{"id":90851,"date":"2026-09-02T16:29:05","date_gmt":"2026-09-02T13:29:05","guid":{"rendered":"https:\/\/twodots.gr\/?p=90851"},"modified":"2026-09-02T16:29:05","modified_gmt":"2026-09-02T13:29:05","slug":"why2speak-exigisi-ai-agent-allazei-apofasi","status":"publish","type":"post","link":"https:\/\/twodots.gr\/en\/why2speak-exigisi-ai-agent-allazei-apofasi\/","title":{"rendered":"Why2Speak: \u03b3\u03b9\u03b1\u03c4\u03af \u03b7 \u03b5\u03be\u03ae\u03b3\u03b7\u03c3\u03b7 \u03b5\u03bd\u03cc\u03c2 AI agent \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03ae \u03c4\u03bf\u03c5"},"content":{"rendered":"<div class=\"td-article-lede\">\n<p><strong>\u03a4\u03bf Why2Speak \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03b1\u03c0\u03b1\u03af\u03c4\u03b7\u03c3\u03b7 \u00ab\u03b5\u03be\u03ae\u03b3\u03b7\u03c3\u03b5 \u03c0\u03c1\u03b9\u03bd \u03b5\u03bd\u03b5\u03c1\u03b3\u03ae\u03c3\u03b5\u03b9\u03c2\u00bb \u03b4\u03b5\u03bd \u03b1\u03bd\u03bf\u03af\u03b3\u03b5\u03b9 \u03b1\u03c0\u03bb\u03ce\u03c2 \u03ad\u03bd\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b8\u03c5\u03c1\u03bf \u03c3\u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7: \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae \u03b4\u03c1\u03ac\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03b1\u03c0\u03bf\u03c7\u03ae\u03c2.<\/strong> \u03a3\u03c4\u03bf Qwen3-8B \u03bf \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf\u03c2 direct classifier \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 macro-F1 0,620, \u03b5\u03bd\u03ce \u03b7 inspectable think-mode \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae 0,536. \u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7, \u03bb\u03bf\u03b9\u03c0\u03cc\u03bd, \u03b7 \u03b5\u03bb\u03b5\u03b3\u03be\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c0\u03b5\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc rationale\u00b7 \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03b1\u03ba\u03c1\u03b9\u03b2\u03bf\u03cd\u03c2 deployed mode, \u03c4\u03c9\u03bd \u03c0\u03b5\u03c1\u03b9\u03c4\u03c4\u03ce\u03bd \u03c0\u03b1\u03c1\u03b5\u03bc\u03b2\u03ac\u03c3\u03b5\u03c9\u03bd \u03ba\u03b1\u03b9 \u03c4\u03c9\u03bd \u03c7\u03b1\u03bc\u03ad\u03bd\u03c9\u03bd \u03b5\u03c5\u03ba\u03b1\u03b9\u03c1\u03b9\u03ce\u03bd.<\/p>\n<\/div>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">Contents<\/div>\n<ul>\n<li><a href=\"#i-siopi-einai-apofasi\">\u0397 \u03c3\u03b9\u03c9\u03c0\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7, \u03cc\u03c7\u03b9 \u03b1\u03c0\u03bf\u03c5\u03c3\u03af\u03b1 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7\u03c2<\/a><\/li>\n<li><a href=\"#peirama-why2speak-when2speak\">\u03a4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03c4\u03bf\u03c5 Why2Speak \u03ba\u03b1\u03b9 \u03c4\u03bf When2Speak<\/a><\/li>\n<li><a href=\"#ikanotita-elegksimotita\">\u0399\u03ba\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03b5\u03bb\u03b5\u03b3\u03be\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b4\u03b5\u03bd \u03c3\u03c5\u03bc\u03c0\u03af\u03c0\u03c4\u03bf\u03c5\u03bd<\/a><\/li>\n<li><a href=\"#pote-schimatizetai-apofasi\">\u03a0\u03cc\u03c4\u03b5 \u03c3\u03c7\u03b7\u03bc\u03b1\u03c4\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03b7 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7<\/a><\/li>\n<li><a href=\"#afairesi-chain-of-thought\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03b1\u03c6\u03b1\u03af\u03c1\u03b5\u03c3\u03b7 \u03c4\u03bf\u03c5 chain of thought \u03c0\u03b1\u03c1\u03b1\u03c0\u03bb\u03b1\u03bd\u03ac<\/a><\/li>\n<li><a href=\"#sostes-exigiseis-pistotita\">\u039f\u03b9 \u00ab\u03c3\u03c9\u03c3\u03c4\u03ad\u03c2\u00bb \u03b5\u03be\u03b7\u03b3\u03ae\u03c3\u03b5\u03b9\u03c2 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03b1\u03c4\u2019 \u03b1\u03bd\u03ac\u03b3\u03ba\u03b7 \u03c0\u03b9\u03c3\u03c4\u03ad\u03c2<\/a><\/li>\n<li><a href=\"#eksi-elegchoi-axiologisis\">\u0388\u03be\u03b9 \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03b9 \u03b3\u03b9\u03b1 \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03b7 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7<\/a><\/li>\n<li><a href=\"#business-ai-agents-roes\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1 AI agents \u03c3\u03b5 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ad\u03c2 \u03c1\u03bf\u03ad\u03c2<\/a><\/li>\n<li><a href=\"#plaisio-ai-oversight\">\u0388\u03bd\u03b1 \u03c0\u03b9\u03bf \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03bf \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf AI oversight<\/a><\/li>\n<li><a href=\"#oria-sosto-symperasma\">\u03a4\u03b1 \u03cc\u03c1\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03c4\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1<\/a><\/li>\n<\/ul>\n<\/div>\n<p>\u039f\u03b9 Shreya Mendi \u03ba\u03b1\u03b9 Brinnae Bent, \u03b1\u03c0\u03cc \u03c4\u03bf Pratt School of Engineering \u03c4\u03bf\u03c5 Duke University, \u03bc\u03b5\u03bb\u03b5\u03c4\u03bf\u03cd\u03bd \u03c4\u03b7\u03bd \u03c0\u03b9\u03c3\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03b7\u03c2 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae\u03c2 \u03c3\u03b5 \u03ad\u03bd\u03b1 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf \u03b1\u03bb\u03bb\u03ac \u03b9\u03b4\u03b9\u03b1\u03af\u03c4\u03b5\u03c1\u03b1 \u03b1\u03bd\u03c4\u03b9\u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b5\u03c5\u03c4\u03b9\u03ba\u03cc \u03c0\u03b5\u03c1\u03b9\u03b2\u03ac\u03bb\u03bb\u03bf\u03bd: \u03ad\u03bd\u03b1\u03bd \u03b2\u03bf\u03b7\u03b8\u03cc \u03c0\u03bf\u03c5 \u03c0\u03b1\u03c1\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03b5\u03af \u03bc\u03b9\u03b1 \u03c0\u03bf\u03bb\u03c5\u03bc\u03b5\u03c1\u03ae \u03c3\u03c5\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b9\u03bb\u03ad\u03b3\u03b5\u03b9 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 \u03b4\u03cd\u03bf \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b5\u03c2, \u03bd\u03b1 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b5\u03b9 \u03c4\u03ce\u03c1\u03b1 \u03ae \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bc\u03b5\u03af\u03bd\u03b5\u03b9 \u03c3\u03b9\u03c9\u03c0\u03b7\u03bb\u03cc\u03c2. \u0397 \u03c0\u03b5\u03c1\u03af\u03c0\u03c4\u03c9\u03c3\u03b7 \u03b1\u03c5\u03c4\u03ae \u03c6\u03c9\u03c4\u03af\u03b6\u03b5\u03b9 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ae\u03bc\u03b1\u03c4\u03b1 \u03c0\u03bf\u03c5 \u03b8\u03b1 \u03c3\u03c5\u03bd\u03b1\u03bd\u03c4\u03ae\u03c3\u03bf\u03c5\u03bd \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03cc\u03c4\u03b1\u03bd \u03b1\u03bd\u03b1\u03b8\u03ad\u03c4\u03bf\u03c5\u03bd \u03c3\u03b5 <a href=\"https:\/\/twodots.gr\/ai-agents-behavioral-testing\/\">agents \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 behavioral testing<\/a> \u03c4\u03b7\u03bd \u03c0\u03b1\u03c1\u03b1\u03ba\u03bf\u03bb\u03bf\u03cd\u03b8\u03b7\u03c3\u03b7 \u03c3\u03c5\u03bd\u03bf\u03bc\u03b9\u03bb\u03b9\u03ce\u03bd, \u03b4\u03b9\u03b1\u03b4\u03b9\u03ba\u03b1\u03c3\u03b9\u03ce\u03bd, \u03b5\u03be\u03c5\u03c0\u03b7\u03c1\u03ad\u03c4\u03b7\u03c3\u03b7\u03c2 \u03ae \u03ac\u03bb\u03bb\u03c9\u03bd \u03c1\u03bf\u03ce\u03bd \u03cc\u03c0\u03bf\u03c5 \u03b7 \u03b1\u03b4\u03c1\u03ac\u03bd\u03b5\u03b9\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae.<\/p>\n<h2 id=\"i-siopi-einai-apofasi\">\u0397 \u03c3\u03b9\u03c9\u03c0\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7, \u03cc\u03c7\u03b9 \u03b1\u03c0\u03bf\u03c5\u03c3\u03af\u03b1 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7\u03c2<\/h2>\n<p>\u03a3\u03c4\u03b1 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b1 benchmarks \u03b5\u03c1\u03c9\u03c4\u03ae\u03c3\u03b5\u03c9\u03bd \u03ba\u03b1\u03b9 \u03b1\u03c0\u03b1\u03bd\u03c4\u03ae\u03c3\u03b5\u03c9\u03bd, \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ba\u03b1\u03bb\u03b5\u03af\u03c4\u03b1\u03b9 \u03bf\u03c0\u03c9\u03c3\u03b4\u03ae\u03c0\u03bf\u03c4\u03b5 \u03bd\u03b1 \u03b4\u03ce\u03c3\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7. \u03a3\u03c4\u03b7\u03bd \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03af\u03b1 \u03b5\u03bd\u03cc\u03c2 agent, \u03cc\u03bc\u03c9\u03c2, \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03c5\u03bd \u03c3\u03c4\u03b9\u03b3\u03bc\u03ad\u03c2 \u03cc\u03c0\u03bf\u03c5 \u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bd\u03b1 \u03bc\u03b7\u03bd \u03ba\u03ac\u03bd\u03b5\u03b9 \u03c4\u03af\u03c0\u03bf\u03c4\u03b1. \u0388\u03bd\u03b1\u03c2 \u03ba\u03bb\u03b9\u03bd\u03b9\u03ba\u03cc\u03c2 \u03b2\u03bf\u03b7\u03b8\u03cc\u03c2 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c3\u03b7\u03bc\u03ac\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1\u03bd \u03ba\u03af\u03bd\u03b4\u03c5\u03bd\u03bf \u03ae \u03bd\u03b1 \u03bc\u03b7 \u03b4\u03b9\u03b1\u03ba\u03cc\u03c8\u03b5\u03b9 \u03c4\u03bf\u03bd \u03b5\u03c0\u03b1\u03b3\u03b3\u03b5\u03bb\u03bc\u03b1\u03c4\u03af\u03b1. \u0388\u03bd\u03b1 \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b4\u03b9\u03b4\u03b1\u03c3\u03ba\u03b1\u03bb\u03af\u03b1\u03c2 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b2\u03bf\u03b7\u03b8\u03ae\u03c3\u03b5\u03b9 \u03ae \u03bd\u03b1 \u03b1\u03c6\u03ae\u03c3\u03b5\u03b9 \u03c4\u03bf\u03bd \u03bc\u03b1\u03b8\u03b7\u03c4\u03ae \u03bd\u03b1 \u03c3\u03c5\u03bd\u03b5\u03c7\u03af\u03c3\u03b5\u03b9 \u03c4\u03b7\u03bd \u03c0\u03c1\u03bf\u03c3\u03c0\u03ac\u03b8\u03b5\u03b9\u03b1. \u0388\u03bd\u03b1\u03c2 \u03b2\u03bf\u03b7\u03b8\u03cc\u03c2 meeting \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03b9\u03bf\u03c1\u03b8\u03ce\u03c3\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03ae \u03bd\u03b1 \u03bc\u03b7\u03bd \u03b5\u03c0\u03b9\u03b2\u03b1\u03c1\u03cd\u03bd\u03b5\u03b9 \u03c4\u03b7 \u03c3\u03c5\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03bc\u03b5 \u03c0\u03b5\u03c1\u03b9\u03c4\u03c4\u03ae \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7.<\/p>\n<p>\u0391\u03c5\u03c4\u03cc \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03b1\u03c3\u03cd\u03bc\u03bc\u03b5\u03c4\u03c1\u03bf \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2. \u039c\u03b9\u03b1 \u03ac\u03c3\u03ba\u03bf\u03c0\u03b7 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03b9\u03b1\u03ba\u03cc\u03c8\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03b2\u03c1\u03ce\u03c3\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b5\u03bc\u03c0\u03b9\u03c3\u03c4\u03bf\u03c3\u03cd\u03bd\u03b7. \u039c\u03b9\u03b1 \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b7 \u03b1\u03c0\u03bf\u03c7\u03ae \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c6\u03ae\u03c3\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03c3\u03c6\u03ac\u03bb\u03bc\u03b1 \u03ae \u03bc\u03b9\u03b1 \u03b5\u03c5\u03ba\u03b1\u03b9\u03c1\u03af\u03b1 \u03c7\u03c9\u03c1\u03af\u03c2 \u03b1\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03ce\u03c0\u03b9\u03c3\u03b7. \u0395\u03c0\u03bf\u03bc\u03ad\u03bd\u03c9\u03c2, \u03b7 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b1\u03c6\u03bf\u03c1\u03ac \u03b1\u03c0\u03bb\u03ce\u03c2 \u03c4\u03bf \u03b1\u03bd \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u00ab\u03be\u03ad\u03c1\u03b5\u03b9\u00bb \u03ba\u03ac\u03c4\u03b9. \u0391\u03c6\u03bf\u03c1\u03ac \u03c4\u03b7\u03bd \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae \u03bc\u03b5 \u03c4\u03b7\u03bd \u03bf\u03c0\u03bf\u03af\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03c4\u03b7\u03bd \u03ba\u03b1\u03c4\u03b1\u03bd\u03cc\u03b7\u03c3\u03b7 \u03c3\u03b5 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1.