{"id":96262,"date":"2026-09-10T09:59:40","date_gmt":"2026-09-10T06:59:40","guid":{"rendered":"https:\/\/twodots.gr\/?p=96262"},"modified":"2026-09-10T09:59:42","modified_gmt":"2026-09-10T06:59:42","slug":"xkv-ai-agents-latent-cache-communication","status":"publish","type":"post","link":"https:\/\/twodots.gr\/fr\/xkv-ai-agents-latent-cache-communication\/","title":{"rendered":"XKV : quand les agents IA communiquent sans \u00e9crire de messages"},"content":{"rendered":"<div class=\"td-article-lede\">\n<p><strong>\u03a4\u03bf XKV \u03b4\u03bf\u03ba\u03b9\u03bc\u03ac\u03b6\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03c1\u03b9\u03b6\u03b9\u03ba\u03ac \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03c3\u03c4\u03bf \u03b1\u03ba\u03c1\u03b9\u03b2\u03cc handoff \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 AI agents:<\/strong> \u03b1\u03bd\u03c4\u03af \u03bf \u03ad\u03bd\u03b1\u03c2 agent \u03bd\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03ba\u03b1\u03b9 \u03bf \u03ac\u03bb\u03bb\u03bf\u03c2 \u03bd\u03b1 \u03c4\u03bf \u03b4\u03b9\u03b1\u03b2\u03ac\u03b6\u03b5\u03b9 \u03be\u03b1\u03bd\u03ac, \u03ad\u03bd\u03b1\u03c2 \u03bc\u03b9\u03ba\u03c1\u03cc\u03c2 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c3\u03b9\u03bc\u03bf\u03c2 translator \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03b1\u03c0\u03b5\u03c5\u03b8\u03b5\u03af\u03b1\u03c2 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03c4\u03b9\u03c2 key-value caches \u03b4\u03cd\u03bf \u03c0\u03b1\u03b3\u03c9\u03bc\u03ad\u03bd\u03c9\u03bd \u03b3\u03bb\u03c9\u03c3\u03c3\u03b9\u03ba\u03ce\u03bd \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03c9\u03bd.<\/p>\n<p>\u03a3\u03c4\u03b1 45 dataset\u2013model-pair settings \u03c4\u03b7\u03c2 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1\u03c2, \u03c4\u03bf XKV \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 \u03c4\u03bf \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc macro score, \u03bc\u03b5\u03c4\u03ad\u03c6\u03c1\u03b1\u03b6\u03b5 \u03ad\u03bd\u03b1 \u03b6\u03b5\u03cd\u03b3\u03bf\u03c2 caches \u03c3\u03b5 5,8 ms \u03b1\u03bd\u03c4\u03af 59,9 ms \u03b3\u03b9\u03b1 \u03c4\u03bf LCF-X \u03ba\u03b1\u03b9 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b5 76,1% \u03bb\u03b9\u03b3\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c3\u03b9\u03bc\u03b5\u03c2 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03bf\u03c5\u03c2. \u03a4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03b9\u03ba\u03cc \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b5\u03bd\u03cc: \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03b7 latent \u03b5\u03c0\u03b9\u03ba\u03bf\u03b9\u03bd\u03c9\u03bd\u03af\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03b7 \u03ba\u03b1\u03b9 \u03b1\u03bd\u03c4\u03b1\u03b3\u03c9\u03bd\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03c3\u03b5 split-evidence reasoning, \u03cc\u03c7\u03b9 \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 multi-agent workflow \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03b3\u03ba\u03b1\u03c4\u03b1\u03bb\u03b5\u03af\u03c8\u03b5\u03b9 \u03c4\u03b1 \u03b1\u03bd\u03b1\u03b3\u03bd\u03ce\u03c3\u03b9\u03bc\u03b1 \u03bc\u03b7\u03bd\u03cd\u03bc\u03b1\u03c4\u03b1.<\/p>\n<\/div>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">Sommaire<\/div>\n<ul>\n<li><a href=\"#akrivo-text-handoff\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf text handoff \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03ba\u03c1\u03b9\u03b2\u03cc<\/a><\/li>\n<li><a href=\"#c2c-lcfx-xkv\">\u0391\u03c0\u03cc \u03c4\u03bf C2C \u03ba\u03b1\u03b9 \u03c4\u03bf LCF-X \u03c3\u03c4\u03bf XKV<\/a><\/li>\n<li><a href=\"#ti-kai-pou\">\u0397 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03c4\u03bf\u03c5 XKV: \u00ab\u03c4\u03b9\u00bb \u03ba\u03b1\u03b9 \u00ab\u03c0\u03bf\u03cd\u00bb<\/a><\/li>\n<li><a href=\"#eterogeni-montela\">\u03a0\u03ce\u03c2 \u03b3\u03b5\u03c6\u03c5\u03c1\u03ce\u03bd\u03b5\u03b9 \u03b5\u03c4\u03b5\u03c1\u03bf\u03b3\u03b5\u03bd\u03ae \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1<\/a><\/li>\n<li><a href=\"#peirama-45-settings\">\u03a4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03c4\u03c9\u03bd 45 settings<\/a><\/li>\n<li><a href=\"#apotelesmata-poiotitas\">\u03a4\u03b9 \u03b4\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03c4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2<\/a><\/li>\n<li><a href=\"#latency-parametroi\">Latency \u03ba\u03b1\u03b9 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c3\u03b9\u03bc\u03b5\u03c2 \u03c0\u03b1\u03c1\u03ac\u03bc\u03b5\u03c4\u03c1\u03bf\u03b9<\/a><\/li>\n<li><a href=\"#ablations\">Ce que r\u00e9v\u00e8lent les ablations<\/a><\/li>\n<li><a href=\"#enterprise-architektoniki\">\u03a4\u03b9 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1 enterprise AI<\/a><\/li>\n<li><a href=\"#auditability-asfaleia\">Auditability, \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1 \u03ba\u03b1\u03b9 fallback<\/a><\/li>\n<li><a href=\"#pilot-axiologisis\">\u0388\u03bd\u03b1 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc pilot \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7\u03c2<\/a><\/li>\n<li><a href=\"#ti-kratame\">\u03a4\u03b9 \u03ba\u03c1\u03b1\u03c4\u03ac\u03bc\u03b5 \u03b1\u03c0\u03cc \u03c4\u03bf XKV<\/a><\/li>\n<\/ul>\n<\/div>\n<h2 id=\"akrivo-text-handoff\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf text handoff \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03ba\u03c1\u03b9\u03b2\u03cc<\/h2>\n<p>\u0397 \u03c6\u03c5\u03c3\u03b9\u03ba\u03ae \u03b3\u03bb\u03ce\u03c3\u03c3\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf \u03c0\u03b9\u03bf \u03c6\u03bf\u03c1\u03b7\u03c4\u03cc interface \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1. \u0394\u03b5\u03bd \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af \u03ba\u03bf\u03b9\u03bd\u03cc tokenizer, \u03af\u03b4\u03b9\u03bf \u03b1\u03c1\u03b9\u03b8\u03bc\u03cc layers \u03ae \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7 \u03c3\u03c4\u03b9\u03c2 \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ad\u03c2 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03b9\u03c2. \u0395\u03af\u03bd\u03b1\u03b9 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03b1\u03bd\u03b1\u03b3\u03bd\u03ce\u03c3\u03b9\u03bc\u03b7 \u03b1\u03c0\u03cc \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03bf\u03c5\u03c2, \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03c5\u03c4\u03b5\u03af \u03c3\u03b5 logs \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b9\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03c3\u03b5 \u03bc\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03bd\u03b1 \u03b5\u03be\u03b5\u03c4\u03ac\u03c3\u03b5\u03b9 \u03c4\u03b9 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ce\u03c2 \u03b9\u03c3\u03c7\u03c5\u03c1\u03af\u03c3\u03c4\u03b7\u03ba\u03b5 \u03bf sender.<\/p>\n<p>\u0397 \u03b5\u03c5\u03ba\u03bf\u03bb\u03af\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2. \u039f sender \u03ba\u03ac\u03bd\u03b5\u03b9 autoregressive decoding \u03c4\u03bf\u03c5 \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf\u03c5 \u03bc\u03b7\u03bd\u03cd\u03bc\u03b1\u03c4\u03bf\u03c2, \u03bf receiver \u03ba\u03ac\u03bd\u03b5\u03b9 tokenization \u03ba\u03b1\u03b9 prefill \u03c4\u03bf\u03c5 \u03af\u03b4\u03b9\u03bf\u03c5 \u03c0\u03b5\u03c1\u03b9\u03b5\u03c7\u03bf\u03bc\u03ad\u03bd\u03bf\u03c5 \u03ba\u03b1\u03b9 \u03c4\u03bf \u03bc\u03ae\u03bd\u03c5\u03bc\u03b1 \u03b3\u03c1\u03ac\u03c6\u03b5\u03c4\u03b1\u03b9 \u03c7\u03c9\u03c1\u03af\u03c2 \u03ac\u03bc\u03b5\u03c3\u03b7 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1 \u03c4\u03bf\u03c5 \u03c4\u03b9 \u03ad\u03c7\u03b5\u03b9 \u03ae\u03b4\u03b7 \u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b5\u03b9 \u03bf \u03c0\u03b1\u03c1\u03b1\u03bb\u03ae\u03c0\u03c4\u03b7\u03c2. \u038c\u03c4\u03b1\u03bd \u03c4\u03bf handoff \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf \u03bc\u03bf\u03bd\u03bf\u03c0\u03ac\u03c4\u03b9 \u03bc\u03b9\u03b1\u03c2 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7\u03c2, \u03b1\u03c5\u03c4\u03cc \u03c4\u03bf decode\u2013encode loop \u03c0\u03c1\u03bf\u03c3\u03b8\u03ad\u03c4\u03b5\u03b9 latency \u03c0\u03c1\u03b9\u03bd \u03be\u03b5\u03ba\u03b9\u03bd\u03ae\u03c3\u03b5\u03b9 \u03b7 \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ae \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03c4\u03bf\u03c5 \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf\u03c5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5.<\/p>\n<aside class=\"td-article-note\">\n<p><strong>\u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03b5\u03b9 cross-context \u03c3\u03c5\u03bd\u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1:<\/strong> \u03bf Sharer \u03ba\u03b1\u03b9 \u03bf Receiver \u03ad\u03c7\u03bf\u03c5\u03bd \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03c9\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac, \u03bc\u03b7 \u03b5\u03c0\u03b9\u03ba\u03b1\u03bb\u03c5\u03c0\u03c4\u03cc\u03bc\u03b5\u03bd\u03b1 \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03b1 \u03ba\u03b1\u03b9 \u03ba\u03b1\u03bd\u03ad\u03bd\u03b1 context \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf\u03c5. \u0397 \u03b1\u03bd\u03b1\u03bb\u03bf\u03b3\u03af\u03b1 \u03bc\u03b5 CRM, policies, \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03b1 \u03ae \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03b5\u03c1\u03b3\u03b1\u03bb\u03b5\u03af\u03c9\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03c3\u03c4\u03b9\u03ba\u03ae \u03b5\u03c1\u03bc\u03b7\u03bd\u03b5\u03af\u03b1\u00b7 \u03c4\u03b1 \u03c0\u03b5\u03b9\u03c1\u03ac\u03bc\u03b1\u03c4\u03b1 \u03ad\u03b3\u03b9\u03bd\u03b1\u03bd \u03c3\u03b5 \u03c0\u03ad\u03bd\u03c4\u03b5 split-evidence reasoning datasets.