\u97e9\u56fdLG AI Research\u4e0eKAIST AI\u8054\u5408\u63d0\u51fa\u5408\u6210\u81ea\u84b8\u998f\uff08Synthetic Self-Distillation\uff09\u65b9\u6cd5\uff0c\u8ba90.5B\u81f33B\u7ea7\u5c0f\u6a21\u578b\u5728\u6301\u7eed\u540e\u8bad\u7ec3\u4e2d\u65e2\u83b7\u5f97\u65b0\u80fd\u529b\u53c8\u4e0d\u4e22\u6389\u65e7\u80fd\u529b\u3002\u8bba\u65872026\u5e7410\u67089\u65e5\u63d0\u4ea4\u81f3\u9884\u5370\u672c\u5e73\u53f0\uff0c\u7b2c\u4e00\u4f5c\u8005Woosung Koh\u6765\u81eaKAIST AI\uff0c\u8bba\u6587\u6ce8\u660e\u5176\u5de5\u4f5c\u5b8c\u6210\u4e8e\u5728LG AI Research\u5b9e\u4e60\u671f\u95f4\uff0c\u901a\u8baf\u4f5c\u8005Moontae Lee\u4efb\u804c\u4e8e\u8be5\u7814\u7a76\u9662\u3002<\/p>\n
\u95ee\u9898\u6e90\u81ea\u4e0a\u4e00\u5e74\u63d0\u51fa\u7684\u81ea\u84b8\u998f\u5fae\u8c03\uff08SDFT\uff09\uff1a\u8ba9\u6a21\u578b\u5148\u770b\u7740\u6807\u51c6\u7b54\u6848\u7528\u81ea\u5df1\u7684\u8bdd\u590d\u8ff0\u4e00\u904d\uff0c\u518d\u62ff\u590d\u8ff0\u7248\u672c\u8bad\u7ec3\u81ea\u5df1\uff0c\u597d\u628a\u8bad\u7ec3\u6570\u636e\u62c9\u56de\u6a21\u578b\u81ea\u5df1\u7684\u8868\u8fbe\u5206\u5e03\u3002\u8bba\u6587\u6307\u51fa\uff0c\u8fd9\u5957\u505a\u6cd5\u4f9d\u8d56\u4e0a\u4e0b\u6587\u5b66\u4e60\u4e0e\u6307\u4ee4\u9075\u5faa\u80fd\u529b\uff0c\u800c\u5c0f\u6a21\u578b\u6070\u5728\u8fd9\u4e24\u9879\u4e0a\u6700\u5f31\uff0c\u4f4e\u5ef6\u8fdf\u3001\u7aef\u4fa7\u90e8\u7f72\u3001\u8349\u7a3f\u6a21\u578b\u4e0e\u9a8c\u8bc1\u5668\u7b49\u573a\u666f\u8981\u7684\u53c8\u6b63\u662f\u5c0f\u6a21\u578b\u3002\u5de5\u5177\u4f7f\u7528\u57df\u7684\u5b9e\u6d4b\u663e\u793a\uff0c\u6a21\u578b\u81ea\u5df1\u6539\u5199\u793a\u8303\u65f6\u5e73\u5747\u51c6\u786e\u7387\u964d\u523060%\uff0c\u601d\u7ef4\u94fe\u67098.2%\u88ab\u4e22\u5f03\u3002<\/p>\n
\u4f5c\u8005\u5148\u505a\u8bca\u65ad\u3002\u56e2\u961f\u752823\u4e2a\u6570\u5b66\u8bad\u7ec3\u6570\u636e\u96c6\u505a\u540e\u8bad\u7ec3\uff0c\u6bcf\u4e2a\u89c4\u6a21\u5339\u914d\u5230\u7ea6250\u4e07completion token\uff0c\u518d\u7528GPQA-Diamond\u3001ARC-C\u3001MMLU-Pro\u3001HumanEval\u3001MBPP\u4e94\u4e2a\u975e\u6570\u5b66\u57fa\u51c6\u7684\u5e73\u5747\u8dcc\u5e45\u8861\u91cf\u9057\u5fd8\uff0c\u7ed3\u8bba\u662f\u540e\u8bad\u7ec3\u9020\u6210\u7684\u5206\u5e03\u504f\u79fb\u4e0e\u9057\u5fd8\u5e45\u5ea6\u5448\u7ebf\u6027\u5173\u7cfb\uff0cSpearman\u76f8\u5173\u7cfb\u6570\u4e3a\u8d1f0.85\u3002\u7b2c\u4e8c\u7ec4\u5b9e\u9a8c\u628a\u53d8\u91cf\u6536\u5f97\u66f4\u7d27\uff1a\u540c\u4e00\u4efd\u7ea6200\u4e07completion