mlsanityData to 14 Oct 2026

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

arXiv 2506.04975 ↗First posted 5 Jun 2025cs.CYcited by 2

On the map: LLM safety and security → Safety alignment, refusal and guardrails

Abstract (arXiv)

Recent research has highlighted that assigning specific personas to large language models (LLMs) can significantly increase harmful content generation. However, limited attention has been given to persona-driven toxicity in non-Western contexts, particularly in Chinese-based LLMs. In this paper, we perform a large-scale, cross-model analysis of refusal behavior and persona-driven toxicity amplification across four Chinese LLMs, leveraging a comprehensive dataset of over 1,400,000 generated texts. We identify significant disparities in persona-driven refusal behavior, including systematic gender differences in refusal triggering across the evaluated Chinese LLMs. Furthermore, we provide quantitative evidence of persona-driven toxicity amplification with respect to model default baselines. We show that this amplification--whose magnitude varies substantially across models--is driven by interactions across several factors, involving persona conditioning, prompting strategy, target social group, and model-specific safety mechanisms. Leveraging model-specific regression analyses, we systematically characterize how persona categories, target social groups, and prompt templates independently and jointly shape both refusal behavior and output toxicity. As a complementary case study, we further explore an iterative, evaluator-guided mitigation strategy based on model feedback with an external LLM evaluator, demonstrating that highly toxic outputs can be substantially reduced without costly model retraining. Overall, our findings highlight the importance of culturally contextualized safety evaluations for Chinese-language LLMs and provide a structured framework for assessing persona-induced risks and exploratory mitigation strategies in LLM-generated content.

Key facts. Cited by 2 papers in the map (data through 14 Oct 2026). Placed in Safety alignment, refusal and guardrails, one of 12 directions in LLM safety and security. Builds on 39 of its 63 references that are in the map. First posted on arXiv 5 Jun 2025.

Builds on

39 of 63 references.

Closest ancestors · 2

Same direction, newest first

  1. 2024Refusal in Language Models Is Mediated by a Single Direction17 Jun
  2. 2024SafetyPrompts: a Systematic Review of Open Datasets for Evaluating and Improving Large Language Model Safety8 Apr

Other references · 37

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,399
  2. 2023Generative Agents: Interactive Simulacra of Human Behaviorcited by 3,568
  3. 2023Jailbroken: How Does LLM Safety Training Fail?cited by 1,651
  4. 2020Don't Stop Pretraining: Adapt Language Models to Domains and Taskscited by 1,565
  5. 2020RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Modelscited by 1,255
  6. 2023Bias and Fairness in Large Language Models: A Surveycited by 639
  7. 2020GeDi: Generative Discriminator Guided Sequence Generationcited by 378
  8. 2023Toxicity in ChatGPT: Analyzing Persona-assigned Language Modelscited by 360
  9. 2022Social Simulacra: Creating Populated Prototypes for Social Computing Systemscited by 301
  10. 2021Towards Understanding and Mitigating Social Biases in Language Modelscited by 296
  11. 2024Large Language Models for Mathematical Reasoning: Progresses and Challengescited by 284
  12. 2023From Pretraining Data to Language Models to Downstream Tasks: Tracking the Trails of Political Biases Leading to Unfair NLP Modelscited by 260
25 more
  1. 2024Two Tales of Persona in LLMs: A Survey of Role-Playing and Personalizationcited by 208
  2. 2023Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMscited by 199
  3. 2023Gender bias and stereotypes in Large Language Modelscited by 187
  4. 2024Agent AI: Surveying the Horizons of Multimodal Interactioncited by 170
  5. 2023Don't Trust ChatGPT when Your Question is not in English: A Study of Multilingual Abilities and Types of LLMscited by 104
  6. 2024Survey of Cultural Awareness in Language Models: Text and Beyondcited by 103
  7. 2023CoMPosT: Characterizing and Evaluating Caricature in LLM Simulationscited by 97
  8. 2023Conversational Health Agents: A Personalized LLM-Powered Agent Frameworkcited by 92
  9. 2022Solving Math Word Problems via Cooperative Reasoning induced Language Modelscited by 77
  10. 2022COLD: A Benchmark for Chinese Offensive Language Detectioncited by 76
  11. 2023Can GPT-3 Perform Statutory Reasoning?cited by 73
  12. 2024Against The Achilles' Heel: A Survey on Red Teaming for Generative Modelscited by 59
  13. 2021A Plug-and-Play Method for Controlled Text Generationcited by 56
  14. 2023Are LLMs All You Need for Task-Oriented Dialogue?cited by 54
  15. 2024Towards Implicit Bias Detection and Mitigation in Multi-Agent LLM Interactionscited by 51
  16. 2023Large Language Models are Built-in Autoregressive Search Enginescited by 27
  17. 2023Toxic Bias: Perspective API Misreads German as More Toxiccited by 21
  18. 2022Reward Modeling for Mitigating Toxicity in Transformer-based Language Modelscited by 18
  19. 2024Realistic Evaluation of Toxicity in Large Language Modelscited by 15
  20. 2024Efficient Detection of Toxic Prompts in Large Language Modelscited by 13
  21. 2024How Are LLMs Mitigating Stereotyping Harms? Learning from Search Engine Studiescited by 10
  22. 2024DeMod: A Holistic Tool with Explainable Detection and Personalized Modification for Toxicity Censorshipcited by 7
  23. 2024Chinese SafetyQA: A Safety Short-form Factuality Benchmark for Large Language Modelscited by 7
  24. 2025Characterizing Bias: Benchmarking Large Language Models in Simplified versus Traditional Chinesecited by 7
  25. 2024Comparing diversity, negativity, and stereotypes in Chinese-language AI technologies: an investigation of Baidu, Ernie and Qwencited by 4

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 1; 2026: 1.

2 papers.

  1. 2025Synthetic Socratic Debates: Examining Persona Effects on Moral Decision and Persuasion Dynamicscited by 19
  2. 2026The Shadow Self: Intrinsic Value Misalignment in Large Language Model Agentscited by 1

Similar, not cited

Close in wording, no citation either way

  1. 2026Persona Prompting as a Lens on LLM Social Reasoning
  2. 2026Persona Non Grata: Single-Method Safety Evaluation Is Incomplete for Persona-Imbued LLMs
  3. 2025Characterizing Selective Refusal Bias in Large Language Models
  4. 2026What Models Express, Suppress, and Resist: Auditing Open-Weight LLMs with Persona Vectors
  5. 2026Do Linear Probes Generalize Better in Persona Coordinates?
  6. 2026The Refusal--Compliance Tradeoff: A Large-Scale Safety Behavior Audit of Large Language Models
  7. 2025Ideology-Based LLMs for Content Moderation
  8. 2026How Narrative Wrapping Affects LLM Refusal: A Cross-Language Benchmark and Defense