{"id":3372,"date":"2024-01-02T18:45:54","date_gmt":"2024-01-02T10:45:54","guid":{"rendered":"http:\/\/www.dama.org.cn\/wordpress\/?p=3372"},"modified":"2024-01-02T18:45:54","modified_gmt":"2024-01-02T10:45:54","slug":"%e5%a4%a7%e6%a8%a1%e5%9e%8b%e7%9f%a5%e8%af%86%e8%92%b8%e9%a6%8f%e6%a6%82%e8%bf%b0","status":"publish","type":"post","link":"https:\/\/www.dama.org.cn\/wordpress\/2024\/01\/02\/%e5%a4%a7%e6%a8%a1%e5%9e%8b%e7%9f%a5%e8%af%86%e8%92%b8%e9%a6%8f%e6%a6%82%e8%bf%b0\/","title":{"rendered":"\u5927\u6a21\u578b\u77e5\u8bc6\u84b8\u998f\u6982\u8ff0"},"content":{"rendered":"\n<p>\u5bfc\u8bed<\/p>\n\n\n\n<p>\u8fd1\u5e74\u6765\uff0c\u968f\u7740Transformer\u3001MOE\u67b6\u6784\u7684\u63d0\u51fa\uff0c\u4f7f\u5f97\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u8f7b\u677e\u7a81\u7834\u4e0a\u4e07\u4ebf\u89c4\u6a21\u53c2\u6570\uff0c\u4ece\u800c\u5bfc\u81f4\u6a21\u578b\u53d8\u5f97\u8d8a\u6765\u8d8a\u5927\uff0c\u56e0\u6b64\uff0c\u6211\u4eec\u9700\u8981\u4e00\u4e9b\u5927\u6a21\u578b\u538b\u7f29\u6280\u672f\u6765\u964d\u4f4e\u6a21\u578b\u90e8\u7f72\u7684\u6210\u672c\uff0c\u5e76\u63d0\u5347\u6a21\u578b\u7684\u63a8\u7406\u6027\u80fd\u3002\u800c\u5927\u6a21\u578b\u538b\u7f29\u4e3b\u8981\u5206\u4e3a\u5982\u4e0b\u51e0\u7c7b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u526a\u679d\uff08Pruning\uff09<\/li>\n\n\n\n<li>\u77e5\u8bc6\u84b8\u998f\uff08Knowledge Distillation\uff09<\/li>\n\n\n\n<li>\u91cf\u5316\uff08Quantization\uff09<\/li>\n\n\n\n<li>\u4f4e\u79e9\u5206\u89e3\uff08Low-Rank Factorization\uff09<\/li>\n<\/ul>\n\n\n\n<p>\u672c\u6587\u5c06\u8bb2\u8ff0\u5f53\u524d\u5927\u6a21\u578b\u84b8\u998f\u76f8\u5173\u7684\u4e00\u4e9b\u5de5\u4f5c\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<h4 class=\"wp-block-heading\">\u77e5\u8bc6\u84b8\u998f\u7b80\u4ecb<\/h4>\n\n\n\n<p>\u77e5\u8bc6\u84b8\u998f\u662f\u4e00\u79cd\u673a\u5668\u5b66\u4e60\u6a21\u578b\u538b\u7f29\u65b9\u6cd5\uff0c\u5b83\u7528\u4e8e\u5c06\u5927\u578b\u6a21\u578b\uff08\u6559\u5e08\u6a21\u578b\uff09\u7684\u77e5\u8bc6\u8fc1\u79fb\u5230\u8f83\u5c0f\u7684\u6a21\u578b\uff08\u5b66\u751f\u6a21\u578b\uff09\u4e2d\u3002<\/p>\n\n\n\n<p>\u77e5\u8bc6\u84b8\u998f\uff08KD\uff09\uff0c\u4e5f\u88ab\u79f0\u4e3a\u6559\u5e08-\u5b66\u751f\u795e\u7ecf\u7f51\u7edc\u5b66\u4e60\u7b97\u6cd5\uff0c\u662f\u4e00\u79cd\u6709\u4ef7\u503c\u7684\u673a\u5668\u5b66\u4e60\u6280\u672f\uff0c\u65e8\u5728\u63d0\u9ad8\u6a21\u578b\u6027\u80fd\u548c\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u5b83\u901a\u8fc7\u5c06\u77e5\u8bc6\u4ece\u590d\u6742\u7684\u6a21\u578b\uff08\u79f0\u4e3a\u6559\u5e08\u6a21\u578b\uff09\u8f6c\u79fb\u5230\u66f4\u7b80\u5355\u7684\u6a21\u578b\uff08\u79f0\u4e3a\u5b66\u751f\u6a21\u578b\uff09\u6765\u5b9e\u73b0\u8fd9\u4e00\u70b9\u3002KD\u80cc\u540e\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u6559\u5