{"id":98920,"date":"2026-10-18T22:30:06","date_gmt":"2026-10-18T19:30:06","guid":{"rendered":"https:\/\/twodots.gr\/?p=98920"},"modified":"2026-10-18T22:30:08","modified_gmt":"2026-10-18T19:30:08","slug":"step-ktoder-ai-lathos-synartisi","status":"publish","type":"post","link":"https:\/\/twodots.gr\/bg\/step-ktoder-ai-lathos-synartisi\/","title":{"rendered":"STEP-KTODER: \u03b3\u03b9\u03b1\u03c4\u03af \u03c4\u03bf AI \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b1\u03b8\u03b1\u03af\u03bd\u03b5\u03b9 \u03c0\u03bf\u03b9\u03b1 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03ad\u03ba\u03b1\u03bd\u03b5 \u03c4\u03bf \u03bb\u03ac\u03b8\u03bf\u03c2"},"content":{"rendered":"<p>\u0388\u03bd\u03b1 \u03c0\u03c1\u03cc\u03b3\u03c1\u03b1\u03bc\u03bc\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c4\u03c5\u03b3\u03c7\u03ac\u03bd\u03b5\u03b9 \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03bc\u03af\u03b1 \u03bc\u03cc\u03bd\u03bf \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03c7\u03b5\u03b9\u03c1\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03ad\u03bd\u03b1 edge case. \u0391\u03bd \u03cc\u03bc\u03c9\u03c2 \u03c4\u03bf \u03b5\u03ba\u03c0\u03b1\u03b9\u03b4\u03b5\u03c5\u03c4\u03b9\u03ba\u03cc \u03c3\u03ae\u03bc\u03b1 \u03c7\u03b1\u03c1\u03b1\u03ba\u03c4\u03b7\u03c1\u03af\u03b6\u03b5\u03b9 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03b7 \u03c4\u03b7 \u03bb\u03cd\u03c3\u03b7 \u00ab\u03ba\u03b1\u03ba\u03ae\u00bb, \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c4\u03b9\u03bc\u03c9\u03c1\u03b5\u03af\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03c4\u03b1 \u03c4\u03bc\u03ae\u03bc\u03b1\u03c4\u03b1 \u03c0\u03bf\u03c5 \u03ad\u03b3\u03c1\u03b1\u03c8\u03b5 \u03c3\u03c9\u03c3\u03c4\u03ac. \u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 <strong>Function-Level Execution Feedback for Code Preference Optimization<\/strong> \u03c0\u03c1\u03bf\u03c4\u03b5\u03af\u03bd\u03b5\u03b9 \u03b1\u03ba\u03c1\u03b9\u03b2\u03ad\u03c3\u03c4\u03b5\u03c1\u03b7 \u03b5\u03c0\u03bf\u03c0\u03c4\u03b5\u03af\u03b1: \u03ba\u03ac\u03b8\u03b5 module-level function \u03b1\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03c9\u03c0\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c9\u03c2 \u03b2\u03ae\u03bc\u03b1, \u03b5\u03bb\u03ad\u03b3\u03c7\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c3\u03c4\u03bf\u03c7\u03b5\u03c5\u03bc\u03ad\u03bd\u03b1 unit tests \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03b4\u03c5\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c4\u03bf\u03c5 \u03c0\u03bb\u03ae\u03c1\u03bf\u03c5\u03c2 \u03c0\u03c1\u03bf\u03b3\u03c1\u03ac\u03bc\u03bc\u03b1\u03c4\u03bf\u03c2.<\/p>\n<p>\u03a4\u03bf \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf \u03bf\u03bd\u03bf\u03bc\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 <strong>Step-KTOder<\/strong> \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u03bf\u03c5\u03c3\u03b9\u03ac\u03c3\u03c4\u03b7\u03ba\u03b5 \u03b1\u03c0\u03cc \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ad\u03c2 \u03c4\u03c9\u03bd Seoul National University \u03ba\u03b1\u03b9 Konkuk University. \u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1, \u03c0\u03bf\u03c5 \u03ad\u03b3\u03b9\u03bd\u03b5 \u03b4\u03b5\u03ba\u03c4\u03ae \u03c3\u03c4\u03b1 Findings of EMNLP 2026, \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af \u03c4\u03b7 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf \u03c3\u03b5 HumanEval(+), MBPP(+), BigCodeBench \u03ba\u03b1\u03b9 LiveCodeBench. \u0393\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b5\u03c2 \u03c0\u03bf\u03c5 \u03c7\u03c4\u03af\u03b6\u03bf\u03c5\u03bd coding copilots, agents \u03ae \u03b5\u03c3\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03ac \u03b5\u03c1\u03b3\u03b1\u03bb\u03b5\u03af\u03b1 \u03b1\u03bd\u03ac\u03c0\u03c4\u03c5\u03be\u03b7\u03c2, \u03c4\u03bf \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03cc \u03bc\u03ae\u03bd\u03c5\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03bb\u03ce\u03c2 \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1 \u03bd\u03ad\u03bf objective \u03b1\u03bd\u03ad\u03b2\u03b1\u03c3\u03b5 \u03ad\u03bd\u03b1 benchmark. \u0395\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03b7 \u03b1\u03be\u03af\u03b1 \u03c4\u03bf\u03c5 feedback \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03b4\u03cd\u03bf \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ad\u03c2: \u03c3\u03b5 \u03c0\u03bf\u03b9\u03bf \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03bf \u03b4\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03b1\u03bd \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03c5\u03c4\u03b5\u03af \u03bc\u03b5 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7.<\/p>\n<div class=\"td-note\">\n<p class=\"td-note-title\">\u0397 \u03c3\u03cd\u03bd\u03c4\u03bf\u03bc\u03b7 \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7<\/p>\n<p>\u03a4\u03bf Step-KTOder \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03bf AI \u03bc\u03b1\u03b8\u03b1\u03af\u03bd\u03b5\u03b9 \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b1 \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1 \u03c3\u03cd\u03bd\u03b8\u03b5\u03c4\u03bf \u03c0\u03c1\u03cc\u03b3\u03c1\u03b1\u03bc\u03bc\u03b1 \u03cc\u03c4\u03b1\u03bd \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03bf\u03cd\u03bc\u03b5 \u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac \u03c4\u03b7 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1 \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03ac\u03b8\u03b5 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03b7\u03c2 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7\u03c2. \u03a4\u03bf execution-based feedback \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03b5\u03b9 \u03c3\u03c5\u03c7\u03bd\u03cc\u03c4\u03b5\u03c1\u03b1 \u03c4\u03bf \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc \u03c3\u03b7\u03bc\u03b5\u03af\u03bf \u03b2\u03bb\u03ac\u03b2\u03b7\u03c2, \u03b5\u03bd\u03ce \u03bc\u03b9\u03b1 \u03bb\u03b5\u03ba\u03c4\u03b9\u03ba\u03ae \u03ba\u03c1\u03af\u03c3\u03b7 \u03b1\u03c0\u03cc \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf LLM \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03bb\u03bb\u03bf\u03b9\u03ce\u03c3\u03b5\u03b9 \u03c4\u03b1 \u03b8\u03b5\u03c4\u03b9\u03ba\u03ac labels.<\/p>\n<\/div>\n<div class=\"td-article-toc\">\n<div class=\"td-toc-title\">\u0421\u044a\u0434\u044a\u0440\u0436\u0430\u043d\u0438\u0435<\/div>\n<ul>\n<li><a href=\"#giati-pass-fail-chontro-sima\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc pass \u03ae fail \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03bf\u03bb\u03cd \u03c7\u03bf\u03bd\u03c4\u03c1\u03cc \u03c3\u03ae\u03bc\u03b1<\/a><\/li>\n<li><a href=\"#pos-kataskeuazetai-step-ktoder\">\u03a0\u03ce\u03c2 \u03ba\u03b1\u03c4\u03b1\u03c3\u03ba\u03b5\u03c5\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c4\u03bf Step-KTOder<\/a><\/li>\n<li><a href=\"#local-global-orthotita-sygkrousi\">\u038c\u03c4\u03b1\u03bd \u03c4\u03bf\u03c0\u03b9\u03ba\u03ae \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c3\u03c5\u03b3\u03ba\u03c1\u03bf\u03cd\u03bf\u03bd\u03c4\u03b1\u03b9<\/a><\/li>\n<li><a href=\"#ti-edeixan-benchmarks\">\u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03c4\u03b1 benchmarks<\/a><\/li>\n<li><a href=\"#function-labels-entopizoun-vlaves\">\u03a4\u03b1 function labels \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03bf\u03c5\u03bd \u03c4\u03b9\u03c2 \u03b2\u03bb\u03ac\u03b2\u03b5\u03c2;<\/a><\/li>\n<li><a href=\"#execution-vs-llm-judge\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf\u03bd LLM judge<\/a><\/li>\n<li><a href=\"#poiotita-unit-tests\">\u03a0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 unit tests: \u03c4\u03bf \u03b1\u03b8\u03ad\u03b1\u03c4\u03bf \u03b8\u03b5\u03bc\u03ad\u03bb\u03b9\u03bf<\/a><\/li>\n<li><a href=\"#coding-agents-production\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1 coding agents \u03ba\u03b1\u03b9 production workflows<\/a><\/li>\n<li><a href=\"#oria-step-ktoder\">\u03a4\u03b1 \u03cc\u03c1\u03b9\u03b1 \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03b3\u03bd\u03bf\u03b7\u03b8\u03bf\u03cd\u03bd<\/a><\/li>\n<li><a href=\"#praktiki-diadromi-efarmogis\">\u039c\u03b9\u03b1 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03ae \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae\u03c2<\/a><\/li>\n<li><a href=\"#ti-kratame-step-ktoder\">\u03a4\u03b9 \u03ba\u03c1\u03b1\u03c4\u03ac\u03bc\u03b5 \u03b1\u03c0\u03cc \u03c4\u03bf Step-KTOder<\/a><\/li>\n<\/ul>\n<\/div>\n<h2 id=\"giati-pass-fail-chontro-sima\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc pass \u03ae fail \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03bf\u03bb\u03cd \u03c7\u03bf\u03bd\u03c4\u03c1\u03cc \u03c3\u03ae\u03bc\u03b1<\/h2>\n<p>\u039f\u03b9 \u03c3\u03c5\u03bd\u03b7\u03b8\u03b9\u03c3\u03bc\u03ad\u03bd\u03b5\u03c2 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03b9 preference optimization \u03b1\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03c9\u03c0\u03af\u03b6\u03bf\u03c5\u03bd \u03c3\u03c5\u03c7\u03bd\u03ac \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03b7 \u03c4\u03b7\u03bd \u03b1\u03c0\u03ac\u03bd\u03c4\u03b7\u03c3\u03b7 \u03c9\u03c2 \u03bc\u03af\u03b1 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b1. \u03a4\u03bf Direct Preference Optimization, \u03ae DPO, \u03bc\u03b1\u03b8\u03b1\u03af\u03bd\u03b5\u03b9 \u03b1\u03c0\u03cc \u03b6\u03b5\u03cd\u03b3\u03b7 \u03c0\u03c1\u03bf\u03c4\u03b9\u03bc\u03ce\u03bc\u03b5\u03bd\u03c9\u03bd \u03ba\u03b1\u03b9 \u03bc\u03b7 \u03c0\u03c1\u03bf\u03c4\u03b9\u03bc\u03ce\u03bc\u03b5\u03bd\u03c9\u03bd outputs. \u03a4\u03bf Kahneman-Tversky Optimization, \u03ae KTO, \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03bc\u03ac\u03b8\u03b5\u03b9 \u03b1\u03c0\u03cc \u03bc\u03b5\u03bc\u03bf\u03bd\u03c9\u03bc\u03ad\u03bd\u03b1 outputs \u03bc\u03b5 \u03b4\u03c5\u03b1\u03b4\u03b9\u03ba\u03cc desirable \u03ae undesirable label. \u038c\u03c4\u03b1\u03bd \u03c4\u03bf output \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1\u03c2, \u03b1\u03c5\u03c4\u03cc \u03c4\u03bf label \u03c0\u03c1\u03bf\u03ba\u03cd\u03c0\u03c4\u03b5\u03b9 \u03c3\u03c5\u03bd\u03ae\u03b8\u03c9\u03c2 \u03b1\u03c0\u03cc \u03c4\u03bf \u03b1\u03bd \u03b7 \u03c0\u03bb\u03ae\u03c1\u03b7\u03c2 \u03bb\u03cd\u03c3\u03b7 \u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf \u03b4\u03b9\u03b1\u03b8\u03ad\u03c3\u03b9\u03bc\u03bf test suite.<\/p>\n<p>\u0397 \u03b1\u03c0\u03bb\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2. \u039c\u03b9\u03b1 \u03bb\u03cd\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03c3\u03c9\u03c3\u03c4\u03ad\u03c2 \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 \u03bc\u03af\u03b1 \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b7 \u03bb\u03b1\u03bc\u03b2\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03b1\u03c1\u03bd\u03b7\u03c4\u03b9\u03ba\u03cc outcome label \u03bc\u03b5 \u03bc\u03b9\u03b1 \u03bb\u03cd\u03c3\u03b7 \u03c0\u03bf\u03c5 \u03b1\u03c0\u03bf\u03c4\u03c5\u03b3\u03c7\u03ac\u03bd\u03b5\u03b9 \u03c0\u03b1\u03bd\u03c4\u03bf\u03cd. \u03a4\u03bf feedback \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 \u03c0\u03bf\u03b9\u03bf \u03c4\u03bc\u03ae\u03bc\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b7\u03b8\u03b5\u03af \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03bf \u03bd\u03b1 \u03b4\u03b9\u03bf\u03c1\u03b8\u03c9\u03b8\u03b5\u03af. \u03a3\u03c4\u03b1 \u03bc\u03b1\u03b8\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac, \u03b7 process supervision \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03ae\u03c3\u03b5\u03b9 \u03c4\u03b1 \u03b5\u03bd\u03b4\u03b9\u03ac\u03bc\u03b5\u03c3\u03b1 \u03b2\u03ae\u03bc\u03b1\u03c4\u03b1 \u03b5\u03bd\u03cc\u03c2 \u03c3\u03c5\u03bb\u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03bf\u03cd. \u03a3\u03c4\u03bf\u03bd \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1 \u03b4\u03b5\u03bd \u03c5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ac \u03b1\u03c0\u03bf\u03b4\u03b5\u03ba\u03c4\u03cc \u00ab\u03b2\u03ae\u03bc\u03b1\u00bb: \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b3\u03c1\u03b1\u03bc\u03bc\u03ae, block, \u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7\u03c2, reasoning trace \u03ae \u03ac\u03bb\u03bb\u03bf \u03b4\u03bf\u03bc\u03b9\u03ba\u03cc \u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03af\u03bf.