Benchmarking Personality Inference in Large Language Models Using Real-World Conversations

Evaluation and psychometric validity2025DOIApproved editorial review

Authors: Jianfeng Zhu, Xinyu Li, Julina Maharjan, Karin G. Coifman, Ruoming Jin

Keywords: Large Language Models, Personality Inference, Benchmarking, Big Five Inventory, Zero-Shot Prompting, Chain-of-Thought Prompting

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

5
Authors
30
Findings
65
Limitations
16
Evidence

Editorial summary

English

The published article evaluates whether four LLMs can infer BFI-10 item scores and Big Five traits from semi-structured interviews. This review uses the definitive 25-page article and its official six-page supplement, retaining arXiv v1 only as a superseded source. The journal version materially expands the preprint: it adds GPT-5-Mini, adds Xinyu Li and Julina Maharjan to the author list, reorders the authors, and reports supplementary confidence intervals. Neither code nor participant data are available.

The dataset starts with 555 U.S. adults recruited across several studies of adjustment to adverse or transitional life events; 518 participants with complete and valid BFI-10 data enter the analyses. The conversations are not spontaneous everyday dialogue: they are scripted research interviews about daily activities, a personal challenge, emotion regulation, and recent positive and negative events. Text is lowercased, punctuation and English stopwords are removed, and repetitions or fillers are truncated by an unspecified procedure. This preprocessing removes some stylistic, function-word, negation, and conversational-dynamics signals that could matter for personality, and no raw-text ablation is reported.

GPT-4.1-Mini, GPT-5-Mini, Meta-LLaMA-3.3-70B-Instruct-Turbo, and DeepSeek-R1-Distill-70B predict ten BFI items with direct prompting and five traits with direct or chain-of-thought prompting. The reference is same-session BFI-10 self-report, not objective personality ground truth. In the 518 analyzed cases, alignment is weak: item correlations range from −0.18 to 0.27 and all trait correlations remain below 0.30. The best trait result is Conscientiousness for zero-shot GPT-4.1-Mini, r=0.25 (95% CI 0.167–0.329). Many intervals include zero, and some statistically non-zero effects are negative. MAE/RMSE often exceed one scale point on the 1–5 scale except in several Openness conditions. Models shift predictions toward moderate or high values and categorical agreement with self-report is minimal, with κ approximately −0.07 to 0.11.

Off-by-one rates from 0.57 to 1.00 do not contradict that failure. With only three ordered categories, a Moderate prediction is at most one category away from both Low and High; the metric penalizes only a jump between the extremes. Meta-LLaMA with CoT reaches 1.00 on three traits by concentrating predictions centrally while κ remains near zero. This review therefore does not present off-by-one as clinical or psychometric accuracy. There is also no mean-prediction baseline, human-reader baseline, simple statistical model, or demographic baseline; low MAE can result from regression to the center without valid individual inference.

Repeatability is tested only for GPT-4.1-Mini, on 50 participants and three runs, while input-segment order is randomized. It cannot support a claim that all four models are stable and it is not validity. The published version does not identify the exact within-model ICC specification; the supplement merely lists coefficients. Inter-model agreement excludes GPT-5-Mini and is low at item level (ICC 0.02–0.26); at trait level it is 0.54–0.76 under zero-shot and falls to 0.11–0.34 under CoT. The length ablation compares the first 100, first 1,000, and full transcript, confounding length with question order and content; methods call the units words while results call them tokens. Medium context produces the highest observed correlations, while full context increases several RMSE values.

The study usefully documents a limitation of current LLMs, but it does not isolate its cause or validate a personality-assessment system. BFI-10 has only two items per trait; the paper does not report reliability in this sample, the exact aggregation/reverse-scoring algorithm, multiplicity correction, or direct statistical comparisons between models or prompts. Cleaning, exclusions, model/API versions, dates, output parsing, and the length χ² test are insufficiently specified for reproduction. CoT rationales are shown as examples but their faithfulness is not evaluated, so they are not validated interpretability evidence. The defensible conclusion is narrow: under these interviews, prompts, models, and brief self-report reference, predictions show limited repeatability evidence but weak individual alignment and poor categorical agreement. The paper does not establish reliable personality inference, superiority to baselines, generalization to everyday conversations, or fitness for decisions about people.

