{"id":663,"date":"2024-03-06T15:22:01","date_gmt":"2024-03-06T07:22:01","guid":{"rendered":"http:\/\/lingbo.online\/?p=663"},"modified":"2024-03-06T15:23:37","modified_gmt":"2024-03-06T07:23:37","slug":"cot-finetune","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/algorithm_learning\/llms\/cot-finetune\/","title":{"rendered":"[\u8bba\u6587\u7b14\u8bb0]Large Language Models Are Reasoning Teachers"},"content":{"rendered":"<h1>MetaData<\/h1>\n<ul>\n<li>\u53d1\u8868\u65e5\u671f\uff1a2023\u5e741\u670813\u65e5<\/li>\n<li>\u4f5c\u8005\uff1aNamgyu Ho, Laura Schmid, Se-Young Yun<\/li>\n<li>\u53d1\u8868\u5e73\u53f0\uff1aarXiv<\/li>\n<li>\u8bba\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/arxiv.org\/pdf\/2212.10071.pdf\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/arxiv.org\/pdf\/2212.10071.pdf<\/a><\/li>\n<li>\u9879\u76ee\u94fe\u63a5\uff1a<a href=\"https:\/\/github.com\/itsnamgyu\/reasoning-teacher\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/itsnamgyu\/reasoning-teacher<\/a><br \/>\n<h1>\u6458\u8981<\/h1>\n<p>\u6700\u8fd1\u7684\u7814\u7a76\u8868\u660e\uff0c\u601d\u7ef4\u94fe(CoT)\u63d0\u793a\u53ef\u4ee5\u5f15\u5bfc\u5927\u6a21\u578b\uff0c\u9010\u6b65\u89e3\u51b3\u590d\u6742\u7684\u63a8\u7406\u4efb\u52a1\u3002\u7136\u800c\uff0c\u57fa\u4e8e\u63d0\u793a\u8bcd\u7684CoT\u65b9\u6cd5\u4f9d\u8d56\u4e8e\u53c2\u6570\u975e\u5e38\u5927\u7684\u6a21\u578b\uff0c\u5982GPT-3 175B\uff0c\u8fd9\u662f\u5f88\u96be\u5927\u89c4\u6a21\u90e8\u7f72\u7684\u3002<br \/>\n\u5728\u672c\u6587\u4e2d\uff0c\u4f7f\u7528\u8fd9\u4e9b\u5927\u6a21\u578b\u4f5c\u4e3a\u63a8\u7406\u8001\u5e08\uff08reasoning teachers\uff09\uff0c\u5728\u8f83\u5c0f\u6a21\u578b\u4e2d\u5b9e\u73b0\u590d\u6742\u63a8\u7406\uff0c\u5e76\u5c06\u6a21\u578b\u5c3a\u5bf8\u8981\u6c42\u4e0b\u964d\u51e0\u4e2a\u6570\u91cf\u7ea7\u3002\u6587\u7ae0\u63d0\u51fa\u4e86fine-tune-CoT\uff0c\u8fd9\u662f\u4e00\u4e2a\u4ece\u975e\u5e38\u5927\u7684\u5728\u6559\u5ba4\u6a21\u578b\u751f\u6210\u63a8\u7406\u6837\u672c\u4ee5\u5fae\u8c03\u8f83\u5c0f\u6a21\u578b\u7684\u65b9\u6cd5\u3002\u6587\u7ae0\u5728\u5e7f\u6cdb\u7684\u516c\u5171\u6a21\u578b\u548c\u590d\u6742\u4efb\u52a1\u4e0a\u8bc4\u4ef7\u63d0\u51fa\u7684\u65b9\u6cd5\u3002\u7ed3\u679c\u8868\u660e\uff0cFine-tuneCoT\u5728\u5c0f\u578b\u6a21\u578b\u4e2d\u5b9e\u73b0\u4e86\u5927\u91cf\u7684\u63a8\u7406\u80fd\u529b\uff0c\u5728\u8bb8\u591a\u4efb\u52a1\u91cd\u8fdc\u8fdc\u4f18\u4e8e\u57fa\u4e8e\u63d0\u793a\u8bcd\u7684baseline\uff0c\u751a\u81f3\u4f18\u4e8e\u6559\u5e08\u6a21\u578b\u3002\u6b64\u5916\uff0c\u5229\u7528\u6559\u5e08\u6a21\u578b\u7684\u80fd\u529b\u62d3\u5c55\u6211\u4eec\u7684\u65b9\u6cd5\uff0c\u4e3a\u6bcf\u4e2a\u539f\u59cb\u6837\u672c\u751f\u6210\u591a\u4e2a\u4e0d\u540c\u7684\u57fa\u672c\u63a8\u7406\u3002\u4f7f\u7528\u591a\u6837\u5316\u7684\u63a8\u7406\u6765\u4e30\u5bcc\u5fae\u8c03\u6570\u636e\uff0c\u53ef\u4ee5\u5728\u6570\u636e\u96c6\u4e0a\u5927\u5e45\u5ea6\u63d0\u5347\u6027\u80fd\uff0c\u5373\u4fbf\u5bf9\