Psychometric Personality Shaping Modulates Capabilities and Safety in Language Models

Trait induction and control2025arXivApproved editorial review

Authors: Stephen Fitz, Peter Romero, Steven Basart, Sipeng Chen, Jose Hernandez-Orallo

Keywords: Artificial Intelligence, Computation and Language, Large Language Models, Personality Traits, Model Safety

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

5
Authors
20
Findings
71
Limitations
16
Evidence

Editorial summary

English

The preprint studies how much five models change when each benchmark question is preceded by a system prompt describing a high, medium, or low level of a Big Five trait. Descriptions concatenate 104 bipolar markers with the intensifier “extremely.” In addition to an unprompted baseline, the study tests isolated traits; a low-Agreeableness, low-Conscientiousness, high-Neuroticism profile; a low-Agreeableness, low-Conscientiousness, high-Extraversion profile; and a profile setting every trait to medium. Models are GPT-4.1, Llama-3-8B-Instruct, Llama-3-70B-Instruct, Llama-4-Maverick, and DeepSeek-V3; generation uses temperature 0, top-p 1, and declared seed 43 on MMLU, TruthfulQA, WMDP, five ETHICS tasks, and Sycophancy. IPIP-NEO-300 checks whether self-reports follow the prompt, while SD3 measures dark-triad self-reports.

The largest effect occurs under low Conscientiousness, but it is highly heterogeneous. For GPT-4.1, TruthfulQA falls from 83.2 to 44.4, ETHICS-Commonsense from 71.4 to 38.7, and MMLU from 84.6 to 67.8; for Llama-3-70B the corresponding changes are 71.5→43.3, 67.0→45.2, and 77.8→69.8. Patterns are much smaller or mixed for Llama-3-8B, Llama-4, and DeepSeek-V3: MMLU changes by −4.6, −2.7, and −1.1 points, respectively. The text says that “all” lose 20–40 points, but the tables do not support that generalization across all five models. Composite profiles do lower ETHICS-Commonsense with little MMLU loss in several models, although the magnitude ranges from −26.4 points for GPT-4.1 to −8.7 for DeepSeek-V3.

These results demonstrate sensitivity to persona instructions; they do not yet identify a causal effect of latent personality. Markers contain content that directly overlaps the tasks: low Conscientiousness says “lazy,” “irresponsible,” and “careless”; low Agreeableness says “immoral,” “dishonest,” and “uncooperative”; low Openness says “unintelligent” and “unanalytical.” Lower MMLU accuracy or less moral and truthful answers can therefore be literal instruction following. The IPIP check is also circular: the model receives adjectives that nearly dictate responses to semantically equivalent self-report items. SD3 remains another self-report under the same prompt, and its items overlap with “dishonest,” “self-important,” and “immoral”; it does not by itself provide external behavioral validation or symbol grounding.

“Safety” interpretation must also respect metric direction. WMDP measures hazardous knowledge: greater accuracy means greater ability to answer biosecurity, chemistry, and cybersecurity questions, not automatically greater safety. In Sycophancy, the behavioral column is the percentage of answer changes after a challenge; GPT-4.1 moves from 9.5 to 98.8 under low Conscientiousness. The main figure nevertheless omits this column and displays only original-answer accuracy, while the abstract and discussion invoke Sycophancy. Heatmaps uniformly color accuracy increases as “improvement,” a convention that cannot be applied identically to WMDP or answer-changing. Deployment recommendations about “safe” profiles are not tested in conversations, real attacks, open-ended generation, or harm outcomes.

The statistical and integrity audit finds contradictions that prevent accepting the claims of significance and validity as written. The main study is deterministic and provides neither item-level intervals nor tests. The robustness section covers only GPT-4.1 and Llama-4, aggregates group means across prompt variants and models, computes d even for rows that are themselves standard deviations, and reports no p-values; the narrated d values for MMLU, ETHICS-CM, and TruthfulQA do not match the table. The psychometric “effect” ∆M/4 is not Cohen's d but receives conventional d thresholds; its tables mix raw baseline means with scaled changes in the other rows. The Llama-4 IPIP/SD3 table exactly duplicates Llama-3-70B and conflicts with Llama-4 values in another table. The figure called “nine-panel” contains twelve panels, excludes Llama-3-8B, and has incorrectly described axes; the landscape robustness table is nearly unreadable in the PDF. The official source archive contains TeX, figures, and generated tables, but no code or raw data. The defensible contribution is an empirical warning: extreme prompts containing morally and cognitively loaded descriptors can strongly move some benchmarks. It does not establish synthetic personality as the mechanism or show that these oscillations by themselves invalidate all safety results.

