{"id":37211,"date":"2025-01-24T15:00:16","date_gmt":"2025-01-24T07:00:16","guid":{"rendered":"https:\/\/17aitech.com\/?p=37211"},"modified":"2025-01-24T15:00:16","modified_gmt":"2025-01-24T07:00:16","slug":"neurips-2024-%e5%8f%af%e4%bf%a1%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%96%b0%e6%8c%91%e6%88%98%ef%bc%9a%e5%99%aa%e5%a3%b0%e6%80%9d%e7%bb%b4%e9%93%be%e6%8f%90%e7%a4%ba%e4%b8%8b%e7%9a%84%e9%b2%81%e6%a3%92","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=37211","title":{"rendered":"NeurIPS 2024 | \u53ef\u4fe1\u5927\u6a21\u578b\u65b0\u6311\u6218\uff1a\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u4e0b\u7684\u9c81\u68d2\u63a8\u7406\uff0c\u51c6\u786e\u7387\u76f4\u964d40%"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-12-12-2\" target=\"_blank\">NeurIPS 2024 | \u53ef\u4fe1\u5927\u6a21\u578b\u65b0\u6311\u6218\uff1a\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u4e0b\u7684\u9c81\u68d2\u63a8\u7406\uff0c\u51c6\u786e\u7387\u76f4\u964d40%<\/a><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8a58fd8142ee302c5fa15b5475a5859e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8a58fd8142ee302c5fa15b5475a5859e.png\"><\/a><\/section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"166\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>AIxiv\u4e13\u680f\u662f\u673a\u5668\u4e4b\u5fc3\u53d1\u5e03\u5b66\u672f\u3001\u6280\u672f\u5185\u5bb9\u7684\u680f\u76ee\u3002\u8fc7\u53bb\u6570\u5e74\uff0c\u673a\u5668\u4e4b\u5fc3AIxiv\u4e13\u680f\u63a5\u6536\u62a5\u9053\u4e862000\u591a\u7bc7\u5185\u5bb9\uff0c\u8986\u76d6\u5168\u7403\u5404\u5927\u9ad8\u6821\u4e0e\u4f01\u4e1a\u7684\u9876\u7ea7\u5b9e\u9a8c\u5ba4\uff0c\u6709\u6548\u4fc3\u8fdb\u4e86\u5b66\u672f\u4ea4\u6d41\u4e0e\u4f20\u64ad\u3002\u5982\u679c\u60a8\u6709\u4f18\u79c0\u7684\u5de5\u4f5c\u60f3\u8981\u5206\u4eab\uff0c\u6b22\u8fce\u6295\u7a3f\u6216\u8005\u8054\u7cfb\u62a5\u9053\u3002\u6295\u7a3f\u90ae\u7bb1\uff1aliyazhou@jiqizhixin.com\uff1bzhaoyunfeng@jiqizhixin.com<\/section>\n<\/section>\n<\/blockquote>\n<section><\/section>\n<section>\u5f53\u524d\uff0c\u5927\u8bed\u8a00\u6a21\u578b\uff08Large Language Model, LLM\uff09\u501f\u52a9\u4e0a\u4e0b\u6587\u5b66\u4e60\uff08In-context Learning\uff09\u548c\u601d\u7ef4\u94fe\u63d0\u793a\uff08Chain of Thoughts Prompting\uff09\uff0c\u5728\u8bb8\u591a\u590d\u6742\u63a8\u7406\u4efb\u52a1\u4e0a\u5c55\u73b0\u51fa\u4e86\u5f3a\u5927\u7684\u80fd\u529b\u3002<\/section>\n<section><\/section>\n<section>\u7136\u800c\uff0c\u73b0\u6709\u7814\u7a76\u8868\u660e\uff0c<strong>LLM \u5728\u5e94\u5bf9\u566a\u58f0\u8f93\u5165\u65f6\u5b58\u5728\u660e\u663e\u4e0d\u8db3<\/strong>\uff1a\u5f53\u8f93\u5165\u7684\u95ee\u9898\u5305\u542b\u65e0\u5173\u5185\u5bb9\uff0c\u6216\u8005\u906d\u5230\u8f7b\u5fae\u4fee\u6539\u65f6\uff0c\u6a21\u578b\u6781\u5bb9\u6613\u53d7\u5230\u5e72\u6270\uff0c\u8fdb\u800c\u504f\u79bb\u6b63\u786e\u7684\u63a8\u7406\u65b9\u5411\u3002\u5982\u56fe 1 \u5de6\u6240\u793a\uff0cQ1 \u4e2d\u7684\u300cWe know 6+6=12 and 3+7=10 in base 10\u300d \u662f\u5173\u4e8e base-9 \u8ba1\u7b97\u7684\u566a\u58f0\u4fe1\u606f\uff0c\u8be5\u4fe1\u606f\u5bb9\u6613\u8bef\u5bfc\u6a21\u578b\u8f93\u51fa\u9519\u8bef\u7684\u7ed3\u679c\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-50b1ae4ba9ed2e4255a833659d2bffd7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-50b1ae4ba9ed2e4255a833659d2bffd7.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 1. \u566a\u58f0\u95ee\u9898\uff08Noisy Questions\uff09\u548c\u566a\u58f0\u601d\u7ef4\u94fe\uff08Noisy Rationales\uff09\u7684\u793a\u4f8b<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5df2\u6709\u7684\u9c81\u68d2\u7814\u7a76\u5927\u591a\u4fa7\u91cd\u4e8e\u566a\u58f0\u95ee\u9898\uff08Noisy Questions\uff09\uff0c\u7136\u800c\uff0cLLM \u5728<strong>\u566a\u58f0\u601d\u7ef4\u94fe\uff08Noisy Rationales\uff09<\/strong>\u4e0b\u7684\u63a8\u7406\u8fd8\u6ca1\u6709\u5f97\u5230\u5145\u5206\u7684\u63a2\u7a76\u3002\u5728\u672c\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u5c06\u566a\u58f0\u601d\u7ef4\u94fe\u5b9a\u4e49\u4e3a\uff1a\u5305\u542b<strong>\u4e0d\u76f8\u5173\u6216\u8005\u4e0d\u51c6\u786e\u63a8\u7406\u6b65\u9aa4<\/strong>\u7684\u601d\u7ef4\u94fe\uff0c\u5982\u56fe 1 \u53f3\u4fa7 R1 \u4e2d\u7684\u300c13 + 8 = 21\u300d\u6b65\u9aa4\uff0c\u5bf9\u4e8e base-9 \u8ba1\u7b97\u6765\u8bf4\uff0c\u662f\u9519\u8bef\u7684\u63a8\u7406\u6b65\u9aa4\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u4e9b\u566a\u58f0\u601d\u7ef4\u94fe\u901a\u5e38\u6e90\u81ea LLM \u7684\u5b9e\u9645\u5e94\u7528\uff0c\u6bd4\u5982<strong>\u4f17\u5305\u5e73\u53f0\u3001\u5bf9\u8bdd\u7cfb\u7edf\u3001\u673a\u5668\u751f\u6210\u6570\u636e<\/strong>\u7b49\u573a\u666f\uff0c\u4eba\u7c7b\u548c\u673a\u5668\u5728\u63a8\u7406\u4e2d\u90fd\u4f1a\u4e0d\u53ef\u907f\u514d\u5730\u72af\u9519\uff0c\u4ece\u800c\u4ea7\u751f\u566a\u58f0\u601d\u7ef4\u94fe\u3002\u56e0\u6b64\uff0c\u566a\u58f0\u601d\u7ef4\u94fe\u7684\u5b9e\u9645\u5f71\u54cd\u548c\u6280\u672f\u6311\u6218\u4e0d\u5bb9\u5c0f\u89d1\u3002\u5f53\u524d\uff0c\u6211\u4eec\u4ecd\u7136\u4e0d\u6e05\u695a LLM \u5728\u9762\u5bf9\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u65f6\u7684\u9c81\u68d2\u6027\u80fd\u5982\u4f55\uff0c\u4e5f\u7f3a\u5c11\u6709\u6548\u7684\u5e94\u5bf9\u7b56\u7565\u3002\u56e0\u6b64\uff0c\u975e\u5e38\u6709\u5fc5\u8981\u6784\u5efa\u4e00\u4e2a\u65b0\u7684\u6570\u636e\u96c6\uff0c\u7528\u4e8e\u7cfb\u7edf\u8bc4\u4f30\u5f53\u524d LLM \u5728\u566a\u58f0\u601d\u7ef4\u94fe\u573a\u666f\u4e0b\u7684\u9c81\u68d2\u6027\uff0c\u4ee5\u53ca\u9a8c\u8bc1\u76f8\u5e94\u7684\u9c81\u68d2\u63a8\u7406\u7b56\u7565\u3002<\/section>\n<section><\/section>\n<section>\u5bf9\u6b64\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u00a0<strong>NoRa \u6570\u636e\u96c6<\/strong>\uff0c\u5e76\u8fdb\u884c\u4e86\u5927\u91cf\u7684\u5b9e\u9a8c\u8bc4\u6d4b\u3002\u7ed3\u679c\u8868\u660e\uff0cGPT-3.5-Turbo\u3001Gemini-Pro\u3001Llama2-70B \u548c Mixtral-8x7B \u7b49\u5f00\u6e90\u6216\u95ed\u6e90 LLM \u90fd\u6781\u5bb9\u6613\u53d7\u5230\u566a\u58f0\u601d\u7ef4\u94fe\u7684\u5f71\u54cd\u3002\u5176\u4e2d\uff0cGPT-3.5-Turbo \u7684\u51c6\u786e\u7387\u81f3\u591a\u53ef\u964d\u4f4e\u00a0<strong>40.4%<\/strong>\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u4e5f\u547c\u5401\u5927\u5bb6\u66f4\u591a\u5730\u5173\u6ce8\u5927\u6a21\u578b\u63a8\u7406\u7684\u9c81\u68d2\u6027\u95ee\u9898\u3002<\/section>\n<section><\/section>\n<section>\u6211\u4eec\u7684\u4e3b\u8981\u8d21\u732e\u6709\u5982\u4e0b\u4e09\u70b9\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<p><strong>\u65b0\u95ee\u9898<\/strong>\uff1a\u5bf9\u5f53\u524d\u6d41\u884c\u7684\u601d\u7ef4\u94fe\u63d0\u793a\u6280\u672f\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u5c1a\u672a\u5145\u5206\u63a2\u7a76\u7684\u566a\u58f0\u601d\u7ef4\u94fe\u95ee\u9898\uff08Noisy Rationales\uff09\uff0c\u5e76\u7ed9\u51fa\u4e86\u8be6\u7ec6\u7684\u95ee\u9898\u5b9a\u4e49\u548c\u7edf\u4e00\u7684\u95ee\u9898\u5efa\u6a21\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u65b0\u6570\u636e\u96c6<\/strong>\uff1a\u6211\u4eec\u6784\u5efa\u4e86 NoRa \u6570\u636e\u96c6\uff0c\u7528\u4e8e\u8bc4\u6d4b LLM \u5728\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u4e0b\u7684\u63a8\u7406\u9c81\u68d2\u6027\u3002\u6211\u4eec\u4f7f\u7528 NoRa \u6570\u636e\u96c6\u5bf9 LLM \u5c55\u5f00\u7cfb\u7edf\u8bc4\u4f30\uff0c\u63ed\u9732\u4e86 LLM \u63a8\u7406\u7684\u9c81\u68d2\u6027\u4e0d\u8db3\uff0c\u6570\u636e\u53bb\u566a\u80fd\u529b\u975e\u5e38\u6709\u9650\u7684\u95ee\u9898\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u65b0\u65b9\u6cd5<\/strong>\uff1a\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u79cd\u7b80\u5355\u6709\u6548\u7684\u65b9\u6cd5\uff08CD-CoT\uff09\uff0c\u57fa\u4e8e\u5355\u4e2a\u6b63\u786e\u7684\u601d\u7ef4\u94fe\u793a\u4f8b\uff0c\u53bb\u7ea0\u6b63\u566a\u58f0\u601d\u7ef4\u94fe\u5e76\u5b8c\u6210\u63a8\u7406\uff0c\u5e76\u901a\u8fc7\u5927\u91cf\u5b9e\u9a8c\u9a8c\u8bc1\u4e86\u65b9\u6cd5\u7684\u6709\u6548\u6027\u3002<\/p>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u63a5\u4e0b\u6765\u5c06\u4ece<strong>\u65b0\u95ee\u9898\u3001\u65b0\u6570\u636e\u96c6\u3001\u65b0\u65b9\u6cd5<\/strong>\u8fd9\u4e09\u4e2a\u89d2\u5ea6\uff0c\u7b80\u8981\u5730\u4ecb\u7ecd\u6211\u4eec\u5173\u4e8e\u5927\u6a21\u578b\u63a8\u7406\u9c81\u68d2\u6027\u7684\u7814\u7a76\u7ed3\u679c\uff0c\u76f8\u5173\u8bba\u6587\u5df2\u53d1\u8868\u4e8e NeurIPS 2024 \u4f1a\u8bae\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-51f7b2936466ae5fdfa02de16ebbbdf3.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-51f7b2936466ae5fdfa02de16ebbbdf3.png\"><\/a><\/section>\n<ul>\n<li>\n<p>\u8bba\u6587\u6807\u9898\uff1aCan Language Models Perform Robust Reasoning in Chain-of-thought Prompting with Noisy Rationales?