e08\u6a21\u578b\u7684\u7efc\u5408\u77e5\u8bc6\u8f6c\u5316\u4e3a\u66f4\u7cbe\u7b80\u3001\u66f4\u6709\u6548\u7684\u8868\u793a\u3002<\/p>\n\n\n\n<p>\u672c\u6587\uff0c\u6211\u4eec\u5c06\u6982\u8ff0\u5229\u7528LLM\u4f5c\u4e3a\u6559\u5e08\u7684\u84b8\u998f\u65b9\u6cd5\u3002\u6839\u636e\u8fd9\u4e9b\u65b9\u6cd5\u662f\u5426\u5c06LLM\u7684\u6d8c\u73b0\u80fd\u529b\uff08EA\uff09\u63d0\u70bc\u6210\u5c0f\u8bed\u8a00\u6a21\u578b\uff08SLM\uff09\u6765\u5bf9\u8fd9\u4e9b\u65b9\u6cd5\u8fdb\u884c\u5206\u7c7b\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5c06\u8fd9\u4e9b\u65b9\u6cd5\u5206\u4e3a\u4e24\u4e2a\u4e0d\u540c\u7684\u7c7b\u522b\uff1a\u6807\u51c6 KD \u548c\u57fa\u4e8e EA \u7684 KD\u3002\u4e3a\u4e86\u76f4\u89c2\u5730\u8868\u793a\uff0c\u4e0b\u56fe\u63d0\u4f9b\u4e86LLM\u77e5\u8bc6\u84b8\u998f\u7684\u7b80\u8981\u5206\u7c7b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1018\" height=\"722\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/ae114426212bdc0869721b4cbad97f09_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3373\" \/><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\">\u6807\u51c6\u77e5\u8bc6\u84b8\u998f<\/h4>\n\n\n\n<p>Standard KD\u65e8\u5728\u4f7f\u5b66\u751f\u6a21\u578b\u5b66\u4e60LLM\u6240\u62e5\u6709\u7684\u5e38\u89c1\u77e5\u8bc6\uff0c\u5982\u8f93\u51fa\u5206\u5e03\u548c\u7279\u5f81\u4fe1\u606f\u3002\u8fd9\u79cd\u65b9\u6cd5\u7c7b\u4f3c\u4e8e\u4f20\u7edf\u7684KD\uff0c\u4f46\u533a\u522b\u5728\u4e8e\u6559\u5e08\u6a21\u578b\u662fLLM\u3002\u6bd4\u5982\uff1aMINILLM \u548c GKD\u3002<\/p>\n\n\n\n<p>MINILLM \uff08\u8bba\u6587\uff1a<strong>Knowledge Distillation of Large Language Models<\/strong>\uff09\u6df1\u5165\u7814\u7a76\u4e86\u767d\u76d2\u751f\u6210LLM\u7684\u84b8\u998f\u3002\u5b83\u89c2\u5bdf\u5230\u6700\u5c0f\u5316\u524d\u5411 Kullback-Leibler \u6563\u5ea6 (KLD) \u7684\u6311\u6218\uff08\u8fd9\u53ef\u80fd\u4f1a\u5bfc\u81f4\u6559\u5e08\u5206\u5e03\u4e2d\u4e0d\u592a\u53ef\u80fd\u7684\u533a\u57df\u51fa\u73b0\u6982\u7387\u8fc7\u9ad8\uff0c\u4ece\u800c\u5728\u81ea\u7531\u8fd0\u884c\u751f\u6210\u8fc7\u7a0b\u4e2d\u5bfc\u81f4\u4e0d\u53ef\u80fd\u7684\u6837\u672c\uff09\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0cMINILLM \u9009\u62e9\u6700\u5c0f\u5316\u9006\u5411 KLD\u3002\u8fd9\u79cd\u65b9\u6cd5\u53ef\u4ee5<strong>\u9632\u6b62\u5b66\u751f\u9ad8\u4f30\u6559\u5e08\u5206\u5e03\u4e2d\u7684\u4f4e\u6982\u7387\u533a\u57df\uff0c\u4ece\u800c\u63d0\u9ad8\u751f\u6210\u6837\u672c\u7684\u8d28\u91cf<\/strong>\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>GitHub\uff1ahttps:\/\/github.com\/microsoft\/LMOps\/tree\/main\/minillm<\/li>\n\n\n\n<li>\u6559\u5e08\/\u5b66\u751f\u6a21\u578b\uff1aGPT2\u3001GPT-J\u3001OPT\u3001LLaMA<\/li>\n\n\n\n<li>\u6765\u6e90\uff1a\u6e05\u534e\u548c\u5fae\u8f6f\u7814\u7a76\u9662<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