<\/p>\n<p>\u0397 \u03c0\u03b9\u03c3\u03c4\u03ae \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03ad\u03c7\u03b5\u03b9 \u03b5\u03b4\u03ce \u03b1\u03c5\u03c3\u03c4\u03b7\u03c1\u03cc\u03c4\u03b5\u03c1\u03b7 \u03ad\u03bd\u03bd\u03bf\u03b9\u03b1 \u03b1\u03c0\u03cc \u03bc\u03b9\u03b1 <a href=\"https:\/\/twodots.gr\/ai-exigei-lathos-chain-of-thought-steering-vectors\/\">\u03c0\u03b5\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03b5\u03be\u03ae\u03b3\u03b7\u03c3\u03b7<\/a>. \u0397 \u03b5\u03be\u03ae\u03b3\u03b7\u03c3\u03b7 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03bd\u03c4\u03b1\u03bd\u03b1\u03ba\u03bb\u03ac \u03c4\u03bf\u03bd \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03cc \u03c0\u03bf\u03c5 \u03b5\u03c0\u03b7\u03c1\u03ad\u03b1\u03c3\u03b5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03c4\u03b7\u03bd \u03c0\u03c1\u03ac\u03be\u03b7 \u03ae \u03c4\u03b7\u03bd \u03b1\u03c0\u03bf\u03c7\u03ae. \u0388\u03bd\u03b1 \u03bb\u03b5\u03ba\u03c4\u03b9\u03ba\u03ac \u03b5\u03cd\u03bb\u03bf\u03b3\u03bf \u03c3\u03ba\u03b5\u03c0\u03c4\u03b9\u03ba\u03cc \u03c0\u03bf\u03c5 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c6\u03bf\u03cd \u03b7 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03ad\u03c7\u03b5\u03b9 \u03ae\u03b4\u03b7 \u03ba\u03bb\u03b5\u03b9\u03b4\u03ce\u03c3\u03b5\u03b9 \u03b5\u03af\u03bd\u03b1\u03b9 rationalization, \u03cc\u03c7\u03b9 \u03b4\u03b9\u03b1\u03c6\u03ac\u03bd\u03b5\u03b9\u03b1.<\/p>\n<h2 id=\"peirama-why2speak-when2speak\">\u03a4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03c4\u03bf\u03c5 Why2Speak \u03ba\u03b1\u03b9 \u03c4\u03bf When2Speak<\/h2>\n<p>\u039f\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ad\u03c2 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd \u03c4\u03bf benchmark When2Speak. \u03a0\u03b5\u03c1\u03b9\u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 16.000 \u03c3\u03c5\u03bd\u03b8\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03c3\u03c5\u03bd\u03bf\u03bc\u03b9\u03bb\u03af\u03b5\u03c2, \u03bf\u03b9 \u03bf\u03c0\u03bf\u03af\u03b5\u03c2 \u03b4\u03af\u03bd\u03bf\u03c5\u03bd 173.325 \u03c3\u03b7\u03bc\u03b5\u03af\u03b1 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2 \u03c3\u03b5 \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03bf token. \u039c\u03cc\u03bd\u03bf \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c4\u03bf 13% \u03c4\u03c9\u03bd \u03c3\u03b7\u03bc\u03b5\u03af\u03c9\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b5\u03c5\u03ba\u03b1\u03b9\u03c1\u03af\u03b5\u03c2 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7\u03c2. \u0391\u03c5\u03c4\u03ae \u03b7 \u03ad\u03bd\u03c4\u03bf\u03bd\u03b7 \u03b1\u03bd\u03b9\u03c3\u03bf\u03c1\u03c1\u03bf\u03c0\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b8\u03b5\u03bc\u03b5\u03bb\u03b9\u03ce\u03b4\u03b7\u03c2: \u03ad\u03bd\u03b1\u03c2 \u03c4\u03b1\u03be\u03b9\u03bd\u03bf\u03bc\u03b7\u03c4\u03ae\u03c2 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03b9 \u03c5\u03c8\u03b7\u03bb\u03ae \u03b1\u03c0\u03bb\u03ae \u03b1\u03ba\u03c1\u03af\u03b2\u03b5\u03b9\u03b1 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c0\u03bf\u03bd\u03c4\u03b1\u03c2 \u03c3\u03c5\u03c7\u03bd\u03ac \u03c4\u03b7 \u03c3\u03b9\u03c9\u03c0\u03ae, \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03b5\u03c2 \u03c3\u03c4\u03b9\u03b3\u03bc\u03ad\u03c2.<\/p>\n<p>\u039f\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c0\u03c4\u03ce\u03c3\u03b5\u03b9\u03c2 \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc 11.498 \u03c3\u03b5\u03bd\u03ac\u03c1\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03bf\u03c5\u03bd \u03c0\u03ad\u03bd\u03c4\u03b5 \u03ba\u03b1\u03c4\u03b7\u03b3\u03bf\u03c1\u03af\u03b5\u03c2 \u03b3\u03bd\u03c9\u03c3\u03c4\u03b9\u03ba\u03bf\u03cd \u03ba\u03b5\u03bd\u03bf\u03cd: \u03b4\u03b9\u03cc\u03c1\u03b8\u03c9\u03c3\u03b7 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03bf\u03bb\u03bf\u03b3\u03b9\u03ba\u03bf\u03cd \u03bb\u03ac\u03b8\u03bf\u03c5\u03c2, \u03bf\u03c1\u03b9\u03c3\u03bc\u03cc \u03ad\u03bd\u03bd\u03bf\u03b9\u03b1\u03c2, \u03c0\u03b1\u03c1\u03bf\u03c7\u03ae \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd, \u03c4\u03b1\u03c5\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03c0\u03b7\u03b3\u03ae\u03c2 \u03ba\u03b1\u03b9 \u03c3\u03cd\u03bd\u03b8\u03b5\u03c3\u03b7 \u03ae \u03b1\u03bd\u03b1\u03c0\u03bb\u03b1\u03b9\u03c3\u03af\u03c9\u03c3\u03b7. \u039f\u03b9 \u03ba\u03b1\u03c4\u03b7\u03b3\u03bf\u03c1\u03af\u03b5\u03c2 \u03b1\u03bd\u03b1\u03ba\u03b1\u03c4\u03b1\u03c3\u03ba\u03b5\u03c5\u03ac\u03c3\u03c4\u03b7\u03ba\u03b1\u03bd \u03b1\u03c0\u03cc \u03c4\u03b1 \u03b1\u03c1\u03c7\u03b9\u03ba\u03ac \u03c3\u03b5\u03bd\u03ac\u03c1\u03b9\u03b1 \u03bc\u03cc\u03bd\u03bf \u03b3\u03b9\u03b1 \u03c4\u03b9\u03c2 \u03b1\u03bd\u03b1\u03bb\u03cd\u03c3\u03b5\u03b9\u03c2 \u03b5\u03c1\u03bc\u03b7\u03bd\u03b5\u03c5\u03c3\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2\u00b7 \u03b4\u03b5\u03bd \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03b8\u03b7\u03ba\u03b1\u03bd \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03c5\u03c4\u03b5\u03af \u03b7 \u03c0\u03c1\u03cc\u03b2\u03bb\u03b5\u03c8\u03b7 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7\u03c2.<\/p>\n<p>\u03a4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf Qwen3-8B, \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03c4\u03b1 \u03af\u03b4\u03b9\u03b1 \u03b2\u03ac\u03c1\u03b7 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03b1\u03c0\u03bf\u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03b7\u03b8\u03bf\u03cd\u03bd \u03c3\u03b5 \u03b4\u03cd\u03bf \u03c4\u03c1\u03cc\u03c0\u03bf\u03c5\u03c2. \u03a3\u03c4\u03bf <em>no-think<\/em> mode \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03b5\u03c5\u03b8\u03b5\u03af\u03b1\u03c2 \u03c4\u03bf token \u03c4\u03b7\u03c2 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2. \u03a3\u03c4\u03bf <em>think<\/em> mode \u03c0\u03c1\u03bf\u03b7\u03b3\u03b5\u03af\u03c4\u03b1\u03b9 \u03b1\u03bb\u03c5\u03c3\u03af\u03b4\u03b1 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae\u03c2 \u03ba\u03b1\u03b9 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03b5\u03af \u03c4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc token. \u0388\u03c4\u03c3\u03b9, \u03b7 \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03b4\u03b5\u03bd \u03bc\u03c0\u03b5\u03c1\u03b4\u03b5\u03cd\u03b5\u03c4\u03b1\u03b9 \u03b5\u03be\u03b1\u03c1\u03c7\u03ae\u03c2 \u03b1\u03c0\u03cc \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae.<\/p>\n<h2 id=\"ikanotita-elegksimotita\">\u0399\u03ba\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03b5\u03bb\u03b5\u03b3\u03be\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b4\u03b5\u03bd \u03c3\u03c5\u03bc\u03c0\u03af\u03c0\u03c4\u03bf\u03c5\u03bd<\/h2>\n<p>\u03a4\u03bf \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03ac\u03bc\u03b5\u03c3\u03bf policy \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf \u03c0\u03b9\u03bf \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03bf. \u039f decision-token classifier, \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03c5\u03bc\u03ad\u03bd\u03bf\u03c2 \u03bc\u03b5 LoRA \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c0\u03b5\u03b9 \u03b1\u03c0\u03b5\u03c5\u03b8\u03b5\u03af\u03b1\u03c2 \u03c4\u03b7\u03bd \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1, \u03c6\u03c4\u03ac\u03bd\u03b5\u03b9 macro-F1 0,620 \u03c3\u03c4\u03bf matched evaluation. \u0394\u03b5\u03bd \u03b5\u03ba\u03b8\u03ad\u03c4\u03b5\u03b9, \u03cc\u03bc\u03c9\u03c2, \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03c0\u03bf\u03c5 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03c0\u03b9\u03b8\u03b5\u03c9\u03c1\u03b7\u03b8\u03b5\u03af. \u0391\u03bd \u03c4\u03bf \u03af\u03b4\u03b9\u03bf checkpoint \u03b1\u03bd\u03b1\u03b3\u03ba\u03b1\u03c3\u03c4\u03b5\u03af \u03bd\u03b1 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03ae\u03c3\u03b5\u03b9 \u03c3\u03b5 think mode, \u03c4\u03bf macro-F1 \u03c0\u03ad\u03c6\u03c4\u03b5\u03b9 \u03c3\u03c4\u03bf 0,367.<\/p>\n<p>\u03a4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc Qwen3-8B \u03c0\u03b1\u03c1\u03bf\u03c5\u03c3\u03b9\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03c1\u03bf\u03c6\u03b7 \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1. \u03a7\u03c9\u03c1\u03af\u03c2 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03c0\u03b1\u03c1\u03b5\u03bc\u03b2\u03b1\u03af\u03bd\u03b5\u03b9 \u03c3\u03c7\u03b5\u03b4\u03cc\u03bd \u03c0\u03ac\u03bd\u03c4\u03b1: macro-F1 0,096 \u03ba\u03b1\u03b9 false-intervention rate 0,989. \u039c\u03b5 think mode \u03c4\u03bf macro-F1 \u03b1\u03bd\u03b5\u03b2\u03b1\u03af\u03bd\u03b5\u03b9 \u03c3\u03c4\u03bf 0,536 \u03ba\u03b1\u03b9 \u03bf\u03b9 \u03c8\u03b5\u03c5\u03b4\u03b5\u03af\u03c2 \u03c0\u03b1\u03c1\u03b5\u03bc\u03b2\u03ac\u03c3\u03b5\u03b9\u03c2 \u03bc\u03b5\u03b9\u03ce\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf 0,232. \u03a0\u03b1\u03c1\u03ac\u03bb\u03bb\u03b7\u03bb\u03b1, \u03cc\u03bc\u03c9\u03c2, \u03b1\u03c5\u03be\u03ac\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bf\u03b9 \u03c7\u03b1\u03bc\u03ad\u03bd\u03b5\u03c2 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b5\u03c5\u03ba\u03b1\u03b9\u03c1\u03af\u03b5\u03c2 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7\u03c2, \u03bc\u03b5 missed-intervention rate 0,579.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-title\">\u03a4\u03bf Why2Speak \u03c3\u03b5 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03b5\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03c5\u03bc\u03ad\u03bd\u03b5\u03c2 \u03bc\u03b5\u03c4\u03c1\u03b9\u03ba\u03ad\u03c2<\/p>\n<p class=\"td-chart-subtitle\">\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 matched evaluation \u03c4\u03bf\u03c5 paper \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc benchmark \u03b3\u03b9\u03b1 \u03cc\u03bb\u03b1 \u03c4\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03ae \u03cc\u03bb\u03b1 \u03c4\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ac workloads.