<\/p>\n<p>\u039f\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03b5\u03c0\u03b9\u03bb\u03ad\u03b3\u03bf\u03c5\u03bd \u03c4\u03b7 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03c9\u03c2 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03ae \u03bc\u03bf\u03bd\u03ac\u03b4\u03b1. \u0395\u03af\u03bd\u03b1\u03b9 \u03c3\u03b7\u03bc\u03b1\u03c3\u03b9\u03bf\u03bb\u03bf\u03b3\u03b9\u03ba\u03ac \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ae, \u03ad\u03c7\u03b5\u03b9 interface \u03ba\u03b1\u03b9 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af, \u03c4\u03bf\u03c5\u03bb\u03ac\u03c7\u03b9\u03c3\u03c4\u03bf\u03bd \u03c3\u03b5 \u03b1\u03c1\u03ba\u03b5\u03c4\u03ad\u03c2 \u03c0\u03b5\u03c1\u03b9\u03c0\u03c4\u03ce\u03c3\u03b5\u03b9\u03c2, \u03bd\u03b1 \u03b4\u03bf\u03ba\u03b9\u03bc\u03b1\u03c3\u03c4\u03b5\u03af \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b1. \u0388\u03c4\u03c3\u03b9 \u03b7 \u03b5\u03c0\u03bf\u03c0\u03c4\u03b5\u03af\u03b1 \u03b4\u03b5\u03bd \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03b1\u03c0\u03cc \u03bc\u03b9\u03b1 \u03bb\u03b5\u03ba\u03c4\u03b9\u03ba\u03ae \u03ba\u03c1\u03af\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03c4\u03bf \u03b1\u03bd \u03bf \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1\u03c2 \u00ab\u03c6\u03b1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c3\u03c9\u03c3\u03c4\u03cc\u03c2\u00bb, \u03b1\u03bb\u03bb\u03ac \u03b1\u03c0\u03cc \u03c0\u03b1\u03c1\u03b1\u03c4\u03b7\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ac.<\/p>\n<div class=\"td-comparison td-comparison-cards--horizontal\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-kicker\">\u03a4\u03c1\u03af\u03b1 \u03b5\u03c0\u03af\u03c0\u03b5\u03b4\u03b1 feedback<\/p>\n<p class=\"td-chart-title\">\u03a4\u03bf \u03af\u03b4\u03b9\u03bf failure \u03b4\u03af\u03bd\u03b5\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ae \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03b1\u03bd\u03ac\u03bb\u03bf\u03b3\u03b1 \u03bc\u03b5 \u03c4\u03bf \u03c3\u03b7\u03bc\u03b5\u03af\u03bf \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7\u03c2<\/p>\n<p class=\"td-chart-intro\">\u0397 \u03b4\u03b9\u03ac\u03ba\u03c1\u03b9\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03b9\u03b1\u03ba\u03bf\u03c3\u03bc\u03b7\u03c4\u03b9\u03ba\u03ae: \u03ba\u03b1\u03b8\u03bf\u03c1\u03af\u03b6\u03b5\u03b9 \u03c0\u03bf\u03b9\u03b1 tokens \u03b5\u03bd\u03b9\u03c3\u03c7\u03cd\u03bf\u03bd\u03c4\u03b1\u03b9, \u03c0\u03bf\u03b9\u03b1 \u03c4\u03b9\u03bc\u03c9\u03c1\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03b1 \u03b1\u03c3\u03c4\u03bf\u03c7\u03af\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03c7\u03b8\u03b5\u03af.<\/p>\n<\/div>\n<div class=\"td-comparison-grid\">\n<div class=\"td-platform-card\">\n<p class=\"td-platform-name\">Outcome label<\/p>\n<p>\u039b\u03ad\u03b5\u03b9 \u03b1\u03bd \u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c4\u03bf \u03c0\u03c1\u03cc\u03b3\u03c1\u03b1\u03bc\u03bc\u03b1, \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03c5\u03c0\u03b5\u03cd\u03b8\u03c5\u03bd\u03b7 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7.<\/p>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-platform-name\">Function label<\/p>\n<p>\u0395\u03bb\u03ad\u03b3\u03c7\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b7 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03bc\u03b5 unit tests \u03ba\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03c4\u03bf \u03c3\u03ae\u03bc\u03b1 \u03c4\u03c9\u03bd \u03c3\u03c9\u03c3\u03c4\u03ce\u03bd \u03bc\u03b5\u03c1\u03ce\u03bd.<\/p>\n<\/div>\n<div class=\"td-platform-card\">\n<p class=\"td-platform-name\">Audit evidence<\/p>\n<p>\u039a\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 input, assertion \u03ba\u03b1\u03b9 \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 candidate \u03ce\u03c3\u03c4\u03b5 \u03b7 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03bd\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03c0\u03b1\u03bd\u03b5\u03bb\u03b5\u03b3\u03c7\u03b8\u03b5\u03af.<\/p>\n<\/div>\n<\/div>\n<\/div>\n<h2 id=\"pos-kataskeuazetai-step-ktoder\">\u03a0\u03ce\u03c2 \u03ba\u03b1\u03c4\u03b1\u03c3\u03ba\u03b5\u03c5\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c4\u03bf Step-KTOder<\/h2>\n<p>\u0397 \u03b4\u03b9\u03b1\u03b4\u03b9\u03ba\u03b1\u03c3\u03af\u03b1 \u03ad\u03c7\u03b5\u03b9 \u03c4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b1 \u03b2\u03b1\u03c3\u03b9\u03ba\u03ac \u03c3\u03c4\u03ac\u03b4\u03b9\u03b1. \u03a0\u03c1\u03ce\u03c4\u03b1, \u03ad\u03bd\u03b1 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc instruction-tuned code model \u03be\u03b1\u03bd\u03b1\u03b3\u03c1\u03ac\u03c6\u03b5\u03b9 \u03c4\u03b7 reference solution \u03c9\u03c2 \u03c0\u03c1\u03cc\u03b3\u03c1\u03b1\u03bc\u03bc\u03b1 \u03c0\u03bf\u03bb\u03bb\u03ce\u03bd \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03c9\u03bd \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03bf \u03b5\u03be\u03c9\u03c4\u03b5\u03c1\u03b9\u03ba\u03cc interface. \u0397 \u03b1\u03bd\u03b1\u03b4\u03b9\u03b1\u03c4\u03c5\u03c0\u03c9\u03bc\u03ad\u03bd\u03b7 \u03bb\u03cd\u03c3\u03b7 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03b1\u03bd \u03b5\u03be\u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03b5\u03af \u03bd\u03b1 \u03c0\u03b5\u03c1\u03bd\u03ac \u03c4\u03bf \u03b1\u03c1\u03c7\u03b9\u03ba\u03cc test suite. \u0391\u03c5\u03c4\u03cc \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03ad\u03bd\u03b1\u03bd \u03c3\u03ba\u03b5\u03bb\u03b5\u03c4\u03cc \u03cc\u03c0\u03bf\u03c5 \u03ba\u03ac\u03b8\u03b5 module-level function \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03bd\u03c4\u03b9\u03bc\u03b5\u03c4\u03c9\u03c0\u03b9\u03c3\u03c4\u03b5\u03af \u03c9\u03c2 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03cc step.<\/p>\n<p>\u0394\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf\u03bd, \u03c0\u03b1\u03c1\u03ac\u03b3\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c3\u03c4\u03bf\u03c7\u03b5\u03c5\u03bc\u03ad\u03bd\u03b1 unit tests \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03c0\u03b9\u03b8\u03b1\u03bd\u03ac edge cases. \u0394\u03b5\u03bd \u03b3\u03af\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b4\u03b5\u03ba\u03c4\u03ac \u03cc\u03bb\u03b1 \u03c4\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03cc\u03bc\u03b5\u03bd\u03b1 tests. \u03a0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03ba\u03ac\u03bd\u03bf\u03c5\u03bd parse, \u03bd\u03b1 \u03b5\u03ba\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03ce\u03c2 \u03c0\u03ac\u03bd\u03c9 \u03c3\u03c4\u03b7 reference implementation, \u03bd\u03b1 \u03ba\u03b1\u03bb\u03bf\u03cd\u03bd \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03c4\u03b7 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7-\u03c3\u03c4\u03cc\u03c7\u03bf \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03c0\u03b5\u03c1\u03b9\u03ad\u03c7\u03bf\u03c5\u03bd \u03bc\u03b7 \u03c4\u03b5\u03c4\u03c1\u03b9\u03bc\u03bc\u03ad\u03bd\u03b1 assertions \u03c4\u03c9\u03bd \u03bf\u03c0\u03bf\u03af\u03c9\u03bd \u03c4\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf output \u03c4\u03b7\u03c2. \u0391\u03bd \u03b4\u03b5\u03bd \u03c5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 \u03ad\u03b3\u03ba\u03c5\u03c1\u03bf \u03c4\u03bf\u03c0\u03b9\u03ba\u03cc test, \u03b7 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03c0\u03b1\u03af\u03c1\u03bd\u03b5\u03b9 null label \u03ba\u03b1\u03b9 \u03b1\u03c0\u03bf\u03ba\u03bb\u03b5\u03af\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf step loss.<\/p>\n<p>\u03a4\u03c1\u03af\u03c4\u03bf\u03bd, \u03c4\u03bf \u03c5\u03c0\u03cc \u03b5\u03ba\u03c0\u03b1\u03af\u03b4\u03b5\u03c5\u03c3\u03b7 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03ce\u03bd\u03b5\u03b9 on-policy \u03c4\u03bf\u03bd \u03c3\u03ba\u03b5\u03bb\u03b5\u03c4\u03cc \u03c4\u03c9\u03bd \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03c9\u03bd. \u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03bf\u03ba\u03c4\u03ce candidates \u03b1\u03bd\u03ac skeleton \u03bc\u03b5 \u03b8\u03b5\u03c1\u03bc\u03bf\u03ba\u03c1\u03b1\u03c3\u03af\u03b1 0,4 \u03ba\u03b1\u03b9 top-p 0,95. \u03a4\u03ad\u03c4\u03b1\u03c1\u03c4\u03bf\u03bd, \u03ba\u03ac\u03b8\u03b5 \u03c5\u03c0\u03bf\u03c8\u03ae\u03c6\u03b9\u03b1 \u03bb\u03cd\u03c3\u03b7 \u03b5\u03ba\u03c4\u03b5\u03bb\u03b5\u03af\u03c4\u03b1\u03b9 \u03c4\u03cc\u03c3\u03bf \u03b1\u03c0\u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c3\u03c4\u03b1 \u03c4\u03bf\u03c0\u03b9\u03ba\u03ac tests \u03cc\u03c3\u03bf \u03ba\u03b1\u03b9 \u03b1\u03c0\u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c3\u03c4\u03bf test suite \u03c4\u03bf\u03c5 dataset. \u0388\u03c4\u03c3\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03c0\u03b1\u03c1\u03ac\u03b4\u03b5\u03b9\u03b3\u03bc\u03b1 \u03b1\u03c0\u03bf\u03ba\u03c4\u03ac \u03ad\u03bd\u03b1 global outcome label \u03ba\u03b1\u03b9 \u03bc\u03b9\u03b1 \u03b1\u03ba\u03bf\u03bb\u03bf\u03c5\u03b8\u03af\u03b1 \u03b1\u03c0\u03cc function-level labels.<\/p>\n<p>\u0397 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae objective \u03c3\u03c5\u03bd\u03b4\u03c5\u03ac\u03b6\u03b5\u03b9 outcome-level KTO \u03bc\u03b5 masked step-level KTO. \u039f\u03b9 \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03bc\u03b5 \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03bf label \u03c3\u03c5\u03bc\u03bc\u03b5\u03c4\u03ad\u03c7\u03bf\u03c5\u03bd \u03c3\u03c4\u03bf \u03c4\u03bf\u03c0\u03b9\u03ba\u03cc loss, \u03b5\u03bd\u03ce \u03cc\u03c3\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b4\u03b5\u03af\u03b3\u03bc\u03b1\u03c4\u03b1 \u03b4\u03b5\u03bd \u03ad\u03c7\u03bf\u03c5\u03bd \u03b1\u03c1\u03ba\u03b5\u03c4\u03ac \u03b5\u03c0\u03bf\u03c0\u03c4\u03b5\u03c5\u03cc\u03bc\u03b5\u03bd\u03b1 steps \u03b5\u03c0\u03b9\u03c3\u03c4\u03c1\u03ad\u03c6\u03bf\u03c5\u03bd \u03c3\u03b5 outcome-only KTO. \u0397 \u03c5\u03bb\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03b1\u03c0\u03b1\u03b9\u03c4\u03b5\u03af \u03c4\u03bf\u03c5\u03bb\u03ac\u03c7\u03b9\u03c3\u03c4\u03bf\u03bd \u03b4\u03cd\u03bf supervised functions \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03b5\u03bd\u03b5\u03c1\u03b3\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b5\u03b9 \u03c4\u03bf step loss, \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03c3\u03b5 \u03bc\u03af\u03b1 \u03bc\u03cc\u03bd\u03bf \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03c4\u03bf \u03c4\u03bf\u03c0\u03b9\u03ba\u03cc \u03c3\u03ae\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ac \u03bc\u03b9\u03b1 \u03b1\u03bd\u03b1\u03ba\u03bb\u03b9\u03bc\u03ac\u03ba\u03c9\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03bf\u03cd.