{"id":11763,"date":"2026-09-01T16:14:26","date_gmt":"2026-09-01T08:14:26","guid":{"rendered":"https:\/\/infernews.com\/blog\/official-implementation-of-quot-layerrecall-a-state-conditioned-memory-router-fo\/"},"modified":"2026-09-01T16:14:26","modified_gmt":"2026-09-01T08:14:26","slug":"official-implementation-of-quot-layerrecall-a-state-conditioned-memory-router-fo","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/official-implementation-of-quot-layerrecall-a-state-conditioned-memory-router-fo\/","title":{"rendered":"LayerRecall \u8b93\u9577\u5f71\u7247\u751f\u6210\u8a18\u5f97\u8fd4\u4e4b\u524d\u5605\u89d2\u8272"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/09\/pasted-db44e6d8a572.jpg\" alt=\"Memory Cue Recall animated preview\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u7528 autoregressive \u65b9\u5f0f\u9010\u6bb5\u780c\u9577\u7247\uff0c\u8996\u7a97\u4e00\u6ed1\u904e\u53bb\u5c31\u6703\u9023\u89d2\u8272\u540c\u5834\u666f\u90fd\u4e00\u9f4a\u5514\u898b\u2014\u2014LayerRecall \u5c31\u4fc2\u91dd\u5c0d\u5462\u500b\u75db\u9ede\u3002\u4f62\u5c6c\u65bc\u8a18\u61b6\u8def\u7531\u6a21\u7d44\uff0c\u5514\u4fc2\u65b0\u6a21\u578b\uff0c\u5b9a\u4f4d\u4fc2\u52a0\u55ba\u73fe\u6709\u5f71\u7247 DiT \u4e0a\u9762\u505a\u88dc\u5f37\u3002\u7576\u4e00\u6bb5 chunk \u5165\u9762\u51fa\u73fe\u9700\u8981\u56de\u61b6\u5605\u5143\u7d20\uff08\u4f8b\u5982\u4e3b\u89d2\u3001\u670d\u98fe\u3001\u5834\u666f\uff09\u6642\uff0c\u4f62\u6703\u7528 current-conditioned \u5605\u65b9\u5f0f\u53bb historical candidate pool \u505a cosine retrieval\uff0c\u6435\u5230\u76f8\u95dc\u5605 K\/V state \u4e4b\u5f8c\u518d\u6ce8\u5165\u8fd4\u53bb\u63c0\u904e\u5605\u8a18\u61b6\u654f\u611f\u5c64\uff0c\u4f8b\u5982 4\u30019\u300110\u300112\u300113\u300115\u300116\u300117\u300118\u300126 \u7b49\u3002\u5176\u9918\u672c\u5730 attention \u5c31\u7167\u820a\u884c\uff0c\u5514\u6703\u56e0\u70ba\u88dc\u6b77\u53f2\u8cc7\u6599\u800c\u64fe\u4e82\u5c40\u90e8\u9023\u8cab\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u60f3\u7406\u89e3\u9ede\u89e3\u8981\u5206\u5c64\u6ce8\u5165\uff0c\u7814\u7a76\u767c\u73fe\u5514\u540c layer \u5c0d\u300c\u7576\u4e0b\u300d\u300c\u8fd1\u671f\u300d\u300c\u9060\u8ddd\u300dcontext \u5605\u4f9d\u8cd1\u5dee\u597d\u9060\uff0c\u6240\u4ee5\u4e82\u9935\u6b77\u53f2 token \u53cd\u800c\u6703\u5bb3\u5497\u672c\u5730\u4e00\u81f4\u6027\u3002\u914d\u5408\u5605 KV cache budget \u5927\u7d04\u4fc2 80 \u500b latent frames\uff0c\u672c\u5730 attention 32 \u5e40\uff0c\u7576\u4e2d 8 \u5e40\u505a sink\u30018 \u5e40\u505a current chunk\u3002\u8a13\u7df4\u4e0a\u4ef2\u7528\u5497 Cross-Horizon Prediction Matching\uff08CHPM\uff09\uff0c\u7531\u4e00\u6bb5\u9577 context reference \u55ba prediction space \u76e3\u7763\u5462\u500b\u6709\u754c\u5605 router\uff0c\u7b49\u4f62\u55ba\u7a00\u7f3a\u9577\u7247\u6578\u64da\u4e0b\u90fd\u5b78\u5230\u5e7e\u6642\u53bb recall \u4e5c\u5622\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">100 \u689d\u591a shot prompt \u8a55\u4f30\u5165\u9762\uff0cLayerRecall \u55ba MemoBench \u540c MovieBench \u651e\u5230\u6700\u4f73\u7d9c\u5408\u7d50\u679c\uff0cVBench-Long \u7dad\u6301\u8fd4 backbone \u6c34\u6e96\uff0c\u53ef\u898b\u9577\u8ddd\u96e2\u56de\u5fa9\u505a\u5f97\u597d\u800c\u672c\u5730\u9023\u8cab\u7121\u53d7\u72a7\u7272\u3002\u8ad6\u6587\u4ef2\u5c55\u793a\u5497 memory-guided self-correction\uff0c\u6a21\u578b\u6703\u4e3b\u52d5\u7cfe\u6b63\u8fd4\u4e4b\u524d\u5605\u932f\u8655\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u91cd\u9ede\u6458\u8981<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5c6c\u65bc\u8a18\u61b6\u8def\u7531\u6a21\u7d44\uff0c\u8a2d\u8a08\u4e0a\u8981\u55ba\u73fe\u6709\u5f71\u7247 DiT \u4e0a\u52a0\u639b<\/li>\n<li>\u53ea\u55ba memory-sensitive \u5c64\u6ce8\u5165\u6b77\u53f2 K\/V\uff0c\u5176\u4ed6\u5c64\u4fdd\u6301\u672c\u5730 attention<\/li>\n<li>\u8a13\u7df4\u7528 CHPM\uff0c\u7528\u9577 context reference \u55ba prediction space \u76e3\u7763 router<\/li>\n<li>MemoBench \u540c MovieBench \u7d9c\u5408\u8868\u73fe\u9818\u5148\uff0cVBench-Long \u7dad\u6301 backbone \u6c34\u6e96<\/li>\n<li>\u9069\u5408\u505a\u591a\u93e1\u982d\u3001\u8de8\u6642\u9593\u89d2\u8272\u540c\u5834\u666f\u4e00\u81f4\u6027\u5605\u9577\u5f71\u7247\u751f\u6210\u5de5\u4f5c<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5462\u985e\u5de5\u5177\u5c0d\u505a\u9577\u7247\u5275\u4f5c\u3001\u5ee3\u544a\u5206\u93e1\u6216\u8005\u89d2\u8272\u9700\u8981\u53cd\u8986\u51fa\u5834\u5605\u5718\u968a\u6703\u5e7e\u5be6\u7528\uff1b\u7814\u7a76\u4eba\u54e1\u4ea6\u53ef\u4ee5\u7747\u4f62\u9ede\u6a23\u5c07\u300cselective injection\u300d\u5462\u500b\u53d6\u6368\u7531\u555f\u767c\u5f0f\u8b8a\u6210\u53ef\u5b78\u7fd2\u8def\u7531\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/yixuan-ding-zju.github.io\/LayerRecall_Web\/\" rel=\"noopener noreferrer\"><strong>\u9805\u76ee\u4e3b\u9801<\/strong><\/a> \u00b7 <a href=\"https:\/\/github.com\/Yixuan-Ding-ZJU\/LayerRecall\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u9577\u5f71\u7247 AI \u6700\u6015\u9694\u5497\u4e00\u6bb5\u5c31\u8a8d\u5514\u8fd4\u89d2\u8272\u540c\u5834\u666f\uff0cLayerRecall \u7528\u8a18\u61b6\u8def\u7531\u52a0\u9078\u64c7\u6027\u6ce8\u5165\uff0c\u53ea\u55ba\u654f\u611f\u5c64\u88dc\u8fd4\u6b77\u53f2 K\/V\uff0c\u6548\u679c\u8d0f\u904e MemoBench \u540c MovieBench\u3002<\/p>\n","protected":false},"author":8,"featured_media":11762,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"wpai_generated_summary":"","footnotes":""},"categories":[133,163,169,141],"tags":[],"class_list":["post-11763","post","type-post","status-publish","format-standard","hentry","category-133","category-163","category-169","category-141"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11763","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=11763"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/11763\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/11762"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=11763"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=11763"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=11763"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}