<\/p>\n<\/li>\n<li>\n<p>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2410.23856<\/p>\n<\/li>\n<li>\n<p>\u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/tmlr-group\/NoisyRationales<\/p>\n<\/li>\n<li>\n<p>slides \u94fe\u63a5\uff1ahttps:\/\/andrewzhou924.github.io\/_pages\/data\/slides-NoRa.pdf<\/p>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u65b0\u95ee\u9898\uff1aNoisy Rationales<\/strong><\/section>\n<section><\/section>\n<section>\u601d\u7ef4\u94fe\u53ef\u4ee5\u6709\u6548\u63d0\u5347\u5927\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b [1]\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u901a\u8fc7\u7ed9\u51fa\u5e26\u6709\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u7684\u793a\u4f8b\uff0cLLM \u53ef\u4ee5\u5f88\u5feb\u5730\u9002\u5e94\u5230\u65b0\u4efb\u52a1\u4e0a\uff0c\u800c\u65e0\u9700\u8fdb\u884c\u53c2\u6570\u4fee\u6539\uff08\u89c1\u56fe 2 \u53f3\u4e0a\u89d2\uff09\u3002\u73b0\u6709\u5de5\u4f5c\u4e2d\uff0c\u901a\u5e38\u5047\u8bbe\u601d\u7ef4\u94fe\u5305\u542b\u6e05\u695a\u4e14\u6b63\u786e\u7684\u63a8\u7406\u6b65\u9aa4\uff0c\u4f46\u5b9e\u9645\u60c5\u51b5\u53ef\u80fd\u5e76\u975e\u5982\u6b64\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-c89e06040c92260a3c250100c3b1d6fb.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-c89e06040c92260a3c250100c3b1d6fb.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 2. \u5404\u79cd setting \u4e0b\u7684\u6a21\u578b\u8f93\u5165<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u76ee\u524d\uff0c\u5df2\u7ecf\u6709\u8bb8\u591a\u5de5\u4f5c\u63a2\u7d22\u4e86 Noisy Questions \u5bf9 LLM \u63a8\u7406\u6027\u80fd\u7684\u5f71\u54cd\uff08\u89c1\u56fe 2 \u5de6\u4e0b\u89d2\uff09\uff0c\u63ed\u793a\u4e86 LLM \u5bf9\u8f93\u5165\u4e2d\u5fae\u5c0f\u4fee\u6539\u7684\u654f\u611f\u6027 [2,3]\u3002<\/section>\n<section><\/section>\n<section>\u7136\u800c\uff0c\u5728\u4eba\u5de5\u6807\u6ce8\u6216\u673a\u5668\u751f\u6210\u7684\u601d\u7ef4\u94fe\u4e2d\uff0c\u53ef\u80fd\u4f1a\u5305\u542b\u4e00\u4e9b<strong>\u4e0e\u95ee\u9898\u4e0d\u76f8\u5173\u6216\u4e0d\u51c6\u786e\u7684\u63a8\u7406\u6b65\u9aa4<\/strong>\uff08\u89c1\u56fe 2 \u53f3\u4e0b\u89d2\uff09\uff0c\u8fd9\u4e9b\u566a\u58f0\u601d\u7ef4\u94fe\u53ef\u80fd\u4f1a\u5bf9\u63a8\u7406\u6027\u80fd\u4ea7\u751f\u8d1f\u9762\u5f71\u54cd\uff0c\u4f46\u76ee\u524d LLM \u5bf9\u566a\u58f0\u601d\u7ef4\u94fe\uff08Noisy Rationales\uff09\u7684\u9c81\u68d2\u6027\u4f9d\u65e7\u672a\u77e5\u3002<\/section>\n<section><\/section>\n<section>\u56e0\u6b64\uff0c<strong>\u672c\u6587\u63d0\u51fa\u4e86\u65b0\u7684\u7814\u7a76\u95ee\u9898 Noisy Rationales\uff1a\u5f53\u793a\u4f8b\u7684\u63a8\u7406\u6b65\u9aa4\u4e2d\u5305\u542b\u65e0\u5173\u7684\u6216\u8005\u4e0d\u51c6\u786e\u7684\u5185\u5bb9\u65f6\uff0cLLM \u7684\u63a8\u7406\u9c81\u68d2\u6027\u5982\u4f55<\/strong>\uff1f\u5bf9\u8fd9\u4e00\u95ee\u9898\u7684\u63a2\u7d22\uff0c\u6709\u52a9\u4e8e\u6df1\u5165\u7406\u89e3\u548c\u63d0\u5347 LLM \u5728\u975e\u5b8c\u5907\u573a\u666f\u4e2d\u7684\u63a8\u7406\u80fd\u529b\u3002<\/section>\n<section><\/section>\n<section><strong>\u65b0\u6570\u636e\u96c6\uff1aNoRa<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u8bc4\u4f30 LLM \u5728\u566a\u58f0\u601d\u7ef4\u94fe\u4e0b\u7684\u9c81\u68d2\u6027\uff0c\u6211\u4eec\u6784\u5efa\u4e86 NoRa\uff08Noisy Rationales\uff09\u6570\u636e\u96c6\uff0cNoRa \u6db5\u76d6\u4e86 3 \u79cd\u63a8\u7406\u4efb\u52a1\u7c7b\u578b\uff1a\u6570\u5b66\u63a8\u7406\u3001\u7b26\u53f7\u63a8\u7406\u548c\u5e38\u8bc6\u63a8\u7406\uff0c\u5171\u5305\u542b\u00a0<strong>26391<\/strong> \u4e2a\u95ee\u9898\u4ee5\u53ca\u00a0<strong>5<\/strong> \u79cd\u5b50\u4efb\u52a1\u3002<\/section>\n<section><\/section>\n<section>\u4e00\u6761\u601d\u7ef4\u94fe\uff08Rationale\uff09\u5305\u542b\u591a\u4e2a\u8fde\u7eed\u7684\u63a8\u7406\u6b65\u9aa4\uff08Thoughts\uff09\uff1b<strong>\u566a\u58f0\u601d\u7ef4\u94fe\uff08Noisy Rationale\uff09<\/strong>\u5305\u542b\u7684<strong>\u566a\u58f0\u63a8\u7406\u6b65\u9aa4\uff08Noisy Thoughts\uff09<\/strong>\u88ab\u5b9a\u4e49\u4e3a\u5982\u4e0b\u4e24\u7c7b\uff08\u793a\u4f8b\u89c1\u56fe 3\uff09\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<p><strong>\u4e0d\u76f8\u5173\u566a\u58f0\uff08Irrelevant Thoughts\uff09<\/strong>\u662f\u6307\u5bf9\u89e3\u51b3\u95ee\u9898\u65e0\u7528\u7684\u4fe1\u606f\uff0c\u5982\u5728\u63a8\u65ad\u4eb2\u5c5e\u5173\u7cfb\u65f6\u8ba8\u8bba\u63a2\u8ba8\u5144\u5f1f\u59d0\u59b9\u4e4b\u95f4\u7684\u57fa\u56e0\u91cd\u53e0\u60c5\u51b5\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u4e0d\u51c6\u786e\u566a\u58f0\uff08Inaccurate Thoughts\uff09<\/strong>\u5219\u662f\u63a8\u7406\u4e2d\u7684\u4e8b\u5b9e\u6027\u9519\u8bef\uff0c\u5982\u5728\u7279\u5b9a\u8fdb\u5236\u8ba1\u7b97\u4e2d\u4f7f\u7528\u9519\u8bef\u7684\u8ba1\u7b97\u89c4\u5219\u3002<\/p>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-76e6a452aca49b8ca8e9145bf9116f2f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-76e6a452aca49b8ca8e9145bf9116f2f.