{"id":36024,"date":"2025-01-04T17:00:20","date_gmt":"2025-01-04T09:00:20","guid":{"rendered":"https:\/\/17aitech.com\/?p=36024"},"modified":"2025-01-04T17:00:20","modified_gmt":"2025-01-04T09:00:20","slug":"%e4%b8%80%e6%96%87%e7%9c%8b%e6%87%82llm%e6%8e%a8%e7%90%86%ef%bc%8cucl%e6%b1%aa%e5%86%9b%e6%95%99%e6%8e%88%e8%a7%a3%e8%af%bbopenai-%ce%bf1%e7%9a%84%e7%9b%b8%e5%85%b3%e6%96%b9%e6%b3%95","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=36024","title":{"rendered":"\u4e00\u6587\u770b\u61c2LLM\u63a8\u7406\uff0cUCL\u6c6a\u519b\u6559\u6388\u89e3\u8bfbOpenAI \u03bf1\u7684\u76f8\u5173\u65b9\u6cd5"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-10-11-4\" target=\"_blank\">\u4e00\u6587\u770b\u61c2LLM\u63a8\u7406\uff0cUCL\u6c6a\u519b\u6559\u6388\u89e3\u8bfbOpenAI \u03bf1\u7684\u76f8\u5173\u65b9\u6cd5<\/a><\/p>\n<p>OpenAI \u6700\u8fd1\u53d1\u5e03\u7684 o1 \u7cfb\u5217\u6a21\u578b\u582a\u79f0\u8fc8\u5411\u5f3a\u4eba\u5de5\u667a\u80fd\u7684\u4e00\u6b21\u98de\u8dc3\uff0c\u5176\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\u4e3a\u6211\u4eec\u63cf\u7ed8\u51fa\u4e86\u4e0b\u4e00\u4ee3\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u7684\u672a\u6765\u56fe\u666f\u3002\u8fd1\u65e5\uff0c\u4f26\u6566\u5927\u5b66\u5b66\u9662\uff08UCL\uff09\u4eba\u5de5\u667a\u80fd\u4e2d\u5fc3\u6c6a\u519b\u6559\u6388\u64b0\u5199\u4e86\u4e00\u4efd\u300cLLM \u63a8\u7406\u6559\u7a0b\u300d\uff0c\u6df1\u5165\u8be6\u7ec6\u5730\u4ecb\u7ecd\u4e86 OpenAI \u03bf1 \u6a21\u578b\u80cc\u540e\u7684\u76f8\u5173\u65b9\u6cd5\u3002<\/p>\n<p><strong>\u4ed6\u5c06\u5728 10 \u6708 12 \u53f7\u672c\u5468\u661f\u671f\u516d\u65e9\u4e0a\u4e8e\u9999\u6e2f\u79d1\u6280\u5927\u5b66\uff08\u5e7f\u5dde\uff09RLChina 2024 \u5927\u4f1a(http:\/\/rlchina.org\/rlchina_2024\/)\u4e0a\u4f5c\u76f8\u5173\u5185\u5bb9\u7684\u4e3b\u9898\u62a5\u544a\uff0c\u5e76\u53d1\u5e03\u5176\u56e2\u961f\u5f00\u53d1\u7684 LLM \u63a8\u7406\u5f00\u6e90\u6846\u67b6\u4ee5\u63a8\u52a8 o1 \u76f8\u5173\u6a21\u578b\u7684\u53d1\u5c55\u3002<\/strong><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ca83ca16a797821b782c07f24651568a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ca83ca16a797821b782c07f24651568a.png\"><\/a><\/p>\n<section>\u94fe\u63a5\uff1ahttps:\/\/github.com\/openreasoner\/openr\/blob\/main\/reports\/Tutorial-LLM-Reasoning-Wang.pdf<\/section>\n<section><\/section>\n<section>o1 \u7684\u8bad\u7ec3\u4f7f\u7528\u4e86\u5f3a\u5316\u5b66\u4e60\u6280\u672f\uff0c\u901a\u8fc7\u663e\u5f0f\u5730\u5d4c\u5165\u4e00\u4e2a\u539f\u751f\u300c\u601d\u7ef4\u94fe\u300d\uff08NCoT\uff09\u8fc7\u7a0b\uff0c\u53ef\u51fa\u8272\u5730\u5b8c\u6210\u590d\u6742\u7684\u63a8\u7406\u4efb\u52a1\u3002\u4e5f\u5c31\u662f\u8bf4\uff0co1 \u5728\u751f\u6210\u54cd\u5e94\u4e4b\u524d\u53ef\u901a\u8fc7\u4e00\u6b65\u6b65\u5730\u63a8\u7406\u5b9e\u73b0\u300c\u6df1\u5ea6\u601d\u8003\u300d\u3002<\/section>\n<section><\/section>\n<section>\u4ece OpenAI \u53d1\u5e03\u7684\u6570\u636e\u770b\uff0c\u76f8\u6bd4\u4e8e\u4e4b\u524d\u7684 ChatGPT 4o\uff0co1 \u5728\u6570\u5b66\u548c\u7f16\u7a0b\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u8981\u5f3a 5 \u500d\u3002\u5b83\u5728\u7ade\u4e89\u6027\u7f16\u7a0b\u4e2d\u6392\u540d\u7b2c 89 \u4f4d\uff0c\u5728\u7f8e\u56fd\u8457\u540d\u7684\u6570\u5b66\u5965\u6797\u5339\u514b\u8d44\u683c\u8d5b\u4e2d\u540d\u5217\u524d 500 \u540d\uff0c\u5e76\u5728\u7269\u7406\u3001\u751f\u7269\u548c\u5316\u5b66\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8d85\u8d8a\u4e86\u4eba\u7c7b\u535a\u58eb\u7ea7\u7684\u51c6\u786e\u5ea6\u3002<\/section>\n<section><\/section>\n<section>o1 \u7684\u4e00\u4e2a\u5173\u952e\u521b\u65b0\u662f\u5b83\u5141\u8bb8\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u82b1\u8d39\u66f4\u591a\u65f6\u95f4\u8fdb\u884c\u63a8\u7406\uff0c\u8fd9\u6807\u5fd7\u7740\u4e00\u79cd\u8303\u5f0f\u8f6c\u53d8\uff1a\u4ece\u5feb\u901f\u3001\u76f4\u63a5\u7684\u53cd\u5e94\u8f6c\u5411\u7f13\u6162\u3001\u6df1\u601d\u719f\u8651\u3001\u591a\u6b65\u9aa4\u7684\u63a8\u7406\u65f6\u95f4\u8ba1\u7b97\u3002\u89c1\u56fe 1\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-9af68d41204a1e015cc7ba45283bd90c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-9af68d41204a1e015cc7ba45283bd90c.png\"><\/a><\/p>\n<p><em><sup>\u56fe 1\uff1a\u63a8\u7406\u65f6\u95f4\u8ba1\u7b97\u3002(a) \u81ea\u56de\u5f52 LLM \u662f\u76f4\u63a5\u57fa\u4e8e\u7ed9\u5b9a\u95ee\u9898\u751f\u6210\u7b54\u6848\u3002(b) \u601d\u7ef4\u94fe\u548c\u9010\u6b65\u601d\u8003\u7684\u6982\u5ff5\u5219\u6d89\u53ca\u5230\u5728\u5f97\u5230\u6700\u7ec8\u7b54\u6848\u4e4b\u524d\uff0c\u6574\u5408\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u3002\u8fd9\u4e9b\u91cd\u590d\u6b65\u9aa4\u64cd\u4f5c\u5141\u8bb8 1) \u4e0d\u65ad\u91cd\u590d\u8bbf\u95ee\u4e4b\u524d\u7684\u8f93\u51fa\uff0c2) \u9010\u6b65\u63a8\u8fdb\u5230\u540e\u7eed\u63a8\u7406\u9636\u6bb5\uff0c3) \u63a2\u7d22\u591a\u4e2a\u63a8\u7406\u8def\u5f84\u6216\u8f68\u8ff9\u3002<\/sup><\/em><\/p>\n<section>\u6709\u8da3\u7684\u662f\uff0c\u5728\u6307\u5bfc\u4eba\u7c7b\u51b3\u7b56\u548c\u884c\u4e3a\u65b9\u9762\uff0c\u4eba\u7c7b\u8ba4\u77e5\u4e2d\u5b58\u5728\u4e24\u79cd\u76f8\u5173\u4f46\u4e0d\u540c\u7684\u8ba4\u77e5\u5904\u7406\u6a21\u5f0f\uff0c\u5176\u4e2d\u6bcf\u79cd\u90fd\u6709\u5404\u81ea\u4e0d\u540c\u7684\u5927\u8111\u56de\u8def\u548c\u795e\u7ecf\u901a\u8def\uff0c\u89c1\u56fe 2\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0344444818d0e7fb0be47a57e8f8f739.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0344444818d0e7fb0be47a57e8f8f739.png\"><\/a><\/p>\n<p><em><sup>\u56fe 2\uff1a\u4eba\u7c7b\u8ba4\u77e5\u548c LLM \u7684\u7c7b\u6bd4\u3002(a) \u548c (b) \u4eba\u7c7b\u6709\u610f\u8bc6\u6216\u65e0\u610f\u8bc6\u63a7\u5236\u7684\u884c\u4e3a\u4f9d\u8d56\u4e8e\u90e8\u5206\u4e0d\u540c\u7684\u5927\u8111\u56de\u8def\u3002(a) \u4eba\u7c7b\u7684\u65e0\u610f\u8bc6\u63a7\u5236\u7531\u4e00\u4e9b\u4e13\u95e8\u7684\u5927\u8111\u533a\u57df\u7ef4\u6301\uff0c\u4f8b\u5982\u524d\u8111\u5c9b\u548c\u524d\u8865\u5145\u8fd0\u52a8\u533a\uff08pre-SMA\uff09\u3002(b) \u800c\u81ea\u4e3b\u63a7\u5236\u5219\u6d89\u53ca\u66f4\u5927\u7684\u7f51\u7edc\uff0c\u6fc0\u6d3b\u9876\u53f6\u548c\u524d\u989d\u53f6\u5185\u7684\u8bb8\u591a\u533a\u57df\u3002\u65e0\u610f\u8bc6\u63a7\u5236\u901a\u5e38\u5feb\u901f\u800c\u672c\u80fd\uff0c\u901a\u5e38\u7531\u81ea\u52a8\u8fc7\u7a0b\u9a71\u52a8\uff0c\u800c\u6709\u610f\u8bc6\u63a7\u5236\u5f80\u5f80\u6d89\u53ca\u66f4\u5ba1\u614e\u3001\u8ba1\u7b97\u548c\u6df1\u5165\u7684\u601d\u8003\uff0c\u9700\u8981\u4ed4\u7ec6\u7684\u53cd\u601d\u548c\u900f\u5f7b\u7684\u5206\u6790\u3002<\/sup><\/em><\/p>\n<section><strong>\u7cfb\u7edf 1 \u601d\u7ef4<\/strong>\u5feb\u901f\u3001\u81ea\u52a8\u4e14\u76f4\u89c2\uff0c\u6beb\u4e0d\u8d39\u529b\u4e14\u901a\u5e38\u662f\u65e0\u610f\u8bc6\u7684\u3002\u5b83\u4f9d\u8d56\u4e8e\u80fd\u591f\u5feb\u901f\u5904\u7406\u7684\u795e\u7ecf\u901a\u8def\uff0c\u5c24\u5176\u662f\u5728\u9700\u8981\u5feb\u901f\u53cd\u5e94\u6216\u8ba4\u77e5\u8d44\u6e90\u53d7\u9650\u7684\u60c5\u51b5\u4e0b\u3002<\/section>\n<section><\/section>\n<section><strong>\u7cfb\u7edf 2 \u601d\u7ef4<\/strong>\u662f\u6df1\u601d\u719f\u8651\u7684\u3001\u8d39\u529b\u7684\u548c\u6709\u610f\u8bc6\u7684\uff0c\u6d89\u53ca\u96c6\u4e2d\u6ce8\u610f\u529b\u548c\u5206\u6790\u5f0f\u63a8\u7406\u3002\u5b83\u5904\u7406\u4fe1\u606f\u7684\u901f\u5ea6\u66f4\u6162\uff0c\u5e38\u7528\u4e8e\u590d\u6742\u7684\u95ee\u9898\u6c42\u89e3\u3001\u903b\u8f91\u63a8\u7406\u548c\u51b3\u7b56\u4efb\u52a1\u3002<\/section>\n<section><\/section>\n<section>o1 \u7684\u8bde\u751f\u975e\u5e38\u6fc0\u52a8\u4eba\u5fc3\uff0c\u56e0\u4e3a \u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09 \u73b0\u5728\u4e0d\u4ec5\u80fd\u4f7f\u7528\u5b66\u4e60\u5230\u7684\u6a21\u5f0f\u8fdb\u884c\u5feb\u901f\u54cd\u5e94\uff0c\u800c\u4e14\u8fd8\u80fd\u901a\u8fc7\u601d\u7ef4\u94fe\u6216\u5176\u5b83\u5f62\u5f0f\u7684\u641c\u7d22\u7b49\u673a\u5236\u6a21\u62df\u590d\u6742\u7684\u63a8\u7406\u8fc7\u7a0b\u3002\u8fd9\u5c31\u7c7b\u4f3c\u4e8e\u4eba\u7c7b\u7684\u66f4\u6df1\u5ea6\u7684\u3001\u6b65\u6b65\u6267\u884c\u7684\u601d\u8003\u65b9\u5f0f\u3002<\/section>\n<section><\/section>\n<section>\u4e5f\u6b63\u56e0\u5982\u6b64\uff0cOpenAI \u03bf1 \u5728\u79d1\u5b66\u3001\u7f16\u7a0b\u548c\u6570\u5b66\u9886\u57df\u90fd\u53d6\u5f97\u4e86\u76f8\u5f53\u5353\u8d8a\u7684\u8868\u73b0\u3002\u6b64\u5916\uff0co1 \u5728 AI \u5b89\u5168\u548c\u5bf9\u9f50\u65b9\u9762\u4e5f\u53d6\u5f97\u4e86\u8fdb\u5c55\u3002\u8be5\u6a21\u578b\u7684\u601d\u7ef4\u94fe\u63a8\u7406\u4e3a\u6574\u5408\u4eba\u7c7b\u4ef7\u503c\u89c2\u548c\u539f\u5219\u63d0\u4f9b\u4e86\u65b0\u7684\u673a\u4f1a\uff0c\u4ece\u800c\u53ef\u63d0\u9ad8\u5b89\u5168\u8bc4\u4f30\u548c\u8d8a\u72f1\u6d4b\u8bd5\u7684\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section>\u4e8b\u5b9e\u4e0a\uff0c\u5728 LLM \u9886\u57df\uff0c\u601d\u7ef4\u94fe\u63a8\u7406\u548c\u5206\u6b65\u9aa4\u601d\u8003\u65b9\u6cd5\u5e76\u975e\u65b0\u6280\u672f\u3002\u4e4b\u524d\u5df2\u6709\u7814\u7a76\u8868\u660e\uff0c\u5982\u679c\u5728\u8f93\u5165\u4e2d\u6dfb\u52a0\u300cdescribe your reasoning in steps\u300d\u6216\u300cexplain your answer step by step\u300d\u8fd9\u6837\u7684\u6307\u4ee4\u6216\u63d0\u4f9b\u5c11\u6837\u672c\u793a\u4f8b\uff0c\u5c31\u53ef\u4ee5\u8ba9 LLM \u751f\u6210\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\uff0c\u8fdb\u800c\u63d0\u5347\u5176\u89e3\u51b3\u95ee\u9898\u7684\u80fd\u529b\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u6570\u5b66\u548c\u7f16\u7a0b\u4efb\u52a1\u3002<\/section>\n<section><\/section>\n<section>\u4f46\u662f\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u90fd\u57fa\u4e8e\u5df2\u6709\u7684 LLM\uff0c\u5e76\u6ca1\u6709\u5c06\u601d\u7ef4\u94fe\u5d4c\u5165\u5230\u6a21\u578b\u672c\u8eab\u4e4b\u4e2d\u3002\u56e0\u6b64\uff0cLLM \u65e0\u6cd5\u5185\u5316\u8fd9\u79cd\u5b66\u4e60\u80fd\u529b\uff0c\u5bfc\u81f4\u8bb8\u591a\u7814\u7a76\u8005\u5728\u63a2\u7d22\u5982\u4f55\u5c06\u5176\u76f4\u63a5\u6574\u5408\u8fdb\u6a21\u578b\u8bad\u7ec3\u4e2d\u3002\u4e4b\u524d\u4eba\u4eec\u63d0\u51fa\u7684\u65b9\u6cd5\u5305\u62ec\u6536\u96c6\u4e13\u95e8\u7684\u8bad\u7ec3\u6570\u636e\u3001\u6784\u5efa\u5956\u52b1\u6a21\u578b\u548c\u589e\u52a0\u89e3\u7801\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u4f46\u76ee\u524d\u8fd8\u6ca1\u6709\u4e00\u79cd\u65b9\u6cd5\u80fd\u5927\u89c4\u6a21\u5730\u5728\u6027\u80fd\u4e0a\u53d6\u5f97\u91cd\u5927\u7a81\u7834\u3002<\/section>\n<section><\/section>\n<section>\u6c6a\u519b\u6559\u6388\u8868\u793a\uff0c\u6211\u4eec\u76ee\u524d\u5c1a\u4e0d\u6e05\u695a OpenAI \u7684 o1 \u521b\u65b0\u662f\u5426\u690d\u6839\u4e8e\u6a21\u578b\u672c\u8eab\uff0c\u8fd8\u662f\u4f9d\u7136\u4f9d\u8d56\u4e8e\u5916\u90e8\u63d0\u793a\u7cfb\u7edf\u3002\u5982\u679c\u5b83\u786e\u5b9e\u6d89\u53ca\u5728\u67b6\u6784\u4e2d\u660e\u786e\u5d4c\u5165\u5206\u6b65\u63a8\u7406\uff0c\u90a3\u4e48\u8fd9\u5c06\u662f\u4e00\u4e2a\u91cd\u5927\u7a81\u7834\u3002<\/section>\n<section><\/section>\n<section>\u5728\u5927\u5e45\u63d0\u9ad8\u6027\u80fd\u7684\u57fa\u7840\u4e0a\uff0cOpenAI