u4e8e\u975e\u5e38\u5c0f\u7684\u6a21\u578b\u4e5f\u662f\u5982\u6b64\u3002\u6211\u4eec\u8fdb\u884c\u4e86\u6d88\u878d\uff08ablations\uff09\u548c\u6837\u672c\u7814\u7a76\uff0c\u4ee5\u4e86\u89e3\u5b66\u751f\u6a21\u578b\u63a8\u7406\u80fd\u529b\u7684\u8868\u73b0\u3002<\/p>\n<\/li>\n<\/ul>\n<h1>\u5f15\u8a00<\/h1>\n<p>\u867d\u7136\u5927\u8bed\u8a00\u6a21\u578b\u5df2\u7ecf\u5c55\u793a\u51fa\u4e0a\u4e0b\u6587\u6cdb\u5316\u80fd\u529b\uff0c\u4f46\u662f\u6700\u5927\u7684\u5927\u8bed\u8a00\u6a21\u578b\u4e5f\u96be\u4ee5\u5904\u7406\u9700\u8981\u591a\u4e2a\u63a8\u7406\u6b65\u9aa4\u7684\u590d\u6742\u4efb\u52a1\u3002\u4e3a\u4e86\u89e3\u51b3\u590d\u6742\u7684\u4efb\u52a1\uff0c\u6700\u8fd1\u7684\u7814\u7a76\u8868\u660e\uff0c\u4f7f\u7528\u601d\u7ef4\u94fe\uff08chain-of-thought CoT\uff09\u7684\u63d0\u793a\u8bcd\u65b9\u6cd5\uff0c\u751f\u6210\u4e00\u7cfb\u5217\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\uff0c\u53ef\u4ee5\u6fc0\u53d1\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002<br \/>\n\u7136\u800c\uff0c\u57fa\u4e8e\u63d0\u793a\u8bcd\u7684CoT\u63a8\u7406\u65b9\u6cd5\u7684\u4e3b\u8981\u7f3a\u70b9\u662f\u4f9d\u8d56\u4e8e\u8d85\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u8fd9\u4e9b\u6a21\u578b\u96be\u4ee5\u5927\u89c4\u6a21\u90e8\u7f72\u3002\u56e0\u6b64\uff0c\u6587\u7ae0\u52aa\u529b\u4f7f\u66f4\u9002\u5408\u5927\u89c4\u6a21\u90e8\u7f72\u7684\u5c0f\u6a21\u578b\u80fd\u591f\u8fdb\u884c\u590d\u6742\u63a8\u7406\u3002<br \/>\n\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6587\u7ae0\u63d0\u51fa\u4e86Fine-tune-CoT\u7684\u65b9\u6cd5\uff0c\u5229\u7528\u8d85\u5927\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u6765\u6559\u5c0f\u6a21\u578b\u89e3\u51b3\u590d\u6742\u4efb\u52a1\u3002\u5e94\u7528\u73b0\u6709\u7684zero-shot CoT\u63d0\u793a\u8bcd\u4ece\u8d85\u5927\u6559\u5e08\u6a21\u578b\u4e2d\u751f\u6210\u57fa\u672c\u63a8\u7406\uff0c\u5e76\u4f7f\u7528\u5b83\u4eec\u5bf9\u8f83\u5c0f\u7684\u5b66\u751f\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u6d41\u7a0b\u5982\u4e0b\u56fe\u6240\u793a\u3002\u4f5c\u8005\u6ce8\u610f\u5230\uff0c\u6ca1\u6709\u63a8\u7406\u7684\u6807\u51c6\u5fae\u8c03\u5df2\u88ab\u8bc1\u660e\u4e0d\u8db3\u4ee5\u89e3\u51b3\u5c0f\u578b\u6a21\u578b\u7684\u63a8\u7406\u4efb\u52a1\u3002\u867d\u7136\u6709\u4eba\u5c1d\u8bd5\u7528\u624b\u5de5\u6807\u6ce8\u7684\u63a8\u7406\u6b65\u9aa4\u8fdb\u884c\u5c0f\u6a21\u578bSFT\uff0c\u4ed6\u4eec\u9700\u8981\u7279\u5b9a\u4efb\u52a1\u7684\u8bad\u7ec3\u8bbe\u7f6e\u548c\u9ad8\u8d28\u91cf\u7684\u63a8\u7406\uff0c\u6ce8\u91ca\u6210\u672c\u5f88\u9ad8\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u4f5c\u8005\u63d0\u51fa\u7684\u65b9\u6cd5\u53ef\u4ee5\u5f88\u5bb9\u6613\u7684\u5e94\u7528\u4e8e\u65b0\u7684\u4e0b\u6e38\u4efb\u52a1\uff0c\u800c\u4e0d\u9700\u8981\u624b\u5de5\u5236\u4f5c\u63a8\u7406\u6216\u8005\u4efb\u52a1\u5de5\u7a0b\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/lingbo.online\/wp-content\/uploads\/2024\/03\/cot_finetune_fg1.jpg\" alt=\"\" \/><br \/>\n\u4f5c\u8005\u8fd8\u63d0\u51fa\u4e86\u4e00\u4e2a\u65b0\u7684\u62d3\u5c55\u65b9\u6cd5\uff0c\u57ce\u4e3a\u591a\u6837\u5316\u63a8\u7406\uff08diverse reasoning\uff09\uff0c\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8SFT-CoT\u7684\u6559\u5b66\u6548\u679c\u3002\u76f4\u89c9\u4e0a\u6765\u8bf4\uff0c\u590d\u6742\u7684\u4efb\u52a1\u53ef\u4ee5\u5177\u6709\u591a\u4e2a\u4e0d\u540c\u63a8\u7406\u8def\u5f84\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u53ef\u4ee5\u4f7f\u7528\u968f\u673a\u62bd\u6837\u4ece\u6559\u5e08\u6a21\u578b\u4e2d\u58f0\u6d4b\u4f1a\u7ed9\u4f60\u591a\u4e2a\u63a8\u7406\u89e3\u51b3\u65b9\u6848\uff0c\u4ee5\u589e\u5f3a\u5b66\u751f\u6a21\u578b\u7684\u8bad\u7ec3\u6570\u636e\u3002\u8fd9\u79cd\u65b9\u6cd5\u7b80\u5355\u800c\u9ad8\u6548\uff0c\u53ef\u4ee5\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8\u5b66\u751f\u6210\u7ee9\uff0c\u4f46\u5728\u4f7f\u7528CoT\u63a8\u7406\u8fdb\u884c\u5fae\u8c03\u7684\u5e76\u884c\u5de5\u4f5c\u4e2d\uff0c\u8fd9\u4e00\u70b9\u5c1a\u672a\u5f97\u5230\u660e\u786e\u7684\u8ba4\u53ef\u3002<br \/>\n\u6211\u4eec\u4f7f\u7528\u5e7f\u6cdb\u7684\u516c\u5f00\u53ef\u7528\u6a21\u578b\u572812\u4e2a\u4efb\u52a1\u4e0a\u8bc4\u4f30\u6211\u4eec\u7684\u65b9\u6cd5\u3002\u6211\u4eec\u53d1\u73b0Fine-tune-CoT\u53ef\u4ee5\u5728\u5c0f\u6a21\u578b\u4e2d\u83b7\u5f97\u663e\u8457\u7684\u63a8\u7406\u6027\u80fd\uff0c\u540c\u65f6\u4fdd\u7559\u4e86\u57fa\u4e8e\u63d0\u793a\u7684CoT\u63a8\u7406\u7684\u5927\u90e8\u5206\u901a\u7528\u6027\uff0c\u800c\u4ee5\u524d\u9700\u8981&gt;100B\u4e2a\u53c2\u6570\u6a21\u578b(Wei et al, 2022b)\u3002\u901a\u8fc7\u5229\u7528\u6211\u4eec\u72ec\u7279\u7684\u5b66\u4e60\u8bbe\u7f6e\uff0c\u591a\u6837\u5316\u7684\u63a8\u7406\u53ef\u4ee5\u5728\u5f00\u53d1\u65f6\u4ee5\u989d\u5916\u7684\u6559\u5e08\u63a8\u7406\u4e3a\u4ee3\u4ef7\u83b7\u5f97\u663e\u7740\u7684\u6027\u80fd\u6536\u76ca\u3002\u8fd9\u4f7f\u5f97\u5c0f\u52300.3B\u7684\u6a21\u578b\u5728\u67d0\u4e9b\u4efb\u52a1\u4e2d\u8868\u73b0\u4f18\u4e8e\u8f83\u5927\u7684\u5b66\u751f\uff0c\u751a\u81f3\u8d85\u8fc7175B\u7684\u6559\u5e08\u6a21\u578b\u3002\u6211\u4eec\u7684\u6d88\u878d\u8868\u660e\uff0c\u6027\u80fd\u5728\u6240\u6709\u8003\u8651\u7684\u8f74\u4e0a\u90fd\u662f\u4e00\u81f4\u7684\u53ef\u6269\u5c55:\u4e0d\u540c\u7684\u63a8\u7406\u3001\u6570\u636e\u96c6\u5927\u5c0f\u3001\u6559\u5e08\u8868\u73b0\u548c\u5b66\u751f\u6a21\u578b\u5927\u5c0f\u3002\u8fd9\u663e\u793a\u4e86\u6211\u4eec\u7684\u65b9\u6cd5\u5728\u5c0f\u578b\u6a21\u578b\u4e2d\u5b9e\u73b0\u53ef\u9760\u6027\u80fd\u7684\u6f5c\u529b\uff0c\u8fd9\u4e9b\u6a21\u578b\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u662f\u53ef\u884c\u7684\u3002\u6700\u540e\uff0c\u6211\u4eec\u8fdb\u884c\u4e86\u5f7b\u5e95\u7684\u6837\u672c\u7814\u7a76\u548c\u5206\u6790\uff0c\u63ed\u793a\u4e86\u4ee5\u524d\u5728CoT\u5fae\u8c03\u4e2d\u5ffd\u7565\u7684\u5173\u952e\u7ec6\u8282\uff0c\u5e76\u63d0\u4f9b\u4e86\u5c0f\u6a21\u578b\u4e2d\u63a8\u7406\u80fd\u529