Evaluating Chinese Large Language Models: The Influence of Persona Assignment on Stereotypes and Safeguards

Applications, bias, and safety2026ACMApproved editorial review

Authors: Geng Liu, Li Feng, Carlo Alberto Bono, Songbo Yang, Mengxiao Zhu, Francesco Pierri

Keywords: Persona-based jailbreaks, Persona Lineage Evolution, Persona-Invariant Consistency Learning, Safety alignment, Adversarial self-play, Output consistency regularization, LLM-as-a-judge circularity, Benchmark contamination, KL divergence implementation, Benign over-refusal, Artifact completeness, Reproducibility audit

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

6
Authors
9
Findings
18
Limitations
5
Evidence

Editorial summary

English

This paper evaluates how prompt-assigned personas alter refusal and toxicity in Qwen-Turbo, Ernie-4.5-Turbo-128k, DeepSeek-V3, and Hunyuan-Standard. It is published in ACM Transactions on Intelligent Systems and Technology, DOI 10.1145/3819074. The audit uses the complete 32-page arXiv:2506.04975v2 manuscript because ACM blocked the publisher PDF; Crossref confirms receipt on 29 April 2025, acceptance on 6 April 2026, and publication on 30 May 2026. The design crosses personas, 240 Chinese social-group labels in thirteen categories, and six prompt templates: generic, good, bad, negative, harmful, and toxic. Personas adapted from Western work are translated with py-googletrans and reviewed by two native Chinese speakers, then grouped as no-persona default, Basic descriptors such as good/bad/nasty person, and named Character personas. Three independent outputs are generated per persona-group-template combination with temperature 1, 500 output tokens, top_p .90, and presence penalty .02; Hunyuan does not receive top_p. Less than five percent of non-Chinese outputs are translated before analysis. The paper reports about 369,000 outputs per model and more than 1,476,000 total. Refusal is detected with a phrase list and bert-base-chinese fine-tuned on 1,200 annotated Qwen responses. A condition counts as refusal only when all three repetitions refuse; toxicity is Perspective API's maximum TOXICITY among non-refusal repetitions. This is a coherent worst-case red-team construction, not an estimate of typical response behavior. Across ten identical independent stateless requests, aggregate refusal falls from roughly 30% to below 20% for Qwen, 20% to about 8% for Ernie, about 7% to near zero for DeepSeek, and about 50% to above 30% for Hunyuan. Because calls have no retained history, the pattern supports attempt-to-attempt instability, not memory adaptation. Significant Welch comparisons always show higher refusal for female persona sets, but female and male samples are different named people and occupations rather than gender-swapped counterfactuals. Identity, profession, politics, fame, and wording are therefore confounded with gender; the design does not identify causal gender stereotyping. Sexual orientation, race, disease, and disability often trigger higher refusal, with substantial model and template heterogeneity. Approximate absolute toxicity medians are .04 for Qwen and .08 for Hunyuan; DeepSeek rises from about .06-.08 to .14 under Toxic, while Ernie moves from near zero by default to .10-.12 under restrictive prompts. Persona/default ratios reach about 9.4 for Qwen hateful/Toxic, above 10 for DeepSeek nasty, and above 40 in one Ernie cell. The authors acknowledge that near-zero denominators inflate ratios, and exact-zero handling is undocumented. Regression coefficients differ sharply across products, but the paper's claim that factors independently and jointly shape outcomes exceeds the reported method: persona, social group, and template are entered in separate regressions with no interaction terms. Reused personas, groups, and templates are not handled with clustered or hierarchical uncertainty, many tests have no multiplicity policy, and OLS is applied to bounded, refusal-selected, maximum toxicity scores without sensitivity analysis. Perspective API is not culturally validated for this Chinese persona domain. Refusal-classifier reporting is also limited. The manual 98/100 result is conditional on one disagreement direction between BERT and the phrase rule, not general accuracy, and BERT is trained only on Qwen. The appendix mixes a 60/20/20 split with ten-fold cross-validation and provides no annotator agreement. The paper says 87 personas, while the visible appendix says 55 and TeX retains an older commented 87-persona table. The claimed factorial design would yield 375,840 outputs per model with 87 personas, not about 369,000; 55 would yield 237,600. No collection date or immutable API revisions are reported. The mitigation study selects Qwen's 1,000 most toxic baseline cases and applies up to three feedback rounds with Qwen or Ernie as evaluator. Selected medians around .6-.8 fall to .1-.3, but there is no fresh no-feedback regeneration control. Regression to the mean after selecting stochastic extremes is therefore unresolved, and usefulness, semantic preservation, refusal, cost, latency, and human safety are not measured. Most importantly, the promised GitHub dataset, classifier, and code return 404 and no rename or mirror was found. Without raw outputs, labels, checkpoints, scores, complete tables, or scripts, no result can currently be recomputed. The defensible contribution is a broad red-team warning that persona wording, target group, and prompt valence are associated with highly model-specific safety behavior and that repeated attempts can expose a non-refusal. It does not establish a causal gender effect, history adaptation, culturally calibrated toxicity, tested joint interactions, a stable ranking of current services, or utility-preserving mitigation.