<\/p>\n<\/aside>\n<p>\u03a4\u03bf \u03c0\u03c1\u03cc\u03b2\u03bb\u03b7\u03bc\u03b1 \u03c3\u03c5\u03bd\u03b4\u03ad\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03bf\u03bd \u03b5\u03c5\u03c1\u03cd\u03c4\u03b5\u03c1\u03bf \u03c3\u03c7\u03b5\u03b4\u03b9\u03b1\u03c3\u03bc\u03cc <a href=\"https:\/\/twodots.gr\/multi-agent-ai-concurrency-control\/\">multi-agent \u03c3\u03c5\u03c3\u03c4\u03b7\u03bc\u03ac\u03c4\u03c9\u03bd \u03bc\u03b5 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf \u03ba\u03bf\u03b9\u03bd\u03cc state<\/a>. \u0397 \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03b7 \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03bc\u03cc\u03bd\u03bf \u03cc\u03c4\u03b1\u03bd ownership, \u03c7\u03c1\u03bf\u03bd\u03b9\u03ba\u03ae \u03c3\u03b5\u03b9\u03c1\u03ac \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b9\u03c4\u03c1\u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03b5\u03c2 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b5\u03c2 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03bf\u03c5\u03bd \u03c3\u03b1\u03c6\u03ae.<\/p>\n<h2 id=\"c2c-lcfx-xkv\">\u0391\u03c0\u03cc \u03c4\u03bf C2C \u03ba\u03b1\u03b9 \u03c4\u03bf LCF-X \u03c3\u03c4\u03bf XKV<\/h2>\n<p>\u039f\u03b9 \u03c0\u03c1\u03bf\u03b7\u03b3\u03bf\u03cd\u03bc\u03b5\u03bd\u03b5\u03c2 cache-to-cache \u03c0\u03c1\u03bf\u03c3\u03b5\u03b3\u03b3\u03af\u03c3\u03b5\u03b9\u03c2 \u03b1\u03c0\u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03cc\u03c4\u03b9 \u03b4\u03cd\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03b1\u03bd\u03c4\u03b1\u03bb\u03bb\u03ac\u03be\u03bf\u03c5\u03bd \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc tokens. \u03a4\u03bf C2C \u03c5\u03c0\u03bf\u03c3\u03c4\u03b7\u03c1\u03af\u03b6\u03b5\u03b9 \u03b5\u03c4\u03b5\u03c1\u03bf\u03b3\u03b5\u03bd\u03ae \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1, \u03b1\u03bb\u03bb\u03ac \u03b5\u03c5\u03b8\u03c5\u03b3\u03c1\u03b1\u03bc\u03bc\u03af\u03b6\u03b5\u03b9 \u03b8\u03ad\u03c3\u03b5\u03b9\u03c2 \u03c0\u03bf\u03c5 \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03c5\u03c0\u03bf\u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf input. \u0391\u03c5\u03c4\u03cc \u03b4\u03b5\u03bd \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 \u03ad\u03bd\u03b1 handoff \u03cc\u03c0\u03bf\u03c5 \u03bf\u03b9 agents \u03ad\u03c7\u03bf\u03c5\u03bd \u03b4\u03b5\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac documents.<\/p>\n<p>\u03a4\u03bf LCF-X \u03b1\u03c6\u03b1\u03af\u03c1\u03b5\u03c3\u03b5 \u03c4\u03b7\u03bd \u03b1\u03c0\u03b1\u03af\u03c4\u03b7\u03c3\u03b7 \u03ba\u03bf\u03b9\u03bd\u03bf\u03cd context \u03bc\u03ad\u03c3\u03c9 position-free pooling: \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd cache \u03c4\u03bf\u03c5 Sharer \u03c3\u03b5 summary \u03ba\u03b1\u03b9 \u03c4\u03b7 \u03c3\u03c5\u03bd\u03b4\u03c5\u03ac\u03b6\u03b5\u03b9 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03c4\u03bf\u03c5 Receiver. \u03a3\u03cd\u03bc\u03c6\u03c9\u03bd\u03b1 \u03bc\u03b5 \u03c4\u03bf\u03c5\u03c2 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03c4\u03bf\u03c5 XKV, \u03c4\u03c1\u03b5\u03b9\u03c2 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03bf\u03af \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03bf\u03c5\u03bd. \u03a4\u03bf summary \u03b1\u03c0\u03bf\u03c6\u03b1\u03c3\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd cache \u03c4\u03bf\u03c5 Sharer, \u03ba\u03ac\u03b8\u03b5 \u03b8\u03ad\u03c3\u03b7 \u03b5\u03bd\u03cc\u03c2 receiver layer \u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03b5\u03be\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03cc summary \u03ba\u03b1\u03b9 \u03b7 \u03b1\u03c1\u03c7\u03b9\u03ba\u03ae \u03b4\u03b9\u03b1\u03c4\u03cd\u03c0\u03c9\u03c3\u03b7 \u03c5\u03c0\u03bf\u03b8\u03ad\u03c4\u03b5\u03b9 \u03af\u03b4\u03b9\u03bf \u03b1\u03c1\u03b9\u03b8\u03bc\u03cc layers \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bc\u03b2\u03b1\u03c4\u03ae KV \u03b3\u03b5\u03c9\u03bc\u03b5\u03c4\u03c1\u03af\u03b1.<\/p>\n<div class=\"td-comparison td-comparison-cards td-comparison-cards--horizontal\">\n<div class=\"td-comparison-grid td-comparison-grid--three\">\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">Text-to-text<\/p>\n<p>\u039f Sharer \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03b1\u03bd\u03b1\u03b3\u03bd\u03ce\u03c3\u03b9\u03bc\u03bf \u03bc\u03ae\u03bd\u03c5\u03bc\u03b1 \u03ba\u03b1\u03b9 \u03bf Receiver \u03c4\u03bf \u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03be\u03b1\u03bd\u03ac. \u0395\u03af\u03bd\u03b1\u03b9 \u03c6\u03bf\u03c1\u03b7\u03c4\u03cc \u03ba\u03b1\u03b9 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03bf, \u03b1\u03bb\u03bb\u03ac \u03c0\u03bb\u03b7\u03c1\u03ce\u03bd\u03b5\u03b9 autoregressive decoding \u03c3\u03c4\u03bf \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf \u03bc\u03bf\u03bd\u03bf\u03c0\u03ac\u03c4\u03b9.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">\u0391\u03bd\u03b1\u03b3\u03bd\u03ce\u03c3\u03b9\u03bc\u03bf trace<\/span><span class=\"td-badge\">\u03a5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf latency<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">LCF-X<\/p>\n<p>\u039c\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 position-free summary \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd cache \u03c4\u03bf\u03c5 Sharer \u03c7\u03c9\u03c1\u03af\u03c2 \u03ba\u03bf\u03b9\u03bd\u03cc context. \u03a0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 sender-centric \u03ba\u03b1\u03b9 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af layer-local single-summary conditioning.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Latent channel<\/span><span class=\"td-badge\">Sharer-only summary<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card td-platform-card--navy\">\n<p class=\"td-comparison-title\">XKV<\/p>\n<p>\u039a\u03ac\u03bd\u03b5\u03b9 pooling \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b9\u03c2 \u03b4\u03cd\u03bf caches, \u03b5\u03c5\u03b8\u03c5\u03b3\u03c1\u03b1\u03bc\u03bc\u03af\u03b6\u03b5\u03b9 \u03c4\u03b1 layers \u03c3\u03b5 \u03ba\u03bf\u03b9\u03bd\u03ae \u03bc\u03bd\u03ae\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03b1\u03c6\u03ae\u03bd\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b8\u03ad\u03c3\u03b7 \u03c4\u03bf\u03c5 Receiver \u03bd\u03b1 \u03b1\u03bd\u03b1\u03ba\u03c4\u03ae\u03c3\u03b5\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc residual.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">Receiver-aware<\/span><span class=\"td-badge\">Native KV geometry<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u03a4\u03bf XKV \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03ac\u03bb\u03bb\u03b7 \u03bc\u03af\u03b1 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03c3\u03c5\u03bc\u03c0\u03af\u03b5\u03c3\u03b7\u03c2 \u03b3\u03b9\u03b1 long context. \u03a3\u03c4\u03cc\u03c7\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03c9\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae\u03c2 \u03b3\u03bd\u03ce\u03c3\u03b7\u03c2 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 \u03b4\u03cd\u03bf frozen language models, \u03ce\u03c3\u03c4\u03b5 \u03bf Receiver \u03bd\u03b1 \u03b1\u03c0\u03b1\u03bd\u03c4\u03ae\u03c3\u03b5\u03b9 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ce\u03bd\u03c4\u03b1\u03c2 \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03c0\u03bf\u03c5 \u03ad\u03c7\u03b5\u03b9 \u03ae\u03b4\u03b7 \u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b5\u03b9 \u03bf Sharer.<\/p>\n<h2 id=\"ti-kai-pou\">\u0397 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03c4\u03bf\u03c5 XKV: \u00ab\u03c4\u03b9\u00bb \u03ba\u03b1\u03b9 \u00ab\u03c0\u03bf\u03cd\u00bb<\/h2>\n<p>\u0397 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03c7\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03bf \u03c0\u03c1\u03cc\u03b2\u03bb\u03b7\u03bc\u03b1 \u03c3\u03b5 \u03b4\u03cd\u03bf \u03b5\u03c1\u03c9\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2. \u0397 \u03c0\u03c1\u03ce\u03c4\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 <em>\u03c4\u03b9 \u03b1\u03be\u03af\u03b6\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03c0\u03b9\u03ba\u03bf\u03b9\u03bd\u03c9\u03bd\u03b7\u03b8\u03b5\u03af<\/em>. \u03a4\u03bf XKV \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af learned-query attention \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b9\u03c2 \u03b4\u03cd\u03bf KV caches, \u03ba\u03c1\u03b1\u03c4\u03ce\u03bd\u03c4\u03b1\u03c2 \u03c0\u03bf\u03bb\u03bb\u03b1\u03c0\u03bb\u03ac candidate summaries \u03b1\u03bd\u03ac layer \u03ba\u03b1\u03b9 KV head. \u0388\u03c4\u03c3\u03b9 \u03b7 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03c4\u03bf\u03c5 Sharer \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03c3\u03b5 \u03c3\u03c7\u03ad\u03c3\u03b7 \u03bc\u03b5 \u03cc\u03c3\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03ae\u03b4\u03b7 \u03bf Receiver, \u03b1\u03bd\u03c4\u03af \u03bd\u03b1 \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c9\u03c2 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03bf sender-authored \u03bc\u03ae\u03bd\u03c5\u03bc\u03b1.