PACIFIC: Can LLMs Discern the Traits Influencing Your Preferences? Evaluating Personality-Driven Preference Alignment in LLMs

Evaluation and psychometric validity2026arXivApproved editorial review

Authors: Tianyu Zhao, Siqi Li, Yasser Shoukry, Salma Elmalaki

Keywords: Large Language Models, Personality, Persona, AI Safety

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

4
Authors
7
Findings
31
Limitations
13
Evidence

Editorial summary

English

PACIFIC is a synthetic multiple-choice benchmark for testing whether organizing preferences by Big Five labels helps an LLM choose a congruent option; it is neither a collection of observed human preferences nor a psychometric test. Gemini 2.5 Pro generates 1,200 preference-question-four-choice triples across 20 topics and ten trait directions, high/low O, C, E, A, and N. Another LLM, GPT-4o-mini, scores every text from 1 to 7 per trait and assigns confidence. A .7 filter leaves 803 candidate preferences; the main experiments use 200 balanced questions, 20 per direction.

For Gemma-3-4B-IT, Table 1 reports 25.75% without preferences, 29.25% with five mixed preferences, 63% with five preferences from the correct trait, and 61.75% after adding two distractors. Adding ground-truth labels to the preferences reaches 76%, while an unlabeled reminder reaches 67%. Labeling choices as well drops to 57.25%, and labels without text to 37.75%, with collapse on several low traits. Llama-3-8B-Instruct reaches 88.75% with aligned preferences; Gemini-2.5-Pro reaches 99.25% and GPT-4o-mini 97.5%, near ceiling on a corpus generated to expose the same rule.

The 29.25→76% headline combines two interventions: ground-truth selection of aligned raw preferences raises Gemma to 63%, then adding oracle labels raises it to 76%. This is not the automatic system's result. Without labels, pretrained DPR scores 30.25% and the contrastively fine-tuned retriever 43%. End-to-end personality inference is also weak: Gemma reaches 54.5% from preferences, with 95–100% on high traits but only 0–40% on low traits; prediction from choices reaches 85.69%. The paper attributes this asymmetry to RLHF-induced social-desirability bias, but it does not compare a base and RLHF model or eliminate prompt and dataset artifacts, so the causal attribution remains a hypothesis.

The human check has 15 participants complete an unnamed personality assessment and evaluate 25 cases routed to their dominant traits. It reports 78.22% accuracy, Fleiss' κ=.8599, Cohen's κ=.9170 between GPT-4o-mini and consensus, and 67.11% personal resonance. The instrument, allocation of people and items, full instructions, recruitment, compensation, uncertainty, mismatched-trait control, and ethics review are absent. A 67.11% approval rate is not a psychometric correlation.

The official Hugging Face artifact at commit dee6c7144ca0efa1132eebfdd538ecbde610b927 has 1,200 rows in one train split: exactly 120 per trait direction and 60 per topic, despite the paper's claim that it does not impose uniform distributions. Every row has ground_truth_choice_id=0 because the generation prompt requires the correct response first. No shuffling is documented; the non-100% reported scores therefore require an undisclosed preprocessing step or a different protocol that cannot be checked. The release also lacks the 2–4-turn conversations the paper says it adds; its card says they will come later. No code, 200-item sample IDs, retriever train/evaluation split, model outputs, or runnable results are public. PACIFIC is useful for showing how constructed OCEAN labels can organize context, but its evidence measures consistency with a synthetic taxonomy rather than real people's preferences.

Español

PACIFIC es un benchmark sintético de elección múltiple para estudiar si organizar preferencias por etiquetas Big Five ayuda a un LLM a elegir una opción coherente; no es una colección de preferencias humanas observadas ni una prueba psicométrica. Gemini 2.5 Pro genera 1.200 triples de preferencia, pregunta y cuatro opciones en 20 temas y diez direcciones de rasgo, alto/bajo para O, C, E, A y N. Otro LLM, GPT-4o-mini, puntúa cada texto de 1 a 7 por rasgo y asigna confianza. Un filtro a 0,7 deja 803 preferencias candidatas; los experimentos principales toman 200 preguntas balanceadas, 20 por dirección.

En Gemma-3-4B-IT, la Tabla 1 informa 25,75 % sin preferencias, 29,25 % con cinco preferencias mezcladas, 63 % con cinco preferencias del rasgo correcto y 61,75 % al añadir dos distractores. Etiquetar con el rasgo verdadero solo las preferencias eleva el resultado a 76 %, mientras un recordatorio sin etiquetas da 67 %. Etiquetar también las opciones baja a 57,25 % y usar solo etiquetas a 37,75 %, con colapso en varios rasgos bajos. Llama-3-8B-Instruct alcanza 88,75 % con preferencias alineadas; Gemini-2.5-Pro 99,25 % y GPT-4o-mini 97,5 %, casi techo en un corpus generado para hacer visible la misma regla.