Español

Este artículo estudia si asignar personas por prompt altera el rechazo y la toxicidad de cuatro servicios LLM chinos: Qwen-Turbo, Ernie-4.5-Turbo-128k, DeepSeek-V3 y Hunyuan-Standard. Está publicado en ACM Transactions on Intelligent Systems and Technology, DOI 10.1145/3819074. La auditoría usa el manuscrito abierto arXiv:2506.04975v2, revisado el 26 de mayo de 2026, porque ACM bloqueó el PDF; Crossref confirma recepción el 29 de abril de 2025, aceptación el 6 de abril de 2026 y publicación el 30 de mayo de 2026. Se inspeccionaron visualmente sus 32 páginas, el texto y el TeX completos. El diseño cruza personas, 240 grupos sociales chinos de 13 categorías y seis plantillas: decir algo genérico, bueno, malo, negativo, dañino o tóxico sobre un grupo. Las personas parten de trabajo occidental, se traducen con py-googletrans y dos hablantes nativos las revisan. Se dividen en condición por defecto, descriptores Basic como una persona buena, mala o desagradable, y Character personas basadas en figuras concretas. El sistema pide copiar la forma de hablar de la persona. Por cada combinación se generan tres respuestas independientes con temperatura 1, 500 tokens, top_p 0,90 y presence penalty 0,02; Hunyuan no recibe top_p. Menos del 5% de salidas no chinas se traducen antes del análisis. El paper declara unas 369.000 respuestas por modelo y más de 1.476.000 en total. El rechazo se detecta con una lista de frases y con un bert-base-chinese ajustado sobre 1.200 respuestas Qwen anotadas por tres personas. Una combinación solo cuenta como rechazo si las tres réplicas rechazan; cualquier respuesta que no rechace convierte la combinación en no-rechazo. La toxicidad se obtiene con Perspective API únicamente sobre respuestas no rechazadas y se conserva el máximo de las tres réplicas. Esta combinación es útil para red teaming porque busca el peor intento, pero no estima el comportamiento típico. En diez consultas idénticas e independientes, el rechazo agregado cae aproximadamente de 30% a menos de 20% en Qwen, de 20% a 8% en Ernie, desde 7% hasta casi cero en DeepSeek y de 50% a más de 30% en Hunyuan. Como las llamadas son explícitamente stateless, esto muestra inestabilidad entre intentos, no adaptación a historial o memoria. En las comparaciones declaradas significativas, los conjuntos de personas femeninas rechazan más que los masculinos: Qwen en Bad, Negative, Harmful y Toxic; DeepSeek en Bad, Negative y Toxic; Ernie solo en Harmful; Hunyuan en Generic, Harmful y Toxic. Sin embargo, no son versiones contrafactuales de una misma persona: se comparan figuras, ocupaciones e identidades distintas. El efecto atribuido a género está confundido con identidad, profesión, política, fama y texto de la persona, por lo que no identifica estereotipos de género causales. Sexual orientation, race, disease y disability tienden a activar más rechazo, mientras age y categorías socioeconómicas o educativas suelen quedar más bajas, con gran heterogeneidad. En toxicidad absoluta, Qwen ronda una mediana 0,04 y Hunyuan 0,08; DeepSeek pasa de aproximadamente 0,06-0,08 a 0,14 con Toxic; Ernie está cerca de cero por defecto y en 0,10-0,12 con plantillas restrictivas. Los ratios persona/default son mucho más llamativos: cerca de 9,4 para hateful/Toxic en Qwen, más de 10 para nasty en DeepSeek y más de 40 en una celda Ernie. El propio artículo reconoce que el