<\/p>\n<p>\u0397 \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03b7 \u03b5\u03c1\u03ce\u03c4\u03b7\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 <em>\u03c0\u03bf\u03cd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b3\u03c1\u03b1\u03c6\u03c4\u03b5\u03af \u03b7 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1<\/em>. \u039a\u03ac\u03b8\u03b5 raw \u03b8\u03ad\u03c3\u03b7 \u03c4\u03b7\u03c2 receiver cache \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03c9\u03c2 query \u03c0\u03c1\u03bf\u03c2 \u03bc\u03b9\u03b1 \u03c3\u03c5\u03bc\u03c0\u03b1\u03b3\u03ae \u03ba\u03bf\u03b9\u03bd\u03ae \u03bc\u03bd\u03ae\u03bc\u03b7. \u0388\u03bd\u03b1\u03c2 shared decoder \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 gated residuals \u03c3\u03c4\u03bf native KV geometry \u03c4\u03bf\u03c5 Receiver. \u0397 cache \u03b4\u03b5\u03bd \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03b8\u03af\u03c3\u03c4\u03b1\u03c4\u03b1\u03b9\u00b7 \u03b5\u03bd\u03b7\u03bc\u03b5\u03c1\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 residual correction, \u03ce\u03c3\u03c4\u03b5 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b8\u03ad\u03c3\u03b5\u03b9\u03c2 \u03bd\u03b1 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03b1\u03bd\u03b1\u03ba\u03c4\u03ae\u03c3\u03bf\u03c5\u03bd \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc \u03c3\u03c5\u03bd\u03b4\u03c5\u03b1\u03c3\u03bc\u03cc \u03b1\u03c0\u03cc \u03c4\u03b1 \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03b1 summaries.<\/p>\n<div class=\"td-decision-band\">\n<div class=\"td-decision-band-content\">\n<p class=\"td-decision-band-kicker\">\u039f \u03c0\u03c5\u03c1\u03ae\u03bd\u03b1\u03c2 \u03c4\u03bf\u03c5 XKV<\/p>\n<p class=\"td-decision-band-title\">\u03a4\u03bf \u00ab\u03c4\u03b9 \u03b8\u03b1 \u03c0\u03b5\u03c1\u03ac\u03c3\u03b5\u03b9\u00bb \u03ba\u03bf\u03b9\u03c4\u03ac \u03ba\u03b1\u03b9 \u03c4\u03b9\u03c2 \u03b4\u03cd\u03bf caches\u00b7 \u03c4\u03bf \u00ab\u03c0\u03bf\u03cd \u03b8\u03b1 \u03b3\u03c1\u03b1\u03c6\u03c4\u03b5\u03af\u00bb \u03c4\u03bf \u03b1\u03c0\u03bf\u03c6\u03b1\u03c3\u03af\u03b6\u03b5\u03b9 \u03bf Receiver<\/p>\n<p>\u0397 \u03b4\u03b9\u03ac\u03ba\u03c1\u03b9\u03c3\u03b7 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 joint communication memory \u03ba\u03b1\u03b9 position-specific retrieval \u03b1\u03c6\u03b1\u03b9\u03c1\u03b5\u03af \u03c4\u03bf single-summary bottleneck \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b1 \u03b2\u03ac\u03c1\u03b7 \u03c4\u03c9\u03bd \u03b4\u03cd\u03bf base models.<\/p>\n<\/div>\n<\/div>\n<p>\u03a4\u03b1 output heads \u03b3\u03b9\u03b1 keys \u03ba\u03b1\u03b9 values \u03b1\u03c1\u03c7\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03bc\u03b7\u03b4\u03ad\u03bd. \u03a0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7, \u03bf translator \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ce\u03c2 \u03c4\u03b7 receiver-only \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ac. \u039a\u03b1\u03c4\u03ac \u03c4\u03b7\u03bd \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03bc\u03b1\u03b8\u03b1\u03af\u03bd\u03b5\u03b9 \u03bc\u03cc\u03bd\u03bf \u03c4\u03b7 \u03b4\u03b9\u03cc\u03c1\u03b8\u03c9\u03c3\u03b7 \u03c4\u03b7\u03c2 cache, \u03b5\u03bd\u03ce \u03c4\u03b1 \u03b4\u03cd\u03bf LLMs \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03bf\u03c5\u03bd frozen. \u03a3\u03c4\u03bf inference \u03b7 \u03b5\u03bd\u03b7\u03bc\u03ad\u03c1\u03c9\u03c3\u03b7 \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03af\u03b1 \u03c6\u03bf\u03c1\u03ac \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c4\u03bf \u03ba\u03b1\u03bd\u03bf\u03bd\u03b9\u03ba\u03cc greedy decoding \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03c0\u03c1\u03bf\u03c3\u03c4\u03af\u03b8\u03b5\u03c4\u03b1\u03b9 \u03bd\u03ad\u03b1 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03c0\u03b1\u03c1\u03b1\u03b3\u03cc\u03bc\u03b5\u03bd\u03bf token.<\/p>\n<h2 id=\"eterogeni-montela\">\u03a0\u03ce\u03c2 \u03b3\u03b5\u03c6\u03c5\u03c1\u03ce\u03bd\u03b5\u03b9 \u03b5\u03c4\u03b5\u03c1\u03bf\u03b3\u03b5\u03bd\u03ae \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1<\/h2>\n<p>\u039f\u03b9 \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ad\u03c2 \u03b4\u03b9\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03b9\u03c2 \u03b4\u03cd\u03bf LLMs \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c5\u03c4\u03bf\u03bd\u03cc\u03b7\u03c4\u03b1 \u03c3\u03c5\u03bc\u03b2\u03b1\u03c4\u03ad\u03c2. \u039c\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c6\u03ad\u03c1\u03bf\u03c5\u03bd \u03c3\u03b5 \u03bf\u03b9\u03ba\u03bf\u03b3\u03ad\u03bd\u03b5\u03b9\u03b1, \u03b2\u03ac\u03b8\u03bf\u03c2, \u03b1\u03c1\u03b9\u03b8\u03bc\u03cc KV heads, head dimension \u03ba\u03b1\u03b9 tokenizer. \u03a4\u03bf XKV \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af learned layer map \u03ce\u03c3\u03c4\u03b5 \u03c4\u03b1 pooled summaries \u03c4\u03bf\u03c5 Sharer \u03bd\u03b1 \u03b5\u03c5\u03b8\u03c5\u03b3\u03c1\u03b1\u03bc\u03bc\u03b9\u03c3\u03c4\u03bf\u03cd\u03bd \u03c3\u03c4\u03bf\u03bd \u03b1\u03c1\u03b9\u03b8\u03bc\u03cc layers \u03c4\u03bf\u03c5 Receiver. \u038c\u03c4\u03b1\u03bd \u03c4\u03b1 \u03b2\u03ac\u03b8\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03af\u03b4\u03b9\u03b1, \u03b5\u03c6\u03b1\u03c1\u03bc\u03cc\u03b6\u03b5\u03c4\u03b1\u03b9 identity mapping.<\/p>\n<p>\u039c\u03b5\u03c4\u03ac \u03c4\u03b7\u03bd \u03b5\u03c5\u03b8\u03c5\u03b3\u03c1\u03ac\u03bc\u03bc\u03b9\u03c3\u03b7, \u03c4\u03b1 summaries \u03c4\u03c9\u03bd \u03b4\u03cd\u03bf \u03c0\u03bb\u03b5\u03c5\u03c1\u03ce\u03bd \u03c3\u03c5\u03bd\u03b5\u03bd\u03ce\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9, \u03c0\u03b5\u03c1\u03bd\u03bf\u03cd\u03bd \u03b1\u03c0\u03cc bottleneck \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b1\u03bc\u03b5\u03b9\u03b3\u03bd\u03cd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b5 self-attention. \u0397 \u03ba\u03bf\u03b9\u03bd\u03ae \u03bc\u03bd\u03ae\u03bc\u03b7 \u03ad\u03c7\u03b5\u03b9 \u03ad\u03bd\u03b1 token \u03b1\u03bd\u03ac receiver layer, \u03ac\u03c1\u03b1 \u03c4\u03bf \u03bc\u03ae\u03ba\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf \u03b2\u03ac\u03b8\u03bf\u03c2 \u03c4\u03bf\u03c5 Receiver \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf \u03bc\u03ae\u03ba\u03bf\u03c2 \u03c4\u03c9\u03bd prompts. \u039f decoder, \u03c4\u03bf gate \u03ba\u03b1\u03b9 \u03c4\u03b1 output heads \u03bc\u03bf\u03b9\u03c1\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b5\u03c4\u03b1\u03be\u03cd layers, \u03b5\u03bd\u03ce learned layer embeddings \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03bf\u03cd\u03bd \u03c4\u03b7\u03bd \u03c4\u03b1\u03c5\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03ac\u03b8\u03b5 layer.<\/p>\n<p>\u0391\u03c5\u03c4\u03ae \u03b7 \u03b5\u03c5\u03b5\u03bb\u03b9\u03be\u03af\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc \u03cc\u03c1\u03b9\u03bf: \u03b7 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7 \u03c3\u03c4\u03b9\u03c2 KV caches \u03ba\u03b1\u03b9 \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 translator \u03b3\u03b9\u03b1 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf ordered model pair. \u0388\u03bd\u03b1 \u03ba\u03bb\u03b5\u03b9\u03c3\u03c4\u03cc API \u03c0\u03bf\u03c5 \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03ad\u03c6\u03b5\u03b9 \u03bc\u03cc\u03bd\u03bf \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03b4\u03b5\u03bd \u03c0\u03b1\u03c1\u03ad\u03c7\u03b5\u03b9 \u03c4\u03bf \u03b1\u03c0\u03b1\u03c1\u03b1\u03af\u03c4\u03b7\u03c4\u03bf \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03cc state.<\/p>\n<h2 id=\"peirama-45-settings\">\u03a4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03c4\u03c9\u03bd 45 settings<\/h2>\n<p>\u0397 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b5 Qwen3-0.6B, Gemma-3-1B \u03ba\u03b1\u03b9 Llama-3.2-3B \u03c9\u03c2 Sharer \u03ba\u03b1\u03b9 Receiver. \u03a4\u03bf \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 3\u00d73 ordered grid \u03ad\u03b4\u03c9\u03c3\u03b5 \u03b5\u03bd\u03bd\u03ad\u03b1 \u03b6\u03b5\u03cd\u03b3\u03b7: \u03ad\u03be\u03b9 \u03b5\u03c4\u03b5\u03c1\u03bf\u03b3\u03b5\u03bd\u03ae \u03ba\u03b1\u03b9 \u03c4\u03c1\u03af\u03b1 \u03af\u03b4\u03b9\u03b1\u03c2 \u03bf\u03b9\u03ba\u03bf\u03b3\u03ad\u03bd\u03b5\u03b9\u03b1\u03c2. \u039f\u03b9 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ad\u03c2 \u03ba\u03ac\u03bb\u03c5\u03c8\u03b1\u03bd ROPES, MuSiQue \u03ba\u03b1\u03b9 HotpotQA-bridge \u03b3\u03b9\u03b1 generative multi-hop QA, \u03ba\u03b1\u03b8\u03ce\u03c2 \u03ba\u03b1\u03b9 QASC \u03ba\u03b1\u03b9 StrategyQA \u03b3\u03b9\u03b1 classification-style reasoning. \u0388\u03c4\u03c3\u03b9 \u03c0\u03c1\u03bf\u03ad\u03ba\u03c5\u03c8\u03b1\u03bd 45 dataset\u2013model-pair cells.