<\/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\">173.325<\/span><span class=\"td-metric-label\">token-level \u03c3\u03b7\u03bc\u03b5\u03af\u03b1 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2 \u03c3\u03c4\u03bf dataset \u03c4\u03b7\u03c2 \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7\u03c2<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">\u224813%<\/span><span class=\"td-metric-label\">\u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b5\u03c5\u03ba\u03b1\u03b9\u03c1\u03af\u03b5\u03c2 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7\u03c2<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">0,620<\/span><span class=\"td-metric-label\">macro-F1 \u03c4\u03bf\u03c5 direct token-SFT classifier<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">0,536<\/span><span class=\"td-metric-label\">macro-F1 \u03c4\u03b7\u03c2 base think-mode \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae\u03c2<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<p>\u03a4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b4\u03b5\u03bd \u03bb\u03ad\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ac \u03ba\u03b1\u03ba\u03ae \u03ae \u03cc\u03c4\u03b9 \u03b7 \u03ac\u03bc\u03b5\u03c3\u03b7 \u03c0\u03c1\u03cc\u03b2\u03bb\u03b5\u03c8\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03ac\u03bd\u03c4\u03b1 \u03c0\u03c1\u03bf\u03c4\u03b9\u03bc\u03cc\u03c4\u03b5\u03c1\u03b7. \u0394\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03ba\u03ac\u03c4\u03b9 \u03c0\u03b9\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf: \u03c3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf abstaining policy, \u03b7 \u03b4\u03c5\u03bd\u03b1\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1 \u03bd\u03b1 \u03b5\u03c0\u03b9\u03b8\u03b5\u03c9\u03c1\u03ae\u03c3\u03bf\u03c5\u03bc\u03b5 \u03ad\u03bd\u03b1 reasoning trace \u03b1\u03b3\u03bf\u03c1\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c3\u03c4\u03b7\u03bd \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2. \u0395\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd, \u03b7 \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03c4\u03bf\u03c5 reasoning mode \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03bf operating point \u03c4\u03bf\u03c5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5, \u03ac\u03c1\u03b1 \u03b4\u03b5\u03bd \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03c3\u03b1\u03bd \u03bf\u03c5\u03b4\u03ad\u03c4\u03b5\u03c1\u03bf \u03c0\u03b1\u03c1\u03ac\u03b8\u03c5\u03c1\u03bf \u03c0\u03c1\u03bf\u03c2 \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae.<\/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\">Direct policy<\/p>\n<p>\u039f token-SFT classifier \u03c6\u03c4\u03ac\u03bd\u03b5\u03b9 macro-F1 0,620 \u03ba\u03b1\u03b9 \u03c3\u03c7\u03b7\u03bc\u03b1\u03c4\u03af\u03b6\u03b5\u03b9 \u03bd\u03c9\u03c1\u03af\u03c2 \u03c4\u03b7\u03bd \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 reasoning trace \u03b3\u03b9\u03b1 \u03b5\u03c0\u03b9\u03b8\u03b5\u03ce\u03c1\u03b7\u03c3\u03b7.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">\u039a\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03bf operating point<\/span><span class=\"td-badge\">\u03a7\u03c9\u03c1\u03af\u03c2 trace<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card td-platform-card--navy\">\n<p class=\"td-comparison-title\">Think-mode policy<\/p>\n<p>\u0397 base \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae \u03b4\u03af\u03bd\u03b5\u03b9 inspectable trace, \u03cc\u03bc\u03c9\u03c2 \u03c0\u03ad\u03c6\u03c4\u03b5\u03b9 \u03c3\u03b5 macro-F1 0,536 \u03ba\u03b1\u03b9 \u03c7\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf 57,9% \u03c4\u03c9\u03bd \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ce\u03bd \u03b5\u03c5\u03ba\u03b1\u03b9\u03c1\u03b9\u03ce\u03bd \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7\u03c2.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">\u039f\u03c1\u03b1\u03c4\u03ae \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae<\/span><span class=\"td-badge\">\u03a5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf MIR<\/span><\/div>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-comparison-title\">Controlled oversight<\/p>\n<p>\u0395\u03bb\u03ad\u03b3\u03c7\u03b5\u03b9 deployed mode, calibration, false \u03ba\u03b1\u03b9 missed interventions, probes \u03ba\u03b1\u03b9 content-controlled perturbations \u03c9\u03c2 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ac \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1.<\/p>\n<div class=\"td-badge-row\"><span class=\"td-badge\">\u03a0\u03bf\u03bb\u03bb\u03b1\u03c0\u03bb\u03bf\u03af \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03b9<\/span><span class=\"td-badge\">\u03a7\u03c9\u03c1\u03af\u03c2 proxy shortcuts<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<aside class=\"td-article-note\">\n<p><strong>\u039c\u03b7\u03bd \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03c4\u03b5 \u03c4\u03bf 0,620 \u03ae \u03c4\u03bf 0,536 \u03c9\u03c2 product SLA.<\/strong> \u039f\u03b9 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc 400 matched evaluation examples \u03c4\u03bf\u03c5 When2Speak, seed 7, \u03c3\u03b5 \u03c3\u03c5\u03bd\u03b8\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03c3\u03c5\u03bd\u03bf\u03bc\u03b9\u03bb\u03af\u03b5\u03c2. \u03a4\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03bc\u03bf\u03c4\u03af\u03b2\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ae \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae\u03c2, \u03cc\u03c7\u03b9 \u03b7 \u03b1\u03bd\u03c4\u03b9\u03b3\u03c1\u03b1\u03c6\u03ae \u03c4\u03bf\u03c5 score \u03c3\u03b5 \u03ac\u03bb\u03bb\u03bf customer-support \u03ae monitoring \u03c0\u03b5\u03c1\u03b9\u03b2\u03ac\u03bb\u03bb\u03bf\u03bd.<\/p>\n<\/aside>\n<h3>\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf supervised fine-tuning \u03b4\u03b5\u03bd \u03b3\u03b5\u03c6\u03c5\u03c1\u03ce\u03bd\u03b5\u03b9 \u03c4\u03bf \u03c7\u03ac\u03c3\u03bc\u03b1<\/h3>\n<p>\u0397 \u03ac\u03bc\u03b5\u03c3\u03b7 \u03b5\u03c0\u03bf\u03c0\u03c4\u03b5\u03c5\u03cc\u03bc\u03b5\u03bd\u03b7 \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03c4\u03b5\u03bb\u03b9\u03ba\u03bf\u03cd token \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03c4\u03b1\u03be\u03b9\u03bd\u03cc\u03bc\u03b7\u03c3\u03b7, \u03b1\u03bb\u03bb\u03ac \u03b1\u03c0\u03bf\u03b4\u03c5\u03bd\u03b1\u03bc\u03ce\u03bd\u03b5\u03b9 \u03c4\u03bf reasoning behavior. \u0397 \u03b5\u03bd\u03b1\u03bb\u03bb\u03b1\u03ba\u03c4\u03b9\u03ba\u03ae masked SFT \u03ba\u03c1\u03b1\u03c4\u03ac \u03c4\u03b9\u03c2 \u03b1\u03bb\u03c5\u03c3\u03af\u03b4\u03b5\u03c2 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae\u03c2 \u03c4\u03bf\u03c5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5, \u03b5\u03bd\u03ce \u03b5\u03c6\u03b1\u03c1\u03bc\u03cc\u03b6\u03b5\u03b9 loss \u03bc\u03cc\u03bd\u03bf \u03c3\u03c4\u03b7\u03bd \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7. \u03a4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c3\u03c5\u03bd\u03b5\u03c7\u03af\u03b6\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 reasoning \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 \u03c3\u03c4\u03bf 99% \u03c4\u03c9\u03bd \u03c0\u03b1\u03c1\u03b1\u03b4\u03b5\u03b9\u03b3\u03bc\u03ac\u03c4\u03c9\u03bd \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7\u03c2, \u03cc\u03bc\u03c9\u03c2 \u03b4\u03b5\u03bd \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03b4\u03cd\u03c3\u03ba\u03bf\u03bb\u03b5\u03c2 \u03c0\u03b5\u03c1\u03b9\u03c0\u03c4\u03ce\u03c3\u03b5\u03b9\u03c2 \u03c0\u03ad\u03c1\u03b1 \u03b1\u03c0\u03cc \u03c4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc reasoning policy.<\/p>\n<p>\u0397 \u03b1\u03b9\u03c4\u03af\u03b1 \u03b2\u03c1\u03af\u03c3\u03ba\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c4\u03b7\u03bd \u03ba\u03b1\u03c4\u03b1\u03c3\u03ba\u03b5\u03c5\u03ae \u03c4\u03c9\u03bd \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7\u03c2. \u0393\u03b9\u03b1 \u03bd\u03b1 \u03bc\u03b7\u03bd \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03c5\u03c4\u03b5\u03af \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 \u03b5\u03ba \u03c4\u03c9\u03bd \u03c5\u03c3\u03c4\u03ad\u03c1\u03c9\u03bd \u03b5\u03be\u03b7\u03b3\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c0\u03bf\u03c5 \u03ad\u03c7\u03bf\u03c5\u03bd \u03b3\u03c1\u03b1\u03c6\u03c4\u03b5\u03af \u03b3\u03bd\u03c9\u03c1\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03c2 \u03ae\u03b4\u03b7 \u03c4\u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7, \u03bf\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ad\u03c2 \u03ba\u03c1\u03b1\u03c4\u03bf\u03cd\u03bd \u03bc\u03cc\u03bd\u03bf reasoning traces \u03c0\u03bf\u03c5 \u03c4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c0\u03b1\u03c1\u03ae\u03b3\u03b1\u03b3\u03b5 \u03c6\u03c5\u03c3\u03b9\u03ba\u03ac \u03ba\u03b1\u03b9 \u03bf\u03b4\u03ae\u03b3\u03b7\u03c3\u03b1\u03bd \u03ae\u03b4\u03b7 \u03c3\u03c4\u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7. \u039c\u03b5 \u03b1\u03c5\u03c4\u03cc\u03bd \u03c4\u03bf\u03bd \u03b1\u03c3\u03c6\u03b1\u03bb\u03ae \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03cc, \u03b7 \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03b5\u03bd\u03b9\u03c3\u03c7\u03cd\u03b5\u03b9 \u03ba\u03c5\u03c1\u03af\u03c9\u03c2 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ad\u03c2 \u03c0\u03bf\u03c5 \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03ba\u03b1\u03c4\u03ad\u03c7\u03b5\u03b9. \u0394\u03b5\u03bd \u03c4\u03bf\u03c5 \u03b4\u03b9\u03b4\u03ac\u03c3\u03ba\u03b5\u03b9 \u03c0\u03ce\u03c2 \u03bd\u03b1 \u03b4\u03b9\u03bf\u03c1\u03b8\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03c0\u03b5\u03c1\u03b9\u03c0\u03c4\u03ce\u03c3\u03b5\u03b9\u03c2 \u03cc\u03c0\u03bf\u03c5 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b2\u03ad\u03b2\u03b1\u03b9\u03bf \u03b1\u03bb\u03bb\u03ac \u03bb\u03ac\u03b8\u03bf\u03c2.