"350\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/f3abc2ed5feb5b5967b00687dae008be_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3374\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"483\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/16ba46b41e2d0fddee4d3cd6a4965e5c_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3375\" \/><\/figure>\n\n\n\n<p>GKD\uff08\u8bba\u6587\uff1a<strong>GKD: Generalized Knowledge Distillation for Auto-regressive Sequence Models<\/strong>\uff09 \u63a2\u7d22\u4e86\u81ea\u56de\u5f52\u6a21\u578b\u7684\u84b8\u998f\uff0c\u8fd9\u91cc\u767d\u76d2\u751f\u6210 LLM \u662f\u4e00\u4e2a\u5b50\u96c6\u3002\u8be5\u65b9\u6cd5\u786e\u5b9a\u4e86\u4e24\u4e2a\u5173\u952e\u95ee\u9898\uff1a\u8bad\u7ec3\u671f\u95f4\u7684\u8f93\u51fa\u5e8f\u5217\u4e0e\u5b66\u751f\u5728\u90e8\u7f72\u671f\u95f4\u751f\u6210\u7684\u8f93\u51fa\u5e8f\u5217\u4e4b\u95f4\u7684\u5206\u5e03\u4e0d\u5339\u914d\uff0c\u4ee5\u53ca\u6a21\u578bunder-specification\uff0c\u5176\u4e2d\u5b66\u751f\u6a21\u578b\u53ef\u80fd\u7f3a\u4e4f\u4e0e\u6559\u5e08\u5206\u5e03\u76f8\u5339\u914d\u7684\u8868\u8fbe\u80fd\u529b\u3002GKD \u901a\u8fc7\u5728\u8bad\u7ec3\u671f\u95f4\u5bf9\u5b66\u751f\u7684\u8f93\u51fa\u5e8f\u5217\u8fdb\u884c\u91c7\u6837\u6765\u5904\u7406\u5206\u5e03\u4e0d\u5339\u914d\uff0c\u5b83\u8fd8\u901a\u8fc7\u4f18\u5316\u66ff\u4ee3\u6563\u5ea6\uff08\u9006\u5411 KL\uff09\u6765\u89e3\u51b3\u6a21\u578bunder-specification\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6765\u6e90\uff1aGoogle DeepMind<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u57fa\u4e8e\u6d8c\u73b0\u80fd\u529b\u7684\u77e5\u8bc6\u84b8\u998f<\/h4>\n\n\n\n<p>\u57fa\u4e8e EA \u7684 KD \u4e0d\u4ec5\u4ec5\u8fc1\u79fb LLM \u7684\u5e38\u8bc6\uff0c\u8fd8\u5305\u62ec\u84b8\u998f\u4ed6\u4eec\u7684\u6d8c\u73b0\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u4e0e BERT\uff08330M\uff09\u548c GPT-2\uff081.5B\uff09\u7b49\u8f83\u5c0f\u6a21\u578b\u76f8\u6bd4\uff0cGPT-3\uff08175B\uff09\u548c PaLM\uff08540B\uff09\u7b49 LLM \u5c55\u793a\u4e86\u72ec\u7279\u7684\u884c\u4e3a\u3002\u8fd9\u4e9bLLM\u5728\u5904\u7406\u590d\u6742\u7684\u4efb\u52a1\u65f6\u8868\u73b0\u51fa\u4ee4\u4eba\u60ca\u8bb6\u7684\u80fd\u529b\uff0c\u79f0\u4e3a\u201c\u6d8c\u73b0\u80fd\u529b\u201d\u3002\u6d8c\u73b0\u80fd\u529b\u5305\u542b\u4e09\u4e2a\u65b9\u9762\uff0c\u5305\u62ec\u4e0a\u4e0b\u6587\u5b66\u4e60 (ICL)\u3001\u601d\u7ef4\u94fe (CoT) \u548c\u6307\u4ee4\u9075\u5faa (IF)\u3002\u5982\u56fe\u4e09\u6240\u793a\uff0c\u5b83\u63d0\u4f9b\u4e86\u57fa\u4e8eEA\u7684\u77e5\u8bc6\u84b8\u998f\u6982\u5ff5\u7684\u7b80\u660e\u8868\u793a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"970\" height=\"600\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/e6995ced7c5986aed3e4e544cab8e9b5_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3376\" \/><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\">\u4e0a\u4e0b\u6587\u5b66\u4e60\u84b8\u998f<\/h4>\n\n\n\n<p>ICL \u91c7\u7528\u7ed3\u6784\u5316\u81ea\u7136\u8bed\u8a00\u63d0\u793a\uff0c\u5176\u4e2d\u5305\u542b\u4efb\u52a1\u63cf\u8ff0\u4ee5\u53ca\u53ef\u80fd\u7684\u4e00\u4e9b\u4efb\u52a1\u793a\u4f8b\u4f5c\u4e3a\u6f14\u793a\u3002<strong>\u901a\u8fc7\u8fd9\u4e9b\u4efb\u52a1\u793a\u4f8b\uff0cLLM\u53ef\u4ee5\u638c\u63e1\u5e76\u6267\u884c\u65b0\u4efb\u52a1\uff0c\u800c\u65e0\u9700\u663e\u5f0f\u68af\u5ea6\u66f4\u65b0<\/strong>\u3002<\/p>\n\n\n\n<p><strong>In-context Learning Distillation: Transferring Few-shot Learning Ability of Pre-trained Language Models<\/strong>&nbsp;\u8bba\u6587\u4e2d\u5f15\u5165\u4e86 ICL \u84b8\u998f\uff0c\u5b83\u5c06\u4e0a\u4e0b\u6587\u5c0f\u6837\u672c\u5b66\u4e60\u548c\u8bed\u8a00\u5efa\u6a21\u529f\u80fd\u4ece LLM \u8f6c\u79fb\u5230 SLM\u3002\u8fd9\u662f\u901a\u8fc7\u5c06\u4e0a\u4e0b\u6587\u5b66\u4e60\u76ee\u6807\u4e0e\u4f20\u7edf\u8bed\u8a00\u5efa\u6a21\u76ee\u6807\u76f8\u7ed3\u5408\u6765\u5b9e\u73b0\u7684\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u4ed6\u4eec\u5728\u4e24\u79cd\u5c0f\u6837\u672c\u5b66\u4e60\u8303\u5f0f\u4e0b\u63a2\u7d22\u4e86 ICL \u84b8\u998f\uff1a\u5143\u4e0a\u4e0b\u6587\u8c03\u4f18 (Meta-ICT) \u548c\u591a\u4efb\u52a1\u4e0a\u4e0b\u6587\u8c03\u4f18 (Multitask-ICT)\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"454\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/5333c57a86336ddc3af83ca007bd9da4_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3377\" \/><\/figure>\n\n\n\n<p>\u5728 Meta-ICT \u4e2d\uff0c\u8bed\u8a00\u6a21\u578b\u4f7f\u7528\u4e0a\u4e0b\u6587\u5b66\u4e60 objectives \u5728\u4e0d\u540c\u4efb\u52a1\u4e2d\u8fdb\u884c\u5143\u8bad\u7ec3\u3002\u8fd9\u4f7f\u5176\u80fd\u591f\u901a\u8fc7\u4e0a\u4e0b\u6587\u5b66\u4e60\u6765\u9002\u5e94\u770b\u4e0d\u89c1\u7684\u4efb\u52a1\uff0c\u4ece\u800c\u6269\u5c55\u5176\u89e3\u51b3\u95ee\u9898\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u53e6\u4e00\u65b9\u9762\uff0cMultitask-ICT \u4f7f\u7528 ICL objectives \u548c target \u4efb\u52a1\u4e2d\u7684\u4e00\u4e9b\u793a\u4f8b\u5bf9\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u3002\u968f\u540e\uff0c\u5b83\u91c7\u7528\u4e0a\u4e0b\u6587\u5b66\u4e60\u6765\u5bf9\u8fd9\u4e9b\u4efb\u52a1\u8fdb\u884c\u9884\u6d4b\u3002<\/p>\n\n\n\n<p>\u6bd4\u8f83\u8fd9\u4e24\u79cd\u8303\u5f0f\uff0cMultitask-ICT \u8868\u73b0\u51fa\u4f18\u4e8e Meta-ICT \u7684\u6027\u80fd\u3002\u7136\u800c\uff0c\u5b83\u5728\u4efb\u52a1\u9002\u5e94\u8fc7\u7a0b\u4e2d\u786e\u5b9e\u9700\u8981\u66f4\u591a\u7684\u8ba1\u7b97\u8d44\u6e90\uff0c\u4f7f\u5176\u8ba1\u7b97\u52a0\u5f3a\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6765\u6e90\uff1a\u54e5\u4f26\u6bd4\u4e9a\u5927\u5b66<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">\u601d\u7ef4\u94fe\u84b8\u998f<\/h4>\n\n\n\n<p>\u4e0e