token\u7684\u6570\u5b66\u8bed\u6599\u3001\u540c\u4e00\u5957\u63d0\u793a\uff0c\u53ea\u6539\u76ee\u6807\u6587\u672c\u2014\u2014\u9ad8\u504f\u79fb\u7248\u4fdd\u7559\u5b8c\u6574\u957f\u63a8\u7406\u94fe\uff0c\u4e2d\u504f\u79fb\u53ea\u7559\u6700\u7ec8\u89e3\u7b54\uff0c\u4f4e\u504f\u79fb\u662f\u6a21\u578b\u81ea\u5df1\u751f\u6210\u5e76\u9a8c\u8bc1\u901a\u8fc7\u7684\u89e3\u7b54\uff1b\u504f\u79fb\u5355\u8c03\u4e0b\u964d\u65f6\uff0c\u76ee\u6807\u57df\u6570\u5b66\u8868\u73b0\u3001\u76ee\u6807\u57df\u4fdd\u6301\u4e0e\u8de8\u9886\u57df\u4ee3\u7801\u4fdd\u6301\u4e09\u9879\u540c\u65f6\u5355\u8c03\u53d8\u597d\u3002\u636e\u6b64\uff0c\u65b9\u6cd5\u628a\u300c\u590d\u8ff0\u300d\u5916\u5305\u51fa\u53bb\uff1a\u4ea4\u7ed9\u53c2\u6570\u66f4\u5927\u4f46\u540c\u5c5e\u4e00\u4e2a\u5bb6\u65cf\u7684\u5bf9\u9f50\u6a21\u578b\uff08\u5b9e\u9a8c\u7528Qwen2.5\u76847B\u4e0e14B\uff09\u6539\u5199\u6807\u51c6\u7b54\u6848\uff0c\u6539\u5199\u79bb\u7ebf\u751f\u6210\u3001\u4e0e\u6807\u51c6\u7b54\u6848\u9010\u6761\u6838\u5bf9\uff0c\u4e0d\u4e00\u81f4\u5373\u4e22\u5f03\uff0c\u5269\u4e0b\u7684\u518d\u7528\u5e38\u89c4\u76d1\u7763\u5fae\u8c03\u8bad\u7ec3\u5b66\u751f\u6a21\u578b\u3002\u4f5c\u8005\u79f0\u8fd9\u65e2\u62ff\u5230\u9ad8\u8d28\u91cf\u76d1\u7763\uff0c\u53c8\u6ca1\u7834\u574f\u5b66\u751f\u539f\u6709\u5206\u5e03\uff0c\u4e14\u4e0d\u9700\u989d\u5916\u8c03\u53c2\u3002<\/p>\n
\u8bc4\u6d4b\u8986\u76d6Qwen2.5-Instruct\u76840.5B\u30011.5B\u30013B\u4e09\u4e2a\u89c4\u6a21\uff0c4\u4e2a\u83b7\u53d6\u4efb\u52a1\uff08\u542b\u56db\u5b66\u79d1\u6df7\u5408\u79d1\u5b66\u4efb\u52a1\u3001\u5de5\u5177\u4f7f\u7528\u4e0e\u533b\u5b66\u95ee\u7b54\uff09\uff0c\u4ee5\u53ca\u6a2a\u8de8\u6307\u4ee4\u9075\u5faa\u3001\u5e38\u8bc6\u3001\u591a\u6837\u77e5\u8bc6\u3001\u4ee3\u7801\u4e0e\u6570\u5b66\u768414\u4e2a\u4fdd\u6301\u57fa\u51c6\uff0c\u5bf9\u6bd4SFT\u3001SDFT\u3001DFT\u3001ASFT\u4e0e\u79bb\u7b56\u7565\u3001\u5728\u7b56\u7565\u84b8\u998f\u3002\u6c47\u603b\u6765\u770b\uff0c\u65b0\u80fd\u529b\u4e60\u5f97\u4e0aSynthSelf\u5e73\u574756.4%\uff0c\u9ad8\u4e8eSFT\u768452.2%\u3001SDFT\u768444.1%\u3001DFT\u768439.8%\u548cASFT\u768446.1%\uff1b\u65e7\u80fd\u529b\u4fdd\u630141.7\u81f341.8\uff0c\u4e0e\u6700\u597d\u7684\u57fa\u7ebfASFT\u768441.8\u57fa\u672c\u6301\u5e73\u3002\u51e0\u4e2a\u7ec6\u8282\u6bd4\u603b\u8868\u66f4\u80fd\u8bf4\u660e\u673a\u5236\uff1a\u540c\u5bb6\u65cf7B\u4e0e14B\u6539\u5199\u51fa\u7684\u6837\u672c\u5728\u57fa\u7840\u6a21\u578b\u4e0b\u7684\u8d1f\u5bf9\u6570\u4f3c\u7136\u66f4\u4f4e\uff0c\u8de8\u5bb6\u65cf\u66f4\u5f3a\u6a21\u578b\u7684\u6570\u636e\u51c6\u786e\u7387\u66f4\u9ad8\u4f46\u504f\u79fb\u66f4\u5927\uff0c\u540c\u5bb6\u65cf\u65b9\u6848\u8fd8\u7701\u4e0b62.9%\u81f376.9%\u7684\u6570\u636e\u751f\u6210\u7b97\u529b\uff1b\u62d2\u7edd\u91c7\u6837\u8ba9\u5de5\u5177\u4f7f\u7528\u4efb\u52a1\u7684\u8bad\u7ec3\u6837\u672c\u51cf\u5c1117.1%\uff0c\u4f46\u76f8\u5bf9\u666e\u901a\u76d1\u7763\u5fae\u8c03\u83b7\u53d6\u4ecd\u63d0\u53474.2\u4e2a\u767e\u5206\u70b9\u3002<\/p>\n
\u628a\u8fd9\u5957\u505a\u6cd5\u653e\u56de\u7aef\u4fa7AI\u7684\u73b0\u5b9e\u91cc\uff0c\u610f\u4e49\u624d\u5b8c\u6574\u3002\u56fd\u5185\u516c\u5f00\u7684\u4ea7\u4e1a\u89c2\u5bdf\u663e\u793a\uff0c\u7aef\u4fa7\u6a21\u578b\u6b63\u4ece\u300c\u80fd\u8dd1\u300d\u5f80\u300c\u80fd\u529e\u4e8b\u300d\u63a8\uff1a\u300a\u4eba\u5de5\u667a\u80fd\u7ec8\u7aef\u667a\u80fd\u5316\u5206\u7ea7\u300b\u56fd\u5bb6\u6807\u51c6\u843d\u5730\u540e\u7684\u9996\u6279\u6d4b\u8bc4\u5df2\u6709\u4ea7\u54c1\u53d6\u5f97L3\u7ea7\u8ba4\u8bc1\uff1b\u672c\u5730\u6a21\u578b\u5904\u7406\u9ad8\u9891\u3001\u8f7b\u91cf\u4e0e\u6d89\u9690\u79c1\u4efb\u52a1\u3001\u590d\u6742\u4efb\u52a1\u8131\u654f\u540e\u4e0a\u4e91\u7684\u667a\u80fd\u8def\u7531\uff0c\u5728\u81ea\u52a8\u5316\u8bbe\u5907\u6d4b\u8bd5\u573a\u666f\u628a\u5b8c\u5168\u672c\u5730\u6267\u884c\u7684\u4efb\u52a1\u62c6\u89e3\u6210\u529f\u7387\u505a\u5230\u7ea680%\uff0c\u914d\u5408\u8def\u7531\u540e\u6574\u4f53\u51c6\u786e\u7387\u62ac\u523095%\u4ee5\u4e0a\u3002\u6301\u7eed\u540e\u8bad\u7ec3\u5904\u7406\u7684\u6b63\u662f\u5c0f\u6a21\u578b\u88c5\u8fdb\u8bbe\u5907\u4e4b\u540e\u8fd8\u80fd\u4e0d\u80fd\u7ee7\u7eed\u53d8\u597d\u7684\u95ee\u9898\u3002\u8bba\u6587\u5e76\u672a\u6d4b\u91cf\u7aef\u4fa7\u5ef6\u8fdf\u3001\u529f\u8017\u4e0e\u5185\u5b58\u5360\u7528\uff0c\u300c\u88c5\u5f97\u4e0b\u300d\u4e0e\u300c\u80fd\u6301\u7eed\u5b66\u300d\u4ecd\u6709\u8ddd\u79bb\u3002<\/p>\n