<\/p>\n<h2 id=\"local-global-orthotita-sygkrousi\">\u038c\u03c4\u03b1\u03bd \u03c4\u03bf\u03c0\u03b9\u03ba\u03ae \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c3\u03c5\u03b3\u03ba\u03c1\u03bf\u03cd\u03bf\u03bd\u03c4\u03b1\u03b9<\/h2>\n<p>\u0388\u03bd\u03b1 \u03b1\u03c0\u03cc \u03c4\u03b1 \u03c0\u03b9\u03bf \u03b5\u03bd\u03b4\u03b9\u03b1\u03c6\u03ad\u03c1\u03bf\u03bd\u03c4\u03b1 \u03b5\u03c5\u03c1\u03ae\u03bc\u03b1\u03c4\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u00ab\u03ba\u03b1\u03b8\u03b1\u03c1\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9\u00bb \u03b2\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03ac \u03bf\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c0\u03c4\u03ce\u03c3\u03b5\u03b9\u03c2 \u03cc\u03c0\u03bf\u03c5 \u03c4\u03b1 labels \u03b4\u03b9\u03b1\u03c6\u03c9\u03bd\u03bf\u03cd\u03bd. \u039c\u03b9\u03b1 \u03bb\u03cd\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03b5\u03c1\u03bd\u03ac \u03c4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc test suite \u03b1\u03bb\u03bb\u03ac \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03bc\u03ad\u03c1\u03bf\u03c5\u03c2 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c4\u03c5\u03b3\u03c7\u03ac\u03bd\u03b5\u03b9 \u03c3\u03b5 \u03c0\u03b9\u03bf \u03c3\u03c4\u03bf\u03c7\u03b5\u03c5\u03bc\u03ad\u03bd\u03b1 tests. \u0391\u03c5\u03c4\u03cc \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c0\u03b1\u03c1\u03b1\u03af\u03c4\u03b7\u03c4\u03b1 \u03b8\u03cc\u03c1\u03c5\u03b2\u03bf\u03c2. \u039c\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03b1 \u03b1\u03c1\u03c7\u03b9\u03ba\u03ac tests \u03b4\u03b5\u03bd \u03ba\u03ac\u03bb\u03c5\u03c8\u03b1\u03bd \u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf edge case \u03ae \u03cc\u03c4\u03b9 \u03c4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 \u03c0\u03b1\u03c1\u03ac \u03bc\u03b9\u03b1 \u03c4\u03bf\u03c0\u03b9\u03ba\u03ae \u03b1\u03b4\u03c5\u03bd\u03b1\u03bc\u03af\u03b1.<\/p>\n<p>\u03a3\u03c4\u03b1 datasets \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2 \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03c3\u03c4\u03b7\u03ba\u03b1\u03bd 1.194 \u03c4\u03ad\u03c4\u03bf\u03b9\u03b5\u03c2 \u03c3\u03c5\u03b3\u03ba\u03c1\u03bf\u03cd\u03c3\u03b5\u03b9\u03c2 \u03b3\u03b9\u03b1 \u03c4\u03bf Qwen2.5-Coder-1.5B, 1.397 \u03b3\u03b9\u03b1 \u03c4\u03bf Qwen2.5-Coder-3B \u03ba\u03b1\u03b9 1.428 \u03b3\u03b9\u03b1 \u03c4\u03bf DeepSeek-Coder-6.7B. \u038c\u03c4\u03b1\u03bd \u03bf\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03ad\u03c2 \u03b4\u03b9\u03b1\u03c4\u03ae\u03c1\u03b7\u03c3\u03b1\u03bd \u03b1\u03c5\u03c4\u03ac \u03c4\u03b1 conflict labels, \u03c4\u03bf Qwen 3B \u03bc\u03b5 Step-KTOder \u03ad\u03c6\u03c4\u03b1\u03c3\u03b5 0,128 \u03c3\u03c4\u03bf BigCodeBench Hard \u03ba\u03b1\u03b9 0,127 \u03c3\u03c4\u03bf LiveCodeBench. \u039c\u03b5 masking \u03c4\u03c9\u03bd \u03c3\u03c5\u03b3\u03ba\u03c1\u03bf\u03cd\u03c3\u03b5\u03c9\u03bd \u03bf\u03b9 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b5\u03c2 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03ae\u03c4\u03b1\u03bd 0,115 \u03ba\u03b1\u03b9 0,119.<\/p>\n<p>\u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03ae \u03bf\u03bc\u03ac\u03b4\u03b1, \u03b7 \u03b5\u03c1\u03bc\u03b7\u03bd\u03b5\u03af\u03b1 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c0\u03c1\u03bf\u03c3\u03bf\u03c7\u03ae. \u0394\u03b5\u03bd \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03c3\u03cd\u03b3\u03ba\u03c1\u03bf\u03c5\u03c3\u03b7 \u03bc\u03b5\u03c4\u03b1\u03be\u03cd tests \u03b5\u03af\u03bd\u03b1\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03ae \u03cc\u03c4\u03b9 \u03cc\u03bb\u03b1 \u03c4\u03b1 test suites \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c9\u03c3\u03c4\u03ac. \u0394\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03bc\u03c9\u03c2 \u03cc\u03c4\u03b9 \u03b7 \u03b4\u03b9\u03b1\u03c6\u03c9\u03bd\u03af\u03b1 \u03b1\u03bd\u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 local \u03ba\u03b1\u03b9 global checks \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1 \u03c0\u03c1\u03bf\u03c2 \u03b4\u03b9\u03b5\u03c1\u03b5\u03cd\u03bd\u03b7\u03c3\u03b7. \u0391\u03bd \u03b5\u03be\u03b1\u03c6\u03b1\u03bd\u03b9\u03c3\u03c4\u03b5\u03af \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03ba\u03c4\u03ae\u03c3\u03bf\u03c5\u03bc\u03b5 \u03ad\u03bd\u03b1 \u00ab\u03ba\u03b1\u03b8\u03b1\u03c1\u03cc\u00bb label, \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c7\u03b1\u03b8\u03b5\u03af \u03b1\u03ba\u03c1\u03b9\u03b2\u03ce\u03c2 \u03c4\u03bf \u03c3\u03ae\u03bc\u03b1 \u03c0\u03bf\u03c5 \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03b5\u03b9 \u03ba\u03c1\u03c5\u03c6\u03ad\u03c2 \u03b1\u03c3\u03c4\u03bf\u03c7\u03af\u03b5\u03c2.<\/p>\n<div class=\"td-decision-band\">\n<p class=\"td-decision-kicker\">\u039a\u03b1\u03bd\u03cc\u03bd\u03b1\u03c2 \u03b3\u03b9\u03b1 training data<\/p>\n<p class=\"td-decision-title\">\u039c\u03b7\u03bd \u03b5\u03c0\u03b9\u03bb\u03cd\u03b5\u03c4\u03b5 \u03c4\u03b7 \u03b4\u03b9\u03b1\u03c6\u03c9\u03bd\u03af\u03b1 \u03c0\u03c1\u03b9\u03bd \u03ba\u03b1\u03c4\u03b1\u03bb\u03ac\u03b2\u03b5\u03c4\u03b5 \u03c4\u03b9 \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 \u03ba\u03ac\u03b8\u03b5 test suite<\/p>\n<p>\u039a\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 \u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ac \u03c4\u03bf local result, \u03c4\u03bf end-to-end result \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 \u03c4\u03c9\u03bd tests. \u039c\u03b9\u03b1 \u03c3\u03cd\u03b3\u03ba\u03c1\u03bf\u03c5\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 \u03b5\u03bb\u03bb\u03b9\u03c0\u03ae \u03ba\u03ac\u03bb\u03c5\u03c8\u03b7, \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b7 \u03c5\u03c0\u03cc\u03b8\u03b5\u03c3\u03b7 \u03ae \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03cc bug \u03c0\u03bf\u03c5 \u03ba\u03c1\u03cd\u03b2\u03b5\u03c4\u03b1\u03b9 \u03c0\u03af\u03c3\u03c9 \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc pass.<\/p>\n<\/div>\n<h2 id=\"ti-edeixan-benchmarks\">\u03a4\u03b9 \u03ad\u03b4\u03b5\u03b9\u03be\u03b1\u03bd \u03c4\u03b1 benchmarks<\/h2>\n<p>\u0397 \u03bc\u03ad\u03b8\u03bf\u03b4\u03bf\u03c2 \u03b4\u03bf\u03ba\u03b9\u03bc\u03ac\u03c3\u03c4\u03b7\u03ba\u03b5 \u03c3\u03b5 \u03c4\u03c1\u03b5\u03b9\u03c2 code-specialized \u03b2\u03ac\u03c3\u03b5\u03b9\u03c2: Qwen2.5-Coder-1.5B-Instruct, Qwen2.5-Coder-3B-Instruct \u03ba\u03b1\u03b9 DeepSeek-Coder-6.7B-Instruct. \u039f\u03b9 \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03c3\u03b5\u03b9\u03c2 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 outcome-only KTO \u03ae\u03c4\u03b1\u03bd \u03bc\u03b9\u03ba\u03c1\u03ad\u03c2 \u03ae \u03bc\u03b7\u03b4\u03b5\u03bd\u03b9\u03ba\u03ad\u03c2 \u03c3\u03c4\u03b1 \u03b5\u03c5\u03ba\u03bf\u03bb\u03cc\u03c4\u03b5\u03c1\u03b1 HumanEval \u03ba\u03b1\u03b9 MBPP, \u03b1\u03bb\u03bb\u03ac \u03c0\u03b9\u03bf \u03b5\u03bc\u03c6\u03b1\u03bd\u03b5\u03af\u03c2 \u03c3\u03c4\u03b1 \u03b4\u03c5\u03c3\u03ba\u03bf\u03bb\u03cc\u03c4\u03b5\u03c1\u03b1 BigCodeBench Hard \u03ba\u03b1\u03b9 LiveCodeBench.<\/p>\n<p>\u0393\u03b9\u03b1 \u03c4\u03bf Qwen 1.5B, \u03c4\u03bf BigCodeBench Hard \u03b1\u03c5\u03be\u03ae\u03b8\u03b7\u03ba\u03b5 \u03b1\u03c0\u03cc 0,061 \u03bc\u03b5 KTO \u03c3\u03b5 0,068 \u03bc\u03b5 Step-KTOder, \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ae \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b7 11,5%, \u03b5\u03bd\u03ce \u03c4\u03bf LiveCodeBench \u03b1\u03bd\u03ad\u03b2\u03b7\u03ba\u03b5 \u03b1\u03c0\u03cc 0,052 \u03c3\u03b5 0,056. \u0393\u03b9\u03b1 \u03c4\u03bf Qwen 3B, \u03bf\u03b9 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03b1\u03c5\u03be\u03ae\u03b8\u03b7\u03ba\u03b1\u03bd \u03b1\u03c0\u03cc 0,115 \u03c3\u03b5 0,128 \u03ba\u03b1\u03b9 \u03b1\u03c0\u03cc 0,116 \u03c3\u03b5 0,127 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b1. \u0393\u03b9\u03b1 \u03c4\u03bf DeepSeek 6.7B, \u03c4\u03bf BigCodeBench Hard \u03bc\u03b5\u03c4\u03b1\u03ba\u03b9\u03bd\u03ae\u03b8\u03b7\u03ba\u03b5 \u03b1\u03c0\u03cc 0,122 \u03c3\u03b5 0,128 \u03ba\u03b1\u03b9 \u03c4\u03bf LiveCodeBench \u03b1\u03c0\u03cc 0,127 \u03c3\u03b5 0,131.<\/p>\n<p>\u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b1\u03bd\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03cc\u03c4\u03b9, \u03c3\u03b5 \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03b1 \u03bc\u03b7 post-trained base models, \u03c4\u03bf Step-KTOder \u03b2\u03b5\u03bb\u03c4\u03af\u03c9\u03c3\u03b5 \u03ad\u03c9\u03c2 26,7% \u03c4\u03bf BigCodeBench Hard \u03ba\u03b1\u03b9 \u03ad\u03c9\u03c2 27,0% \u03c4\u03bf LiveCodeBench. \u0391\u03c5\u03c4\u03ad\u03c2 \u03bf\u03b9 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c3\u03c5\u03b3\u03c7\u03ad\u03bf\u03bd\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03b7 \u03c3\u03cd\u03b3\u03ba\u03c1\u03b9\u03c3\u03b7 \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 KTO: \u03bc\u03b5\u03c4\u03c1\u03bf\u03cd\u03bd \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03cc baseline. \u03a4\u03bf \u03bf\u03c5\u03c3\u03b9\u03b1\u03c3\u03c4\u03b9\u03ba\u03cc pattern \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03bf \u03bb\u03b5\u03c0\u03c4\u03bf\u03bc\u03b5\u03c1\u03ad\u03c3\u03c4\u03b5\u03c1\u03bf execution-grounded feedback \u03b1\u03c0\u03bf\u03b4\u03af\u03b4\u03b5\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf \u03b5\u03ba\u03b5\u03af \u03cc\u03c0\u03bf\u03c5 \u03c4\u03bf task \u03b5\u03af\u03bd\u03b1\u03b9 \u03b4\u03c5\u03c3\u03ba\u03bf\u03bb\u03cc\u03c4\u03b5\u03c1\u03bf \u03ba\u03b1\u03b9 \u03b7 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1 \u03b1\u03c0\u03cc \u03bc\u03cc\u03bd\u03b7 \u03c4\u03b7\u03c2 \u03b4\u03af\u03bd\u03b5\u03b9 \u03c6\u03c4\u03c9\u03c7\u03cc\u03c4\u03b5\u03c1\u03bf \u03c3\u03ae\u03bc\u03b1.<\/p>\n<p>\u0397 \u03b5\u03c0\u03b9\u03c6\u03cd\u03bb\u03b1\u03be\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03be\u03af\u03c3\u03bf\u03c5 \u03c3\u03b7\u03bc\u03b1\u03bd\u03c4\u03b9\u03ba\u03ae \u03bc\u03b5 \u03c4\u03bf \u03ba\u03ad\u03c1\u03b4\u03bf\u03c2. \u03a4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 pass rates \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03c9\u03bd \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03c9\u03bd \u03ba\u03b1\u03b9 \u03c0\u03c1\u03c9\u03c4\u03bf\u03ba\u03cc\u03bb\u03bb\u03c9\u03bd, \u03cc\u03c7\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae \u03c0\u03c1\u03cc\u03b2\u03bb\u03b5\u03c8\u03b7 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 repository \u03ae coding agent. \u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03c0\u03cc\u03c4\u03b5 \u03b1\u03be\u03af\u03b6\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03bf\u03ba\u03b9\u03bc\u03b1\u03c3\u03c4\u03b5\u03af \u03b7 \u03b9\u03b4\u03ad\u03b1, \u03cc\u03c7\u03b9 \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03b8\u03b1 \u03b4\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03af\u03b4\u03b9\u03b5\u03c2 \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03b9\u03b1\u03af\u03b5\u03c2 \u03b1\u03bb\u03bb\u03b1\u03b3\u03ad\u03c2.