Evaluating LLM Alignment on Personality Inference from Real-World Interview Data

Evaluation and psychometric validity2026arXivApproved editorial review

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

Keywords: Computation and Language, Large Language Models, Personality Inference, Big Five Personality Traits, Machine Learning

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

5
Authors
7
Findings
12
Limitations
8
Evidence

Editorial summary

English

This preprint evaluates whether several language-model-based methods can predict continuous Big Five scores from semi-structured interviews. The manuscript reports 518 adults, approximately 15 minutes of speech per participant, and BFI-10 self-reports as targets. Its sample size conflicts with its own demographics: 275 male + 278 female and 431 White + 122 non-White/other both total 553. It reports no trait means, standard deviations, ranges, two-item-scale reliability, missingness, or explanation of how 553 becomes 518 if both counts are meaningful. The BFI-10 is a validated human measure, but a brief self-report score is not observable personality “ground truth,” and the paper does not establish that a fifteen-minute interview contains enough signal to recover each trait. Four configurations are compared. GPT-4.1 Mini receives the transcript and simultaneously returns five 1–5 scores in half-point steps through either zero-shot prompting or a prompt asking it to summarize tone, identify cues, and justify scores. The paper calls temperature .2 deterministic, although .2 still samples and no repeated runs are reported. A second family adapts Llama-3.1-8B-Instruct separately for each trait using LoRA ranks 8, 16, and 32. For RoBERTa-base, overlapping transcript chunks are encoded, mean-pooled, and passed to a regression head; however, the method, Table 4, and Figure 3 alternate among RoBERTa, BERT, “embedding+LoRA,” and “LoRA-based” without clearly specifying which parameters are adapted. Finally, all-MiniLM-L6-v2 and text-embedding-3-small embeddings feed a multi-output Ridge regression with an 80/20 split. These are said to embed full transcripts without explaining truncation or aggregation despite MiniLM's context limit. The paper does not report LoRA/RoBERTa splits, seeds, stratification, validation selection, regularization, epochs, batch size, confidence intervals, or statistical comparisons. The strongest finding is negative: no Pearson correlation exceeds .255. For GPT-4.1 Mini zero-shot, Conscientiousness reaches r=.250 and Agreeableness .132; Neuroticism, Openness, and Extraversion are .065, .020, and .041. Chain-of-thought does not improve the set: .236, .133, −.009, .051, and .005, respectively. In the encoder/“BERT” table, the best values by trait are .197 for Extraversion, .173 for Agreeableness, .219 for Conscientiousness, .236 for Neuroticism, and .255 for Openness. Llama-LoRA peaks at .164 for Conscientiousness and produces several negative correlations. Static embedding regression also remains at or below .119: MiniLM ranges from −.069 to .112 and OpenAI from −.050 to .119. MAE tells a different story because some models predict near the scale center. GPT-4 has MAE below .6 only for Openness and above 1.2 for Conscientiousness and Agreeableness. OpenAI embeddings achieve .776 for Openness and .916 for Agreeableness, but 1.080 for Conscientiousness and above 1.2 for Extraversion and Neuroticism. In Figure 3, low values such as .431 on Openness coexist with null or negative r, which is compatible with regression to the mean rather than valid ranking of individuals. The text also attributes .431 at rank 8 and .493 at rank 16 to Llama and embedding+LoRA in ways inconsistent with the plotted curves, preventing a clean architecture comparison. No mean-prediction baseline is provided, although it is essential for deciding whether central MAE adds value, and no uncertainty is shown for correlations from a test set of roughly one hundred cases. Trait-level stories, such as attributing Conscientiousness to RLHF, Openness to imagination, or poor Extraversion to absent interaction, are speculative rather than tested. Overall, the paper offers a useful comparison and a defensible main conclusion: these methods recover BFI-10 scores poorly from the interviews. It does not show that the failure reflects a general lack of LLM psychological understanding or alignment rather than weak textual signal, BFI-10 noise, small samples, incomplete protocols, or calibration problems.

