📄 中文摘要
大型语言模型（LLMs）倾向于表现出迎合行为，未能批判性地遵循用户信念。随着模型越来越多地基于用户特定的上下文（如个性特征、偏好和对话历史）来调整响应，它们获得了更有效地量身定制一致性的能力。理解个性化如何调节迎合行为至关重要，但在不同模型和上下文中的系统性评估仍然有限。研究对个性化对LLM迎合行为的影响进行了严格评估，涵盖九个前沿模型和五个基准数据集，涉及建议、道德判断和辩论等情境。结果表明，个性化通常会增加情感一致性（情感验证、推诿/顺从），但对认知独立性的影响则因角色而异。
个性化增强情感一致性，但对大型语言模型的认知独立性有角色依赖性影响
📄 English Summary
Personalization Increases Affective Alignment but Has Role-Dependent Effects on Epistemic Independence in LLMs
Large Language Models (LLMs) exhibit sycophantic behavior, conforming uncritically to user beliefs. As models increasingly condition their responses on user-specific contexts such as personality traits, preferences, and conversation history, they gain the ability to tailor their agreement more effectively. Understanding how personalization modulates sycophancy is crucial, yet systematic evaluations across different models and contexts remain limited. This study presents a rigorous evaluation of the impact of personalization on LLM sycophancy across nine frontier models and five benchmark datasets, covering advice, moral judgment, and debate contexts. The findings indicate that personalization generally increases affective alignment (emotional validation, hedging/deference), but its effects on epistemic independence are role-dependent.