ICL \u76f8\u6bd4\uff0cCoT \u91c7\u7528\u4e86\u4e0d\u540c\u7684\u65b9\u6cd5\uff0c\u5b83\u5c06\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\uff08\u53ef\u4ee5\u5bfc\u81f4\u6700\u7ec8\u8f93\u51fa\uff09\u5408\u5e76\u5230\u63d0\u793a\u4e2d\uff0c\u800c\u4e0d\u662f\u4f7f\u7528\u7b80\u5355\u7684\u8f93\u5165\u8f93\u51fa\u5bf9\u3002<\/p>\n\n\n\n<p>MT-COT \uff08\u8bba\u6587\uff1a<strong>Explanations from Large Language Models Make Small Reasoners Better<\/strong>\uff09 \u65e8\u5728\u5229\u7528 LLM \u4ea7\u751f\u7684\u89e3\u91ca\u6765\u52a0\u5f3a\u5c0f\u578b\u63a8\u7406\u673a\u7684\u8bad\u7ec3\u3002\u5b83\u5229\u7528\u591a\u4efb\u52a1\u5b66\u4e60\u6846\u67b6\u4f7f\u8f83\u5c0f\u7684\u6a21\u578b\u5177\u6709\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\u4ee5\u53ca\u751f\u6210\u89e3\u91ca\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6765\u6e90\uff1a\u52a0\u5dde\u5927\u5b66\u5723\u5854\u82ad\u82ad\u62c9\u5206\u6821\u3001\u817e\u8baf\u4eba\u5de5\u667a\u80fd\u5b9e\u9a8c\u5ba4\u3001\u5fae\u8f6f\u7b49<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"211\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/ad43fa006de53d9bbeac074f450c2a3d_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3378\" \/><\/figure>\n\n\n\n<p>Fine-tune CoT \uff08\u8bba\u6587\uff1a<strong>Large language models are reasoning teachers<\/strong>\uff09\u66f4\u8fdb\u4e00\u6b65\uff0c\u901a\u8fc7\u968f\u673a\u91c7\u6837\u4ece LLM \u751f\u6210\u591a\u4e2a\u63a8\u7406\u89e3\u51b3\u65b9\u6848\u3002\u8bad\u7ec3\u6570\u636e\u7684\u589e\u5f3a\u6709\u52a9\u4e8e\u5b66\u751f\u6a21\u578b\u7684\u5b66\u4e60\u8fc7\u7a0b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>GitHub\uff1ahttps:\/\/github.com\/itsnamgyu\/reasoning-teacher<\/li>\n\n\n\n<li>\u6765\u6e90\uff1a\u97e9\u56fd\u79d1\u5b66\u6280\u672f\u9662<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"738\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/9534c3ebb4e43eeb17eb23666ff62e3f_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3379\" \/><\/figure>\n\n\n\n<p>Fu \u7b49\u4eba\u7684\u7814\u7a76\uff08\u8bba\u6587\uff1a<strong>Specializing Smaller Language Models towards Multi-Step Reasoning<\/strong>\uff09\u53d1\u73b0\u8bed\u8a00\u6a21\u578b\u591a\u7ef4\u80fd\u529b\u4e4b\u95f4\u7684\u6743\u8861\uff0c\u5e76\u63d0\u51fa\u5fae\u8c03\u6307\u4ee4\u8c03\u6574\u6a21\u578b\u3002\u4ed6\u4eec<strong>\u4ece\u5927\u578b\u6559\u5e08\u6a21\u578b\u4e2d\u63d0\u53d6 CoT \u63a8\u7406\u8def\u5f84\uff0c\u4ee5\u63d0\u9ad8\u5206\u5e03\u5916\u6cdb\u5316\u80fd\u529b<\/strong>\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6765\u6e90\uff1a\u7231\u4e01\u5821\u5927\u5b66\u3001\u827e\u4f26AI\u7814\u7a76\u6240<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"545\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/48a0eace2e2155fa88d617766b52ca69_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3380\" \/><\/figure>\n\n\n\n<p>Hsieh \u7b49\u4eba\u7684\u7814\u7a76\uff08\u8bba\u6587\uff1a<strong>Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes<\/strong>\uff09\u4f7f\u7528 LLM \u8bba\u636e\u4f5c\u4e3a\u5728\u591a\u4efb\u52a1\u6846\u67b6\u5185\u8bad\u7ec3\u8f83\u5c0f\u6a21\u578b\u7684\u989d\u5916\u6307\u5bfc\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6765\u6e90\uff1a\u534e\u76db\u987f\u5927\u5b66\u3001\u8c37\u6b4c<\/li>\n\n\n\n<li>GitHub\uff1ahttps:\/\/github.com\/google-research\/distilling-step-by-step<\/li>\n\n\n\n<li>\u5b66\u751f\u6a21\u578b\uff1aT5<\/li>\n\n\n\n<li>\u6559\u5e08\u6a21\u578b\uff1aPaLM<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"560\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/40c3681fb5623ed145e668592d888b88_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3381\" style=\"width:690px;height:auto\" \/><\/figure>\n\n\n\n<p>SOCRATIC CoT\uff08\u8bba\u6587\uff1a<strong>Distilling Reasoning Capabilities into Smaller Language Models<\/strong>\uff09\u8bad\u7ec3\u4e24\u4e2a\u84b8\u998f\u6a21\u578b\uff1a\u95ee\u9898\u5206\u89e3\u5668\u548c\u5b50\u95ee\u9898\u6c42\u89e3\u5668\u3002\u5206\u89e3\u5668\u5c06\u539f\u59cb\u95ee\u9898\u5206\u89e3\u4e3a\u4e00\u7cfb\u5217\u5b50\u95ee\u9898\uff0c\u800c\u5b50\u95ee\u9898\u6c42\u89e3\u5668\u8d1f\u8d23\u89e3\u51b3\u8fd9\u4e9b\u5b50\u95ee\u9898\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6765\u6e90\uff1a\u82cf\u9ece\u4e16\u8054\u90a6\u7406\u5de5\u5b66\u9662\u8ba1\u7b97\u673a\u79d1\u5b66\u7cfb<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"787\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/f1846f307c0535bfa6faf9b9e452054d_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3382\" \/><\/figure>\n\n\n\n<p>DISCO\uff08\u8bba\u6587\uff1a<strong>DISCO: Distilling Counterfactuals with Large Language Models<\/strong>\uff09\u5f15\u5165\u4e86\u4e00\u79cd\u57fa\u4e8e LLM \u7684<strong>\u5168\u81ea\u52a8\u53cd\u4e8b\u5b9e<\/strong>\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\u3002\u5b83\u901a\u8fc7\u5de5\u7a0b\u5316\u7684\u63d0\u793a\u4f7f\u7528 LLM \u751f\u6210\u77ed\u8bed\u6270\u52a8\uff0c\u7136\u540e\u901a\u8fc7\u7279\u5b9a\u4e8e\u4efb\u52a1\u7684\u6559\u5e08\u6a21\u578b\u8fc7\u6ee4\u8fd9\u4e9b\u6270\u52a8\uff0c\u4ee5\u63d0\u53d6\u9ad8\u8d28\u91cf\u7684\u53cd\u4e8b\u5b9e\u6570\u636e\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6765\u6e90\uff1a\u6d1b\u6851\u8054\u90a6\u7406\u5de5\u5b66\u9662\u81ea\u7136\u8bed\u8a00\u5904\u7406\u5b9e\u9a8c\u5ba4\u3001\u827e\u4f26\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u6240<\/li>\n\n\n\n<li>GitHub\uff1ahttps:\/\/github.com\/eric11eca\/disco<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"972\" height=\"1164\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/0697e9921ff741b1be487225dad01048_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3383\" \/><\/figure>\n\n\n\n<p>\u4e3a\u4e86\u5f62\u6210\u66f4\u597d\u7684\u76d1\u7763\uff0cSCOTT\uff08\u8bba\u6587\uff1a<strong>SCOTT: Self-Consistent Chain-of-Thought Distillation<\/strong>\uff09\u901a\u8fc7<strong>\u5bf9\u6bd4\u89e3\u7801<\/strong>\u4ece\u5927\u578b LM\uff08\u8001\u5e08\uff09\u90a3\u91cc\u5f97\u5230\u652f\u6301\u6807\u51c6\u7b54\u6848\u7684\u8bba\u636e\uff0c\u8fd9\u9f13\u52b1\u8001\u5e08\u751f\u6210\u53ea\u6709\u5728\u8003\u8651\u7b54\u6848\u65f6\u624d\u53d8\u5f97\u66f4\u52a0\u53ef\u4fe1\u7684Token\u3002\u4e3a\u4e86\u786e\u4fdd\u53ef\u4fe1\u7684\u84b8\u998f\uff0c\u6211\u4eec\u4f7f\u7528\u6559\u5e08\u751f\u6210\u7684\u57fa\u672c\u8bba\u636e\u6765\u5b66\u4e60\u5177\u6709\u53cd\u4e8b\u5b9e\u63a8\u7406\u76ee\u6807\u7684\u5b66\u751f LM\uff0c\u8fd9\u53ef\u4ee5\u9632\u6b62\u5b66\u751f\u5ffd\u7565\u57fa\u672c\u8bba\u636e\u800c\u505a\u51fa\u4e0d\u4e00\u81f4\u7684\u9884\u6d4b\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6765\u6e90\uff1a\u5357\u52a0\u5dde\u5927\u5b66\u8ba1\u7b97\u673a\u79d1\u5b66\u7cfb\uff0c\u4e9a\u9a6c\u900a\u516c\u53f8<\/li>\n\n\n\n<li>GitHub\uff1ahttps:\/\/github.com\/wangpf3\/consistent-CoT-distillation<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"428\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/819b12b76d7ff9d378ff11d56a37a39b_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3384\" \/><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\">\u6307\u4ee4\u9075\u5faa\u84b8\u998f<\/h4>\n\n\n\n<p>IF \u81f4\u529b\u4e8e\u4ec5\u57fa\u4e8e\u9605\u8bfb\u4efb\u52a1\u63cf\u8ff0\u6765\u589e\u5f3a\u8bed\u8a00\u6a21\u578b\u6267\u884c\u65b0\u4efb\u52a1\u7684\u80fd\u529b\uff0c\u800c\u4e0d\u4f9d\u8d56\u4e8e\u5c11\u6570\u6837\u672c\u3002\u901a\u8fc7\u4f7f\u7528\u4e00\u7cfb\u5217\u4ee5\u6307\u4ee4\u8868\u793a\u7684\u4efb\u52a1\u8fdb\u884c\u5fae\u8c03\uff0c\u8bed\u8a00\u6a21\u578b\u5c55\u793a\u4e86\u51c6\u786e\u6267\u884c\u4ee5\u524d\u672a\u89c1\u8fc7\u7684\u6307\u4ee4\u4e2d\u63cf\u8ff0\u7684\u4efb\u52a1\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u4f8b\u5982\uff0cLion \uff08\u8bba\u6587\uff1a<strong>Lion: Adversarial Distillation of Closed-Source Large Language