Stable and Explainable Personality Trait Evaluation in Large Language Models with Internal Activations

Evaluation and psychometric validity2026arXivApproved editorial review

Authors: Xiaoxu Ma, Xiangbo Zhang, Zhenyu Weng

Keywords: Large Language Models, Personality, Persona, Personality Control, Model Evaluation

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

3
Authors
6
Findings
30
Limitations
11
Evidence

Editorial summary

English

PVNI proposes converting internal activations into five OCEAN scores intended to be prompt-neutral. For each trait, it generates responses under positive, negative, and neutral instructions; averages hidden states at a fixed layer and probe position; defines the axis as positive minus negative activation; projects the neutral vector onto that axis; clips the coefficient to [0,1]; and uses it to interpolate between two extreme scores assigned by GPT-4.1-mini. GPT-5.2 generates contrastive prompts and questions. The geometric intuition is clear: locate a “neutral” response between two explicitly constructed endpoints.

The experiment compares PVNI with IPIP-BFFM-50, IPIP-NEO-120, and API-judged open-ended evaluation on Qwen-2.5-7B-Instruct, Llama-3-8B-Instruct, and Mistral-7B-v0.1-Instruct. Each protocol uses ten prompt sets in two families: question rewrites and role-play variants. The favorable result is consistent: PVNI has the lowest standard deviation in all 30 model-by-trait-by-variant-family combinations. For Qwen question rewrites, for example, PVNI reports O 83.55±.82, C 87.63±.73, E 42.89±2.49, A 93.39±.68, and N 36.45±.83, whereas IPIP-BFFM standard deviations range from 5.9 to 18.1. On Mistral, PVNI's least stable case is extraversion at 60.15±5.89 versus 36.98±24.86 for open-ended evaluation. The evidence supports lower sensitivity to this controlled set of prompt reformulations.

It does not show that PVNI scores are truer, psychometrically valid, or behaviorally predictive. There is no ground truth, human panel, behavioral evaluation, convergent or discriminant validity, factor structure, or external criterion. Method means disagree sharply: for Mistral extraversion, IPIP-BFFM gives 75, IPIP-NEO 70.83, open-ended evaluation 36.98, and PVNI 60.15. Without a criterion, the study cannot determine which describes the model better. A constant estimator would also have zero variance; stability can be necessary for measurement but is not sufficient evidence of validity.

Some stability is structural. PVNI averages prompts, outputs, and activations, projects onto one axis, clips coefficients outside [0,1], and interpolates between averaged extremes. It does not report clipping frequency or ablate clipping, prompt count, layers, or probe positions. Positive and negative prompts already define the trait the judge is asked to detect; the method measures where neutral falls between constructed endpoints. Calling this “internal personality” conflates a coordinate relative to a prompt ensemble with an intrinsic, task-independent disposition.

The formal theory does not empirically validate PVNI. It assumes local score linearity, a well-trained persona parameter update, approximately rank-one amplification, and, when discussing generalization, an orthonormal OCEAN basis. The experiment uses prompting rather than the persona-adapted models in Assumption 2 and tests neither rank-one structure, orthogonality, error bounds, composition, negation, nor out-of-domain synthesis. The limitations themselves acknowledge correlated axes and that neutrality may not lie on the positive-negative axis. The theorems therefore state consequences conditional on assumptions, not evidence that those assumptions hold.

“Explainable” means the computation can be described as projection and interpolation. The paper does not identify neurons, tokens, layers, or mechanisms that cause behavior, test explanation fidelity, or even report the actual layer and probe position used. There is no layer/token sensitivity, temporal stability, repeated decoding, long-context, dialogue, multilingual, scale, or larger-model evaluation. Requiring activations restricts PVNI to white-box models.

The preprint links no code, artifact checkpoints, complete five-trait prompts, extraction/evaluation splits, outputs, judge scores, or executable configuration. GPT-4.1-mini and GPT-5.2 snapshots, dates, temperatures, seeds, layer, probe position, and the reproducible procedure for a logit-weighted average over integer tokens 0–100 are absent. Tables and partial examples are insufficient to reconstruct the results. PVNI is a promising representation-robustness proposal, but it is not yet a validated LLM personality assessment or an operational benchmark for model comparison.

