{"id":988,"date":"2024-12-06T16:42:49","date_gmt":"2024-12-06T08:42:49","guid":{"rendered":"http:\/\/tamanegi.xyz\/?p=988"},"modified":"2024-12-06T16:42:49","modified_gmt":"2024-12-06T08:42:49","slug":"how-is-chatgpts-behavior-changing-over-time","status":"publish","type":"post","link":"http:\/\/tamanegi.xyz\/?p=988","title":{"rendered":"How Is ChatGPT\u2019s Behavior Changing over Time"},"content":{"rendered":"<h1>Abstract<\/h1>\n<p>GPT-3.5\u548cGPT-4\u662f\u4e24\u79cd\u6700\u5e7f\u6cdb\u4f7f\u7528\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u670d\u52a1\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u6a21\u578b\u4f55\u65f6\u4ee5\u53ca\u5982\u4f55\u968f\u7740\u65f6\u95f4\u66f4\u65b0\u662f\u4e0d\u900f\u660e\u7684\u3002\u5728\u8fd9\u91cc\uff0c\u6211\u4eec\u8bc4\u4f30\u4e862023\u5e743\u6708\u548c2023\u5e746\u6708\u7248\u672c\u7684GPT-3.5\u548cGPT-4\u5728\u51e0\u4e2a\u591a\u6837\u5316\u7684\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\uff1a1) \u6570\u5b66\u95ee\u9898\uff0c2) \u654f\u611f\/\u5371\u9669\u7684\u95ee\u9898\uff0c3) \u610f\u89c1\u8c03\u67e5\uff0c4) \u591a\u6b65\u77e5\u8bc6\u5bc6\u96c6\u578b\u95ee\u9898\uff0c5) \u751f\u6210\u4ee3\u7801\uff0c6) \u7f8e\u56fd\u533b\u5b66\u6267\u7167\u8003\u8bd5\uff0c\u4ee5\u53ca7) \u89c6\u89c9\u63a8\u7406\u3002\u6211\u4eec\u53d1\u73b0\uff0cGPT-3.5\u548cGPT-4\u7684\u8868\u73b0\u548c\u884c\u4e3a\u968f\u65f6\u95f4\u53d8\u5316\u5f88\u5927\u3002\u4f8b\u5982\uff0c2023\u5e743\u6708\u7248\u672c\u7684GPT-4\u5728\u8bc6\u522b\u8d28\u6570\u4e0e\u5408\u6570\u95ee\u9898\u4e0a\u8868\u73b0\u51fa\u8272\uff08\u51c6\u786e\u7387\u4e3a84%\uff09\uff0c\u4f46\u5230\u4e862023\u5e746\u6708\u7248\u672c\uff0c\u5b83\u5728\u8fd9\u7c7b\u95ee\u9898\u4e0a\u7684\u8868\u73b0\u5374\u8f83\u5dee\uff08\u51c6\u786e\u7387\u964d\u81f351%\uff09\u3002\u8fd9\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u53ef\u4ee5\u89e3\u91ca\u4e3aGPT-4\u8ddf\u968f\u94fe\u5f0f\u601d\u7ef4\u63d0\u793a\u7684\u80fd\u529b\u4e0b\u964d\u3002\u6709\u8da3\u7684\u662f\uff0c\u5728\u8fd9\u4e2a\u4efb\u52a1\u4e2d\uff0cGPT-3.5\u57282023\u5e746\u6708\u7684\u8868\u73b0\u6bd42023\u5e743\u6708\u8981\u597d\u5f97\u591a\u3002GPT-4\u5728\u56de\u7b54\u654f\u611f\u95ee\u9898\u548c\u610f\u89c1\u8c03\u67e5\u65f6\u7684\u610f\u613f\u57282023\u5e746\u6708\u76f8\u6bd42023\u5e743\u6708\u6709\u6240\u4e0b\u964d\u3002GPT-4\u57282023\u5e746\u6708\u5904\u7406\u591a\u6b65\u95ee\u9898\u4e0a\u7684\u8868\u73b0\u4f18\u4e8e2023\u5e743\u6708\uff0c\u800cGPT-3.5\u5728\u8fd9\u9879\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u5219\u6709\u6240\u4e0b\u6ed1\u3002GPT-4\u548cGPT-3.5\u5728\u751f\u6210\u4ee3\u7801\u65f6\u7684\u683c\u5f0f\u9519\u8bef\u6570\u91cf\u57282023\u5e746\u6708\u76f8\u6bd42023\u5e743\u6708\u90fd\u6709\u6240\u589e\u52a0\u3002\u6211\u4eec\u63d0\u4f9b\u8bc1\u636e\u8868\u660e\uff0cGPT-4\u9075\u5faa\u7528\u6237\u6307\u4ee4\u7684\u80fd\u529b\u968f\u65f6\u95f4\u4e0b\u964d\uff0c\u8fd9\u662f\u5bfc\u81f4\u8bb8\u591a\u884c\u4e3a\u6f02\u79fb\u7684\u4e00\u4e2a\u5171\u540c\u56e0\u7d20\u3002\u603b\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u7684\u53d1\u73b0\u663e\u793a\uff0c\u201c\u76f8\u540c\u201d\u7684LLM\u670d\u52a1\u7684\u884c\u4e3a\u5728\u76f8\u5bf9\u8f83\u77ed\u7684\u65f6\u95f4\u5185\u53ef\u4ee5\u53d1\u751f\u663e\u8457\u53d8\u5316\uff0c\u8fd9\u5f3a\u8c03\u4e86\u6301\u7eed\u76d1\u6d4bLLMs\u7684\u5fc5\u8981\u6027\u3002<\/p>\n<h1>Introduction<\/h1>\n<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08Large Language Models\uff0cLLMs\uff09\u5982GPT-3.5\u548cGPT-4\u6b63\u88ab\u5e7f\u6cdb\u4f7f\u7528\u3002\u50cfGPT-4\u8fd9\u6837\u7684LLM\u53ef\u4ee5\u6839\u636e\u7528\u6237\u6570\u636e\u53cd\u9988\u4ee5\u53ca\u8bbe\u8ba1\u53d8\u66f4\u8fdb\u884c\u66f4\u65b0\u3002\u7136\u800c\uff0c\u76ee\u524d\u5c1a\u4e0d\u6e05\u695aGPT-3.5\u548cGPT-4\u5177\u4f53\u5728\u4f55\u65f6\u4f55\u5730\u8fdb\u884c\u4e86\u54ea\u4e9b\u66f4\u65b0\uff0c\u4e5f\u65e0\u6cd5\u660e\u786e\u8fd9\u4e9b\u66f4\u65b0\u5982\u4f55\u5f71\u54cd\u6a21\u578b\u7684\u884c\u4e3a\u8868\u73b0\u3002\u8fd9\u79cd\u4e0d\u786e\u5b9a\u6027\u4f7f\u5f97\u5c06LLMs\u7a33\u5b9a\u96c6\u6210\u5230\u66f4\u5e7f\u6cdb\u7684\u6d41\u7a0b\u4e2d\u53d8\u5f97\u6781\u5177\u6311\u6218\u6027\uff1a\u5982\u679cLLM\u5bf9\u67d0\u4e2a\u63d0\u793a\u7684\u54cd\u5e94\uff08\u4f8b\u5982\u51c6\u786e\u6027\u6216\u683c\u5f0f\uff09\u7a81\u7136\u53d1\u751f\u53d8\u5316\uff0c\u5219\u53ef\u80fd\u4f1a\u7834\u574f\u4e0b\u6e38\u7684\u5de5\u4f5c\u6d41\u3002\u6b64\u5916\uff0c\u8fd9\u4e5f\u4f7f\u5f97\u4ece\u201c\u540c\u4e00\u201dLLM\u4e2d\u91cd\u73b0\u7ed3\u679c\u53d8\u5f97\u66f4\u52a0\u56f0\u96be\uff0c\u751a\u81f3\u4e0d\u53ef\u80fd\u3002<br \/>\n\u9664\u4e86\u8fd9\u4e9b\u96c6\u6210\u6311\u6218\u4e4b\u5916\uff0c\u4e00\u4e2a\u6709\u8da3\u7684\u95ee\u9898\u662f\u50cfGPT-4\u8fd9\u6837\u7684LLM\u670d\u52a1\u662f\u5426\u5728\u968f\u7740\u65f6\u95f4\u6301\u7eed\u6539\u8fdb\u3002\u91cd\u8981\u7684\u662f\u8981\u77e5\u9053\u5bf9\u6a21\u578b\u8fdb\u884c\u7684\u66f4\u65b0\u662f\u5426\u65e8\u5728\u6539\u5584\u67d0\u4e9b\u65b9\u9762\u7684\u540c\u65f6\u53ef\u80fd\u524a\u5f31\u4e86\u5176\u4ed6\u65b9\u9762\u7684\u6027\u80fd\u3002<br \/>\n\u9274\u4e8e\u8fd9\u4e9b\u95ee\u9898\uff0c\u6211\u4eec\u8bc4\u4f30\u4e862023\u5e743\u6708\u548c2023\u5e746\u6708\u7248\u672c\u7684GPT-3.5\u53caGPT-4\u5728\u591a\u4e2a\u4efb\u52a1\u4e2d\u7684\u884c\u4e3a\u8868\u73b0\uff1a1\uff09\u89e3\u51b3\u6570\u5b66\u95ee\u9898\uff1b2\uff09\u56de\u7b54\u654f\u611f\u6216\u5371\u9669\u7684\u95ee\u9898\uff1b3\uff09\u56de\u7b54\u610f\u89c1\u8c03\u67e5\uff1b4\uff09\u56de\u7b54\u591a\u8df3\u77e5\u8bc6\u5bc6\u96c6\u578b\u95ee\u9898\uff1b5\uff09\u751f\u6210\u4ee3\u7801\uff1b6\uff09\u7f8e\u56fd\u533b\u5b66\u6267\u7167\u8003\u8bd5\uff1b7\uff09\u89c6\u89c9\u63a8\u7406\u3002\u8fd9\u4e9b\u4efb\u52a1\u88ab\u9009\u7528\u6765\u8bc4\u4f30\u8fd9\u4e9bLLM\u7684\u591a\u6837\u6027\u548c\u5b9e\u7528\u6027\u80fd\u529b\u3002\u6211\u4eec\u53d1\u73b0\uff0cGPT-3.5\u548cGPT-4\u5728\u8fd9\u4e24\u6b21\u53d1\u5e03\u7248\u672c\u4e2d\u7684\u6027\u80fd\u548c\u884c\u4e3a\u8868\u73b0\u5dee\u5f02\u663e\u8457\uff0c\u5e76\u4e14\u5b83\u4eec\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u968f\u7740\u65f6\u95f4\u7684\u63a8\u79fb\u660e\u663e\u53d8\u5dee\uff0c\u800c\u5728\u5176\u4ed6\u95ee\u9898\u4e0a\u5219\u6709\u6240\u6539\u5584\uff08\u5982\u56fe1\u6240\u793a\uff09\u3002<br \/>\n\u5982\u4f55\u89e3\u91ca\u8fd9\u4e9b\u6027\u80fd\u548c\u884c\u4e3a\u7684\u53d8\u5316\uff1f\u6211\u4eec\u5047\u8bbeChatGPT\u9075\u5faa\u7528\u6237\u6307\u4ee4\u7684\u80fd\u529b\u53ef\u80fd\u662f\u5bfc\u81f4\u5404\u4efb\u52a1\u4e2d\u6f02\u79fb\u7684\u5171\u540c\u56e0\u7d20\u3002\u4e3a\u4e86\u9a8c\u8bc1\u8fd9\u4e00\u5047\u8bbe\uff0c\u6211\u4eec\u9996\u5148\u6574\u7406\u4e86\u4e00\u5957\u4e0e\u5177\u4f53\u4efb\u52a1\u65e0\u5173\u7684\u6307\u4ee4\u96c6\uff0c\u5e76\u8bc4\u4f30\u4e862023\u5e743\u6708\u548c6\u6708\u7248\u672c\u7684GPT-4\u53caGPT-3.5\u5728\u8fd9\u5957\u6307\u4ee4\u4e0a\u7684\u8868\u73b0\u3002\u603b\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u89c2\u5bdf\u5230GPT-4\u9075\u5faa\u8bb8\u591a\u6307\u4ee4\u7684\u80fd\u529b\u663e\u8457\u4e0b\u964d\u3002\u57283\u6708\u4efd\uff0cGPT-4\u901a\u5e38\u80fd\u591f\u5f88\u597d\u5730\u9075\u5faa\u7528\u6237\u7684\u6307\u4ee4\uff08\u4f8b\u5982\u6309\u7167\u6307\u5b9a\u683c\u5f0f\u751f\u6210\u54cd\u5e94\uff09\uff0c\u4f46\u57286\u6708\u4efd\uff0c\u5b83\u5374\u65e0\u6cd5\u9075\u5faa\u5927\u591a\u6570\u8fd9\u4e9b\u7b80\u5355\u7684\u6307\u4ee4\uff08\u5982\u56fe1\u6240\u793a\uff09\u3002<br \/>\n![[How Is ChatGPTs Behavior Changing over Time-Fiigure 1 