Stick to your Role! Stability of Personal Values Expressed in Large Language Models

Personas, identity, and agents2024journals.plos.orgApproved editorial review

Original title: Stick to your role! Stability of personal values expressed in large language models

Authors: Grgur Kovač, Rémy Portelas, Masataka Sawayama, Peter Ford Dominey, Pierre-Yves Oudeyer

Keywords: Computation and Language, Artificial Intelligence, Machine Learning, Value Stability, Personal Values

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

5
Authors
15
Findings
26
Limitations
14
Evidence

Editorial summary

English

Kovač and coauthors study a narrower question than whether an LLM “has values”: whether the ordering of the values it expresses is preserved when an apparently unrelated conversational context changes. They administer the 40-item Portrait Values Questionnaire (PVQ-40) to 21 models from six families after conversations about grammar, jokes, poetry, history, or chess. In one condition, each model role-plays 60 Tolkien characters or 50 famous people; in another, it receives no persona. The tested model converses with a second instance of the same model, and each questionnaire response is the highest-scoring A–F next-token option. Scores are centered within participant and aggregated into Schwartz's ten values.

Rank-order stability asks whether differences among simulated people retain the same ordering across two topics: for example, whether personas expressing more Tradition in one context remain relatively high in another. Ipsative stability asks whether the ordering of the ten values within one participant is preserved across contexts. These are different properties. A model may retain a similar within-profile ordering while failing to preserve differences among personas; it may also obtain a high correlation while being systematically wrong about every persona. The results therefore concern PVQ response stability under this protocol, not evidence for internal values.

With fictional personas, the highest mean rank-order result is Mixtral-8x7B-Instruct (r=0.43), followed by its 4-bit variant (0.30), Mistral-7B-Instruct-v0.2 (0.28), Qwen-72B (0.24), and GPT-3.5-1106 (0.20). With real-world personas, Mixtral-Instruct reaches 0.50 and Qwen-72B 0.46. Llama-2 and Phi models remain near zero. Without personas, ipsative stability is higher: Mixtral-Instruct reaches 0.84, its quantized variant 0.82, Qwen-72B 0.73, and Zephyr 0.62. This does not establish human equivalence: the human references describe longitudinal change over years, whereas the model scores correlate synthetic answers after topic changes. The authors themselves restrict the comparison to identifying models clearly below those references.

In the only conversation-length experiment, conducted with Mixtral-8x7B-Instruct and fictional characters, rank-order stability falls from 0.42 at three messages to 0.15 at 43, while ipsative stability stays high. The neutral-profile analysis suggests convergence toward a default profile: preserving within-person ordering is not the same as preserving differentiated identities. Three author-created behavioral tasks partly reproduce the family ranking but have very different ceilings. Religion is the most stable (maxima of 0.66–0.68), Donation is moderate (maximum 0.31), and Stealing reaches only 0.16. Expected PVQ–Donation directions appear for Universalism, Benevolence, Power, and Achievement, but none of the correlations exceeds 0.3.

The paper's strongest contribution is methodological: it demonstrates why a minimal-context psychological score cannot by itself characterize later behavior and separates interpersonal from intrapersonal stability. The comparative evidence is nevertheless exploratory. Topics, personas, and tasks are convenience samples; contexts are generated by the tested model family; sampling configurations differ across families; PVQ measurement structure is not validated for LLMs; and explanations involving scale, data, SFT, DPO, RLHF, quantization, or alignment are observational and confounded. The code audit also finds independent-sample t-tests applied to matched units, raw averaging of Spearman correlations, and a BH implementation that includes 21 diagonal entries in addition to the 210 declared pairwise comparisons. The official v1.0 tag contains neither data nor results; the later repository restores input stimuli but still lacks model outputs. The study therefore supports strong contextual sensitivity and descriptive checkpoint differences under this protocol. It does not establish internal values, training causality, general behavioral stability, or independent reproducibility of the published numbers.