Español

Este preprint evalúa si varios métodos basados en modelos de lenguaje pueden predecir las puntuaciones continuas del Big Five a partir de entrevistas semiestructuradas. El manuscrito declara 518 adultos, unos 15 minutos de habla por participante y autoinforme BFI-10 como objetivo. La cifra de muestra no cuadra con sus propios datos demográficos: 275 hombres + 278 mujeres y 431 personas blancas + 122 no blancas/otras suman 553 en ambos casos. Tampoco informa medias, desviaciones, rango, fiabilidad de las escalas de dos ítems, datos faltantes ni cómo pasa de 553 a 518 si ambas cifras fueran correctas. El BFI-10 es una medida humana validada, pero una puntuación breve autoinformada no es «ground truth» observable de la personalidad y el paper no demuestra que el contenido de una entrevista de quince minutos contenga información suficiente para recuperar cada rasgo. Se comparan cuatro configuraciones. GPT-4.1 Mini recibe la transcripción completa y devuelve simultáneamente cinco valores entre 1 y 5, en pasos de 0,5, mediante zero-shot o un prompt que pide resumir tono, identificar indicios y justificar la puntuación. El texto llama determinista a temperatura 0,2, aunque 0,2 permite muestreo y no se reportan repeticiones. La segunda familia adapta Llama-3.1-8B-Instruct por separado para cada rasgo mediante LoRA de rango 8, 16 o 32. Para RoBERTa-base se trocean las transcripciones con solapamiento, se promedian representaciones y se entrena una cabeza de regresión; sin embargo, método, Tabla 4 y Figura 3 alternan los nombres RoBERTa, BERT, «embedding+LoRA» y «LoRA-based» sin especificar con precisión qué parámetros se adaptan. Finalmente, embeddings all-MiniLM-L6-v2 y text-embedding-3-small alimentan una regresión Ridge multi-output con split 80/20. Para estos dos se dice que se incrusta la transcripción completa, sin aclarar truncamiento o agregación pese al límite de contexto de MiniLM. No se publican particiones para LoRA/RoBERTa, semillas, estratificación, conjunto de validación, selección de hiperparámetros, regularización, epochs, batch size, intervalos o pruebas entre modelos. El hallazgo más sólido es negativo: ninguna correlación de Pearson supera 0,255. En GPT-4.1 Mini zero-shot, responsabilidad alcanza r=0,250 y amabilidad 0,132; neuroticismo, apertura y extraversión quedan en 0,065, 0,020 y 0,041. CoT no mejora el conjunto: 0,236, 0,133, −0,009, 0,051 y 0,005, respectivamente. En la tabla de la variante encoder/«BERT», los mejores valores por rasgo son 0,197 en extraversión, 0,173 en amabilidad, 0,219 en responsabilidad, 0,236 en neuroticismo y 0,255 en apertura. Llama-LoRA llega como máximo a 0,164 en responsabilidad y arroja varias correlaciones negativas. La regresión con embeddings estáticos tampoco supera 0,119: MiniLM va de −0,069 a 0,112 y OpenAI de −0,050 a 0,119. Las MAE cuentan una historia distinta porque algunos modelos predicen cerca del centro de la escala. GPT-4 tiene MAE inferior a 0,6 solo en apertura y superior a 1,2 en responsabilidad y amabilidad. OpenAI embeddings logra 0,776 en apertura y 0,916 en amabilidad, pero 1,080 en responsabilidad y más de 1,2 en extraversión y neuroticismo. En la Figura 3, valores bajos como 0,431 en apertura coexisten con r nulo o negativo; eso es compatible con regresión a la media y no con ordenación válida de personas. El propio texto atribuye 0,431 en rango 8 y 0,493 en rango 16 tanto al grupo Llama como al grupo embedding+LoRA de formas incompatibles con las curvas, de modo que esas cifras no permiten una comparación fiable de arquitectura. El artículo tampoco incluye un baseline de predecir la media, necesario para juzgar si una MAE central aporta valor, ni incertidumbre para las correlaciones de un test de aproximadamente cien casos. Su interpretación de diferencias entre rasgos, por ejemplo atribuir responsabilidad a RLHF, apertura a imaginación o extraversión a falta de interacción, es especulativa y no se contrasta. En conjunto, el trabajo ofrece una comparación útil y prudente en su conclusión principal: estos métodos no recuperan bien las puntuaciones BFI-10 desde las entrevistas. No demuestra, sin embargo, que el fallo sea una carencia general de «comprensión psicológica» o alineamiento del LLM en lugar de baja señal del texto, ruido del BFI-10, muestra pequeña, protocolo incompleto o mala calibración.

Research question

To what extent can GPT-4.1 Mini, models adapted with LoRA, and regressors over embeddings recover continuous BFI-10 scores from real semi-structured interviews?

Method

Pearson correlation and MAE per trait are compared across four configurations: GPT-4.1 Mini zero-shot and with CoT justification; Llama-3.1-8B-Instruct adapted per trait with LoRA of ranks 8/16/32; a RoBERTa/BERT encoder with chunks, mean pooling, and a regression head presented as embedding+LoRA; and multi-output Ridge regression over MiniLM or OpenAI embeddings with an 80/20 split. The targets are BFI-10 self-reports of interviewed participants; the protocol does not sufficiently document the partitions and the training of all variants.

Sample: The text declares 518 adults and approximately 15 minutes of interview per person. The published counts by sex and race sum to 553, without reconciliation. The indicated age is M=39.39, SD=16.33. No BFI-10 distributions, exclusions, missingness, or exact train/validation/test sizes for each paradigm are published. Only Ridge regression specifies 80/20, which would imply about 104 test cases if n=518. There is no external validation sample.