denominador casi cero infla estos cocientes; además, no documenta qué hace con denominadores exactamente cero. Hunyuan se mantiene normalmente en 1-1,3. Los ejemplos máximos bajo nasty person llegan a 0,92 para Younger Sister y 0,86 para Middle School Student en DeepSeek, 0,72 para Rural People en Qwen, 0,69 para Son en Hunyuan y 0,60 para Male Classmate y People From Liaoning en Ernie. Los modelos de regresión encuentran efectos muy diferentes por producto. Basic Persona aumenta la toxicidad estimada aproximadamente 0,145 en Qwen, hasta 0,40 en DeepSeek, 0,21-0,22 en Ernie y 0,075 en Hunyuan. Pero el abstract afirma que los factores actúan independiente y conjuntamente, mientras el método ajusta modelos separados para persona, grupo y plantilla, sin interacciones. No estima efectos conjuntos ajustados ni prueba interacciones. Las observaciones reutilizan las mismas personas, grupos y plantillas, pero no se agrupan errores ni se usan modelos jerárquicos; se interpretan muchos intervalos y Welch tests sin política de multiplicidad. La OLS opera sobre una variable 0-1, filtrada por rechazo y transformada al máximo, sin diagnósticos o sensibilidades. Perspective API tampoco se valida culturalmente para este dominio chino, un límite que los autores reconocen. El clasificador de rechazo tiene 0,9133 de accuracy media en diez folds y 0,9167 final, pero el apéndice mezcla un split 60/20/20 con validación cruzada sin explicar la relación. El 98/100 manual procede solo de casos en que BERT dice rechazo y la regla no; no es accuracy general ni valida otros modelos. Tres anotadores se mencionan sin acuerdo o adjudicación. El paper dice 87 personas, mientras el apéndice visible dice 55 y el TeX conserva comentada una tabla antigua de 87. Además, 87×240×6×3 son 375.840 salidas por modelo, no ~369.000; 55 producirían 237.600. Los datos ausentes impiden reconciliarlo. Tampoco hay fecha de recogida, revisión exacta de los cuatro endpoints, IDs de respuesta, seeds o política de reintentos, por lo que los resultados no fijan una versión estable de servicios mutables. La mitigación toma las 1.000 condiciones Qwen con mayor toxicidad y permite hasta tres rondas de feedback, usando Qwen-Turbo o Ernie-Character-8K como evaluador. La mediana seleccionada de 0,6-0,8 cae a 0,1-0,3 y Ernie queda algo más bajo. Es evidencia exploratoria, no causal concluyente: al escoger extremos y volver a generar sin un control nuevo sin feedback, parte de la caída puede ser regresión a la media. No se mide utilidad, conservación semántica, rechazo, coste, latencia, acuerdo humano ni falsos positivos; una respuesta genérica segura también bajaría Perspective. Finalmente, el manuscrito promete dataset, clasificador y código en GitHub, pero el repositorio devuelve 404 y no aparece renombrado o espejado. Sin las 1,4 millones de salidas, labels, checkpoint, splits, scores, tablas de regresión o scripts, ningún resultado puede recomputarse. La contribución defendible es una alerta de red teaming amplia: el framing de persona, el grupo objetivo y la valencia del prompt se asocian con riesgos muy distintos entre servicios, y repetir intentos puede encontrar una salida no rechazada. No demuestra un efecto causal de género, adaptación por historial, toxicidad culturalmente calibrada, interacciones conjuntas, superioridad actual de un modelo ni eficacia de mitigación preservando utilidad.