Overview.png]]<\/p>\n<blockquote>\n<p><strong>\u8bf4\u660e\uff1a<\/strong> \u56fe1\uff1aGPT-4\uff08\u5de6\u680f\uff09\u548cGPT-3.5\uff08\u53f3\u680f\uff09\u57282023\u5e743\u6708\u81f32023\u5e746\u6708\u4e4b\u95f4\u7684\u6027\u80fd\u6f02\u79fb\uff08a\uff09\u548c\u9075\u5faa\u6307\u4ee4\u80fd\u529b\u7684\u53d8\u5316\uff08b\uff09\u3002\u8bc4\u4ef7\u6307\u6807\u8d8a\u9ad8\u8d8a\u597d\u3002<br \/>\n\u5728\u516b\u4e2a\u591a\u6837\u5316\u7684\u4efb\u52a1\u4e0a\uff08\u8be6\u89c1\u4e0b\u6587\uff09\uff0c\u6a21\u578b\u7684\u6027\u80fd\u968f\u65f6\u95f4\u663e\u8457\u53d8\u5316\uff0c\u6709\u65f6\u751a\u81f3\u53d8\u5dee\u3002GPT-4\u9075\u5faa\u6307\u4ee4\u7684\u80fd\u529b\u968f\u65f6\u95f4\u4e0b\u964d\u4e0e\u5b83\u7684\u884c\u4e3a\u6f02\u79fb\u76f8\u5339\u914d\uff0c\u5e76\u90e8\u5206\u89e3\u91ca\u4e86\u76f8\u5e94\u7684\u6027\u80fd\u4e0b\u6ed1\u3002<\/p>\n<\/blockquote>\n<p>\u6211\u4eec\u7684\u7814\u7a76\u7ed3\u679c\u7a81\u663e\u4e86\u6301\u7eed\u76d1\u6d4b\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u884c\u4e3a\u53d8\u5316\u7684\u5fc5\u8981\u6027\u3002\u6211\u4eec\u5728\u672c\u6587\u4e2d\u6574\u7406\u7684\u6240\u6709\u63d0\u793a\u4ee5\u53caGPT-4\u548cGPT-3.5\u57282023\u5e743\u6708\u548c6\u6708\u7684\u54cd\u5e94\u6570\u636e\u5df2\u6536\u96c6\u5e76\u53d1\u5e03\u5728LLMDrift[<a href=\"https:\/\/github.com\/lchen001\/LLMDrift\">https:\/\/github.com\/lchen001\/LLMDrift<\/a>] \u3002\u6211\u4eec\u7684\u5206\u6790\u548c\u53ef\u89c6\u5316\u4ee3\u7801\u4e5f\u5df2\u7ecf\u5f00\u6e90\u3002\u6211\u4eec\u5e0c\u671b\u6211\u4eec\u7684\u5de5\u4f5c\u80fd\u591f\u6fc0\u53d1\u66f4\u591a\u5173\u4e8eLLM\u6f02\u79fb\u7684\u7814\u7a76\uff0c\u4ee5\u5b9e\u73b0\u503c\u5f97\u4fe1\u8d56\u4e14\u53ef\u9760\u7684LLM\u5e94\u7528\u3002<\/p>\n<h2>Related Work<\/h2>\n<p>\u5df2\u6709\u591a\u4e2a\u9488\u5bf9\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08\u5305\u62ecGPT-3.5\u548cGPT-4\uff09\u7684\u57fa\u51c6\u6d4b\u8bd5\u548c\u8bc4\u4f30\u7814\u7a76[LBL+22, ZPM+23, LNT+23, BCL+23]\u3002\u73b0\u6709\u7814\u7a76\u8868\u660e\uff0cLLM\u5728\u4f20\u7edf\u7684\u8bed\u8a00\u4efb\u52a1\u5982\u9605\u8bfb\u7406\u89e3[dW23]\u3001\u7ffb\u8bd1[JWH+23]\u548c\u6458\u8981\u751f\u6210[GLD22]\u4e0a\u8868\u73b0\u51fa\u5408\u7406\u7684\u6027\u80fd\u3002\u6700\u8fd1\u7684\u7814\u7a76\u8868\u660e\uff0cGPT-4\u5728\u533b\u5b66[NKM+23]\u548c\u6cd5\u5f8b\u9886\u57df[KBGA23]\u7b49\u4e13\u4e1a\u9886\u57df\u7684\u56f0\u96be\u8003\u8bd5\u4e2d\u4e5f\u53d6\u5f97\u4e86\u6210\u529f\u3002\u636e\u6211\u4eec\u6240\u77e5\uff0c\u5927\u591a\u6570\u8fd9\u4e9b\u7814\u7a76\u5e76\u6ca1\u6709\u7cfb\u7edf\u5730\u76d1\u6d4b\u5e7f\u6cdb\u4f7f\u7528\u7684LLM\u670d\u52a1\u968f\u65f6\u95f4\u7684\u957f\u671f\u6f02\u79fb\u6216\u62a5\u544a\u663e\u8457\u7684\u53d8\u5316\u3002<\/p>\n<p>ChatLog [TLY+23] \u63d0\u51fa\u4e86\u81ea\u52a8\u8bb0\u5f55\u5e76\u76d1\u63a7ChatGPT\u54cd\u5e94\u7684\u65b9\u6cd5\uff0c\u5e76\u5728\u4e00\u4e9b\u5e38\u89c1\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u62a5\u544a\u4e86\u8f83\u5c0f\u7684\u6027\u80fd\u53d8\u5316\uff08\u591a\u6570\u4f4e\u4e8e5%\uff09\u3002\u5176\u4ed6\u8bba\u6587[AAKA23, SKNM23]\u4e5f\u62a5\u544a\u4e86\u4e00\u4e9b\u7279\u5b9a\u95ee\u9898\u4e0a\u7684\u53d8\u5316\u3002\u76d1\u6d4b\u6a21\u578b\u6027\u80fd\u7684\u53d8\u5316\u662f\u673a\u5668\u5b66\u4e60\u4f5c\u4e3a\u670d\u52a1\uff08MLaaS\uff09\u9886\u57df\u7684\u65b0\u5174\u7814\u7a76\u9886\u57df\u3002[CJE+22] \u63d0\u4f9b\u4e86\u4e00\u4e2a\u5927\u89c4\u6a21\u7684\u957f\u671f\u6570\u636e\u96c6\uff0c\u8bb0\u5f55\u4e86\u5404\u79cd\u8bc4\u4f30\u4efb\u52a1\u4e2d\u5546\u4e1aML\u670d\u52a1\u7684\u54cd\u5e94\uff0c\u800c[CCZZ21] \u5219\u7814\u7a76\u5982\u4f55\u9ad8\u6548\u4f30\u8ba1ML\u670d\u52a1\u6027\u80fd\u7684\u53d8\u5316\u3002\u8fd9\u4e9b\u8bba\u6587\u4e3b\u8981\u5173\u6ce8\u7b80\u5355\u7684\u5206\u7c7b\u4efb\u52a1\uff0c\u5982\u60c5\u611f\u5206\u6790\u7b49ML\u670d\u52a1\uff0c\u800c\u672c\u5de5\u4f5c\u5219\u4e13\u6ce8\u4e8e\u751f\u6210\u5f0fLLM\u670d\u52a1\u3002<br \/>\n![[How Is ChatGPTs Behavior Changing over Time-Figure 2.png]]<\/p>\n<blockquote>\n<p>\u56fe2\uff1aGPT-4\u548cGPT-3.5\u57282023\u5e743\u6708\u548c2023\u5e746\u6708\u7248\u672c\u4e0a\u7684\u516b\u9879\u4efb\u52a1\u6027\u80fd\uff1a<br \/>\n(a,b) \u89e3\u51b3\u6570\u5b66\u95ee\u9898\uff08\u8d28\u6570\u4e0e\u5408\u6570\u3001\u5feb\u4e50\u6570\u5b57\uff09\uff0c (c) \u56de\u7b54\u654f\u611f\u95ee\u9898\uff0c (d) \u610f\u89c1\u8c03\u67e5\uff0c (e) \u8fd0\u884cLangChain\u5e94\u7528\u8fdb\u884c\u591a\u8df3\u95ee\u7b54\uff0c (f) \u751f\u6210\u53ef\u6267\u884c\u4ee3\u7801\uff0c (g) \u7f8e\u56fd\u6267\u4e1a\u533b\u5e08\u8d44\u683c\u8003\u8bd5(USMLE)\uff0c\u4ee5\u53ca (h) \u89c6\u89c9\u63a8\u7406\u3002<br \/>\n\u5bf9\u4e8e\u6bcf\u4e00\u9879\u4efb\u52a1\uff0c\u56fe\u4e2d\u7528\u7d2b\u8272\u6846\u5c55\u793a\u4e86\u4e00\u4e2a\u793a\u4f8b\uff0c\u5e76\u5728\u6807\u9898\u4e2d\u6807\u6ce8\u4e86\u6837\u672c\u6570\u91cfn\u3002\u6a21\u578b\u7684\u6027\u80fd\u968f\u65f6\u95f4\u53d8\u5316\u663e\u8457\uff0c\u6709\u65f6\u751a\u81f3\u4f1a\u53d8\u5dee\u3002<\/p>\n<\/blockquote>\n<h1>Overview: LLM Services, Tasks and Metrics<\/h1>\n<p>\u672c\u6587\u7814\u7a76\u4e86\u4e0d\u540c\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u884c\u4e3a\u968f\u65f6\u95f4\u7684\u53d8\u5316\u3002\u4e3a\u4e86\u5b9a\u91cf\u56de\u7b54\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u9700\u8981\u660e\u786e\u4ee5\u4e0b\u51e0\u70b9\uff1a (i) \u76d1\u63a7\u54ea\u4e9bLLM\u670d\u52a1\uff0c (ii) \u5173\u6ce8\u54ea\u4e9b\u5e94\u7528\u573a\u666f\uff0c\u4ee5\u53ca (iii) \u5982\u4f55\u5728\u6bcf\u4e2a\u573a\u666f\u4e2d\u6d4b\u91cfLLM\u6f02\u79fb\u3002<\/p>\n<p><strong>LLM Services<\/strong><br \/>\n\u672c\u6587\u76d1\u63a7\u7684LLM\u670d\u52a1\u662fGPT-4\u548cGPT-3.5\uff0c\u5b83\u4eec\u6784\u6210\u4e86ChatGPT\u7684\u6838\u5fc3\u3002\u7531\u4e8eChatGPT\u7684\u6d41\u884c\uff0cGPT-4\u548cGPT-3.5\u5df2\u7ecf\u88ab\u5e7f\u5927\u4e2a\u4eba\u7528\u6237\u548c\u8bb8\u591a\u4f01\u4e1a\u5e7f\u6cdb\u91c7\u7528\u3002\u56e0\u6b64\uff0c\u53ca\u65f6\u4e14\u7cfb\u7edf\u5730\u76d1\u6d4b\u8fd9\u4e24\u79cd\u670d\u52a1\u6709\u52a9\u4e8e\u5927\u91cf\u7528\u6237\u66f4\u597d\u5730\u7406\u89e3\u548c\u5229\u7528LLM\u4ee5\u6ee1\u8db3\u81ea\u5df1\u7684\u9700\u6c42\u6848\u4f8b\u3002\u5728\u64b0\u5199\u672c\u6587\u65f6\uff0c\u901a\u8fc7OpenAI\u7684API\u53ef\u4ee5\u83b7\u53d6\u5230GPT-4\u548cGPT-3.5\u4e24\u4e2a\u4e3b\u8981\u7248\u672c\uff0c\u4e00\u4e2a\u662f\u57282023\u5e743\u6708\u5feb\u7167\u7684\u65f6\u95f4\u70b9\uff0c\u53e6\u4e00\u4e2a\u662f\u57282023\u5e746\u6708\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u91cd\u70b9\u5173\u6ce8\u8fd9\u4e24\u4e2a\u65f6\u95f4\u70b9\u4e4b\u95f4\u7684\u6f02\u79fb\u53d8\u5316\u3002<\/p>\n<p><strong>Evaluation Tasks<\/strong><br \/>\n\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u91cd\u70b9\u5173\u6ce8\u516b\u4e2a\u7ecf\u5e38\u7528\u4e8e\u6027\u80fd\u548c\u5b89\u5168\u57fa\u51c6\u6d4b\u8bd5\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u4efb\u52a1\uff1a\u89e3\u51b3\u6570\u5b66\u95ee\u9898\uff08\u5305\u62ec\u4e24\u79cd\u7c7b\u578b\u7684\u95ee\u9898\uff09\u3001\u56de\u7b54\u654f\u611f\u95ee\u9898\u3001\u56de\u7b54OpinionQA\u8c03\u67e5\u3001LangChain HotpotQA