<\/p>\n<p>\u03a3\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03c0\u03b1\u03c1\u03ac\u03b4\u03b5\u03b9\u03b3\u03bc\u03b1, \u03c4\u03b1 \u03c5\u03c0\u03bf\u03c3\u03c4\u03b7\u03c1\u03b9\u03ba\u03c4\u03b9\u03ba\u03ac \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03b1 \u03c7\u03c9\u03c1\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03bd \u03bd\u03c4\u03b5\u03c4\u03b5\u03c1\u03bc\u03b9\u03bd\u03b9\u03c3\u03c4\u03b9\u03ba\u03ac \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03c4\u03bf\u03c5\u03c2 \u03b4\u03cd\u03bf \u03c1\u03cc\u03bb\u03bf\u03c5\u03c2. \u0397 \u03b5\u03c1\u03ce\u03c4\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9, \u03cc\u03c0\u03bf\u03c5 \u03c5\u03c0\u03ae\u03c1\u03c7\u03b1\u03bd, \u03bf\u03b9 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ad\u03c2 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7\u03c2 \u03ae\u03c4\u03b1\u03bd \u03bf\u03c1\u03b1\u03c4\u03ad\u03c2 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b1 \u03b4\u03cd\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1, \u03b1\u03bb\u03bb\u03ac \u03c4\u03b1 evidence contexts \u03ae\u03c4\u03b1\u03bd \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03c9\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac. \u0393\u03b9\u03b1 \u03c4\u03b1 generative datasets \u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b1\u03bd\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 exact match \u03ba\u03b1\u03b9 F1\u00b7 \u03b3\u03b9\u03b1 \u03c4\u03b1 classification datasets accuracy.<\/p>\n<p>\u039f\u03b9 baselines \u03ae\u03c4\u03b1\u03bd text-to-text \u03ba\u03b1\u03b9 \u03bc\u03b9\u03b1 \u03b5\u03bd\u03b9\u03c3\u03c7\u03c5\u03bc\u03ad\u03bd\u03b7 \u03b5\u03c4\u03b5\u03c1\u03bf\u03b3\u03b5\u03bd\u03ae\u03c2 \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 \u03c4\u03bf\u03c5 LCF-X. \u03a3\u03c4\u03b9\u03c2 latent \u03bc\u03b5\u03b8\u03cc\u03b4\u03bf\u03c5\u03c2 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c4\u03b7\u03ba\u03b5 \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf communication module. \u039a\u03ac\u03b8\u03b5 \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03b5\u03ba\u03c4\u03b5\u03bb\u03ad\u03c3\u03c4\u03b7\u03ba\u03b5 \u03bc\u03af\u03b1 \u03c6\u03bf\u03c1\u03ac \u03c3\u03b5 Nvidia A100 80GB, Ubuntu 22.04 \u03ba\u03b1\u03b9 CUDA 12.2, \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03bf \u03c0\u03bf\u03c5 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03bd\u03bf\u03b4\u03b5\u03cd\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03bd\u03ac\u03b3\u03bd\u03c9\u03c3\u03b7 \u03c4\u03c9\u03bd latency \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03c3\u03bc\u03ac\u03c4\u03c9\u03bd.<\/p>\n<h2 id=\"apotelesmata-poiotitas\">\u03a4\u03b9 \u03b4\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03c4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2<\/h2>\n<p>\u03a3\u03c4\u03b7 \u03c3\u03c5\u03b3\u03ba\u03ad\u03bd\u03c4\u03c1\u03c9\u03c3\u03b7 \u03c4\u03c9\u03bd 45 cells, \u03c4\u03bf XKV \u03b5\u03af\u03c7\u03b5 macro native score 52,37, \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 49,76 \u03b3\u03b9\u03b1 \u03c4\u03bf LCF-X \u03ba\u03b1\u03b9 48,86 \u03b3\u03b9\u03b1 \u03c4\u03bf text-to-text. \u0389\u03c4\u03b1\u03bd \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03ae \u03b9\u03c3\u03cc\u03b2\u03b1\u03b8\u03bc\u03b1 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03c3\u03b5 31 \u03b1\u03c0\u03cc \u03c4\u03b1 45 cells \u03ba\u03b1\u03b9 \u03b5\u03af\u03c7\u03b5 \u03bc\u03ad\u03c3\u03b7 \u03ba\u03b1\u03c4\u03ac\u03c4\u03b1\u03be\u03b7 1,37. \u03a4\u03bf LCF-X \u03ae\u03c4\u03b1\u03bd \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf baseline \u03c3\u03c4\u03b1 \u03c4\u03c1\u03af\u03b1 generative QA datasets, \u03b5\u03bd\u03ce \u03c4\u03bf text-to-text \u03c3\u03c4\u03b1 \u03b4\u03cd\u03bf classification datasets.<\/p>\n<p>\u0397 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03bd\u03af\u03ba\u03b7. \u03a4\u03bf XKV \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03ba\u03b1\u03c4\u03ac \u03bc\u03ad\u03c3\u03bf \u03cc\u03c1\u03bf \u03c4\u03bf LCF-X \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b1 \u03c0\u03ad\u03bd\u03c4\u03b5 datasets, \u03b1\u03bb\u03bb\u03ac \u03c4\u03bf text-to-text \u03c0\u03b1\u03c1\u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03bf\u03c1\u03b9\u03b1\u03ba\u03ac \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03c3\u03c4\u03bf QASC \u03ba\u03b1\u03c4\u03ac 0,27 \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03b9\u03b1\u03af\u03b5\u03c2 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2. \u03a3\u03b5 \u03b5\u03c0\u03b9\u03bc\u03ad\u03c1\u03bf\u03c5\u03c2 model pairs \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03c5\u03bd \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 cells \u03cc\u03c0\u03bf\u03c5 \u03ac\u03bb\u03bb\u03bf \u03c0\u03c1\u03c9\u03c4\u03cc\u03ba\u03bf\u03bb\u03bb\u03bf \u03ba\u03b5\u03c1\u03b4\u03af\u03b6\u03b5\u03b9. \u03a4\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03b5\u03cd\u03c1\u03b7\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03bf XKV \u03c0\u03b1\u03c1\u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03b1\u03bd\u03c4\u03b1\u03b3\u03c9\u03bd\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b1 \u03b4\u03cd\u03bf \u03b5\u03af\u03b4\u03b7 task, \u03cc\u03c7\u03b9 \u03cc\u03c4\u03b9 \u03b7 latent \u03b5\u03c0\u03b9\u03ba\u03bf\u03b9\u03bd\u03c9\u03bd\u03af\u03b1 \u03c5\u03c0\u03b5\u03c1\u03ad\u03c7\u03b5\u03b9 \u03c0\u03ac\u03bd\u03c4\u03b1.<\/p>\n<p>\u03a3\u03c4\u03bf ROPES, \u03b7 \u03bc\u03ad\u03c3\u03b7 \u03b5\u03c0\u03af\u03b4\u03bf\u03c3\u03b7 \u03c4\u03bf\u03c5 XKV \u03ae\u03c4\u03b1\u03bd 50,48 exact match \u03ba\u03b1\u03b9 56,99 F1, \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 45,86\/52,79 \u03b3\u03b9\u03b1 \u03c4\u03bf LCF-X. \u0397 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03b1\u03bd\u03c4\u03b9\u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af \u03c3\u03b5 4,62 EM \u03ba\u03b1\u03b9 4,20 F1 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2 \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf experimental setup. \u0391\u03c5\u03c4\u03ac \u03c4\u03b1 scores \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc benchmark \u03b3\u03b9\u03b1 customer support, marketing, e-commerce \u03ae tool-use agents.<\/p>\n<aside class=\"td-article-note\">\n<p><strong>\u03a4\u03bf scope \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c7\u03b1\u03b8\u03b5\u03af:<\/strong> \u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 arXiv v1, \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c4\u03c1\u03af\u03b1 \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ac \u03bc\u03b9\u03ba\u03c1\u03ac open models, \u03c0\u03ad\u03bd\u03c4\u03b5 split-evidence datasets \u03ba\u03b1\u03b9 \u03bc\u03af\u03b1 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03b1\u03bd\u03ac \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1. \u0394\u03b5\u03bd \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af \u03bc\u03b5\u03b3\u03ac\u03bb\u03b1 \u03b5\u03bc\u03c0\u03bf\u03c1\u03b9\u03ba\u03ac \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1, \u03bc\u03b1\u03ba\u03c1\u03bf\u03c7\u03c1\u03cc\u03bd\u03b9\u03b5\u03c2 \u03c3\u03c5\u03bd\u03bf\u03bc\u03b9\u03bb\u03af\u03b5\u03c2, adversarial inputs, \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac permissions \u03ae agents \u03c0\u03bf\u03c5 \u03b1\u03bb\u03bb\u03ac\u03b6\u03bf\u03c5\u03bd \u03b5\u03be\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ac \u03c3\u03c5\u03c3\u03c4\u03ae\u03bc\u03b1\u03c4\u03b1.<\/p>\n<\/aside>\n<h2 id=\"latency-parametroi\">Latency \u03ba\u03b1\u03b9 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c3\u03b9\u03bc\u03b5\u03c2 \u03c0\u03b1\u03c1\u03ac\u03bc\u03b5\u03c4\u03c1\u03bf\u03b9<\/h2>\n<p>\u03a4\u03bf \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03b4\u03b9\u03b1\u03cd\u03bb\u03bf\u03c5. \u03a3\u03b5 \u03cc\u03bb\u03b1 \u03c4\u03b1 45 cells, \u03bf XKV translator \u03b5\u03af\u03c7\u03b5 \u03bc\u03ad\u03c3\u03bf \u03c7\u03c1\u03cc\u03bd\u03bf 5,8 ms, \u03b5\u03bd\u03ce \u03bf LCF-X fusor 59,9 ms: \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac 10,3\u00d7. \u03a4\u03bf \u03bc\u03ad\u03c3\u03bf end-to-end latency \u03bc\u03b5\u03b9\u03ce\u03b8\u03b7\u03ba\u03b5 \u03b1\u03c0\u03cc 227,9 ms \u03c3\u03b5 167,6 ms, \u03b4\u03b7\u03bb\u03b1\u03b4\u03ae \u03ba\u03b1\u03c4\u03ac 26,4%. \u03a3\u03b5 \u03c3\u03c7\u03ad\u03c3\u03b7 \u03bc\u03b5 text-to-text, \u03c4\u03bf XKV \u03ae\u03c4\u03b1\u03bd 6,8\u00d7 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf end to end, \u03ba\u03b1\u03b8\u03ce\u03c2 \u03b1\u03c0\u03ad\u03c6\u03b5\u03c5\u03b3\u03b5 \u03c4\u03b7\u03bd autoregressive \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf\u03c5 \u03bc\u03b7\u03bd\u03cd\u03bc\u03b1\u03c4\u03bf\u03c2.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-title\">XKV \u03c3\u03c4\u03b1 45 dataset\u2013model-pair cells<\/p>\n<p class=\"td-chart-subtitle\">\u039c\u03ad\u03c3\u03b5\u03c2 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03c4\u03b7\u03c2 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1\u03c2 \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf A100 setup\u00b7 \u03c4\u03bf macro score \u03c3\u03c5\u03bd\u03b4\u03c5\u03ac\u03b6\u03b5\u03b9 F1 \u03b3\u03b9\u03b1 generative QA \u03ba\u03b1\u03b9 accuracy \u03b3\u03b9\u03b1 classification.