Español

El artículo publicado evalúa si cuatro LLM pueden inferir puntuaciones BFI-10 y rasgos Big Five a partir de entrevistas semiestructuradas. La revisión usa el artículo definitivo de 25 páginas, su suplemento oficial de 6 páginas y conserva la prepublicación arXiv v1 solo como antecedente. La versión de revista amplía el estudio original: añade GPT-5-Mini, incorpora a Xinyu Li y Julina Maharjan a la autoría, reorganiza los autores y publica resultados suplementarios con intervalos de confianza. No hay código ni datos de participantes disponibles.

El conjunto parte de 555 adultos de Estados Unidos reclutados en varios estudios sobre adaptación a acontecimientos adversos o de transición; 518 con BFI-10 completo y válido entran en los análisis. Las conversaciones no son diálogos cotidianos espontáneos: son entrevistas de investigación guionizadas sobre actividades diarias, una dificultad personal, regulación emocional y acontecimientos positivos y negativos recientes. El texto se pasa a minúsculas, se elimina puntuación y stopwords en inglés, y se truncan repeticiones o muletillas mediante un procedimiento no especificado. Ese preprocesamiento elimina precisamente parte de las señales de estilo, función gramatical, negación y dinámica conversacional que podrían ser pertinentes para personalidad, y no se compara con el texto sin procesar.

GPT-4.1-Mini, GPT-5-Mini, Meta-LLaMA-3.3-70B-Instruct-Turbo y DeepSeek-R1-Distill-70B predicen diez ítems BFI con prompt directo y cinco rasgos mediante prompt directo o chain-of-thought. El referente es el autoinforme BFI-10 de la misma sesión, no una observación objetiva de la personalidad. En los 518 casos analizados, la relación con el referente es débil: las correlaciones por ítem van de −0,18 a 0,27 y las de rasgo permanecen por debajo de 0,30. El mejor resultado de rasgo es Conscientiousness con GPT-4.1-Mini en zero-shot, r=0,25 (IC95% 0,167–0,329). Muchos intervalos incluyen cero, y algunos efectos significativos son negativos. Los errores MAE/RMSE suelen superar un punto en la escala 1–5 salvo varias condiciones de Openness. Los modelos desplazan las predicciones hacia valores moderados o altos y casi no concuerdan por categorías con el autoinforme: κ va aproximadamente de −0,07 a 0,11.

Las tasas “off-by-one” de 0,57 a 1,00 no contradicen ese fracaso. Con solo tres categorías ordenadas, una predicción Moderate está a como máximo una categoría tanto de Low como de High; la métrica solo penaliza el salto entre extremos. Meta-LLaMA con CoT logra 1,00 en tres rasgos al concentrar predicciones en el centro, mientras κ sigue cerca de cero. Por eso el resumen no interpreta off-by-one como precisión clínica o psicométrica. Tampoco hay comparación con un predictor de la media, humanos que lean las mismas entrevistas, modelos estadísticos sencillos o baselines demográficos; MAE bajo puede reflejar regresión al centro sin inferencia individual válida.

La repetibilidad se estudia únicamente con GPT-4.1-Mini, 50 participantes y tres ejecuciones, y además se aleatoriza el orden de segmentos: no justifica decir que los cuatro modelos tienen alta estabilidad ni equivale a validez. El tipo exacto de ICC de esta prueba no se declara en la versión publicada y el suplemento solo lista los coeficientes. La concordancia entre modelos excluye GPT-5-Mini y es baja por ítem (ICC 0,02–0,26); a nivel de rasgo es 0,54–0,76 en zero-shot y cae a 0,11–0,34 con CoT. El análisis de longitud usa los primeros 100, los primeros 1.000 y el texto completo, por lo que confunde longitud con posición y contenido de las preguntas; además, el método dice “palabras” y resultados dice “tokens”. El contexto medio maximiza las correlaciones observadas, mientras el contexto largo aumenta varios RMSE.