Agent\u3001\u4ee3\u7801\u751f\u6210\u3001\u53c2\u52a0USMLE\u533b\u5b66\u8003\u8bd5\u4ee5\u53ca\u89c6\u89c9\u63a8\u7406\uff0c\u5982\u56fe1\u6240\u793a\u3002\u8fd9\u4e9b\u4efb\u52a1\u4e4b\u6240\u4ee5\u88ab\u9009\u4e2d\uff0c\u6709\u4e24\u4e2a\u539f\u56e0\u3002\u9996\u5148\uff0c\u5b83\u4eec\u662f\u7ecf\u5e38\u5728\u6587\u732e\u4e2d\u7528\u4e8e\u8bc4\u4f30LLM\u7684\u591a\u6837\u5316\u4efb\u52a1[WWS+22, ZPM+23, CTJ+21]\u3002\u5176\u6b21\uff0c\u8fd9\u4e9b\u4efb\u52a1\u76f8\u5bf9\u5ba2\u89c2\u4e14\u6613\u4e8e\u8bc4\u4ef7\u3002<\/p>\n<p>\u5bf9\u4e8e\u6bcf\u4e2a\u4efb\u52a1\uff0c\u6211\u4eec\u4f7f\u7528\u4ece\u73b0\u6709\u6570\u636e\u96c6\u4e2d\u62bd\u6837\u6216\u7531\u6211\u4eec\u6784\u5efa\u7684\u95ee\u9898\u8fdb\u884c\u67e5\u8be2\u3002\u6211\u4eec\u627f\u8ba4\u8fd9\u91cc\u4f7f\u7528\u7684\u5177\u4f53\u57fa\u51c6\u6570\u636e\u96c6\u5e76\u4e0d\u80fd\u5168\u9762\u6db5\u76d6ChatGPT\u7684\u590d\u6742\u884c\u4e3a\u3002\u6211\u4eec\u7684\u76ee\u6807\u4e0d\u662f\u63d0\u4f9b\u4e00\u4e2a\u5168\u9762\u7684\u8bc4\u4f30\uff0c\u800c\u662f\u5c55\u793a\u5728\u7b80\u5355\u4efb\u52a1\u4e0a\u5b58\u5728\u663e\u8457\u7684ChatGPT\u6027\u80fd\u6f02\u79fb\u73b0\u8c61\u3002\u5728\u672a\u6765\u7684\u7814\u7a76\u4e2d\uff0c\u6211\u4eec\u5c06\u589e\u52a0\u66f4\u591a\u7684\u57fa\u51c6\u6d4b\u8bd5\u4f5c\u4e3a\u957f\u671f\u7814\u7a76\u7684\u4e00\u90e8\u5206\u3002<\/p>\n<p>\u6211\u4eec\u5728\u4e0b\u4e00\u8282\u4e2d\u8be6\u7ec6\u8ba8\u8bba\u6bcf\u4e2a\u4efb\u52a1\u3002<\/p>\n<p><strong>Metrics<\/strong><br \/>\n\u4e3a\u4e86\u5b9a\u91cf\u5efa\u6a21\u548c\u6d4b\u91cf\u4e0d\u540c\u4efb\u52a1\u4e2d\u7684LLM\u6f02\u79fb\uff0c\u6211\u4eec\u4e3a\u6bcf\u4e2a\u4efb\u52a1\u786e\u5b9a\u4e00\u4e2a\u4e3b\u8981\u6027\u80fd\u6307\u6807\uff0c\u5e76\u4e3a\u6240\u6709\u4efb\u52a1\u8bbe\u5b9a\u4e24\u4e2a\u5e38\u89c1\u7684\u8f85\u52a9\u6307\u6807\u3002\u524d\u8005\u6355\u6349\u7279\u5b9a\u573a\u666f\u4e0b\u7684\u6027\u80fd\u5ea6\u91cf\uff0c\u800c\u540e\u8005\u8986\u76d6\u8de8\u4e0d\u540c\u5e94\u7528\u7684\u901a\u7528\u8865\u5145\u5ea6\u91cf\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\uff1a<\/p>\n<ul>\n<li>\u5bf9\u4e8e\u6570\u5b66\u95ee\u9898\u548cUSMLE\u95ee\u9898\uff0c\u6211\u4eec\u5c06\u51c6\u786e\u6027\uff08LLM\u670d\u52a1\u751f\u6210\u6b63\u786e\u7b54\u6848\u7684\u9891\u7387\uff09\u4f5c\u4e3a\u6211\u4eec\u7684\u4e3b\u8981\u6307\u6807\u3002<\/li>\n<li>\u5728\u56de\u7b54\u654f\u611f\u6027\u548c\u610f\u89c1\u6027\u95ee\u9898\u65f6\uff0c\u6211\u4eec\u4f7f\u7528\u54cd\u5e94\u7387\uff0c\u5373LLM\u670d\u52a1\u76f4\u63a5\u56de\u7b54\u95ee\u9898\u7684\u9891\u7387\u3002<\/li>\n<li>\u5bf9\u4e8e\u4ee3\u7801\u751f\u6210\u4efb\u52a1\uff0c\u4e3b\u8981\u6307\u6807\u662f\u8f93\u51fa\u4e2d\u53ef\u4ee5\u76f4\u63a5\u6267\u884c\u7684\u6bd4\u4f8b\uff08\u5982\u679c\u4ee3\u7801\u53ef\u4ee5\u5728\u7f16\u7a0b\u73af\u5883\u4e2d\u76f4\u63a5\u8fd0\u884c\u5e76\u901a\u8fc7\u5355\u5143\u6d4b\u8bd5\uff09\u3002<\/li>\n<li>\u5bf9\u4e8e\u89c6\u89c9\u63a8\u7406\u548cLangChain\u4efb\u52a1\uff0c\u4e3b\u8981\u6307\u6807\u662f\u7cbe\u786e\u5339\u914d\uff08\u6700\u7ec8\u56de\u590d\u662f\u5426\u4e0e\u771f\u5b9e\u7ed3\u679c\u5b8c\u5168\u4e00\u81f4\uff09\u3002<\/li>\n<\/ul>\n<p>\u6211\u4eec\u7684\u7b2c\u4e00\u4e2a\u5e38\u89c1\u8f85\u52a9\u6307\u6807\u662f\u5197\u957f\u5ea6\uff0c\u5373\u751f\u6210\u5185\u5bb9\u7684\u5b57\u7b26\u957f\u5ea6\u3002\u7b2c\u4e8c\u4e2a\u6307\u6807\u662f\u4e0d\u4e00\u81f4\u6027\uff0c\u5373\u5bf9\u4e8e\u76f8\u540c\u7684\u63d0\u793a\uff0c\u5728\u4e24\u4e2a\u7248\u672c\u7684LLM\u670d\u52a1\u4e2d\u63d0\u53d6\u7684\u7b54\u6848\u4e0d\u5339\u914d\u7684\u9891\u7387\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u4e2a\u6307\u6807\u4ec5\u6bd4\u8f83\u7b54\u6848\u4e4b\u95f4\u7684\u5dee\u5f02\uff0c\u800c\u4e0d\u662f\u539f\u59cb\u751f\u6210\u7684\u5185\u5bb9\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8e\u6570\u5b66\u95ee\u9898\uff0c\u5982\u679c\u751f\u6210\u7684\u7b54\u6848\u76f8\u540c\uff0c\u5219\u5373\u4f7f\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u4e0d\u540c\uff0c\u4e5f\u4e0d\u4e00\u81f4\u503c\u4e3a0\u3002<\/p>\n<p>\u5bf9\u4e8e\u6bcf\u4e2aLLM\u670d\u52a1\uff0c\u6211\u4eec\u4f7f\u7528\u6574\u4e2a\u6837\u672c\u7fa4\u4f53\u4e2d\u7684\u4e0d\u4e00\u81f4\u6027\u7ecf\u9a8c\u5e73\u5747\u503c\u6765\u91cf\u5316\u5176\u6240\u9700\u529f\u80fd\u968f\u65f6\u95f4\u7684\u53d8\u5316\u60c5\u51b5\uff08\u800c\u975e\u6587\u672c\u8f93\u51fa\uff09\u3002\u66f4\u5927\u7684\u4e0d\u4e00\u81f4\u6027\u610f\u5473\u7740\u66f4\u5927\u7684\u6f02\u79fb\u3002\u5bf9\u4e8e\u5176\u4ed6\u6307\u6807\uff0c\u6211\u4eec\u5c063\u6708\u548c6\u6708\u7248\u672c\u7684\u7fa4\u4f53\u5747\u503c\u8fdb\u884c\u6bd4\u8f83\uff0c\u5e76\u5229\u7528\u8fd9\u4e9b\u5dee\u5f02\u6765\u8861\u91cf\u6f02\u79fb\u5927\u5c0f\u3002<\/p>\n<h1>3 Monitoring Reveals Substantial LLM Drifts<\/h1>\n<h2>3.1 \u6570\u5b66\u95ee\u9898I\uff08\u8d28\u6570\u4e0e\u5408\u6570\uff09\uff1a\u63a8\u7406\u94fe\u53ef\u80fd\u5931\u6548<\/h2>\n<p>GPT-4 \u548c GPT-3.5 \u7684\u6570\u5b66\u89e3\u9898\u80fd\u529b\u968f\u65f6\u95f4\u5982\u4f55\u6f14\u53d8\uff1f\u4f5c\u4e3a\u4e00\u9879\u7ecf\u5178\u7814\u7a76\uff0c\u6211\u4eec\u63a2\u7d22\u4e86\u8fd9\u4e9b\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5728\u5224\u65ad\u7ed9\u5b9a\u6574\u6570\u662f\u8d28\u6570\u8fd8\u662f\u5408\u6570\u65b9\u9762\u7684\u53d8\u5316\u3002\u6211\u4eec\u5173\u6ce8\u8fd9\u4e00\u4efb\u52a1\u662f\u56e0\u4e3a\u5b83\u5bf9\u4eba\u7c7b\u6765\u8bf4\u6613\u4e8e\u7406\u89e3\u4f46\u4ecd\u7136\u9700\u8981\u63a8\u7406\u8fc7\u7a0b\uff0c\u7c7b\u4f3c\u4e8e\u8bb8\u591a\u6570\u5b66\u95ee\u9898\u3002\u6570\u636e\u96c6\u5305\u542b1,000\u4e2a\u95ee\u9898\uff0c\u5176\u4e2d500\u4e2a\u8d28\u6570\u4ece[ZPM+23]\u4e2d\u63d0\u53d6\uff0c\u53e6\u5916500\u4e2a\u5408\u6570\u5219\u5728\u533a\u95f4[1,000, 20,000]\u5185\u5747\u5300\u91c7\u6837\u3002\u4e3a\u4e86\u5e2e\u52a9LLM\u8fdb\u884c\u63a8\u7406\uff0c\u6211\u4eec\u4f7f\u7528\u4e86Chain-of-Thought (CoT) \u63d0\u793a\u6cd5[WWS+22]\uff0c\u8fd9\u662f\u4e00\u79cd\u5904\u7406\u9700\u8981\u5927\u91cf\u63a8\u7406\u4efb\u52a1\u7684\u6807\u51c6\u65b9\u6cd5\u3002<br \/>\n\u6216\u8bb8\u4ee4\u4eba\u60ca\u8bb6\u7684\u662f\uff0c\u5728\u8fd9\u4e2a\u7b80\u5355\u7684\u4efb\u52a1\u4e0a\uff0c\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u8868\u73b0\u51fa\u73b0\u4e86\u663e\u8457\u53d8\u5316\u3002\u5982\u56fe3(a)\u6240\u793a\uff0cGPT-4\u7684\u51c6\u786e\u7387\u4ece3\u6708\u4efd\u768484.0%\u4e0b\u964d\u52306\u6708\u4efd\u768451.1%\uff0c\u800cGPT-3.5\u7684\u51c6\u786e\u7387\u5219\u4ece49.6%\u63d0\u5347\u5230\u4e8676.2%\u3002\u6b64\u5916\uff0cGPT-4\u7684\u56de\u7b54\u53d8\u5f97\u66f4\u52a0\u7b80\u6d01\uff1a\u5b83\u751f\u6210\u7684\u5b57\u7b26\u6570\u91cf\u5e73\u5747\u503c\u4ece3\u6708\u4efd\u7684638.3\u51cf\u5c11\u52306\u6708\u4efd\u76843.9\u3002\u53e6\u4e00\u65b9\u9762\uff0cGPT-3.5\u7684\u56de\u7b54\u957f\u5ea6\u589e\u957f\u4e86\u7ea622.2%\u3002\u4e24\u4e2a\u7248\u672c\uff083\u6708\u548c6\u6708\uff09\u4e4b\u95f4\u7684\u7b54\u6848\u4e0d\u5339\u914d\u60c5\u51b5\u5728\u4e24\u79cd\u670d\u52a1\u4e2d\u90fd\u975e\u5e38\u663e\u8457\u3002<br \/>\n\u4e3a\u4ec0\u4e48\u5dee\u8ddd\u5982\u6b64\u4e4b\u5927\uff1f\u4e00\u4e2a\u53ef\u80fd\u7684\u89e3\u91ca\u662f\u94fe\u5f0f\u601d\u8003\uff08CoT\uff09\u884c\u4e3a\u7684\u53d8\u5316\u3002\u56fe3(b)\u7ed9\u51fa\u4e86\u4e00\u4e2a\u8bf4\u660e\u6027\u4f8b\u5b50\u3002\u4e3a\u4e86\u786e\u5b9a17077\u662f\u5426\u4e3a\u8d28\u6570\uff0cGPT-4\u7684\u4e09\u6708\u7248\u672c\u5f88\u597d\u5730\u9075\u5faa\u4e86CoT\u6307\u4ee4\u3002\u5b83\u9996\u5148\u5c06\u4efb\u52a1\u5206\u89e3\u6210\u56db\u4e2a\u6b65\u9aa4\uff1a\u68c0\u67e517077\u662f\u5426\u4e3a\u5076\u6570\uff0c\u627e\u523017077\u7684\u5e73\u65b9\u6839\uff0c\u83b7\u53d6\u5c0f\u4e8e\u5b83\u7684\u6240\u6709\u7d20\u6570\uff0c\u5e76\u68c0\u67e517077\u662f\u5426\u53ef\u4ee5\u88ab\u8fd9\u4e9b\u6570\u4e2d\u7684\u4efb\u4f55\u4e00\u4e2a\u6574\u9664\u3002\u7136\u540e\u5b83\u6267\u884c\u4e86\u6bcf\u4e2a\u6b65\u9aa4\uff0c\u6700\u7ec8\u5f97\u51fa\u6b63\u786e\u7684\u7ed3\u8bba\uff0c\u537317077\u786e\u5b9e\u662f\u4e00\u4e2a\u8d28\u6570\u3002\u7136\u800c\uff0c\u94fe\u5f0f\u601d\u8003\u5728\u516d\u6708\u7248\u672c\u4e2d\u6ca1\u6709\u8d77\u4f5c\u7528\uff1a\u5c3d\u7ba1\u63d0\u793a\u8981\u6c42\u9010\u6b65\u601d\u8003\uff0c\u670d\u52a1\u5e76\u6ca1\u6709\u751f\u6210\u4efb\u4f55\u4e2d\u95f4\u6b65\u9aa4\uff0