<\/p>\n<h2 id=\"function-labels-entopizoun-vlaves\">\u03a4\u03b1 function labels \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03bf\u03c5\u03bd \u03c4\u03b9\u03c2 \u03b2\u03bb\u03ac\u03b2\u03b5\u03c2;<\/h2>\n<p>\u039f\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03ad\u03ba\u03b1\u03bd\u03b1\u03bd \u03bc\u03b9\u03b1 \u03c3\u03c4\u03bf\u03c7\u03b5\u03c5\u03bc\u03ad\u03bd\u03b7 repair intervention. \u03a3\u03b5 \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b5\u03c2 \u03bb\u03cd\u03c3\u03b5\u03b9\u03c2 \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03c4\u03ad\u03c3\u03c4\u03b7\u03c3\u03b1\u03bd \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03bc\u03b5 \u03c4\u03b9\u03c2 \u03c3\u03c9\u03c3\u03c4\u03ad\u03c2 reference implementations \u03ba\u03b1\u03b9 \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b1\u03bd \u03c0\u03cc\u03c3\u03bf \u03c3\u03c5\u03c7\u03bd\u03ac \u03b5\u03c0\u03b9\u03c3\u03ba\u03b5\u03c5\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bf\u03bb\u03cc\u03ba\u03bb\u03b7\u03c1\u03bf \u03c4\u03bf \u03c0\u03c1\u03cc\u03b3\u03c1\u03b1\u03bc\u03bc\u03b1. \u038c\u03c4\u03b1\u03bd \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03c4\u03ad\u03c3\u03c4\u03b7\u03c3\u03b1\u03bd locally negative functions, \u03b7 \u03bb\u03cd\u03c3\u03b7 \u03b4\u03b9\u03bf\u03c1\u03b8\u03ce\u03b8\u03b7\u03ba\u03b5 \u03c3\u03c4\u03bf 74,9% \u03c4\u03c9\u03bd \u03c0\u03b5\u03c1\u03b9\u03c0\u03c4\u03ce\u03c3\u03b5\u03c9\u03bd. \u039c\u03b5 \u03c4\u03c5\u03c7\u03b1\u03af\u03b5\u03c2 \u03c0\u03b1\u03c1\u03b1\u03c4\u03b7\u03c1\u03b7\u03bc\u03ad\u03bd\u03b5\u03c2 \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c4\u03bf repair rate \u03ae\u03c4\u03b1\u03bd 32,9%, \u03b5\u03bd\u03ce \u03bc\u03b5 locally positive functions \u03bc\u03cc\u03bb\u03b9\u03c2 1,2%.<\/p>\n<p>\u03a4\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b1\u03c0\u03bf\u03b4\u03b5\u03b9\u03ba\u03bd\u03cd\u03b5\u03b9 \u03cc\u03c4\u03b9 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03c1\u03bd\u03b7\u03c4\u03b9\u03ba\u03cc label \u03b1\u03c0\u03bf\u03bc\u03bf\u03bd\u03ce\u03bd\u03b5\u03b9 \u03c4\u03ad\u03bb\u03b5\u03b9\u03b1 \u03c4\u03b7 \u03bc\u03bf\u03bd\u03b1\u03b4\u03b9\u03ba\u03ae \u03b1\u03b9\u03c4\u03af\u03b1. \u0394\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03cc\u03bc\u03c9\u03c2 \u03cc\u03c4\u03b9 \u03c4\u03b1 function-level tests \u03c3\u03c5\u03b3\u03ba\u03b5\u03bd\u03c4\u03c1\u03ce\u03bd\u03bf\u03c5\u03bd \u03c4\u03bf \u03c3\u03ae\u03bc\u03b1 \u03c3\u03b5 \u03c4\u03bc\u03ae\u03bc\u03b1\u03c4\u03b1 \u03c0\u03bf\u03c5 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03bf\u03bb\u03cd \u03c0\u03b9\u03bf \u03c0\u03b9\u03b8\u03b1\u03bd\u03cc \u03bd\u03b1 \u03b5\u03c5\u03b8\u03cd\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b3\u03b9\u03b1 \u03c4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc failure. \u0397 \u03c0\u03c1\u03cc\u03c3\u03b8\u03b5\u03c4\u03b7 mutation-sensitivity \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03b5\u03bd\u03b9\u03c3\u03c7\u03cd\u03b5\u03b9 \u03b1\u03c5\u03c4\u03ae \u03c4\u03b7\u03bd \u03b1\u03bd\u03ac\u03b3\u03bd\u03c9\u03c3\u03b7: \u03c4\u03b1 tests \u03b1\u03c0\u03ad\u03c1\u03c1\u03b9\u03c8\u03b1\u03bd 25.880 \u03b1\u03c0\u03cc 27.958 \u03ad\u03b3\u03ba\u03c5\u03c1\u03b1 \u03c3\u03c5\u03bd\u03c4\u03b1\u03ba\u03c4\u03b9\u03ba\u03ac mutants, mutation kill rate 92,6%.<\/p>\n<div class=\"td-chart td-chart--metrics\">\n<div class=\"td-chart-head\">\n<p class=\"td-chart-title\">\u03a4\u03ad\u03c3\u03c3\u03b5\u03c1\u03b9\u03c2 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b5\u03b9\u03c2 \u03b3\u03b9\u03b1 \u03c4\u03b7\u03bd \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03bf\u03c5 execution feedback<\/p>\n<p class=\"td-chart-subtitle\">\u039f\u03b9 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03b1\u03c6\u03bf\u03c1\u03bf\u03cd\u03bd \u03c4\u03b9\u03c2 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b5\u03c2 \u03c0\u03b1\u03c1\u03b5\u03bc\u03b2\u03ac\u03c3\u03b5\u03b9\u03c2 \u03ba\u03b1\u03b9 audits \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2, \u03cc\u03c7\u03b9 \u03b3\u03b5\u03bd\u03b9\u03ba\u03ae \u03b1\u03ba\u03c1\u03af\u03b2\u03b5\u03b9\u03b1 debugging.<\/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\">74,9%<\/span><span class=\"td-metric-label\">repair \u03bc\u03b5 negative function<\/span><span class=\"td-metric-note\">\u0391\u03bd\u03c4\u03b9\u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7 \u03c4\u03bf\u03c0\u03b9\u03ba\u03ac \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03c9\u03bd \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03c9\u03bd<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">32,9%<\/span><span class=\"td-metric-label\">repair \u03bc\u03b5 \u03c4\u03c5\u03c7\u03b1\u03af\u03b1 function<\/span><span class=\"td-metric-note\">Control \u03c0\u03ac\u03bd\u03c9 \u03c3\u03c4\u03b9\u03c2 \u03af\u03b4\u03b9\u03b5\u03c2 failing \u03bb\u03cd\u03c3\u03b5\u03b9\u03c2<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">1,2%<\/span><span class=\"td-metric-label\">repair \u03bc\u03b5 positive function<\/span><span class=\"td-metric-note\">\u0388\u03bb\u03b5\u03b3\u03c7\u03bf\u03c2 \u03cc\u03c4\u03b9 \u03c4\u03bf \u03b8\u03b5\u03c4\u03b9\u03ba\u03cc label \u03b4\u03b5\u03bd \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03c4\u03b7 \u03b2\u03bb\u03ac\u03b2\u03b7<\/span><\/div>\n<div class=\"td-metric-card\"><span class=\"td-metric-value\">92,6%<\/span><span class=\"td-metric-label\">mutation kill rate<\/span><span class=\"td-metric-note\">25.880 \u03b1\u03c0\u03cc 27.958 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03b1 mutants<\/span><\/div>\n<\/div>\n<\/div>\n<\/div>\n<h2 id=\"execution-vs-llm-judge\">\u0393\u03b9\u03b1\u03c4\u03af \u03b7 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03be\u03b5\u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf\u03bd LLM judge<\/h2>\n<p>\u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03c4\u03ad\u03c3\u03c4\u03b7\u03c3\u03b5 \u03c4\u03b1 execution-based step labels \u03bc\u03b5 \u03ba\u03c1\u03af\u03c3\u03b5\u03b9\u03c2 \u03b1\u03c0\u03cc GPT-5.4 mini. \u0397 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03c3\u03c5\u03bc\u03c6\u03c9\u03bd\u03af\u03b1 \u03ae\u03c4\u03b1\u03bd 73,2%, \u03b1\u03bb\u03bb\u03ac \u03bf \u03bc\u03ad\u03c3\u03bf\u03c2 \u03cc\u03c1\u03bf\u03c2 \u03ad\u03ba\u03c1\u03c5\u03b2\u03b5 \u03ad\u03bd\u03c4\u03bf\u03bd\u03b7 \u03b1\u03c3\u03c5\u03bc\u03bc\u03b5\u03c4\u03c1\u03af\u03b1: \u03bf LLM judge \u03c3\u03c5\u03bc\u03c6\u03ce\u03bd\u03b7\u03c3\u03b5 \u03bc\u03b5 \u03c4\u03b7\u03bd \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03c3\u03c4\u03bf 94,1% \u03c4\u03c9\u03bd failing functions, \u03b1\u03bb\u03bb\u03ac \u03bc\u03cc\u03bd\u03bf \u03c3\u03c4\u03bf 52,2% \u03c4\u03c9\u03bd passing functions. \u039c\u03b5 \u03ac\u03bb\u03bb\u03b1 \u03bb\u03cc\u03b3\u03b9\u03b1, \u03b5\u03af\u03c7\u03b5 \u03c4\u03ac\u03c3\u03b7 \u03bd\u03b1 \u03c0\u03c1\u03bf\u03b2\u03bb\u03ad\u03c0\u03b5\u03b9 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03af\u03b5\u03c2 \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03ba\u03b1\u03b8\u03b1\u03c1\u03ac \u03b8\u03b5\u03c4\u03b9\u03ba\u03ac labels \u03c3\u03b5 \u03b1\u03c1\u03bd\u03b7\u03c4\u03b9\u03ba\u03ac.<\/p>\n<p>\u03a4\u03bf \u03c0\u03c1\u03cc\u03b2\u03bb\u03b7\u03bc\u03b1 \u03b4\u03b5\u03bd \u03ad\u03bc\u03b5\u03b9\u03bd\u03b5 \u03c3\u03c4\u03bf annotation. \u039c\u03b5 execution labels, \u03c4\u03bf Qwen 3B Step-KTOder \u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 0,128 \u03c3\u03c4\u03bf BigCodeBench Hard \u03ba\u03b1\u03b9 0,127 \u03c3\u03c4\u03bf LiveCodeBench. \u039c\u03b5 LLM-as-a-judge labels \u03bf\u03b9 \u03c4\u03b9\u03bc\u03ad\u03c2 \u03ad\u03c0\u03b5\u03c3\u03b1\u03bd \u03c3\u03b5 0,095 \u03ba\u03b1\u03b9 0,112, \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ae \u03b1\u03bb\u03bb\u03b1\u03b3\u03ae -25,8% \u03ba\u03b1\u03b9 -11,8% \u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c4\u03b7\u03c2 execution \u03b5\u03ba\u03b4\u03bf\u03c7\u03ae\u03c2. \u03a3\u03c4\u03bf MBPP Plus \u03c5\u03c0\u03ae\u03c1\u03be\u03b5 \u03bc\u03b9\u03ba\u03c1\u03ae \u03ac\u03bd\u03bf\u03b4\u03bf\u03c2, \u03b1\u03c0\u03cc 0,630 \u03c3\u03b5 0,635, \u03b1\u03bb\u03bb\u03ac \u03b7 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03ae \u03b5\u03b9\u03ba\u03cc\u03bd\u03b1 \u03c3\u03c4\u03b1 \u03c0\u03b9\u03bf \u03b1\u03c0\u03b1\u03b9\u03c4\u03b7\u03c4\u03b9\u03ba\u03ac benchmarks \u03b5\u03c0\u03b9\u03b4\u03b5\u03b9\u03bd\u03ce\u03b8\u03b7\u03ba\u03b5.<\/p>\n<p>\u03a4\u03bf \u03c3\u03c5\u03bc\u03c0\u03ad\u03c1\u03b1\u03c3\u03bc\u03b1 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03ad\u03bd\u03b1\u03c2 LLM judge \u03b4\u03b5\u03bd \u03ad\u03c7\u03b5\u03b9 \u03b8\u03ad\u03c3\u03b7 \u03c3\u03b5 code workflows. \u039c\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b2\u03bf\u03b7\u03b8\u03ae\u03c3\u03b5\u03b9 \u03c3\u03b5 \u03c0\u03c1\u03cc\u03b8\u03b5\u03c3\u03b7, \u03b1\u03bd\u03b1\u03b3\u03bd\u03c9\u03c3\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1, \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7 \u03ae \u03c3\u03b5 \u03c0\u03b5\u03c1\u03b9\u03c0\u03c4\u03ce\u03c3\u03b5\u03b9\u03c2 \u03cc\u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03c5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 \u03b5\u03ba\u03c4\u03b5\u03bb\u03ad\u03c3\u03b9\u03bc\u03bf oracle. \u038c\u03c4\u03b1\u03bd \u03cc\u03bc\u03c9\u03c2 \u03c4\u03bf \u03b6\u03b7\u03c4\u03bf\u03cd\u03bc\u03b5\u03bd\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03b9\u03ba\u03ae \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03c5\u03c0\u03ac\u03c1\u03c7\u03bf\u03c5\u03bd \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03b1 tests, \u03b7 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03c0\u03b1\u03c1\u03ad\u03c7\u03b5\u03b9 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf ground truth. \u0397 \u03b3\u03bb\u03c9\u03c3\u03c3\u03b9\u03ba\u03ae \u03ba\u03c1\u03af\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03c9\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae, \u03cc\u03c7\u03b9 \u03b1\u03c5\u03c4\u03cc\u03bc\u03b1\u03c4\u03b7 \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03c4\u03ac\u03c3\u03c4\u03b1\u03c3\u03b7.<\/p>\n<div class=\"td-note\">\n<p class=\"td-note-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03c4\u03bf 73,2% \u03b4\u03b5\u03bd \u03b1\u03c1\u03ba\u03b5\u03af<\/p>\n<p>\u0388\u03bd\u03b1 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc agreement score \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c6\u03b1\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b9\u03ba\u03b1\u03bd\u03bf\u03c0\u03bf\u03b9\u03b7\u03c4\u03b9\u03ba\u03cc \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u2019 \u03cc\u03bb\u03b1 \u03b1\u03c5\u03c4\u03ac \u03bd\u03b1 \u03c0\u03b1\u03c1\u03ac\u03b3\u03b5\u03b9 \u03ba\u03b1\u03ba\u03cc training signal. \u0395\u03b4\u03ce \u03c4\u03bf \u03ba\u03c1\u03af\u03c3\u03b9\u03bc\u03bf \u03c0\u03c1\u03cc\u03b2\u03bb\u03b7\u03bc\u03b1 \u03ae\u03c4\u03b1\u03bd \u03c4\u03bf \u03b5\u03af\u03b4\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03bb\u03ac\u03b8\u03bf\u03c5\u03c2: \u03bf judge \u03b1\u03bb\u03bb\u03bf\u03af\u03c9\u03bd\u03b5 \u03c3\u03c5\u03c3\u03c4\u03b7\u03bc\u03b1\u03c4\u03b9\u03ba\u03ac \u03c4\u03b1 \u03b8\u03b5\u03c4\u03b9\u03ba\u03ac labels \u03c0\u03bf\u03c5 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c4\u03bf step-level objective \u03b3\u03b9\u03b1 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03c3\u03c9\u03c3\u03c4\u03ad\u03c2 \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2.