arXiv · cs.AI · Tamunotonye Harry, Yuanyuan Feng · 5 小时前模型评分52
论文指出固定人设的合成人群无法复现情境相关行为
Personas Are Not People: Identity as State, Not Attribute
导读
Tamunotonye Harry 和 Yuanyuan Feng 的论文指出，用固定人设（persona，即写死的性格、人口特征和态度档案）构建的合成人群，无法复现人群中随情境变化的那部分行为，人设再详细、底层模型再大也不行。
分类置信度
中 · 待人工复核
- 只有单一媒体来源，需要人工复核。
来源级别：媒体与个人
事件类型：研究发布
证据摘录
- “We propose a new evaluation criterion based on this result and a person-in-context design that can meet it in principle.”
- “We then prove, under two stated assumptions, that any simulation built on fixed personas, no matter how detailed the personas or how large the underlying model, cannot reproduce the part of a population's behavior that is specific to it and its circumstances.”
- “The design does produce context-dependent behavior when tested directly, but in neither setting does it fully match the real population, and we show where it falls short.”
来源：arXiv · cs.AI · arxiv.org