Español

Kovač y coautores estudian una pregunta más precisa que si un LLM «posee valores»: hasta qué punto conserva el orden de los valores que expresa cuando cambia un contexto conversacional aparentemente ajeno. Administran el Portrait Values Questionnaire de 40 ítems (PVQ-40) a 21 modelos de seis familias, después de conversaciones sobre gramática, chistes, poesía, historia o ajedrez. En una condición cada modelo representa a 60 personajes de Tolkien o 50 personas famosas; en otra no recibe persona. El modelo evaluado conversa con otra instancia del mismo modelo y, para cada pregunta, se elige la letra A–F con mayor puntuación de siguiente token. Los scores se centran por participante y se agregan en los diez valores de Schwartz.

La estabilidad rank-order pregunta si las diferencias entre personas simuladas se ordenan igual entre dos temas: por ejemplo, si quienes expresan más tradición en un contexto siguen ocupando posiciones altas en otro. La estabilidad ipsativa pregunta si el orden de los diez valores dentro de una misma respuesta se conserva entre contextos. Son propiedades distintas. Un modelo puede mantener un perfil interno parecido y, a la vez, representar mal las diferencias entre personas; también puede obtener una correlación alta aunque el perfil sea sistemáticamente incorrecto. Por eso el resultado debe leerse como estabilidad de respuestas PVQ bajo este protocolo, no como evidencia de valores internos.

Con personas ficticias, el mayor rank-order medio es Mixtral-8x7B-Instruct (r=0,43), seguido por su variante 4-bit (0,30), Mistral-7B-Instruct-v0.2 (0,28), Qwen-72B (0,24) y GPT-3.5-1106 (0,20). Con personas reales, Mixtral-Instruct alcanza 0,50 y Qwen-72B 0,46. Los Llama-2 y Phi quedan cerca de cero. Sin persona, la estabilidad ipsativa es mayor: Mixtral-Instruct llega a 0,84, su variante cuantizada a 0,82, Qwen-72B a 0,73 y Zephyr a 0,62. Esto no implica equivalencia humana: los referentes humanos proceden de cambios longitudinales durante años, mientras aquí se correlacionan respuestas sintéticas tras cambios de tema. Los propios autores limitan la comparación a identificar modelos claramente por debajo de esos referentes.

En el único experimento de longitud, realizado con Mixtral-8x7B-Instruct y personajes ficticios, el rank-order cae de 0,42 con tres mensajes a 0,15 con 43; la estabilidad ipsativa permanece alta. El análisis de perfiles neutrales sugiere que las personas convergen hacia un perfil por defecto: conservar el orden intrapersonal no significa conservar identidades diferenciadas. Tres tareas conductuales creadas por los autores reproducen parcialmente la jerarquía de familias, pero con techos muy distintos. Religión es la más estable (máximos de 0,66–0,68), donación es moderada (máximo 0,31) y robo apenas llega a 0,16. Las correlaciones esperadas entre PVQ y donación aparecen para universalismo, benevolencia, poder y logro, pero ninguna supera 0,3.

La contribución sólida es metodológica: hace visible que una puntuación psicológica de contexto mínimo no caracteriza por sí sola el comportamiento posterior y separa estabilidad interpersonal de intrapersonal. La evidencia comparativa, sin embargo, es exploratoria. Los temas, personas y tareas son muestras de conveniencia; los contextos son generados por el propio modelo; las configuraciones de muestreo difieren entre familias; no se valida la estructura psicométrica del PVQ en LLM; y las explicaciones sobre tamaño, datos, SFT, DPO, RLHF, cuantización o alignment son observacionales y confundidas. La auditoría del código encuentra además tests t independientes sobre unidades emparejadas, promedios directos de correlaciones de Spearman y una corrección BH que incluye 21 diagonales además de las 210 comparaciones declaradas. La etiqueta oficial v1.0 no contiene data ni results; el repositorio posterior recupera los estímulos, pero no los outputs de los modelos. Por tanto, el estudio apoya que este protocolo detecta una fuerte sensibilidad contextual y diferencias descriptivas entre checkpoints; no demuestra valores internos, causalidad de entrenamiento, estabilidad conductual general ni reproducibilidad independiente de las cifras publicadas.

Research question

To what extent do 21 LLMs maintain stable the interpersonal and intrapersonal order of the values they express when the topic or length of a conversation changes, with and without simulated persons, and do those differences transfer to three behavioral tasks?