<\/p>\n<p>\u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03b5\u03b9 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 mismatch \u03c3\u03c4\u03bf prior. \u038c\u03c4\u03b1\u03bd \u03c4\u03bf \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03cc \u03c0\u03b1\u03c1\u03b5\u03bc\u03b2\u03ac\u03c3\u03b5\u03c9\u03bd \u03c3\u03c4\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1 \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c5\u03c8\u03b7\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03cc \u03c3\u03c4\u03bf \u03c0\u03b5\u03c1\u03b9\u03b2\u03ac\u03bb\u03bb\u03bf\u03bd \u03c7\u03c1\u03ae\u03c3\u03b7\u03c2, \u03b7 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03c3\u03c4\u03bf \u03c6\u03c5\u03c3\u03b9\u03ba\u03cc operating point \u03c6\u03b1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03ba\u03b1\u03c4\u03b1\u03c1\u03c1\u03ad\u03b5\u03b9. \u038c\u03c4\u03b1\u03bd \u03c4\u03b1 priors \u03c4\u03b1\u03b9\u03c1\u03b9\u03ac\u03b6\u03bf\u03c5\u03bd, \u03c4\u03bf \u03c0\u03c1\u03cc\u03b2\u03bb\u03b7\u03bc\u03b1 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03c4\u03b1\u03b9, \u03b5\u03bd\u03ce \u03b7 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae reasoning \u03ba\u03b1\u03b9 \u03b7 threshold-free \u03b4\u03b9\u03ac\u03ba\u03c1\u03b9\u03c3\u03b7 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03bf\u03c5\u03bd \u03c3\u03c7\u03b5\u03b4\u03cc\u03bd \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03ad\u03c2. \u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7 \u03b1\u03c5\u03c4\u03cc \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03b5\u03bd\u03cc\u03c2 KPI \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bf\u03c6\u03b5\u03af\u03bb\u03b5\u03c4\u03b1\u03b9 \u03c3\u03b5 \u03bb\u03ac\u03b8\u03bf\u03c2 calibration \u03ae threshold \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03c3\u03b5 \u03b1\u03c0\u03ce\u03bb\u03b5\u03b9\u03b1 \u03b2\u03b1\u03c3\u03b9\u03ba\u03ae\u03c2 \u03b9\u03ba\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2.<\/p>\n<div class=\"td-decision-band\">\n<p class=\"td-decision-band-kicker\">\u0395\u03c1\u03ce\u03c4\u03b7\u03c3\u03b7 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc deployment<\/p>\n<h3>\u0395\u03bb\u03ad\u03b3\u03c7\u03b5\u03c4\u03b5 \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae \u03c0\u03bf\u03c5 \u03b8\u03b1 \u03b5\u03bd\u03b5\u03c1\u03b3\u03b5\u03af \u03c3\u03c4\u03b7\u03bd \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae;<\/h3>\n<p>\u03a3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 think \u03ba\u03b1\u03b9 no-think mode, deployment-matched priors, deployed threshold, false-intervention \u03ba\u03b1\u03b9 missed-intervention costs, calibration, AUROC \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03bf override. \u0391\u03bd \u03c4\u03bf rationale \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03c4\u03b1\u03b9 \u03b5\u03ba \u03c4\u03c9\u03bd \u03c5\u03c3\u03c4\u03ad\u03c1\u03c9\u03bd, \u03bc\u03b7\u03bd \u03c4\u03bf \u03b2\u03b1\u03c6\u03c4\u03af\u03b6\u03b5\u03c4\u03b5 audit \u03c4\u03bf\u03c5 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03bf\u03cd action path.<\/p>\n<\/div>\n<h3>\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf group-relative reinforcement learning \u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03c3\u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03bf<\/h3>\n<p>\u039f\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ac\u03b6\u03bf\u03c5\u03bd GRPO \u03c3\u03b5 think mode \u03bc\u03b5 \u03b4\u03cd\u03bf rewards. \u03a4\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf \u03b5\u03c0\u03b9\u03b2\u03c1\u03b1\u03b2\u03b5\u03cd\u03b5\u03b9 \u03c4\u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7. \u03a4\u03bf \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf \u03c0\u03c1\u03bf\u03c3\u03b8\u03ad\u03c4\u03b5\u03b9 bonus \u03cc\u03c4\u03b1\u03bd \u03b7 \u03bf\u03bb\u03bf\u03ba\u03bb\u03ae\u03c1\u03c9\u03c3\u03b7 \u03c4\u03b7\u03c2 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae\u03c2 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03c3\u03b5 \u03c3\u03c7\u03ad\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03c0\u03c1\u03cc\u03c9\u03c1\u03b7 \u03b4\u03b9\u03b1\u03ba\u03bf\u03c0\u03ae \u03c4\u03b7\u03c2. \u0397 \u03c0\u03b1\u03c1\u03ac\u03bc\u03b5\u03c4\u03c1\u03bf\u03c2 \u03c4\u03bf\u03c5 bonus \u03b5\u03af\u03bd\u03b1\u03b9 0,3. \u03a0\u03b1\u03c1\u2019 \u03cc\u03bb\u03b1 \u03b1\u03c5\u03c4\u03ac, \u03bf\u03cd\u03c4\u03b5 \u03c4\u03bf accuracy reward \u03bf\u03cd\u03c4\u03b5 \u03c4\u03bf faithfulness-aware reward \u03be\u03b5\u03c0\u03b5\u03c1\u03bd\u03ac \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ac \u03c4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc reasoning policy.<\/p>\n<p>\u03a4\u03bf base policy \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 macro-F1 0,536, \u03c4\u03bf RL-accuracy 0,520 \u03ba\u03b1\u03b9 \u03c4\u03bf RL-faithfulness 0,549. \u039f\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ad\u03c2 \u03b2\u03c1\u03af\u03c3\u03ba\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03ad\u03c3\u03b1 \u03c3\u03c4\u03b7 \u03b4\u03b9\u03b1\u03ba\u03cd\u03bc\u03b1\u03bd\u03c3\u03b7 \u03bc\u03b5\u03c4\u03b1\u03be\u03cd \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03c9\u03bd runs, \u03b3\u03b9\u2019 \u03b1\u03c5\u03c4\u03cc \u03bf\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ad\u03c2 \u03b4\u03b5\u03bd \u03c4\u03b9\u03c2 \u03b1\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03c9\u03c0\u03af\u03b6\u03bf\u03c5\u03bd \u03c9\u03c2 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03ba\u03b1\u03c4\u03ac\u03c4\u03b1\u03be\u03b7. \u0391\u03ba\u03cc\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03ac \u03b1\u03c0\u03cc \u03b5\u03ba\u03c4\u03b9\u03bc\u03ce\u03bc\u03b5\u03bd\u03b7 \u03b1\u03cd\u03be\u03b7\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03c3\u03bc\u03b1\u03c4\u03b9\u03ba\u03bf\u03cd gradient \u03ba\u03b1\u03c4\u03ac 5 \u03ad\u03c9\u03c2 10 \u03c6\u03bf\u03c1\u03ad\u03c2, \u03b7 held-out \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03bf\u03c0\u03bf\u03b9\u03b5\u03af\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03bf.<\/p>\n<p>\u039f \u03bc\u03b7\u03c7\u03b1\u03bd\u03b9\u03c3\u03bc\u03cc\u03c2 \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03af\u03b1\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03cc\u03c2 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af group-relative objectives. \u0391\u03bd \u03cc\u03bb\u03b1 \u03c4\u03b1 rollouts \u03bc\u03b9\u03b1\u03c2 \u03bf\u03bc\u03ac\u03b4\u03b1\u03c2 \u03b5\u03c0\u03b9\u03bb\u03ad\u03b3\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1, \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b7 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1, \u03bf\u03b9 \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b1\u03bd\u03c4\u03b1\u03bc\u03bf\u03b9\u03b2\u03ad\u03c2 \u03b4\u03b5\u03bd \u03ad\u03c7\u03bf\u03c5\u03bd \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03bf\u03c0\u03bf\u03af\u03b1 \u03bd\u03b1 \u03c0\u03c1\u03bf\u03ba\u03cd\u03c8\u03b5\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf learning signal. \u03a4\u03bf policy \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03bc\u03b5 \u03b1\u03c5\u03c4\u03bf\u03c0\u03b5\u03c0\u03bf\u03af\u03b8\u03b7\u03c3\u03b7 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ce\u03c2 \u03c3\u03c4\u03b1 prompts \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03b4\u03b9\u03cc\u03c1\u03b8\u03c9\u03c3\u03b7.<\/p>\n<h2 id=\"pote-schimatizetai-apofasi\">\u03a0\u03cc\u03c4\u03b5 \u03c3\u03c7\u03b7\u03bc\u03b1\u03c4\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03b7 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7<\/h2>\n<p>\u0393\u03b9\u03b1 \u03bd\u03b1 \u03b5\u03be\u03b5\u03c4\u03ac\u03c3\u03bf\u03c5\u03bd \u03b1\u03bd \u03c4\u03bf reasoning \u03b1\u03c0\u03bb\u03ce\u03c2 \u03bd\u03c4\u03cd\u03bd\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03c0\u03c1\u03bf\u03cb\u03c0\u03ac\u03c1\u03c7\u03bf\u03c5\u03c3\u03b1 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae, \u03bf\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ad\u03c2 \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03bf\u03c5\u03bd \u03b3\u03c1\u03b1\u03bc\u03bc\u03b9\u03ba\u03ac probes \u03c0\u03ac\u03bd\u03c9 \u03c3\u03b5 hidden activations \u03c0\u03c1\u03b9\u03bd \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03ac \u03c4\u03b7\u03bd \u03b1\u03bb\u03c5\u03c3\u03af\u03b4\u03b1 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae\u03c2. \u03a3\u03c4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc reasoning policy, \u03b7 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03bf\u03bb\u03cd \u03c0\u03b9\u03bf \u03b1\u03c0\u03bf\u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b9\u03bc\u03b7 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c4\u03bf chain of thought \u03b1\u03c0\u03cc \u03cc\u03c3\u03bf \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c0\u03b5\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03b1\u03c0\u03cc \u03c4\u03bf \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03b5\u03b9\u03c3\u03cc\u03b4\u03bf\u03c5: AUROC 0,664 \u00b1 0,034 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 0,631 \u00b1 0,053 \u03b3\u03b9\u03b1 text-only baseline. \u039c\u03b5\u03c4\u03ac \u03c4\u03bf reasoning, \u03b7 \u03b1\u03c0\u03bf\u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03b7\u03c3\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03bd\u03b5\u03b2\u03b1\u03af\u03bd\u03b5\u03b9 \u03c3\u03b5 0,976 \u00b1 0,045.<\/p>\n<p>\u03a3\u03c4\u03bf\u03bd \u03ac\u03bc\u03b5\u03c3\u03bf classifier, \u03b1\u03bd\u03c4\u03af\u03b8\u03b5\u03c4\u03b1, \u03b7 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03ae\u03b4\u03b7 \u03c3\u03c7\u03b5\u03b4\u03cc\u03bd \u03c0\u03bb\u03ae\u03c1\u03c9\u03c2 \u03b1\u03c0\u03bf\u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b9\u03bc\u03b7 \u03c0\u03c1\u03b9\u03bd \u03c0\u03b1\u03c1\u03b1\u03c7\u03b8\u03b5\u03af reasoning, \u03bc\u03b5 AUROC 0,997. \u0391\u03c5\u03c4\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03b1\u03bd\u03b1\u03bc\u03b5\u03bd\u03cc\u03bc\u03b5\u03bd\u03b7 \u03c5\u03c0\u03bf\u03b3\u03c1\u03b1\u03c6\u03ae \u03bc\u03b9\u03b1\u03c2 \u03ac\u03bc\u03b5\u03c3\u03b7\u03c2 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae\u03c2: \u03b7 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03c3\u03c7\u03b7\u03bc\u03b1\u03c4\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bd\u03c9\u03c1\u03af\u03c2 \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03ac \u03b1\u03c0\u03bb\u03ce\u03c2 \u03b5\u03ba\u03c0\u03ad\u03bc\u03c0\u03b5\u03c4\u03b1\u03b9.