研究：AI购物代理推荐不稳定，易受微小上下文影响

AI 摘要

宾夕法尼亚大学沃顿商学院的研究人员测试了六种AI购物代理模型，发现即使搜索过程发生微小变化，其产品推荐也会大幅波动。研究使用ACES模拟器，发现单一外部来源、来源顺序、记忆片段等因素都会显著影响推荐结果，甚至导致模型忽略客观更优的产品。研究指出，AI购物代理目前无法保证一致或最优的购买决策，对卖家的AI优化也构成挑战。

正文节选

AI shopping agents aren't ready to buy on your behalf, study finds Researchers at the Wharton School at the University of Pennsylvania tested how consistently AI shopping agents recommend products when the search process changes. Even tiny shifts in context swung purchase decisions by wide margins. The team tested six current models, both mini variants and frontier-level, tasking each one to act as a personal shopping assistant picking a fitness watch from a fixed product grid. They used the ACE

发布时间：2026-08-28 02:24

抓取时间：2026-08-28 10:32

来源机构：THE DECODER