<\/p>\n<\/div>\n<h2 id=\"poiotita-unit-tests\">\u03a0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 unit tests: \u03c4\u03bf \u03b1\u03b8\u03ad\u03b1\u03c4\u03bf \u03b8\u03b5\u03bc\u03ad\u03bb\u03b9\u03bf<\/h2>\n<p>\u0397 \u03b5\u03c0\u03b9\u03c4\u03c5\u03c7\u03af\u03b1 \u03c4\u03bf\u03c5 \u03c0\u03bb\u03b1\u03b9\u03c3\u03af\u03bf\u03c5 \u03b5\u03be\u03b1\u03c1\u03c4\u03ac\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03c4\u03bf \u03b1\u03bd \u03c4\u03b1 unit tests \u03b5\u03af\u03bd\u03b1\u03b9 \u03ad\u03b3\u03ba\u03c5\u03c1\u03b1 \u03ba\u03b1\u03b9 \u03b4\u03b9\u03b1\u03ba\u03c1\u03b9\u03c4\u03b9\u03ba\u03ac. \u03a3\u03c4\u03b7 \u03b4\u03b9\u03b1\u03b4\u03b9\u03ba\u03b1\u03c3\u03af\u03b1 \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2, \u03c4\u03bf 82,0% \u03c4\u03c9\u03bd testable functions \u03b1\u03c0\u03cc TACO \u03ba\u03b1\u03b9 \u03c4\u03bf 84,5% \u03b1\u03c0\u03cc APPS \u03b1\u03c0\u03ad\u03ba\u03c4\u03b7\u03c3\u03b1\u03bd \u03c4\u03bf\u03c5\u03bb\u03ac\u03c7\u03b9\u03c3\u03c4\u03bf\u03bd \u03ad\u03bd\u03b1 \u03ad\u03b3\u03ba\u03c5\u03c1\u03bf test. \u039f \u03bc\u03ad\u03c3\u03bf\u03c2 \u03b1\u03c1\u03b9\u03b8\u03bc\u03cc\u03c2 valid tests \u03b1\u03bd\u03ac \u03ba\u03b1\u03bb\u03c5\u03bc\u03bc\u03ad\u03bd\u03b7 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03ae\u03c4\u03b1\u03bd 5,46 \u03ba\u03b1\u03b9 5,59 \u03b1\u03bd\u03c4\u03af\u03c3\u03c4\u03bf\u03b9\u03c7\u03b1. \u0397 \u03bc\u03ad\u03c3\u03b7 \u03b4\u03b9\u03b1\u03b8\u03b5\u03c3\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 tests \u03b1\u03bd\u03ac \u03c0\u03c1\u03cc\u03b2\u03bb\u03b7\u03bc\u03b1 \u03ae\u03c4\u03b1\u03bd 83,4% \u03b3\u03b9\u03b1 TACO \u03ba\u03b1\u03b9 85,4% \u03b3\u03b9\u03b1 APPS.<\/p>\n<p>\u039f\u03b9 \u03b1\u03c1\u03b9\u03b8\u03bc\u03bf\u03af \u03b4\u03b5\u03af\u03c7\u03bd\u03bf\u03c5\u03bd \u03cc\u03c4\u03b9 \u03b7 \u03b1\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03ba\u03b1\u03bb\u03cd\u03c8\u03b5\u03b9 \u03bc\u03b5\u03b3\u03ac\u03bb\u03bf \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03bf\u03c5 \u03c5\u03bb\u03b9\u03ba\u03bf\u03cd, \u03cc\u03c7\u03b9 \u03cc\u03c4\u03b9 \u03b5\u03be\u03b1\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03bd\u03ac\u03b3\u03ba\u03b7 \u03b5\u03bb\u03ad\u03b3\u03c7\u03bf\u03c5. \u039f\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03c6\u03b9\u03bb\u03c4\u03c1\u03ac\u03c1\u03bf\u03c5\u03bd tests \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03ba\u03ac\u03bd\u03bf\u03c5\u03bd parse, \u03b4\u03b5\u03bd \u03b5\u03ba\u03c4\u03b5\u03bb\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9, \u03b4\u03b5\u03bd \u03ba\u03b1\u03bb\u03bf\u03cd\u03bd \u03c4\u03bf\u03bd \u03c3\u03c4\u03cc\u03c7\u03bf \u03ae \u03c0\u03b5\u03c1\u03b9\u03ad\u03c7\u03bf\u03c5\u03bd \u03c4\u03b5\u03c4\u03c1\u03b9\u03bc\u03bc\u03ad\u03bd\u03b1 assertions. \u0395\u03c0\u03b9\u03c0\u03bb\u03ad\u03bf\u03bd, \u03b7 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b1\u03bd\u03b1\u03b3\u03bd\u03c9\u03c1\u03af\u03b6\u03b5\u03b9 \u03c1\u03b7\u03c4\u03ac \u03cc\u03c4\u03b9 \u03c4\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03cc\u03bc\u03b5\u03bd\u03b1 tests \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03bb\u03bb\u03b9\u03c0\u03ae. \u039c\u03b9\u03b1 \u03bb\u03ac\u03b8\u03bf\u03c2 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03c0\u03b5\u03c1\u03ac\u03c3\u03b5\u03b9 \u03ad\u03bd\u03b1 \u03b1\u03b4\u03cd\u03bd\u03b1\u03bc\u03bf local suite, \u03b5\u03bd\u03ce \u03bc\u03b9\u03b1 \u03c3\u03c9\u03c3\u03c4\u03ae \u03c5\u03bb\u03bf\u03c0\u03bf\u03af\u03b7\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c4\u03cd\u03c7\u03b5\u03b9 \u03b1\u03bd \u03c4\u03bf test \u03b5\u03bd\u03c3\u03c9\u03bc\u03b1\u03c4\u03ce\u03bd\u03b5\u03b9 \u03bb\u03b1\u03bd\u03b8\u03b1\u03c3\u03bc\u03ad\u03bd\u03b7 \u03c5\u03c0\u03cc\u03b8\u03b5\u03c3\u03b7.<\/p>\n<p>\u03a3\u03c4\u03b7\u03bd \u03c0\u03c1\u03ac\u03be\u03b7, \u03c4\u03bf test-generation pipeline \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b4\u03b9\u03ba\u03ac \u03c4\u03bf\u03c5 metrics: reference pass, mutation sensitivity, \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03ba\u03bb\u03ae\u03c3\u03b7 \u03c4\u03bf\u03c5 \u03c3\u03c4\u03cc\u03c7\u03bf\u03c5, \u03ba\u03ac\u03bb\u03c5\u03c8\u03b7 edge cases \u03ba\u03b1\u03b9 \u03c0\u03bf\u03c3\u03bf\u03c3\u03c4\u03cc \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03c9\u03bd \u03c7\u03c9\u03c1\u03af\u03c2 \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03bf oracle. \u03a4\u03bf Step-KTOder \u03b5\u03c6\u03b1\u03c1\u03bc\u03cc\u03b6\u03b5\u03b9 reliability masking \u03b3\u03b9\u03b1 steps \u03c7\u03c9\u03c1\u03af\u03c2 \u03ad\u03b3\u03ba\u03c5\u03c1\u03bf label, \u03bc\u03b9\u03b1 \u03c3\u03c5\u03bd\u03b5\u03c4\u03ae \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03c0\u03bf\u03c5 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03c1\u03bf\u03c4\u03b9\u03bc\u03cc\u03c4\u03b5\u03c1\u03b7 \u03b1\u03c0\u03cc \u03c4\u03bf \u03bd\u03b1 \u03b2\u03b1\u03c6\u03c4\u03b9\u03c3\u03c4\u03b5\u03af \u03b7 \u03b1\u03b2\u03b5\u03b2\u03b1\u03b9\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c9\u03c2 \u03b1\u03c1\u03bd\u03b7\u03c4\u03b9\u03ba\u03cc \u03c0\u03b1\u03c1\u03ac\u03b4\u03b5\u03b9\u03b3\u03bc\u03b1.<\/p>\n<h2 id=\"coding-agents-production\">\u03a4\u03b9 \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03b3\u03b9\u03b1 coding agents \u03ba\u03b1\u03b9 production workflows<\/h2>\n<p>\u0397 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03b1\u03c6\u03bf\u03c1\u03ac offline preference optimization \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03ad\u03c4\u03bf\u03b9\u03bc\u03bf \u03bc\u03b7\u03c7\u03b1\u03bd\u03b9\u03c3\u03bc\u03cc \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03ae\u03c2 \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7\u03c2 agents. \u03a0\u03b1\u03c1\u2019 \u03cc\u03bb\u03b1 \u03b1\u03c5\u03c4\u03ac, \u03c0\u03c1\u03bf\u03c4\u03b5\u03af\u03bd\u03b5\u03b9 \u03bc\u03b9\u03b1 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03b1\u03c1\u03c7\u03b9\u03c4\u03b5\u03ba\u03c4\u03bf\u03bd\u03b9\u03ba\u03ae \u03b1\u03c1\u03c7\u03ae: \u03cc\u03c4\u03b1\u03bd \u03ad\u03bd\u03b1 \u03c3\u03cd\u03bd\u03b8\u03b5\u03c4\u03bf output \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c3\u03c5\u03bd\u03c4\u03b5\u03b8\u03b5\u03af \u03c3\u03b5 \u03b5\u03ba\u03c4\u03b5\u03bb\u03ad\u03c3\u03b9\u03bc\u03b5\u03c2 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2, \u03c4\u03bf feedback \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03b5\u03af \u03b1\u03c5\u03c4\u03ae \u03c4\u03b7 \u03b4\u03bf\u03bc\u03ae. \u03a3\u03b5 \u03ad\u03bd\u03b1\u03bd coding agent, \u03b1\u03bd\u03c4\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03cd\u03b5\u03c4\u03b1\u03b9 \u03bc\u03cc\u03bd\u03bf \u03c4\u03bf \u03b1\u03bd \u03c0\u03ad\u03c1\u03b1\u03c3\u03b5 \u03c4\u03bf \u03c4\u03b5\u03bb\u03b9\u03ba\u03cc CI, \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03bd\u03b1 \u03ba\u03b1\u03c4\u03b1\u03b3\u03c1\u03ac\u03c6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c4\u03b1 \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1 \u03b1\u03bd\u03ac function, module \u03ae \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03bf contract.<\/p>\n<p>\u0391\u03c5\u03c4\u03cc \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03c0\u03b1\u03c1\u03b1\u03c4\u03b7\u03c1\u03b7\u03c3\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1. \u039c\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b4\u03b9\u03b1\u03ba\u03c1\u03af\u03bd\u03b5\u03b9 \u03b1\u03bd \u03c4\u03bf \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf \u03b1\u03c0\u03bf\u03c4\u03c5\u03b3\u03c7\u03ac\u03bd\u03b5\u03b9 \u03c3\u03b5 parsing, business rules, data transformations \u03ae integration boundaries. \u039c\u03c0\u03bf\u03c1\u03b5\u03af \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03bd\u03b1 \u03b4\u03b7\u03bc\u03b9\u03bf\u03c5\u03c1\u03b3\u03ae\u03c3\u03b5\u03b9 \u03c0\u03b9\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03b1 training examples, \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03bf\u03b9 \u03c3\u03c9\u03c3\u03c4\u03ad\u03c2 \u03c5\u03c0\u03bf\u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2 \u03b4\u03b5\u03bd \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c1\u03c1\u03b9\u03c6\u03b8\u03bf\u03cd\u03bd \u03bc\u03b1\u03b6\u03af \u03bc\u03b5 \u03c4\u03bf \u03bb\u03ac\u03b8\u03bf\u03c2 \u03c3\u03cd\u03bd\u03bf\u03bb\u03bf. \u0397 \u03af\u03b4\u03b9\u03b1 \u03b1\u03bd\u03ac\u03b3\u03ba\u03b7 \u03b3\u03b9\u03b1 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7 \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03cc\u03c4\u03b1\u03bd \u03b7 <a href=\"https:\/\/twodots.gr\/documentation-ai-agents-metrisimo-bug\/\">documentation \u03c4\u03c9\u03bd AI agents \u03bc\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c0\u03b5\u03c4\u03b1\u03b9 \u03b1\u03c0\u03cc \u03ba\u03b5\u03bd\u03cc \u03b3\u03bd\u03ce\u03c3\u03b7\u03c2 \u03c3\u03b5 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03bf bug<\/a>.<\/p>\n<p>\u03a5\u03c0\u03ac\u03c1\u03c7\u03b5\u03b9 \u03ba\u03b1\u03b9 governance \u03cc\u03c6\u03b5\u03bb\u03bf\u03c2. \u03a4\u03b1 \u03b5\u03ba\u03c4\u03b5\u03bb\u03ad\u03c3\u03b9\u03bc\u03b1 checks \u03b1\u03c6\u03ae\u03bd\u03bf\u03c5\u03bd audit trail: \u03c0\u03bf\u03b9\u03bf input \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03b8\u03b7\u03ba\u03b5, \u03c0\u03bf\u03b9\u03bf assertion \u03b1\u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 \u03ba\u03b1\u03b9 \u03c0\u03bf\u03b9\u03b1 \u03ad\u03ba\u03b4\u03bf\u03c3\u03b7 \u03c4\u03bf\u03c5 candidate \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03ae\u03b8\u03b7\u03ba\u03b5. \u0391\u03c5\u03c4\u03cc \u03b4\u03b5\u03bd \u03b5\u03b3\u03b3\u03c5\u03ac\u03c4\u03b1\u03b9 \u03c0\u03bb\u03ae\u03c1\u03b7 \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1, \u03b1\u03bb\u03bb\u03ac \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b9\u03bf \u03b5\u03c0\u03b9\u03b8\u03b5\u03c9\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03b1\u03c0\u03cc \u03bc\u03b9\u03b1 \u03b1\u03b4\u03b9\u03b1\u03c6\u03b1\u03bd\u03ae \u03b2\u03b1\u03b8\u03bc\u03bf\u03bb\u03bf\u03b3\u03af\u03b1 \u03b5\u03bd\u03cc\u03c2 \u03b4\u03b5\u03cd\u03c4\u03b5\u03c1\u03bf\u03c5 LLM. \u03a3\u03b5 \u03c0\u03b5\u03c1\u03b9\u03b2\u03ac\u03bb\u03bb\u03bf\u03bd\u03c4\u03b1 \u03bc\u03b5 \u03c0\u03bf\u03bb\u03bb\u03bf\u03cd\u03c2 agents, \u03b7 \u03af\u03b4\u03b9\u03b1 \u03b1\u03c1\u03c7\u03ae \u03c3\u03c5\u03bc\u03c0\u03bb\u03b7\u03c1\u03ce\u03bd\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b1\u03bd\u03ac\u03b3\u03ba\u03b7 \u03b3\u03b9\u03b1 <a href=\"https:\/\/twodots.gr\/agentroom-ai-coding-agents-syntoni-parallilismos\/\">\u03c3\u03c5\u03bd\u03c4\u03bf\u03bd\u03b9\u03c3\u03bc\u03cc \u03c4\u03c9\u03bd AI coding agents<\/a>: \u03c4\u03bf handoff \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03bc\u03b5\u03c4\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1, \u03b1\u03bb\u03bb\u03ac \u03ba\u03b1\u03b9 \u03c4\u03b5\u03ba\u03bc\u03ae\u03c1\u03b9\u03b1 \u03b1\u03bd\u03ac \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03b7 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b1.