El trabajo documenta con utilidad una limitación real de los LLM actuales, pero no aísla su causa ni valida un sistema de evaluación de personalidad. El BFI-10 tiene solo dos ítems por rasgo; no se informa su fiabilidad en esta muestra, el algoritmo de agregación/reverse scoring, la corrección por comparaciones múltiples ni tests directos entre modelos o prompts. La limpieza, exclusiones, modelos/API, fechas, parsing de respuestas y prueba χ² de longitud no están suficientemente especificados para reproducir los resultados. Las explicaciones CoT se muestran como ejemplos pero no se evalúa su fidelidad, por lo que no son evidencia interpretable. La conclusión defendible es estrecha: bajo estas entrevistas, prompts, modelos y autoinforme breve, las predicciones son repetibles en una prueba limitada pero tienen poca alineación individual y mala concordancia categórica. No demuestra inferencia fiable de personalidad, superioridad sobre baselines, generalización a conversaciones cotidianas ni aptitud para decisiones sobre personas.

Research question

To what extent can four contemporary LLMs produce BFI-10 scores and Big Five traits aligned with self-reports from semistructured interviews, and how do their results change with direct prompting or CoT, repetitions, model, and input length?

Method

Retrospective observational benchmark. Transcripts of scripted interviews from several studies are preprocessed and 518 of 555 participants with complete BFI-10 are retained. Four LLMs predict BFI-10 items via zero-shot and Big Five traits via zero-shot and CoT. Pearson with 95% CI, Spearman, MAE, RMSE, exact match, off-by-one, and kappa are computed. GPT-4.1-Mini is repeated three times on 50 participants with randomized segment order. Intermodel agreement uses only GPT-4.1-Mini, Meta-LLaMA, and DeepSeek. An ablation with GPT-4.1-Mini compares the start of 100, 1,000, and all words/transcript.

Sample: The initial base includes 555 adults from the United States from several behavioral studies. 275 men, 278 women, and two cases without sex are reported; mean age 39.4 years (SD 16.33); 431 White participants and 122 non-White/Other, leaving two racial values unclarified. Only 518 with complete and valid BFI-10 are analyzed; the pattern and exact criterion of the 37 exclusions are not described. The sample comes from occupational stress, health, and life transition contexts, not from a representative population sample.

Findings

  • The definitive source is the open article published on 31 December 2025, not the arXiv preprint from July.
  • The journal version adds GPT-5-Mini, two authors, and a supplement with confidence intervals.
  • The benchmark is presented with 555 interviews, but all analytical results use 518 participants.
  • The interviews are semistructured, scripted, and focused on recent experiences; they are not spontaneous everyday conversations.
  • The referent is BFI-10 self-report from the same session and should not be called objective ground truth or true profile.
  • Item-level correlations span approximately -0.18 to 0.27.
  • GPT-4.1-Mini obtains the maximum per item on BFI-3, r=0.272, with 95% CI 0.190-0.350.
  • The best result per trait is GPT-4.1-Mini zero-shot on Conscientiousness, r=0.250, 95% CI 0.167-0.329.
  • No trait correlation reaches 0.30 and many conditions have intervals that include zero.
  • There are statistically significant negative correlations distinct from zero, for example Meta-LLaMA zero-shot on Extraversion and GPT-5-Mini zero-shot on Extraversion.
  • MAE and RMSE usually exceed one point on the 1-5 scale; Openness presents smaller errors due to its distribution and centrality.
  • Models underrepresent low values and shift many predictions toward moderate or high categories.
  • Categorical agreement is almost null: kappa sits approximately between -0.07 and 0.11.
  • Off-by-one reaches 1.00 on three traits for Meta-LLaMA CoT without kappa ceasing to be almost zero.
  • On a three-category scale, off-by-one only penalizes confusing Low with High and automatically rewards any Moderate against any referent.
  • The published repeatability corresponds only to GPT-4.1-Mini, 50 participants, and three executions.
  • The repeatability coefficients of GPT-4.1-Mini range from 0.61 to 0.93 per item and from 0.81 to 1.00 per trait.
  • The segment order changes between repetitions, so the test mixes stochasticity with order sensitivity.
  • Intermodel agreement per item is low, ICC(2,1)=0.02-0.26.
  • Intermodel agreement per trait is 0.54-0.76 in zero-shot and 0.11-0.34 in CoT.
  • GPT-5-Mini is excluded from the intermodel ICCs although it is part of the four-model benchmark.
  • The medium context obtains the highest correlations of the ablation: r=0.219 in Agreeableness and 0.214 in Conscientiousness.
  • Full context increases RMSE for Agreeableness, Conscientiousness, and Neuroticism compared with short context.
  • The ablation shows no monotonic improvement with more context nor a consistent advantage of full text.
  • CoT does not systematically improve correlation or error and markedly reduces agreement between models.
  • The supplement publishes a single CoT example; its explanations are not scored against human evidence nor subjected to fidelity tests.
  • Figure 2 is titled comparison of three LLMs although it contains four model rows.
  • The text states that kappa of the ablation remains below 0.01, but the same section reports up to 0.088; it probably meant 0.1.
  • Methods defines lengths in words and Results redefines them as tokens.
  • The central empirical conclusion is supported: individual alignment with the BFI-10 is weak under the studied conditions.