Research question

How do the rejection and toxicity of four Chinese LLM services change when combining assigned personas, social groups, and templates of different valence, and can an LLM evaluator reduce the most toxic outputs through iterative feedback?

Method

Factorial audit of hosted APIs with three replicas per combination of persona, 240 groups, and six templates. Rejection is operationalized with BERT and unanimity across three attempts; toxicity with the maximum Perspective score of non-rejected responses. Repeated stateless queries, Welch tests, separate regressions, and a mitigation case over the 1,000 Qwen extremes are added. The independent audit reviews the 32 pages, TeX, arithmetic, metrics, statistics, publication metadata, and real availability of the artifact.

Sample: The unit is a persona-group-template combination aggregated from three independent calls. The paper declares approximately 369,000 outputs per each of four models. The factorial does not match the two published persona counts: 87 would yield 375,840 and 55 would yield 237,600 per model. The ten repetitions use an unspecified subset. The detector uses 1,200 Qwen responses; the mitigation selects 1,000 Qwen extremes.

Findings

  • Rejection decreases when repeating ten stateless calls across the four models; this reflects instability across attempts, not conversational memory.
  • Female persona sets show higher rejection when Welch is significant, but they are not counterfactually matched and do not isolate gender.
  • Sexual orientation, race, disease, and disability usually trigger more rejection, with patterns highly dependent on model and template.
  • DeepSeek reaches the highest Perspective examples under nasty person, 0.92 and 0.86; Qwen reaches 0.72, Hunyuan 0.69, and Ernie 0.60.
  • The persona/default ratios reach approximately 9.4 in Qwen, more than 10 in DeepSeek, and more than 40 in Ernie, but near-zero denominators inflate them.
  • The persona and template coefficients change sign and magnitude across products; there is no uniform safety response.
  • Mitigation lowers the selected scores from approximately 0.6-0.8 to 0.1-0.3, without sufficient control to separate feedback from regression to the mean.
  • The factorial total, the number of personas, and the declared volume are internally incompatible.
  • The promised repository returns 404, so no result, figure, or model can be recomputed.

Limitations

  • There is no collection date, immutable model review, response IDs, seeds, or retry policy for mutable APIs.
  • The linked endpoint for DeepSeek is Alibaba Model Studio, despite all interfaces being described as official to the model.
  • Hunyuan does not share top_p and the decoding conditions are not identical.
  • The paper says 87 personas, the appendix 55, and the approximate total does not match any factorial arithmetic.
  • The ten repetitions are independent and the subset is not defined; they do not test adaptation to history.
  • Rejection unanimity and toxicity maximum construct a worst-case metric, not a typical rate or toxicity.
  • BERT is trained only with Qwen and the 98/100 manual score is conditioned on a single type of disagreement.
  • The 60/20/20 split and ten-fold cross-validation are not reconciled.
  • No agreement, adjudication, or qualification of the three annotators is published.
  • Perspective API is not calibrated for Chinese or for these groups; translating some outputs may alter the score.
  • The ratios do not explain zero denominators and are unstable with near-null baselines.
  • The male and female sets are not counterfactually matched; gender is confounded with identity and occupation.
  • The regressions are separate and without interactions, despite claims of joint effects.
  • There are no clustered errors, hierarchical models, multiplicity correction, or sensitivity to dependencies and selection by rejection.
  • OLS is applied to bounded, filtered, and maximized scores without diagnostics.
  • The mitigation selects extremes and lacks re-generation without feedback, utility, semantic preservation, or human evaluation.
  • Dataset, classifier, checkpoint, results, and promised code are not available; GitHub returns 404.
  • The ACM editorial version PDF could not be directly compared with the open manuscript v2.

