多维偏好建模中的多维偏差：评估合成代理替代人类参与者的能力
原标题：Multi-dimensional Bias in Modeling Multi-dimensional Preferences: Evaluating the Ability of Synthetic Agents to Replace Human Participants in Conjoint Experiments
AI 摘要
该研究通过复现已发表的联合分析实验，评估了合成代理（LLM）在再现多维人类偏好方面的能力。结果显示，合成代理在边际分布和方向一致性上表现尚可，但在联合分布、个体选择对齐和效应量精确度上表现不佳，表明其有效性依赖于具体声明且具有层级性，不能作为人类样本的可靠替代。
正文节选
[datatype=bibtex] \map \step[fieldsource=doi, final] \step[fieldset=url, null] Multi-dimensional Bias in Modeling Multi-dimensional Preferences: Evaluating the Ability of Synthetic Agents to Replace Human Participants in Conjoint Experiments Thanks: The authors would like to thank Professor Ray Duch, Dr. Jamie Cummins, Dr. Alice Malmberg, and panel participants of IMEBESS 2026 for their valuable feedback. We would also like to thank Professor Ben Ansell, Dr. Noah Bacine, and Anne-Charlotte Gimen