Limitations

  • The sample pools several studies of adversity and transition without detailing sizes, criteria, and protocols per cohort.
  • Representativeness of the United States population or of ordinary-use conversations is not demonstrated.
  • The real-world label may lead one to think of spontaneous natural dialogue, but the material comes from standardized research interviews.
  • Of 555 participants, 37 are excluded for incomplete or invalid BFI-10 without flow, reasons, or comparison of excluded cases.
  • The racial counts sum to 553 and the two remaining values are not explained.
  • No breakdowns by study, age, sex, race, education, or other variables are offered to evaluate heterogeneity or fairness.
  • The private data contain sensitive narratives; there is no privacy audit of prompts, provider, retention, or data transfer.
  • The original consent for behavioral studies does not clarify specific authorization for personality inference via third-party APIs.
  • Ethical review was exempted for the secondary analysis, but identifiers of the original studies or approvals are not provided.
  • The BFI-10 uses only two items per trait and may limit the reliability and coverage of the construct.
  • The reliability of the BFI-10 in this sample and the measurement error of the referent are not reported.
  • It is not specified how BFI-10 items are inverted and aggregated to produce the five reference traits.
  • Self-report from a single session is not an objective trait truth and may contain response bias, transient state, and error.
  • There are no observer reports, behavioral outcomes, clinical measures, or external criteria for triangulation.
  • The instruction asks to score only signals present in state narratives and then compares them with stable traits, creating a target mismatch.
  • The maximum attenuation expected from imperfect reliability of the referent and limited signal availability is not estimated.
  • Capitalization, punctuation, and stopwords are removed, which may contain style, pronouns, negation, and other relevant signals.
  • NLTK includes stopword lists whose treatment of negations must be specified; the version and the exact list are not published.
  • The truncation of repeated words and filler words has no definition, threshold, code, or audit.
  • There is no ablation of raw text versus preprocessed text.
  • Audio, prosody, pauses, tone, interviewer interaction, or conversational dynamics are not analyzed.
  • The models and endpoints are not fixed with versioned identifiers or reproducible snapshots.
  • Execution dates, region, provider of Meta-LLaMA/DeepSeek, seed, and update policies are not reported.
  • GPT-5-Mini uses default decoding, so it does not share a controlled configuration with the other models.
  • Top-p/top-k, output limit, system prompt, retries, rate limits, or API failures are not reported.
  • The score parser, the rate of invalid formats, repairs, or exclusions of outputs are not described.
  • No code, complete BFI-10 prompts, outputs, derived data tables, or analysis environment are published.
  • Privacy may justify not releasing transcripts, but it does not prevent releasing code, deidentified outputs, or reproducible aggregates.
  • There is no baseline of mean, median, category frequency, linear regression, embeddings, or classic NLP method.
  • There is no human baseline with readers exposed to exactly the same preprocessed text.
  • Without a central baseline, MAE and exact match do not demonstrate that the model uses individual information.
  • Dependent correlations between models, prompts, or lengths are not formally compared.
  • Dozens of correlations are evaluated without declaring multiplicity correction.
  • The study emphasizes significance of some small correlations without separating statistical significance from individual utility.
  • There are no intervals for MAE, RMSE, exact match, off-by-one, or kappa.
  • Kappa is not weighted even though the categories are ordinal, and it is not accompanied by prevalence analysis.
  • The binning thresholds of integer self-report and continuous prediction are different and are not psychometrically justified.
  • Off-by-one is almost vacuous with three categories and can reach 100% through central predictions.