<\/p>\n<\/div>\n<div class=\"td-chart-body\">\n<div class=\"td-metric-grid\">\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">52,37<\/span><span class=\"td-metric-label\">Macro native score<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">5,8 ms<\/span><span class=\"td-metric-label\">Translator latency<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">167,6 ms<\/span><span class=\"td-metric-label\">End-to-end latency<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">4,55 \u03b5\u03ba\u03b1\u03c4.<\/span><span class=\"td-metric-label\">Trainable parameters<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u0397 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03c0\u03c1\u03bf\u03ba\u03cd\u03c0\u03c4\u03b5\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae. \u03a4\u03bf direct pooling \u03c3\u03c5\u03bc\u03c0\u03b9\u03ad\u03b6\u03b5\u03b9 \u03c4\u03b9\u03c2 positions \u03c3\u03b5 \u03bb\u03af\u03b3\u03b1 slots \u03bc\u03b5 \u03bc\u03af\u03b1 vectorized \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1, \u03b5\u03bd\u03ce \u03c4\u03bf span pooling \u03c4\u03bf\u03c5 LCF-X \u03b5\u03c0\u03b5\u03be\u03b5\u03c1\u03b3\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5\u03c4\u03b1\u03b2\u03bb\u03b7\u03c4\u03bf\u03cd \u03bc\u03ae\u03ba\u03bf\u03c5\u03c2 spans \u03ba\u03b1\u03b9 \u03bc\u03b5\u03b3\u03b1\u03bb\u03ce\u03bd\u03b5\u03b9 \u03bc\u03b5 \u03c4\u03bf context. \u03a3\u03c4\u03b1 \u03c0\u03b5\u03b9\u03c1\u03ac\u03bc\u03b1\u03c4\u03b1, \u03bf translator \u03c4\u03bf\u03c5 XKV \u03ba\u03b9\u03bd\u03bf\u03cd\u03bd\u03c4\u03b1\u03bd \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c3\u03c4\u03b1 5\u201310 ms \u03ba\u03b1\u03b9 \u03bf LCF-X fusor \u03b1\u03c0\u03cc 31 \u03ad\u03c9\u03c2 123 ms.<\/p>\n<p>\u03a4\u03bf XKV \u03b5\u03af\u03c7\u03b5 4,55 \u03b5\u03ba\u03b1\u03c4\u03bf\u03bc\u03bc\u03cd\u03c1\u03b9\u03b1 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c3\u03b9\u03bc\u03b5\u03c2 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03c4\u03c1\u03bf\u03c5\u03c2 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 19,03 \u03b5\u03ba\u03b1\u03c4\u03bf\u03bc\u03bc\u03c5\u03c1\u03af\u03c9\u03bd \u03b3\u03b9\u03b1 \u03c4\u03bf LCF-X, \u03bc\u03b5\u03af\u03c9\u03c3\u03b7 76,1%. \u0391\u03c5\u03c4\u03cc \u03b4\u03b5\u03bd \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c4\u03bf multi-agent \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03b9\u03ba\u03c1\u03cc \u03ae \u03c6\u03b8\u03b7\u03bd\u03cc: \u03ba\u03b1\u03b9 \u03c4\u03b1 \u03b4\u03cd\u03bf frozen \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03b5\u03be\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03bf\u03cd\u03bd \u03bd\u03b1 \u03ba\u03ac\u03bd\u03bf\u03c5\u03bd prefill \u03ba\u03b1\u03b9 \u03bf Receiver \u03b5\u03be\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03b5\u03af \u03bd\u03b1 \u03ba\u03ac\u03bd\u03b5\u03b9 decoding \u03c4\u03b7\u03c2 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae\u03c2 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7\u03c2.<\/p>\n<h2 id=\"ablations\">Ce que r\u00e9v\u00e8lent les ablations<\/h2>\n<p>\u03a4\u03b1 ablations \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03bf\u03c5\u03bd QASC, ROPES \u03ba\u03b1\u03b9 StrategyQA \u03c3\u03b5 27 \u03b1\u03bd\u03c4\u03b9\u03c3\u03c4\u03bf\u03b9\u03c7\u03b9\u03c3\u03bc\u03ad\u03bd\u03b1 cells. \u0397 \u03b1\u03c6\u03b1\u03af\u03c1\u03b5\u03c3\u03b7 \u03c4\u03bf\u03c5 receiver pooling, \u03ce\u03c3\u03c4\u03b5 \u03b7 \u03bc\u03bd\u03ae\u03bc\u03b7 \u03bd\u03b1 \u03c7\u03c4\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf\u03bd Sharer, \u03bc\u03b5\u03af\u03c9\u03c3\u03b5 \u03c4\u03b7 \u03bc\u03ad\u03c3\u03b7 \u03b2\u03b1\u03b8\u03bc\u03bf\u03bb\u03bf\u03b3\u03af\u03b1 \u03ba\u03b1\u03c4\u03ac 0,46 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2 \u03ba\u03b1\u03b9 \u03b5\u03be\u03bf\u03b9\u03ba\u03bf\u03bd\u03cc\u03bc\u03b7\u03c3\u03b5 1,1 ms. \u0391\u03c5\u03c4\u03cc \u03c5\u03c0\u03bf\u03c3\u03c4\u03b7\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03c5\u03c0\u03cc\u03b8\u03b5\u03c3\u03b7 \u03cc\u03c4\u03b9 \u03c4\u03bf \u00ab\u03c4\u03b9 \u03b8\u03b1 \u03c3\u03c4\u03b1\u03bb\u03b5\u03af\u00bb \u03c9\u03c6\u03b5\u03bb\u03b5\u03af\u03c4\u03b1\u03b9 \u03cc\u03c4\u03b1\u03bd \u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 \u03c5\u03c0\u03cc\u03c8\u03b7 \u03ba\u03b1\u03b9 \u03c4\u03bf state \u03c4\u03bf\u03c5 \u03c0\u03b1\u03c1\u03b1\u03bb\u03ae\u03c0\u03c4\u03b7.<\/p>\n<p>\u0397 \u03b1\u03c6\u03b1\u03af\u03c1\u03b5\u03c3\u03b7 \u03c4\u03bf\u03c5 position-specific receiver cross-attention \u03b5\u03af\u03c7\u03b5 \u03bc\u03ad\u03c3\u03b7 \u03c0\u03c4\u03ce\u03c3\u03b7 0,83 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2 \u03ba\u03b1\u03b9 \u03c0\u03c4\u03ce\u03c3\u03b7 1,98 \u03c3\u03c4\u03bf QASC. \u0397 \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03c4\u03bf\u03c5 direct pooling \u03bc\u03b5 \u03c4\u03bf span pooling \u03c4\u03bf\u03c5 LCF-X \u03b1\u03cd\u03be\u03b7\u03c3\u03b5 \u03c4\u03bf component latency \u03b1\u03c0\u03cc 5,8 \u03c3\u03b5 39,1 ms \u03ba\u03b1\u03b9 \u03bc\u03b5\u03af\u03c9\u03c3\u03b5 \u03c4\u03b7 \u03bc\u03ad\u03c3\u03b7 \u03b2\u03b1\u03b8\u03bc\u03bf\u03bb\u03bf\u03b3\u03af\u03b1 \u03ba\u03b1\u03c4\u03ac 0,94. \u039f\u03b9 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ad\u03c2 \u03b1\u03c0\u03bf\u03bc\u03cc\u03bd\u03c9\u03c3\u03b7\u03c2 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03bf\u03c5\u03bd \u03ba\u03ac\u03b8\u03b5 \u03b1\u03b9\u03c4\u03b9\u03ce\u03b4\u03b7 \u03b5\u03c1\u03bc\u03b7\u03bd\u03b5\u03af\u03b1, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03cc\u03c4\u03b9 \u03ba\u03b1\u03b9 \u03bf\u03b9 \u03c4\u03c1\u03b5\u03b9\u03c2 \u03c3\u03c7\u03b5\u03b4\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ad\u03c2 \u03c3\u03c5\u03bd\u03b5\u03b9\u03c3\u03c6\u03ad\u03c1\u03bf\u03c5\u03bd.<\/p>\n<p>\u0397 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b1\u03bd\u03ac\u03b3\u03bd\u03c9\u03c3\u03b7 \u03bc\u03bf\u03b9\u03ac\u03b6\u03b5\u03b9 \u03bc\u03b5 \u03c4\u03bf <a href=\"https:\/\/twodots.gr\/aces-ai-skill-evaluation-skill-lift\/\">paired evaluation \u03b5\u03bd\u03cc\u03c2 AI skill<\/a>: \u03b1\u03bb\u03bb\u03ac\u03b6\u03bf\u03c5\u03bc\u03b5 \u03ad\u03bd\u03b1 \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03bf, \u03ba\u03c1\u03b1\u03c4\u03ac\u03bc\u03b5 \u03c4\u03bf baseline \u03bf\u03c1\u03b1\u03c4\u03cc \u03ba\u03b1\u03b9 \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03bf\u03c5\u03bc\u03b5 \u03bc\u03b1\u03b6\u03af \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1, latency \u03ba\u03b1\u03b9 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7\u03c2.<\/p>\n<h2 id=\"enterprise-architektoniki\">\u03a4\u03b9 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1 enterprise AI<\/h2>\n<p>\u03a4\u03bf XKV \u03c0\u03b1\u03c1\u03bf\u03c5\u03c3\u03b9\u03ac\u03b6\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03c0\u03b9\u03b8\u03b1\u03bd\u03ae \u03ba\u03b1\u03c4\u03b5\u03cd\u03b8\u03c5\u03bd\u03c3\u03b7 \u03b3\u03b9\u03b1 workloads \u03cc\u03c0\u03bf\u03c5 \u03b7 \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03c9\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03ae\u03b4\u03b7 \u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03b7\u03bc\u03ad\u03bd\u03b7 \u03c3\u03b5 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03ba\u03b1\u03b9 \u03c4\u03bf text decoding \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03af \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc bottleneck. \u0388\u03bd\u03b1\u03c2 research agent \u03b8\u03b1 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03c3\u03b5 \u03b8\u03b5\u03c9\u03c1\u03b7\u03c4\u03b9\u03ba\u03ac \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 latent state \u03c3\u03b5 verifier, \u03ae \u03ad\u03bd\u03b1\u03c2 agent \u03bc\u03b5 \u03b5\u03c0\u03b9\u03c4\u03c1\u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03b7 \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7 \u03c3\u03b5 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b1 \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03b1 \u03bd\u03b1 \u03c4\u03c1\u03bf\u03c6\u03bf\u03b4\u03bf\u03c4\u03ae\u03c3\u03b5\u03b9 \u03ad\u03bd\u03b1\u03bd Receiver \u03c7\u03c9\u03c1\u03af\u03c2 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf memo.<\/p>\n<p>\u0397 \u03b4\u03c5\u03bd\u03b1\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03c5\u03c4\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b9\u03bf \u03ba\u03bf\u03bd\u03c4\u03ac \u03c3\u03b5 \u03b5\u03be\u03b5\u03b9\u03b4\u03b9\u03ba\u03b5\u03c5\u03bc\u03ad\u03bd\u03b7 \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae inference \u03c0\u03b1\u03c1\u03ac \u03c3\u03b5 prompt engineering. \u0391\u03c0\u03b1\u03b9\u03c4\u03b5\u03af \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7 \u03c3\u03c4\u03b9\u03c2 caches, \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ce\u03bd interfaces, pair-specific \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b1\u03bd\u03ad\u03bb\u03b5\u03b3\u03c7\u03bf \u03cc\u03c4\u03b1\u03bd \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 model version \u03ae cache geometry. \u0397 \u03b5\u03bc\u03c0\u03b5\u03b9\u03c1\u03af\u03b1 \u03b1\u03c0\u03cc <a href=\"https:\/\/twodots.gr\/fl-maestro-multi-agent-federated-learning\/\">multi-agent orchestration \u03c3\u03b5 federated learning<\/a> \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03b1\u03c1\u03c7\u03ae: \u03bf learned \u03c3\u03c5\u03bd\u03c4\u03bf\u03bd\u03b9\u03c3\u03bc\u03cc\u03c2 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 deterministic validation \u03ba\u03b1\u03b9 resource controls \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03bf\u03ba\u03c4\u03ae\u03c3\u03b5\u03b9 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03cc \u03c1\u03cc\u03bb\u03bf.