Too Nice to Tell the Truth: Quantifying Agreeableness-Driven Sycophancy in Role-Playing Language Models

Applications, bias, and safety2026arXivApproved editorial review

Authors: Arya Shah, Deepali Mishra, Chaklam Silpasuwanchai

Keywords: Personality, Persona conditioning, Role-playing agents, Safety and bias

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

3
Authors
7
Findings
22
Limitations
6
Evidence

Editorial summary

English

Shah, Mishra and Silpasuwanchai study whether attributed agreeableness across 275 synthetic personas is associated with the tendency of thirteen open models, from 0.6B to 20B, to accept user-stated opinions. Each model scores the same personas with forty adapted NEO-IPIP items and answers 4,950 opinion prompts across 33 categories, first as a generic assistant and then under each persona. Responses are reduced to agreement 1, disagreement 0, or partial .5. Analysis combines correlations, regression, and high-versus-low comparisons over 275 persona means per model.

The reported pattern is real within that design: nine of thirteen models show a positive association between their own agreeableness scores and agreement rates. Llama 3.1 8B reaches r=.868 and OLMo 3 7B r=.853. The headline effect size, however, has a material contradiction: the abstract and introduction claim Cohen's d=2.33 for SmolLM3, while the tables assign it d=.455 and the study-wide maximum is d=1.282 for OLMo. No public result artifact resolves the discrepancy.

The baseline comparison changes the interpretation. Adopting a persona lowers agreement for most models: Llama falls from .36 to .05, SmolLM3 from .41 to .17, and Phi-4 Mini from .36 to .14. Gemma 3 1B is the clear exception, rising from .37 to .59; Yi and GPT-OSS barely change. The study therefore primarily finds variation among personas within each model, not a general increase in acquiescence caused by persona use.

The decisive limitation is construct validity. The appendix explicitly says prompts are subjective opinions rather than verifiable claims. Accepting a debatable stance is not necessarily lying, deception, or sacrificing factual accuracy. There are no truth labels, independent references, matched neutral phrasings, or evaluation of reasons. Trait-Truthfulness Gap merely multiplies the agreement shift by measured agreeableness; its unvalidated plus-or-minus .1 deception and truth zones do not turn subjective agreement into factuality.

Measurement also does not isolate agreeableness. The same model interprets and scores each persona and generates the outcome, creating shared-method variance. Personas simultaneously vary in occupation, ideology, ethics, and style, dimensions overlapping the prompt topics. There are no human or external ratings, reliability estimates, factor structure, measurement invariance, or randomized single-trait manipulation. The six tests are also two dependent analysis families over the same observations rather than six independent replications; multiplicity is uncontrolled and the median split discards information.

The manual-validation claim is unsupported. The paper points to Appendix D, but that appendix only gives deterministic extraction rules and reports no sample, annotators, agreement, or accuracy. ACL's official checklist says no human annotators were used. The repository releases code and complete inputs, 275 personas, 4,950 prompts, and the questionnaire, but not generations, matrices, breakdowns, figures, or hypothesis-test JSON that Appendix E says are available. The Hugging Face raw files exist, although its Viewer fails and the card describes inputs as results. This is a large study of persona-conditioned agreement sensitivity, not evidence of truth, deception, human personality, or an isolated causal effect of agreeableness.

Español

Shah, Mishra y Silpasuwanchai estudian si la amabilidad atribuida a 275 personas sintéticas se asocia con la tendencia de 13 modelos abiertos, de 0,6B a 20B, a aceptar opiniones planteadas por el usuario. Cada modelo puntúa las mismas personas con 40 ítems adaptados de NEO-IPIP y responde 4.950 prompts de opinión en 33 categorías, primero como asistente genérico y después bajo cada persona. Las respuestas se reducen a acuerdo 1, desacuerdo 0 o parcial .5. Sobre 275 medias por modelo, el artículo combina correlaciones, regresión y comparaciones alto-bajo.

El patrón publicado es real dentro de ese diseño: nueve de trece modelos muestran asociación positiva entre su propia puntuación de amabilidad y su tasa de acuerdo. Llama 3.1 8B alcanza r=.868 y OLMo 3 7B r=.853. Pero el titular de tamaño de efecto contiene una contradicción material: abstract e introducción anuncian Cohen d=2,33 para SmolLM3, mientras las tablas le asignan d=.455 y el máximo del estudio es d=1,282 para OLMo. Ningún resultado o artefacto público resuelve la diferencia.

La comparación con el baseline cambia la lectura. Adoptar una persona reduce el acuerdo en la mayoría de modelos: Llama pasa de .36 a .05, SmolLM3 de .41 a .17 y Phi-4 Mini de .36 a .14. Gemma 3 1B es la excepción clara, de .37 a .59; Yi y GPT-OSS apenas cambian. Por tanto, el estudio encuentra sobre todo variación entre personas dentro de cada modelo, no un aumento general de aquiescencia causado por usar personas.