o1 \u8fd8\u8868\u660e\uff0c\u4f20\u7edf\u4e0a\u5728\u8bad\u7ec3\u671f\u95f4\u5e94\u7528\u7684\u6269\u5c55\u539f\u5219\u73b0\u5728\u4e5f\u4e0e\u63a8\u7406\u9636\u6bb5\u76f8\u5173\u4e86\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u6837\u4e00\u6765\uff0c\u5c31\u9700\u8981\u8003\u8651\u7ed9\u63a8\u7406\u9636\u6bb5\u591a\u5206\u914d\u4e00\u4e9b\u7b97\u529b\u4e86\u3002\u5982\u679c\u80fd\u8ba9 LLM \u901a\u8fc7\u589e\u52a0\u6d4b\u8bd5\u65f6\u95f4\u8ba1\u7b97\u6765\u63d0\u5347\u8f93\u51fa\uff0c\u90a3\u4fbf\u662f\u671d\u7740<strong>\u81ea\u6211\u6539\u8fdb\u5f0f\u667a\u80fd\u4f53\uff08self-improving agent\uff09<\/strong>\u8fc8\u51fa\u7684\u91cd\u8981\u4e00\u6b65\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u4e2a\u7814\u7a76\u65b9\u5411\u88ab\u6c6a\u519b\u6559\u6388\u6682\u4e14\u79f0\u4e3a\u00a0<strong>LLM \u539f\u751f\u601d\u7ef4\u94fe\uff08LLM-Native Chain-of-Thought\/NativeCoT\uff09<\/strong>\uff0c\u5176\u5e94\u5f53\u80fd\u591f\u56fa\u6709\u5730\u53cd\u6620\u4eba\u7c7b\u7cfb\u7edf 2 \u601d\u7ef4\u6240\u5177\u6709\u7684\u6df1\u601d\u719f\u8651\u7684\u5206\u6790\u8fc7\u7a0b\u3002<\/section>\n<section><\/section>\n<section>\u4e0d\u8fc7\uff0c\u8003\u8651\u5230 o1 \u662f\u4e00\u4e2a\u95ed\u6e90\u7cfb\u7edf\uff0c\u56e0\u6b64\u5b83\u7a76\u7adf\u662f\u5982\u4f55\u5b9e\u73b0\u5982\u6b64\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\u7684\u8fd8\u4f9d\u7136\u662f\u4e00\u4e2a\u8c1c\u3002<\/section>\n<section><\/section>\n<section>\u5728\u672c\u6587\u4e2d\uff0c\u6c6a\u519b\u6559\u6388\u5168\u9762\u56de\u987e\u4e86\u53ef\u80fd\u7684\u76f8\u5173\u6587\u732e\uff0c\u5e76\u63a2\u8ba8\u4e86\u8fd9\u4e00\u7a81\u7834\u80cc\u540e\u53ef\u80fd\u7684\u6838\u5fc3\u6280\u672f\u548c\u65b9\u6cd5\u3002\u6b64\u5916\uff0c\u4ed6\u8fd8\u63d0\u51fa\u4e86\u57fa\u4e8e\u8fd1\u671f\u7814\u7a76\u6210\u679c\u5b9e\u73b0\u76f8\u5e94\u5f00\u6e90\u7248\u672c\u7684\u65b9\u6cd5\uff0c\u4ee5\u52a0\u901f\u8be5\u9886\u57df\u7684\u7814\u7a76\u3002<\/section>\n<section><\/section>\n<section>\u4e0b\u9762\u9996\u5148\u5c06\u4ecb\u7ecd\u5178\u578b\u81ea\u56de\u5f52 LLM \u5e38\u9047\u5230\u7684\u4e24\u4e2a\u6311\u6218\uff0c\u4ee5\u5f3a\u8c03\u5bf9\u4e16\u754c\u6a21\u578b\u548c\u601d\u7ef4\u94fe\u673a\u5236\u7684\u9700\u6c42\u3002\u7136\u540e\u5c06\u7ed9\u51fa\u4e00\u4e2a MDP \u516c\u5f0f\uff0c\u7528\u4e8e\u5c06\u539f\u751f CoT \u6574\u5408\u8fdb LLM\uff08\u8fdb\u800c\u5f97\u5230\u7c7b\u4f3c o1 \u7684\u63a8\u7406\u6a21\u578b\uff09\uff1b\u540c\u65f6\u8fd8\u4f1a\u63a2\u7d22\u5176\u5b9e\u73b0\u7ec6\u8282\u3002\u6700\u540e\u4f1a\u4ee5\u6587\u732e\u8bc4\u8bba\u4f5c\u7ed3\uff0c\u5e76\u7ed9\u51fa\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\u3002<\/section>\n<section><\/section>\n<section><strong>\u81ea\u56de\u5f52 LLM \u9762\u4e34\u7684\u6311\u6218<\/strong><\/section>\n<section><\/section>\n<section>\u8fd9\u91cc\u5c31\u7565\u8fc7\u81ea\u56de\u5f52 LLM \u7684\u57fa\u7840\u4ecb\u7ecd\uff0c\u91cd\u70b9\u6765\u770b\u6c6a\u519b\u6559\u6388\u63d0\u5230\u7684\u4e24\u4e2a\u6311\u6218\u3002<\/section>\n<section><\/section>\n<section>\u7b2c\u4e00\u4e2a\u6311\u6218\u662f\u9884\u6d4b\u4e0b\u4e00 token \u7684\u76ee\u6807\u3002\u867d\u7136\u6709\u4e9b\u4eba\u8ba4\u4e3a\u9884\u6d4b\u4e0b\u4e00 token \u53ef\u80fd\u4f1a\u9020\u5c31\u901a\u7528\u667a\u80fd\uff08AGI\uff09\uff0c\u4f46\u6c6a\u519b\u6559\u6388\u8868\u793a\uff0c\u4ec5\u4ec5\u4e13\u6ce8\u4e8e\u9884\u6d4b\u4e0b\u4e00\u4e2a\u8bcd\u4f1a\u9650\u5236\u667a\u80fd\u7684\u6f5c\u529b\u3002\u4e3a\u4e86\u5f97\u5230\u66f4\u6df1\u5c42\u6b21\u7684\u667a\u80fd\uff0c\u53ef\u80fd\u9700\u8981\u4e0d\u540c\u7684\u4f18\u5316\u76ee\u6807\u548c\u5b66\u4e60\u8303\u5f0f\u3002<\/section>\n<section><\/section>\n<section>\u4f20\u7edf\u81ea\u56de\u5f52 