Español

PVNI propone convertir activaciones internas en cinco puntuaciones OCEAN supuestamente neutras al prompt. Para cada rasgo, genera respuestas bajo instrucciones positivas, negativas y neutrales; promedia el hidden state de una capa y posición fijas; define el eje como activación positiva menos negativa; proyecta el vector neutral sobre ese eje; recorta el coeficiente a [0,1] y lo usa para interpolar entre dos puntuaciones extremas asignadas por GPT-4.1-mini. GPT-5.2 genera los prompts contrastivos y las preguntas. La intuición geométrica es clara: ubicar la respuesta “neutral” entre dos extremos creados explícitamente.

El experimento compara PVNI con IPIP-BFFM-50, IPIP-NEO-120 y evaluación abierta juzgada por API en Qwen-2.5-7B-Instruct, Llama-3-8B-Instruct y Mistral-7B-v0.1-Instruct. Para cada protocolo usa diez conjuntos de prompts en dos familias: reescrituras de preguntas y variantes de role-play. El resultado favorable es consistente: la desviación estándar de PVNI es la menor en las 30 combinaciones modelo×rasgo×familia. Por ejemplo, en Qwen bajo reescritura, PVNI da O 83,55±0,82, C 87,63±0,73, E 42,89±2,49, A 93,39±0,68 y N 36,45±0,83, mientras IPIP-BFFM alcanza desviaciones entre 5,9 y 18,1. En Mistral, el caso más inestable de PVNI es extraversión, 60,15±5,89, frente a 36,98±24,86 con evaluación abierta. La evidencia sostiene que el pipeline es menos sensible a este conjunto controlado de reformulaciones.

No demuestra que las puntuaciones sean más verdaderas, psicométricamente válidas o conductualmente predictivas. No existe ground truth, panel humano, evaluación de comportamiento, validez convergente/discriminante, estructura factorial ni criterio externo. Los métodos discrepan ampliamente en medias: para extraversión de Mistral, IPIP-BFFM da 75, IPIP-NEO 70,83, evaluación abierta 36,98 y PVNI 60,15. Sin criterio no puede decidirse cuál describe mejor al modelo. Una salida constante también tendría varianza cero; estabilidad es una propiedad necesaria para algunas mediciones, no evidencia suficiente de validez.

Parte de la estabilidad es estructural. PVNI promedia preguntas, respuestas y activaciones, proyecta sobre un único eje, recorta coeficientes fuera de [0,1] e interpola entre promedios extremos. No se publica cuántos casos son recortados ni una ablación sin clipping, con menos prompts o con otras capas. Los prompts positivos y negativos ya definen el rasgo que el juez debe detectar; el método mide dónde cae el neutral entre esos endpoints construidos. Llamarlo “personalidad interna” confunde una coordenada relativa al conjunto de prompts con una disposición intrínseca e independiente de la tarea.

La teoría formal no valida empíricamente PVNI. Parte de supuestos fuertes: linealidad local del score, persona bien adaptada mediante un cambio de parámetros, actualización aproximadamente de rango uno y, para generalización, una base OCEAN ortonormal. El experimento usa prompting, no los modelos persona-adaptados del supuesto 2; no verifica rango uno, ortogonalidad, límites de error, composición, negación ni síntesis fuera de dominio. La propia sección de limitaciones admite que los ejes están correlacionados y que lo neutral puede no estar sobre el eje positivo-negativo. Por tanto, los teoremas expresan consecuencias condicionales de supuestos, no una demostración de que esos supuestos se cumplan.

“Explainable” significa que el cálculo puede describirse como proyección e interpolación. No se identifica qué neuronas, tokens, capas o mecanismos causan conducta, no se prueba fidelidad de la explicación y ni siquiera se especifican en el paper la capa l o posición de probe empleadas. No hay análisis de sensibilidad por capa/token, estabilidad temporal, decodificación repetida, contexto largo, diálogo, idiomas, tamaños o modelos mayores. El requisito de activaciones limita PVNI a modelos white-box.