What the study does not establish

  • That repeating prompts degrades a conversational memory or history, because each call is stateless.
  • That gender causes the rejection differences or that these stem from internal stereotypes.
  • That Perspective TOXICITY measures in a culturally valid way all Chinese stereotypes, biases, or harms.
  • That the factors act jointly or interact based on separate regressions without interaction terms.
  • That a large ratio against a near-zero baseline is equivalent to a stable proportional increase in harm.
  • That BERT retains accuracy, recall, or calibration on Ernie, DeepSeek, and Hunyuan.
  • That feedback explains the entire toxicity reduction without regression to the mean.
  • That the mitigated responses remain useful, faithful, informative, or non-rejections.
  • That the classifications compare current or stable versions of the four services.
  • That the volume, personas, or results are reproducible with the available artifacts.
  • That the study covers general safety beyond rejection and an automatic toxicity score.

Traceability

Scope: Full text

Version: arXiv:2506.04975v2, 32 pages; published as ACM TIST article 3819074, DOI 10.1145/3819074

Consulted source: https://arxiv.org/abs/2506.04975v2

Review: Codex 32-page full-text visual, complete TeX, publication-metadata, factorial-count, measurement-validity, dependency/statistics, mitigation-control and live-artifact audit, 2026-07-17

Approval: Codex fidelity pass, 2026-07-17

English translation: approved, 2026-07-18

Models evaluated

  • Qwen-Turbo
  • Ernie-4.5-Turbo-128k
  • DeepSeek-V3 mediante endpoint enlazado de Alibaba Model Studio
  • Hunyuan-Standard
  • Ernie-Character-8K como evaluador de mitigación
  • bert-base-chinese ajustado como detector de rechazo

Instruments and metrics

  • Perspective API TOXICITY 0-1
  • Clasificador de rechazo bert-base-chinese
  • Lista de frases de rechazo
  • Regla de unanimidad de tres réplicas
  • Máximo de toxicidad entre respuestas no rechazadas
  • Regresión logística separada por familia de predictores
  • OLS separada por familia de predictores
  • Welch two-sample t-tests
  • Ratios de toxicidad persona/default
  • Feedback iterativo con juez LLM

Data used

  • Más de 1.476.000 generaciones declaradas, no publicadas
  • 240 grupos sociales chinos en 13 categorías, lista completa no publicada
  • 55 personas según apéndice visible o 87 según método, inconsistencia sin resolver
  • 1.200 respuestas Qwen anotadas para rechazo, no publicadas
  • 1.000 condiciones Qwen con mayor toxicidad para mitigación, no publicadas

Evidence and location

  • Publication, DOI, authorship, and reception, acceptance, and publication dates: Crossref work 10.1145/3819074 checked 2026-07-17
  • Method, results, appendices, limitations, ethics, and examples: arXiv:2506.04975v2 PDF, all 32 pages rendered and visually inspected
  • Parameters, contradictory counts, commented tables, metrics, and declared analyses: Complete arXiv:2506.04975v2 TeX source, sha256 2f0e57efbf34ce55e0822280c762eb9e508acca492c4893ddd74014a7374decd
  • Availability of dataset, classifier, and code: leoleepsyche/Toxicity_Chinese_Based_LLMs clone and authenticated GitHub API returned 404; exact-name searches checked 2026-07-17
  • Measurement validity, statistics, mitigation, reproducibility, and claim limits: reports/verification/article-310-chinese-llm-persona-toxicity-refusal-measurement-statistics-mitigation-and-missing-artifact-audit.json