El titular 29,25→76 % mezcla dos intervenciones: seleccionar preferencias mediante etiquetas verdaderas eleva Gemma a 63 %, y añadir esas etiquetas oracle al prompt lo eleva después a 76 %. No es el resultado del sistema automático. Cuando no se dispone de etiquetas, el DPR preentrenado obtiene 30,25 % y el retriever ajustado contrastivamente 43 %. Tampoco se infiere bien la personalidad de extremo a extremo: Gemma alcanza 54,5 % desde preferencias, con 95–100 % en rasgos altos y solo 0–40 % en rasgos bajos; desde opciones llega a 85,69 %. El paper atribuye la asimetría a sesgo de deseabilidad social inducido por RLHF, pero el diseño no compara un modelo base con uno RLHF ni descarta artefactos del prompt o del corpus, por lo que es una hipótesis, no una causa demostrada.

La validación humana reúne 15 participantes con una prueba de personalidad no identificada y 25 casos dirigidos a sus rasgos dominantes. Se informa 78,22 % de acierto, Fleiss κ=0,8599, Cohen κ=0,9170 entre GPT-4o-mini y el consenso, y 67,11 % de resonancia personal. Faltan instrumento, distribución de participantes y casos, instrucciones completas, reclutamiento, compensación, intervalos, comparación con rasgos no coincidentes y revisión ética. Una tasa de aprobación del 67,11 % no es una correlación psicométrica.

La auditoría del artefacto oficial en Hugging Face, commit dee6c7144ca0efa1132eebfdd538ecbde610b927, encuentra 1.200 filas en un único split train: exactamente 120 por dirección de rasgo y 60 por tema, aunque el paper afirma no imponer distribuciones uniformes. En todas las filas ground_truth_choice_id es 0 porque el prompt de generación exige colocar primero la respuesta correcta. No se documenta barajado; que los resultados publicados no sean 100 % implica un preprocesado no descrito o un protocolo distinto que no puede verificarse. El dataset no incluye las conversaciones de 2–4 turnos que el paper dice añadir; la ficha avisa que llegarán después. Tampoco hay código, IDs de la muestra de 200, split de ajuste/evaluación del retriever, respuestas de modelos o resultados ejecutables. PACIFIC es útil para revelar cómo etiquetas OCEAN construidas pueden organizar contexto, pero la evidencia mide consistencia con una taxonomía sintética y no preferencias reales de personas.

Research question

Does grouping, labeling, or retrieving synthetic preferences by Big Five directions help different LLMs select the answer option defined as coherent with the same trait, even when the exact preference does not appear in the question?

Method

Gemini 2.5 Pro generates preferences, questions, and four answers according to ten OCEAN directions and 20 topics. GPT-4o-mini scores traits and confidence; a filter retains high-confidence pairs. Over 200 balanced questions, the following are compared: absence of preferences, mixed preferences, aligned preferences, contaminated preferences, ground-truth labels, reminder, and RAG with DPR. Gemma-3-4B-IT, Llama-3-8B-Instruct, Gemini-2.5-Pro, and GPT-4o-mini are evaluated using accuracy. A small study compares synthetic options and preferences with participants of supposedly matching traits.

Sample: The release contains 1,200 rows in English, 120 for each of the ten OCEAN directions and 60 for each of 20 topics. The filter τ=0.7 produces 803 preferences according to the paper. Each main evaluation states it samples 200 questions: two per direction repeated ten times, with five preferences per question unless otherwise indicated. The human validation uses 15 participants and 25 instances, without sufficient breakdown to reconstruct assignments or denominators.

Findings

  • For Gemma, five aligned preferences per true label raise accuracy from 29.25% with mixed preferences to 63%; adding the ground-truth label of the preferences reaches 76%.
  • A reminder without labels reaches 67% in Gemma. Labeling preferences and options drops to 57.25%, and providing only labels drops to 37.75%, with strong asymmetry against Low traits.
  • Llama reaches 88.75%, Gemini 99.25%, and GPT-4o-mini 97.5% with aligned preferences; the two closed models are near ceiling, compatible with abundant textual signals of the generated trait.
  • The automatic system without labels falls well below the oracle: pretrained DPR 30.25% and fine-tuned DPR 43% in PACIFIC. The paper calls these values retrieval accuracy, but the table measures the final answer by trait.
  • Trait prediction from preferences obtains 54.5%: the five High traits are between 95% and 100%, while Low remains between 0% and 40%. From options it reaches 85.69%.
  • The human validation reports 78.22% accuracy and 67.11% resonance, with high κ; the available documentation does not allow auditing how they were calculated or interpreting 67.11% as a correlation.
  • The published artifact fixes the correct answer in the first position in the 1,200 rows and is exactly balanced. Without documented code or shuffling, the released benchmark has a deterministic positional shortcut.