b\u51fa\u73b0\u7684\u76f4\u89c9\u3002<\/p>\n<h1>\u76f8\u5173\u5de5\u4f5c<\/h1>\n<ul>\n<li><strong>\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u4e0b\u6e38\u8fc1\u79fb\uff1a<\/strong>\u4e4b\u524d\u7684\u8bb8\u591a\u5de5\u4f5c\u5efa\u7acb\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u8303\u5f0f\uff0c\u7528\u4e8e\u63d0\u9ad8LLM\u5728\u4e0b\u6e38\u4efb\u52a1\u7684\u6027\u80fd\u3002\u7136\u800c\uff0c\u5fae\u8c03\u5e76\u4e0d\u603b\u662f\u5bb9\u6613\u88ab\u63a5\u53d7\u7684\u3002\u6700\u8fd1\u7684\u6587\u732e\u663e\u793a\u51fa\u4e00\u79cd\u8303\u5f0f\u7816\u9762\uff0c\u5373\u63d0\u793a\u6a21\u578b\u9884\u6d4b\u671f\u671b\u8f93\u51fa\u3002LLMs\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\u8868\u73b0\u51fa\u5f3a\u5927\u6027\u80fd\u3002\u5bf9\u4e8e\u80fd\u591f\u6267\u884c\u7c7b\u4f3c\u64cd\u4f5c\u7684\u5c0f\u578b\u6a21\u578b\uff0c\u901a\u5e38\u9700\u8981\u989d\u5916\u7684\u5de5\u7a0b\u3002\u5bf9\u4e8e\u66f4\u590d\u6742\u7684\u4efb\u52a1\uff0c\u4f7f\u7528\u5e26\u6709\u660e\u786e\u63a8\u7406\u6b65\u9aa4\u7684\u6837\u672c\u6765\u5fae\u8c03\u6a21\u578b\u7684\u60f3\u6cd5\u5148\u4e8e\u601d\u7ef4\u94fe\uff08CoT\uff09\u63d0\u793a\u65b9\u6cd5\uff0c\u5728\u975e\u5e38\u5927\u7684LLMs\u4e2d\u80fd\u591f\u8868\u73b0\u826f\u597d\u3002<\/li>\n<li>CoT\u63a8\u7406\uff1a\u5728few-shot\u7684CoT\u63d0\u793a\u8bcd\u4e2d\uff0c\u6a21\u578b\u5f97\u5230\u9010\u6b65\u63a8\u7406\u7684\u6848\u4f8b\u540e\uff0c\u5b66\u4f1a\u751f\u6210\u95ee\u9898\u89e3\u51b3\u65b9\u6848\u7684\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u3002\u8fd9\u5728\u5404\u79cd\u4efb\u52a1\u4e0a\u90fd\u6709\u5f88\u597d\u5730\u8868\u73b0\u3002\u6b64\u5916\uff0cllm\u53ef\u4ee5\u5728\u65e0\u76d1\u7763\u7684\u4efb\u52a1\u4e2d\u4f7f\u7528Zero-shot-CoT\u83b7\u5f97\u826f\u597d\u7684\u8868\u73b0\u3002\u8fd9\u4e0d\u9700\u8981\u8fdb\u884c\u5fae\u8c03\u6216\u8005\u8bbe\u7f6e\u7279\u5b9a\u7684\u4efb\u52a1\u6761\u4ef6\uff0c\u5728\u5927\u91cf\u7684\u4efb\u52a1\u91cd\u7ed3\u679c\u5927\u5927\u7531\u4e8e\u6807\u51c6\u7684zero-shot learning\uff0c\u6709\u65f6\u5019\u751a\u81f3\u4f18\u4e8efew-shot learning\u3002\u7136\u800c\uff0c\u5148\u524d\u7684\u5de5\u4f5c\u8868\u660e\uff0cCoT\u9700\u8981\u5728\u975e\u5e38\u5927\u7684\u6a21\u578b\u4e0a\u624d\u80fd\u83b7\u5f97\u6700\u4f73\u6027\u55ef\u5462\u3002\u5728\u6b64\u7bc7\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528\u8d85\u5927\u6a21\u578b\u751f\u6210\u7684\u63a8\u7406\u7ed3\u679c\u5bf9\u5c0f\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u3002\u4f5c\u8005\u63d0\u4f9b\u4e86\u4e00\u5957\u4e30\u5bcc\u7684\u7ed3\u679c\u4ee5\u53ca\u5b9a\u6027\/\u5b9a\u91cf\u5206\u6790\u7684\u6570\u636e\u96c6\u3002<\/li>\n<li>\u77e5\u8bc6\u84b8\u998f\uff08Knowledge distillation\uff09\uff1a\u901a\u5e38\u60c5\u51b5\u4e0b\uff0c\u77e5\u8bc6\u84b8\u998f\uff08Knowledge distillation