研究发现LLM具有模型专属的偏好人名：'Elena Vasquez'和'Marcus Chen'成为Claude生成内容的强特征
英文摘要
Researchers discovered that large language models exhibit strong, model-specific and version-specific priors over character names. The names 'Elena Vasquez' and 'Marcus Chen' frequently appear as a correlated ensemble across dozens of websites in diverse roles, including volcano experts, podcast hosts, thriller protagonists, and authors of 1,000+ papers published in two months, making them a reliable signal that content was generated by Claude. The team identified a third name in the ensemble, further solidifying the fingerprint. The finding emerged as a side observation from a model diffing method (CDD) and grew into a standalone paper (arXiv:2606.02184).
中文摘要
研究人员发现大型语言模型对角色名称有强烈的、模型和版本特定的先验偏好。名字'Elena Vasquez'和'Marcus Chen'作为一个关联集合频繁出现在数十个网站上，扮演火山专家、播客主持人、惊悚小说主角以及两个月内发表1000多篇论文的作者等多样角色，成为Claude生成内容的可靠信号。团队还发现了该集合中的第三个名字，进一步强化了指纹特征。这一发现源于一种模型差异方法（CDD）的副产品，并发展为独立论文（arXiv:2606.02184）。
关键要点
LLMs generate model-specific name ensembles that act as hidden fingerprints.
LLM生成模型专属的名字集合，成为隐藏的指纹。
'Elena Vasquez' and 'Marcus Chen' strongly correlate with Claude outputs across many websites.
'Elena Vasquez'和'Marcus Chen'在大量网站上与Claude输出强相关。
A third name in the ensemble was later identified.
后来识别出该集合中的第三个名字。
The discovery originated from a model diffing method (CDD) as a side finding.
发现来自一种模型差异方法（CDD）的副产品。