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 3. NoRa \u6570\u636e\u96c6\u7684\u6837\u672c<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5728\u6784\u5efa\u6570\u636e\u96c6\u65f6\uff0c\u6211\u4eec\u901a\u8fc7\u63d2\u5165 Noisy Thoughts \u6765\u751f\u6210\u566a\u58f0\u601d\u7ef4\u94fe\uff0c\u8fd9\u4e9b\u566a\u58f0\u4ec5\u5f71\u54cd\u63a8\u7406\u94fe\u7684\u7ec6\u8282\uff0c\u800c\u4e0d\u6539\u53d8\u95ee\u9898\u548c\u6700\u7ec8\u7b54\u6848\u7684\u6b63\u786e\u6027\u3002\u6b64\u5916\uff0c\u6211\u4eec\u4f7f\u7528\u4e0d\u540c\u7684\u566a\u58f0\u6bd4\u4f8b\uff08Noise Ratio\uff0c\u5373 Noisy Thoughts \u5360\u6240 Thoughts \u7684\u6bd4\u4f8b\uff0c\u5982 0.3\u30010.5\u30010.8\uff09\u6765\u63a7\u5236\u4efb\u52a1\u7684\u56f0\u96be\u7a0b\u5ea6\uff0cNoise Ratio \u8d8a\u5927\u4efb\u52a1\u96be\u5ea6\u4e5f\u8d8a\u5927\u3002NoRa \u6570\u636e\u96c6\u7684\u7edf\u8ba1\u4fe1\u606f\u5982\u56fe 4 \u6240\u793a\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-525eb93188b7c226634e8e21272305bc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-525eb93188b7c226634e8e21272305bc.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 4. NoRa \u6570\u636e\u96c6\u7684\u7edf\u8ba1\u4fe1\u606f<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>NoRa \u6570\u636e\u96c6 \u6d4b\u8bc4\u7ed3\u679c<\/strong><\/section>\n<section><\/section>\n<section>\u6211\u4eec\u4ee5 GPT-3.5-Turbo \u4e3a base model\uff0c\u6d4b\u8bd5\u4e86\u5176\u5728 NoRa \u4e0a\u7684\u8868\u73b0\uff0c\u5e76\u4e14\u5bf9\u6bd4\u4e86\u591a\u79cd<strong>\u53bb\u566a\u65b9\u6cd5<\/strong>\u3002\u8fd9\u4e9b\u53bb\u566a\u65b9\u6cd5\u53ef\u4ee5\u5206\u4e3a\u4e24\u7c7b\uff1a<\/section>\n<ul>\n<li>\n<p><strong>\u81ea\u6211\u7ea0\u6b63\u65b9\u6cd5\uff08Self-correction\uff09<\/strong>\uff1a\u5305\u62ec Intrinsic Self-correction (ISC) [4] \u548c Self-polish (SP) [5]\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u81ea\u6211\u4e00\u81f4\u6027\u65b9\u6cd5\uff08Self-consistency\uff09<\/strong>\uff1a\u5305\u62ec SmoothLLM (SM) [6]\uff0cSelf-denoise (SD) [7] \u548c Self-consistency (SC) [8]\u3002<\/p>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-db978c14f1fce8eb57ba7851cf85ebf4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-db978c14f1fce8eb57ba7851cf85ebf4.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 5. \u5404\u79cd\u53bb\u566a\u65b9\u6cd5 \u5728 NoRa \u6570\u636e\u96c6\u4e0a\u7684\u6d4b\u8bc4\u7ed3\u679c<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5b9e\u9a8c\u7ed3\u679c\uff08\u56fe 5\uff09\u8868\u660e\uff1a<\/section>\n<ol>\n<li>\n<p><strong>\u65e0\u8bba\u91c7\u53d6\u54ea\u79cd\u73b0\u6709\u65b9\u6cd5\uff0cLLM \u90fd\u4f1a\u53d7\u5230\u566a\u58f0\u601d\u7ef4\u94fe\u7684\u4e25\u91cd\u5f71\u54cd<\/strong>\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5b58\u5728\u4e0d\u76f8\u5173\u566a\u58f0\u65f6\uff0c\u5404\u65b9\u6cd5\u7684\u6027\u80fd\u4e0b\u964d\u4e86 0.2% &#8211; 25.3%\uff1b\u5b58\u5728\u4e0d\u51c6\u786e\u566a\u58f0\u65f6\uff0c\u5404\u65b9\u6cd5\u7684\u6027\u80fd\u4e0b\u964d\u4e86 0.1% &#8211; 54.0%\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u5728 NoRa \u7684\u5927\u591a\u6570\u4efb\u52a1\u4e2d\uff0c\u81ea\u6211\u7ea0\u6b63\u65b9\u6cd5\u7684\u8868\u73b0\u4e0d\u4f73<\/strong>\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u81ea\u4e00\u81f4\u6027\u65b9\u6cd5\u53ef\u4ee5\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u7f13\u89e3\u566a\u58f0\u7684\u5f71\u54cd\uff0c\u4f46\u65e0\u6cd5\u505a\u5230\u771f\u6b63\u7684\u6570\u636e\u53bb\u566a<\/strong>\u3002<\/p>\n<\/li>\n<\/ol>\n<section>\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u8fdb\u884c\u4e86\u5404\u79cd\u6d88\u878d\u7814\u7a76\uff0c\u6765\u63a2\u7d22<strong>\u4e0d\u540c\u56e0\u7d20<\/strong>\u5bf9 NoRa \u6570\u636e\u96c6\u8bc4\u4f30\u7ed3\u679c\u7684\u5f71\u54cd\uff08\u89c1\u56fe 6\uff09\uff0c\u6211\u4eec\u53d1\u73b0\uff1a<\/section>\n<ol>\n<li>\n<p><strong>\u8c03\u6574\u6e29\u5ea6\u7cfb\u6570<\/strong>\u53ef\u4ee5\u6539\u5584\u6a21\u578b\u5728\u566a\u58f0\u601d\u7ef4\u94fe\u4e0b\u7684\u63a8\u7406\u6027\u80fd\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u4f7f\u7528\u66f4\u591a\u7684\u566a\u58f0\u793a\u4f8b<\/strong>\u53ef\u4ee5\u63d0\u9ad8\u5927\u591a\u6570\u4efb\u52a1\u7684\u63a8\u7406\u6027\u80fd\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u4e0d\u540c\u7684\u5927\u8bed\u8a00\u6a21\u578b<\/strong>\u666e\u904d\u5bb9\u6613\u53d7\u5230\u566a\u58f0\u601d\u7ef4\u94fe\u7684\u5f71\u54cd\u3002<\/p>\n<\/li>\n<\/ol>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-07efab737a22b775da646fb0596699bf.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-07efab737a22b775da646fb0596699bf.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 6. \u6d88\u878d\u5b9e\u9a8c\uff1a(\u5de6) \u6e29\u5ea6\u7cfb\u6570\u5bf9\u6027\u80fd\u7684\u5f71\u54cd\uff1b(\u4e2d) \u793a\u4f8b\u4e2a\u6570\u5bf9\u6027\u80fd\u7684\u5f71\u54cd\uff1b(\u53f3) \u5404\u79cd\u6a21\u578b\u7684\u6027\u80fd<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u65b0\u65b9\u6cd5\uff1aCD-CoT<\/strong><\/section>\n<section><\/section>\n<section>\u6839\u636e\u6d4b\u8bc4\u7ed3\u679c\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u5728\u5e94\u5bf9\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u65f6\uff0c\u5176\u81ea\u8eab\u7684\u53bb\u566a\u80fd\u529b\u975e\u5e38\u6709\u9650\uff1b\u5373\u4fbf\u4f7f\u7528\u81ea\u6211\u7ea0\u6b63\u6216\u81ea\u4e00\u81f4\u6027\u65b9\u6cd5\uff0c\u6548\u679c\u4ecd\u4e0d\u7406\u60f3\u3002<\/section>\n<section><\/section>\n<section>\u56e0\u6b64\uff0c\u6211\u4eec\u8ba4\u4e3a\u6709\u5fc5\u8981\u5f15\u5165<strong>\u5916\u90e8\u76d1\u7763\u4fe1\u53f7<\/strong>\u6765\u589e\u5f3a\u6a21\u578b\u9c81\u68d2\u6027\uff0c\u4e14\u8fd9\u79cd\u76d1\u7763\u4fe1\u53f7\u65e2\u8981\u8db3\u4ee5\u5b9e\u73b0\u53bb\u566a\uff0c\u53c8\u8981\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u5207\u5b9e\u53ef\u884c\u3002\u5bf9\u6b64\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u7b80\u5355\u6709\u6548\u7684\u53bb\u566a\u63a8\u7406\u65b9\u6cd5\uff0c\u00a0<strong>CD-CoT\uff08Contrastive Denoising with Noisy Chain of Thoughts\uff09<\/strong>\u3002<\/section>\n<section><\/section>\n<section>CD-CoT \u501f\u9274\u4e86\u5bf9\u6bd4\u5b66\u4e60\u7684\u601d\u8def\uff0c\u901a\u8fc7\u8ba9 LLM \u663e\u5f0f\u5730\u5bf9\u6bd4\u6709\u566a\u548c\u5e72\u51c0\u7684\u601d\u7ef4\u94fe\uff0c\u4ece\u800c\u8bc6\u522b\u51fa\u566a\u58f0\u4fe1\u606f\u3002\u65b9\u6cd5\u4e3b\u8981\u5305\u62ec\u56db\u4e2a\u5173\u952e\u6b65\u9aa4\uff0c\u6b65\u9aa4 1&amp;2 \u8fdb\u884c\u663e\u5f0f\u7684\u53bb\u566a\uff0c\u6b65\u9aa4 3&amp;4 \u8fdb\u884c\u7cbe\u7ec6\u63a8\u7406\u5e76\u83b7\u5f97\u6700\u7ec8\u7b54\u6848\u3002<\/section>\n<section><\/section>\n<section>\u56db\u4e2a\u6b65\u9aa4\u5177\u4f53\u5982\u4e0b\uff1a<\/section>\n<ol>\n<li>\n<p><strong>\u6539\u5199\u601d\u7ef4\u94fe<\/strong>\uff1a\u501f\u52a9\u4e00\u4e2a\u5e72\u51c0\u7684\u601d\u7ef4\u94fe\u793a\u4f8b\uff0c\u5f15\u5bfc LLM \u901a\u8fc7\u5bf9\u6bd4\u6539\u5199\u548c\u7ea0\u6b63\u566a\u58f0\u601d\u7ef4\u94fe\uff0c\u5e76\u751f\u6210\u591a\u4e2a\u6539\u5199\u7684\u601d\u7ef4\u94fe\uff08\u89c1\u56fe 7 step1\uff09\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u9009\u62e9\u601d\u7ef4\u94fe<\/strong>\uff1a\u901a\u8fc7\u7b54\u6848\u5339\u914d\uff0c\u7b5b\u9009\u51fa\u6539\u5199\u540e\u7b54\u6848\u4e0d\u53d8\u7684\u601d\u7ef4\u94fe\uff0c\u5f62\u6210\u7cbe\u70bc\u7684\u5019\u9009\u96c6\uff1b\u518d\u4ece\u4e2d\u968f\u673a\u9009\u53d6\u4e00\u5b9a\u6570\u91cf\u7684\u601d\u7ef4\u94fe\uff0c\u7528\u4e8e\u540e\u7eed\u7684\u63a8\u7406\uff08\u89c1\u56fe 7 step2\uff09\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u63a2\u7d22\u63a8\u7406\u94fe<\/strong>\uff1a\u5c06\u9009\u53d6\u7684\u601d\u7ef4\u94fe\u6392\u5217\u6210\u4e0d\u540c\u7684\u4e0a\u4e0b\u6587\uff0c\u4e0e\u76ee\u6807\u95ee\u9898\u4e00\u540c\u8f93\u5165\u7ed9 LLM\uff0c\u5e76\u91c7\u7528\u8f83\u9ad8\u7684\u6e29\u5ea6\u53c2\u6570\u8fdb\u884c\u591a\u6b21\u91cd\u590d\u63a8\u7406\uff0c\u4ee5\u63a2\u7d22\u591a\u6837\u7684\u63a8\u7406\u8def\u5f84\uff08\u89c1\u56fe 8 step3\uff09\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u7b54\u6848\u6295\u7968<\/strong>\uff1a\u5c06\u6240\u6709\u5f97\u5230\u7684\u7b54\u6848\u8fdb\u884c\u6295\u7968\uff0c\u5f97\u5230\u6700\u7ec8\u7b54\u6848\uff08\u89c1\u56fe 8 step4\uff09\u3002<\/p>\n<\/li>\n<\/ol>\n<section>\u5b8c\u6574\u7684 CD-CoT \u7b97\u6cd5\u8bf7\u89c1\u56fe 9\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-064d384491f3b0c02e5a4ea474ce17c6.