KD\uff09\u662f\u6307\u8bad\u7ec3\u4ece\u5927\u6a21\u578b\u884d\u751f\u7684\u5c0f\u6a21\u578b\uff0c\u4ee5\u51cf\u5c0f\u6a21\u578b\u5c3a\u5bf8\u548c\u5c42\u6570\uff0c\u540c\u65f6\u4ecd\u7136\u4fdd\u6301\u51c6\u786e\u6027\u548c\u6cdb\u5316\u80fd\u529b\u3002\u672c\u8d28\u4e0a\uff0cKD\u662f\u4e00\u79cd\u6a21\u578b\u538b\u7f29\u5f62\u5f0f\uff0c\u4f7f\u5f97\u6709\u6548\u90e8\u7f72\u5230\u5bb9\u91cf\u6709\u9650\u7684\u8bbe\u5907\u4e0a\u6210\u4e3a\u53ef\u80fd\u3002\u4f5c\u8005\u63d0\u51fa\u7684\u65b9\u6cd5\u4e5f\u53ef\u4ee5\u8ba4\u4e3a\u662fKD\u7684\u53d8\u4f53\uff0c\u7c7b\u4f3c\u4e0e\u6539\u8fdb\u4f9d\u636e\u63d0\u793a\u7684\u65b9\u6cd5\u7684\u5de5\u4f5c\uff0c\u6216\u81f4\u529b\u4e8e\u65e0\u6570\u636e\u84b8\u998f\uff0c\u5176\u4e2d\u8fc1\u79fb\u6570\u636e\u662f\u7531\u4e00\u4e2a\u5927\u578b\u6559\u5e08\u6a21\u578b\u7efc\u5408\u751f\u6210\u7684\u3002\u4f46\u5728\u4f5c\u8005\u7684\u65b9\u6cd5\u4e2d\uff0c\u6559\u5e08\u6a21\u578b\u7684\u4f5c\u7528\u662f\u6559\u6388\u4e2d\u95f4\u63a8\u7406\u7684\u6982\u5ff5\u3002\u5177\u4f53\u7684\u8f93\u51fa\u4e0d\u662f\u63a8\u7406\u7684\u4e3b\u8981\u76d1\u7763\u4fe1\u53f7\uff0c\u800c\u662f\u751f\u6210\u7684\u7ed3\u6784\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u4e0d\u4f7f\u7528\u53cd\u6620\u8bd5\u56fe\u5339\u914d\u6559\u5e08\u8f93\u51fa\u7684\u6807\u51c6KD\u635f\u5931\u51fd\u6570\u3002<\/li>\n<\/ul>\n<h1>CoT SFT<\/h1>\n<p>\u6211\u4eec\u63d0\u51fa\u4e86\u5fae\u8c03- cot\uff0c\u8fd9\u662f\u4e00\u79cd\u4efb\u52a1\u4e0d\u53ef\u77e5(task-agnostic)\u7684\u65b9\u6cd5\uff0c\u53ef\u4ee5\u5728\u5c0f\u578bLMs\u4e2d\u5b9e\u73b0\u601d\u7ef4\u94fe\u63a8\u7406\u3002\u5176\u6838\u5fc3\u601d\u60f3\u662f\u4f7f\u7528CoT\u63d0\u793a\u4ece\u975e\u5e38\u5927\u7684\u6559\u5e08\u6a21\u578b\u4e2d\u751f\u6210\u63a8\u7406\u6837\u672c\uff0c\u7136\u540e\u4f7f\u7528\u751f\u6210\u7684\u6837\u672c\u5bf9\u5c0f\u7684\u5b66\u751f\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u3002\u8fd9\u79cd\u65b9\u6cd5\u4fdd\u7559\u4e86\u57fa\u4e8e\u63d0\u793a\u7684CoT\u65b9\u6cd5\u7684\u901a\u7528\u6027\uff0c\u540c\u65f6\u514b\u670d\u4e86\u5b83\u4eec\u5bf9\u8fc7\u4e8e\u5e9e\u5927\u7684\u6a21\u578b\u7684\u4f9d\u8d56\u3002\u4e3a\u4e86\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8\u901a\u7528\u6027\u548c\u6700\u5c0f\u5316\u6559\u5e08\u63a8\u7406\u6210\u672c\uff0c\u6211\u4eec\u5728\u6559\u5e08\u6a21\u578b\u4e0a\u4f7f\u7528\u4e86\u4efb\u52a1\u4e0d\u53ef\u77e5\u7684Zero-shot-CoT\u63d0\u793a\u65b9\u6cd5(Kojima\u7b49\u4eba\uff0c2022)\uff0c\u56e0\u4e3a\u5b83\u4e0d\u9700\u8981\u4efb\u4f55\u63a8\u7406\u793a\u4f8b\u6216\u957f\u63a8\u7406\u4e0a\u4e0b\u6587\u3002\u6211\u4eec\u5c06\u57287.3\u8282\u4e2d\u8ba8\u8bba\u6211\u4eec\u5bf9\u6559\u5e08CoT\u63d0\u793a\u65b9\u6cd5\u7684\u9009\u62e9\u3002\u5728\u4e0b\u9762\uff0c\u6211\u4eec\u7528\u4e09\u4e2a\u4e0d\u540c\u7684\u6b65\u9aa4\u6765\u63cf\u8ff0Fine-tune-CoT\u3002\u6211\u4eec\u8fd8\u5728\u56fe2\u4e2d\u63d0\u4f9b\u4e86\u4e00\u4e2a\u53ef\u89c6\u5316\u7684\u6982\u8ff0\u3002<\/p>\n<ul>\n<li>\u6b65\u9aa41\uff1a\u63a8\u7406\u751f\u6210\uff1a\u9996\u5148\u4f7f\u7528\u5927\u578b\u6559\u5e08\u6a21\u578b\u4e3a\u7ed9\u5b9a\u4efb\u52a1\u751f\u6210CoT\u63a8\u7406\u89e3\u91ca\u3002\u8003\u8651\u4e00\u4e2a\u95ee\u9898<span