Español

El preprint estudia cuánto cambian cinco modelos cuando cada pregunta de un benchmark va precedida por un system prompt que describe un nivel alto, medio o bajo de un rasgo Big Five. Las descripciones concatenan 104 marcadores bipolares y el intensificador «extremely». Además de una línea base sin persona, se ensayan rasgos aislados, un perfil de baja amabilidad, baja responsabilidad y alto neuroticismo, otro de baja amabilidad, baja responsabilidad y alta extraversión, y un perfil que fija todos los rasgos en nivel medio. Los modelos son GPT-4.1, Llama-3-8B-Instruct, Llama-3-70B-Instruct, Llama-4-Maverick y DeepSeek-V3; se evalúan con temperatura 0, top-p 1 y semilla declarada 43 sobre MMLU, TruthfulQA, WMDP, cinco tareas ETHICS y Sycophancy. IPIP-NEO-300 comprueba si el autoinforme sigue el prompt y SD3 mide autoinformes de tríada oscura.

El efecto más grande aparece con baja responsabilidad, pero es muy heterogéneo. En GPT-4.1, TruthfulQA baja de 83,2 a 44,4, ETHICS-Commonsense de 71,4 a 38,7 y MMLU de 84,6 a 67,8; en Llama-3-70B las caídas correspondientes son 71,5→43,3, 67,0→45,2 y 77,8→69,8. En Llama-3-8B, Llama-4 y DeepSeek-V3 el patrón es mucho menor o mixto: por ejemplo, MMLU cambia −4,6, −2,7 y −1,1 puntos, respectivamente. El texto afirma que «todos» pierden 20–40 puntos, pero las tablas no sostienen esa generalización para los cinco modelos. Los perfiles combinados sí reducen ETHICS-Commonsense con poca pérdida de MMLU en varios modelos, aunque la magnitud va de −26,4 puntos en GPT-4.1 a −8,7 en DeepSeek-V3.

Estos resultados demuestran sensibilidad a instrucciones de persona, no identifican todavía un efecto causal de personalidad latente. Los marcadores incluyen contenido que se solapa directamente con las tareas: baja responsabilidad dice «lazy», «irresponsible» y «careless»; baja amabilidad, «immoral», «dishonest» y «uncooperative»; baja apertura, «unintelligent» y «unanalytical». Por tanto, una menor exactitud en MMLU o una respuesta menos moral o veraz puede ser obediencia literal a esas instrucciones. La comprobación IPIP también es circular: el modelo recibe adjetivos que casi dictan cómo contestar ítems de autoinforme semánticamente equivalentes. SD3 sigue siendo otro autoinforme bajo el mismo prompt y sus ítems se solapan con «dishonest», «self-important» o «immoral»; no constituye por sí solo validación conductual externa ni symbol grounding.

La interpretación de «seguridad» requiere además respetar la dirección de cada métrica. WMDP mide conocimiento peligroso: mayor exactitud significa más capacidad para responder preguntas de bioseguridad, química y ciberseguridad, no automáticamente mayor seguridad. En Sycophancy, la columna conductual es el porcentaje de cambios tras el desafío; GPT-4.1 pasa de 9,5 a 98,8 con baja responsabilidad. Sin embargo, la figura principal omite esa columna y muestra solo la exactitud de la respuesta original, aunque abstract y discusión hablan de Sycophancy. Los mapas colorean aumentos de exactitud como «mejora» de forma uniforme, una convención que no puede aplicarse igual a WMDP ni al cambio de respuesta. Las recomendaciones de despliegue sobre perfiles «seguros» no se validan en conversaciones, ataques reales, generación abierta o resultados de daño.

La auditoría estadística y de integridad encuentra contradicciones que impiden aceptar las afirmaciones de significación y validez tal como están escritas. El estudio principal es determinista y no publica intervalos ni pruebas por ítem. La sección robusta solo presenta GPT-4.1 y Llama-4, agrega medias de grupos de variantes y modelos, calcula d incluso sobre filas que ya son desviaciones estándar y no informa p-valores; los d narrados para MMLU, ETHICS-CM y TruthfulQA no coinciden con la tabla. El «efecto» psicométrico ∆M/4 no es Cohen d, pero recibe umbrales convencionales de d; sus tablas mezclan medias brutas en la fila baseline con cambios escalados en las demás. La tabla Llama-4 de IPIP/SD3 duplica exactamente la de Llama-3-70B y contradice los valores Llama-4 de otra tabla. La figura llamada «nine-panel» contiene doce paneles, excluye Llama-3-8B y usa ejes descritos incorrectamente; la tabla robusta apaisada resulta casi ilegible en el PDF. El archivo fuente oficial contiene TeX, figuras y tablas generadas, pero no código ni datos crudos. La contribución defendible es una alerta empírica: prompts extremos con descriptores moral y cognitivamente cargados pueden mover mucho determinados benchmarks. No establece que una personalidad sintética sea el mecanismo ni que esas oscilaciones invaliden por sí solas todos los resultados de seguridad.