<\/p>\n<p>\u03a4\u03b1 probes \u03b3\u03b9\u03b1 \u03c4\u03bf \u03b5\u03af\u03b4\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7\u03c2 \u03b4\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03c1\u03bf\u03c6\u03b7 \u03c7\u03c1\u03bf\u03bd\u03b9\u03ba\u03ae \u03c0\u03bf\u03c1\u03b5\u03af\u03b1. \u0397 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03b3\u03b9\u03b1 \u03c4\u03bf \u03b5\u03af\u03b4\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03b3\u03bd\u03c9\u03c3\u03c4\u03b9\u03ba\u03bf\u03cd \u03ba\u03b5\u03bd\u03bf\u03cd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03b7 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc \u03c4\u03bf reasoning \u03ba\u03b1\u03b9 \u03b5\u03be\u03b1\u03c3\u03b8\u03b5\u03bd\u03b5\u03af \u03ba\u03b1\u03b8\u03ce\u03c2 \u03c3\u03c7\u03b7\u03bc\u03b1\u03c4\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b7 \u03b4\u03c5\u03b1\u03b4\u03b9\u03ba\u03ae \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7. \u0397 \u03c0\u03c1\u03bf\u03c3\u03b5\u03ba\u03c4\u03b9\u03ba\u03ae \u03b5\u03c1\u03bc\u03b7\u03bd\u03b5\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03b1\u03bd\u03b1\u03b3\u03bd\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03c0\u03c1\u03ce\u03c4\u03b1 \u03c4\u03b9 \u03b5\u03af\u03b4\u03bf\u03c5\u03c2 \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03b1\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03c9\u03c0\u03af\u03b6\u03b5\u03b9 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b7 \u03c3\u03c5\u03bd\u03ad\u03c7\u03b5\u03b9\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03b1\u03c5\u03c4\u03ae \u03c4\u03b7\u03bd \u03ba\u03b1\u03c4\u03b1\u03bd\u03cc\u03b7\u03c3\u03b7 \u03c3\u03b5 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03b4\u03c1\u03ac\u03c3\u03b7\u03c2. \u0397 \u03b1\u03c0\u03bf\u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03b7\u03c3\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1, \u03c0\u03ac\u03bd\u03c4\u03c9\u03c2, \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03b7 \u03c4\u03b7\u03c2 \u03b1\u03b9\u03c4\u03b9\u03ce\u03b4\u03b7 \u03c7\u03c1\u03ae\u03c3\u03b7.<\/p>\n<h2 id=\"afairesi-chain-of-thought\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03b1\u03c6\u03b1\u03af\u03c1\u03b5\u03c3\u03b7 \u03c4\u03bf\u03c5 chain of thought \u03c0\u03b1\u03c1\u03b1\u03c0\u03bb\u03b1\u03bd\u03ac<\/h2>\n<p>\u0397 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03b9 \u03c0\u03bb\u03ae\u03c1\u03b7 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae, \u03ba\u03bf\u03bc\u03bc\u03ad\u03bd\u03b7 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03ba\u03b1\u03b9 \u03bf\u03c5\u03b4\u03ad\u03c4\u03b5\u03c1\u03bf filler \u03af\u03b4\u03b9\u03bf\u03c5 \u03bc\u03ae\u03ba\u03bf\u03c5\u03c2. \u0397 \u03b1\u03c0\u03bb\u03ae \u03b1\u03c6\u03b1\u03af\u03c1\u03b5\u03c3\u03b7 \u03c4\u03bf\u03c5 reasoning \u03ba\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03bd\u03b1 \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03ad\u03c6\u03b5\u03b9 \u03c3\u03c4\u03b7\u03bd \u03b5\u03ba\u03c6\u03c5\u03bb\u03b9\u03c3\u03bc\u03ad\u03bd\u03b7 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ac \u00ab\u03c3\u03c7\u03b5\u03b4\u03cc\u03bd \u03c0\u03ac\u03bd\u03c4\u03b1 \u03c0\u03b1\u03c1\u03b5\u03bc\u03b2\u03b1\u03af\u03bd\u03c9\u00bb. \u039c\u03b9\u03b1 \u03b1\u03c6\u03b5\u03bb\u03ae\u03c2 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 \u03b8\u03b1 \u03b8\u03b5\u03c9\u03c1\u03bf\u03cd\u03c3\u03b5 \u03c4\u03b9\u03c2 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ad\u03c2 \u03b1\u03c0\u03bf\u03c6\u03ac\u03c3\u03b5\u03c9\u03bd \u03b1\u03c0\u03cc\u03b4\u03b5\u03b9\u03be\u03b7 \u03cc\u03c4\u03b9 \u03c4\u03bf \u03c0\u03b5\u03c1\u03b9\u03b5\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf \u03c4\u03b7\u03c2 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03bb\u03ae\u03c1\u03c9\u03c2 \u03b1\u03b9\u03c4\u03b9\u03ce\u03b4\u03b5\u03c2.<\/p>\n<p>\u038c\u03bc\u03c9\u03c2 \u03c4\u03bf \u03bf\u03c5\u03b4\u03ad\u03c4\u03b5\u03c1\u03bf filler \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03b1\u03ba\u03cc\u03bc\u03b7 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03b1\u03c0\u03bf\u03c6\u03ac\u03c3\u03b5\u03b9\u03c2. \u0386\u03c1\u03b1 \u03ad\u03bd\u03b1 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03b5\u03c0\u03af\u03b4\u03c1\u03b1\u03c3\u03b7\u03c2 \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03c4\u03bf\u03c5 inference mode \u03ae \u03c4\u03b7\u03c2 \u03bc\u03bf\u03c1\u03c6\u03ae\u03c2, \u03cc\u03c7\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf \u03bd\u03cc\u03b7\u03bc\u03b1 \u03c4\u03c9\u03bd reasoning tokens. \u038c\u03c4\u03b1\u03bd \u03bf\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ad\u03c2 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03ba\u03bf\u03bc\u03bc\u03ad\u03bd\u03b7 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03bc\u03b5 filler \u03af\u03b4\u03b9\u03bf\u03c5 \u03bc\u03ae\u03ba\u03bf\u03c5\u03c2, \u03c4\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf \u03bc\u03b9\u03c3\u03cc \u03c4\u03bf\u03c5 \u03b1\u03c5\u03b8\u03b5\u03bd\u03c4\u03b9\u03ba\u03bf\u03cd reasoning \u03c4\u03c1\u03b1\u03b2\u03ac \u03c4\u03b9\u03c2 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c8\u03b5\u03b9\u03c2 \u03c0\u03c1\u03bf\u03c2 \u03c4\u03bf \u03c0\u03bb\u03ae\u03c1\u03b5\u03c2 policy. \u03a4\u03bf \u03c0\u03b5\u03c1\u03b9\u03b5\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf \u03ad\u03c7\u03b5\u03b9 \u03bb\u03bf\u03b9\u03c0\u03cc\u03bd \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03b1\u03bb\u03bb\u03ac \u03bc\u03ad\u03c4\u03c1\u03b9\u03b1 \u03b1\u03b9\u03c4\u03b9\u03ce\u03b4\u03b7 \u03c3\u03c5\u03bc\u03b2\u03bf\u03bb\u03ae.<\/p>\n<p>\u03a4\u03bf \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03b1\u03c0\u03bf\u03c6\u03b5\u03cd\u03b3\u03b5\u03b9 \u03ba\u03b1\u03b9 \u03c4\u03b1 \u03b4\u03cd\u03bf \u03ac\u03ba\u03c1\u03b1. \u0397 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bb\u03ae \u03b4\u03b9\u03b1\u03ba\u03cc\u03c3\u03bc\u03b7\u03c3\u03b7, \u03b1\u03bb\u03bb\u03ac \u03bf\u03cd\u03c4\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03bc\u03b5\u03c4\u03ac \u03b1\u03c0\u03cc ablation \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03b5\u03be\u03ae\u03b3\u03b7\u03c3\u03b7 \u03ae\u03c4\u03b1\u03bd \u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc\u03c2 \u03bc\u03b7\u03c7\u03b1\u03bd\u03b9\u03c3\u03bc\u03cc\u03c2. \u03a7\u03c9\u03c1\u03af\u03c2 content control, \u03c4\u03b1 flip tests \u03c5\u03c0\u03b5\u03c1\u03b5\u03ba\u03c4\u03b9\u03bc\u03bf\u03cd\u03bd \u03c4\u03b7\u03bd \u03c0\u03b9\u03c3\u03c4\u03cc\u03c4\u03b7\u03c4\u03b1.<\/p>\n<aside class=\"td-article-note\">\n<p><strong>\u03a4\u03bf chain of thought \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 production audit log.<\/strong> \u03a3\u03c4\u03bf Why2Speak \u03c4\u03bf neutral filler \u03ac\u03bb\u03bb\u03b1\u03be\u03b5 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03b1\u03c0\u03bf\u03c6\u03ac\u03c3\u03b5\u03b9\u03c2 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b1\u03c0\u03bb\u03ae truncation, \u03ac\u03c1\u03b1 \u03ad\u03bd\u03b1 raw flip rate \u03bc\u03c0\u03ad\u03c1\u03b4\u03b5\u03c8\u03b5 \u03c4\u03bf \u03c0\u03b5\u03c1\u03b9\u03b5\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf \u03c4\u03b7\u03c2 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae\u03c2 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03c4\u03bf\u03c5 inference format. \u039a\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b1 action logs, inputs, thresholds \u03ba\u03b1\u03b9 outcomes.<\/p>\n<\/aside>\n<h2 id=\"sostes-exigiseis-pistotita\">\u039f\u03b9 \u00ab\u03c3\u03c9\u03c3\u03c4\u03ad\u03c2\u00bb \u03b5\u03be\u03b7\u03b3\u03ae\u03c3\u03b5\u03b9\u03c2 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03b1\u03c4\u2019 \u03b1\u03bd\u03ac\u03b3\u03ba\u03b7 \u03c0\u03b9\u03c3\u03c4\u03ad\u03c2<\/h2>\n<p>\u039c\u03b5\u03c4\u03ac \u03c4\u03bf reinforcement learning, \u03bf\u03b9 \u03bb\u03b5\u03ba\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b5\u03be\u03b7\u03b3\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c3\u03c5\u03bc\u03c6\u03c9\u03bd\u03bf\u03cd\u03bd \u03c3\u03c5\u03c7\u03bd\u03cc\u03c4\u03b5\u03c1\u03b1 \u03bc\u03b5 \u03c4\u03b7 \u03c3\u03c9\u03c3\u03c4\u03ae \u03ba\u03b1\u03c4\u03b7\u03b3\u03bf\u03c1\u03af\u03b1 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7\u03c2: \u03b7 \u03c3\u03c5\u03bc\u03c6\u03c9\u03bd\u03af\u03b1 \u03b1\u03c5\u03be\u03ac\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc 0,378 \u03c3\u03c4\u03bf base policy \u03c3\u03b5 0,425 \u03bc\u03b5 accuracy objective \u03ba\u03b1\u03b9 0,474 \u03bc\u03b5 faithfulness objective. \u0391\u03c5\u03c4\u03cc \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03c0\u03b5\u03c1\u03b9\u03b3\u03c1\u03b1\u03c6\u03ae \u03c4\u03b7\u03c2 \u03ba\u03b1\u03c4\u03b7\u03b3\u03bf\u03c1\u03af\u03b1\u03c2, \u03cc\u03c7\u03b9 \u03c5\u03c0\u03bf\u03c7\u03c1\u03b5\u03c9\u03c4\u03b9\u03ba\u03ac \u03c0\u03b9\u03c3\u03c4\u03cc\u03c4\u03b5\u03c1\u03b7 \u03c0\u03c1\u03cc\u03c3\u03b2\u03b1\u03c3\u03b7 \u03c3\u03c4\u03bf\u03bd \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03cc \u03bc\u03b7\u03c7\u03b1\u03bd\u03b9\u03c3\u03bc\u03cc.<\/p>\n<p>\u03a3\u03c4\u03bf \u03c4\u03ad\u03bb\u03bf\u03c2 \u03c4\u03bf\u03c5 chain of thought, \u03c4\u03b1 internal representations \u03b5\u03c5\u03b8\u03c5\u03b3\u03c1\u03b1\u03bc\u03bc\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03bc\u03b5 \u03c4\u03b7 \u03b4\u03b7\u03bb\u03c9\u03bc\u03ad\u03bd\u03b7 \u03b5\u03be\u03ae\u03b3\u03b7\u03c3\u03b7. \u0395\u03ba\u03b5\u03af, \u03cc\u03bc\u03c9\u03c2, \u03c4\u03bf residual stream \u03c0\u03b5\u03c1\u03b9\u03ad\u03c7\u03b5\u03b9 \u03ae\u03b4\u03b7 \u03c4\u03bf \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03c0\u03bf\u03c5 \u03bc\u03cc\u03bb\u03b9\u03c2 \u03c0\u03b1\u03c1\u03ae\u03c7\u03b8\u03b7. \u0388\u03bd\u03b1 probe \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03b1\u03c0\u03bb\u03ce\u03c2 \u03bd\u03b1 \u03be\u03b1\u03bd\u03b1\u03b4\u03b9\u03b1\u03b2\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7 \u03b4\u03b9\u03b1\u03c4\u03cd\u03c0\u03c9\u03c3\u03b7 \u03b1\u03bd\u03c4\u03af \u03bd\u03b1 \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b7 \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ae \u03b1\u03b9\u03c4\u03af\u03b1. \u03a0\u03c1\u03b9\u03bd \u03b1\u03c1\u03c7\u03af\u03c3\u03b5\u03b9 \u03c4\u03bf reasoning, \u03b1\u03c5\u03c4\u03ae \u03b7 \u03b5\u03c5\u03b8\u03c5\u03b3\u03c1\u03ac\u03bc\u03bc\u03b9\u03c3\u03b7 \u03c3\u03c7\u03b5\u03b4\u03cc\u03bd \u03b5\u03be\u03b1\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03c4\u03b1\u03b9.<\/p>\n<p>\u0393\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b5\u03c2 AI governance, \u03b1\u03c5\u03c4\u03cc \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf: \u03b7 \u03b1\u03cd\u03be\u03b7\u03c3\u03b7 \u03c4\u03b7\u03c2 \u03bb\u03b5\u03ba\u03c4\u03b9\u03ba\u03ae\u03c2 \u03c3\u03c5\u03bc\u03c6\u03c9\u03bd\u03af\u03b1\u03c2 \u03bc\u03b5 \u03ad\u03bd\u03b1 gold label \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03bf\u03c5\u03c3\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c9\u03c2 \u03b1\u03cd\u03be\u03b7\u03c3\u03b7 \u03c4\u03b7\u03c2 reasoning faithfulness. \u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b4\u03b9\u03ac\u03ba\u03c1\u03b9\u03c3\u03b7 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 explanation quality, internal representation \u03ba\u03b1\u03b9 causal influence.