La limitación decisiva es de constructo. El propio apéndice dice que los prompts son opiniones subjetivas, no afirmaciones verificables. Aceptar una postura debatible no equivale necesariamente a mentir, engañar o sacrificar precisión factual. No hay etiquetas de verdad, referencias externas, formulaciones neutrales emparejadas ni evaluación de las razones. Trait-Truthfulness Gap solo multiplica el cambio de acuerdo por la amabilidad ya medida; sus zonas de «engaño» y «verdad», fijadas en ±.1 sin validación, no convierten acuerdo subjetivo en factualidad.

La medición tampoco aísla amabilidad. El mismo modelo interpreta la persona, la puntúa y genera el resultado, lo que introduce método compartido. Las personas varían a la vez en ocupación, ideología, ética y estilo, dimensiones que se solapan con los temas de los prompts. No hay ratings humanos o externos, fiabilidad, estructura factorial, invariancia ni manipulación aleatoria de un único rasgo. Además, los seis tests son dos familias de análisis dependientes sobre los mismos datos, no seis réplicas independientes; no se corrige multiplicidad y el median split pierde información.

La afirmación de validación manual tampoco está respaldada. El texto remite al Appendix D, pero este solo describe reglas deterministas de extracción y no informa muestra, anotadores, acuerdo ni accuracy. El checklist oficial de ACL marca que no se usaron anotadores humanos. El repositorio publica código y los inputs completos, 275 personas, 4.950 prompts y cuestionario, pero no generaciones, matrices, breakdowns, figuras ni JSON de tests que el Appendix E dice disponibles. El dataset de Hugging Face existe como ficheros crudos, aunque su Viewer falla y la tarjeta los llama resultados cuando son inputs. Es un estudio amplio y útil sobre sensibilidad de acuerdo a persona, no una demostración de verdad, engaño, personalidad humana o causalidad de la amabilidad.

Research question

Is the agreeableness that each LLM attributes to a synthetic person associated with its agreement rate in response to solicited opinions, and how does that relationship change across models and against a generic assistant?

Method

Observational study within 13 open models. Each model scores 275 descriptions with 40 items adapted from NEO-IPIP and responds to 4,950 subjective prompts under baseline and under each person. Responses are converted to 0, .5 or 1; per model, 275 agreement means are correlated with agreeableness scores, a regression is fitted, and high-low groups are compared by median split with three related tests and effect sizes. TTG weights the change relative to baseline by agreeableness.

Sample: Thirteen models, 275 persons and 4,950 prompts produce 17,903,600 reported queries: 143,000 to score agreeableness, 64,350 for baseline and 17,696,250 under person. The primary analytical unit is each of the 275 persons per model, after averaging up to 4,950 valid responses; there is no population sampling of persons, checkpoint replicas, or immutable versions of models.

Findings

  • Nine of thirteen models show a significant positive association between self-rated agreeableness and agreement rate; Llama 3.1 8B reaches r=.868 and OLMo 3 7B r=.853.
  • The tables report a maximum Cohen d=1.282 for OLMo and d=.455 for SmolLM3, contradicting the d=2.33 for SmolLM3 in the abstract and the introduction.
  • Agreement rates under person are lower than baseline in the majority of models; Gemma 3 1B is the substantive exception, .59 versus .37.
  • Llama goes from baseline .36 to a person mean of .05; SmolLM3 from .41 to .17; Phi-4 Mini from .36 to .14.
  • TTG is negative for the majority because persona reduces agreement; Llama appears in the zone called truthful and Gemma in the one called deceptive, labels that have no external truth.
  • No clear relationship is observed between model size and the published association.
  • The public inputs contain exactly 275 persons and 4,950 prompts, with 150 per each of 33 categories; the outputs needed to verify the results are not public.

Limitations

  • The prompts are defined as subjective opinions without verifiable truth; agreement does not equate to factual error, deception, or lying.
  • There are no truth labels, independent sources, matched neutral prompts, or evaluation of content or reasons.
  • The Don't you agree? formulation is leading and is not compared with neutral, inverted, or counterbalanced versions.
  • Partial agree and partial disagree both receive .5, eliminating a directional distinction and any explanatory nuance.
  • Invalid responses are excluded without reporting rates per model or condition; the keyword fallback may match quoted or negated terms.
  • The same model scores agreeableness and generates the outcome, introducing shared method variance.
  • There are no human or external ratings, test-retest, internal consistency, factor structure, invariance, convergent, discriminant, or predictive validity.
  • The 275 persons simultaneously change occupation, ethics, ideology, history, and style; agreeableness is not manipulated in isolation.
  • The persons are designed and ordered to traverse low-to-high agreeableness, so they are not a random sample from a defined population.
  • Pearson, Spearman, and regression reuse the same association; Welch, Mann-Whitney, and permutation reuse the same median split. They are not six independent confirmations.
  • The majority-of-six criterion is ad hoc, uses one-sided hypotheses, and does not correct for multiplicity across models, facets, categories, and tests.
  • The median split loses information and its groups depend on the specific score distribution of each model.
  • Responses are nested within persons, prompts, and categories, but uncertainty at the item and category level is not propagated to the main conclusions.
  • TTG incorporates agreeableness in its formula and therefore is not an independent validation of its effect.
  • The ±.1 thresholds of TTG are arbitrary and have no calibration against observed deception or truth.
  • The d=2.33 in the headline contradicts d=.455 for SmolLM3 and the maximum d=1.282 in the tables.
  • The paper says it validates against manual annotations in Appendix D, but the appendix does not contain that validation and the official checklist declares that there were no human annotators.
  • Exact checkpoint revisions, Python, PyTorch, CUDA, Transformers, or a complete executable environment are not fixed.
  • The repository duplicates runners per model, and lacks tests, CI, release, lock, and execution manifest.
  • GitHub and Hugging Face do not publish generations or statistical outputs, even though Appendix E claims they are available.
  • The Hugging Face card describes five splits and results, but the Viewer fails and the public files are inputs without responses.
  • Deployment recommendations are not tested in multi-turn conversation, with real users, on factual tasks, or in operational contexts.

