking.
  • TraitsOnly and RAG-only produce small or negative improvements depending on model and condition.
  • There are no statistical tests supporting the word significantly.
  • The Qwen rows imported from EmoBench use five repetitions per order, versus three declared by PTEI.
  • Personality is inferred from the same scenario and may transmit emotional cues from the annotator LLM.
  • The 10% human audit contradicts the claim of human validation of all retained cases.
  • The κ=0.92 is not interpretable without definition of raters, sample, decision and contingencies.
  • No code, memory, labels, checkpoint, outputs or PTEI scripts are published.
  • The comparison with humans is not run on the same 200 items and does not appear in the PTEI table.

Limitations

  • It is an arXiv v1 preprint with no indicated archival acceptance.
  • It only uses the 200 English items of one task from one benchmark.
  • Emotional responses may admit plausible alternative interpretations.
  • MBTI has limited empirical robustness and test-retest reliability.
  • MBTI and OCEAN are inferred from a brief vignette and are not ground truth.
  • Personality labels may directly encode cues about emotion and cause.
  • The prompts receive benchmark category metadata.
  • The memory replicates the taxonomy and category distribution of the test.
  • There are no controls with shuffled, erroneous traits or neutral summaries.
  • The OCEAN cosine is undefined for the all-Low traits vector if there is no special treatment.
  • The retrieval k is not reported.
  • Exact snapshots and generation parameters for several models are missing.
  • No validation split, epochs, projection head or seeds are published.
  • The demonstrations and the predefined list of emotions are not shown in full.
  • Imported baselines and PTEI use a different number of repetitions.
  • There are no per-item outputs, intervals, paired tests or variation across runs.
  • Several cells contradict the claims of consistent improvement.
  • Only the filtered 10% receives human evaluation according to the method.
  • The size, recruitment, training, compensation or ethics of annotators is not reported.
  • The κ=0.92 lacks information to be reproduced or interpreted.
  • The synthetic correlations do not include significance or correction for multiple comparisons.
  • There is no public code, data, checkpoint or logs for PTEI.
  • Generalization to other languages, benchmarks, models or real people is not demonstrated.
  • The profile is static and omits dynamic and contextual personality.

What the study does not establish

  • That PTEI improves all models, categories or metrics
  • That the differences are statistically significant
  • That there is a universal 4% gain
  • That PTEI directly and in a measured way reduces the gap with humans
  • That inferred MBTI or OCEAN are valid measures of personality
  • That personality causes the observed improvements
  • That LLMs possess genuine personality, emotions or emotional intelligence
  • That the synthetic correlations describe human psychology
  • That the 500 cases were validated by human annotators
  • That κ=0.92 validates the automatic judge at scale
  • That the method is independent of EmoBench
  • That it generalizes to other languages, cultures, domains or tasks
  • That the study is reproducible from public artifacts
  • That the preprint has passed peer review or been accepted at a conference

Traceability

Scope: Full text

Version: arXiv:2607.10245v1; 15-page preprint; TeX source, OpenReview record, public artifact availability and official EmoBench benchmark/code audited 2026-07-16

Consulted source: https://arxiv.org/abs/2607.10245v1

Review: Codex 15-page full-text visual, TeX-source, arXiv/OpenReview metadata, official EmoBench paper/code/protocol and public artifact audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4o
  • GPT-4o-mini
  • GPT-4 no especificado para generación
  • Qwen-7B-Chat
  • Qwen-14B-Chat
  • Llama-3.1-8B-Instruct
  • Gemma2 no especificado
  • Claude no especificado
  • all-mpnet-base-v2

Instruments and metrics

  • EmoBench Emotional Understanding
  • MBTI inferido por prompt
  • OCEAN inferido en niveles Low/Medium/High
  • Memoria sintética de escenarios
  • NT-Xent
  • Similitud MBTI/OCEAN compuesta
  • KNN
  • Few-shot prompting
  • Chain-of-Thought
  • Exactitud conjunta de emoción y causa
  • Cohen's kappa

Data used

  • EmoBench EU: 200 escenarios ingleses; 49 CE, 56 PBE, 67 PT y 28 EC
  • Memoria PTEI: 500 escenarios sintéticos con emoción, causa y personalidad inferida; no publicada
  • Muestra humana ciega: 10% del conjunto filtrado; tamaño y anotadores no informados

Evidence and location

  • Method, tables, figures, prompts, limitations and ethics statement: arXiv:2607.10245v1 PDF, 15 pages; every page rendered and visually inspected
  • Date, version, authors, categories, license and absence of direct links to artifacts: Official arXiv record, export API, HTML and TeX source bundle for 2607.10245v1
  • Submission record, not equivalent to archival acceptance: Official OpenReview forum rIh6fjH7sX checked 2026-07-16
  • Size, distribution, joint scoring and five-repetition protocol of EmoBench: Sahandfer/EmoBench commit 3b4f84312541f8469e909965e1fff3691ef85c62 and ACL 2024 paper; all 19 paper pages visually inspected
  • Absence of locatable PTEI implementation and data: Exact-title and code searches on GitHub and Hugging Face checked 2026-07-16
  • Consolidated audit of validity, results, claims and reproducibility: reports/verification/article-276-ptei-personality-emotional-intelligence-validity-reproducibility-and-claims-audit.json