\u4e89\u8bae\u4e5f\u5728\u540c\u4e00\u5904\u3002\u56f4\u7ed5\u84b8\u998f\u7684\u4e89\u8bba\u57282026\u5e74\u6301\u7eed\u5347\u6e29\uff0c\u6709\u6cd5\u5f8b\u5b66\u8005\u64b0\u6587\u4e3b\u5f20\u7406\u6027\u770b\u5f85\uff0c\u6307\u51fa\u82f1\u4f1f\u8fbe\u4e0eMeta\u3001\u5fae\u8f6f\u3001\u6234\u5c14\u3001GitHub\u53caHugging Face\u7b49270\u4f59\u5bb6\u4f01\u4e1a\u548c\u7ec4\u7ec7\u5df2\u8054\u540d\u516c\u5f00\u4fe1\u4e3a\u8fd9\u9879\u6280\u672f\u6b63\u540d\uff1b\u4ea7\u4e1a\u4fa7\u4e5f\u6709\u4f01\u4e1a\u8d1f\u8d23\u4eba\u628a\u4f7f\u7528\u4e91\u7aef\u5927\u6a21\u578b\u7684\u6700\u5927\u98ce\u9669\u8bf4\u6210\u662f\u4f01\u4e1a\u6570\u636e\u4e0e\u7ecf\u9a8c\u88ab\u300c\u84b8\u998f\u300d\u3002\u672c\u6587\u7528\u81ea\u6709\u5bb6\u65cf\u66f4\u5927\u6a21\u578b\u751f\u6210\u6539\u5199\u6570\u636e\uff0c\u4e0e\u300c\u4ece\u4ed6\u65b9\u6a21\u578b\u8f93\u51fa\u6784\u9020\u6570\u636e\u300d\u7684\u5f62\u6001\u4e0d\u540c\uff0c\u4f46\u540c\u6837\u4f9d\u8d56\u300c\u66f4\u5927\u7684\u6a21\u578b\u66ff\u66f4\u5c0f\u7684\u6a21\u578b\u6539\u5199\u300d\u8fd9\u4e00\u524d\u63d0\u3002<\/p>\n
\u8bba\u6587\u5212\u51fa\u7684\u8fb9\u754c\u540c\u6837\u6e05\u695a\uff1a\u65b9\u6cd5\u9700\u8981\u4e00\u4e2a\u66f4\u5927\u7684\u540c\u5bb6\u65cf\u5bf9\u9f50\u6a21\u578b\uff0c\u624b\u4e0a\u6ca1\u6709\u540c\u7cfb\u5217\u7684\u5927\u4e00\u53f7\u7248\u672c\u5c31\u65e0\u4ece\u501f\u7528\uff1b\u4f5c\u8005\u628a\u300c\u628a\u5b66\u751f\u6a21\u578b\u81ea\u5df1\u7684\u7b54\u6848\u653e\u8fdb\u4e0a\u4e0b\u6587\u3001\u8ba9\u8de8\u5bb6\u65cf\u6a21\u578b\u751f\u6210\u66f4\u8d34\u8fd1\u5b66\u751f\u5206\u5e03\u7684\u6539\u5199\u300d\u5217\u4e3a\u5c1a\u672a\u9a8c\u8bc1\u7684\u65b9\u5411\uff1b\u533b\u5b66\u9886\u57df\u76f8\u5bf9\u79bb\u7b56\u7565\u4e0e\u5728\u7b56\u7565\u84b8\u998f\u6ca1\u6709\u660e\u663e\u4f18\u52bf\uff0c\u4f5c\u8005\u89e3\u91ca\u4e3a\u5bf9\u9f50\u6a21\u578b\u5df2\u638c\u63e1\u8bad\u7ec3\u6570\u636e\u3001\u793a\u8303\u6ce8\u5165\u7684\u8fb9\u9645\u4ef7\u503c\u6709\u9650\uff1b\u4e3b\u5b9e\u9a8c\u6bcf\u4e2a\u5355\u5143\u53ea\u8dd1\u4e86\u4e00\u4e2a\u968f\u673a\u79cd\u5b50\uff0c\u7a33\u5b9a\u6027\u5c1a\u672a\u5145\u5206\u8bc4\u4f30\u3002<\/p>\n
