arXiv · cs.AI · Kainan Zhou, Gangzhen Qian, Janet Sung, Zhi Jing, Rong Bao, Qianru Lao · 10 小时前模型评分43
SearchInject：控制检索暴露以防御语言模型搜索中的间接提示注入
SearchInject: Retrieval Exposure Control for Indirect Prompt Injection in Language-Model Search
导读
SearchInject 通过综合检索相关性、搜索滥用信号、指令可能性及两项相关性交互的联合评分，控制携带间接提示注入的页面是否进入模型上下文。在 300 条 SciFact 查询上，BM25 的 top-10 目标暴露率从 0.9408 降至 0.1583，三种检索器的相对降幅为 83.2%-83.7%，而干净检索的 nDCG@10 最大绝对变化仅 0.0018。
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
- “SearchInject controls this admission decision with a joint score combining retrieval relevance, search-abuse cues, instruction likelihood, and two relevance interactions.”
- “A quarantine threshold is selected under an empirical 0.5% clean-candidate false-quarantine constraint on a calibration split.”
- “The frozen threshold yields clean test false-quarantine rates of 0.59-0.65%, above the empirical calibration constraint.”
- “The evidence supports retrieval exposure control within the evaluated construction families; the benefit of the interaction terms at matched clean cost remains unresolved.”
来源：arXiv · cs.AI · arxiv.org