Limitations

  • All preferences, questions, options, explanations, and first labels come from an LLM. The benchmark largely tests consistency with rules that the generator received explicitly.
  • The generator receives an OCEAN direction and must create an adherent answer first and three violations afterward. The construction favors evident lexical markers and artificial separation of options.
  • In the release, ground_truth_choice_id=0 for the 1,200 rows. No shuffling of options, seed, or map between published order and evaluated order is documented.
  • The published distribution is exactly uniform: 120 rows per trait direction and 60 per topic. This contradicts the claim of covering the spectrum without imposing uniform distributions.
  • The paper alternates between 1,200 curated pairs, 803 preferences after τ=0.7, and samples of 200 without publishing the exact IDs or a flag of the evaluated subset.
  • Only one train split exists. It is not described how the data used to fine-tune DPR are separated from the 200 evaluation questions; overlap or leakage cannot be excluded.
  • The published dataset does not contain the scenario context or the 2–4 turn dialogues described in the paper. The card itself indicates that full conversations will arrive later.
  • The headline 29.25→76% does not isolate selection by personality: 63% corresponds to selecting aligned text and 76% to adding true labels to the prompt.
  • The user's personality is not inferred in the 76% condition; the experiment assumes ground-truth labels of question and preferences, inaccessible in normal deployment.
  • The proposed automatic method reaches 43%, not 76%, and lacks a direct measure of retrieval quality such as recall@k, precision@k, or nDCG.
  • The end-to-end trait prediction fails almost completely in several Low categories. Therefore, the real bottleneck contradicts the presentation of inferred personality as a reliable signal.
  • The attribution of Low failure to RLHF and social desirability is post hoc. There is no base/instruction comparison, RLHF ablation, neutral reformulation, or causal experiment.
  • The taxonomy hardcodes debatable associations: high neuroticism with safety, low with efficiency and risk tolerance; low agreeableness, openness, or conscientiousness may receive normative or stereotyped descriptions.
  • Saying that the dataset formalizes a causal link between personality and preference confuses a rule imposed during generation with causal evidence about people.
  • The 32 profiles are factorial High/Low combinations, not measured individuals. Figure 4 analyzes constructed synthetic persons, not observed human diversity.
  • The human validation uses only 15 participants and 25 instances directed at their dominant traits, which may inflate agreement. There is no non-matching trait condition or real human preferences.
  • The personality questionnaire, its scales or cutoffs are not identified, nor are demographics, recruitment, compensation, consent, ethical approval, or annotation data published.
  • The 67.11% resonance is an approval proportion, not a correlation between measured personality and preference. No baseline, interval, or contrast is reported.
  • GPT-4o-mini reduces the bias of using exactly the generator Gemini as evaluator, but it remains automatic annotation based on the same definition and OCEAN prompts.
  • The words “significantly” and “confirms” are not accompanied by tests, intervals, inference repetitions, or multiple correction.
  • The protocol says 20 observations per direction, but the per-trait accuracies change in steps of 2.5 points and the global averages in 0.25; aggregation or additional runs need explanation.
  • Temperature, top-p, seeds, API dates, dated snapshots of Gemini/GPT, retries, or variability are not reported. Closed models may change.
  • The DPR fine-tuning reports some hyperparameters, but not exact checkpoint, top-k, complete negative construction, split, convergence criterion, seed, or validation metric.
  • The PrefEval replica uses derived labels and lacks several trait directions; it is not an independent psychometric control. Its results also do not show that adding labels improves over unlabeled preferences.
  • There is no evaluation of free generation, longitudinal dialogue, changing preferences, conflict between traits, response utility, satisfaction, or harm; the future work itself defers generation.
  • The corpus is English and uses 20 predefined topics. Cultural, linguistic, or out-of-synthetic-taxonomy generalization is not demonstrated.
  • There are high-risk domains (medicine, safety, finance, relationships) where obeying a risk preference can be harmful. The impact statement claims not to foresee specific risks without analyzing them.
  • Inferring personality from history raises consent, privacy, sensitive profiling, manipulation, and right to correct the profile. No user controls, minimization, or deletion are proposed.
  • The Hugging Face repository has an Apache-2.0 license and the data, but a minimal card without schema, detailed provenance, intended uses, limitations, risks, splits, or reproducible instructions.
  • The paper promises code, but no official repository was found. Without scripts, supplementary full prompts, samples, outputs, and logs, the tables cannot be regenerated.
  • Internal references are inconsistent: the setup section refers Gemma results and strategies to Table 11, which in the appendix contains only DPR; the main results are in Tables 1–3.