\u4e09\u4e2a\u6a21\u578b\u89c4\u6a21\u3001\u56db\u7c7b\u83b7\u53d6\u4efb\u52a1\u300114\u9879\u4fdd\u6301\u57fa\u51c6\u4e0a\u7684\u7ed3\u679c\u8bf4\u660e\uff0c\u5c0f\u6a21\u578b\u6301\u7eed\u540e\u8bad\u7ec3\u5df2\u4ece\u300c\u81ea\u5df1\u6559\u81ea\u5df1\u300d\u7684\u8bbe\u60f3\u53d8\u6210\u53ef\u590d\u73b0\u7684\u5de5\u7a0b\u6d41\u7a0b\u3002\u4f46\u843d\u5230\u5177\u4f53\u4ea7\u54c1\uff0c\u8fd8\u9700\u56de\u7b54\u540c\u5bb6\u65cf\u5bf9\u9f50\u6a21\u578b\u4ece\u54ea\u91cc\u6765\u3001\u62d2\u7edd\u91c7\u6837\u5728\u771f\u5b9e\u566a\u58f0\u6570\u636e\u4e0b\u8fd8\u5269\u591a\u5c11\u6536\u76ca\u3001\u628a\u6301\u7eed\u540e\u8bad\u7ec3\u7684\u5c0f\u6a21\u578b\u88c5\u8fdb\u7ec8\u7aef\u65f6\u6210\u672c\u4e0e\u529f\u8017\u600e\u4e48\u7b97\u2014\u2014\u8fd9\u4e09\u70b9\u5728\u73b0\u6709\u516c\u5f00\u6750\u6599\u91cc\u90fd\u6ca1\u6709\u7b54\u6848\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u97e9\u56fdLG AI Research\u4e0eKAI<\/p>\n","protected":false},"author":8,"featured_media":22412,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"colormag_page_container_layout":"default_layout","colormag_page_sidebar_layout":"default_layout","footnotes":""},"categories":[1],"tags":[],"class_list":["post-22413","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"acf":[],"_links":{"self":[{"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/posts\/22413","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/comments?post=22413"}],"version-history":[{"count":1,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/posts\/22413\/revisions"}],"predecessor-version":[{"id":22463,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/posts\/22413\/revisions\/22463"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/media\/22412"}],"wp:attachment":[{"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/media?parent=22413"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/categories?post=22413"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/dguan.com.cn\/ai\/wp-json\/wp\/v2\/tags?post=22413"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}