<\/p>\n<p>\u0393\u03b9\u03b1 marketing \u03ae e-commerce \u03bf\u03bc\u03ac\u03b4\u03b5\u03c2, \u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b4\u03b5\u03bd \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03ad\u03c4\u03bf\u03b9\u03bc\u03bf plug-in. \u03a0\u03c1\u03bf\u03c3\u03c6\u03ad\u03c1\u03b5\u03b9 \u03cc\u03bc\u03c9\u03c2 \u03ba\u03c1\u03b9\u03c4\u03ae\u03c1\u03b9\u03bf \u03c3\u03c7\u03b5\u03b4\u03b9\u03b1\u03c3\u03bc\u03bf\u03cd: \u03b4\u03b5\u03bd \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03ba\u03ac\u03b8\u03b5 model-to-model handoff \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03ad\u03bd\u03b1 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf \u03c6\u03c5\u03c3\u03b9\u03ba\u03bf\u03b3\u03bb\u03c9\u03c3\u03c3\u03b9\u03ba\u03cc \u03bc\u03ae\u03bd\u03c5\u03bc\u03b1, \u03b5\u03b9\u03b4\u03b9\u03ba\u03ac \u03cc\u03c4\u03b1\u03bd \u03bf \u03c7\u03c1\u03cc\u03bd\u03bf\u03c2 \u03b1\u03c0\u03cc\u03ba\u03c1\u03b9\u03c3\u03b7\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf\u03c2. \u0391\u03c0\u03cc \u03c4\u03b7\u03bd \u03ac\u03bb\u03bb\u03b7, \u03cc\u03c4\u03b1\u03bd \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c0\u03c1\u03cc\u03ba\u03b5\u03b9\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03b5\u03c0\u03b7\u03c1\u03b5\u03ac\u03c3\u03b5\u03b9 \u03c0\u03b5\u03bb\u03ac\u03c4\u03b7, \u03c0\u03b1\u03c1\u03b1\u03b3\u03b3\u03b5\u03bb\u03af\u03b1, campaign budget \u03ae \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03cc \u03b1\u03c1\u03c7\u03b5\u03af\u03bf, \u03c4\u03bf \u03b1\u03bd\u03b1\u03b3\u03bd\u03ce\u03c3\u03b9\u03bc\u03bf trace \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03ae \u03b1\u03c0\u03b1\u03af\u03c4\u03b7\u03c3\u03b7.<\/p>\n<h2 id=\"auditability-asfaleia\">Auditability, \u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1 \u03ba\u03b1\u03b9 fallback<\/h2>\n<p>\u0397 latent \u03bc\u03bd\u03ae\u03bc\u03b7 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b1 \u03b1\u03bd\u03b1\u03b3\u03bd\u03ce\u03c3\u03b9\u03bc\u03bf \u03bc\u03ae\u03bd\u03c5\u03bc\u03b1. \u0391\u03c5\u03c4\u03cc \u03b4\u03c5\u03c3\u03ba\u03bf\u03bb\u03b5\u03cd\u03b5\u03b9 debugging, attribution \u03ba\u03b1\u03b9 \u03b1\u03c0\u03cc\u03b4\u03b5\u03b9\u03be\u03b7 \u03c4\u03b7\u03c2 \u03c0\u03c1\u03bf\u03ad\u03bb\u03b5\u03c5\u03c3\u03b7\u03c2 \u03bc\u03b9\u03b1\u03c2 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae\u03c2 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7\u03c2. \u0397 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c3\u03c4\u03b7 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03bc\u03b5\u03c4\u03c1\u03ae\u03b8\u03b7\u03ba\u03b5 \u03b1\u03c0\u03cc downstream answers, \u03cc\u03c7\u03b9 \u03b1\u03c0\u03cc \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b7 \u03b5\u03c0\u03b9\u03b8\u03b5\u03ce\u03c1\u03b7\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03c0\u03b5\u03c1\u03b9\u03b5\u03c7\u03bf\u03bc\u03ad\u03bd\u03bf\u03c5 \u03c0\u03bf\u03c5 \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b8\u03b7\u03ba\u03b5 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03c4\u03b9\u03c2 caches.<\/p>\n<p>\u0388\u03bd\u03b1 production design \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 policy \u03b3\u03b9\u03b1 \u03c4\u03bf \u03c0\u03bf\u03b9\u03bf context \u03b5\u03c0\u03b9\u03c4\u03c1\u03ad\u03c0\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03c0\u03b5\u03c1\u03ac\u03c3\u03b5\u03b9, tests \u03b3\u03b9\u03b1 data leakage \u03ba\u03b1\u03b9 isolation, versioning \u03c4\u03bf\u03c5 translator, monitoring \u03b1\u03bd\u03ac model pair \u03ba\u03b1\u03b9 fallback \u03c3\u03b5 \u03b1\u03bd\u03b1\u03b3\u03bd\u03ce\u03c3\u03b9\u03bc\u03b7 text \u03b5\u03c0\u03b9\u03ba\u03bf\u03b9\u03bd\u03c9\u03bd\u03af\u03b1. \u0397 <a href=\"https:\/\/twodots.gr\/agentic-ai-security-stack-controls-technologikes-ependyseis\/\">\u03b1\u03c3\u03c6\u03ac\u03bb\u03b5\u03b9\u03b1 \u03b5\u03bd\u03cc\u03c2 agentic AI stack<\/a> \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 \u03c4\u03cc\u03c3\u03bf \u03c4\u03b1 \u03b5\u03c1\u03b3\u03b1\u03bb\u03b5\u03af\u03b1 \u03ba\u03b1\u03b9 \u03c4\u03b9\u03c2 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b5\u03c2 \u03cc\u03c3\u03bf \u03ba\u03b1\u03b9 \u03c4\u03b1 \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03b1 states, \u03c4\u03b1 logs \u03ba\u03b1\u03b9 \u03c4\u03bf retention \u03c4\u03bf\u03c5\u03c2.<\/p>\n<p>\u03a4\u03bf fallback \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03ad\u03bd\u03b4\u03b5\u03b9\u03be\u03b7 \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03af\u03b1\u03c2. \u0395\u03af\u03bd\u03b1\u03b9 \u03bf \u03c4\u03c1\u03cc\u03c0\u03bf\u03c2 \u03bc\u03b5 \u03c4\u03bf\u03bd \u03bf\u03c0\u03bf\u03af\u03bf \u03c4\u03bf \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03ad\u03c6\u03b5\u03b9 \u03c3\u03b5 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03b7 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1 \u03cc\u03c4\u03b1\u03bd \u03bb\u03b5\u03af\u03c0\u03b5\u03b9 cache access, \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03bf model pair, \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 distribution shift \u03ae \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af\u03c4\u03b1\u03b9 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b7 \u03ad\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7. \u03a4\u03bf \u03af\u03b4\u03b9\u03bf \u03b9\u03c3\u03c7\u03cd\u03b5\u03b9 \u03c3\u03c4\u03b7 <a href=\"https:\/\/twodots.gr\/ai-agents-behavioral-testing\/\">test comportemental des agents IA<\/a>: \u03b7 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1 \u03c4\u03b7\u03c2 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae\u03c2 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7\u03c2 \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af \u03b1\u03bd \u03b7 \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae \u03c0\u03b1\u03c1\u03b1\u03ba\u03ac\u03bc\u03c0\u03c4\u03b5\u03b9 permissions, verification \u03ae human handoff.<\/p>\n<h2 id=\"pilot-axiologisis\">\u0388\u03bd\u03b1 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc pilot \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7\u03c2<\/h2>\n<p>\u0397 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03c0\u03c1\u03ce\u03c4\u03b1 \u03bd\u03b1 \u03b2\u03c1\u03b5\u03b9 handoffs \u03cc\u03c0\u03bf\u03c5 \u03c4\u03bf text decoding \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03b5\u03c4\u03c1\u03b7\u03bc\u03ad\u03bd\u03bf bottleneck \u03ba\u03b1\u03b9 \u03cc\u03c0\u03bf\u03c5 \u03ba\u03b1\u03b9 \u03c4\u03b1 \u03b4\u03cd\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ac \u03bd\u03b1 \u03b5\u03ba\u03b8\u03ad\u03c3\u03bf\u03c5\u03bd \u03c4\u03b9\u03c2 \u03b1\u03c0\u03b1\u03c1\u03b1\u03af\u03c4\u03b7\u03c4\u03b5\u03c2 caches. \u0388\u03c0\u03b5\u03b9\u03c4\u03b1 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 paired benchmark \u03bc\u03b5 \u03c4\u03b1 \u03af\u03b4\u03b9\u03b1 tasks, \u03c4\u03b1 \u03af\u03b4\u03b9\u03b1 ordered model pairs, \u03af\u03b4\u03b9\u03bf hardware \u03ba\u03b1\u03b9 \u03c3\u03b1\u03c6\u03ad\u03c2 text-to-text baseline.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">\u0395\u03c0\u03c4\u03ac \u03b2\u03ae\u03bc\u03b1\u03c4\u03b1 \u03b3\u03b9\u03b1 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 XKV pilot<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">\u00c9tape 1<\/span><strong>\u039f\u03c1\u03af\u03c3\u03c4\u03b5 \u03c4\u03bf cross-context handoff<\/strong>\n<p>\u039a\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 \u03c0\u03bf\u03b9\u03b1 \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03c9\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03ba\u03c1\u03b1\u03c4\u03ac \u03bf Sharer, \u03c4\u03b9 \u03b3\u03bd\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03bf Receiver \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1\u03c4\u03af \u03ba\u03b1\u03bd\u03ad\u03bd\u03b1 context \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf\u03c5.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u00c9tape 2<\/span><strong>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 \u03c4\u03bf text baseline<\/strong>\n<p>\u03a3\u03c5\u03bb\u03bb\u03ad\u03be\u03c4\u03b5 \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7\u03c2, p50 \u03ba\u03b1\u03b9 p95 latency, tokens, prefill, decoding \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc workload.