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-064d384491f3b0c02e5a4ea474ce17c6.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 7. CD-CoT \u7b97\u6cd5\u7684\u6b65\u9aa4 1&amp;2<\/sup><\/em><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-0ac799e02e1b3ec4286f03b35fd56f3f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-0ac799e02e1b3ec4286f03b35fd56f3f.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 8. CD-CoT \u7b97\u6cd5\u7684\u6b65\u9aa4 3&amp;4<\/sup><\/em><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-cc163e3f92f0ed6fabd9b455e2bff79f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-cc163e3f92f0ed6fabd9b455e2bff79f.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 9. \u5b8c\u6574\u7684 CD-CoT \u7b97\u6cd5<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>CD-CoT \u5b9e\u9a8c\u7ed3\u679c<\/strong><\/section>\n<section><\/section>\n<section>\u6211\u4eec\u5728 NoRa \u6570\u636e\u96c6\u4e0a\u5168\u9762\u6d4b\u8bd5\u4e86 CD-CoT\uff0c\u5e76\u5bf9\u6bd4\u4e86\u591a\u4e2a\u9700\u8981\u989d\u5916\u76d1\u7763\u4fe1\u606f\u7684\u53bb\u566a\u65b9\u6cd5\uff08\u89c1\u56fe 10\uff09\uff0c\u6211\u4eec\u53d1\u73b0\uff1a<\/section>\n<section><\/section>\n<ol>\n<li>\n<p><strong>\u5f53\u9762\u5bf9\u566a\u58f0\u601d\u7ef4\u94fe\u65f6\uff0c\u4e0e base model \u76f8\u6bd4\uff0cCD-CoT \u5728\u6240\u6709\u6570\u636e\u96c6\u4e0a\u7684\u6027\u80fd\u5747\u6709\u663e\u8457\u63d0\u5347\uff0c\u51c6\u786e\u7387\u5e73\u5747\u63d0\u5347\u5e45\u5ea6\u8fbe\u5230 17.8%<\/strong>\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>CD-CoT \u5bf9\u9ad8\u566a\u58f0\u8868\u73b0\u51fa\u663e\u8457\u7684\u62b5\u6297\u529b\uff0c\u5c24\u5176\u5728\u66f4\u5177\u6311\u6218\u7684\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u4e2d<\/strong>\u3002<\/p>\n<\/li>\n<\/ol>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-142be7ab7f862e96ebe2c00d6632b559.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-142be7ab7f862e96ebe2c00d6632b559.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 10. \u5404\u79cd\u9700\u8981\u989d\u5916\u76d1\u7763\u4fe1\u606f\u7684\u65b9\u6cd5 \u5728 NoRa \u6570\u636e\u96c6\u4e0a\u7684\u6d4b\u8bc4\u7ed3\u679c<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c\u901a\u8fc7\u8bf8\u591a\u6d88\u878d\u5b9e\u9a8c\uff0c\u6211\u4eec\u53d1\u73b0\uff1a<\/section>\n<ol>\n<li>\n<p>\u5173\u4e8e CD-CoT \u8d85\u53c2\u6570\u7684\u6d88\u878d\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\uff0c<strong>\u5e72\u51c0\u7684\u601d\u7ef4\u94fe\u793a\u4f8b\u5728 CD-CoT \u4e2d\u626e\u6f14\u7740\u5173\u952e\u7684\u89d2\u8272<\/strong>\uff1b\u5f53\u53d8\u5316 N,M,C \u7b49\u8d85\u53c2\u6570\u7684\u53d6\u503c\u65f6\uff0c\u51c6\u786e\u6027\u4ec5\u5448\u73b0\u51fa\u7ec6\u5fae\u7684\u53d8\u5316\uff08\u89c1\u56fe 11\uff09\u3002\u5728\u8bba\u6587\u4e2d\uff0c\u6211\u4eec\u9ed8\u8ba4\u91c7\u7528 M \u8bbe\u4e3a 2 \u7684 CD-CoT \u793a\u4f8b\uff0c\u4ee5\u5728\u6548\u7387\u548c\u6548\u679c\u4e4b\u95f4\u53d6\u5f97\u5e73\u8861\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>CD-CoT \u5728\u4e0d\u540c\u7684 LLM \u4e0a\u5177\u6709\u826f\u597d\u7684\u6cdb\u5316\u80fd\u529b<\/strong>\uff0c\u4e0e base model\uff08GPT-3.5-Turbo \u548c Gemini-Pro\uff09\u76f8\u6bd4\uff0c\u5176\u51c6\u786e\u7387\u5206\u522b\u63d0\u9ad8\u4e86<strong>\u00a023.4%\u00a0<\/strong>\u548c <strong>21.6%<\/strong>\uff0c\u5e76\u8d85\u8d8a\u4e86\u6240\u6709\u57fa\u7ebf\u65b9\u6cd5\uff08\u89c1\u56fe 12\uff09\u3002<\/p>\n<\/li>\n<\/ol>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-772825b494c0fab1f71c952c75465702.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-772825b494c0fab1f71c952c75465702.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 11. \u5173\u4e8e CD-CoT \u8d85\u53c2\u6570\u7684\u6d88\u878d\u7814\u7a76<\/sup><\/em><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-0bb6579f6f1d205f11d4aa8103040f6f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-0bb6579f6f1d205f11d4aa8103040f6f.