  • Calibration by trait level or error conditioned on extremes is not evaluated.
  • Histograms and aggregate MAE may hide distinct performance by subgroup or cohort.
  • Stability is studied only on one model, 50 of 518 cases, and three executions.
  • The exact type of ICC, single versus average measure, is not identified in the definitive article or supplement.
  • Calling output repeatability internal consistency confounds different psychometric properties.
  • Randomizing segments means the three executions do not receive exactly the same ordered stimulus.
  • Intervals of the repeatability ICCs or sensitivity to seed/temperature are not provided.
  • Intermodel agreement excludes GPT-5-Mini without justification.
  • Treating three models as a random sample of raters for ICC(2,1) is a strong generalization with very few selected raters.
  • The p-values of some ICCs do not intuitively agree with intervals that include zero and the procedure is not explained.
  • The ablation takes prefixes, so length, question order, and content type change at the same time.
  • The article alternates words and tokens for 100/1,000 without resolving which was the real unit.
  • Fragments, questions, or windows are not randomized to isolate amount of context.
  • The chi-square test of length does not identify test, degrees of freedom, repeated structure, post-hoc, or effect size.
  • Testing that the distribution of predictions changes does not prove that accuracy changes.
  • The multiple length tests for five traits are not corrected.
  • Visible CoT does not constitute interpretability by itself and may be a non-faithful rationalization.
  • The CoT example contains unvalidated causal and trait inferences; its reasons are not systematically evaluated.
  • No factorial, convergent, discriminant, incremental, or predictive analysis is performed to validate the construct.
  • There is no preregistration, exploratory/confirmatory separation, or independent development set for prompts and analytical decisions.
  • Sensitivity to prompt formulations, language, transcription, or question order is not studied.
  • Only English and a United States sample are evaluated.
  • Harms, operational consent, explainability for affected parties, or misuse in employment, education, insurance, or health are not evaluated.
  • The published version retains editorial errors: Figure 2 says three models although it shows four; the kappa limit <0.01 contradicts 0.088; Table Y remains unresolved in the supplement.
  • The claim of ecological validity is broader than the design: several research cohorts do not equate to deployment contexts.
  • There is no external validation in another sample, another interview protocol, free conversation, or temporal period.
  • Drift of proprietary models or reproducibility after API updates is not evaluated.

What the study does not establish

  • It does not establish that LLMs reliably infer a person's personality.
  • It does not establish that self-reported BFI-10 is objective ground truth.
  • It does not demonstrate clinical, diagnostic, occupational, educational, or forensic validity.
  • It does not demonstrate superiority over a mean predictor or any simple baseline.
  • It does not demonstrate superiority over human raters who read the same information.
  • It does not demonstrate that high off-by-one represents useful accuracy.
  • It does not demonstrate high stability of the four models; it only repeats GPT-4.1-Mini on 50 cases.
  • It does not demonstrate that CoT improves accuracy or produces faithful and interpretable reasoning.
  • It does not demonstrate that more context monotonically improves inference.
  • It does not isolate input length from position, question, and content.
  • It does not demonstrate generalization from scripted interviews to spontaneous conversations.
  • It does not demonstrate generalization outside of United States adults or adverse/transitional contexts.
  • It does not demonstrate fairness across sexes, ages, racial groups, education, or cohorts.
  • It does not allow reproducing the benchmark without data, code, outputs, and versioned models.
  • It does not establish that CoT explanations are causal evidence of how the model obtained the score.
  • It does not justify using these predictions to make decisions about people.