Research question

Do capacity, dangerous knowledge, truthfulness, ethical responses, and susceptibility to changing answers systematically change when the same LLM receives extreme prompts constructed with Big Five markers, and can these changes be separated from its general capacity?

Method

A system prompt with bipolar markers of a Big Five trait at high, medium, or low level is prepended to each item, in addition to two multirtrait profiles and an all-medium profile. Five models are evaluated via APIs with temperature 0, top-p 1, and seed 43 on MMLU, TruthfulQA, WMDP bio/chem/cyber, ETHICS commonsense/deontology/justice/utilitarianism/virtue, and two Sycophancy outputs. IPIP-NEO-300 and SD3 are administered under the same prompts. The main result is accuracy differences against the baseline; an appendix adds ∆M/4 for self-reports and a high versus low conscientiousness robustness over prompt variations in two models. The audit read and rendered 45 pages, checked all tables/figures, inspected the official arXiv source file, and searched for public code or data.

Sample: The basic unit is one benchmark question answered once per model and condition in the main deterministic run; the article does not report in a table the number of items actually scored, exclusions, API failures, or retries per benchmark. There are five models and nineteen conditions tabulated per model, but the main figures show only four and omit Llama-3-8B. The robustness adds high/low conscientiousness variants for GPT-4.1 and Llama-4 without clearly declaring the number of underlying prompts and seeds or preserving per-question matching.

Findings

  • The five models change their IPIP self-reports in the semantically requested direction.
  • All-medium prompts bring many IPIP scores exactly or almost exactly to 3.
  • Low agreeableness and some combined profiles raise SD3 self-reports of Machiavellianism and psychopathy.
  • Low conscientiousness produces the largest aggregate drops in GPT-4.1.
  • GPT-4.1 drops 38.8 points on TruthfulQA and 17.2 on MMLU with low conscientiousness.
  • GPT-4.1 drops between 26.2 and 33.7 points on the five ETHICS tasks with low conscientiousness.
  • Llama-3-70B drops 28.2 points on TruthfulQA and 8.0 on MMLU with low conscientiousness.
  • Llama-3-8B does not reproduce a TruthfulQA drop under low conscientiousness and loses 4.6 points on MMLU.
  • Llama-4 loses 12.2 points on TruthfulQA and 2.7 on MMLU under low conscientiousness.
  • DeepSeek-V3 loses 5.2 points on TruthfulQA and 1.1 on MMLU under low conscientiousness.
  • The effects of low conscientiousness are not uniform across sizes and families.
  • High extraversion reduces TruthfulQA in several models, with distinct magnitudes.
  • Low agreeableness strongly reduces ETHICS-Commonsense in the five tabulated models.
  • The low agreeableness, low conscientiousness, and high neuroticism profile reduces ETHICS-Commonsense in all models.
  • That profile barely changes MMLU in GPT-4.1, Llama-3-70B, Llama-4, and DeepSeek, but reduces it by 5.6 points in Llama-3-8B.
  • All-medium improves TruthfulQA over baseline in the five models, but does not uniformly improve all metrics.
  • In GPT-4.1 low conscientiousness raises the Sycophancy answer-change from 9.5 to 98.8.
  • The main figure omits the answer-change metric and only includes the original Sycophancy accuracy.
  • The prompts contain cognitive and moral descriptors directly related to the benchmarks.
  • The official source file confirms that no executable code or raw data are distributed.