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 12. \u5173\u4e8e CD-CoT \u5728\u4e0d\u540c LLM \u4e0a\u7684\u6548\u679c\u7684\u6d88\u878d\u7814\u7a76<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u66f4\u591a\u7684\u5b9e\u9a8c\u5206\u6790\u548c\u6280\u672f\u7ec6\u8282\uff0c\u8bf7\u79fb\u6b65\u53c2\u9605\u6211\u4eec\u7684\u8bba\u6587\u53ca\u6e90\u7801\uff0c\u6211\u4eec\u4e5f\u5c06\u6301\u7eed\u66f4\u65b0\u672c\u5de5\u4f5c\u7684\u5185\u5bb9\u3002<\/section>\n<section><\/section>\n<section>\u6211\u4eec\u5e0c\u671b\u901a\u8fc7\u8fd9\u9879\u5de5\u4f5c\uff0c\u547c\u5401\u4eba\u4eec\u66f4\u591a\u5730\u5173\u6ce8 LLM \u63a8\u7406\u7684\u9c81\u68d2\u6027\u95ee\u9898\uff0c\u5e76\u5f00\u5c55\u5173\u4e8e\u5927\u6a21\u578b\u63a8\u7406\u9c81\u68d2\u6027\u7684\u63a2\u8ba8\u4e0e\u7814\u7a76\u3002\u975e\u5e38\u611f\u8c22\u5927\u5bb6\u5173\u6ce8\u6211\u4eec\u7684\u5de5\u4f5c\uff01<\/section>\n<section><\/section>\n<section><em><sup>\u53c2\u8003\u6587\u732e<\/sup><\/em><\/section>\n<section><sup><em>[1] Wei J, Wang X, Schuurmans D, et al. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS 2022.<\/em><\/sup><\/section>\n<section><sup><em>[2] Shi F, Chen X, Misra K, et al. Large language models can be easily distracted by irrelevant context. ICML 2023.<\/em><\/sup><\/section>\n<section><sup><em>[3] Tian Q, Zhu H, Wang L, et al. R3 Prompting: Review, Rephrase and Resolve for Chain-of-Thought Reasoning in Large Language Models under Noisy Context. EMNLP 2023.<\/em><\/sup><\/section>\n<section><sup><em>[4] Huang J, Chen X, Mishra S, et al. Large language models cannot self-correct reasoning yet. ICLR 2024.<\/em><\/sup><\/section>\n<section><sup><em>[5] Xi Z, Jin S, Zhou Y, et al. Self-polish: Enhance reasoning in large language models via problem refinement. EMNLP 2023.<\/em><\/sup><\/section>\n<section><sup><em>[6] Robey A, Wong E, Hassani H, et al. Smoothllm: Defending large language models against jailbreaking attacks. Arxiv 2023.<\/em><\/sup><\/section>\n<section><sup><em>[7] Zhang Z, Zhang G, Hou B, et al. Certified robustness for large language models with self-denoising. Arxiv 2023.<\/em><\/sup><\/section>\n<section><em><sup>[8] Wang X, Wei J, Schuurmans D, et al. Self-Consistency Improves Chain of Thought Reasoning in Language Models. ICLR 2023.<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u8bfe\u9898\u7ec4\u4ecb\u7ecd<\/strong><\/section>\n<section><\/section>\n<section>\u9999\u6e2f\u6d78\u4f1a\u5927\u5b66\u53ef\u4fe1\u673a\u5668\u5b66\u4e60\u548c\u63a8\u7406\u8bfe\u9898\u7ec4 (TMLR Group) \u7531\u591a\u540d\u9752\u5e74\u6559\u6388\u3001\u535a\u58eb\u540e\u7814\u7a76\u5458\u3001\u535a\u58eb\u751f\u3001\u8bbf\u95ee\u535a\u58eb\u751f\u548c\u7814\u7a76\u52a9\u7406\u5171\u540c\u7ec4\u6210\uff0c\u8bfe\u9898\u7ec4\u96b6\u5c5e\u4e8e\u7406\u5b66\u9662\u8ba1\u7b97\u673a\u7cfb\u3002\u8bfe\u9898\u7ec4\u4e13\u653b\u53ef\u4fe1\u8868\u5f81\u5b66\u4e60\u3001\u53ef\u4fe1\u57fa\u7840\u6a21\u578b\u3001\u57fa\u4e8e\u56e0\u679c\u63a8\u7406\u7684\u53ef\u4fe1\u5b66\u4e60\u7b49\u76f8\u5173\u7684\u7b97\u6cd5\uff0c\u7406\u8bba\u548c\u7cfb\u7edf\u8bbe\u8ba1\u4ee5\u53ca\u5728\u81ea\u7136\u79d1\u5b66\u4e0a\u7684\u5e94\u7528\uff0c\u5177\u4f53\u7814\u7a76\u65b9\u5411\u548c\u76f8\u5173\u6210\u679c\u8be6\u89c1\u672c\u7ec4 GitHub (https:\/\/github.com\/tmlr-group)\u3002<\/section>\n<section><\/section>\n<section>\u8bfe\u9898\u7ec4\u7531\u653f\u5e9c\u79d1\u7814\u57fa\u91d1\u4ee5\u53ca\u5de5\u4e1a\u754c\u79d1\u7814\u57fa\u91d1\u8d44\u52a9\uff0c\u5982\u9999\u6e2f\u7814\u7a76\u8d44\u52a9\u5c40\u6770\u51fa\u9752\u5e74\u5b66\u8005\u8ba1\u5212\uff0c\u56fd\u5bb6\u81ea\u7136\u79d1\u5b66\u57fa\u91d1\u9762\u4e0a\u9879\u76ee\u548c\u9752\u5e74\u9879\u76ee\uff0c\u4ee5\u53ca\u5fae\u8f6f\u3001\u82f1\u4f1f\u8fbe\u3001\u5b57\u8282\u8df3\u52a8\u3001\u767e\u5ea6\u3001\u963f\u91cc\u3001\u817e\u8baf\u7b49\u4f01\u4e1a\u7684\u79d1\u7814\u57fa\u91d1\u3002\u9752\u5e74\u6559\u6388\u548c\u8d44\u6df1\u7814\u7a76\u5458\u624b\u628a\u624b\u5e26\uff0cGPU \u8ba1\u7b97\u8d44\u6e90\u5145\u8db3\uff0c\u957f\u671f\u62db\u6536\u591a\u540d\u535a\u58eb\u540e\u7814\u7a76\u5458\u3001\u535a\u58eb\u751f\u3001\u7814\u7a76\u52a9\u7406\u548c\u7814\u7a76\u5b9e\u4e60\u751f\u3002\u6b64\u5916\uff0c\u672c\u7ec4\u4e5f\u6b22\u8fce\u81ea\u8d39\u7684\u8bbf\u95ee\u535a\u58eb\u540e\u7814\u7a76\u5458\u3001\u535a\u58eb\u751f\u548c\u7814\u7a76\u52a9\u7406\u7533\u8bf7\uff0c\u8bbf\u95ee\u81f3\u5c11 