<\/p>\n<p>\u0397 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03ae \u03bc\u03b5\u03c4\u03b1\u03c6\u03bf\u03c1\u03ac \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bd\u03b1 \u03c0\u03b1\u03c1\u03b1\u03bc\u03b5\u03af\u03bd\u03b5\u03b9 \u03b1\u03bd\u03b1\u03bb\u03bf\u03b3\u03af\u03b1 \u03c3\u03c7\u03b5\u03b4\u03b9\u03b1\u03c3\u03bc\u03bf\u03cd. \u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b4\u03b5\u03bd \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b5 production repositories, \u03bf\u03c1\u03b3\u03b1\u03bd\u03c9\u03c4\u03b9\u03ba\u03ac workflows \u03ae SLA. \u039c\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03cc\u03bc\u03c9\u03c2 \u03bd\u03b1 \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03b5\u03b9 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03c3\u03c7\u03ae\u03bc\u03b1 \u03b3\u03b9\u03b1 pilot: coarse outcome, local executable contracts, \u03b4\u03b9\u03b1\u03c4\u03ae\u03c1\u03b7\u03c3\u03b7 conflicts \u03ba\u03b1\u03b9 \u03be\u03b5\u03c7\u03c9\u03c1\u03b9\u03c3\u03c4\u03ae \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7 label quality. \u0391\u03c5\u03c4\u03cc \u03c3\u03c5\u03bd\u03b4\u03ad\u03b5\u03c4\u03b1\u03b9 \u03ac\u03bc\u03b5\u03c3\u03b1 \u03bc\u03b5 \u03c4\u03b7\u03bd <a href=\"https:\/\/twodots.gr\/sdad-prodiagrafi-ai-anaptyxi-logismikou\/\">\u03c0\u03c1\u03bf\u03b4\u03b9\u03b1\u03b3\u03c1\u03b1\u03c6\u03ae \u03c9\u03c2 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03cc \u03c3\u03cd\u03c3\u03c4\u03b7\u03bc\u03b1 \u03c4\u03b7\u03c2 AI \u03b1\u03bd\u03ac\u03c0\u03c4\u03c5\u03be\u03b7\u03c2 \u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03b9\u03ba\u03bf\u03cd<\/a>, \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03c4\u03b1 contracts \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c1\u03ba\u03b5\u03c4\u03ac \u03c3\u03b1\u03c6\u03ae \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03b3\u03af\u03bd\u03bf\u03bd\u03c4\u03b1\u03b9 tests.<\/p>\n<h2 id=\"oria-step-ktoder\">\u03a4\u03b1 \u03cc\u03c1\u03b9\u03b1 \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b1\u03b3\u03bd\u03bf\u03b7\u03b8\u03bf\u03cd\u03bd<\/h2>\n<p>\u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03b5\u03af Python \u03c0\u03c1\u03bf\u03b2\u03bb\u03ae\u03bc\u03b1\u03c4\u03b1 \u03b1\u03c0\u03cc TACO \u03ba\u03b1\u03b9 APPS \u03b3\u03b9\u03b1 \u03c4\u03b7\u03bd \u03ba\u03b1\u03c4\u03b1\u03c3\u03ba\u03b5\u03c5\u03ae \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd \u03ba\u03b1\u03b9 \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03b1 benchmark suites \u03b3\u03b9\u03b1 \u03c4\u03b7\u03bd \u03c4\u03b5\u03bb\u03b9\u03ba\u03ae \u03bc\u03ad\u03c4\u03c1\u03b7\u03c3\u03b7. \u039f\u03b9 \u03af\u03b4\u03b9\u03bf\u03b9 \u03bf\u03b9 \u03c3\u03c5\u03b3\u03b3\u03c1\u03b1\u03c6\u03b5\u03af\u03c2 \u03c3\u03b7\u03bc\u03b5\u03b9\u03ce\u03bd\u03bf\u03c5\u03bd \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2 \u03c9\u03c2 \u03c0\u03c1\u03bf\u03c2 \u03c4\u03b1 model families, \u03c4\u03b1 datasets \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03ba\u03b1\u03c4\u03b1\u03bd\u03bf\u03bc\u03ae \u03b4\u03c5\u03c3\u03ba\u03bf\u03bb\u03af\u03b1\u03c2. \u0397 pipeline \u03ba\u03c1\u03b1\u03c4\u03ac \u03c0\u03c1\u03bf\u03b2\u03bb\u03ae\u03bc\u03b1\u03c4\u03b1 \u03bc\u03b5 \u03c3\u03b1\u03c6\u03ad\u03c2 function-style entry point, \u03bc\u03b5 \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03c4\u03b1 \u03c4\u03b5\u03bb\u03b9\u03ba\u03ac training sets \u03bd\u03b1 \u03b3\u03ad\u03c1\u03bd\u03bf\u03c5\u03bd \u03c0\u03c1\u03bf\u03c2 \u03b5\u03c5\u03ba\u03bf\u03bb\u03cc\u03c4\u03b5\u03c1\u03b1 \u03ba\u03b1\u03b9 \u03c0\u03b9\u03bf decomposition-friendly tasks.<\/p>\n<p>\u0397 \u03b5\u03c0\u03b9\u03bb\u03bf\u03b3\u03ae \u03c4\u03b7\u03c2 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7\u03c2 \u03c9\u03c2 step \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b5\u03af \u03ba\u03b1\u03bb\u03ac \u03cc\u03c4\u03b1\u03bd \u03b7 \u03bb\u03b5\u03b9\u03c4\u03bf\u03c5\u03c1\u03b3\u03b9\u03ba\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03c0\u03bf\u03bc\u03bf\u03bd\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9. \u0394\u03b5\u03bd \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 \u03b5\u03be\u03af\u03c3\u03bf\u03c5 \u03b5\u03cd\u03ba\u03bf\u03bb\u03b1 bugs \u03c0\u03bf\u03c5 \u03c0\u03c1\u03bf\u03ba\u03cd\u03c0\u03c4\u03bf\u03c5\u03bd \u03bc\u03cc\u03bd\u03bf \u03b1\u03c0\u03cc \u03b1\u03bb\u03bb\u03b7\u03bb\u03b5\u03c0\u03af\u03b4\u03c1\u03b1\u03c3\u03b7 \u03c0\u03bf\u03bb\u03bb\u03ce\u03bd modules, concurrency, performance regressions, stateful services \u03ae \u03b1\u03bb\u03bb\u03b1\u03b3\u03ad\u03c2 \u03c3\u03c4\u03bf \u03c0\u03b5\u03c1\u03b9\u03b2\u03ac\u03bb\u03bb\u03bf\u03bd. \u0395\u03ba\u03b5\u03af \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2 feedback, \u03cc\u03c0\u03c9\u03c2 API contracts, integration tests \u03ae trace-level invariants.<\/p>\n<p>\u0397 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae decompositions \u03ba\u03b1\u03b9 tests \u03ad\u03c7\u03b5\u03b9 \u03b5\u03c0\u03af\u03c3\u03b7\u03c2 \u03ba\u03cc\u03c3\u03c4\u03bf\u03c2. \u0393\u03b9\u03b1 \u03c4\u03b7 \u03b4\u03b9\u03b1\u03bc\u03cc\u03c1\u03c6\u03c9\u03c3\u03b7 \u03c4\u03c9\u03bd \u03b4\u03b5\u03b4\u03bf\u03bc\u03ad\u03bd\u03c9\u03bd \u03c4\u03bf\u03c5 Qwen2.5-Coder-3B-Instruct, \u03b7 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7 \u03b1\u03bd\u03b1\u03c6\u03ad\u03c1\u03b5\u03b9 \u03c0\u03b5\u03c1\u03af\u03c0\u03bf\u03c5 78 GPU-hours: 24 \u03b3\u03b9\u03b1 decomposition, 43 \u03b3\u03b9\u03b1 generation \u03c4\u03c9\u03bd function-level tests \u03ba\u03b1\u03b9 11 \u03b3\u03b9\u03b1 \u03c4\u03b7\u03bd \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae \u03bf\u03ba\u03c4\u03ce candidates. \u039f\u03b9 decompositions \u03ba\u03b1\u03b9 \u03c4\u03b1 tests \u03b5\u03c0\u03b1\u03bd\u03b1\u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9, \u03b1\u03bb\u03bb\u03ac \u03bc\u03b9\u03b1 \u03b5\u03c0\u03b9\u03c7\u03b5\u03af\u03c1\u03b7\u03c3\u03b7 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03b5\u03c4\u03b1\u03b9 \u03b4\u03b9\u03ba\u03cc \u03c4\u03b7\u03c2 \u03c5\u03c0\u03bf\u03bb\u03bf\u03b3\u03b9\u03c3\u03bc\u03cc \u03c3\u03b5 \u03c3\u03c7\u03ad\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03bf\u03bd \u03cc\u03b3\u03ba\u03bf, \u03c4\u03b7\u03bd \u03ba\u03c1\u03b9\u03c3\u03b9\u03bc\u03cc\u03c4\u03b7\u03c4\u03b1 \u03c4\u03c9\u03bd failures \u03ba\u03b1\u03b9 \u03c4\u03b7\u03bd \u03b1\u03be\u03af\u03b1 \u03c4\u03bf\u03c5 \u03b2\u03b5\u03bb\u03c4\u03b9\u03c9\u03bc\u03ad\u03bd\u03bf\u03c5 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf\u03c5.<\/p>\n<p>\u03a4\u03ad\u03bb\u03bf\u03c2, \u03c4\u03b1 benchmark \u03ba\u03ad\u03c1\u03b4\u03b7 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03bf\u03bc\u03bf\u03b9\u03cc\u03bc\u03bf\u03c1\u03c6\u03b1. \u03a3\u03b5 \u03b1\u03c1\u03ba\u03b5\u03c4\u03ad\u03c2 \u03c3\u03c4\u03ae\u03bb\u03b5\u03c2 \u03b7 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ac \u03b1\u03c0\u03cc KTO \u03b5\u03af\u03bd\u03b1\u03b9 \u03bc\u03b7\u03b4\u03b5\u03bd\u03b9\u03ba\u03ae \u03ae \u03bc\u03b9\u03ba\u03c1\u03ae. \u03a4\u03bf \u03c0\u03bb\u03b5\u03bf\u03bd\u03ad\u03ba\u03c4\u03b7\u03bc\u03b1 \u03c3\u03c5\u03b3\u03ba\u03b5\u03bd\u03c4\u03c1\u03ce\u03bd\u03b5\u03c4\u03b1\u03b9 \u03ba\u03c5\u03c1\u03af\u03c9\u03c2 \u03c3\u03b5 \u03c0\u03b9\u03bf \u03b4\u03cd\u03c3\u03ba\u03bf\u03bb\u03b5\u03c2 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03ae\u03c3\u03b5\u03b9\u03c2. \u0397 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03b4\u03b5\u03bd \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03c5\u03b9\u03bf\u03b8\u03b5\u03c4\u03ae\u03c3\u03b5\u03b9 \u03c0\u03b5\u03c1\u03af\u03c0\u03bb\u03bf\u03ba\u03bf pipeline \u03bc\u03cc\u03bd\u03bf \u03b5\u03c0\u03b5\u03b9\u03b4\u03ae \u03b7 \u03b9\u03b4\u03ad\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b5\u03bb\u03ba\u03c5\u03c3\u03c4\u03b9\u03ba\u03ae, \u03b1\u03bb\u03bb\u03ac \u03bd\u03b1 \u03b5\u03bb\u03ad\u03b3\u03be\u03b5\u03b9 \u03b1\u03bd \u03c4\u03b1 \u03b4\u03b9\u03ba\u03ac \u03c4\u03b7\u03c2 failures \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03c1\u03ac\u03b3\u03bc\u03b1\u03c4\u03b9 \u03c4\u03bf\u03c0\u03b9\u03ba\u03ac \u03ba\u03b1\u03b9 \u03b1\u03bd \u03c4\u03bf function-level signal \u03b1\u03bb\u03bb\u03ac\u03b6\u03b5\u03b9 \u03c4\u03b9\u03c2 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03bc\u03b5\u03c4\u03c1\u03b9\u03ba\u03ad\u03c2 \u03c4\u03b7\u03c2.<\/p>\n<h2 id=\"praktiki-diadromi-efarmogis\">\u039c\u03b9\u03b1 \u03c0\u03c1\u03b1\u03ba\u03c4\u03b9\u03ba\u03ae \u03b4\u03b9\u03b1\u03b4\u03c1\u03bf\u03bc\u03ae \u03b5\u03c6\u03b1\u03c1\u03bc\u03bf\u03b3\u03ae\u03c2<\/h2>\n<p>\u03a4\u03bf \u03c0\u03c1\u03ce\u03c4\u03bf \u03b2\u03ae\u03bc\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03bd\u03b1 \u03bf\u03c1\u03b9\u03c3\u03c4\u03b5\u03af \u03b7 \u03ba\u03b1\u03c4\u03ac\u03bb\u03bb\u03b7\u03bb\u03b7 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b1 feedback. \u03a3\u03b5 \u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03b1 repositories \u03b8\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b7 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7, \u03b1\u03bb\u03bb\u03bf\u03cd \u03ad\u03bd\u03b1 API contract, \u03ad\u03bd\u03b1 transformation stage \u03ae \u03ad\u03bd\u03b1 \u03b5\u03c1\u03b3\u03b1\u03bb\u03b5\u03af\u03bf \u03c0\u03bf\u03c5 \u03ba\u03b1\u03bb\u03b5\u03af \u03bf agent. \u0397 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b1 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03b1\u03c1\u03ba\u03b5\u03c4\u03ac \u03b1\u03c5\u03c4\u03cc\u03bd\u03bf\u03bc\u03b7 \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03b5\u03bb\u03b5\u03b3\u03c7\u03b8\u03b5\u03af, \u03b1\u03bb\u03bb\u03ac \u03cc\u03c7\u03b9 \u03c4\u03cc\u03c3\u03bf \u03bc\u03b9\u03ba\u03c1\u03ae \u03ce\u03c3\u03c4\u03b5 \u03c4\u03bf label \u03bd\u03b1 \u03c7\u03ac\u03bd\u03b5\u03b9 \u03c4\u03bf \u03bd\u03cc\u03b7\u03bc\u03ac \u03c4\u03bf\u03c5.<\/p>\n<p>\u0388\u03c0\u03b5\u03b9\u03c4\u03b1 \u03b4\u03b9\u03b1\u03c7\u03c9\u03c1\u03af\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03c4\u03bf local \u03ba\u03b1\u03b9 \u03c4\u03bf global result, \u03bc\u03b1\u03b6\u03af \u03bc\u03b5 \u03c4\u03b9\u03c2 \u03c3\u03c5\u03b3\u03ba\u03c1\u03bf\u03cd\u03c3\u03b5\u03b9\u03c2 \u03c4\u03bf\u03c5\u03c2. \u03a0\u03c1\u03b9\u03bd \u03c4\u03b1 labels \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b7\u03b8\u03bf\u03cd\u03bd \u03b3\u03b9\u03b1 training, \u03bc\u03b5\u03c4\u03c1\u03b9\u03ad\u03c4\u03b1\u03b9 \u03b7 \u03b1\u03be\u03b9\u03bf\u03c0\u03b9\u03c3\u03c4\u03af\u03b1 \u03c4\u03c9\u03bd tests. \u03a4\u03ad\u03bb\u03bf\u03c2, \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03b5\u03bb\u03b5\u03b3\u03c7\u03cc\u03bc\u03b5\u03bd\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03b1\u03c0\u03ad\u03bd\u03b1\u03bd\u03c4\u03b9 \u03c3\u03b5 outcome-only baseline, \u03bc\u03b5 \u03ad\u03bc\u03c6\u03b1\u03c3\u03b7 \u03c3\u03c4\u03b1 \u03b4\u03cd\u03c3\u03ba\u03bf\u03bb\u03b1, \u03b1\u03bd\u03c4\u03b9\u03c0\u03c1\u03bf\u03c3\u03c9\u03c0\u03b5\u03c5\u03c4\u03b9\u03ba\u03ac tasks \u03ba\u03b1\u03b9 \u03cc\u03c7\u03b9 \u03bc\u03cc\u03bd\u03bf \u03c3\u03b5 \u03b5\u03cd\u03ba\u03bf\u03bb\u03bf\u03c5\u03c2 \u03bc\u03ad\u03c3\u03bf\u03c5\u03c2 \u03cc\u03c1\u03bf\u03c5\u03c2. \u0391\u03bd \u03b7 \u03bf\u03bc\u03ac\u03b4\u03b1 \u03b4\u03bf\u03c5\u03bb\u03b5\u03cd\u03b5\u03b9 \u03bc\u03b5 \u03b1\u03c5\u03c4\u03cc\u03bd\u03bf\u03bc\u03bf\u03c5\u03c2 agents, \u03bf\u03b9 \u03af\u03b4\u03b9\u03bf\u03b9 \u03ad\u03bb\u03b5\u03b3\u03c7\u03bf\u03b9 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03bd\u03c4\u03b1\u03c7\u03b8\u03bf\u03cd\u03bd \u03c3\u03b5 \u03ad\u03bd\u03b1 <a href=\"https:\/\/twodots.gr\/ai-agents-hugging-face-hack-asfaleia-agentic-ai\/\">\u03b1\u03c3\u03c6\u03b1\u03bb\u03ad\u03c2 agentic AI workflow<\/a> \u03bc\u03b5 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03b1 \u03b4\u03b9\u03ba\u03b1\u03b9\u03ce\u03bc\u03b1\u03c4\u03b1 \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03cd\u03c3\u03b9\u03bc\u03b1 artifacts.