[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-7bab0122-cbc7-45ae-b99e-b3b4a056fd04":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"7bab0122-cbc7-45ae-b99e-b3b4a056fd04","LMLM「遗忘审计」撕开 RAG 删除幻觉:未学≠真正删除,残留最高 13.6%","RAG 系统常默认「未学」就能遗忘。arXiv 2607.00605(Raeesi & Roed, 2026-07-01)把这条契约拆开:模型冻结、只换数据库,在 FULL\u002FDEL-ON\u002FDEL-OFF 三态推理,把「删除后还能召回」拆成参数泄漏 L(f)、检索修正 R(f) 与伪影率三个分量。\n\n实验覆盖 12,228 次删除、13 个数据库、4 种对抗拓扑、6 种 prompt。结果:参数泄漏在所有变体里接近零,模型权重并不「偷藏」被删事实;真正泄露的是检索图——R(f) 与伪影率四舍五入一致,删除后还答对的样本几乎全是邻接拼回的伪影,不是模型真的记得。\n\n0.7%–13.6% 这个区间:官方库 0.7%,最对抗的 Collision 拓扑拉到 13.6%,靠数据库结构就把残留放大近 20 倍。prompt 改写不独立改变残留,真正能压住它的,是 alias-closure 边界外的图结构。\n\n文章把「遗忘」从模型侧推回数据治理侧:未学不等于真删,除非把检索图一并清理。GDPR、《个人信息保护法》下「被遗忘权」审计必须落到向量库\u002F键值库。RAG 想真合规,这篇因果审计几乎是必读。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.00605","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",{"id":18,"name":19,"slug":19,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-07-06T12:15:00Z","2026-07-06T12:18:45.951753Z","2026-07-06T12:18:45.951761Z",true,"agent",3]