<\/p>\n<h2 id=\"eksi-elegchoi-axiologisis\">\u0388\u03be\u03b9 \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03b9 \u03b3\u03b9\u03b1 \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03b7 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7<\/h2>\n<p>\u03a4\u03bf Why2Speak \u03b4\u03b5\u03bd \u03b4\u03af\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03bc\u03bf\u03bd\u03b1\u03b4\u03b9\u03ba\u03cc \u00abfaithfulness score\u00bb. \u03a0\u03c1\u03bf\u03c4\u03b5\u03af\u03bd\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03c3\u03cd\u03bd\u03bf\u03bb\u03bf \u03b4\u03b9\u03ba\u03bb\u03af\u03b4\u03c9\u03bd \u03c0\u03bf\u03c5 \u03b1\u03c0\u03bf\u03c6\u03b5\u03cd\u03b3\u03bf\u03c5\u03bd \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c8\u03bf\u03c5\u03bd \u03c4\u03bf reasoning trace \u03c3\u03b5 \u03b5\u03cd\u03ba\u03bf\u03bb\u03bf \u03b1\u03bb\u03bb\u03ac \u03c0\u03b1\u03c1\u03b1\u03c0\u03bb\u03b1\u03bd\u03b7\u03c4\u03b9\u03ba\u03cc proxy.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">\u0391\u03c0\u03cc \u03c4\u03bf \u03c0\u03b5\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc rationale \u03c3\u03b5 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf action policy<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">Test 1<\/span><strong>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 \u03c4\u03bf \u03c6\u03c5\u03c3\u03b9\u03ba\u03cc operating point<\/strong>\n<p>\u039a\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c8\u03c4\u03b5 \u03c0\u03cc\u03c3\u03bf \u03c3\u03c5\u03c7\u03bd\u03ac \u03bf agent \u03bc\u03b9\u03bb\u03ac \u03ae \u03b1\u03c0\u03ad\u03c7\u03b5\u03b9 \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac false-intervention \u03ba\u03b1\u03b9 missed-intervention rates. \u0397 \u03b1\u03c0\u03bb\u03ae accuracy \u03ba\u03c1\u03cd\u03b2\u03b5\u03b9 \u03c4\u03b7\u03bd \u03ba\u03bb\u03ac\u03c3\u03b7 \u03c4\u03bf\u03c5 13%.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Check 2<\/span><strong>\u03a7\u03c9\u03c1\u03af\u03c3\u03c4\u03b5 calibration \u03b1\u03c0\u03cc discrimination<\/strong>\n<p>\u03a3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 macro-F1 \u03c3\u03c4\u03bf deployed threshold \u03bc\u03b5 AUROC. \u0391\u03bd \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf operating point, \u03b4\u03b9\u03bf\u03c1\u03b8\u03ce\u03c3\u03c4\u03b5 priors \u03ba\u03b1\u03b9 threshold \u03c0\u03c1\u03b9\u03bd \u03bc\u03b9\u03bb\u03ae\u03c3\u03b5\u03c4\u03b5 \u03b3\u03b9\u03b1 \u03bd\u03ad\u03b1 \u03b9\u03ba\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Check 3<\/span><strong>\u0392\u03ac\u03bb\u03c4\u03b5 controls \u03c3\u03c4\u03b1 activation probes<\/strong>\n<p>\u03a7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03c4\u03b5 balanced metrics, text-only baselines, shuffled labels, within-family evaluation \u03ba\u03b1\u03b9 nested layer selection. \u0397 decodability \u03b4\u03b5\u03bd \u03b9\u03c3\u03bf\u03b4\u03c5\u03bd\u03b1\u03bc\u03b5\u03af \u03bc\u03b5 \u03b1\u03b9\u03c4\u03b9\u03ce\u03b4\u03b7 \u03c7\u03c1\u03ae\u03c3\u03b7.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Check 4<\/span><strong>\u0395\u03bb\u03ad\u03b3\u03be\u03c4\u03b5 \u03c4\u03b7 \u03b8\u03ad\u03c3\u03b7 \u03c4\u03bf\u03c5 probe<\/strong>\n<p>\u0388\u03bd\u03b1 end-of-reasoning probe \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03b1\u03c0\u03bb\u03ce\u03c2 \u03bd\u03b1 \u03be\u03b1\u03bd\u03b1\u03b4\u03b9\u03b1\u03b2\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b5\u03be\u03ae\u03b3\u03b7\u03c3\u03b7 \u03c0\u03bf\u03c5 \u03bc\u03cc\u03bb\u03b9\u03c2 \u03b3\u03c1\u03ac\u03c6\u03c4\u03b7\u03ba\u03b5. \u03a3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 pre-CoT \u03ba\u03b1\u03b9 post-CoT states \u03ba\u03b1\u03b9 \u03b4\u03b7\u03bb\u03ce\u03c3\u03c4\u03b5 \u03c4\u03b9 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ce\u03c2 \u03c5\u03c0\u03bf\u03c3\u03c4\u03b7\u03c1\u03af\u03b6\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Test 5<\/span><strong>\u039a\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03cc \u03c4\u03bf inference format<\/strong>\n<p>\u03a3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 truncated reasoning \u03bc\u03b5 neutral filler \u03af\u03b4\u03b9\u03bf\u03c5 \u03bc\u03ae\u03ba\u03bf\u03c5\u03c2. \u0388\u03c4\u03c3\u03b9 \u03b4\u03b9\u03b1\u03c7\u03c9\u03c1\u03af\u03b6\u03b5\u03c4\u03b5 \u03c4\u03b7\u03bd \u03b5\u03c0\u03af\u03b4\u03c1\u03b1\u03c3\u03b7 \u03c4\u03bf\u03c5 reasoning content \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b1\u03c0\u03bb\u03ae \u03bc\u03b5\u03c4\u03ac\u03b2\u03b1\u03c3\u03b7 \u03c0\u03c1\u03bf\u03c2 \u03ac\u03bb\u03bb\u03bf decoding mode.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Test 6<\/span><strong>\u0395\u03c0\u03b1\u03bd\u03b1\u03bb\u03ac\u03b2\u03b5\u03c4\u03b5 \u03c4\u03b1 RL runs<\/strong>\n<p>\u0391\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03c9\u03c0\u03af\u03c3\u03c4\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03bf optimization run \u03c9\u03c2 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b1 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03ac\u03c3\u03bc\u03b1\u03c4\u03bf\u03c2. \u03a3\u03c4\u03b7 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03b7 \u03b4\u03b9\u03b1\u03ba\u03cd\u03bc\u03b1\u03bd\u03c3\u03b7 \u03bc\u03b5\u03c4\u03b1\u03be\u03cd seeds \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03b9\u03c2 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ad\u03c2 \u03c4\u03c9\u03bd reward objectives.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<h2 id=\"business-ai-agents-roes\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1 AI agents \u03c3\u03b5 \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ad\u03c2 \u03c1\u03bf\u03ad\u03c2<\/h2>\n<p>\u0397 \u03c0\u03c1\u03bf\u03c3\u03b5\u03ba\u03c4\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03ae \u03b5\u03c1\u03bc\u03b7\u03bd\u03b5\u03af\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 agent \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 chain of thought. \u03a4\u03bf paper \u03b4\u03b5\u03bd \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b5\u03bc\u03c0\u03bf\u03c1\u03b9\u03ba\u03ad\u03c2 \u03b5\u03b3\u03ba\u03b1\u03c4\u03b1\u03c3\u03c4\u03ac\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c3\u03c5\u03bd\u03b8\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03c3\u03c5\u03bd\u03bf\u03bc\u03b9\u03bb\u03af\u03b5\u03c2. \u0394\u03b5\u03af\u03c7\u03bd\u03b5\u03b9, \u03cc\u03bc\u03c9\u03c2, \u03b3\u03b9\u03b1\u03c4\u03af \u03b7 <a href=\"https:\/\/twodots.gr\/otan-to-ai-peithei-ton-elegkti-tou-chain-of-thought-monitoring\/\">\u03b5\u03bd\u03c4\u03bf\u03bb\u03ae \u00ab\u03b5\u03be\u03ae\u03b3\u03b7\u03c3\u03b5 \u03c0\u03c1\u03b9\u03bd \u03b5\u03bd\u03b5\u03c1\u03b3\u03ae\u03c3\u03b5\u03b9\u03c2\u00bb \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03b5\u03af \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03b7 \u03c4\u03b7\u03c2 \u03bc\u03b7\u03c7\u03b1\u03bd\u03b9\u03c3\u03bc\u03cc \u03b5\u03bb\u03ad\u03b3\u03c7\u03bf\u03c5<\/a>. \u039c\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc policy, \u03bc\u03b5 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03c3\u03c5\u03c7\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b4\u03c1\u03ac\u03c3\u03b7\u03c2 \u03ba\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc \u03c0\u03c1\u03bf\u03c6\u03af\u03bb \u03bb\u03b1\u03b8\u03ce\u03bd.<\/p>\n<p>\u03a3\u03b5 <a href=\"https:\/\/twodots.gr\/6-ai-agents-customer-support-gia-epicheiriseis\/\">customer support \u03bc\u03b5 AI agents<\/a>, social listening, moderation, meeting assistance \u03ae \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03bf\u03cd\u03c2 copilots, \u03b7 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03b8\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c1\u03b7\u03c4\u03ac \u03c4\u03bf \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2 \u03c0\u03b5\u03c1\u03b9\u03c4\u03c4\u03ae\u03c2 \u03ba\u03b1\u03b9 \u03c7\u03b1\u03bc\u03ad\u03bd\u03b7\u03c2 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7\u03c2. \u03a4\u03bf macro-F1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03c3\u03b5 \u03b1\u03bd\u03b9\u03c3\u03cc\u03c1\u03c1\u03bf\u03c0\u03b1 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03b1, \u03b1\u03bb\u03bb\u03ac \u03b4\u03b5\u03bd \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03b8\u03b9\u03c3\u03c4\u03ac \u03c4\u03b7\u03bd \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ae \u03c0\u03b1\u03c1\u03b1\u03ba\u03bf\u03bb\u03bf\u03cd\u03b8\u03b7\u03c3\u03b7 false-intervention \u03ba\u03b1\u03b9 missed-intervention rates. \u03a4\u03bf \u03ba\u03b1\u03c4\u03ac\u03bb\u03bb\u03b7\u03bb\u03bf operating point \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf \u03c1\u03af\u03c3\u03ba\u03bf \u03c4\u03b7\u03c2 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7\u03c2 \u03c1\u03bf\u03ae\u03c2 \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03cc benchmark.<\/p>\n<p>\u0395\u03c0\u03af\u03c3\u03b7\u03c2, \u03ad\u03bd\u03b1 <a href=\"https:\/\/twodots.gr\/enterprise-ai-harnesses-diakyvernisi-architektoniki\/\">enterprise AI harness<\/a> \u03ba\u03b1\u03b9 \u03c4\u03bf audit \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03bf\u03c5\u03bd \u03c4\u03bf deployed mode, \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03bc\u03b9\u03b1 \u03b5\u03c1\u03b3\u03b1\u03c3\u03c4\u03b7\u03c1\u03b9\u03b1\u03ba\u03ae \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 \u03c0\u03bf\u03c5 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03b5\u03be\u03b7\u03b3\u03ae\u03c3\u03b5\u03b9\u03c2. \u0391\u03bd \u03c4\u03bf production \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03b1\u03c0\u03bf\u03c6\u03b1\u03c3\u03af\u03b6\u03b5\u03b9 \u03ac\u03bc\u03b5\u03c3\u03b1 \u03b1\u03bb\u03bb\u03ac \u03c4\u03bf audit \u03b6\u03b7\u03c4\u03ac \u03b5\u03ba \u03c4\u03c9\u03bd \u03c5\u03c3\u03c4\u03ad\u03c1\u03c9\u03bd rationale, \u03c4\u03cc\u03c4\u03b5 \u03b5\u03bb\u03ad\u03b3\u03c7\u03b5\u03c4\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03b4\u03b9\u03b1\u03b4\u03b9\u03ba\u03b1\u03c3\u03af\u03b1. \u0391\u03bd \u03c4\u03bf production \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03c5\u03c0\u03bf\u03c7\u03c1\u03b5\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 reason \u03c0\u03c1\u03b9\u03bd \u03b5\u03bd\u03b5\u03c1\u03b3\u03ae\u03c3\u03b5\u03b9, \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03c1\u03b7\u03b8\u03b5\u03af \u03b5\u03ba \u03bd\u03ad\u03bf\u03c5 \u03b7 \u03af\u03b4\u03b9\u03b1 \u03b7 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7, \u03b3\u03b9\u03b1\u03c4\u03af \u03c4\u03bf reasoning \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03ba\u03b9\u03bd\u03b5\u03af \u03c4\u03b7\u03bd \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae.<\/p>\n<h2 id=\"plaisio-ai-oversight\">\u0388\u03bd\u03b1 \u03c0\u03b9\u03bf \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03bf \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf AI oversight<\/h2>\n<p>\u039c\u03b5 \u03b2\u03ac\u03c3\u03b7 \u03c4\u03b1 \u03b5\u03c5\u03c1\u03ae\u03bc\u03b1\u03c4\u03b1 \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2, \u03bc\u03b9\u03b1 <a href=\"https:\/\/twodots.gr\/ai-agent-demo-paragogi-pente-teixi-epicheirisis\/\">\u03c5\u03c0\u03b5\u03cd\u03b8\u03c5\u03bd\u03b7 \u03bc\u03b5\u03c4\u03ac\u03b2\u03b1\u03c3\u03b7 \u03c4\u03bf\u03c5 AI agent \u03c3\u03c4\u03b7\u03bd \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae<\/a> \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bf\u03c1\u03b3\u03b1\u03bd\u03ce\u03c3\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03c3\u03b5 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b1 \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03b1. \u03a0\u03c1\u03ce\u03c4\u03b1 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03c4\u03b1\u03b9 \u03b7 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ac \u03c3\u03c4\u03bf \u03c6\u03c5\u03c3\u03b9\u03ba\u03cc operating point: \u03c0\u03cc\u03c3\u03bf \u03c3\u03c5\u03c7\u03bd\u03ac \u03bf agent \u03b4\u03c1\u03b1, \u03b1\u03c0\u03ad\u03c7\u03b5\u03b9, \u03c0\u03b1\u03c1\u03b5\u03bc\u03b2\u03b1\u03af\u03bd\u03b5\u03b9 \u03ac\u03c3\u03ba\u03bf\u03c0\u03b1 \u03ba\u03b1\u03b9 \u03c7\u03ac\u03bd\u03b5\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03b5\u03c2 \u03b5\u03c5\u03ba\u03b1\u03b9\u03c1\u03af\u03b5\u03c2. \u0388\u03c0\u03b5\u03b9\u03c4\u03b1 \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b7 threshold-free \u03b4\u03b9\u03ac\u03ba\u03c1\u03b9\u03c3\u03b7 \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03c6\u03b1\u03bd\u03b5\u03af \u03b1\u03bd \u03bc\u03b9\u03b1 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae \u03b5\u03af\u03bd\u03b1\u03b9 \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ae \u03bc\u03ac\u03b8\u03b7\u03c3\u03b7 \u03ae \u03b1\u03c0\u03bb\u03ae \u03bc\u03b5\u03c4\u03b1\u03c4\u03cc\u03c0\u03b9\u03c3\u03b7 calibration.<\/p>\n<p>\u03a3\u03c4\u03bf \u03c4\u03c1\u03af\u03c4\u03bf \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03bf, \u03b7 \u03b5\u03c1\u03bc\u03b7\u03bd\u03b5\u03c5\u03c3\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03b5\u03c2 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03c3\u03b5\u03b9\u03c2: probes \u03c0\u03c1\u03b9\u03bd \u03ba\u03b1\u03b9 \u03bc\u03b5\u03c4\u03ac \u03c4\u03bf reasoning, text baselines, class-balanced metrics \u03ba\u03b1\u03b9 tests \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03b1\u03bd\u03b1\u03bc\u03b9\u03b3\u03bd\u03cd\u03bf\u03c5\u03bd \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03bf\u03b9\u03ba\u03bf\u03b3\u03ad\u03bd\u03b5\u03b9\u03b5\u03c2 \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd. \u03a3\u03c4\u03bf \u03c4\u03ad\u03c4\u03b1\u03c1\u03c4\u03bf, \u03b7 \u03b1\u03b9\u03c4\u03b9\u03ce\u03b4\u03b7\u03c2 \u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03b5\u03be\u03b5\u03c4\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 perturbations \u03c0\u03bf\u03c5 \u03ba\u03c1\u03b1\u03c4\u03bf\u03cd\u03bd \u03c3\u03c4\u03b1\u03b8\u03b5\u03c1\u03ae \u03c4\u03b7 \u03bc\u03bf\u03c1\u03c6\u03ae \u03ba\u03b1\u03b9 \u03b1\u03bb\u03bb\u03ac\u03b6\u03bf\u03c5\u03bd \u03c4\u03bf \u03c0\u03b5\u03c1\u03b9\u03b5\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf. \u039a\u03b1\u03bc\u03af\u03b1 \u03bc\u03b5\u03bc\u03bf\u03bd\u03c9\u03bc\u03ad\u03bd\u03b7 \u03ad\u03bd\u03b4\u03b5\u03b9\u03be\u03b7 \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03c7\u03b1\u03c1\u03b1\u03ba\u03c4\u03b7\u03c1\u03b9\u03c3\u03c4\u03b5\u03af \u03ad\u03bd\u03b1 explanation \u03c0\u03b9\u03c3\u03c4\u03cc.<\/p>\n<p>\u0391\u03c5\u03c4\u03cc \u03c4\u03bf \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf \u03b4\u03b5\u03bd \u03b5\u03b3\u03b3\u03c5\u03ac\u03c4\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03b8\u03b1 \u03b2\u03c1\u03b5\u03b8\u03b5\u03af \u03bc\u03b9\u03b1 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae \u03bc\u03b5 \u03bc\u03ad\u03b3\u03b9\u03c3\u03c4\u03b7 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03c0\u03bb\u03ae\u03c1\u03b7 \u03b5\u03bb\u03b5\u03b3\u03be\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1. \u039a\u03ac\u03bd\u03b5\u03b9 \u03cc\u03bc\u03c9\u03c2 \u03c3\u03b1\u03c6\u03ad\u03c3\u03c4\u03b5\u03c1\u03bf \u03c0\u03bf\u03b9\u03b1 \u03b9\u03b4\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03bc\u03b5\u03c4\u03c1\u03ac \u03ba\u03ac\u03b8\u03b5 test \u03ba\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03c0\u03b9\u03b8\u03b1\u03bd\u03cc\u03c4\u03b7\u03c4\u03b1 \u03bd\u03b1 \u03b2\u03b1\u03c6\u03c4\u03b9\u03c3\u03c4\u03b5\u03af \u00ab\u03b4\u03b9\u03b1\u03c6\u03ac\u03bd\u03b5\u03b9\u03b1\u00bb \u03bc\u03b9\u03b1 \u03c0\u03b5\u03b9\u03c3\u03c4\u03b9\u03ba\u03ae \u03bb\u03b5\u03ba\u03c4\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03c6\u03ac\u03bd\u03b5\u03b9\u03b1.<\/p>\n<h2 id=\"oria-sosto-symperasma\">\u03a4\u03b1 \u03cc\u03c1\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03c4\u03bf \u03c3\u03c9\u03c3\u03c4\u03cc \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1<\/h2>\n<p>\u039f\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03c4\u03bf\u03bd\u03af\u03b6\u03bf\u03c5\u03bd \u03cc\u03c4\u03b9 \u03c4\u03b1 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03ac\u03c3\u03bc\u03b1\u03c4\u03b1 \u03c0\u03c1\u03bf\u03ad\u03c1\u03c7\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf task, \u03ba\u03c5\u03c1\u03af\u03c9\u03c2 \u03ad\u03bd\u03b1 model family \u03ba\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf \u03b5\u03cd\u03c1\u03bf\u03c2 \u03b1\u03bb\u03b3\u03bf\u03c1\u03af\u03b8\u03bc\u03c9\u03bd. \u039f\u03b9 \u03c3\u03c5\u03bd\u03bf\u03bc\u03b9\u03bb\u03af\u03b5\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c5\u03bd\u03b8\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2, \u03b5\u03bd\u03ce \u03bf\u03b9 \u03c3\u03c5\u03b3\u03ba\u03c1\u03af\u03c3\u03b5\u03b9\u03c2 GPT-4o, GPT-5.1, Qwen3-30B, Nemotron-3-Nano \u03ba\u03b1\u03b9 Llama-3.2-3B \u03b4\u03b5\u03bd \u03ad\u03c7\u03bf\u03c5\u03bd \u03cc\u03bb\u03b5\u03c2 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03bc\u03ad\u03b3\u03b5\u03b8\u03bf\u03c2 \u03b4\u03b5\u03af\u03b3\u03bc\u03b1\u03c4\u03bf\u03c2 \u03ae \u03c4\u03bf\u03bd \u03af\u03b4\u03b9\u03bf \u03c1\u03cc\u03bb\u03bf. \u03a4\u03b1 \u03b1\u03c0\u03cc\u03bb\u03c5\u03c4\u03b1 scores \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03b2\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03b1\u03bd \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae \u03ba\u03b1\u03c4\u03ac\u03c4\u03b1\u03be\u03b7 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03c9\u03bd.<\/p>\n<p>\u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03c0\u03c1\u03bf\u03c4\u03b5\u03af\u03bd\u03b5\u03b9 \u03c9\u03c2 \u03b5\u03c0\u03cc\u03bc\u03b5\u03bd\u03b1 \u03b2\u03ae\u03bc\u03b1\u03c4\u03b1 multi-round self-distillation \u03bc\u03b5 priors \u03c0\u03bf\u03c5 \u03c4\u03b1\u03b9\u03c1\u03b9\u03ac\u03b6\u03bf\u03c5\u03bd \u03c3\u03c4\u03bf deployment, objectives \u03c0\u03bf\u03c5 \u03b4\u03af\u03bd\u03bf\u03c5\u03bd \u03c3\u03ae\u03bc\u03b1 \u03b1\u03ba\u03cc\u03bc\u03b7 \u03ba\u03b1\u03b9 \u03cc\u03c4\u03b1\u03bd \u03cc\u03bb\u03b1 \u03c4\u03b1 rollouts \u03c3\u03c5\u03bc\u03c6\u03c9\u03bd\u03bf\u03cd\u03bd \u03c3\u03c4\u03b7 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1, rewards \u03b2\u03b1\u03c3\u03b9\u03c3\u03bc\u03ad\u03bd\u03b1 \u03c3\u03b5 content-controlled perturbations \u03ba\u03b1\u03b9 \u03c4\u03b1\u03c5\u03c4\u03cc\u03c7\u03c1\u03bf\u03bd\u03b7 \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7 \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03c9\u03bd \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03b9\u03ba\u03ce\u03bd \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03c3\u03c4\u03b1\u03c4\u03b9\u03ba\u03ce\u03bd audits.<\/p>\n<p>\u03a4\u03bf \u03c0\u03b9\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03b3\u03b9\u03b1 \u03c4\u03b7\u03bd \u03b1\u03b3\u03bf\u03c1\u03ac \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bb\u03cc \u03b1\u03bb\u03bb\u03ac \u03b1\u03c0\u03b1\u03b9\u03c4\u03b7\u03c4\u03b9\u03ba\u03cc: \u03b7 \u03b5\u03be\u03ae\u03b3\u03b7\u03c3\u03b7 \u03b5\u03bd\u03cc\u03c2 agent \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03bf\u03c5 interface \u03ba\u03b1\u03b9 \u03c0\u03b9\u03b8\u03b1\u03bd\u03cc \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03bf\u03c5 policy. \u0394\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b8\u03b5\u03c9\u03c1\u03b5\u03af\u03c4\u03b1\u03b9 \u03b1\u03c5\u03c4\u03cc\u03bc\u03b1\u03c4\u03b1 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ae\u03c2 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03b1\u03c6\u03ae \u03bc\u03b9\u03b1\u03c2 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2 \u03c0\u03bf\u03c5 \u03b8\u03b1 \u03ae\u03c4\u03b1\u03bd \u03af\u03b4\u03b9\u03b1 \u03c7\u03c9\u03c1\u03af\u03c2 \u03b1\u03c5\u03c4\u03ae\u03bd. \u03a4\u03bf <a href=\"https:\/\/twodots.gr\/agentic-ai-security-stack-controls-technologikes-ependyseis\/\">agentic AI security stack<\/a> \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03b5\u03b9\u03c1\u03ac\u03bc\u03b1\u03c4\u03b1 \u03c0\u03bf\u03c5 \u03b4\u03b9\u03b1\u03c7\u03c9\u03c1\u03af\u03b6\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03b7\u03c2 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7\u03c2, \u03c4\u03b7 \u03bb\u03b5\u03ba\u03c4\u03b9\u03ba\u03ae \u03c0\u03b5\u03b9\u03c3\u03c4\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1, \u03c4\u03b7\u03bd \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ae \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03b1\u03b9\u03c4\u03b9\u03ce\u03b4\u03b7 \u03c3\u03c5\u03bc\u03b2\u03bf\u03bb\u03ae \u03c4\u03bf\u03c5 reasoning.<\/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 rationale \u03c3\u03b5 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03b7 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7<\/p>\n<h3>\u03a3\u03c7\u03b5\u03b4\u03b9\u03ac\u03c3\u03c4\u03b5 AI agents \u03c0\u03bf\u03c5 \u03be\u03ad\u03c1\u03bf\u03c5\u03bd \u03c0\u03cc\u03c4\u03b5 \u03bd\u03b1 \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03cd\u03bd, \u03c0\u03cc\u03c4\u03b5 \u03bd\u03b1 \u03b1\u03c0\u03ad\u03c7\u03bf\u03c5\u03bd \u03ba\u03b1\u03b9 \u03c0\u03cc\u03c4\u03b5 \u03bd\u03b1 \u03b6\u03b7\u03c4\u03bf\u03cd\u03bd \u03ac\u03bd\u03b8\u03c1\u03c9\u03c0\u03bf<\/h3>\n<p>\u0397 TWO DOTS \u03c7\u03b1\u03c1\u03c4\u03bf\u03b3\u03c1\u03b1\u03c6\u03b5\u03af intervention policies, risk thresholds, action logs, quality gates \u03ba\u03b1\u03b9 human handoffs \u03ce\u03c3\u03c4\u03b5 \u03b7 \u03b5\u03be\u03ae\u03b3\u03b7\u03c3\u03b7 \u03bd\u03b1 \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03ce\u03bd\u03b5\u03b9 \u2014 \u03cc\u03c7\u03b9 \u03bd\u03b1 \u03c5\u03c0\u03bf\u03ba\u03b1\u03b8\u03b9\u03c3\u03c4\u03ac \u2014 \u03c4\u03b7 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03c4\u03bf\u03c5 agent.