What the study does not establish

  • It does not demonstrate that the models tell falsehoods, deceive, or sacrifice factual accuracy by agreeing with subjective opinions.
  • It does not validate TTG as a measure of truth or deception, nor its zones as psychological or safety categories.
  • It does not identify agreeableness as an isolated cause because the persons contain multiple semantic and social differences.
  • It does not establish that an LLM possesses agreeableness, human personality, intention, belief, or motivation to please.
  • It does not test psychometric equivalence between model responses to the NEO-IPIP and human traits.
  • It does not demonstrate that using personas increases agreement in general; for the majority of models it reduces it relative to baseline.
  • It does not support the effect size d=2.33 announced in the abstract.
  • It does not provide six independent statistical replications or sufficient control of multiple comparisons.
  • It does not substantiate the claim of manual validation with the published materials.
  • It does not allow arithmetic reproduction of the tables, figures, or tests from the public artifacts.

Traceability

Scope: Full text

Version: arXiv:2604.10733v1 primary snapshot; ACL 2026 Long Paper 2026.acl-long.1421, DOI 10.18653/v1/2026.acl-long.1421, and official Responsible NLP Checklist also audited

Consulted source: https://arxiv.org/abs/2604.10733

Review: Codex dual 14-page ACL and arXiv visual full-text, two-page checklist, TeX/source, construct-validity, measurement, stance-scoring, statistical-dependence, TTG, GitHub, Hugging Face data-quality and reproducibility audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Qwen3 0.6B
  • Gemma 3 1B-IT
  • Granite 3.3 2B-Instruct
  • LFM2 2.6B
  • SmolLM3 3B
  • Phi-4 Mini-Instruct
  • Yi 6B-Chat
  • Mistral 7B-Instruct v0.2
  • OLMo 3 7B-Instruct
  • Qwen2.5 7B-Instruct
  • Llama 3.1 8B-Instruct
  • MiniCPM4 8B
  • GPT-OSS 20B

Instruments and metrics

  • Adapted 40-item NEO-IPIP agreeableness instrument
  • Trust, Altruism, Cooperation and Sympathy subscales
  • Structured stance classification scored 0, .5 or 1
  • Pearson and Spearman correlation
  • Linear regression
  • Median-split Welch, Mann-Whitney and permutation comparisons
  • Cohen d and Hedges g
  • Trait-Truthfulness Gap

Data used

  • 275 synthetic persona descriptions
  • 4,950 subjective opinion prompts across 33 categories
  • 64,350 generic-baseline input messages
  • 1,361,250 persona-prompt input messages
  • Public Hugging Face dataset aryashah00/Persona-Induced-Sycophancy

Evidence and location

  • Publication, design, results, limitations, and appendices: ACL Anthology 2026.acl-long.1421, DOI 10.18653/v1/2026.acl-long.1421, 14 pages rendered and inspected
  • Preprint and editable source: arXiv:2604.10733v1; 14 pages rendered and inspected; source SHA-256 5051531e15d1d324faba719fd7f77623136efce3fcb540a0bbd7b8ec9775cde2
  • Statement on human annotators and use of AI assistants: Official Responsible NLP Checklist for 2026.acl-long.1421, both pages rendered and inspected
  • Code, scope, absence of outputs, and reproduction quality: https://github.com/aryashah2k/Quantifying-Agreeableness-Driven-Sycophancy-in-Role-Playing-Language-Models commit 6e84bd619c62dcf641034e9695b2f07316037f93
  • Inputs, counts, Viewer, and absence of results in the dataset: https://huggingface.co/datasets/aryashah00/Persona-Induced-Sycophancy
  • Comprehensive audit of construct, TTG, statistics, artifacts, and reproducibility: reports/verification/article-368-acl-agreeableness-sycophancy-subjective-construct-ttg-circular-measurement-statistics-artifact-and-reproducibility-audit.json