Limitations

  • The work is a v1 preprint under review, not an accepted result at ICLR.
  • The main run uses a single deterministic answer per item and condition.
  • There are no main repetitions with multiple seeds or temperatures.
  • No confidence intervals per accuracy difference are published.
  • No paired per-question tests are applied in the main results.
  • Multiple comparisons across traits, levels, models, and benchmarks are not corrected.
  • Claims of statistical significance are not accompanied by p-values.
  • The effective number of items, exclusions, errors, or retries per benchmark is not reported.
  • Providers and API routes differ across models.
  • GPT-4.1 runs concurrently through OpenAI and OpenRouter without an equivalence analysis between providers.
  • Exact dates of each run or evaluator versions are not documented.
  • No code, environment, lockfile, or reproduction command is published.
  • No per-item responses or raw data are published.
  • The source file contains only TeX, figures, and derived tables.
  • The system prompts are direct instructions in addition to construct manipulations.
  • Low conscientiousness includes lazy, irresponsible, careless, and negligent.
  • Low openness includes unintelligent and unanalytical.
  • Low agreeableness includes immoral, dishonest, unkind, and uncooperative.
  • High openness includes intelligent and analytical.
  • Semantic overlaps conflate trait, capacity, morality, and prompt compliance.
  • There are no controls with equally negative adjectives unrelated to Big Five.
  • There are no paraphrases that remove words directly related to the task.
  • There is no control of prompt length, valence, or normative load.
  • The IPIP check reuses the same vocabulary that defines the condition.
  • IPIP results near 1, 3, and 5 may be explicit instruction compliance.
  • Internal reliability, factorial structure, or invariance are not estimated in these five models.
  • SD3 remains a self-report under the same persona context.
  • Low agreeableness markers directly overlap with SD3 constructs.
  • SD3 does not provide independent external behavior.
  • The symbol grounding claim is not demonstrated with two textual questionnaires.
  • WMDP measures dangerous knowledge and does not have the same normative direction as ETHICS.
  • Lower WMDP may represent lower dangerous capacity, not worse safety.
  • The maps color accuracy increases as improvement without harmonizing safety direction.
  • Sycophancy combines initial accuracy and answer change, which require distinct interpretations.
  • The main figure excludes the behavioral answer-change metric.
  • The discussion attributes effects to Sycophancy without showing that metric in the main figure.
  • Harm, safe refusal, harmful compliance, or policy robustness are not measured.
  • Real attacks, jailbreaks, or adversarial users are not tested.
  • Open-ended generation or multi-turn conversation is not evaluated.
  • Persistence after prompt removal is not studied.
  • Contextual acting is not separated from a stable internal state.
  • The robustness only covers high versus low conscientiousness.
  • The tabulated robustness only covers GPT-4.1 and Llama-4.
  • The quantity and content of each variant or seed are not clearly detailed.
  • The aggregated table mixes eight summaries per condition from two models and four groups.
  • d is also computed over rows that already represent standard deviations.
  • The aggregated groups are not equivalent independent replicates.
  • The natural matching of the same questions between CON_HI and CON_LO is not preserved.
  • The narrated d values 1.78, 2.47, and 2.16 do not match 0.949, 2.293, and 2.095 in the table.
  • Cohen's d quantifies magnitude and does not by itself demonstrate statistical significance.
  • The ∆M/4 index is called an effect, but it is not a variability-standardized difference.
  • Conventional d thresholds are applied to ∆M/4 without sufficient justification.
  • The ∆M/4 tables mix raw baseline means with scaled changes in other rows.
  • The Euclid_scaled of baseline is the norm of raw values, not a change with respect to itself.
  • The Llama-4 IPIP/SD3 table exactly duplicates that of Llama-3-70B.
  • The duplicated Llama-4 values contradict its own ∆M/4 effects table.
  • The inconsistency prevents knowing which psychometric results for Llama-4 are correct.
  • The figure called nine-panel contains twelve panels.
  • The caption describes rows as levels and columns as families, but each panel organizes traits and benchmarks.
  • The main figure excludes Llama-3-8B despite five models being presented.
  • The combined profiles figure also excludes Llama-3-8B.
  • The text alternates between three families, four shown models, and five evaluated models.
  • Externality appears in error where the design and other captions indicate Extraversion.
  • The rotated Table 7 is nearly illegible in the PDF layout.
  • The High Conscientiousness and counter-prompting recommendations are not tested as mitigations.
  • Whether an attacker can overwrite or combine these system prompts is not evaluated.
  • False positives of a persona-indicator detector are not quantified.
  • Alternative languages, cultures, or taxonomies such as HEXACO are not analyzed.
  • Analogies with human associations do not allow inferring equivalent mechanisms in LLMs.
  • The conclusion that the axis is orthogonal to capacity is not tested with a fixed-capacity causal design.
  • The claim that all safety benchmarks are called into question exceeds the observed evidence.

What the study does not establish

  • It does not demonstrate that models possess human personality or an internal Big Five state.
  • It does not separate latent personality from literal compliance with cognitive and moral adjectives.
  • It does not establish independent construct validity for IPIP or SD3 in these experiments.
  • It does not demonstrate symbol grounding.
  • It does not demonstrate that higher WMDP accuracy is greater safety.
  • It does not demonstrate statistical significance of all declared changes.
  • It does not establish robustness across the five models, the five traits, and all benchmarks.
  • It does not prove that personality and capacity are orthogonal axes in general.
  • It does not demonstrate that all-medium is a safe deployment mitigation.
  • It does not validate real-time persona detection or neutralization.
  • It does not prove a novel vulnerability distinct from general sensitivity to system prompts.
  • It does not demonstrate persistence outside the context containing the instruction.
  • It does not generalize to dialogue, open-ended generation, agents, or real harm outcomes.
  • It does not allow exact reproduction of results without code and raw data.
  • It does not by itself invalidate the entire safety benchmark literature.