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u00c9tape 3<\/span><strong>\u039a\u03bb\u03b5\u03b9\u03b4\u03ce\u03c3\u03c4\u03b5 \u03c4\u03bf model pair<\/strong>\n<p>\u039f\u03c1\u03af\u03c3\u03c4\u03b5 \u03b1\u03ba\u03c1\u03b9\u03b2\u03b5\u03af\u03c2 \u03b5\u03ba\u03b4\u03cc\u03c3\u03b5\u03b9\u03c2 Sharer \u03ba\u03b1\u03b9 Receiver, tokenizer, layer count \u03ba\u03b1\u03b9 KV geometry \u03ce\u03c3\u03c4\u03b5 \u03c4\u03bf learned bridge \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03b3\u03ce\u03b3\u03b9\u03bc\u03bf.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u00c9tape 4<\/span><strong>\u0395\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c3\u03c4\u03b5 \u03ba\u03b1\u03b9 \u03b1\u03c0\u03bf\u03bc\u03bf\u03bd\u03ce\u03c3\u03c4\u03b5 \u03c4\u03bf\u03bd translator<\/strong>\n<p>\u039a\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 \u03c4\u03b1 base models frozen, \u03b4\u03b9\u03b1\u03c7\u03c9\u03c1\u03af\u03c3\u03c4\u03b5 training \u03ba\u03b1\u03b9 evaluation evidence \u03ba\u03b1\u03b9 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 version, data lineage \u03ba\u03b1\u03b9 configuration.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u00c9tape 5<\/span><strong>\u0395\u03bb\u03ad\u03b3\u03be\u03c4\u03b5 leakage \u03ba\u03b1\u03b9 \u03b1\u03c3\u03c4\u03bf\u03c7\u03af\u03b5\u03c2<\/strong>\n<p>\u0394\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b5 \u03bc\u03b7 \u03b5\u03c0\u03b9\u03c4\u03c1\u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03bf context, missing evidence, adversarial inputs, stale caches \u03ba\u03b1\u03b9 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ad\u03c2 model version \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03c7\u03c1\u03ae\u03c3\u03b7.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u00c9tape 6<\/span><strong>\u0394\u03b9\u03b1\u03c4\u03b7\u03c1\u03ae\u03c3\u03c4\u03b5 \u03b1\u03bd\u03b1\u03b3\u03bd\u03ce\u03c3\u03b9\u03bc\u03bf observability<\/strong>\n<p>\u039a\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 input provenance, output, latency, model pair, translator version \u03ba\u03b1\u03b9 \u03bb\u03cc\u03b3\u03bf fallback \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b5\u03ba\u03b8\u03ad\u03c4\u03b5\u03c4\u03b5 \u03b5\u03c5\u03b1\u03af\u03c3\u03b8\u03b7\u03c4\u03b1 states.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">\u00c9tape 7<\/span><strong>\u0392\u03ac\u03bb\u03c4\u03b5 text \u03ba\u03b1\u03b9 human fallback<\/strong>\n<p>\u0395\u03c0\u03b9\u03c3\u03c4\u03c1\u03ad\u03c8\u03c4\u03b5 \u03c3\u03b5 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03bf \u03bc\u03ae\u03bd\u03c5\u03bc\u03b1 \u03ae \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b7 \u03ad\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03cc\u03c4\u03b1\u03bd \u03b7 \u03c3\u03c5\u03bc\u03b2\u03b1\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1, \u03b7 \u03b5\u03bc\u03c0\u03b9\u03c3\u03c4\u03bf\u03c3\u03cd\u03bd\u03b7 \u03ae \u03b7 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ae \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae \u03b4\u03b5\u03bd \u03b5\u03c0\u03b1\u03c1\u03ba\u03bf\u03cd\u03bd.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<p>\u0393\u03b9\u03b1 workflows \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 \u03b5\u03c4\u03b1\u03b9\u03c1\u03b9\u03ba\u03ac \u03ad\u03b3\u03b3\u03c1\u03b1\u03c6\u03b1, \u03c4\u03bf <a href=\"https:\/\/twodots.gr\/generative-question-answering-llm-etairika-eggrafa\/\">generative question answering \u03bc\u03b5 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03b5\u03c2 \u03c0\u03b7\u03b3\u03ad\u03c2<\/a> \u03c0\u03c1\u03bf\u03c3\u03c6\u03ad\u03c1\u03b5\u03b9 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03cc \u03c0\u03b1\u03c1\u03ac\u03b4\u03b5\u03b9\u03b3\u03bc\u03b1 \u03b3\u03b9\u03b1 permissions, provenance \u03ba\u03b1\u03b9 human review. \u03a4\u03bf latent channel \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03b5\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c7\u03c1\u03cc\u03bd\u03bf, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03c6\u03b1\u03b9\u03c1\u03b5\u03af \u03c4\u03b1 \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1 \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc\u03c2 \u03c7\u03c1\u03ae\u03c3\u03c4\u03b7\u03c2.<\/p>\n<h2 id=\"ti-kratame\">\u03a4\u03b9 \u03ba\u03c1\u03b1\u03c4\u03ac\u03bc\u03b5 \u03b1\u03c0\u03cc \u03c4\u03bf XKV<\/h2>\n<p>\u03a4\u03bf XKV \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03b5\u03c0\u03b9\u03ba\u03bf\u03b9\u03bd\u03c9\u03bd\u03af\u03b1 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 \u03b5\u03c4\u03b5\u03c1\u03bf\u03b3\u03b5\u03bd\u03ae LLMs \u03b4\u03b5\u03bd \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03ac\u03bd\u03c4\u03b1 \u03bd\u03b1 \u03c0\u03b5\u03c1\u03bd\u03ac \u03b1\u03c0\u03cc \u03c6\u03c5\u03c3\u03b9\u03ba\u03ae \u03b3\u03bb\u03ce\u03c3\u03c3\u03b1. \u039c\u03b5 joint pooling \u03c4\u03c9\u03bd \u03b4\u03cd\u03bf caches, \u03b5\u03c5\u03b8\u03c5\u03b3\u03c1\u03ac\u03bc\u03bc\u03b9\u03c3\u03b7 layers \u03ba\u03b1\u03b9 position-specific residual updates, \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc score \u03ba\u03b1\u03b9 \u03c0\u03bf\u03bb\u03cd \u03c7\u03b1\u03bc\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf communication latency \u03b1\u03c0\u03cc \u03c4\u03b1 \u03b4\u03cd\u03bf baselines \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03b9\u03ba\u03cc setup.<\/p>\n<p>\u03a4\u03bf \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03cc \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03cc \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u00ab\u03ba\u03b1\u03c4\u03b1\u03c1\u03b3\u03ae\u03c3\u03c4\u03b5 \u03c4\u03b1 \u03bc\u03b7\u03bd\u03cd\u03bc\u03b1\u03c4\u03b1\u00bb. \u0395\u03af\u03bd\u03b1\u03b9 \u03bd\u03b1 \u03be\u03b5\u03c7\u03c9\u03c1\u03af\u03c3\u03bf\u03c5\u03bc\u03b5 \u03c4\u03b1 machine-to-machine handoffs \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b7\u03c4\u03b1 \u03b1\u03c0\u03cc \u03c4\u03b1 \u03c3\u03b7\u03bc\u03b5\u03af\u03b1 \u03cc\u03c0\u03bf\u03c5 \u03b1\u03c0\u03b1\u03b9\u03c4\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 auditability, \u03b5\u03be\u03ae\u03b3\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b7 \u03b5\u03c5\u03b8\u03cd\u03bd\u03b7. \u0397 <a href=\"https:\/\/twodots.gr\/consilience-ai-agent-communication-control\/\">\u03b5\u03c0\u03b9\u03ba\u03bf\u03b9\u03bd\u03c9\u03bd\u03af\u03b1 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 AI agents<\/a> \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03c9\u03c2 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf\u03c2 \u03b4\u03af\u03b1\u03c5\u03bb\u03bf\u03c2 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03c9\u03c2 \u03b1\u03cc\u03c1\u03b1\u03c4\u03b7 \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03ae \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03ad\u03c1\u03b5\u03b9\u03b1.<\/p>\n<p>\u0397 <a href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">TWO DOTS \u03c3\u03c7\u03b5\u03b4\u03b9\u03ac\u03b6\u03b5\u03b9 \u0391\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2 \u0395\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03ae\u03c3\u03b5\u03c9\u03bd &amp; AI<\/a> \u03bc\u03b5 \u03c7\u03b1\u03c1\u03c4\u03bf\u03b3\u03c1\u03ac\u03c6\u03b7\u03c3\u03b7 \u03c1\u03bf\u03ae\u03c2, permissions, baselines, error handling, monitoring \u03ba\u03b1\u03b9 human handoff \u03c0\u03c1\u03b9\u03bd \u03ad\u03bd\u03b1 \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b1\u03c0\u03bf\u03ba\u03c4\u03ae\u03c3\u03b5\u03b9 \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7 \u03c3\u03b5 CRM, e-shop, ERP, email \u03ae \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03ac \u03b5\u03c1\u03b3\u03b1\u03bb\u03b5\u03af\u03b1.<\/p>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<p class=\"td-service-cta-kicker\">\u0391\u03c0\u03cc \u03c4\u03bf paper \u03c3\u03c4\u03bf \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf pilot<\/p>\n<p class=\"td-service-cta-title\">\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 \u03c4\u03bf handoff \u03c4\u03c9\u03bd AI agents \u03c0\u03c1\u03b9\u03bd \u03b1\u03bb\u03bb\u03ac\u03be\u03b5\u03c4\u03b5 \u03c4\u03bf\u03bd \u03b4\u03af\u03b1\u03c5\u03bb\u03bf \u03b5\u03c0\u03b9\u03ba\u03bf\u03b9\u03bd\u03c9\u03bd\u03af\u03b1\u03c2<\/p>\n<p>\u0397 TWO DOTS \u03c7\u03b1\u03c1\u03c4\u03bf\u03b3\u03c1\u03b1\u03c6\u03b5\u03af cross-context \u03c1\u03bf\u03ad\u03c2, text baselines, model pairs, permissions, observability \u03ba\u03b1\u03b9 fallback \u03ce\u03c3\u03c4\u03b5 \u03ad\u03bd\u03b1 AI workflow \u03bd\u03b1 \u03ba\u03c1\u03b9\u03b8\u03b5\u03af \u03c3\u03c4\u03bf \u03b4\u03b9\u03ba\u03cc \u03c3\u03b1\u03c2 latency, \u03c3\u03c4\u03b1 \u03b4\u03b9\u03ba\u03ac \u03c3\u03b1\u03c2 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03cc \u03c1\u03af\u03c3\u03ba\u03bf.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">Discutez de votre workflow d'IA<\/a><\/div>\n<\/div>\n<\/section>\n<section id=\"sychnes-erotiseis\" class=\"td-faq-section\">\n<div class=\"td-faq\">\n<p class=\"td-faq-heading\">Foire aux questions (FAQ)<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf XKV;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf XKV \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c3\u03b9\u03bc\u03bf\u03c2 translator \u03c0\u03bf\u03c5 \u03c3\u03c5\u03bd\u03b4\u03c5\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b9\u03c2 KV caches \u03b4\u03cd\u03bf frozen language models \u03c3\u03b5 \u03ba\u03bf\u03b9\u03bd\u03ae latent \u03bc\u03bd\u03ae\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 residual updates \u03c3\u03c4\u03bf native cache format \u03c4\u03bf\u03c5 Receiver.