c\u800c\u662f\u76f4\u63a5\u7ed9\u51fa\u4e86\u201c\u5426\u201d\u7684\u7b54\u6848\u3002\u5bf9\u4e8eGPT-3.5\u6765\u8bf4\uff0cCoT\u7684\u5f71\u54cd\u6709\u4e0d\u540c\u7684\u6f02\u79fb\u6a21\u5f0f\u3002\u4e09\u6708\u4efd\uff0cGPT-3.5\u503e\u5411\u4e8e\u5148\u7ed9\u51fa\u7b54\u6848\u201c\u5426\u201d\uff0c\u7136\u540e\u518d\u8fdb\u884c\u63a8\u7406\u6b65\u9aa4\u3002\u56e0\u6b64\uff0c\u5373\u4f7f\u6b65\u9aa4\u548c\u6700\u7ec8\u7ed3\u8bba\uff08\u201c17077\u662f\u8d28\u6570\u201d\uff09\u90fd\u662f\u6b63\u786e\u7684\uff0c\u5176\u540d\u4e49\u4e0a\u7684\u7b54\u6848\u4ecd\u7136\u662f\u9519\u8bef\u7684\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u516d\u6708\u66f4\u65b0\u4f3c\u4e4e\u89e3\u51b3\u4e86\u8fd9\u4e2a\u95ee\u9898\uff1a\u5b83\u9996\u5148\u5199\u51fa\u4e86\u63a8\u7406\u6b65\u9aa4\uff0c\u6700\u540e\u751f\u6210\u4e86\u6b63\u786e\u7b54\u6848\u201c\u662f\u201d\u3002\u8fd9\u4e00\u6709\u8da3\u7684\u73b0\u8c61\u8868\u660e\uff0c\u5373\u4f7f\u662f\u5e7f\u6cdb\u91c7\u7528\u7684\u94fe\u5f0f\u601d\u8003\u7b56\u7565\uff0c\u5728LLM\u6f02\u79fb\u7684\u5f71\u54cd\u4e0b\u4e5f\u53ef\u80fd\u5bfc\u81f4\u663e\u8457\u4e0d\u540c\u7684\u6027\u80fd\u8868\u73b0\u3002<br \/>\n\u4e3a\u4e86\u8fdb\u4e00\u6b65\u8c03\u67e5\u94fe\u5f0f\u601d\u8003\uff08CoT\uff09\u884c\u4e3a\u53d8\u5316\u7684\u5f71\u54cd\uff0c\u6211\u4eec\u6bd4\u8f83\u4e86GPT-4\u548cGPT-3.5\u5728\u6709\u65e0\u660e\u786e\u7684CoT\u6307\u4ee4\u7684\u60c5\u51b5\u4e0b\u5bf9\u540c\u4e00\u95ee\u9898\u7684\u56de\u7b54\u3002\u5bf9\u4e8e\u540e\u8005\uff0c\u6211\u4eec\u53ea\u662f\u8981\u6c42\u6a21\u578b\u7ed9\u51fa\u4e00\u4e2a\u4e8c\u5143\u751f\u6210\u7ed3\u679c\uff0c\u800c\u6ca1\u6709\u660e\u786e\u8981\u6c42\u5b83\u8fdb\u884c\u9010\u6b65\u63a8\u7406\uff08\u4f8b\u5982\uff0c\u201c17077\u662f\u8d28\u6570\u5417\uff1f\u56de\u7b54\u2018[Yes]\u2019\u6216\u2018[No]\u2019\u3002\u201d\uff09\u3002<br \/>\n\u5982\u88681\u6240\u793a\uff0c\u4f7f\u7528CoT\u5c06GPT-4\u5728\u4e09\u6708\u4efd\u7684\u6027\u80fd\u4ece59.6%\u63d0\u5347\u5230\u4e8684.0%\uff0c\u63d0\u9ad8\u4e8624.4\u4e2a\u767e\u5206\u70b9\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u5bf9\u4e8e\u516d\u6708\u7248\u672c\u7684GPT-4\uff0cCoT\u5e76\u6ca1\u6709\u5e26\u6765\u592a\u591a\u5e2e\u52a9\uff1a\u51c6\u786e\u6027\u5b9e\u9645\u4e0a\u964d\u4f4e\u4e860.1%\u3002\u5982\u524d\u6240\u8ff0\uff0c\u8fd9\u662f\u56e0\u4e3a\u65b0\u7248\u672c\u6ca1\u6709\u9075\u5faaCoT\u6307\u4ee4\u3002\u800c\u5bf9\u4e8eGPT-3.5\u6765\u8bf4\uff0c\u5219\u89c2\u5bdf\u5230\u76f8\u53cd\u7684\u8d8b\u52bf\uff1a\u5728\u4e09\u6708\u4efd\u901a\u8fc7\u6dfb\u52a0CoT\uff0c\u51c6\u786e\u7387\u7565\u6709\u63d0\u9ad8\uff08+6.3%\uff09\uff0c\u4f46\u5728\u516d\u6708\u5219\u663e\u8457\u63d0\u9ad8\u4e86\uff08+15.8%\uff09\u3002\u7531\u4e8eGPT-3.5\u5728\u4e09\u6708\u548c\u516d\u6708\u90fd\u9075\u5faa\u4e86CoT\u6307\u4ee4\u3002\u8fd9\u8868\u660eLLM\u6f02\u79fb\u53ef\u80fd\u4f1a\u6539\u53d8\u6a21\u578b\u662f\u5426\u4ee5\u53ca\u5982\u4f55\u9075\u5faa\u7528\u6237\u6307\u4ee4\u7684\u65b9\u5f0f\u3002<br \/>\n\u5230\u76ee\u524d\u4e3a\u6b62\uff0c\u6211\u4eec\u7684\u5206\u6790\u4e3b\u8981\u57fa\u4e8e\u4e3b\u8981\u6307\u6807\u51c6\u786e\u7387\u7684\u53d8\u5316\uff0c\u4f46\u7ec6\u7c92\u5ea6\u7684\u8c03\u67e5\u53ef\u80fd\u4f1a\u63ed\u793a\u5176\u4ed6\u6709\u8da3\u7684\u6f02\u79fb\u6a21\u5f0f\u3002\u4e00\u4e2a\u89c2\u5bdf\u7ed3\u679c\u662f\uff0cGPT-4\u7684\u516d\u6708\u7248\u672c\u503e\u5411\u4e8e\u5c06\u6574\u6570\u89c6\u4e3a\u5408\u6570\u3002\u4e3a\u4e86\u9a8c\u8bc1\u8fd9\u4e00\u70b9\uff0c\u6211\u4eec\u91cf\u5316\u4e86\u6df7\u6dc6\u77e9\u9635\u968f\u65f6\u95f4\u53d8\u5316\u7684\u60c5\u51b5\u3002\u5982\u56fe4(c)\u6240\u793a\uff0cGPT-4\u7684\u516d\u6708\u7248\u672c\u51e0\u4e4e\u603b\u662f\u5c06\u6574\u6570\u8bc6\u522b\u4e3a\u5408\u6570\uff0849.9%+48.8%=99.7%\uff09\u3002GPT-3.5\u7684\u4e09\u6708\u7248\u672c\u4e5f\u5b58\u5728\u7c7b\u4f3c\u95ee\u9898\u3002GPT-4\u5728\u4e09\u6708\u548cGPT-3.5\u5728\u516d\u6708\u7684\u6df7\u6dc6\u77e9\u9635\u770b\u8d77\u6765\u66f4\u4e3a\u5408\u7406\uff1a\u5b83\u4eec\u5728\u5927\u591a\u6570\u60c5\u51b5\u4e0b\u80fd\u591f\u6b63\u786e\u8bc6\u522b\u8d28\u6570\u548c\u5408\u6570\u3002\u7c7b\u4f3c\u7684\u53d1\u73b0\u6700\u8fd1\u5df2\u88ab\u62a5\u9053[ NK23]\u3002<br \/>\n\u5b9e\u4f8b\u7ea7\u522b\u7684\u8c03\u67e5\u63ed\u793a\u4e86\u66f4\u591a\u89c1\u89e3\u3002\u6211\u4eec\u5747\u5300\u968f\u673a\u62bd\u53d6\u4e868\u4e2a\u6570\u5b57\uff0c\u624b\u52a8\u68c0\u67e5\u4e09\u6708\u7248\u672cGPT-4\u7684\u9010\u6b65\u903b\u8f91\u3002\u5728\u8fd9\u4e9b\u6848\u4f8b\u4e2d\uff0c\u4e09\u6708\u7248\u672c\u7684GPT-4\u7684CoT\u903b\u8f91\u5927\u591a\u662f\u5408\u7406\u7684\uff0c\u5c3d\u7ba1\u5b83\u53ef\u80fd\u4f1a\u72af\u7b97\u672f\u9519\u8bef\uff0c\u57288\u4e2a\u4f8b\u5b50\u4e2d\u67097\u4e2a\u5f97\u51fa\u4e86\u6b63\u786e\u7684\u7ed3\u8bba\u3002\u4ee5\u4e0b\u662f4\u4e2a\u793a\u4f8b\uff0c\u5176\u4f59\u7684\u653e\u5728\u9644\u5f55\u4e2d\u3002<\/p>\n<p>\u4f8b\u5982\uff0c\u4ee5\u4e0b\u7531GPT-4\u751f\u6210\u7684\u56de\u7b54\u7ed9\u51fa\u4e86\u5bf919997\u7684\u6b63\u786e\u63a8\u7406\u6b65\u9aa4\uff1a<\/p>\n<p>$$<br \/>\n\\begin{array}{l}<br \/>\n\\text{\u4e3a\u4e86\u786e\u5b9a } 19997 \\text{ \u662f\u5426\u4e3a\u8d28\u6570\uff0c\u6211\u4eec\u9996\u5148\u68c0\u67e5\u5b83\u662f\u5426\u4e3a\u5076\u6570\uff08\u4e0d\u662f\uff09\u3002} \\<br \/>\n\\text{\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u8ba1\u7b97\u5b83\u7684\u5e73\u65b9\u6839\u5e76\u68c0\u67e5\u5c0f\u4e8e\u8be5\u5e73\u65b9\u6839\u7684\u6240\u6709\u7d20\u6570\uff1a} \\<br \/>\n\\sqrt{19997} \\approx 141.4, \\text{ \u56e0\u6b64\u6211\u4eec\u9700\u8981\u68c0\u67e5\u6240\u6709\u5c0f\u4e8e\u7b49\u4e8e } 141 \\text{ \u7684\u8d28\u6570\u3002} \\<br \/>\n\\text{\u7ecf\u8fc7\u9010\u4e00\u9a8c\u8bc1\uff0c\u6ca1\u6709\u53d1\u73b0\u4efb\u4f55\u80fd\u6574\u9664 } 19997 \\text{ \u7684\u8d28\u6570\uff0c} \\<br \/>\n\\text{\u56e0\u6b64\u5b83\u662f\u4e00\u4e2a\u8d28\u6570\u3002}<br \/>\n\\end{array}<br \/>\n$$<br \/>\n\u591a\u6b21\u8be2\u95ee\u76f8\u540c\u7684\u67e5\u8be2\u4f1a\u5f97\u51fa\u76f8\u540c\u7684\u7ed3\u679c\uff0c\u5c3d\u7ba1\u4e2d\u95f4\u7684\u63a8\u7406\u6b65\u9aa4\u53ef\u80fd\u4f1a\u6709\u6240\u4e0d\u540c\u3002\u4f8b\u5982\uff0c\u5728\u4e09\u6708\u7248\u672c\u4e2d\uff0cGPT-4\u751f\u6210\u7684\u7b2c\u4e8c\u4e2a\u7b54\u6848\u66f4\u52a0\u8be6\u5c3d\uff0c\u56e0\u4e3a\u5b83\u660e\u786e\u68c0\u67e5\u4e86\u6bcf\u4e2a\u5c0f\u4e8e19997\u7684\u8d28\u6570\u662f\u5426\u80fd\u6574\u966419997\uff08\u89c1\u9644\u5f55\uff09\u3002\u6b64\u5916\uff0c\u4e09\u6708\u4efd\u7684GPT-4\u8fd8\u80fd\u6b63\u786e\u8bc6\u522b\u4ee52\u62165\u4e3a\u56e0\u5b50\u7684\u5408\u6570\u3002\u4f8b\u5982\uff0c\u5728\u786e\u5b9a16450\u662f\u5426\u4e3a\u8d28\u6570\u65f6\uff0c\u4e09\u6708\u7248\u672c\u7684GPT-4\u7ed9\u51fa\u4e86\u4ee5\u4e0b\u56de\u7b54\u3002<br \/>\n![[How Is ChatGPTs Behavior Changing over Time-Extract1.png]]<br \/>\n\u7c7b\u4f3c\u5730\uff0c\u5728\u68c0\u67e59175\u65f6\uff0cGPT-4\u5728\u4e09\u6708\u7ed9\u51fa\u4e86\u4ee5\u4e0b\u63a8\u7406\u6b65\u9aa4\u3002<br \/>\n![[How Is ChatGPTs Behavior Changing over Time-Extract2.png]]<br \/>\n\u5728\u8fd9\u4e24\u79cd\u60c5\u51b5\u4e0b\uff0c\u4e09\u6708\u7248\u672c\u7684GPT-4\u901a\u8fc7\u9010\u6b65\u903b\u8f91\u5f97\u51fa\u4e86\u6b63\u786e\u7684\u7b54\u6848\u3002\u7136\u800c\uff0c\u5728\u5904\u7406\u5176\u4ed6\u4e00\u4e9b\u67e5\u8be2\u65f6\uff0c\u5b83\u4e5f\u72af\u4e86\u4e00\u4e9b\u7ec6\u5fae\u7684\u9519\u8bef\u3002\u4f8b\u5982\uff0c\u5728\u68c0\u67e54323\u662f\u5426\u4e3a\u8d28\u6570\u65f6\uff0c\u5b83\u751f\u6210\u4e86\u4ee5\u4e0b\u5185\u5bb9\u3002<br \/>\n![[How Is ChatGPTs Behavior Changing over Time-Extract3.png]]<br \/>\n\u5728\u8fd9\u91cc\uff0c\u5b83\u5b9e\u9645\u4e0a\u4f7f\u7528\u4e86\u6b63\u786e\u7684\u903b\u8f91\uff08\u68c0\u67e54323\u662f\u5426\u80fd\u88ab\u4efb\u4f55\u8d28\u6570\u6574\u9664\uff09\uff0c\u4f46\u5728\u8ba1\u7b97\u65f6\u72af\u4e86\u4e00\u4e2a\u9519\u8bef\uff084323\u9664\u4ee53\u7b49\u4e8e1441\uff0c\u6070\u597d\u6ca1\u6709\u4f59\u6570\uff09\u3002<\/p>\n<h2>3.2 \u6570\u5b66II\uff08\u8ba1\u6570\u5feb\u4e50\u6570\uff09\uff1aCoT\u6548\u679c\u8f83\u5dee<\/h2>\n<p>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u7814\u7a76ChatGPT\u5728\u89e3\u51b3\u6570\u5b66\u95ee\u9898\u548c\u94fe\u5f0f\u601d\u7ef4\u884c\u4e3a\u65b9\u9762\u7684\u80fd\u529b\uff0c\u6211\u4eec\u8981\u6c42\u5b83\u5904\u7406\u4e00\u4e2a\u4e0d\u540c\u7684\u6570\u5b66\u95ee\u9898\uff1a\u8ba1\u7b97\u7ed9\u5b9a\u533a\u95f4\u5185\u7684\u5feb\u4e50\u6570\u7684\u6570\u91cf\u3002\u5982\u679c\u5c06\u6574\u6570\u66ff\u6362\u4e3a\u5176\u6570\u5b57\u5e73\u65b9\u7684\u603b\u548c\uff0c\u91cd\u590d\u8fdb\u884c\u76f4\u5230\u6700\u7ec8\u7ed3\u679c\u4e3a1\uff0c\u5219\u8be5\u6574\u6570\u88ab\u79f0\u4e3a\u5feb\u4e50\u6570\u3002\u4f8b\u5982\uff0c$13$\u662f\u4e00\u4e2a\u5feb\u4e50\u6570\uff0c\u56e0\u4e3a$1^2+3^2=10$\uff0c\u63a5\u7740 $1^2+0^2=1$\u3002<br \/>\n\u8fd9\u9879\u4efb\u52a1\u4e0e\u7d20\u6570\u6d4b\u8bd5\u4e0d\u540c\uff0c\u56e0\u4e3a\u5b83\u8981\u6c42\u63d0\u4f9b\u4e00\u4e2a\u5b9a\u91cf\u7684\u7b54\u6848\uff08\u5feb\u4e50\u6570\u7684\u6570\u91cf\uff09\uff0c\u800c\u4e0d\u662f\u4e8c\u5143\u51b3\u7b56\uff08\u5982\u8d28\u6570\u6216\u5408\u6570\uff09\u3002\u6b64\u5916\uff0c\u5b83\u4ec5\u4f7f\u7528\u7b80\u5355\u7684\u7b97\u672f\u8fd0\u7b97\u3002\u4e3a\u4e86\u8bc4\u4f30LLM\u5728\u8be5\u4efb\u52a1\u4e0a\u7684\u6f02\u79fb\u73b0\u8c61\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u5305\u542b500\u4e2a\u67e5\u8be2\u7684\u6570\u636e\u96c6\u3002\u6bcf\u4e2a\u67e5\u8be2\u8be2\u95ee\u7ed9\u5b9a\u533a\u95f4\u5185\u6709\u591a\u5c11\u4e2a\u5feb\u4e50\u6570\uff0c\u5e76\u7edf\u8ba1LLM\u51c6\u786e\u56de\u7b54\u6b63\u786e\u6570\u91cf\u7684\u9891\u7387\u3002\u533a\u95f4\u7684\u5927\u5c0f\u4ece6\u523010\u4e4b\u95f4\u5747\u5300\u968f\u673a\u9009\u62e9\uff0c\u8d77\u59cb\u70b9\u5219\u4ece500\u523015,000\u4e4b\u95f4\u5747\u5300\u968f\u673a\u9009\u53d6\u3002\u4e3a\u4e86\u9f13\u52b1\u903b\u8f91\u63a8\u7406\u6b65\u9aa4\uff0c\u6211\u4eec\u518d\u6b21\u91c7\u7528\u4e86CoT\u63d0\u793a\u3002<br \/>\n\u6211\u4eec\u8fd8\u89c2\u5bdf\u5230\u5728\u8fd9\u4e00\u4efb\u52a1\u4e0a\u5b58\u5728\u663e\u8457\u7684\u6027\u80fd\u6f02\u79fb\u3002\u5982\u56fe5(a)\u6240\u793a\uff0cGPT-4\u7684\u51c6\u786e\u7387\u4ece\u4e09\u6708\u768483.6%\u4e0b\u964d\u5230\u4e86\u516d\u6708\u768435.2%\u3002\u53e6\u4e00\u65b9\u9762\uff0cGPT-3.5\u7684\u51c6\u786e\u7387\u5219\u4ece30.6%\u4e0a\u5347\u523048.2%\u3002\u751f\u6210\u7684\u56de\u7b54\u4e2d\u8fd8\u5b58\u5728\u5927\u91cf\u5b57\u7b26\u91cf\u7684\u53d8\u5316\uff08verbosity\uff09\u3002GPT-4\u7684\u751f\u6210\u957f\u5ea6\u4ece\u4e09\u6708\u76842163.5\u4e2a\u5b57\u7b26\u4e0b\u964d\u5230\u4e86\u516d\u6708\u768410.0\u4e2a\u5b57\u7b26\uff0c\u800cGPT-3.5\u7684\u957f\u5ea6\u5219\u589e\u52a0\u4e86\u8d85\u8fc760%\u3002\u4e0e\u7d20\u6570\u6d4b\u8bd5\uff08Math I\uff09\u76f8\u6bd4\uff0c\u7531\u4e8e\u9700\u8981\u66f4\u591a\u6b65\u9aa4\u6765\u679a\u4e3e\u533a\u95f4\u5185\u7684\u6570\u5b57\u5e76\u53cd\u590d\u8ba1\u7b97\u5e73\u65b9\u548c\uff0c\u56e0\u6b64\u7b54\u6848\u7684\u5e73\u5747\u957f\u5ea6\u663e\u8457\u589e\u52a0\u3002\u6b64\u5916\uff0c\u5728\u4e09\u6708\u5230\u516d\u6708\u4e4b\u95f4\uff0cGPT-4\u6700\u7ec8\u7b54\u6848\u7684\u53d8\u5316\u7387\u4e3a67.6%\uff0c\u800cGPT-3.5\u5219\u4e3a77.2%\u3002<br \/>\n\u4e3a\u4e86\u8fdb\u4e00\u6b65\u4e86\u89e3\u94fe\u5f0f\u601d\u7ef4\uff08CoT\uff09\u6548\u679c\u7684\u53d8\u5316\uff0c\u6211\u4eec\u8981\u6c42\u6bcf\u4e2a\u670d\u52a1\u5728\u6709\u6216\u6ca1\u6709CoT\u63d0\u793a\u7684\u60c5\u51b5\u4e0b\u56de\u7b54\u76f8\u540c\u7684\u67e5\u8be2\uff0c\u5e76\u7814\u7a76\u4e86\u4f7f\u7528CoT\u5e26\u6765\u7684\u51c6\u786e\u6027\u63d0\u5347\u60c5\u51b5\u3002\u6211\u4eec\u53d1\u73b0CoT\u7684\u76ca\u5904\u4e5f\u53d1\u751f\u4e86\u53d8\u5316\u3002\u4f8b\u5982\uff0c\u5bf9\u4e8eGPT-4\uff0c\u5728\u4e09\u6708\u65f6CoT\u5e26\u6765\u4e8656.6%\u7684\u51c6\u786e\u7387\u63d0\u5347\uff0c\u4f46\u5728\u516d\u6708\u65f6\u4ec5\u63d0\u5347\u4e863.2%\uff0c\u5982\u88682\u6240\u793a\u3002\u800c\u5bf9\u4e8eGPT-3.5\uff0c\u5728\u516d\u6708\u65f6CoT\u5bfc\u81f4\u4e8620.6%\u7684\u6027\u80fd\u63d0\u5347\u3002\u7136\u800c\u5728\u4e09\u6708\u65f6\uff0c\u4f7f\u7528CoT\u53cd\u800c\u5bfc\u81f4\u4e861.6%\u7684\u51c6\u786e\u7387\u4e0b\u964d\u3002<br \/>\nGPT-4 \u548c GPT-3.5 \u7684\u9519\u8bef\u6570\u91cf\u968f\u65f6\u95f4\u53d1\u751f\u4e86\u53d8\u5316\u3002\u4f46\u5b83\u4eec\u5177\u4f53\u51fa\u73b0\u4e86\u54ea\u4e9b\u65b0\u9519\u8bef\u5462\uff1f\u4e3a\u4e86\u56de\u7b54\u8fd9\u4e2a\u95ee\u9898\uff0c\u6211\u4eec\u5bf9\u8fd9\u4e9b\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u6df7\u6dc6\u77e9\u9635\u8fdb\u884c\u4e86\u7ec6\u7c92\u5ea6\u5206\u6790\uff0c\u5982\u56fe 6 \u6240\u793a\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cGPT-4 \u548c GPT-3.5 \u7684\u504f\u5dee\u968f\u65f6\u95f4\u7684\u53d8\u5316\u60c5\u51b5\u3002\u4f8b\u5982\uff0c\u5728\u516d\u6708\uff0cGPT-4 \u575a\u4fe1\u5728\u4e00\u4e2a\u7ed9\u5b9a\u533a\u95f4\u5185\u53ea\u5b58\u5728 0 \u6216 1 \u4e2a\u5feb\u4e50\u6570\uff08happy number\uff09\u3002\u53e6\u4e00\u65b9\u9762\uff0cGPT-3.5 \u5728\u516d\u6708\u503e\u5411\u4e8e\u9ad8\u4f30\u6570\u91cf\uff1a\u5728\u8d85\u8fc7 10% \u7684\u67e5\u8be2\u4e2d\uff0c\u5b83\u56de\u7b54\u8bf4\u6709\u8d85\u8fc7 4 \u4e2a\u5feb\u4e50\u6570\uff0c\u800c\u5b9e\u9645\u4e0a 4 \u662f\u6240\u6709\u6211\u4eec\u67e5\u8be2\u4e2d\u7684\u4e0a\u9650\u3002\u6211\u4eec\u4e5f\u9488\u5bf9\u5feb\u4e50\u6570\u8fdb\u884c\u4e86\u66f4\u5c0f\u533a\u95f4\u8303\u56f4\u7684\u989d\u5916\u5b9e\u9a8c\uff0c\u5e76\u89c2\u5bdf\u5230\u4e86\u7c7b\u4f3c\u7684\u8d8b\u52bf\uff08\u89c1\u9644\u5f55\uff09\u3002<br \/>\n![[How Is ChatGPTs Behavior Changing over Time-Figure6.png]]<\/p>\n<blockquote>\n<p>\u56fe6\uff1a\u8ba1\u7b97\u5feb\u4e50\u6570\u7684\u6df7\u6dc6\u77e9\u9635\u53d8\u5316\u3002GPT-4 \u7684\u4e09\u6708\u7248\u672c\u5728\u5927\u591a\u6570\u67e5\u8be2\u4e2d\u6b63\u786e\u5730\u8ba1\u7b97\u4e86\u5feb\u4e50\u6570\u7684\u6570\u91cf\uff0c\u800c\u5176\u516d\u6708\u7248\u672c\u5219\u5927\u591a\u56de\u7b54\u8bf4\u5728\u4e00\u4e2a\u7ed9\u5b9a\u533a\u95f4\u5185\u53ea\u6709\u4e00\u4e2a\u5feb\u4e50\u6570\u3002<\/p>\n<\/blockquote>\n<h2>3.3 \u56de\u7b54\u654f\u611f\u95ee\u9898\uff1a\u66f4\u5b89\u5168\u4f46\u7406\u7531\u8f83\u5c11<\/h2>\n<h2>3.4 \u610f\u89c1\u95ee\u7b54\u8c03\u67e5\uff1a\u54cd\u5e94\u7387\u8f83\u4f4e<\/h2>\n<h2>3.5 \u4ee3\u7801\u751f\u6210\uff1a\u683c\u5f0f\u5316\u6307\u4ee4\u9075\u5faa\u7a0b\u5ea6\u8f83\u4f4e<\/h2>\n<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u4e00\u4e2a\u4e3b\u8981\u5e94\u7528\u662f\u4ee3\u7801\u751f\u6210[CTJ+21]\u3002\u867d\u7136\u5b58\u5728\u8bb8\u591a\u4ee3\u7801\u751f\u6210\u6570\u636e\u96c6[CTJ+21, YZY+18, AON+21]\uff0c\u4f46\u4f7f\u7528\u8fd9\u4e9b\u6570\u636e\u96c6\u6765\u8bc4\u4f30LLM\u670d\u52a1\u7684\u4ee3\u7801\u751f\u6210\u80fd\u529b\u4f1a\u9762\u4e34\u6570\u636e\u6c61\u67d3\u7684\u95ee\u9898\u3002\u4e3a\u514b\u670d\u8fd9\u4e00\u95ee\u9898\uff0c\u6211\u4eec\u6784\u5efa\u4e86\u4e00\u4e2a\u65b0\u7684\u4ee3\u7801\u751f\u6210\u6570\u636e\u96c6\u3002\u8be5\u6570\u636e\u96c6\u5305\u542b\u64b0\u5199\u65f6LeetCode\u201c\u7b80\u5355\u201d\u7c7b\u522b\u4e2d\u7684\u6700\u65b050\u4e2a\u95ee\u9898\u3002\u6700\u65e9\u7684\u516c\u5f00\u89e3\u51b3\u65b9\u6848\u548c\u8ba8\u8bba\u4e8e2022\u5e7412\u6708\u53d1\u5e03\u3002\u6bcf\u4e2a\u95ee\u9898\u7684\u63d0\u793a\u662f\u539f\u95ee\u9898\u63cf\u8ff0\u4e0e\u76f8\u5e94Python\u4ee3\u7801\u6a21\u677f\u7684\u62fc\u63a5\u3002\u6bcf\u79cdLLM\u7684\u751f\u6210\u7ed3\u679c\u76f4\u63a5\u53d1\u9001\u5230LeetCode\u5728\u7ebf\u88c1\u5224\u8fdb\u884c\u8bc4\u4f30\u3002\u5982\u679c\u5728\u7ebf\u88c1\u5224\u63a5\u53d7\u7b54\u6848\uff08\u5373\uff0c\u7b54\u6848\u4e3a\u6709\u6548Python\u4ee3\u7801\u5e76\u901a\u8fc7\u5176\u6d4b\u8bd5\uff09\uff0c\u5219\u79f0\u5176\u4e3a\u53ef\u76f4\u63a5\u6267\u884c\u7684\u3002<br \/>\n\u603b\u4f53\u800c\u8a00\uff0c\u53ef\u76f4\u63a5\u6267\u884c\u7684\u751f\u6210\u7ed3\u679c\u6570\u91cf\u4ece\u4e09\u6708\u5230\u516d\u6708\u4e0b\u964d\u3002\u5982\u56fe9(a)\u6240\u793a\uff0cGPT-4\u5728\u4e09\u6708\u4efd\u8d85\u8fc750%\u7684\u751f\u6210\u7ed3\u679c\u662f\u53ef\u76f4\u63a5\u6267\u884c\u7684\uff0c\u4f46\u5728\u516d\u6708\u8fd9\u4e00\u6bd4\u4f8b\u964d\u81f3\u4ec510%\u3002\u5bf9\u4e8eGPT-3.5\uff0c\u8d8b\u52bf\u4e5f\u7c7b\u4f3c\u3002\u540c\u65f6\uff0c\u4e24\u4e2a\u6a21\u578b\u7684\u5197\u957f\u7a0b\u5ea6\u90fd\u6709\u5c0f\u5e45\u589e\u52a0\u3002<br \/>\n\u4e3a\u4ec0\u4e48\u53ef\u76f4\u63a5\u6267\u884c\u7684\u751f\u6210\u7ed3\u679c\u6570\u91cf\u4e0b\u964d\u4e86\uff1f\u4e00\u4e2a\u53ef\u80fd\u7684\u89e3\u91ca\u662f\uff0c\u516d\u6708\u7248\u672c\u7684\u6a21\u578b\u5728\u751f\u6210\u4e2d\u6301\u7eed\u6dfb\u52a0\u4e86\u989d\u5916\u7684\u975e\u4ee3\u7801\u6587\u672c\u3002\u56fe9(b)\u7ed9\u51fa\u4e86\u4e00\u4f8b\u8fd9\u79cd\u60c5\u51b5\u3002GPT-4\u4e09\u6708\u4efd\u548c\u516d\u6708\u4efd\u7684\u751f\u6210\u5185\u5bb9\u51e0\u4e4e\u76f8\u540c\uff0c\u9664\u4e86\u4e24\u4e2a\u90e8\u5206\u3002\u9996\u5148\uff0c\u516d\u6708\u7248\u672c\u5728\u4ee3\u7801\u7247\u6bb5\u524d\u540e\u5206\u522b\u589e\u52a0\u4e86\u201c\u2018\u2018\u2018python\u201d\uff08\u53ef\u80fd\u662f\u4e3a\u4e86\u5728UI\u4e2d\u5c06\u5176\u683c\u5f0f\u5316\u4e3aMarkdown\uff09\u3002\u5176\u6b21\uff0c\u5b83\u8fd8\u751f\u6210\u4e86\u66f4\u591a\u7684\u6ce8\u91ca\u3002\u867d\u7136\u8fd9\u4e9b\u53d8\u5316\u5f88\u5c0f\uff0c\u4f46\u989d\u5916\u7684\u4e09\u4e2a\u5f15\u53f7\u4f7f\u4ee3\u7801\u65e0\u6cd5\u6267\u884c\u3002\u8fd9\u79cd\u683c\u5f0f\u884c\u4e3a\u7684\u53d8\u5316\u5728LLM\u751f\u6210\u7684\u4ee3\u7801\u88ab\u7528\u4e8e\u66f4\u5927\u7684\u8f6f\u4ef6\u7ba1\u9053\u5185\u90e8\u65f6\u5c24\u5176\u96be\u4ee5\u68c0\u6d4b\u5230\u3002<br \/>\n\u6211\u4eec\u8fd8\u7814\u7a76\u4e86\u5728\u53bb\u9664\u975e\u4ee3\u7801\u6587\u672c\u7684\u989d\u5916\u540e\u5904\u7406\u4e4b\u540e\uff0c\u751f\u6210\u7684\u4ee3\u7801\u662f\u5426\u80fd\u591f\u901a\u8fc7LeetCode\u6d4b\u8bd5\u3002\u5982\u88684\u6240\u793a\uff0c\u6027\u80fd\u518d\u6b21\u51fa\u73b0\u4e86\u663e\u8457\u53d8\u5316\uff1aGPT-4\u7684\u8868\u73b0\u4ece52%\u63d0\u9ad8\u523070%\uff0c\u800cGPT-3.5\u63d0\u9ad8\u4e862%\u3002\u867d\u7136\u4ee3\u7801\u7684\u6b63\u786e\u6027\u6709\u6240\u63d0\u5347\uff0c\u4f46\u672a\u80fd\u9075\u5faa\u683c\u5f0f\u5316\u6307\u4ee4\uff08\u201c\u4ec5\u751f\u6210\u4ee3\u7801\u201d\uff09\u4ecd\u7136\u662f\u4e24\u4e2aGPT\u6a21\u578b\u7248\u672c\u4e4b\u95f4\u7684\u4e00\u4e2a\u95ee\u9898\u884c\u4e3a\u53d8\u5316\u3002<br \/>\n![[How Is ChatGPTs Behavior Changing over Time-Figure9.png]]<\/p>\n<blockquote>\n<p>\u56fe9\uff1a\u4ee3\u7801\u751f\u6210\u3002(a) \u603b\u4f53\u6027\u80fd\u6f02\u79fb\u3002\u5bf9\u4e8eGPT-4\uff0c\u53ef\u76f4\u63a5\u6267\u884c\u7684\u751f\u6210\u7ed3\u679c\u6bd4\u4f8b\u4ece\u4e09\u6708\u768452.0%\u4e0b\u964d\u5230\u516d\u6708\u768410.0%\u3002GPT-3.5\u7684\u8868\u73b0\u4e5f\u5927\u5e45\u4e0b\u964d\uff08\u4ece22.0%\u964d\u81f32.0%\uff09\u3002\u6b64\u5916\uff0c\u6839\u636e\u751f\u6210\u4ee3\u7801\u4e2d\u7684\u5b57\u7b26\u6570\u91cf\u6d4b\u91cf\uff0cGPT-4\u7684\u5197\u957f\u7a0b\u5ea6\u589e\u52a0\u4e8620%\u3002(b) \u4e00\u4e2a\u793a\u4f8b\u67e5\u8be2\u53ca\u5176\u5bf9\u5e94\u7684\u54cd\u5e94\u3002\u5728\u4e09\u6708\uff0c\u65e0\u8bba\u662fGPT-4\u8fd8\u662fGPT-3.5\u90fd\u9075\u5faa\u4e86\u7528\u6237\u6307\u4ee4\uff08\u201c\u4ec5\u8f93\u51fa\u4ee3\u7801\u201d\uff09\uff0c\u56e0\u6b64\u4ea7\u751f\u7684\u7ed3\u679c\u53ef\u4ee5\u76f4\u63a5\u6267\u884c\u3002\u7136\u800c\uff0c\u5728\u516d\u6708\uff0c\u5b83\u4eec\u5728\u4ee3\u7801\u7247\u6bb5\u524d\u540e\u6dfb\u52a0\u4e86\u989d\u5916\u7684\u4e09\u4e2a\u5f15\u53f7\uff0c\u5bfc\u81f4\u4ee3\u7801\u65e0\u6cd5\u76f4\u63a5\u6267\u884c\u3002<\/p>\n<\/blockquote>\n<h2>3.6 LangChain HotpotQA \u4ee3\u7406\uff1a\u63d0\u793a\u7a33\u5b9a\u6027\u8f83\u5dee<\/h2>\n<p>\u8bb8\u591a\u5b9e\u9645\u5e94\u7528\u8981\u6c42\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u57fa\u4e8e\u5404\u79cd\u6570\u636e\u6e90\u56de\u7b54\u77e5\u8bc6\u5bc6\u96c6\u578b\u95ee\u9898\uff0c\u5305\u62ec\u6d89\u53ca\u591a\u4e2a\u6765\u6e90\u548c\/\u6216\u591a\u6b65\u63a8\u7406\u7684\u201c\u591a\u8df3\u201d\u95ee\u9898\u3002\u56e0\u6b64\uff0c\u76d1\u6d4bLLM\u56de\u7b54\u591a\u8df3\u95ee\u9898\u7684\u80fd\u529b\u968f\u65f6\u95f4\u5982\u4f55\u6f14\u53d8\u662f\u5f88\u81ea\u7136\u7684\u4e8b\u60c5\u3002\u6211\u4eec\u901a\u8fc7\u6d4b\u91cfLangChain HotpotQA\u4ee3\u7406[Tea23]\u7684\u65f6\u95f4\u6f02\u79fb\u6765\u8fc8\u51fa\u7b2c\u4e00\u6b65\uff0c\u8fd9\u662f\u4e00\u4e2a\u7528\u4e8e\u56de\u7b54\u7c7b\u4f3cHotpotQA [YQZ+18]\u590d\u6742\u591a\u8df3\u95ee\u9898\u7684\u7ba1\u9053\u3002\u8be5\u4ee3\u7406\u5229\u7528LLMs\u5728Wikipedia\u6bb5\u843d\u4e2d\u8fdb\u884c\u641c\u7d22\u4ee5\u56de\u7b54\u590d\u6742\u95ee\u9898\u3002\u6211\u4eec\u9009\u62e9\u6b64\u7ba1\u9053\u6709\u4e24\u4e2a\u539f\u56e0\u3002\u9996\u5148\uff0cLangChain\u662f\u4e0eLLM\u5408\u4f5c\u4f7f\u7528\u6700\u53d7\u6b22\u8fce\u7684\u8f6f\u4ef6\u6846\u67b6\u4e4b\u4e00\uff0c\u63d0\u4f9b\u4e86\u7ecf\u8fc7\u201c\u63d0\u793a\u5de5\u7a0b\u201d\u4f18\u5316\u7684\u5f00\u6e90\u6a21\u5757\uff0c\u80fd\u591f\u5f88\u597d\u5730\u6267\u884c\u5404\u79cd\u4efb\u52a1\u3002\u56e0\u6b64\uff0c\u8fd9\u4e9b\u6a21\u5757\u63d0\u793a\u968f\u65f6\u95f4\u7684\u53d8\u5316\u7a33\u5b9a\u6027\u5bf9\u8bb8\u591a\u7528\u6237\u6765\u8bf4\u662f\u6709\u610f\u4e49\u7684\u5173\u6ce8\u70b9\u3002\u5176\u6b21\uff0cHotpotQA\u5e7f\u6cdb\u7528\u4e8e\u8861\u91cfLLM\u56de\u7b54\u591a\u8df3\u95ee\u9898\u7684\u80fd\u529b\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u4f7f\u7528\u4e86LangChain3\u4e2d\u7684\u9ed8\u8ba4ReAct\u4ee3\u7406\uff08\u65e8\u5728\u91cd\u73b0ReAct\u63d0\u793a[YZY+22]\uff09\uff0c\u5e76\u5c06\u5176\u4e0e\u4e0d\u540c\u7684LLMs\uff08GPT-4\u548cGPT-3.5\uff09\u7ed3\u5408\u4f5c\u4e3a\u4ee3\u7801\u7684\u9aa8\u5e72\u3002\u7136\u540e\u8ba9\u8be5\u4ee3\u7406\u56de\u7b54HotpotQA\u6570\u636e\u96c6\u4e2d\u6bcf\u4e2a\u67e5\u8be2\u7684\u95ee\u9898\u3002<br \/>\n![[How Is ChatGPTs Behavior Changing over Time-Figure10.png]]<\/p>\n<blockquote>\n<p>\u56fe10\uff1aLangChain HotpotQA Agent<br \/>\n(a) \u7cbe\u786e\u5339\u914d\u3001\u5197\u957f\u5ea6\u548c\u4e0d\u5339\u914d\u7387\u7684\u53d8\u5316\u3002\u603b\u4f53\u800c\u8a00\uff0cGPT-4 \u7684\u7cbe\u786e\u5339\u914d\u7387\u6709\u6240\u63d0\u9ad8\uff0c\u800c GPT-3.5 \u7684\u6027\u80fd\u5219\u4e0b\u964d\u4e86\u3002<br \/>\n(b) \u793a\u4f8b\u67e5\u8be2\u53ca\u5176\u5bf9\u5e94\u7b54\u6848\u3002LangChain \u65e0\u6cd5\u89e3\u6790\u4e09\u6708\u7684 GPT-4 \u54cd\u5e94\uff0c\u56e0\u4e3a\u5b83\u672a\u80fd\u9075\u5faa LangChain \u63d0\u793a\u4e2d\u6307\u5b9a\u7684\u683c\u5f0f\u3002\u516d\u6708\u7684 GPT-3.5 \u65e0\u6cd5\u627e\u5230\u5b83\u5728\u4e09\u6708\u80fd\u591f\u83b7\u53d6\u7684\u4fe1\u606f\u3002\u8fd9\u4e9b\u95ee\u9898\u7a81\u663e\u4e86\u5c06 LLM \u96c6\u6210\u5230\u66f4\u5927\u7ba1\u9053\u4e2d\u7684\u7a33\u5b9a\u6027\u95ee\u9898\u3002<\/p>\n<\/blockquote>\n<p>\u603b\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u89c2\u5bdf\u5230GPT-4\u548cGPT-3.5\u5728\u8fd9\u9879\u4efb\u52a1\u4e0a\u90fd\u6709\u663e\u8457\u7684\u6f02\u79fb\u3002\u4f8b\u5982\uff0c\u5982\u56fe10(a)\u6240\u793a\uff0cGPT-4\u5728\u4e09\u6708\u7684\u786e\u5207\u5339\u914d\u7387\u4e3a1.2%\uff0c\u4f46\u5728\u516d\u6708\u589e\u52a0\u5230\u4e8637.8%\u3002\u76f8\u53cd\uff0c\u5bf9\u4e8eGPT-3.5\uff0c\u5728\u4ece\u4e09\u6708\u81f3\u516d\u6708\u671f\u95f4\uff0c\u786e\u5207\u5339\u914d\u7387\u4e0b\u964d\u4e86\u8fd19%\u3002\u6b64\u5916\uff0c\u8d85\u8fc780%\u7684\u6700\u7ec8\u7b54\u6848\u5728\u8fd9\u4e24\u4e2a\u6708\u4e4b\u95f4\u5bf9\u4e24\u4e2a\u6a21\u578b\u90fd\u4e0d\u5339\u914d\u3002<br \/>\n\u6211\u4eec\u8fd8\u6ce8\u610f\u5230\uff0cGPT-4\u5728\u516d\u6708\u751f\u6210\u7684\u7b54\u6848\u6bd4\u4e09\u6708\u66f4\u52a0\u7b80\u6d01\uff0c\u800cGPT-3.5\u968f\u7740\u65f6\u95f4\u63a8\u79fb\u53d8\u5f97\u5197\u957f\u4e8630%\u3002<br \/>\n\u8fd9\u79cd\u60c5\u51b5\u4e3a\u4ec0\u4e48\u4f1a\u53d1\u751f\uff1f\u4ed4\u7ec6\u5206\u6790\u90a3\u4e9b\u4e0d\u5339\u914d\u7684\u56de\u7b54\u8868\u660e\uff0c\u63d0\u793a\u7684\u7a33\u5b9a\u6027\u8f83\u5dee\u662f\u5176\u4e2d\u4e00\u4e2a\u89e3\u91ca\u3002\u4e3a\u4e86\u8bf4\u660e\u8fd9\u4e00\u70b9\uff0c\u8bf7\u8003\u8651\u56fe10(b)\u4e2d\u7684\u4f8b\u5b50\u3002\u8be5\u67e5\u8be2\u8be2\u95ee\u4e24\u4eba\u7684\u653f\u6cbb\u503e\u5411\u662f\u5426\u4e3a\u6c11\u4e3b\u515a\u6216\u5171\u548c\u515a\u30023\u6708\u4efd\u7684GPT-4\u5b9e\u9645\u4e0a\u80fd\u591f\u627e\u5230\u6b63\u786e\u7684\u7b54\u6848\uff1a\u4ed6\u4eec\u90fd\u662f\u6c11\u4e3b\u515a\u6210\u5458\u3002\u7136\u800c\uff0cLangChain\u4ee3\u7406\u671f\u671b\u4e00\u4e2a\u7279\u5b9a\u7684\u683c\u5f0f\uff1a\u201c[\u52a8\u4f5c]+\u6587\u672c\u201d\uff0c\u8fd9\u4e2a\u683c\u5f0f\u88ab\u7f16\u7801\u5728\u5176\u63d0\u793a\u4e2d\u3002\u4e0d\u5e78\u7684\u662f\uff0c3\u6708\u4efd\u7684GPT-4\u672a\u80fd\u9075\u5faa\u6b64\u683c\u5f0f\uff0c\u56e0\u6b64LangChain\u4ee3\u7406\u7b80\u5355\u5730\u751f\u6210\u4e86\u4e00\u4e2a\u9519\u8bef\u6d88\u606f\u201c\u65e0\u6cd5\u89e3\u6790LLM\u8f93\u51fa\u201d\u3002\u8fd9\u5728\u5b9e\u9645\u7684LLM\u5e94\u7528\u4e2d\u662f\u4e00\u4e2a\u95ee\u9898\uff0c\u56e0\u4e3a\u5728\u5927\u89c4\u6a21\u7ba1\u9053\u4e2d\u7684\u624b\u52a8\u8c03\u8bd5\u8fd9\u7c7b\u95ee\u9898\u662f\u5177\u6709\u6311\u6218\u6027\u7684\u3002\u6b64\u5916\uff0c\u57283\u6708\u65f6\uff0cGPT-3.5\u627e\u5230\u4e86\u6b63\u786e\u7684\u7b54\u6848\uff1b\u7136\u800c\u5230\u4e866\u6708\uff0c\u5b83\u8868\u793a\u201c\u672a\u80fd\u627e\u5230\u76f8\u5173\u4fe1\u606f\u201d\u3002\u8fd9\u4e9b\u95ee\u9898\u8868\u660e\uff0c\u73b0\u6709\u7684\u63d0\u793a\u65b9\u6cd5\u548c\u5e93\u5728\u9762\u5bf9LLM\u6f02\u79fb\uff08drift\uff09\u8fdb\u884c\u590d\u6742\u4efb\u52a1\u5904\u7406\u65f6\u7684\u8106\u5f31\u6027\u3002<\/p>\n<h2>3.7 USMLE \u533b\u5b66\u8003\u8bd5\uff1aGPT-4 \u8868\u73b0\u7565\u6709\u4e0b\u964d<\/h2>\n<h2>3.8 \u89c6\u89c9\u63a8\u7406\uff1a\u4e24\u4e2a\u6a21\u578b\u5747\u6709\u5c0f\u5e45\u63d0\u5347<\/h2>\n<h1>4 GPT-4 \u7684\u6307\u4ee4\u9075\u5faa\u80fd\u529b\u968f\u65f6\u95f4\u53d8\u5dee\u4e86\u5417\uff1f<\/h1>\n<p>\u5982\u4f55\u89e3\u91ca\u89c2\u5bdf\u5230\u7684\u884c\u4e3a\u6f02\u79fb\uff1f\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\uff0cLLM\u6027\u80fd\u7684\u4e0b\u964d\u901a\u5e38\u4e0e\u8f83\u5dee\u7684\u6307\u4ee4\u9075\u5faa\u80fd\u529b\uff08\u5373\u66f4\u5dee\u5730\u6267\u884c\u7528\u6237\u6307\u4ee4\u7684\u80fd\u529b\uff09\u76f8\u5173\u3002\u4f8b\u5982\uff0c\u5728Math I\u548cMath II\u4efb\u52a1\u4e0a\uff0cGPT-4\u57283\u6708\u6309\u7167\u7528\u6237\u7684\u6307\u4ee4\u8fdb\u884c\u9010\u6b65\u63a8\u7406\u5e76\u56de\u7b54\u95ee\u9898\uff0c\u4f46\u57286\u6708\u62d2\u7edd\u8fd9\u6837\u505a\u3002OpinionQA\u63d0\u4f9b\u4e86\u53e6\u4e00\u4e2a\u4f8b\u5b50\uff1aGPT-4\u57283\u6708\u4efd\u56de\u5e94\u4e86\u7528\u6237\u7684\u95ee\u9898\uff0c\u4f46\u5230\u4e866\u6708\u5219\u6ca1\u6709\u505a\u51fa\u56de\u5e94\u3002<br \/>\n<strong>\u91cf\u5316\u5355\u6307\u4ee4\u6267\u884c\u504f\u5dee<\/strong>  \u5728\u73b0\u6709\u7684\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u91cf\u5316\u6307\u4ee4\u9075\u5faa\u7684\u6f02\u79fb\u5177\u6709\u6311\u6218\u6027\uff1a\u5b83\u4eec\u7684\u4efb\u52a1\u548c\u8bc4\u4f30\u6307\u6807\u5e38\u5e38\u6a21\u7cca\u4e86\u6a21\u578b\u7684\u6307\u4ee4\u5fe0\u5b9e\u5ea6\u4e0e\u5176\u7279\u5b9a\u4efb\u52a1\u80fd\u529b\uff08\u5982\u5199\u4f5c\u548c\u903b\u8f91\u63a8\u7406\uff09\u4ee5\u53ca\u77e5\u8bc6\uff08\u5e38\u8bc6\u3001\u5386\u53f2\u7b49\uff09\u4e4b\u95f4\u7684\u754c\u9650\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u7b56\u5212\u4e86\u4e00\u4e2a\u65b0\u7684\u4e13\u6ce8\u4e8e\u4efb\u52a1\u65e0\u5173\u6307\u4ee4\u7684\u57fa\u51c6\u6d4b\u8bd5\u96c6\u3002\u8fd9\u5305\u62ec\u5b9e\u8df5\u4e2d\u5e38\u89c1\u7684\u56db\u79cd\u7c7b\u578b\u7684\u6307\u4ee4\uff1a\u7b54\u6848\u63d0\u53d6\uff08\u201c\u5728\u65b9\u62ec\u53f7\u5185\u56de\u7b54\u662f\u6216\u5426\u201d\uff09\u3001\u505c\u6b62\u9053\u6b49\uff08\u201c\u4e0d\u8981\u8bf4\u5bf9\u4e0d\u8d77\u6216\u4f5c\u4e3aAI\u6a21\u578b\u201d\uff09\u3001\u5199\u4f5c\u7ea6\u675f\uff08\u201c\u7528\u4ee5Y\u5f00\u5934\/\u7ed3\u5c3e\u7684\u5355\u8bcd\u63cf\u8ff0X\u201d\uff09\u548c\u6587\u672c\u683c\u5f0f\u5316\uff08\u201c\u5c06\u6bcf\u4e2a\u5355\u8bcd\u7684\u7b2c\u4e00\u4e2a\u5b57\u6bcd\u52a0\u4e0a\u65b9\u62ec\u53f7\uff08\u5305\u62ec\u2018the\u2019\u8fd9\u6837\u7684\u51a0\u8bcd\uff09\u201d\uff09\u3002\u6211\u4eec\u5bf9200\u7bc7\u8fd1\u671farxiv\u8bba\u6587\u7684\u6458\u8981\u8fdb\u884c\u4e86\u7b54\u6848\u63d0\u53d6\u548c\u6587\u672c\u683c\u5f0f\u5316\u7684\u5e94\u7528\uff0c\u5e76\u5728sensitiveQA\u6570\u636e\u96c6\u4e0a\u6267\u884c\u4e86\u5185\u5bb9\u8fc7\u6ee4\u3002\u6b64\u5916\uff0c\u6211\u4eec\u624b\u52a8\u521b\u5efa\u4e8620\u4e2a\u98ce\u683c\u4f18\u5316\u67e5\u8be2\u3002<br \/>\n\u5982\u56fe13\u6240\u793a\uff0cGPT-4\u4ece\u4e09\u6708\u5230\u516d\u6708\u7684\u786e\u7ecf\u5386\u4e86\u8f83\u5927\u7684\u6307\u4ee4\u5fe0\u5b9e\u5ea6\u4e0b\u964d\u3002\u4f8b\u5982\uff0c\u5728\u4e09\u6708\u4efd\uff0cGPT-4\u9075\u5faa\u4e8699.5%\u7684\u7b54\u6848\u63d0\u53d6\u67e5\u8be2\uff0c\u800c\u5230\u4e86\u516d\u6708\uff0c\u8fd9\u4e00\u6570\u5b57\u964d\u81f30.5%\u3002\u572874%\u7684\u654f\u611f\u95ee\u9898\u4e0a\uff0cGPT-4\u5728\u4e09\u6708\u4efd\u786e\u5b9e\u6ca1\u6709\u50cf\u6307\u4ee4\u8981\u6c42\u90a3\u6837\u63d0\u5230\u201c\u5bf9\u4e0d\u8d77\u201d\u6216\u201c\u4f5c\u4e3aAI\u6a21\u578b\u201d\u3002\u7136\u800c\uff0c\u5230\u516d\u6708\u4efd\uff0c\u8fd9\u4e2a\u6bd4\u4f8b\u4ec5\u526919%\u3002\u56fe13\u4e2d\u7ed9\u51fa\u7684\u4f8b\u5b50\u8fdb\u4e00\u6b65\u63ed\u793a\u4e86\u5bfc\u81f4\u516d\u6708\u7248\u672c\u4f4e\u5fe0\u5b9e\u5ea6\u7684\u539f\u56e0\u3002\u4f8b\u5982\uff0c\u5728\u516d\u6708\uff0c\u5c3d\u7ba1GPT-4\u5c06\u7b54\u6848\u653e\u5728\u65b9\u62ec\u53f7\u5185\uff0c\u4f46\u5b83\u59cb\u7ec8\u5927\u5199\u7b2c\u4e00\u4e2a\u5b57\u6bcd\u3002\u540c\u6837\u5730\uff0c\u867d\u7136\u7528\u6237\u8981\u6c42\u4e0d\u8981\u8bf4\u201c\u5bf9\u4e0d\u8d77\u201d\uff0c\u4f46GPT-4\u5728\u516d\u6708\u4efd\u4ecd\u7136\u751f\u6210\u4e86\u201c\u5bf9\u4e0d\u8d77\u201d\uff0c\u800c\u5176\u4e09\u6708\u4efd\u7684\u7248\u672c\u5219\u91cd\u65b0\u63aa\u8f9e\u4ee5\u9075\u5faa\u7528\u6237\u7684\u8bf7\u6c42\u3002\u5bf9\u4e8e\u5199\u4f5c\u7ea6\u675f\u7684\u4f8b\u5b50\uff0c\u5728\u4e09\u6708\u65f6\uff0cGPT-4\u4e25\u683c\u9075\u5b88\u4e86\u7528\u6237\u7684\u6307\u4ee4\uff1a\u5b83\u751f\u6210\u4e86\u4e0e\u673a\u5668\u5b66\u4e60\u76f8\u5173\u7684\u4e14\u4ee5\u201cn\u201d\u7ed3\u5c3e\u7684\u5355\u8bcd\u3002\u7136\u800c\u5230\u4e86\u516d\u6708\u7248\u672c\uff0c\u5219\u66f4\u5173\u6ce8\u4e8e\u201c\u673a\u5668\u5b66\u4e60\u201d\u800c\u5ffd\u7565\u4e86\u201c\u4ee5\u2018n\u2019\u7ed3\u5c3e\u201d\u7684\u8981\u6c42\u3002\u6b64\u5916\uff0cGPT-4\u5728\u4e09\u6708\u4efd\u6210\u529f\u5730\u5bf9\u6bcf\u4e2a\u5355\u8bcd\u7684\u7b2c\u4e00\u4e2a\u5b57\u6bcd\u8fdb\u884c\u4e86\u5927\u5199\u5904\u7406\uff0c\u4f46\u5728\u516d\u6708\u4efd\u5374\u9057\u6f0f\u4e86\u4e00\u4e9b\u8bcd\uff08\u5982\u793a\u4f8b\u4e2d\u7684\u201cprovides\u201d\u548c\u201cabout\u201d\uff09\u3002\u603b\u4f53\u800c\u8a00\uff0c\u4ece\u4e09\u6708\u5230\u516d\u6708\uff0cGPT-4\u7684\u6307\u4ee4\u9075\u5faa\u5fe0\u5b9e\u5ea6\u663e\u8457\u4e0b\u964d\uff0c\u8fd9\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u89e3\u91ca\u4e86\u5176\u884c\u4e3a\u6f02\u79fb\u3002<br \/>\n<strong>\u590d\u5408\u6307\u4ee4\u7684\u6267\u884c\u6f02\u79fb<\/strong> \u6211\u4eec\u8fdb\u4e00\u6b65\u7814\u7a76\u4e86GPT-4\u5728\u7ec4\u5408\u6307\u4ee4\u4e0b\u7684\u6307\u4ee4\u6267\u884c\u53d8\u5316\u60c5\u51b5\u3002\u4e3a\u4e86\u91cf\u5316\u8fd9\u4e00\u53d8\u5316\uff0c\u6211\u4eec\u6536\u96c6\u4e86\u4e00\u7ec4\u5355\u6307\u4ee4\uff0c\u5e76\u521b\u5efa\u4e86\u4e00\u4e2a\u590d\u5408\u6307\u4ee4\u5217\u8868\uff0c\u6bcf\u4e2a\u590d\u5408\u6307\u4ee4\u5bf9\u5e94\u4e8e\u5355\u6307\u4ee4\u96c6\u4e2d\u7684\u4e24\u4e2a\u6307\u4ee4\u3002\u6211\u4eec\u5728arxiv\u8bba\u6587\u7684\u7b2c\u4e00\u53e5\u8bdd\u4e0a\u8bc4\u4f30\u4e86GPT-4\u5728\u8fd9\u7c7b\u590d\u5408\u6307\u4ee4\u4e0a\u7684\u8868\u73b0\u3002\u5355\u6307\u4ee4\u96c6\u4e2d\u5305\u542b\u4e09\u4e2a\u6587\u672c\u683c\u5f0f\u5316\u6307\u4ee4\uff1a\u6dfb\u52a0\u9017\u53f7\uff08\u201c\u5728\u6bcf\u4e2a\u5355\u8bcd\u540e\u52a0\u4e00\u4e2a\u9017\u53f7\u201d\uff09\u3001\u5220\u9664\u5f15\u53f7\uff08\u201c\u79fb\u9664\u6240\u6709\u5f15\u53f7\u201d\uff09\u548c\u5927\u5199\u5904\u7406\uff08\u201c\u5c06\u6bcf\u4e2a\u5b57\u6bcd\u5927\u5199\u201d\uff09\u3002\u8fd9\u4e9b\u6307\u4ee4\u6613\u4e8e\u4eba\u7c7b\u7406\u89e3\uff0c\u5e76\u4e14\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u4e5f\u5f88\u5e38\u89c1\u3002<br \/>\n\u6709\u51e0\u4e2a\u6709\u8da3\u7684\u89c2\u5bdf\u7ed3\u679c\u3002\u9996\u5148\uff0cGPT-4\u5728\u4e09\u6708\u548c\u516d\u6708\u90fd\u5f88\u597d\u5730\u9075\u5faa\u4e86\u5355\u6307\u4ee4\u3002\u4e8b\u5b9e\u4e0a\uff0c\u4ece\u4e09\u6708\u81f3\u516d\u6708\uff0c\u5355\u4e00\u6307\u4ee4\u7684\u6267\u884c\u53d8\u5316\u4ec5\u4e3a -2%\u3001+4.0% \u548c -1.0%\uff08\u89c1\u56fe14(a)\uff09\u3002\u5176\u6b21\uff0c\u4e0e\u4e09\u6708\u4efd\u76f8\u6bd4\uff0c\u516d\u6708\u4efd\u7684GPT-4\u5bf9\u590d\u5408\u6307\u4ee4\u66f4\u4e3a\u654f\u611f\u3002\u4f8b\u5982\uff0c\u5728\u88ab\u8981\u6c42\u79fb\u9664\u5f15\u53f7\u4ee5\u53ca\u5728\u6bcf\u4e2a\u5355\u8bcd\u540e\u52a0\u4e00\u4e2a\u9017\u53f7\u65f6\uff0cGPT-4\u7684\u8868\u73b0\u4ece\u4e09\u6708\u5230\u516d\u6708\u4e0b\u964d\u4e8624%\u3002\u540c\u6837\u5730\uff0c\u4ece\u4e09\u6708\u5230\u516d\u6708\uff0c\u6dfb\u52a0\u9017\u53f7\u548c\u5927\u5199\u5b57\u6bcd\u8fd9\u4e00\u7ec4\u5408\u6307\u4ee4\u7684\u51c6\u786e\u6027\u4e0b\u964d\u4e869.2%\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u590d\u5408\u6307\u4ee4\u89e6\u53d1\u7684\u9519\u8bef\u6a21\u5f0f\u5f88\u6709\u610f\u601d\u3002\u5982\u56fe14\u4e2d\u7684\u793a\u4f8b\u6240\u793a\uff0c\u5728\u7ed9\u5b9a\u590d\u5408\u6307\u4ee4\u65f6\uff0c\u516d\u6708\u4efd\u7684GPT-4\u503e\u5411\u4e8e\u5728\u6bcf\u4e2a\u5b57\u7b26\u540e\u52a0\u4e00\u4e2a\u9017\u53f7\u3002\u800c\u4e09\u6708\u4efd\u7684\u7248\u672c\u5219\u5fe0\u5b9e\u5b8c\u6210\u4e86\u7528\u6237\u4efb\u52a1\u3002<br \/>\n\u603b\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u89c2\u5bdf\u5230GPT-4\u968f\u65f6\u95f4\u9010\u6e10\u51cf\u5c11\u4e86\u5bf9\u7528\u6237\u6307\u4ee4\u7684\u9075\u5faa\u3002\u8fd9\u4e0d\u4ec5\u9002\u7528\u4e8e\u5355\u6307\u4ee4\uff0c\u4e5f\u9002\u7528\u4e8e\u590d\u5408\u6307\u4ee4\u3002\u4e0e\u524d\u4e00\u8282\u5206\u6790\u7684\u6027\u80fd\u53d8\u5316\u4e00\u81f4\uff0c\u6307\u4ee4\u6267\u884c\u7684\u53d8\u5316\u4f3c\u4e4e\u662f\u5bfc\u81f4GPT-4\u884c\u4e3a\u6f02\u79fb\u7684\u4e3b\u8981\u56e0\u7d20\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5728\u4e0d\u540c\u65f6\u95f4\u6bb5\u5185\uff0cGPT-3.5\u7684\u6307\u4ee4\u9075\u5faa\u60c5\u51b5\u6ca1\u6709\u51fa\u73b0\u4e00\u81f4\u6027\u7684\u53d8\u5316\uff08\u89c1\u9644\u5f55\u4e2d\u7684\u56fe16\uff09\u3002<\/p>\n<h1>5 Conclusions and Future Work<\/h1>\n<p>\u6211\u4eec\u7684\u7814\u7a76\u7ed3\u679c\u8868\u660e\uff0cGPT-3.5\u548cGPT-4\u7684\u884c\u4e3a\u5728\u76f8\u5bf9\u8f83\u77ed\u7684\u65f6\u95f4\u5185\u53d1\u751f\u4e86\u663e\u8457\u53d8\u5316\u3002\u8fd9\u7a81\u663e\u4e86\u6301\u7eed\u8bc4\u4f30\u548c\u76d1\u6d4b\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u884c\u4e3a\u6f02\u79fb\u7684\u5fc5\u8981\u6027\uff0c\u5c24\u5176\u662f\u8003\u8651\u5230\u50cfChatGPT\u8fd9\u6837\u7684LLM\u968f\u7740\u65f6\u95f4\u63a8\u79fb\u5982\u4f55\u66f4\u65b0\u5e76\u4e0d\u900f\u660e\u3002\u6211\u4eec\u7684\u7814\u7a76\u8fd8\u5f3a\u8c03\u4e86\u7edf\u4e00\u63d0\u9ad8LLM\u591a\u65b9\u9762\u80fd\u529b\u7684\u6311\u6218\u3002\u4f8b\u5982\uff0c\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\u901a\u8fc7\u5728\u66f4\u591a\u6570\u636e\u4e0a\u7684\u5fae\u8c03\u6765\u63d0\u5347\u6a21\u578b\u6027\u80fd\u53ef\u80fd\u4f1a\u5bf9\u5176\u5728\u5176\u4ed6\u4efb\u52a1\u4e2d\u7684\u884c\u4e3a\u4ea7\u751f\u610f\u5916\u7684\u5f71\u54cd\u3002\u6b63\u5982\u6211\u4eec\u6240\u89c2\u5bdf\u5230\u7684\uff0c\u65e0\u8bba\u662fGPT-3.5\u8fd8\u662fGPT-4\uff0c\u5728\u4e00\u4e9b\u4efb\u52a1\u4e0a\u8868\u73b0\u53d8\u5dee\u7684\u540c\u65f6\uff0c\u5728\u5176\u4ed6\u7ef4\u5ea6\u4e0a\u5219\u6709\u6240\u6539\u8fdb\u3002\u6b64\u5916\uff0cGPT-3.5\u548cGPT-4\u7684\u53d1\u5c55\u8d8b\u52bf\u5f80\u5f80\u5b58\u5728\u5206\u6b67\u3002\u9664\u4e86\u6700\u7ec8\u6027\u80fd\u4e4b\u5916\uff0c\u89c2\u5bdf\u6a21\u578b\u7684\u94fe\u5f0f\u601d\u7ef4\u884c\u4e3a\u53d8\u5316\u548c\u5197\u957f\u6027\uff08verbosity\uff09\u4e5f\u9887\u5177\u7814\u7a76\u4ef7\u503c\u3002<\/p>\n<p>\u6211\u4eec\u8ba1\u5212\u901a\u8fc7\u5b9a\u671f\u8bc4\u4f30GPT-3.5\u3001GPT-4\u53ca\u5176\u4ed6LLM\u5728\u5404\u79cd\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u6765\u6301\u7eed\u66f4\u65b0\u8fd9\u91cc\u5c55\u793a\u7684\u7814\u7a76\u6210\u679c\u3002\u5bf9\u4e8e\u4f9d\u8d56LLM\u670d\u52a1\u4f5c\u4e3a\u5176\u65e5\u5e38\u5de5\u4f5c\u6d41\u7a0b\u4e00\u90e8\u5206\u7684\u7528\u6237\u6216\u516c\u53f8\uff0c\u5efa\u8bae\u4ed6\u4eec\u91c7\u7528\u4e0e\u6211\u4eec\u7c7b\u4f3c\u7684\u65b9\u6cd5\u8fdb\u884c\u76d1\u6d4b\u5206\u6790\u3002\u611f\u8c22\u6240\u6709\u4e3a\u6211\u4eec\u7814\u7a76\u63d0\u4f9b\u6709\u76ca\u53cd\u9988\u7684\u4eba\u58eb\u3002<\/p>\n<p>\u4e3a\u4e86\u9f13\u52b1\u8fdb\u4e00\u6b65\u7814\u7a76LLM\u6f02\u79fb\u73b0\u8c61\uff0c\u6211\u4eec\u5c06\u53d1\u5e03\u6211\u4eec\u7684\u8bc4\u4f30\u6570\u636e\u548cChatGPT\u54cd\u5e94\uff0c\u8bf7\u8bbf\u95ee<a href=\"https:\/\/github.com\/lchen001\/LLMDrift\">LLMDrift<\/a>\u83b7\u53d6\u8be6\u7ec6\u4fe1\u606f\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Abstract GPT-3.5\u548cGPT-4\u662f\u4e24\u79cd\u6700\u5e7f\u6cdb\u4f7f\u7528\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u670d\u52a1\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u6a21\u578b\u4f55\u65f6\u4ee5 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[105],"tags":[116],"_links":{"self":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/988"}],"collection":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=988"}],"version-history":[{"count":1,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/988\/revisions"}],"predecessor-version":[{"id":989,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=\/wp\/v2\/posts\/988\/revisions\/989"}],"wp:attachment":[{"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=988"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=988"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/tamanegi.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=988"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}