Model<\/strong>\uff09\u5229\u7528LLM\u7684\u9002\u5e94\u6027\u5f3a\u7684\u7279\u70b9\u6765\u63d0\u9ad8\u5b66\u751f\u6a21\u578b\u7684\u8868\u73b0\u3002\u5b83\u63d0\u793aLLM\u8bc6\u522b\u5e76\u751f\u6210\u201chard\u201d\u6307\u4ee4\uff0c\u7136\u540e\u5229\u7528\u8fd9\u4e9b\u6307\u4ee4\u6765\u589e\u5f3a\u5b66\u751f\u6a21\u578b\u7684\u80fd\u529b\u3002\u8fd9\u79cd\u65b9\u6cd5\u5229\u7528\u4e86LLM\u7528\u9014\u5e7f\u6cdb\u7684\u7279\u6027\u6765\u6307\u5bfc\u5b66\u751f\u6a21\u578b\u7684\u5b66\u4e60\uff0c\u4ee5\u89e3\u51b3\u590d\u6742\u7684\u6307\u4ee4\u548c\u4efb\u52a1\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6765\u6e90\uff1a\u9999\u6e2f\u79d1\u6280\u5927\u5b66<\/li>\n\n\n\n<li>GitHub\uff1ahttps:\/\/github.com\/YJiangcm\/Lion<\/li>\n\n\n\n<li>\u5b66\u751f\u6a21\u578b\uff1aLLaMA-7B<\/li>\n\n\n\n<li>\u6559\u5e08\u6a21\u578b\uff1aChatGPT(GPT-3.5)<\/li>\n<\/ul>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"607\" src=\"http:\/\/www.dama.org.cn\/wordpress\/wp-content\/uploads\/2024\/01\/868cf8e25a1e7bd8443ddfcf3af1015f_640_wx_fmtpngwxfrom5wx_lazy1wx_co1.png\" alt=\"\" class=\"wp-image-3385\" \/><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\">\u7ed3\u8bed<\/h4>\n\n\n\n<p>\u672c\u6587\u7b80\u5355\u8bb2\u8ff0\u4e86\u76ee\u524d\u5927\u6a21\u578b\u77e5\u8bc6\u84b8\u998f\u65b9\u5411\u7684\u4e00\u4e9b\u5de5\u4f5c\uff0c\u5728\u5927\u6a21\u578b\u53c2\u6570\u91cf\u8d8a\u6765\u8d8a\u5927\u7684\u4eca\u5929\uff0c\u8be5\u65b9\u5411\uff08\u901a\u8fc7\u76f8\u5bf9\u8f83\u5c0f\u91cf\u7ea7\u7684\u6a21\u578b\u84b8\u998f\u5927\u6a21\u578b\u7684\u77e5\u8bc6\uff09\u7684\u7814\u7a76\u663e\u5f97\u683c\u5916\u6709\u610f\u4e49\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<p>\u6587\u7ae0\u6765\u6e90\uff1a\u5403\u679c\u51bb\u4e0d\u5410\u679c\u51bb\u76ae\u5fae\u4fe1\u516c\u4f17\u53f7<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5bfc\u8bed \u8fd1\u5e74\u6765\uff0c\u968f\u7740Transformer &hellip;<\/p>\n","protected":false},"author":4,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[24],"tags":[],"class_list":["post-3372","post","type-post","status-publish","format-standard","hentry","category-24"],"_links":{"self":[{"href":"https:\/\/www.dama.org.cn\/wordpress\/wp-json\/wp\/v2\/posts\/3372","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.dama.org.cn\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.dama.org.cn\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.dama.org.cn\/wordpress\/wp-json\/wp\/v2\/users\/4"}],"replies":[{"embeddable":true,"href":"https:\/\/www.dama.org.cn\/wordpress\/wp-json\/wp\/v2\/comments?post=3372"}],"version-history":[{"count":1,"href":"https:\/\/www.dama.org.cn\/wordpress\/wp-json\/wp\/v2\/posts\/3372\/revisions"}],"predecessor-version":[{"id":3386,"href":"https:\/\/www.dama.org.cn\/wordpress\/wp-json\/wp\/v2\/posts\/3372\/revisions\/3386"}],"wp:attachment":[{"href":"https:\/\/www.dama.org.cn\/wordpress\/wp-json\/wp\/v2\/media?parent=3372"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.dama.org.cn\/wordpress\/wp-json\/wp\/v2\/categories?post=3372"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.dama.org.cn\/wordpress\/wp-json\/wp\/v2\/tags?post=3372"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}