class=\"katex-eq\" data-katex-display=\"false\">q_i<\/span>\u548c\u5b83\u7684\u7b54\u6848<span class=\"katex-eq\" data-katex-display=\"false\">a_i<\/span>\u7ec4\u6210\u7684\u6807\u51c6\u6837\u672c<span class=\"katex-eq\" data-katex-display=\"false\">S_i<\/span>\u3002\u4f7f\u7528Zero-shot-CoT\u65b9\u6cd5\u8ba9\u6559\u5e08\u6a21\u578b\u751f\u6210\u4e00\u4e2a\u63a8\u7406\u89e3\u91ca<span class=\"katex-eq\" data-katex-display=\"false\">\\hat{r_i}<\/span>\u6765\u89e3\u51b3\u95ee\u9898<span class=\"katex-eq\" data-katex-display=\"false\">q_i<\/span>\u5e76\u4e14\u7ed9\u51fa\u6700\u7ec8\u7684\u7b54\u6848\u9884\u6d4b<span class=\"katex-eq\" data-katex-display=\"false\">\\hat{a_i}<\/span>\u3002\u6700\u7ec8\u751f\u6210\u7684\u6587\u672c\u5e8f\u5217\u6309\u7167\u5982\u4e0b\u683c\u5f0f\u7ec4\u7ec7\uff1aQ:<<span class=\"katex-eq\" data-katex-display=\"false\">q_i<\/span>>. A\uff1aLet\u2019s think step by step.<<span class=\"katex-eq\" data-katex-display=\"false\">\\hat{r_i}<\/span>>. Threrefor, the answer is <<span class=\"katex-eq\" data-katex-display=\"false\">\\hat{a_i}<\/span>>.<\/li>\n<li>\u6b65\u9aa42: \u683c\u5f0f\u5316: \u63a5\u4e0b\u6765\uff0c\u8fc7\u6ee4\u751f\u6210\u7684\u793a\u4f8b\u5e76\u5c06\u5b83\u4eec\u91cd\u65b0\u683c\u5f0f\u5316\u4e3a\u63d0\u793a\u5b8c\u6210\u5bf9\uff08prompt-completion\uff09\u3002\u5bf9\u4e8e\u8fc7\u6ee4\uff0c\u7b80\u5355\u7684\u5c06\u6559\u5e08\u6a21\u578b\u7684\u6700\u7ec8\u9884\u6d4b<span class=\"katex-eq\" data-katex-display=\"false\">\\hat{a_i}<\/span>\u4e0e\u771f\u5b9e\u7b54\u6848<span class=\"katex-eq\" data-katex-display=\"false\">a_i<\/span>\u6bd4\u8f83\u3002\u6ce8\u610f\uff0c\u8fd9\u4e2a\u8fc7\u6ee4\u73af\u8282\u53ef\u80fd\u4f1a\u5bfc\u81f4\u4e00\u4e9b\u8bad\u7ec3\u6837\u672c\u7684\u635f\u5931\u3002\u5bf9\u4e8e\u6240\u6709<span class=\"katex-eq\" data-katex-display=\"false\">\\hat{a_i}=a_i<\/span>\u7684\u5b9e\u4f8b\uff0c\u5c06\uff08<span class=\"katex-eq\" data-katex-display=\"false\">S_i,\\hat{r_i},\\hat{a_i}<\/span>\uff09\u6253\u5305\u6210\u4e00\u4e2a\u63a8\u7406\u6837\u672c<span class=\"katex-eq\" data-katex-display=\"false\">S_i^&#039;=(p_i,c_i)<\/span>,\u5373\u4e00\u4e2a\u63d0\u793a\u8865\u5168\u5bf9\uff08prompt-completion pair\uff09\u3002\u4e3a\u4e86\u6700\u5927\u5316\u63a8\u7406\u6548\u7387\uff0c\u4f7f\u7528\u57fa\u4e8e\u7279\u6b8a\u5b57\u7b26\u7684\u5206\u9694\u7b26\u6765\u6700\u5c0f\u5316token\u7684\u4f7f\u7528\u3002\u5177\u4f53\u6765\u8bf4\uff1a<span class=\"katex-eq\" data-katex-display=\"false\">p_i<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">c_i<\/span>\u5206\u522b\u91c7\u7528\u201c<<span class=\"katex-eq\" data-katex-display=\"false\">q_i<\/span>>###\u201d\u548c\u201c<<span class=\"katex-eq\" data-katex-display=\"false\">\\hat{r_i}<\/span>>--&gt;<<span