设备端商业意图检索：3 MiB系统与类型化出口边界
原标题：On-Device Commercial Intent Retrieval Under Size, Latency, and Privacy Constraints: A 3 MiB Retrieval System with Typed Egress Boundaries
AI 摘要
该论文研究在设备端进行商业意图检索的可行性，在3 MiB下载体积、20毫秒内推理且不上传原始文本、嵌入或稳定标识符的约束下，构建了基于4比特量化静态嵌入表的检索系统，并测试了联邦学习层。实验显示，在真实韩语商品文本上，当查询包含叶子类目名称时准确率较高，仅含品牌和型号时准确率下降约20个百分点；通过任务适配微调教师模型可部分弥补差距，且设备端可根据排序分数边际拒绝低置信样本以提升准确率。
正文节选
On-Device Commercial Intent Retrieval Under Size, Latency, and Privacy Constraints A 3 MiB Retrieval System with Typed Egress Boundaries Abstract Behavioural ad targeting rests on a stable cross-site identifier, and that assumption is eroding unevenly — blocked in some browsers, consent-gated on mobile, and subject to increasingly strict consent and disclosure requirements. We ask whether the inference can move to the device instead, and report an engineering study of what that costs. Before the