<\/p>\n<div class=\"td-step-list\">\n<p class=\"td-step-list-title\">Pilot function-level feedback \u03c3\u03b5 \u03b5\u03c0\u03c4\u03ac \u03b2\u03ae\u03bc\u03b1\u03c4\u03b1<\/p>\n<ol>\n<li><span class=\"td-step-kicker\">Failure map<\/span><strong>\u03a7\u03b1\u03c1\u03c4\u03bf\u03b3\u03c1\u03b1\u03c6\u03ae\u03c3\u03c4\u03b5 \u03c4\u03b9\u03c2 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03ba\u03b1\u03c4\u03b7\u03b3\u03bf\u03c1\u03af\u03b5\u03c2 \u03b2\u03bb\u03ac\u03b2\u03b7\u03c2<\/strong>\n<p>\u039e\u03b5\u03c7\u03c9\u03c1\u03af\u03c3\u03c4\u03b5 parsing, business logic, data transformations, integrations \u03ba\u03b1\u03b9 \u03c0\u03b5\u03c1\u03b9\u03b2\u03b1\u03bb\u03bb\u03bf\u03bd\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b1\u03c3\u03c4\u03bf\u03c7\u03af\u03b5\u03c2 \u03c0\u03c1\u03b9\u03bd \u03bf\u03c1\u03af\u03c3\u03b5\u03c4\u03b5 labels.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Unit<\/span><strong>\u0395\u03c0\u03b9\u03bb\u03ad\u03be\u03c4\u03b5 \u03c4\u03b7\u03bd \u03b5\u03bb\u03ac\u03c7\u03b9\u03c3\u03c4\u03b7 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03b7 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b1<\/strong>\n<p>\u03a7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03ae\u03c3\u03c4\u03b5 function, module \u03ae API contract \u03bc\u03cc\u03bd\u03bf \u03cc\u03c4\u03b1\u03bd \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03ba\u03c4\u03b5\u03bb\u03b5\u03c3\u03c4\u03b5\u03af \u03bc\u03b5 \u03c3\u03b1\u03c6\u03ad\u03c2 input \u03ba\u03b1\u03b9 \u03c0\u03b1\u03c1\u03b1\u03c4\u03b7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf output.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Oracle<\/span><strong>\u039f\u03c1\u03af\u03c3\u03c4\u03b5 \u03b5\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03cd\u03c3\u03b9\u03bc\u03bf ground truth<\/strong>\n<p>\u03a0\u03c1\u03bf\u03c4\u03b9\u03bc\u03ae\u03c3\u03c4\u03b5 unit tests, reference execution \u03ba\u03b1\u03b9 invariants \u03b1\u03bd\u03c4\u03af \u03b3\u03b9\u03b1 \u03b1\u03bc\u03b9\u03b3\u03ce\u03c2 \u03bb\u03b5\u03ba\u03c4\u03b9\u03ba\u03ae \u03b1\u03be\u03b9\u03bf\u03bb\u03cc\u03b3\u03b7\u03c3\u03b7 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03b9\u03ba\u03ae\u03c2 \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1\u03c2.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Quality<\/span><strong>\u0395\u03bb\u03ad\u03b3\u03be\u03c4\u03b5 \u03c0\u03c1\u03ce\u03c4\u03b1 \u03c4\u03b1 \u03af\u03b4\u03b9\u03b1 \u03c4\u03b1 tests<\/strong>\n<p>\u039c\u03b5\u03c4\u03c1\u03ae\u03c3\u03c4\u03b5 reference pass, mutation kill rate, edge-case coverage \u03ba\u03b1\u03b9 \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c7\u03c9\u03c1\u03af\u03c2 \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03bf oracle.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Conflict<\/span><strong>\u0394\u03b9\u03b1\u03c4\u03b7\u03c1\u03ae\u03c3\u03c4\u03b5 local \u03ba\u03b1\u03b9 global \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1<\/strong>\n<p>\u039c\u03b7\u03bd \u03c3\u03c5\u03b3\u03c7\u03c9\u03bd\u03b5\u03cd\u03b5\u03c4\u03b5 \u03b1\u03c5\u03c4\u03cc\u03bc\u03b1\u03c4\u03b1 \u03c4\u03b9\u03c2 \u03b4\u03b9\u03b1\u03c6\u03c9\u03bd\u03af\u03b5\u03c2\u00b7 \u03b1\u03c0\u03bf\u03b8\u03b7\u03ba\u03b5\u03cd\u03c3\u03c4\u03b5 \u03c0\u03bf\u03b9\u03bf suite \u03c0\u03b1\u03c1\u03ae\u03b3\u03b1\u03b3\u03b5 \u03ba\u03ac\u03b8\u03b5 label \u03ba\u03b1\u03b9 \u03b3\u03b9\u03b1\u03c4\u03af.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Baseline<\/span><strong>\u03a3\u03c5\u03b3\u03ba\u03c1\u03af\u03bd\u03b5\u03c4\u03b5 \u03bc\u03b5 outcome-only training<\/strong>\n<p>\u03a4\u03c1\u03ad\u03be\u03c4\u03b5 \u03c4\u03bf \u03af\u03b4\u03b9\u03bf \u03b4\u03cd\u03c3\u03ba\u03bf\u03bb\u03bf evaluation \u03bc\u03b5 \u03ba\u03b1\u03b9 \u03c7\u03c9\u03c1\u03af\u03c2 function-level signal \u03ce\u03c3\u03c4\u03b5 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03bc\u03bf\u03bd\u03c9\u03b8\u03b5\u03af \u03b7 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ae \u03c3\u03c5\u03bd\u03b5\u03b9\u03c3\u03c6\u03bf\u03c1\u03ac.<\/p>\n<\/li>\n<li><span class=\"td-step-kicker\">Governance<\/span><strong>\u039a\u03c1\u03b1\u03c4\u03ae\u03c3\u03c4\u03b5 audit trail \u03ba\u03b1\u03b9 \u03cc\u03c1\u03b9\u03b1 \u03c7\u03c1\u03ae\u03c3\u03b7\u03c2<\/strong>\n<p>\u03a3\u03c5\u03bd\u03b4\u03ad\u03c3\u03c4\u03b5 candidate, tests, \u03b1\u03c0\u03bf\u03c4\u03b5\u03bb\u03ad\u03c3\u03bc\u03b1\u03c4\u03b1, model version \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03bf review \u03c0\u03c1\u03b9\u03bd \u03c4\u03bf feedback \u03b5\u03c0\u03b7\u03c1\u03b5\u03ac\u03c3\u03b5\u03b9 \u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03b9\u03ba\u03cc \u03ba\u03ce\u03b4\u03b9\u03ba\u03b1.<\/p>\n<\/li>\n<\/ol>\n<\/div>\n<h2 id=\"ti-kratame-step-ktoder\">\u03a4\u03b9 \u03ba\u03c1\u03b1\u03c4\u03ac\u03bc\u03b5 \u03b1\u03c0\u03cc \u03c4\u03bf Step-KTOder<\/h2>\n<p>\u0397 \u03c0\u03b9\u03bf \u03ce\u03c1\u03b9\u03bc\u03b7 \u03b1\u03bd\u03ac\u03b3\u03bd\u03c9\u03c3\u03b7 \u03b5\u03af\u03bd\u03b1\u03b9 \u03cc\u03c4\u03b9 \u03c4\u03b1 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03b1 \u03b4\u03b5\u03bd \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 \u03b1\u03c0\u03bb\u03ce\u03c2 \u03c0\u03b5\u03c1\u03b9\u03c3\u03c3\u03cc\u03c4\u03b5\u03c1\u03bf feedback\u00b7 \u03c7\u03c1\u03b5\u03b9\u03ac\u03b6\u03bf\u03bd\u03c4\u03b1\u03b9 feedback \u03bc\u03b5 \u03c3\u03c9\u03c3\u03c4\u03ae \u03b1\u03bd\u03ac\u03bb\u03c5\u03c3\u03b7 \u03ba\u03b1\u03b9 \u03b5\u03c0\u03b1\u03bb\u03b7\u03b8\u03b5\u03cd\u03c3\u03b9\u03bc\u03b7 \u03c0\u03b7\u03b3\u03ae. \u038c\u03c4\u03b1\u03bd \u03ad\u03bd\u03b1 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc failure \u03b1\u03ba\u03c5\u03c1\u03ce\u03bd\u03b5\u03b9 \u03c3\u03c9\u03c3\u03c4\u03ac components \u03ae \u03cc\u03c4\u03b1\u03bd \u03ad\u03bd\u03b1 \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc pass \u03ba\u03c1\u03cd\u03b2\u03b5\u03b9 \u03b1\u03b4\u03c5\u03bd\u03b1\u03bc\u03af\u03b1 \u03c3\u03b5 edge case, \u03b7 \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7 \u03b1\u03bd\u03ac \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03b1\u03c0\u03bf\u03ba\u03b1\u03b8\u03b9\u03c3\u03c4\u03ac \u03bc\u03ad\u03c1\u03bf\u03c2 \u03c4\u03b7\u03c2 \u03c0\u03bb\u03b7\u03c1\u03bf\u03c6\u03bf\u03c1\u03af\u03b1\u03c2 \u03c0\u03bf\u03c5 \u03c7\u03ac\u03bd\u03b5\u03c4\u03b1\u03b9.<\/p>\n<p>\u0397 \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03b4\u03b5\u03bd \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03b1\u03b8\u03bf\u03bb\u03b9\u03ba\u03ae \u03bb\u03cd\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 codebase. \u0395\u03af\u03bd\u03b1\u03b9 \u03cc\u03bc\u03c9\u03c2 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc \u03c0\u03b1\u03c1\u03ac\u03b4\u03b5\u03b9\u03b3\u03bc\u03b1 \u03bc\u03b9\u03b1\u03c2 \u03b3\u03b5\u03bd\u03b9\u03ba\u03cc\u03c4\u03b5\u03c1\u03b7\u03c2 \u03b1\u03c1\u03c7\u03ae\u03c2: \u03b7 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b1 \u03b5\u03c0\u03bf\u03c0\u03c4\u03b5\u03af\u03b1\u03c2 \u03c0\u03c1\u03ad\u03c0\u03b5\u03b9 \u03bd\u03b1 \u03b5\u03c5\u03b8\u03c5\u03b3\u03c1\u03b1\u03bc\u03bc\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03bc\u03b5 \u03c4\u03b7 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b1 \u03c0\u03bf\u03c5 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c4\u03cd\u03c7\u03b5\u03b9 \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03b5\u03bb\u03b5\u03b3\u03c7\u03b8\u03b5\u03af. \u0393\u03b9\u03b1 \u03bc\u03b9\u03b1 \u03bf\u03bc\u03ac\u03b4\u03b1 software \u03ae AI, \u03b1\u03c5\u03c4\u03cc \u03c3\u03b7\u03bc\u03b1\u03af\u03bd\u03b5\u03b9 \u03c3\u03b1\u03c6\u03ad\u03c3\u03c4\u03b5\u03c1\u03b1 contracts, \u03ba\u03b1\u03bb\u03cd\u03c4\u03b5\u03c1\u03b1 tests, \u03b4\u03b9\u03b1\u03c4\u03ae\u03c1\u03b7\u03c3\u03b7 \u03c4\u03c9\u03bd conflicts \u03ba\u03b1\u03b9 metrics \u03c0\u03bf\u03c5 \u03be\u03b5\u03c7\u03c9\u03c1\u03af\u03b6\u03bf\u03c5\u03bd \u03c4\u03b7\u03bd \u03c4\u03bf\u03c0\u03b9\u03ba\u03ae \u03bf\u03c1\u03b8\u03cc\u03c4\u03b7\u03c4\u03b1 \u03b1\u03c0\u03cc \u03c4\u03bf end-to-end \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1.<\/p>\n<p>\u03a4\u03bf Step-KTOder \u03b4\u03b5\u03bd \u03ba\u03b1\u03c4\u03b1\u03c1\u03b3\u03b5\u03af \u03c4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc test suite \u03ba\u03b1\u03b9 \u03b4\u03b5\u03bd \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03b8\u03b9\u03c3\u03c4\u03ac \u03c4\u03bf human review. \u03a3\u03c5\u03bd\u03b4\u03c5\u03ac\u03b6\u03b5\u03b9 \u03c4\u03bf coarse outcome \u03bc\u03b5 fine-grained execution evidence. \u0391\u03c5\u03c4\u03ae \u03b7 \u03b9\u03c3\u03bf\u03c1\u03c1\u03bf\u03c0\u03af\u03b1 \u03b5\u03af\u03bd\u03b1\u03b9 \u03ba\u03b1\u03b9 \u03c4\u03bf \u03c0\u03b9\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03cc \u03c4\u03bf\u03c5 \u03b4\u03af\u03b4\u03b1\u03b3\u03bc\u03b1: \u03bf \u03b1\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03cc\u03c2 \u03b3\u03af\u03bd\u03b5\u03c4\u03b1\u03b9 \u03c0\u03b9\u03bf \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03bf\u03c2 \u03cc\u03c4\u03b1\u03bd \u03b7 \u03b1\u03c0\u03cc\u03c6\u03b1\u03c3\u03b7 \u03bc\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b5\u03bd\u03c4\u03bf\u03c0\u03b9\u03c3\u03c4\u03b5\u03af, \u03bd\u03b1 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03c7\u03b8\u03b5\u03af \u03ba\u03b1\u03b9 \u03bd\u03b1 \u03b1\u03bc\u03c6\u03b9\u03c3\u03b2\u03b7\u03c4\u03b7\u03b8\u03b5\u03af.<\/p>\n<section class=\"td-service-cta\">\n<div class=\"td-service-cta-content\">\n<p class=\"td-service-cta-eyebrow\">AI workflows \u03bc\u03b5 \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1<\/p>\n<p class=\"td-service-cta-title\">\u039c\u03b5\u03c4\u03b1\u03c4\u03c1\u03ad\u03c8\u03c4\u03b5 \u03c4\u03bf feedback \u03c4\u03c9\u03bd AI agents \u03c3\u03b5 \u03bc\u03b5\u03c4\u03c1\u03ae\u03c3\u03b9\u03bc\u03b7 \u03c5\u03c0\u03bf\u03b4\u03bf\u03bc\u03ae<\/p>\n<p>\u0397 TWO DOTS \u03c3\u03c7\u03b5\u03b4\u03b9\u03ac\u03b6\u03b5\u03b9 \u03b1\u03c5\u03c4\u03bf\u03bc\u03b1\u03c4\u03b9\u03c3\u03bc\u03bf\u03cd\u03c2 \u03bc\u03b5 \u03c3\u03b1\u03c6\u03ae contracts, execution checks, observability \u03ba\u03b1\u03b9 \u03b1\u03bd\u03b8\u03c1\u03ce\u03c0\u03b9\u03bd\u03b1 \u03c3\u03b7\u03bc\u03b5\u03af\u03b1 \u03b5\u03bb\u03ad\u03b3\u03c7\u03bf\u03c5, \u03ce\u03c3\u03c4\u03b5 \u03ba\u03ac\u03b8\u03b5 \u03b1\u03c0\u03bf\u03c4\u03c5\u03c7\u03af\u03b1 \u03bd\u03b1 \u03b1\u03c6\u03ae\u03bd\u03b5\u03b9 \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03c4\u03b5\u03c7\u03bd\u03b9\u03ba\u03cc \u03c3\u03ae\u03bc\u03b1.