class=\"katex-eq\" data-katex-display=\"false\">a_i<\/span>>END\u201d\u3002\u6ce8\u610f\u5230\uff0c\u57fa\u4e8e\u7b54\u6848\u7684\u8fc7\u6ee4\u5e76\u4e0d\u80fd\u4fdd\u969c\u529f\u80fd\u63a8\u7406\u7684\u6b63\u786e\u6027\uff0c\u7279\u522b\u5bf9\u4e8e\u9009\u62e9\u9898\uff0c\u8be5\u7ec6\u8282\u6709\u5f85\u8fdb\u4e00\u6b65\u7684\u5206\u6790\u3002<\/li>\n<li>\u6b65\u9aa43\uff1a\u5fae\u8c03\uff1a\u6700\u540e\uff0c\u4f7f\u7528\u7ec4\u88c5\u7684\u63a8\u7406\u6837\u672c\u5fae\u8c03\u4e00\u4e2a\u5c0f\u7684\u9884\u8bad\u7ec3\u5b66\u751f\u6a21\u578b\u3002\u4f7f\u7528\u4e0e\u9884\u8bad\u7ec3\u671f\u95f4\u76f8\u540c\u7684\u8bad\u7ec3\u76ee\u6807\uff0c\u5373\u81ea\u56de\u5f52\u8bed\u8a00\u5efa\u6a21\u76ee\u6807\uff0c\u6216\u4e0e\u6d4b\u4e0b\u4e00\u4e2a\u4ee4\u724c\u3002<\/li>\n<li>\u591a\u6837\u5316\u63a8\u7406\uff1b\u4e3a\u4e86\u4f7fFine-tune-CoT\u7684\u6559\u5b66\u6548\u679c\u6700\u5927\u5316\uff0c\u53ef\u4ee5\u4e3a\u6bcf\u4e2a\u8bad\u7ec3\u6837\u672c\u751f\u6210\u591a\u79cd\u63a8\u7406\u89e3\u91ca\u3002\u8fd9\u79cd\u65b9\u6cd5\u7684\u52a8\u673a\u662f\u76f4\u89c9\uff0c\u5373\u53ef\u4ee5\u4f7f\u7528\u591a\u79cd\u63a8\u7406\u65b9\u6cd5\u6765\u89e3\u51b3\u590d\u6742\u7684\u4efb\u52a1\u770b\uff0c\u5373\u7b2c\u4e8c\u7c7b\u578b\u4efb\u52a1\uff08type-2 tasks\uff09\u3002\u6211\u4eec\u5047\u8bbe\uff0c\u590d\u6742\u4efb\u52a1\u7684\u72ec\u7279\u7279\u5f81\u4e0e\u6559\u5e08\u6a21\u578b\u7684\u968f\u673a\u751f\u6210\u80fd\u529b\u76f8\u7ed3\u5408\uff0c\u53ef\u4ee5\u901a\u8fc7\u591a\u6837\u5316\u7684\u63a8\u7406\u663e\u8457\u63a8\u7406\u76d1\u7763\u3002\u5177\u4f53\u7684\u6765\u8bf4\u3002\u5bf9\u4e8e\u7ed9\u5b9a\u6837\u672c<span class=\"katex-eq\" data-katex-display=\"false\">S_i<\/span>\uff0c\u6211\u4eec\u91c7\u7528\u968f\u673a\u91c7\u6837\u7b56\u7565\uff0c\u5373\u9ad8\u6e29\u5ea6\uff08large T\uff09\u91c7\u6837\uff0c\u800c\u4e0d\u662f\u4f7f\u7528\u8d2a\u5fc3\u89e3\u7801\u7684zero-shot-CoT\u6765\u83b7\u5f97\u5355\u4e2a\u89e3\u91ca-\u7b54\u6848\u5bf9\uff08<span class=\"katex-eq\" data-katex-display=\"false\">\\hat{e_i},\\hat{a_i}<\/span>\uff09\uff0c\u7528\u6765\u83b7\u5f97<span class=\"katex-eq\" data-katex-display=\"false\">D<\/span>\u4e2a\u4e0d\u540c\u7684\u751f\u6210\u7ed3\u679c<span class=\"katex-eq\" data-katex-display=\"false\">{(\\hat{r_{ij}},\\hat{a_{i,j}})}_j^D<\/span>\u3002\u968f\u540e\u8fdb\u884c\u63a8\u7406\u7ed3\u679c\u8fc7\u6ee4\uff0c\u7136\u540e\u5fae\u8c03\u3002<span class=\"katex-eq\" data-katex-display=\"false\">D<\/span>\u88ab\u79f0\u4e3a\u63a8\u7406\u591a\u6837\u6027\u7a0b\u5ea6\u3002<\/li>\n<\/ul>\n<h1>\u5b9e\u9a8c<\/h1>\n<ul>\n<li>\u4efb\u52a1\u548c\u6570\u636e\u96c6\uff1a\u4f5c\u8005\u5728\u6d89\u53ca\u56db\u7c7b\u590d\u6742\u63a8\u7406\u768412\u4e2a\u6570\u636e\u96c6\u4e0a\u8bc4\u4f30\u4e86\u6240\u63d0\u7684\u65b9\u6cd5\uff0c\u5305\u62ec\u7b97\u672f(SingleEq, AddSub, MultiArith,GSM8K, SVAMP)\u3001\u5176\u4ed6(Date Understanding, Tracking Shuffled Objects)\u3001\u7b26\u53f7\uff08Last Letter Concatenation, Coin