Traceability

Scope: Full text

Version: Journal of Psychiatry and Brain Science 2025;10(6):e250020; DOI 10.20900/jpbs.20250020; published 31 December 2025; 25-page article plus 6-page official supplement; arXiv:2507.14355v1 retained as superseded source

Consulted source: https://jpbs.hapres.com/UpLoad/PdfFile/JPBS_1832.pdf

Review: Codex definitive-version reconciliation, complete bilingual full-text fidelity pass, all-page visual inspection of article, supplement and superseded preprint, supplementary-table audit, construct-validity and data-quality review; summaries written from the full evidence rather than abstract keywords, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4.1-Mini via OpenAI API; temperature 0.2
  • GPT-5-Mini via OpenAI API; provider-default decoding because temperature was not configurable
  • Meta-LLaMA-3.3-70B-Instruct-Turbo; provider and exact endpoint not reported
  • DeepSeek-R1-Distill-70B; provider and exact endpoint not reported

Instruments and metrics

  • BFI-10 same-session self-report reference on a 1–5 Likert scale
  • Direct zero-shot BFI-10 item prediction
  • Direct zero-shot Big Five trait prediction
  • Chain-of-thought Big Five trait prediction with visible rationales
  • Pearson correlation with 95% confidence interval and Spearman rank correlation
  • Mean absolute error and root mean squared error
  • Three-bin exact match, off-by-one rate and unweighted Cohen kappa
  • Within-model ICC coefficients across three GPT-4.1-Mini runs; exact ICC form not reported in the definitive version
  • Inter-model ICC(2,1) across three models, excluding GPT-5-Mini
  • Input-length ablation on the first 100, first 1,000 and full transcript

Data used

  • Private pooled dataset of 555 U.S. adult semi-structured research interviews about adjustment to adverse or transitional life events
  • Analysis subset of 518 participants with complete and valid BFI-10 scores
  • BFI-10 self-report collected in the same session as the interview
  • Random 50-participant subset used only for three-run GPT-4.1-Mini repeatability analysis
  • No public participant-level dataset, model outputs, code, or reproducible analysis package

Evidence and location

  • Definitive publication, DOI, dates, license, and authorship: Publisher article front matter, page 1, and official detail page
  • Sample, interviews, demographics, preprocessing, and 518 cases analyzed: Definitive article Materials and Methods, page 4; Supplement Tables S1-S2
  • Models and decoding configuration: Definitive article LLMs and Prompting Strategies, page 5
  • Direct prompts and CoT: Definitive article pages 5-6; Supplement Boxes S1-S2, pages 5-6
  • Metrics and bins: Definitive article Evaluation Metrics, page 6
  • Item-level correlations with 95% CI, MAE, and RMSE: Supplement Table S3, pages 1-2
  • Trait-level correlations, prompting, and models: Supplement Table S4, pages 2-3
  • Distributions, exact match, off-by-one, and kappa: Definitive article Figures 3-4 and 6-8, pages 9-14
  • Repeatability limited to GPT-4.1-Mini, 50 participants, and three runs: Definitive article page 6 and pages 14-15; Supplement Tables S5-S6, page 3
  • Agreement among three models and exclusion of GPT-5-Mini: Supplement Tables S7-S9, pages 3-4
  • Length ablation and words/tokens contradiction: Definitive article pages 6-7 and 15-17; Supplement Figures S1-S2, pages 4-5
  • Unvalidated CoT reasoning example: Supplement Box 3, page 6
  • Limitations acknowledged by the authors: Definitive article Limitations, page 20
  • Data not available due to privacy restrictions: Definitive article Data Availability, page 21
  • Editorial errors and internal contradictions: Definitive article Figure 2 caption page 8; length units pages 6 and 15; kappa text page 16; Supplement Table S9 note page 4
  • Visual inspection: All 25 definitive article pages, all 6 supplement pages, and all 21 superseded arXiv v1 pages rendered and visually inspected on 15 July 2026