El preprint no enlaza código, checkpoints de artefactos, prompts completos para los cinco rasgos, conjuntos train/evaluation, salidas, scores del juez o configuración ejecutable. Faltan snapshots de GPT-4.1-mini y GPT-5.2, fecha, temperatura, seeds, capa, posición y procedimiento reproducible para el promedio ponderado por logits de enteros 0–100. Con solo tablas y ejemplos parciales no se pueden reconstruir los resultados. PVNI es una propuesta prometedora para robustez de representaciones, pero todavía no es una evaluación validada de personalidad LLM ni un referente operativo listo para comparar modelos.

Research question

Can a projection of the neutral hidden state onto a positive-negative axis constructed with prompts produce OCEAN scores that are less variable under reformulations than questionnaires and open-ended evaluation?

Method

For each trait and model, PVNI generates responses with positive, negative, and neutral prompts, averages activations, projects neutral-negative onto positive-negative, and clips the coefficient to [0,1]. It interpolates between 0-100 anchors judged by GPT-4.1-mini. Standard deviation is compared over ten sets of rewrites or role-play with two IPIP and one open-ended evaluation on three 7B models. GPT-5.2 generates prompts and questions.

Sample: Three instruction-tuned models of approximately 7B, five traits, and two families of ten prompt sets per protocol. Four protocols are compared, producing 30 model×trait×family combinations for the claim of lower standard deviation. The paper reports about two GPU hours per model on RTX 4090, but no exact number of generations, judge rollouts, examples per split, tokens, seeds, or clipping frequency.

Findings

  • PVNI has the lowest standard deviation in all published combinations of model, trait, and variant type.
  • In Qwen with rewrites, PVNI standard deviations are .82, .73, 2.49, .68, and .83 for O, C, E, A, and N, lower than those of the three baselines.
  • In Mistral with rewrites, PVNI is very stable except for E ±5.89; even so, it remains below open-ended E ±24.86 and IPIP-NEO E ±14.91.
  • Role-play variants tend to have lower variability than rewriting questions, even in the baselines.
  • Means differ materially across protocols, so lower variance does not resolve which score is valid.
  • The article provides no inferential tests, intervals on variance differences, ground truth, or correlation with behavior.

Limitations

  • It optimizes and evaluates stability, not validity. A constant estimator would be stable and would not measure the construct.
  • There is no human reference, behavioral benchmark, causal intervention, external criterion, or known profile to evaluate accuracy.
  • Convergent, discriminant, predictive, incremental, factorial validity, or invariance across models are not studied.
  • The means of the four protocols differ widely; the paper offers no rule for choosing the correct measure.
  • The positive and negative extremes are imposed with prompts and are scored by a judge who knows the trait, creating circular calibration.
  • The neutral is also a designed prompt. It does not represent a natural or identified absence of the trait.
  • Projection, averaging, clipping, and interpolation compress variance by construction. No clipping frequency or ablations are published.
  • If neutral falls outside the segment or the axis is curved/multimodal, clipping hides the failure rather than calibrating it; the authors acknowledge the assumption.
  • OCEAN directions may be correlated and not independent axes, so per-trait scores may mix constructs.
  • No correlations between vectors, separation between traits, leakage between prompts, or axis stability across datasets are reported.
  • The extraction layer and fixed probe position are not identified; nor is there sensitivity analysis by layer, token, or aggregation.
  • The claim that response activations are better than prompt tokens is not accompanied by comparative results.
  • The theory assumes local linearity and well-trained parametric adaptation; the prompting experiments do not verify these assumptions.
  • The assumption of an orthonormal OCEAN basis in the generalization theorem is in tension with the admission of correlated axes.
  • Composition, negation, and out-of-domain synthesis are formulated theoretically but not tested empirically.
  • The bounds contain constants and O(.) terms that are not estimated; they do not predict error or calibration of the real experiment.
  • “Explainable” describes a geometric operation, but fidelity, causality, mechanistic localization, or human comprehensibility are not evaluated.
  • GPT-4.1-mini still determines the absolute anchors; judge bias and drift pass directly into the scores.
  • It is not specified how a logit-weighted mean over the 101 possible integers is obtained, nor what happens if tokens are not available.
  • GPT-5.2 generates the prompts used to demonstrate robustness; there is no independent generation, human audit, or adversarial audit of the set.
  • The alternative open-ended questions are not paraphrases but distinct probes, so they also change semantic facets of the trait.
  • There are only ten sets per family, with no justification of coverage, random sampling, or uncertainty intervals for the variance.
  • No variance equality tests, bootstrap, stability effect sizes, or correction for multiple comparisons are reported.
  • Three small models of a similar epoch are compared; there are no larger models, base models, MoE, other families, or closed models.
  • It only covers English, Big Five, single-turn, and controlled prompts; it does not test temporality, long context, conversation, or tool use.
  • It requires white-box access to activations and therefore cannot be applied directly to closed APIs.
  • Cost is not compared in a normalized way with baselines; averaging more samples may buy stability with more compute.
  • Exact checkpoint IDs, transformers versions, decoding, numerical precision, seeds, and hardware/software details are missing.
  • No official repository or artifact was found linked; code, full prompts, splits, outputs, and machine-readable results are missing.
  • The status is arXiv v1 preprint, with no venue or peer review indicated on the official surface consulted.