Method

Exploratory experimental study. For each model, conversations on five topics are simulated with another instance of the same model, and each item of the PVQ-40 or of a downstream task is administered separately. In the conditions with persona, 60 Tolkien characters and 50 famous people are used; in the condition without persona, permutations of the response order are repeated. Each response is the allowed letter with the highest next-token score. The authors compute Spearman rank-order correlations between persons for each value and pair of contexts, and ipsative correlations among the ten values of each person; they then average correlations. They use five seeds for the main experiments, pairwise t comparisons with BH, a length experiment in one model, three behavioral tasks, and an analysis of 14 contexts in two models and one seed. The editorial audit read and rendered the 21 pages, verified PLOS ONE and arXiv v4, inspected the GitLab tag v1.0 and the subsequent repository, compiled the code, and reviewed generation, scoring, and statistical inference.

Sample: Twenty-one checkpoints from six families. In each model, the conditions with persona combine five seeds, five topics, 50 or 60 persons, and 40 PVQ items: 50,000 or 60,000 queries. Without persona, 50 permutations of options, five topics, and 40 items are used: 10,000 queries per model. Donation and theft have 100 prompts each; religion has five prompts in the released data. The length experiment uses only Mixtral-8x7B-Instruct with 3, 7, 11, 19, 27, 35, and 43 messages; the 14-context experiment uses one seed and two models.

Findings

  • With fictional characters, Mixtral-8x7B-Instruct obtains the highest mean rank-order (r=0.43), followed by Mixtral-Instruct 4-bit (0.30), Mistral-Instruct-v0.2 (0.28), Qwen-72B (0.24), GPT-3.5-1106 (0.20), and GPT-3.5-0125 (0.15).
  • With real persons, Mixtral-8x7B-Instruct reaches r=0.50 and Qwen-72B reaches 0.46; the Llama-2 and Phi models remain near zero in both populations.
  • Without persona instruction, ipsative stability is 0.84 for Mixtral-Instruct, 0.82 for its 4-bit version, 0.73 for Qwen-72B, and 0.62 for Zephyr; Mistral-Instruct-v0.2 obtains 0.48 and Llama-70B-chat 0.47.
  • The human longitudinal benchmarks used are rank-order 0.57 between ages 10 and 12 and 0.66 between ages 20 and 28, and ipsative 0.66 and 0.59; the article acknowledges that the comparison favors the LLMs and only allows a one-directional inference.
  • In Mixtral-Instruct with fictional persons, increasing the conversation from 3 to 43 messages reduces the rank-order from 0.42 to 0.15; between 35 and 43 it barely changes from 0.166 to 0.162.
  • Ipsative stability remains high as conversations lengthen; together with the neutral profile in the appendix, the pattern suggests convergence of personas toward a common profile, not conservation of differentiated identities.
  • Religion is the most stable downstream task: Mistral-Instruct-v0.2 obtains 0.66, Mixtral-Instruct 0.67, and Qwen-72B 0.68.
  • In donation, the maxima are 0.31 for Mixtral-Instruct, 0.28 for its 4-bit version, and 0.25 for GPT-3.5-1106.
  • Theft is much more difficult: the maximum reported is approximately 0.16 and the differences between models are compressed.
  • The correlations between PVQ values and donation have the expected directions for universalism and benevolence, and the opposite for power and achievement, but none exceeds 0.3.
  • In 14 contexts and a single seed, the mean stability is 0.215 for Mistral-Instruct-v0.2 and 0.334 for Mixtral-Instruct; contexts with longer initial messages tend to yield lower stability.
  • The descriptive hierarchies frequently favor Mixtral, Mistral, Qwen, and GPT-3.5 over Llama-2 and Phi, but they are not uniform across all tasks and do not identify a cause.
  • The code uses five seed aggregates per model for the main rank-order contrasts and independent t tests, although seeds, topics, and persons are paired across models.
  • The FDR function includes the diagonal of 21 unit p-values: it corrects 231 entries, not the 210 between-model comparisons declared in the text; the effect is conservative but confirms a divergence between the described and executed method.
  • The v1.0 tag compiles, but ignores and contains no data or results; the subsequent state contains the stimuli and persons, not the raw outputs or results needed to recompute the figures.