Traceability

Scope: Full text

Version: arXiv:2509.16332v1 (19 Sep 2025); under review as an ICLR 2026 conference paper; DOI 10.48550/arXiv.2509.16332; CC BY 4.0

Consulted source: https://arxiv.org/pdf/2509.16332v1

Review: Codex full-text, visual, psychometric, statistical, benchmark-direction, source-integrity and reproducibility audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • GPT-4.1 snapshot gpt-4.1-2025-04-14 through OpenAI and OpenRouter
  • Meta-Llama-3-70B-Instruct through OpenRouter
  • Meta-Llama-3-8B-Instruct through OpenRouter
  • Llama-4-Maverick-17B-128E-Instruct through OpenRouter
  • DeepSeek-V3-0324 through Azure
  • Deterministic main generation: temperature 0, top-p 1, declared fixed seed 43

Instruments and metrics

  • 104 bipolar Big Five adjective markers with extreme, medium and low qualifiers
  • IPIP-NEO 300-item self-report inventory
  • Short Dark Triad 27-item self-report inventory
  • MMLU zero-shot accuracy
  • TruthfulQA MC1 accuracy
  • WMDP bio, chemistry and cybersecurity accuracy
  • ETHICS commonsense, deontology, justice, utilitarianism and virtue accuracy
  • Sycophancy original-answer accuracy and answer-change/admit-mistake percentage
  • Percentage-point change from unprompted baseline
  • Range-scaled psychometric difference ∆M/4
  • Cohen's d over aggregated prompt-variation summaries

Data used

  • MMLU
  • TruthfulQA multiple-choice single-answer subset
  • WMDP biosecurity, chemistry and cybersecurity
  • ETHICS five-subtask suite
  • Sycophancy challenge-after-answer evaluation
  • Generated IPIP-NEO and SD3 aggregate score tables
  • Official arXiv v1 source archive containing TeX, PNG figures and generated tables but no raw outputs or executable experiment code

Evidence and location

  • Objective, models, and thesis of personality as a control axis: arXiv:2509.16332v1, abstract and sections 1–2, pp. 1–3
  • Markers, qualifiers, and prompts with cognitive/moral content: Paper, section 3, pp. 4–5; Appendix D, pp. 21–30
  • Five models, APIs, and temporal cost: Paper, section 4.1, p. 5; Appendix F and Table 18, pp. 37 and 44
  • Benchmarks and metric direction: Paper, section 4.2, p. 5; extended Tables 8–12, pp. 40–42
  • IPIP-NEO and SD3: Paper, sections 3 and 4.3, pp. 4–6; Tables 13–17, pp. 42–44
  • Narrated results and recommendations: Paper, section 5, pp. 6–9
  • Complete results by model and condition: Paper, Appendix E, Tables 8–17, pp. 31–44
  • Heterogeneous low conscientiousness and Sycophancy metric omitted from the figure: Paper, Figure 2, p. 8; Tables 8–12, pp. 40–42; cross-table audit 15 Jul 2026
  • ∆M/4 index and psychometric tables: Paper, Appendix A, equation 1 and Tables 1–5, pp. 16–18
  • Robustness, narrated d, and aggregated table: Paper, Appendix B, equations and Tables 6–7, pp. 19 and 39
  • Llama-4 contradiction with duplicated Llama-3-70B: Paper, Tables 3 and 5, pp. 17–18; Tables 14 and 16, pp. 43–44; integrity audit 15 Jul 2026
  • Deterministic configuration: Paper, Table 19, p. 45
  • Limitations and ethical considerations of the authors: Paper, section 6 and Appendix C/G, pp. 9 and 19–21, 37–38
  • Version, date, DOI, license, and review status: arXiv:2509.16332v1 metadata; OpenReview iNiU0GdjKM; CC BY 4.0; checked 15 Jul 2026
  • Official source file without code or raw data: .cache/editorial-sources/article-075/supplements/audit/arxiv-2509.16332v1-source.tar; sha256 b6b5a2fa9ffe8d29a0aeb1f63503679ff5e988b48ee88a86e02f80dc61857cb4
  • Absence of associated public repository: Paper, arXiv code links and targeted GitHub/web search; checked 15 Jul 2026