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c4\u03b1 \u03b4\u03cd\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03af\u03b4\u03b9\u03bf tokenizer \u03ae \u03af\u03b4\u03b9\u03bf \u03b2\u03ac\u03b8\u03bf\u03c2;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9. \u0397 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03c3\u03c7\u03b5\u03b4\u03b9\u03ac\u03c3\u03c4\u03b7\u03ba\u03b5 \u03b3\u03b9\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03c0\u03bf\u03c5 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c6\u03ad\u03c1\u03bf\u03c5\u03bd \u03c3\u03b5 \u03bf\u03b9\u03ba\u03bf\u03b3\u03ad\u03bd\u03b5\u03b9\u03b1, \u03b1\u03c1\u03b9\u03b8\u03bc\u03cc layers, KV heads, head dimension \u03ba\u03b1\u03b9 tokenizer, \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ce\u03bd\u03c4\u03b1\u03c2 learned \u03b5\u03c5\u03b8\u03c5\u03b3\u03c1\u03ac\u03bc\u03bc\u03b9\u03c3\u03b7 \u03c0\u03c1\u03bf\u03c2 \u03c4\u03bf\u03bd Receiver.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf XKV \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b1\u03bd\u03c4\u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03ba\u03b5\u03b9\u03bc\u03ad\u03bd\u03bf\u03c5;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0391\u03c0\u03bf\u03c6\u03b5\u03cd\u03b3\u03b5\u03b9 \u03c4\u03bf autoregressive decoding \u03c4\u03bf\u03c5 \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03bf\u03c5 \u03bc\u03b7\u03bd\u03cd\u03bc\u03b1\u03c4\u03bf\u03c2 \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03b5\u03c0\u03b1\u03bd\u03b5\u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03ae \u03c4\u03bf\u03c5. \u03a3\u03c4\u03b7 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03ae\u03c4\u03b1\u03bd 6,8\u00d7 \u03c4\u03b1\u03c7\u03cd\u03c4\u03b5\u03c1\u03bf end to end \u03b1\u03c0\u03cc \u03c4\u03bf text-to-text baseline.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03cc\u03c3\u03bf \u03b3\u03c1\u03ae\u03b3\u03bf\u03c1\u03bf\u03c2 \u03ae\u03c4\u03b1\u03bd \u03bf translator;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039f \u03bc\u03ad\u03c3\u03bf\u03c2 \u03c7\u03c1\u03cc\u03bd\u03bf\u03c2 \u03ae\u03c4\u03b1\u03bd 5,8 ms \u03c3\u03c4\u03b1 45 dataset\u2013model-pair cells, \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 59,9 ms \u03b3\u03b9\u03b1 \u03c4\u03bf\u03bd LCF-X fusor. \u039f\u03b9 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03b1\u03c6\u03bf\u03c1\u03bf\u03cd\u03bd \u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf A100 setup \u03ba\u03b1\u03b9 \u03bc\u03af\u03b1 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03b1\u03bd\u03ac \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u039a\u03ad\u03c1\u03b4\u03b9\u03c3\u03b5 \u03c4\u03bf XKV \u03c3\u03b5 \u03cc\u03bb\u03b1 \u03c4\u03b1 datasets;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039e\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03ba\u03b1\u03c4\u03ac \u03bc\u03ad\u03c3\u03bf \u03cc\u03c1\u03bf \u03c4\u03bf LCF-X \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b1 \u03c0\u03ad\u03bd\u03c4\u03b5 datasets \u03ba\u03b1\u03b9 \u03c4\u03bf text-to-text \u03c3\u03c4\u03b1 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b1 \u03b1\u03c0\u03cc \u03c4\u03b1 \u03c0\u03ad\u03bd\u03c4\u03b5. \u03a3\u03c4\u03bf QASC \u03c4\u03bf text baseline \u03ae\u03c4\u03b1\u03bd \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf \u03ba\u03b1\u03c4\u03ac 0,27 \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03b9\u03b1\u03af\u03b5\u03c2 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u039c\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b7\u03b8\u03b5\u03af \u03bc\u03b5 \u03ba\u03bb\u03b5\u03b9\u03c3\u03c4\u03ac LLM APIs;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9 \u03cc\u03c4\u03b1\u03bd \u03c4\u03bf API \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03ad\u03c6\u03b5\u03b9 \u03bc\u03cc\u03bd\u03bf \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03b4\u03af\u03bd\u03b5\u03b9 \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7 \u03c3\u03c4\u03b9\u03c2 \u03b1\u03c0\u03b1\u03c1\u03b1\u03af\u03c4\u03b7\u03c4\u03b5\u03c2 KV caches \u03ae \u03c3\u03b5 \u03b9\u03c3\u03bf\u03b4\u03cd\u03bd\u03b1\u03bc\u03bf \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03cc state. \u03a4\u03bf XKV \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 prompt-layer \u03c0\u03c1\u03c9\u03c4\u03cc\u03ba\u03bf\u03bb\u03bb\u03bf.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03bf\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc\u03c2 \u03ba\u03af\u03bd\u03b4\u03c5\u03bd\u03bf\u03c2 \u03c4\u03b7\u03c2 latent \u03b5\u03c0\u03b9\u03ba\u03bf\u03b9\u03bd\u03c9\u03bd\u03af\u03b1\u03c2;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf \u03bc\u03b5\u03c4\u03b1\u03c6\u03b5\u03c1\u03cc\u03bc\u03b5\u03bd\u03bf state \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b1 \u03b1\u03bd\u03b1\u03b3\u03bd\u03ce\u03c3\u03b9\u03bc\u03bf, \u03ac\u03c1\u03b1 debugging, provenance, data isolation \u03ba\u03b1\u03b9 compliance \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03cc observability layer \u03ba\u03b1\u03b9 \u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 fallback \u03c3\u03b5 \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03ae \u03ac\u03bd\u03b8\u03c1\u03c9\u03c0\u03bf.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0395\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf XKV \u03ad\u03c4\u03bf\u03b9\u03bc\u03bf \u03b3\u03b9\u03b1 enterprise \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03b9\u03ba\u03ae \u03b4\u03c5\u03bd\u03b1\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c3\u03b5 \u03bc\u03b9\u03ba\u03c1\u03ac open models \u03ba\u03b1\u03b9 split-evidence QA. \u03a0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03ae \u03c7\u03c1\u03ae\u03c3\u03b7 \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af pair-specific \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7, versioning, \u03b4\u03b9\u03ba\u03ac \u03c3\u03b1\u03c2 benchmarks, security tests, monitoring \u03ba\u03b1\u03b9 human-readable fallback.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">Sources d'information<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.20617\" target=\"_blank\" rel=\"noopener\">Liu et al. \u2014 Dual-Cache Latent Space Communication between Heterogeneous Language Models, arXiv v1<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/html\/2608.20617\" target=\"_blank\" rel=\"noopener\">Liu et al. \u2014 \u03a0\u03bb\u03ae\u03c1\u03b5\u03c2 HTML paper \u03bc\u03b5 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf, \u03c0\u03af\u03bd\u03b1\u03ba\u03b5\u03c2, ablations \u03ba\u03b1\u03b9 technical appendix<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2605.22863\" target=\"_blank\" rel=\"noopener\">Rossi et al. \u2014 Latent Cache Flow: Model-to-Model Communication without Text<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2601.06123\" target=\"_blank\" rel=\"noopener\">Fu et al. \u2014 Latent Space Communication via K-V Cache Alignment<\/a><\/li>\n<li><a href=\"https:\/\/www.nist.gov\/itl\/ai-risk-management-framework\" target=\"_blank\" rel=\"noopener\">NIST \u2014 AI Risk Management Framework \u03b3\u03b9\u03b1 governance, \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c7\u03b5\u03af\u03c1\u03b9\u03c3\u03b7 \u03ba\u03b9\u03bd\u03b4\u03cd\u03bd\u03bf\u03c5<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>Le XKV transf\u00e8re des connaissances entre des agents d'IA via des caches KV, r\u00e9duisant ainsi la latence sans recourir \u00e0 un tokeniseur commun. D\u00e9couvrez les r\u00e9sultats, les limites et les conditions pour un projet pilote s\u00e9curis\u00e9.<\/p>","protected":false},"author":1,"featured_media":97184,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","footnotes":""},"categories":[199],"tags":[6506,20246,20247,20245,20244],"class_list":["post-96262","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-ai-agents","tag-kv-cache","tag-latent-communication","tag-multi-agent-llm","tag-xkv"],"_links":{"self":[{"href":"https:\/\/twodots.gr\/fr\/wp-json\/wp\/v2\/posts\/96262","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/fr\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/fr\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/fr\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/fr\/wp-json\/wp\/v2\/comments?post=96262"}],"version-history":[{"count":0,"href":"https:\/\/twodots.gr\/fr\/wp-json\/wp\/v2\/posts\/96262\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/fr\/wp-json\/wp\/v2\/media\/97184"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/fr\/wp-json\/wp\/v2\/media?parent=96262"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/fr\/wp-json\/wp\/v2\/categories?post=96262"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/fr\/wp-json\/wp\/v2\/tags?post=96262"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}