<\/p>\n<div class=\"td-service-cta-actions\"><a class=\"td-service-cta-button\" href=\"https:\/\/twodots.gr\/aftomatismoi-epicheiriseon-ai\/\">\u0412\u0438\u0436\u0442\u0435 \u0430\u0432\u0442\u043e\u043c\u0430\u0442\u0438\u0437\u0430\u0446\u0438\u044f\u0442\u0430 \u043d\u0430 \u0431\u0438\u0437\u043d\u0435\u0441 \u043f\u0440\u043e\u0446\u0435\u0441\u0438\u0442\u0435 \u0438 \u0438\u0437\u043a\u0443\u0441\u0442\u0432\u0435\u043d\u0438\u044f \u0438\u043d\u0442\u0435\u043b\u0435\u043a\u0442<\/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\">\u0427\u0435\u0441\u0442\u043e \u0437\u0430\u0434\u0430\u0432\u0430\u043d\u0438 \u0432\u044a\u043f\u0440\u043e\u0441\u0438<\/p>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a4\u03b9 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf Step-KTOder;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0395\u03af\u03bd\u03b1\u03b9 \u03c0\u03bb\u03b1\u03af\u03c3\u03b9\u03bf preference optimization \u03b3\u03b9\u03b1 code LLM \u03c0\u03bf\u03c5 \u03c3\u03c5\u03bd\u03b4\u03c5\u03ac\u03b6\u03b5\u03b9 outcome-level KTO \u03bc\u03b5 \u03b4\u03c5\u03b1\u03b4\u03b9\u03ba\u03cc feedback \u03b1\u03bd\u03ac module-level function, \u03b2\u03b1\u03c3\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf \u03c3\u03b5 \u03b1\u03c5\u03c4\u03cc\u03bc\u03b1\u03c4\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03cc\u03bc\u03b5\u03bd\u03b1 \u03ba\u03b1\u03b9 \u03c6\u03b9\u03bb\u03c4\u03c1\u03b1\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03b1 unit tests.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03c7\u03c1\u03b7\u03c3\u03b9\u03bc\u03bf\u03c0\u03bf\u03b9\u03b5\u03af \u03c4\u03b9\u03c2 \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03c9\u03c2 steps;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039f\u03b9 \u03c3\u03c5\u03bd\u03b1\u03c1\u03c4\u03ae\u03c3\u03b5\u03b9\u03c2 \u03b5\u03af\u03bd\u03b1\u03b9 \u03c3\u03b7\u03bc\u03b1\u03c3\u03b9\u03bf\u03bb\u03bf\u03b3\u03b9\u03ba\u03ad\u03c2 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b5\u03c2 \u03bc\u03b5 interface \u03ba\u03b1\u03b9 \u03bc\u03c0\u03bf\u03c1\u03bf\u03cd\u03bd \u03c3\u03c5\u03c7\u03bd\u03ac \u03bd\u03b1 \u03b5\u03bb\u03b5\u03b3\u03c7\u03b8\u03bf\u03cd\u03bd \u03b1\u03bd\u03b5\u03be\u03ac\u03c1\u03c4\u03b7\u03c4\u03b1. \u0391\u03c5\u03c4\u03cc \u03b5\u03c0\u03b9\u03c4\u03c1\u03ad\u03c0\u03b5\u03b9 \u03c0\u03b9\u03bf \u03b5\u03bd\u03c4\u03bf\u03c0\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf feedback \u03b1\u03c0\u03cc \u03ad\u03bd\u03b1 \u03b5\u03bd\u03b9\u03b1\u03af\u03bf label \u03b3\u03b9\u03b1 \u03cc\u03bb\u03bf \u03c4\u03bf \u03c0\u03c1\u03cc\u03b3\u03c1\u03b1\u03bc\u03bc\u03b1.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a5\u03c0\u03b5\u03c1\u03b5\u03af\u03c7\u03b5 \u03c0\u03ac\u03bd\u03c4\u03b1 \u03b1\u03c0\u03cc \u03c4\u03bf outcome-only KTO;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u038c\u03c7\u03b9 \u03c3\u03b5 \u03ba\u03ac\u03b8\u03b5 benchmark \u03ba\u03b1\u03b9 \u03bc\u03bf\u03bd\u03c4\u03ad\u03bb\u03bf. \u039f\u03b9 \u03ba\u03b1\u03b8\u03b1\u03c1\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03c3\u03c7\u03b5\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b2\u03b5\u03bb\u03c4\u03b9\u03ce\u03c3\u03b5\u03b9\u03c2 \u03b5\u03bc\u03c6\u03b1\u03bd\u03af\u03c3\u03c4\u03b7\u03ba\u03b1\u03bd \u03ba\u03c5\u03c1\u03af\u03c9\u03c2 \u03c3\u03c4\u03b1 BigCodeBench Hard \u03ba\u03b1\u03b9 LiveCodeBench, \u03b5\u03bd\u03ce \u03c3\u03b5 \u03b5\u03c5\u03ba\u03bf\u03bb\u03cc\u03c4\u03b5\u03c1\u03b5\u03c2 \u03b1\u03be\u03b9\u03bf\u03bb\u03bf\u03b3\u03ae\u03c3\u03b5\u03b9\u03c2 \u03b1\u03c1\u03ba\u03b5\u03c4\u03ad\u03c2 \u03b4\u03b9\u03b1\u03c6\u03bf\u03c1\u03ad\u03c2 \u03ae\u03c4\u03b1\u03bd \u03bc\u03b9\u03ba\u03c1\u03ad\u03c2 \u03ae \u03bc\u03b7\u03b4\u03b5\u03bd\u03b9\u03ba\u03ad\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0393\u03b9\u03b1\u03c4\u03af \u03b4\u03b9\u03b1\u03c4\u03b7\u03c1\u03bf\u03cd\u03bd\u03c4\u03b1\u03b9 \u03bf\u03b9 \u03c3\u03c5\u03b3\u03ba\u03c1\u03bf\u03cd\u03c3\u03b5\u03b9\u03c2 local \u03ba\u03b1\u03b9 global labels;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039c\u03c0\u03bf\u03c1\u03b5\u03af \u03bd\u03b1 \u03b1\u03c0\u03bf\u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03bf\u03c5\u03bd \u03c4\u03bf\u03c0\u03b9\u03ba\u03ad\u03c2 \u03b1\u03c3\u03c4\u03bf\u03c7\u03af\u03b5\u03c2 \u03c0\u03bf\u03c5 \u03b4\u03b5\u03bd \u03ba\u03b1\u03bb\u03cd\u03c0\u03c4\u03b5\u03b9 \u03c4\u03bf \u03c3\u03c5\u03bd\u03bf\u03bb\u03b9\u03ba\u03cc test suite. \u03a3\u03c4\u03b1 ablations \u03c4\u03b7\u03c2 \u03bc\u03b5\u03bb\u03ad\u03c4\u03b7\u03c2, \u03b7 \u03b4\u03b9\u03b1\u03c4\u03ae\u03c1\u03b7\u03c3\u03b7 \u03b1\u03c5\u03c4\u03ce\u03bd \u03c4\u03c9\u03bd \u03c3\u03c5\u03b3\u03ba\u03c1\u03bf\u03cd\u03c3\u03b5\u03c9\u03bd \u03ad\u03b4\u03c9\u03c3\u03b5 \u03b9\u03c3\u03c7\u03c5\u03c1\u03cc\u03c4\u03b5\u03c1\u03bf \u03b1\u03c0\u03bf\u03c4\u03ad\u03bb\u03b5\u03c3\u03bc\u03b1 \u03b1\u03c0\u03cc \u03c4\u03bf masking.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u039c\u03c0\u03bf\u03c1\u03b5\u03af \u03ad\u03bd\u03b1\u03c2 LLM judge \u03bd\u03b1 \u03b1\u03bd\u03c4\u03b9\u03ba\u03b1\u03c4\u03b1\u03c3\u03c4\u03ae\u03c3\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b5\u03ba\u03c4\u03ad\u03bb\u03b5\u03c3\u03b7;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a3\u03c4\u03bf \u03c3\u03c5\u03b3\u03ba\u03b5\u03ba\u03c1\u03b9\u03bc\u03ad\u03bd\u03bf \u03c0\u03b5\u03af\u03c1\u03b1\u03bc\u03b1 \u03cc\u03c7\u03b9. \u039f judge \u03c5\u03c0\u03b5\u03c1\u03c0\u03c1\u03bf\u03ad\u03b2\u03bb\u03b5\u03c8\u03b5 function failures \u03ba\u03b1\u03b9 \u03c4\u03b1 labels \u03c4\u03bf\u03c5 \u03bc\u03b5\u03af\u03c9\u03c3\u03b1\u03bd \u03c4\u03b7\u03bd \u03b1\u03c0\u03cc\u03b4\u03bf\u03c3\u03b7 \u03c3\u03c4\u03b1 \u03b4\u03c5\u03c3\u03ba\u03bf\u03bb\u03cc\u03c4\u03b5\u03c1\u03b1 benchmarks \u03c3\u03b5 \u03c3\u03c7\u03ad\u03c3\u03b7 \u03bc\u03b5 \u03c4\u03b1 execution-based labels.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03cc\u03c3\u03bf \u03b1\u03be\u03b9\u03cc\u03c0\u03b9\u03c3\u03c4\u03b1 \u03ae\u03c4\u03b1\u03bd \u03c4\u03b1 \u03c0\u03b1\u03c1\u03b1\u03b3\u03cc\u03bc\u03b5\u03bd\u03b1 unit tests;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u03a4\u03bf\u03c5\u03bb\u03ac\u03c7\u03b9\u03c3\u03c4\u03bf\u03bd \u03ad\u03bd\u03b1 valid test \u03b2\u03c1\u03ad\u03b8\u03b7\u03ba\u03b5 \u03b3\u03b9\u03b1 82,0% \u03c4\u03c9\u03bd testable functions \u03c3\u03c4\u03bf TACO \u03ba\u03b1\u03b9 84,5% \u03c3\u03c4\u03bf APPS. \u0397 mutation audit \u03ad\u03b4\u03c9\u03c3\u03b5 92,6% kill rate, \u03b1\u03bb\u03bb\u03ac \u03b7 \u03b1\u03c4\u03b5\u03bb\u03ae\u03c2 \u03ba\u03ac\u03bb\u03c5\u03c8\u03b7 \u03c0\u03b1\u03c1\u03b1\u03bc\u03ad\u03bd\u03b5\u03b9 \u03b1\u03bd\u03b1\u03b3\u03bd\u03c9\u03c1\u03b9\u03c3\u03bc\u03ad\u03bd\u03bf\u03c2 \u03c0\u03b5\u03c1\u03b9\u03bf\u03c1\u03b9\u03c3\u03bc\u03cc\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u0395\u03c6\u03b1\u03c1\u03bc\u03cc\u03b6\u03b5\u03c4\u03b1\u03b9 \u03ac\u03bc\u03b5\u03c3\u03b1 \u03c3\u03b5 production repositories;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u0397 \u03b5\u03c1\u03b3\u03b1\u03c3\u03af\u03b1 \u03c0\u03b1\u03c1\u03ad\u03c7\u03b5\u03b9 \u03b5\u03c1\u03b5\u03c5\u03bd\u03b7\u03c4\u03b9\u03ba\u03ae \u03c4\u03b5\u03ba\u03bc\u03b7\u03c1\u03af\u03c9\u03c3\u03b7, \u03cc\u03c7\u03b9 \u03b5\u03b3\u03b3\u03cd\u03b7\u03c3\u03b7 \u03b3\u03b9\u03b1 \u03ba\u03ac\u03b8\u03b5 repository. \u0391\u03c0\u03b1\u03b9\u03c4\u03b5\u03af\u03c4\u03b1\u03b9 \u03c4\u03bf\u03c0\u03b9\u03ba\u03cc pilot \u03bc\u03b5 \u03c4\u03b9\u03c2 \u03c0\u03c1\u03b1\u03b3\u03bc\u03b1\u03c4\u03b9\u03ba\u03ad\u03c2 \u03b3\u03bb\u03ce\u03c3\u03c3\u03b5\u03c2, dependencies, tests \u03ba\u03b1\u03b9 failure modes \u03c4\u03b7\u03c2 \u03bf\u03bc\u03ac\u03b4\u03b1\u03c2.<\/p>\n<\/div>\n<\/details>\n<details class=\"td-faq-item\">\n<summary class=\"td-faq-title\">\u03a0\u03bf\u03b9\u03bf \u03b5\u03af\u03bd\u03b1\u03b9 \u03c4\u03bf \u03b2\u03b1\u03c3\u03b9\u03ba\u03cc \u03b5\u03c0\u03b9\u03c7\u03b5\u03b9\u03c1\u03b7\u03c3\u03b9\u03b1\u03ba\u03cc \u03b4\u03af\u03b4\u03b1\u03b3\u03bc\u03b1;<\/summary>\n<div class=\"td-faq-content\">\n<p>\u039d\u03b1 \u03bc\u03b7 \u03c3\u03c5\u03b3\u03c7\u03ad\u03b5\u03c4\u03b1\u03b9 \u03b7 \u03c0\u03bf\u03c3\u03cc\u03c4\u03b7\u03c4\u03b1 feedback \u03bc\u03b5 \u03c4\u03b7\u03bd \u03c0\u03bf\u03b9\u03cc\u03c4\u03b7\u03c4\u03ac \u03c4\u03bf\u03c5. \u03a4\u03bf feedback \u03b5\u03af\u03bd\u03b1\u03b9 \u03c0\u03b9\u03bf \u03c7\u03c1\u03ae\u03c3\u03b9\u03bc\u03bf \u03cc\u03c4\u03b1\u03bd \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03b5\u03b9 \u03c4\u03b7\u03bd \u03b5\u03bb\u03ad\u03b3\u03be\u03b9\u03bc\u03b7 \u03bc\u03bf\u03bd\u03ac\u03b4\u03b1 \u03c0\u03bf\u03c5 \u03b1\u03c0\u03ad\u03c4\u03c5\u03c7\u03b5 \u03ba\u03b1\u03b9 \u03c3\u03c4\u03b7\u03c1\u03af\u03b6\u03b5\u03c4\u03b1\u03b9 \u03c3\u03b5 \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03b3\u03ce\u03b3\u03b9\u03bc\u03b7 \u03c3\u03c5\u03bc\u03c0\u03b5\u03c1\u03b9\u03c6\u03bf\u03c1\u03ac.<\/p>\n<\/div>\n<\/details>\n<\/div>\n<\/section>\n<div class=\"td-source-list\">\n<p id=\"piges\" class=\"td-source-list-title\">\u0418\u0437\u0442\u043e\u0447\u043d\u0438\u0446\u0438<\/p>\n<ul>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2608.23632\" target=\"_blank\" rel=\"noopener\">Nechnech et al.: Function-Level Execution Feedback for Code Preference Optimization<\/a><\/li>\n<li><a href=\"https:\/\/github.com\/inechnech\/STEP-KTODER\" target=\"_blank\" rel=\"noopener\">STEP-KTODER: \u03b5\u03c0\u03af\u03c3\u03b7\u03bc\u03bf repository \u03ba\u03b1\u03b9 pipeline \u03b1\u03bd\u03b1\u03c0\u03b1\u03c1\u03b1\u03b3\u03c9\u03b3\u03ae\u03c2<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2402.01306\" target=\"_blank\" rel=\"noopener\">Ethayarajh et al.: KTO &#8211; Model Alignment as Prospect Theoretic Optimization<\/a><\/li>\n<li><a href=\"https:\/\/arxiv.org\/abs\/2305.18290\" target=\"_blank\" rel=\"noopener\">Rafailov et al.: Direct Preference Optimization<\/a><\/li>\n<\/ul>\n<\/div>","protected":false},"excerpt":{"rendered":"<p>\u03a4\u03bf Step-KTOder \u03b4\u03b5\u03af\u03c7\u03bd\u03b5\u03b9 \u03c0\u03ce\u03c2 \u03c4\u03bf execution feedback \u03b1\u03bd\u03ac \u03c3\u03c5\u03bd\u03ac\u03c1\u03c4\u03b7\u03c3\u03b7 \u03b2\u03bf\u03b7\u03b8\u03ac \u03c4\u03b1 code LLM \u03bd\u03b1 \u03b5\u03bd\u03c4\u03bf\u03c0\u03af\u03b6\u03bf\u03c5\u03bd \u03b2\u03bb\u03ac\u03b2\u03b5\u03c2 \u03c7\u03c9\u03c1\u03af\u03c2 \u03bd\u03b1 \u03b1\u03c0\u03bf\u03c1\u03c1\u03af\u03c0\u03c4\u03bf\u03c5\u03bd \u03c4\u03b1 \u03c3\u03c9\u03c3\u03c4\u03ac \u03bc\u03ad\u03c1\u03b7 \u03c4\u03bf\u03c5 \u03c0\u03c1\u03bf\u03b3\u03c1\u03ac\u03bc\u03bc\u03b1\u03c4\u03bf\u03c2.<\/p>","protected":false},"author":1,"featured_media":111646,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_gspb_post_css":"","footnotes":""},"categories":[199],"tags":[21298,21301,21299,21302,21300],"class_list":["post-98920","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-techniti-noimosyni","tag-code-ai","tag-preference-optimization","tag-process-supervision","tag-step-ktoder","tag-unit-testing"],"_links":{"self":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/98920","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/comments?post=98920"}],"version-history":[{"count":0,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/posts\/98920\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/media\/111646"}],"wp:attachment":[{"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/media?parent=98920"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/categories?post=98920"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/twodots.gr\/bg\/wp-json\/wp\/v2\/tags?post=98920"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}