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">Check out Business Automation &amp; AI<\/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\">Frequently Asked Questions (FAQs)<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b5\u03af\u03bd\u03b1\u03b9 \u03ad\u03bd\u03b1 abstaining action policy;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03af\u03bd\u03b1\u03b9 \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae \u03cc\u03c0\u03bf\u03c5 \u03bf agent \u03b5\u03c0\u03b9\u03bb\u03ad\u03b3\u03b5\u03b9 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03c4\u03b7\u03bd \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b7\u03bd \u03b1\u03c0\u03bf\u03c7\u03ae. \u03a3\u03c4\u03bf Why2Speak \u03b1\u03c0\u03bf\u03c6\u03b1\u03c3\u03af\u03b6\u03b5\u03b9 \u03b1\u03bd \u03b8\u03b1 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b5\u03b9 \u03c4\u03ce\u03c1\u03b1 \u03c3\u03b5 \u03bc\u03b9\u03b1 \u03c3\u03c5\u03b6\u03ae\u03c4\u03b7\u03c3\u03b7 \u03ae \u03b1\u03bd \u03b8\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bc\u03b5\u03af\u03bd\u03b5\u03b9 \u03c3\u03b9\u03c9\u03c0\u03b7\u03bb\u03cc\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03bf chain of thought \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03bd\u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03bf;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9. \u0394\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1 reasoning trace \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b8\u03b5\u03c9\u03c1\u03b5\u03af\u03c4\u03b1\u03b9 \u03b1\u03c5\u03c4\u03cc\u03bc\u03b1\u03c4\u03b1 \u03c0\u03b9\u03c3\u03c4\u03cc. \u039c\u03b5\u03c4\u03ac \u03b1\u03c0\u03cc content controls, \u03c4\u03bf reasoning \u03b5\u03af\u03c7\u03b5 \u03bc\u03ad\u03c4\u03c1\u03b9\u03b1 \u03b1\u03bb\u03bb\u03ac \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03c3\u03c5\u03bc\u03b2\u03bf\u03bb\u03ae \u03c3\u03c4\u03b7\u03bd \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 direct policy \u03b5\u03af\u03c7\u03b5 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b7 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039f token-SFT classifier \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03cd\u03c4\u03b7\u03ba\u03b5 \u03b5\u03b9\u03b4\u03b9\u03ba\u03ac \u03b3\u03b9\u03b1 \u03c4\u03b7\u03bd \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03b4\u03c5\u03b1\u03b4\u03b9\u03ba\u03ae \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 macro-F1 0,620. \u038c\u03c4\u03b1\u03bd \u03c4\u03bf \u03af\u03b4\u03b9\u03bf checkpoint \u03bc\u03c0\u03ae\u03ba\u03b5 \u03c3\u03b5 think mode, \u03b7 \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03ad\u03c0\u03b5\u03c3\u03b5 \u03c3\u03c4\u03bf 0,367.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf macro-F1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03cc \u03c3\u03c4\u03bf When2Speak;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039f\u03b9 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b5\u03c5\u03ba\u03b1\u03b9\u03c1\u03af\u03b5\u03c2 \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 13% \u03c4\u03c9\u03bd decision points. \u03a4\u03bf macro-F1 \u03b4\u03af\u03bd\u03b5\u03b9 \u03af\u03c3\u03bf \u03b2\u03ac\u03c1\u03bf\u03c2 \u03c3\u03c4\u03b9\u03c2 \u03b4\u03cd\u03bf \u03ba\u03bb\u03ac\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03c0\u03b1\u03c1\u03b1\u03c0\u03bb\u03ac\u03bd\u03b7\u03c3\u03b7 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03c0\u03bb\u03b5\u03b9\u03bf\u03c8\u03b7\u03c6\u03b9\u03ba\u03ae \u03c3\u03b9\u03c9\u03c0\u03ae.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf GRPO \u03b4\u03b5\u03bd \u03b4\u03b9\u03cc\u03c1\u03b8\u03c9\u03c3\u03b5 \u03c4\u03b1 confidently wrong prompts;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c4\u03b1\u03bd \u03cc\u03bb\u03b1 \u03c4\u03b1 rollouts \u03bc\u03b9\u03b1\u03c2 \u03bf\u03bc\u03ac\u03b4\u03b1\u03c2 \u03b5\u03c0\u03ad\u03bb\u03b5\u03b3\u03b1\u03bd \u03c4\u03b7\u03bd \u03af\u03b4\u03b9\u03b1 \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b7 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b1, \u03b4\u03b5\u03bd \u03c5\u03c0\u03ae\u03c1\u03c7\u03b5 \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ae \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03b1\u03bd\u03c4\u03b1\u03bc\u03bf\u03b9\u03b2\u03ae\u03c2 \u03b3\u03b9\u03b1 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf learning signal. \u03a0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf gradient \u03b4\u03b5\u03bd \u03ac\u03bb\u03bb\u03b1\u03be\u03b5 \u03c4\u03bf plateau.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b5\u03bb\u03ad\u03b3\u03c7\u03b5\u03b9 \u03c4\u03bf neutral-filler control;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0391\u03bd\u03c4\u03b9\u03ba\u03b1\u03b8\u03b9\u03c3\u03c4\u03ac \u03c4\u03bf \u03b1\u03c6\u03b1\u03b9\u03c1\u03b5\u03bc\u03ad\u03bd\u03bf reasoning \u03bc\u03b5 \u03bf\u03c5\u03b4\u03ad\u03c4\u03b5\u03c1\u03bf \u03ba\u03b5\u03af\u03bc\u03b5\u03bd\u03bf \u03af\u03b4\u03b9\u03bf\u03c5 \u03bc\u03ae\u03ba\u03bf\u03c5\u03c2. \u0388\u03c4\u03c3\u03b9 \u03be\u03b5\u03c7\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b5\u03c0\u03af\u03b4\u03c1\u03b1\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03c0\u03b5\u03c1\u03b9\u03b5\u03c7\u03bf\u03bc\u03ad\u03bd\u03bf\u03c5 \u03b1\u03c0\u03cc \u03c4\u03b7\u03bd \u03b5\u03c0\u03af\u03b4\u03c1\u03b1\u03c3\u03b7 \u03c4\u03b7\u03c2 \u03bc\u03bf\u03c1\u03c6\u03ae\u03c2 \u03ae \u03c4\u03b7\u03c2 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae\u03c2 inference mode.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u039c\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03c4\u03b1 activation probes \u03bd\u03b1 \u03b1\u03c0\u03bf\u03b4\u03b5\u03af\u03be\u03bf\u03c5\u03bd \u03b1\u03b9\u03c4\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9 \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03b1 \u03c4\u03bf\u03c5\u03c2. \u0394\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03cc\u03c4\u03b9 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bf\u03ba\u03c9\u03b4\u03b9\u03ba\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b9\u03bc\u03b7 \u03b1\u03c0\u03cc activations. \u03a7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03b9\u03ba\u03ad\u03c2 \u03c0\u03b1\u03c1\u03b5\u03bc\u03b2\u03ac\u03c3\u03b5\u03b9\u03c2, baselines \u03ba\u03b1\u03b9 controls \u03b3\u03b9\u03b1 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03ce\u03c2 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7 \u03bd\u03b1 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af \u03ad\u03bd\u03b1\u03bd AI agent \u03c0\u03bf\u03c5 \u03c0\u03b1\u03c1\u03b5\u03bc\u03b2\u03b1\u03af\u03bd\u03b5\u03b9;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039d\u03b1 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ac\u03b6\u03b5\u03b9 \u03c4\u03bf \u03b1\u03ba\u03c1\u03b9\u03b2\u03ad\u03c2 deployed mode, \u03bd\u03b1 \u03bc\u03b5\u03c4\u03c1\u03ac \u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac \u03c0\u03b5\u03c1\u03b9\u03c4\u03c4\u03ad\u03c2 \u03ba\u03b1\u03b9 \u03c7\u03b1\u03bc\u03ad\u03bd\u03b5\u03c2 \u03c0\u03b1\u03c1\u03b5\u03bc\u03b2\u03ac\u03c3\u03b5\u03b9\u03c2, \u03bd\u03b1 \u03b5\u03bb\u03ad\u03b3\u03c7\u03b5\u03b9 calibration \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03ba\u03c1\u03b1\u03c4\u03ac \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03bf override \u03b3\u03b9\u03b1 \u03b5\u03bd\u03ad\u03c1\u03b3\u03b5\u03b9\u03b5\u03c2 \u03c5\u03c8\u03b7\u03bb\u03bf\u03cd \u03c1\u03af\u03c3\u03ba\u03bf\u03c5.<\/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<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.20670\" target=\"_blank\" rel=\"noopener\">arXiv \u2014 Why2Speak: Faithful Reasoning for Abstaining Action Policies<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2605.05626\" target=\"_blank\" rel=\"noopener\">arXiv \u2014 When2Speak: A Dataset for Temporal Participation and Turn-Taking<\/a><\/li>\n<li><a href=\"https:\/\/www.nist.gov\/publications\/artificial-intelligence-risk-management-framework-generative-artificial-intelligence\" target=\"_blank\" rel=\"noopener\">NIST \u2014 AI Risk Management Framework: Generative AI Profile<\/a><\/li>\n<li><a href=\"https:\/\/genai.owasp.org\/resource\/agentic-ai-threats-and-mitigations\/\" target=\"_blank\" rel=\"noopener\">OWASP \u2014 Agentic AI: Threats and Mitigations<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>\u03a4\u03bf Why2Speak \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1\u03c4\u03af \u03c4\u03bf reasoning \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03c0\u03bf\u03bb\u03b9\u03c4\u03b9\u03ba\u03ae \u03c0\u03b1\u03c1\u03ad\u03bc\u03b2\u03b1\u03c3\u03b7\u03c2 \u03b5\u03bd\u03cc\u03c2 AI agent \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03bf\u03b9 \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03b9 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c0\u03c1\u03b9\u03bd \u03b1\u03c0\u03cc deployment.<\/p>","protected":false},"author":1,"featured_media":90855,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","content-type":"","footnotes":""},"categories":[199],"tags":[6506,18337,18338,7549,7477],"class_list":["post-90851","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-ai-agents","tag-ai-safety","tag-chain-of-thought","tag-explainable-ai","tag-machine-learning"],"blocksy_meta":{"styles_descriptor":{"styles":{"desktop":"","tablet":"","mobile":""},"google_fonts":[],"version":7}},"_links":{"self":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/90851","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/comments?post=90851"}],"version-history":[{"count":1,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/90851\/revisions"}],"predecessor-version":[{"id":90856,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/posts\/90851\/revisions\/90856"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media\/90855"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/media?parent=90851"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/categories?post=90851"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/en\/wp-json\/wp\/v2\/tags?post=90851"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}