Limitations

  • "Value stability" operationalizes correlations of PVQ responses, not the existence, intensity, or stability of internal values.
  • No validation is performed in LLMs of the factorial structure, reliability, invariance, convergent, discriminant, or criterion validity of the PVQ-40.
  • The PVQ is answered through a forced A to F choice by next-token score; it does not measure an open response or spontaneous behavior.
  • Rank-order mixes person differentiation and context sensitivity; a high correlation preserves relative order, but does not demonstrate profile accuracy.
  • Only five main topics are used, brief and hand-written; they do not form a representative sample of deployment contexts.
  • Each conversation is generated by another instance of the same model, so the checkpoints receive different contents and the comparison mixes sensitivity with interlocutor behavior.
  • Configurations are not standardized: GPT uses temperature 1, Zephyr 0.7, and many Hugging Face models use do_sample=true with version-dependent defaults.
  • Seeds control permutations of options, but the code does not fix the global RNGs of Python, NumPy, or Torch for stochastic generation.
  • The Tolkien and famous-person populations are convenience samples without validated gold profiles or representativeness.
  • Templates vary between base models, chat with system, and chat without system; the Llama-2 control does not eliminate all format effects.
  • Spearman correlations are averaged directly without Fisher z transformation, hiding their distribution by value, context, and person.
  • If a single series is constant, the code substitutes the undefined correlation with 0; if both are constant it returns NaN. The paper does not discuss these decisions.
  • The rank-order t tests compare five seed means as independent samples, although models share seeds, topics, and populations; no pairing or hierarchy is modeled.
  • The ipsative contrasts also do not exploit the pairing of the same permutations or persons across models.
  • The BH implementation includes 21 diagonals in addition to 210 comparisons, and the binary matrices do not provide p-values, effects, or numeric intervals.
  • There is no preregistration or clear correction for the multiple exploratory PVQ to donation correlations.
  • The human comparison is not commensurable: years of longitudinal development versus immediate topic changes in synthetic agents.
  • The length experiment is limited to one model and fictional characters; the 14-context experiment uses one seed and two selected models.
  • Donation, religion, and theft are ad hoc instruments without validation, reliability, difficulty analysis, or external criteria.
  • The text says that religion contains six queries, but it lists five and the released CSV has five rows.
  • The explanations about size, data, SFT, DPO, RLHF, quantization, and alignment compare non-equivalent checkpoints and do not identify causality.
  • The GPT-3.5 versions are proprietary; weights, data, and training details are not known.
  • Only English is evaluated, despite the fact that values and persons have strong cultural components.
  • The PLOS statement says that code and data are available, but v1.0 omits data and results; the subsequent repository also does not publish raw outputs.
  • Without responses, run manifests, logs, and intermediate results, bars, errors, tests, or exclusions cannot be independently recomputed.
  • There is no test suite; requirements duplicates termcolor and tiktoken and depends on old versions and backends that are difficult to reconstruct.

What the study does not establish

  • It does not demonstrate that LLMs possess personal values, internal personality, or a persistent self.
  • It does not demonstrate that the PVQ is a psychometrically valid measure of the human construct in an LLM.
  • It does not demonstrate that Mixtral, Mistral, Qwen, or GPT-3.5 are globally more stable outside of these prompts, versions, and tasks.
  • It does not demonstrate that a high rank-order corresponds to a faithful representation of a specific person.
  • It does not demonstrate equivalence or superiority relative to humans; the longitudinal benchmarks are not a test of equivalence.
  • It does not demonstrate that long conversations always destroy the coherence of personas.
  • It does not demonstrate that high ipsative stability preserves differentiated identities; it may reflect a default profile.
  • It does not demonstrate that PVQ responses predict general behavior; the relationships with ad hoc tasks are small and variable.
  • It does not demonstrate that SFT or DPO cause stability, that RLHF reduces it, or that alignment prevents representing controversial characters.
  • It does not demonstrate that dataset size is more important than model size.
  • It does not demonstrate a causal effect of quantization; it only observes two Mixtral pairs.
  • It does not establish a sufficient threshold for deployment, safety, or social simulation.
  • It does not allow independent verification of the published figures from the released artifact.
  • It does not cover other languages, cultures, values, contexts, or subsequent models.
  • It does not replace psychometric validation, longitudinal behavioral evaluation, or audit of a real use case.