3-6 \u4e2a\u6708\uff0c\u652f\u6301\u8fdc\u7a0b\u8bbf\u95ee\u3002\u6709\u5174\u8da3\u7684\u540c\u5b66\u8bf7\u53d1\u9001\u4e2a\u4eba\u7b80\u5386\u548c\u521d\u6b65\u7814\u7a76\u8ba1\u5212\u5230\u90ae\u7bb1 (bhanml@comp.hkbu.edu.hk)\u3002<\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-12-12-2\" target=\"_blank\">NeurIPS 2024 | \u53ef\u4fe1\u5927\u6a21\u578b\u65b0\u6311\u6218\uff1a\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u4e0b\u7684\u9c81\u68d2\u63a8\u7406\uff0c\u51c6\u786e\u7387\u76f4\u964d40%<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:NeurIPS 202 [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68,55],"class_list":["post-37211","post","type-post","status-publish","format-standard","hentry","category-news","tag-68","tag-55"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>NeurIPS 2024 | \u53ef\u4fe1\u5927\u6a21\u578b\u65b0\u6311\u6218\uff1a\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u4e0b\u7684\u9c81\u68d2\u63a8\u7406\uff0c\u51c6\u786e\u7387\u76f4\u964d40% - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=37211\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=37211\",\"url\":\"https:\/\/17aitech.com\/?p=37211\",\"name\":\"NeurIPS 2024 | \u53ef\u4fe1\u5927\u6a21\u578b\u65b0\u6311\u6218\uff1a\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u4e0b\u7684\u9c81\u68d2\u63a8\u7406\uff0c\u51c6\u786e\u7387\u76f4\u964d40% - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=37211#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=37211#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8a58fd8142ee302c5fa15b5475a5859e.png\",\"datePublished\":\"2025-01-24T07:00:16+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=37211#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=37211\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=37211#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8a58fd8142ee302c5fa15b5475a5859e.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8a58fd8142ee302c5fa15b5475a5859e.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=37211#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"NeurIPS 2024 | \u53ef\u4fe1\u5927\u6a21\u578b\u65b0\u6311\u6218\uff1a\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u4e0b\u7684\u9c81\u68d2\u63a8\u7406\uff0c\u51c6\u786e\u7387\u76f4\u964d40%\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"NeurIPS 2024 | \u53ef\u4fe1\u5927\u6a21\u578b\u65b0\u6311\u6218\uff1a\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u4e0b\u7684\u9c81\u68d2\u63a8\u7406\uff0c\u51c6\u786e\u7387\u76f4\u964d40% - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=37211","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=37211","url":"https:\/\/17aitech.com\/?p=37211","name":"NeurIPS 2024 | \u53ef\u4fe1\u5927\u6a21\u578b\u65b0\u6311\u6218\uff1a\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u4e0b\u7684\u9c81\u68d2\u63a8\u7406\uff0c\u51c6\u786e\u7387\u76f4\u964d40% - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=37211#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=37211#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8a58fd8142ee302c5fa15b5475a5859e.png","datePublished":"2025-01-24T07:00:16+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=37211#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=37211"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=37211#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8a58fd8142ee302c5fa15b5475a5859e.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8a58fd8142ee302c5fa15b5475a5859e.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=37211#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"NeurIPS 2024 | \u53ef\u4fe1\u5927\u6a21\u578b\u65b0\u6311\u6218\uff1a\u566a\u58f0\u601d\u7ef4\u94fe\u63d0\u793a\u4e0b\u7684\u9c81\u68d2\u63a8\u7406\uff0c\u51c6\u786e\u7387\u76f4\u964d40%"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/37211","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=37211"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/37211\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=37211"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=37211"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=37211"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}