What the study does not establish

  • It does not establish that Big Five traits cause the synthetic preferences; that relation is imposed in the generation prompt.
  • It does not demonstrate that PACIFIC measures real human personality, psychometric validity, or natural preferences outside the constructed corpus.
  • It does not demonstrate that an automatic system reaches 76%: that figure uses selection and ground-truth labels, while RAG without labels reaches 43%.
  • It does not causally prove that RLHF produces the failure in Low traits or that the asymmetry is social desirability rather than prompt design or data stereotype.
  • It does not demonstrate that more memory is catastrophic in general; it only observes a small drop with mixed synthetic preferences in a specific model and protocol.
  • It does not validate safe, fair, or beneficial personalization in medicine, finance, safety, persuasion, or relationships.
  • It does not evaluate a persistent personality of the LLM; it profiles the user to select multiple-choice answers.
  • It does not allow reproducing the published results from the current official artifacts.

Traceability

Scope: Full text

Version: arXiv:2602.07181v3, submitted 6 February 2026, revised 7 April 2026; preprint under review, 28 pages

Consulted source: https://arxiv.org/pdf/2602.07181v3

Review: Codex full-text, bilingual-fidelity, 28-page visual, arXiv-v3, under-review-status, Hugging-Face-Dataset-Viewer, Parquet-distribution, synthetic-data, answer-position, psychometric-construct, oracle-label, headline-decomposition, retrieval-split, human-grounding, statistical-claim, social-desirability-causality, safety, privacy and reproducibility audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Gemini 2.5 Pro dataset generator and evaluated model
  • GPT-4o-mini dataset evaluator and evaluated model
  • Gemma-3-4B-IT
  • Llama-3-8B-Instruct
  • Dense Passage Retrieval with BERT bi-encoders
  • Contrastively fine-tuned persona-aware DPR

Instruments and metrics

  • Synthetic High/Low OCEAN persona directions
  • LLM-assigned 1–7 trait scores and 0–1 confidence
  • High-violation generation constraint
  • Confidence threshold τ=0.7
  • Four-choice preference-alignment accuracy
  • Ground-truth trait label prompting
  • Instruction-only personality reminder
  • Fleiss kappa and Cohen kappa
  • Unnamed participant personality assessment

Data used

  • PACIFIC, 1,200 released synthetic rows
  • PACIFIC confidence-filtered subset, reported as 803 preferences
  • PACIFIC 200-question balanced experimental sample
  • PrefEval-derived control subset
  • 15-person, 25-instance human grounding sample

Evidence and location

  • Metadata, abstract, scope, and headline: arXiv:2602.07181v3, pp. 1–2, Abstract, Introduction and Contributions
  • Synthetic construction, OCEAN, and human validation: Paper, pp. 3–4, Sections 2.1–2.3
  • Contexts, conditions, and sampling: Paper, pp. 5–6, Sections 3.1–3.2, Table 1 and Figure 2
  • Oracle labels, reminder, and RAG: Paper, pp. 6–8, Sections 3.2–4.2 and Tables 2–3
  • End-to-end prediction and bias attribution: Paper, pp. 8–9, Section 4.3 and Table 3
  • Memory, conclusion, and real scope: Paper, pp. 9–10, Discussion and Conclusion
  • Generation, filter, prompts, and PrefEval: Paper, pp. 13–18, Appendices A.1–A.5 and Prompts 1–2
  • Results of Llama, Gemini, and GPT-4o-mini: Paper, pp. 18–21, Appendix A.7 and Tables 8–10
  • Retriever, profiles, future, and impact: Paper, pp. 22–28, Appendices C–F, Tables 11–12 and Figure 4
  • Integral visual inspection: Paper, all 28 rendered pages, including every figure, table, prompt and appendix page
  • Schema, rows, splits, license, and absence of conversations: Official Hugging Face dataset, commit dee6c7144ca0efa1132eebfdd538ecbde610b927, README.md, PACIFIC.jsonl and Dataset Viewer API
  • Positional shortcut and exact distribution: Official Hugging Face Dataset Viewer and Parquet query: 1,200/1,200 ground_truth_choice_id=0; 120 rows per trait direction; 60 rows per topic
  • Absence of reproducible code: Paper and official arXiv/Hugging Face project surfaces; code release is stated as planned, not present