Findings

  • All correlations are below 0.26; the maximum is openness r=0.255 in the rank 32 encoder variant.
  • GPT-4.1 Mini zero-shot reaches r=0.250 in conscientiousness and 0.132 in agreeableness, but only 0.020–0.065 in the other three traits.
  • The CoT prompt does not improve globally: it reduces conscientiousness, neuroticism, and extraversion, barely raises openness, and leaves agreeableness practically the same.
  • The encoder variant reports better correlations than Llama-LoRA in most cells; Llama reaches at most 0.164 and has several negative values.
  • MiniLM and text-embedding-3-small with Ridge do not exceed r=0.119 and show different signs for agreeableness and neuroticism.
  • Some openness MAEs are low despite correlations near zero or negative, a pattern compatible with central predictions that do not correctly rank individuals.
  • The narrative and Figure 3 do not agree in assigning MAE of 0.431 and 0.493 to the Llama or embedding+LoRA variants.

Limitations

  • The declared size n=518 contradicts two independent demographic breakdowns that sum to 553; exclusions and missing data are not explained.
  • The BFI-10 has only two items per trait; the study does not publish reliability, distributions, or measurement error and treats it excessively as ground truth.
  • There is no mean baseline, linear regression with simple linguistic features, or reliability ceiling to contextualize MAE and r.
  • No intervals, p-values, bootstraps, or paired comparisons are published; with a test set near one hundred cases, small differences between r are unstable.
  • The protocol omits LoRA/RoBERTa partitions, validation, seeds, epochs, batch size, optimizer, learning rate, regularization, and selection criterion.
  • The BERT/RoBERTa/embedding+LoRA nomenclature is inconsistent and the attribution of several MAEs contradicts the figure.
  • Temperature 0.2 is called deterministic without repeating runs; sensitivity to sampling or prompt is not measured.
  • MiniLM reportedly receives full transcripts without explanation of truncation, chunking, or effective length.
  • The single 80/20 split is not repeated or stratified and there is no external, cross-cohort, or demographic generalization evaluation.
  • The paper does not detail recruitment, consent/IRB, distribution by interview, handling of interviewer interventions, or transcription.
  • Qualitative thresholds for r and MAE are presented without specific justification for BFI-10 or for high-risk decisions.
  • Psychological and architectural explanations of trait differences are post hoc and are not subjected to tests.

What the study does not establish

  • It does not demonstrate that GPT-4.1, Llama, RoBERTa, or embeddings infer personality with individual validity.
  • It does not demonstrate that a low isolated MAE corresponds to good prediction; it may arise from approaching the scale mean.
  • It does not demonstrate that CoT is unnecessary for all personality inference; it tests a single template and one model.
  • It does not allow attributing low performance to lack of alignment or psychological understanding rather than low signal, noisy labels, or insufficient design.
  • It does not validate mental health screening, counseling, marketing, or decisions about people and warns against high-risk uses without supervision.
  • It does not establish generalization to other interviews, languages, populations, modalities, long instruments, or behavioral outcomes.

Traceability

Scope: Full text

Version: arXiv:2509.13244v1 (16 September 2025)

Consulted source: https://arxiv.org/pdf/2509.13244

Review: Codex editorial review, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4.1 Mini (gpt-4.1-mini-2025-04-14)
  • Meta-Llama-3.1-8B-Instruct with trait-specific LoRA
  • RoBERTa-base / BERT encoder reported inconsistently
  • all-MiniLM-L6-v2
  • OpenAI text-embedding-3-small (2024-04-15)
  • Multi-output Ridge regression

Instruments and metrics

  • 10-item Big Five Inventory (BFI-10)
  • Semi-structured interview protocol
  • Zero-shot 1–5 Big Five prompt
  • Chain-of-thought-style cue and justification prompt
  • Pearson correlation
  • Mean Absolute Error
  • LoRA ranks 8, 16 and 32
  • Overlapping transcript chunks and mean pooling

Data used

  • Real-world semi-structured interview transcripts described as 518 participants
  • BFI-10 self-report scores paired to transcripts
  • Demographic table counts totaling 553 participants

Evidence and location

  • Objective, paradigms, and global result below 0.26: arXiv v1, pp. 1–2, Abstract and Introduction
  • Sample, interview, BFI-10, and demographic contradiction: arXiv v1, p. 4, Ecologically Valid Interview Dataset and Table 2
  • GPT-4.1 Mini prompt, temperature, and LoRA/encoder: arXiv v1, pp. 4–5, Experimental Design and Figure 1
  • Embeddings, 80/20 split, Pearson, and MAE: arXiv v1, p. 5, Embedding-based Regression and Evaluation Metrics
  • Zero-shot and CoT correlations and their MAE: arXiv v1, p. 5, Table 3 and Figure 2
  • LoRA/encoder correlations and naming inconsistency: arXiv v1, pp. 5–6, Table 4 and Figure 3
  • Embedding regression and calibration: arXiv v1, p. 6, Table 5 and Figure 4
  • Interpretation, declared limits, and high-risk uses: arXiv v1, pp. 6–8, Discussion, Limitations, Conclusion and Ethical Considerations