What the study does not establish

  • It does not establish that PVNI measures true, intrinsic, or stable LLM personality outside of its prompts.
  • It does not demonstrate that lower standard deviation implies greater accuracy or psychometric validity.
  • It does not prove that OCEAN scores correspond to model behavior, decisions, or interaction.
  • It does not demonstrate that the axes are independent, linear, causal, or that they generalize outside the prompt set.
  • It does not empirically validate the theorems of composition, negation, or out-of-domain synthesis.
  • It does not produce a judge-invariant evaluation applicable to closed models.
  • It does not allow reproducing the published numbers from identified public artifacts.

Traceability

Scope: Full text

Version: arXiv:2601.09833v1, submitted 14 January 2026; preprint, 17 pages

Consulted source: https://arxiv.org/pdf/2601.09833v1

Review: Codex full-text, bilingual-fidelity, 17-page visual, arXiv-v1, internal-activation, construct-validity, psychometric-validity, variance-vs-accuracy, clipping, judge-circularity, hidden-layer, theoretical-assumption, generalization-claim, prompt-coverage, compute-fairness and reproducibility audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Qwen-2.5-7B-Instruct
  • Llama-3-8B-Instruct
  • Mistral-7B-v0.1-Instruct
  • GPT-4.1-mini external trait judge, unspecified snapshot
  • GPT-5.2 prompt and question generator, unspecified snapshot

Instruments and metrics

  • Persona-Vector Neutrality Interpolation
  • Positive, negative and neutral contrastive persona prompts
  • Hidden-state mean-difference persona vectors
  • Clipped projection coefficient and anchor interpolation
  • IPIP-BFFM-50 self-report baseline
  • IPIP-NEO-120 self-report baseline
  • Open-ended elicitation with API judge
  • Big Five OCEAN coordinates
  • Standard deviation, IQR and boxplots over prompt sets

Data used

  • Ten generated questionnaire-variant prompt sets per protocol
  • Ten generated role-play-variant prompt sets per protocol
  • Open-ended trait questions split into extraction and evaluation sets, unreleased
  • Five-trait contrastive prompt ensembles, only partial examples published

Evidence and location

  • Metadata, abstract, and status: arXiv:2601.09833v1 abstract page, submitted 14 January 2026
  • Definition, projection, clipping, and interpolation: Paper, pp. 3–4, Section 3 and Algorithm 1
  • Assumptions and linear theory: Paper, pp. 4–6, Section 4, Assumptions 1–2 and Theorems 4.1–4.3
  • Models, judges, artifacts, hardware, and prompt sample: Paper, p. 7, Section 5.1 Experimental Settings
  • Qwen and Llama stability results: Paper, pp. 7–8, Table 1 and Figures 3–4
  • Mistral stability results: Paper, p. 16, Table 2
  • Variants and prompt coverage: Paper, pp. 12–15, Appendix A.1
  • Boxplots and robustness analysis: Paper, pp. 15–17, Appendix A.2–A.3 and Figures 5–7
  • Acknowledged limitations: Paper, p. 9, Section 7 Limitations
  • Integral visual inspection: Paper, all 17 rendered pages, including every equation, table, plot, and appendix page
  • Absence of linked official code: Paper and official arXiv surface checked 15 July 2026; no repository or executable artifact identified