Traceability

Scope: Full text

Version: arXiv v4, 21-page extended version updated 28 August 2024; cross-checked against the PLOS ONE version published 26 August 2024, 19(8): e0309114, DOI 10.1371/journal.pone.0309114

Consulted source: https://arxiv.org/pdf/2402.14846

Review: Codex editorial review and methods/artifact audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • Llama 2 base and chat: 7B, 13B and 70B (6 checkpoints)
  • Mistral-7B base, Mistral-7B-Instruct v0.1 and v0.2, and Zephyr-7B-beta (4 checkpoints)
  • Mixtral-8x7B base and instruct, each at 16-bit and 4-bit precision (4 checkpoints)
  • Phi-1 and Phi-2 (2 checkpoints)
  • Qwen-7B, Qwen-14B and Qwen-72B (3 checkpoints)
  • GPT-3.5-turbo-1106 and GPT-3.5-turbo-0125 (2 checkpoints)

Instruments and metrics

  • Portrait Values Questionnaire, 40-item version (PVQ-40)
  • Schwartz's ten basic personal values
  • Rank-order stability across simulated participants
  • Ipsative stability within simulated participants
  • Author-created Donation, Religion and Stealing tasks

Data used

  • 60 Tolkien characters selected mainly from Wikipedia page length and manually balanced
  • 50 real-world famous people taken from an online influential-people list
  • Five hand-written conversation starters: grammar, joke, poem, history and chess
  • PVQ male and female item variants
  • 100 Donation prompts and 100 Stealing prompts based on Tolkien races and names
  • Five Religion schedule prompts, despite the paper describing six
  • Fourteen-context extension with nine additional hand-written topics

Evidence and location

  • Publication, DOI, date, abstract, and availability: PLOS ONE 19(8): e0309114, published 26 Aug 2024, DOI 10.1371/journal.pone.0309114; official metadata, abstract and Data Availability
  • Design, conversations, PVQ, and two types of stability: arXiv v4 full paper, Figure 1 and sections 3.1 to 3.2, pp. 1 to 4
  • Twenty-one models and six families: arXiv v4 section 4.1, pp. 4 to 5
  • Rank-order, ipsative, and human comparison: arXiv v4 sections 4.2 to 4.3 and Figures 4 to 5, pp. 5 to 6
  • Length and convergence toward neutral profile: arXiv v4 section 4.4, Figures 6 to 7, pp. 5 to 7; Appendix D.2, pp. 16 to 17
  • Religion, donation, theft, and PVQ to donation relationship: arXiv v4 sections 4.5 to 4.6 and Figures 8 to 9, pp. 7 to 9; Appendix C, pp. 15 to 16
  • Fourteen contexts, factors, and limitations: arXiv v4 Figure 10 and sections How additional contexts, What influences, Limitations, pp. 8 to 11
  • Populations, prompts, scoring, and counts: arXiv v4 Appendices B to C, pp. 14 to 16; current official repository input files audited 15 Jul 2026
  • t tests and FDR described: arXiv v4 Appendix E and Figures 16 to 21, pp. 17 and 20 to 21
  • Spearman, averages, and constant series: Official GitLab tag v1.0 commit e237b7ab7e69fe9053e78a49afb87bccf50ad158, visualization_scripts/data_analysis.py, audited 15 Jul 2026
  • Independent tests and BH diagonal: Official GitLab tag v1.0 commit e237b7ab7e69fe9053e78a49afb87bccf50ad158, campaign_data_analysis.py, audited 15 Jul 2026
  • Sampling and seeds: Official GitLab tag v1.0 model configs, evaluate.py and run_campaign_seeds.sh, audited 15 Jul 2026
  • Absence of inputs and results in v1.0: Official GitLab v1.0: 59 tracked files, data_files=0 and result_files=0 because .gitignore excludes data/* and results*; current master has stimuli but no paper result JSONs, audited 15 Jul 2026
  • Visual integrity of the full text: arXiv v4 PDF SHA-256 babd4c09fdb9b6baaa6cc9d589d823abe431db0c153315923dfd84d7f69162b8; all 21 pages rendered and inspected 15 Jul 2026