Flip\uff09\u548c\u540c\u6837\u573a\u666f\u63a8\u7406\u3002<\/li>\n<li>\u6a21\u578b\uff1a\u5bf9\u4e8e\u6559\u5e08\u6a21\u578b\uff0c\u4f7f\u7528OpenAI\u63d0\u4f9b\u7684GPT-3 175B\u7684\u56db\u4e2a\u53d8\u4f53\u3002\u9664\u975e\u53e6\u6709\u8bf4\u660e\uff0c\u4f7f\u7528\u57fa\u4e8eInstructGPT 175B\u7684text-davinci-002\u4f5c\u4e3aFinetune-CoT\u7684\u6559\u5e08\u3002\u5bf9\u4e8e\u5b66\u751f\u6a21\u578b\uff0c\u4f7f\u7528\u4e86\u56db\u79cd\u6d41\u884c\u7684\u6a21\u578b\u5bb6\u65cf\u3002\u5bf9\u4e8e\u4e3b\u8981\u65f6\u95f4\uff0c\u4f7f\u7528GPT-3\uff0c\u4ee5\u4e3a\u4ed6\u4eec\u5f88\u5bb9\u6613\u901a\u8fc7OpenAI API\u8fdb\u884c\u5fae\u8c03\u3002\u7531\u4e8eAPI\u7684\u9ed1\u7bb1\u6027\u8d28\uff0c\u6211\u4eec\u8fd8\u8003\u8651\u5728\u53d7\u63a7\u8bbe\u7f6e\u4e0b\u7684\u5404\u79cd\u5f00\u6e90\u6a21\u578b\u3002\u5206\u522b\u4f7f\u7528GPT-2{small,medium,large}\u548cT5-{small,medium,large}\u4f5c\u4e3a\u4ec5\u89e3\u7801\u5668\u548c\u7f16\u7801\u5668-\u89e3\u7801\u5668\u67b6\u6784\u7684\u4ee3\u8868\u6027\u6a21\u578b\u5bb6\u65cf\u3002<\/li>\n<li>\u57fa\u7ebf\u65b9\u6cd5\uff1a\u5c06Fine-tune-CoT\u548c\u56db\u79cd\u57fa\u7ebf\u65b9\u6cd5\u8fdb\u884c\u6bd4\u8f83\uff1astandard zero-shot prompting\uff0cvanilla fine-tuning\uff0cZero-shot-CoT\u548cFew-shotCoT\u3002\u7ed9\u5b9a\u4e00\u4e2a\u8bad\u7ec3\u6837\u672c<span class=\"katex-eq\" data-katex-display=\"false\">{(q_i,a_i)}_i<\/span>\uff0c\u4f7f\u7528\u4e00\u4e2a\u7b80\u5355\u7684\u683c\u5f0f\u201cQ<<span class=\"katex-eq\" data-katex-display=\"false\">q_i<\/span>>\u201d\u6765\u8fdb\u884czero-shot prompting\u3002\u5bf9\u4e8evanilla fine-tuning\uff0c\u5c06\u63d0\u793a\u8bcd\u548c\u7b54\u6848\u683c\u5f0f\u5316\u4e3a\u201c<<span class=\"katex-eq\" data-katex-display=\"false\">q_i<\/span>>###\u201d\u548c\u201c<<span class=\"katex-eq\" data-katex-display=\"false\">a_i<\/span>>END\u201d.\u5bf9\u4e8e\u6559\u5e08\u7684\u591a\u6837\u5316\u63a8\u7406\uff0c\u91c7\u7528T=0.7\u7684\u6e29\u5ea6\u62bd\u6837\uff0c\u9644\u5f55A\u4e2d\u63d0\u4f9b\u4e86\u5b9e\u9a8c\u7ec6\u8282\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>MetaData \u53d1\u8868\u65e5\u671f\uff1a2023\u5e741\u670813\u65e5 \u4f5c\u8005\uff1aNamgyu Ho, Laura Schmid, Se-Young Yun &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[45],"tags":[],"class_list":["post-663","post","type-post","status-publish","format-standard","hentry","category-llms"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/663","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/comments?post=663"}],"version-history":[{"count":5,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/663\/revisions"}],"predecessor-version":[{"id":670,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/663\/revisions\/670"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=663"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=663"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=663"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}