LLM \u9762\u4e34\u7740\u4e00\u4e2a\u5173\u952e\u6311\u6218\uff1a\u5982\u4f55\u4f7f\u7cfb\u7edf\u8d85\u8d8a\u5176\u8bad\u7ec3\u6570\u636e\u7684\u754c\u9650\u5e76\u5f00\u53d1\u51fa\u65b0\u9896\u7684\u3001\u53ef\u80fd\u66f4\u4f18\u7684\u7b56\u7565\uff1f\u56e0\u4e3a\u667a\u80fd\u4f53\u7684\u8868\u73b0\u5e38\u5e38\u53d7\u9650\u4e8e\u5176\u5b66\u4e60\u7684\u6f14\u793a\u7684\u8d28\u91cf\uff0c\u65e0\u6cd5\u8d85\u8d8a\u5176\u8bad\u7ec3\u6570\u636e\u6240\u4f53\u73b0\u51fa\u7684\u6280\u80fd\u6c34\u5e73\u3002<\/section>\n<section><\/section>\n<section>\u4f46\u662f\uff0c\u5982\u679c\u4f7f\u7528\u6570\u636e\u6765\u5f00\u53d1\u66f4\u6df1\u5ea6\u7684\u7406\u89e3\u6216\u4e16\u754c\u6a21\u578b\uff0c\u5c31\u6709\u53ef\u80fd\u5b9e\u73b0\u590d\u6742\u7b56\u7565\u7684\u6f14\u8fdb\uff0c\u8fdb\u800c\u8d85\u8d8a\u8bad\u7ec3\u6570\u636e\u7684\u9650\u5236\u3002<\/section>\n<section><\/section>\n<section>\u4e16\u754c\u6a21\u578b\uff08world model\uff09\u4ee3\u8868\u4e86\u667a\u80fd\u4f53\u5bf9\u73af\u5883\u7684\u7406\u89e3\u3002\u5b66\u4e60\u548c\u4f18\u5316\u8fd9\u4e2a\u4e16\u754c\u6a21\u578b\uff0c\u518d\u52a0\u4e0a\u6a21\u62df\u6f5c\u5728\u7ed3\u679c\u7684\u80fd\u529b\uff0c\u6709\u671b\u6781\u5927\u63d0\u5347 AI \u667a\u80fd\u4f53\u7684\u80fd\u529b\u3002\u8fd9\u4e9b\u5185\u90e8\u4e16\u754c\u6a21\u578b\u6240\u63d0\u4f9b\u7684\u6a21\u62df\u80fd\u529b\u5c06\u80fd\u5b9e\u73b0\u6df1\u5ea6\u601d\u8003\uff08\u6a21\u62df\uff09\uff0c\u4ece\u800c\u589e\u5f3a\u667a\u80fd\u4f53\u7684\u63a8\u7406\u548c\u6cdb\u5316\u80fd\u529b\u3002\u57fa\u4e8e\u6a21\u578b\u7684\u7b56\u7565\uff08\u5982\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22 \uff08MCTS\uff09\uff09\u662f\u8fd9\u79cd\u65b9\u6cd5\u7684\u7ecf\u5178\u4f8b\u8bc1\u3002\u5411\u7cfb\u7edf 2 \u578b\u63a8\u7406\u7684\u8fc7\u6e21\uff08o1 \u53ef\u80fd\u5c31\u662f\u4e00\u4e2a\u4f8b\u8bc1\uff09\u4f9d\u8d56\u4e8e\u5efa\u7acb\u67d0\u79cd\u7c7b\u578b\u7684\u4e16\u754c\u6a21\u578b\u5e76\u5229\u7528\u5f3a\u5316\u5b66\u4e60\uff08\u5956\u52b1\u6700\u5927\u5316\uff09\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f\u6700\u5c0f\u5316\u9884\u6d4b\u8bef\u5dee\u3002\u8fd9\u79cd\u65b9\u6cd5\u7684\u8f6c\u53d8\u53ef\u80fd\u662f OpenAI o1 \u5f3a\u5927\u63a8\u7406\u80fd\u529b\u80cc\u540e\u7684\u5173\u952e\u8fc7\u6e21\u6280\u672f\u4e4b\u4e00\u3002<\/section>\n<section><\/section>\n<section><strong>\u901a\u8fc7\u5c06 LLM \u7684\u9884\u6d4b\u80fd\u529b\u4e0e\u5f3a\u5316\u5b66\u4e60\u548c\u4e16\u754c\u5efa\u6a21\u7684\u7b56\u7565\u6df1\u5ea6\u76f8\u7ed3\u5408\uff0c\u50cf o1 \u8fd9\u6837\u7684 AI \u7cfb\u7edf\u53ef\u4ee5\u89e3\u51b3\u66f4\u590d\u6742\u7684\u95ee\u9898\u548c\u5b9e\u73b0\u66f4\u590d\u6742\u7684\u51b3\u7b56\u8fc7\u7a0b\u3002\u8fd9\u79cd\u6df7\u5408\u65b9\u6cd5\u65e2\u53ef\u4ee5\u5b9e\u73b0\u5feb\u901f\u6a21\u5f0f\u8bc6\u522b\uff08\u7c7b\u4f3c\u4e8e\u7cfb\u7edf 1 \u601d\u7ef4\uff09\uff0c\u4e5f\u53ef\u4ee5\u5b9e\u73b0\u6df1\u601d\u719f\u8651\u7684\u9010\u6b65\u63a8\u7406\uff08\u7cfb\u7edf 2 \u601d\u7ef4\u7684\u7279\u5f81\uff09\u3002<\/strong>\u8fd9\u4e5f\u8bb8\u80fd\u89e3\u91ca o1 \u8868\u73b0\u51fa\u7684\u5f3a\u5927\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section>\u7b2c\u4e8c\u4e2a\u6311\u6218\u5219\u6765\u81ea\u8ba1\u7b97\u590d\u6742\u6027\u89d2\u5ea6\uff1a<strong>LLM \u8fd0\u884c\u65f6\u53d7\u5230\u4e8c\u6b21\u8ba1\u7b97\u590d\u6742\u6027\u7684\u7ea6\u675f<\/strong>\u3002\u5f53 LLM \u9047\u5230\u591a\u6b65\u6570\u5b66\u96be\u9898\u65f6\uff0c\u8fd9\u79cd\u7ea6\u675f\u4f1a\u53d8\u5f97\u5c24\u4e3a\u660e\u663e\u3002<\/section>\n<section><\/section>\n<section>\u4f46\u662f\uff0c\u601d\u7ef4\u94fe\u5374\u6709\u671b\u51cf\u8f7b\u8fd9\u4e00\u9650\u5236\u3002\u5176\u53ef\u901a\u8fc7\u4e00\u7cfb\u5217\u300c\u601d\u7ef4\u300d\u6b65\u9aa4\u6765\u6269\u5c55\u54cd\u5e94\uff0c\u7531\u6b64\u652f\u6301\u4e00\u5b9a\u6570\u91cf\u7684\u989d\u5916\u8ba1\u7b97\u8d44\u6e90\uff1b\u5b83\u672c\u8d28\u4e0a\u662f\u4e00\u4e2a\u6709\u9650\u7684\u5185\u5b58\uff0c\u652f\u6301\u5199\u5165\u4f46\u7f3a\u4e4f\u5220\u9664\u6216\u8986\u76d6\u7684\u80fd\u529b\u3002\u5c3d\u7ba1\u8be5\u65b9\u6cd5\u9887\u5177\u6f5c\u529b\uff0c\u4f46\u5b83\u4ecd\u7136\u4e0d\u662f\u4e00\u4e2a\u5b8c\u5168\u52a8\u6001\u7684\u5185\u5b58\u7cfb\u7edf\uff0c\u5e76\u4e14\u6ca1\u6709\u539f\u751f\u5730\u878d\u5165\u89e3\u7801\u9636\u6bb5\u3002\u8fd9\u79cd\u5fc5\u8981\u6027\u4f7f\u5f97\u7814\u7a76\u793e\u533a\u4e9f\u9700\u8d85\u8d8a\u5f53\u524d Transformer \u89e3\u7801\u5668\u7f51\u7edc\u80fd\u529b\u7684\u9ad8\u7ea7\u8ba1\u7b97\u67b6\u6784\u3002\u4e8b\u5b9e\u4e0a\uff0c\u5b58\u5728\u8fd9\u6837\u7684\u9700\u6c42\uff1a<strong>\u5728\u63a8\u7406\u548c\u89e3\u7801\u9636\u6bb5\u5b9e\u73b0\u7c7b\u4f3c\u4e8e\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22 \uff08MCTS\uff09\u7684\u57fa\u4e8e\u6a21\u578b\u7684\u590d\u6742\u7b56\u7565\u3002<\/strong><\/section>\n<section><\/section>\n<section><strong>\u8fd9\u79cd\u5148\u8fdb\u7684\u63a8\u7406\u65f6\u95f4\u8ba1\u7b97\u7cfb\u7edf\u5c06\u4f7f AI \u6a21\u578b\u80fd\u591f\u7ef4\u62a4\u548c\u52a8\u6001\u66f4\u65b0\u95ee\u9898\u7a7a\u95f4\u7684\u8868\u5f81\uff0c\u4ece\u800c\u4fc3\u8fdb\u66f4\u590d\u6742\u7684\u63a8\u7406\u8fc7\u7a0b\u3002<\/strong>\u8fd9\u79cd\u65b9\u6cd5\u4e0e\u8ba4\u77e5\u79d1\u5b66\u4e2d\u7684\u5de5\u4f5c\u8bb0\u5fc6\uff08working memory\uff09\u6982\u5ff5\u4e00\u81f4\uff1b\u8fd9\u5bf9\u4e8e\u590d\u6742\u7684\u95ee\u9898\u89e3\u51b3\u548c\u6df1\u5ea6\u601d\u8003\u81f3\u5173\u91cd\u8981\u3002\u901a\u8fc7\u6574\u5408\u8fd9\u4e9b\u529f\u80fd\uff0cAI \u7cfb\u7edf\u53ef\u4ee5\u6a21\u62df\u591a\u4e2a\u6b65\u9aa4\uff0c\u8bc4\u4f30\u4e0d\u540c\u7684\u573a\u666f\uff0c\u5e76\u505a\u51fa\u66f4\u660e\u667a\u7684\u51b3\u7b56 \u2014\u2014 \u7c7b\u4f3c\u4e8e\u4eba\u7c7b\u4e13\u5bb6\u63a8\u7406\u7684\u6df1\u601d\u719f\u8651\u8fc7\u7a0b\u3002<\/section>\n<section><\/section>\n<section><strong>\u5c06 LLM \u63a8\u7406\u770b\u4f5c\u662f\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u5efa\u6a21\u95ee\u7b54\u6216\u95ee\u9898\u89e3\u7b54\u7b49\u4efb\u52a1\u4e2d\u7684\u63a8\u7406\u8fc7\u7a0b\uff0c\u8fd9\u91cc\u8981\u5c06\u63a8\u7406\u7684\u7ed3\u6784\u8c03\u6574\u6210 Q \u2192 {R} \u2192 A \u5e8f\u5217\u7684\u5f62\u5f0f\u3002<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>Q\uff1a\u8868\u793a\u542f\u52a8\u63a8\u7406\u8fc7\u7a0b\u7684\u95ee\u9898\u6216\u63d0\u793a\u8bcd\uff1b<\/section>\n<\/li>\n<li>\n<section>R\uff1a\u8868\u793a\u4e3a\u4e86\u5f97\u5230\u89e3\u7b54\uff0c\u6a21\u578b\u751f\u6210\u7684\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u7684\u5e8f\u5217\uff1b<\/section>\n<\/li>\n<li>\n<section>A\uff1a\u8868\u793a\u63a8\u7406\u6b65\u9aa4\u5b8c\u6210\u540e\u5f97\u5230\u7684\u6700\u7ec8\u7b54\u6848\u6216\u89e3\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u8fd9\u79cd\u7ed3\u6784\u5141\u8bb8 LLM \u751f\u6210\u4e00\u7cfb\u5217\u63a8\u7406\u6b65\u9aa4\uff0c\u4ece\u903b\u8f91\u4e0a\u5c06\u95ee\u9898 Q \u4e0e\u6700\u7ec8\u7b54\u6848 A \u8054\u7cfb\u8d77\u6765\u3002<\/section>\n<section><\/section>\n<section>\u6c6a\u519b\u6559\u6388\u8868\u793a\uff0c\u53ef\u4ee5\u5c06\u8be5\u63a8\u7406\u8fc7\u7a0b\u5b9a\u4e49\u4e3a\u4e00\u4e2a\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b\uff08MDP\uff09\u3002MDP \u80fd\u4e3a\u5efa\u6a21\u63a8\u7406\u63d0\u4f9b\u4e00\u4e2a\u7075\u6d3b\u7684\u6846\u67b6\u3002\u5b83\u5141\u8bb8\u6a21\u578b\u81ea\u56de\u5f52\u5730\u751f\u6210\u8fc8\u5411\u6700\u7ec8\u7b54\u6848\u7684\u987a\u5e8f\u63a8\u7406\u6b65\u9aa4\uff0c\u540c\u65f6\u8fd8\u901a\u8fc7\u5728\u6bcf\u4e2a\u6b65\u9aa4\u91c7\u6837\u591a\u6761\u8def\u5f84\u6765\u5b9e\u73b0\u6811\u7ed3\u6784\u4ee5\u83b7\u5f97\u5907\u9009\u63a8\u7406\u8f68\u8ff9\u3002\u901a\u8fc7\u7ed3\u5408\u987a\u5e8f\u63a8\u7406\u548c\u5206\u652f\u63a8\u7406\u8fd9\u4e24\u79cd\u65b9\u6cd5\uff0c\u8be5\u6a21\u578b\u53ef\u4ee5\u63a2\u7d22\u5404\u79cd\u89e3\u51b3\u65b9\u6848\uff0c\u4ece\u800c\u521b\u5efa\u4e00\u4e2a\u591a\u529f\u80fd\u4e14\u5168\u9762\u7684\u63a8\u7406\u8fc7\u7a0b\u3002<\/section>\n<section><\/section>\n<section>\u73b0\u5728\u53ef\u4ee5\u4f7f\u7528\u72b6\u6001\u3001\u52a8\u4f5c\u3001\u7b56\u7565\u548c\u5956\u52b1\u6765\u63cf\u8ff0\u8fd9\u4e2a\u63a8\u7406\u8fc7\u7a0b\u4e86\u3002\u5176\u4e2d LLM \u7684\u4efb\u52a1\u662f\u9010\u6b65\u751f\u6210\u4e0e\u63a8\u7406\u6b65\u9aa4\u548c\u6700\u7ec8\u7b54\u6848\u76f8\u5bf9\u5e94\u7684\u8fde\u8d2f token \u5e8f\u5217\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0fe911e49119daa01b0832d4acfb4ac3.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0fe911e49119daa01b0832d4acfb4ac3.png\"><\/a><\/p>\n<section><em><sup>\u56fe 3\uff1a\u5728\u8be5\u5f62\u5f0f\u7684\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b\u4e2d\uff0cLLM \u7684\u4efb\u52a1\u662f\u9010\u6b65\u751f\u6210\u63a8\u7406\u6b65\u9aa4\u548c\u95ee\u9898\u7684\u6700\u7ec8\u7b54\u6848\u3002\u5176\u4e2d\uff0cLLM \u7b56\u7565\u7684\u8fd0\u4f5c\u65b9\u5f0f\u662f\u751f\u6210 token\uff0c\u800c\uff0c\u8fd9\u4e9b token \u53ef\u4ee5\u5f62\u6210\u66f4\u9ad8\u7ea7\u7684\u63a8\u7406\u7ed3\u6784\u3002\u72b6\u6001\u8868\u793a\u8fc4\u4eca\u4e3a\u6b62\u7684\u63a8\u7406\u6b65\u9aa4\u5e8f\u5217\uff0c\u52a8\u4f5c\u5bf9\u5e94\u4e8e\u9009\u53d6\u65b0\u63a8\u7406\u6b65\u9aa4\u6216\u6700\u7ec8\u7b54\u6848\u3002LLM \u7b56\u7565\u63a7\u5236\u52a8\u4f5c\u7684\u9009\u62e9\uff0c\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff08PRM\uff09\u7684\u4f5c\u7528\u662f\u63d0\u4f9b\u6709\u5173\u63a8\u7406\u6b65\u9aa4\u548c\u6700\u7ec8\u7b54\u6848\u8d28\u91cf\u7684\u53cd\u9988\u3002\u901a\u8fc7\u4f18\u5316\u7b56\u7565\u4ee5\u6700\u5927\u5316\u5956\u52b1\uff0cLLM \u53ef\u4ee5\u5728 PRM \u7684\u5f15\u5bfc\u4e0b\u751f\u6210\u51c6\u786e\u4e14\u6709\u610f\u4e49\u7684\u63a8\u7406\u8fc7\u7a0b\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5176\u4e2d\u72b6\u6001\u8868\u793a\u8fc4\u4eca\u4e3a\u6b62\u7684\u63a8\u7406\u6b65\u9aa4\u987a\u5e8f\uff0c\u800c\u52a8\u4f5c\u5bf9\u5e94\u4e8e\u9009\u62e9\u65b0\u63a8\u7406\u6b65\u9aa4\u6216\u6700\u7ec8\u7b54\u6848\u3002LLM \u7b56\u7565\u63a7\u5236\u7740\u52a8\u4f5c\u7684\u9009\u62e9\uff0c\u800c\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b \uff08PRM\uff09\u5219\u63d0\u4f9b\u6709\u5173\u63a8\u7406\u6b65\u9aa4\u548c\u6700\u7ec8\u7b54\u6848\u8d28\u91cf\u7684\u53cd\u9988\u3002\u901a\u8fc7\u4f18\u5316\u7b56\u7565\u4ee5\u6700\u5927\u5316\u5956\u52b1\uff0cLLM \u53ef\u4ee5\u5728 PRM \u7684\u5f15\u5bfc\u4e0b\u751f\u6210\u51c6\u786e\u4e14\u6709\u610f\u4e49\u7684\u63a8\u7406\u8fc7\u7a0b\u3002<\/section>\n<section><\/section>\n<section>\u8be6\u7ec6\u7684\u8bba\u8bc1\u8fc7\u7a0b\u8fd9\u91cc\u5c31\u4e0d\u591a\u8c08\u4e86\u3002\u603b\u4e4b\uff0c\u53ef\u4ee5\u57fa\u4e8e\u6b64\u5f97\u5230 LLM \u7684\u4e16\u754c\u6a21\u578b\u7684\u5b9a\u4e49\uff1a<\/section>\n<section><\/section>\n<section>\u5b9a\u4e49 1\uff1a<strong>LLM \u7684\u4e16\u754c\u6a21\u578b<\/strong>\u53ef\u4ee5\u5b9a\u4e49\u4e3a (\ud835\udcaf, \ud835\udcb1)\uff0c\u5176\u4e2d\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\ud835\udcaf (s_t, a_t) \u662f\u8f6c\u6362\u6a21\u578b\uff0c\u5b83\u662f\u786e\u5b9a\u6027\u7684\uff0c\u56e0\u4e3a\u5f53\u524d\u72b6\u6001 s_t \u548c\u52a8\u4f5c a_t \u4ec5\u80fd\u5b9a\u4e49\u552f\u4e00\u4e0b\u4e00\u72b6\u6001 s_(t+1) \uff0c\u56e0\u6b64 s_(t+1) = s_t + a_t\u3002<\/section>\n<\/li>\n<li>\n<section>\ud835\udcb1 (s_t, a_t) \u662f\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff08PRM\uff09\uff0c\u7528\u4e8e\u8bc4\u4f30\u5728\u72b6\u6001 s_t \u4e0b\u52a8\u4f5c a_t \u7684\u8d28\u91cf\u3002\u5b83\u80fd\u53cd\u6620\u751f\u6210\u7684\u63a8\u7406\u6b65\u9aa4\u6216 token \u5728\u5f97\u5230\u6700\u7ec8\u7b54\u6848\u8fc7\u7a0b\u4e2d\u7684\u5408\u9002\u7a0b\u5ea6\u548c\u6709\u6548\u6027\uff1a\ud835\udcb1 (s_t, a_t)=\ud835\udc63_t\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u7531\u4e8e\u8f6c\u6362\u662f\u786e\u5b9a\u6027\u7684\u5e76\u4e14\u76f4\u63a5\u9075\u5faa\u7b56\u7565\uff0c\u56e0\u6b64\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b \ud835\udcb1 (s_t, a_t) \u5c01\u88c5\u4e86 LLM \u4e0e\u5176\u73af\u5883\u4e4b\u95f4\u7684\u6574\u4e2a\u4ea4\u4e92\uff0c\u53ef\u8bc4\u4f30\u6bcf\u4e2a\u63a8\u7406\u6b65\u9aa4\u6216 token \u5bf9\u6240\u5f97\u51fa\u7684\u6700\u7ec8\u7b54\u6848\u7684\u8d21\u732e\u7a0b\u5ea6\u3002<\/section>\n<section><\/section>\n<section><strong>\u5b9e\u9645\u5b9e\u73b0<\/strong><\/section>\n<section><\/section>\n<section>\u63a5\u4e0b\u6765\u5c06\u4ecb\u7ecd\u5982\u4f55\u6536\u96c6\u4e2d\u95f4\u63a8\u7406\u6570\u636e\uff0c\u5e76\u4f7f\u7528\u5b83\u6765\u8bad\u7ec3\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff08PRM\uff09\uff0c\u518d\u5229\u7528 PRM \u6765\u8bad\u7ec3 LLM \u7b56\u7565\uff0c\u5e76\u5728\u89e3\u7801\u9636\u6bb5\u5f15\u5bfc\u63a8\u7406\u8fc7\u7a0b\u3002<\/section>\n<section><\/section>\n<section><strong>\u81ea\u52a8\u83b7\u53d6\u63a8\u7406\u6b65\u9aa4\u6570\u636e<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u6a21\u62df\u51fa\u9ad8\u7ea7\u63a8\u7406\u80fd\u529b\uff0c\u5c31\u9700\u8981\u63a8\u7406\u8f68\u8ff9\u6570\u636e\u3002\u6700\u76f4\u63a5\u7684\u65b9\u6cd5\u5f53\u7136\u662f\u4eba\u5de5\u6807\u6ce8\u63a8\u7406\u6b65\u9aa4\uff0c\u4f46\u8fd9\u79cd\u65b9\u6cd5\u7f3a\u70b9\u4e5f\u5f88\u660e\u663e\u3002<\/section>\n<section><\/section>\n<section>\u4e00\u79cd\u65e0\u9700\u4eba\u7c7b\u76d1\u7763\uff0c\u7279\u522b\u6709\u6548\u7684\u6536\u96c6\u6570\u636e\u548c\u63d0\u5347 LLM \u63a8\u7406\u7684\u65b9\u6cd5\u662f Self-Taught Reasoner\uff08STaR\uff09\u3002<\/section>\n<section><\/section>\n<section>\u4f7f\u7528 STaR \u65b9\u6cd5\u65f6\uff0c\u6a21\u578b\u4f1a\u81ea\u4e3b\u751f\u6210\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u5e76\u4f7f\u7528\u5b83\u4eec\u6765\u9a8c\u8bc1\u5176\u5185\u90e8\u63a8\u7406\u80fd\u529b\u3002\u66f4\u65b9\u6cd5\u7684\u57fa\u7840\u662f LLM \u6709\u80fd\u529b\u901a\u8fc7\u751f\u6210\u4e2d\u95f4\u6b65\u9aa4 {R_1, R_2, . . . , R_n} \u00a0\u4ece\u95ee\u9898 Q \u63a8\u7406\u5230\u6700\u7ec8\u7b54\u6848 A\uff0c\u5e76\u4f7f\u7528\u81ea\u5df1\u7684\u7b56\u7565\u9a8c\u8bc1\u6b63\u786e\u6027\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-51c378d933fdafca16740ffb5e7fc81d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-51c378d933fdafca16740ffb5e7fc81d.png\"><\/a><\/p>\n<section>\u4e5f\u5c31\u662f\u8bf4\uff0c\u8be5\u65b9\u6cd5\u9996\u5148\u4f1a\u91c7\u7528 LLM \u7684\u7b56\u7565 \u03c0_LLM\uff0c\u57fa\u4e8e\u521d\u59cb\u95ee\u9898 Q \u548c\u6700\u7ec8\u7b54\u6848 A \u6765\u751f\u6210\u63a8\u7406\u6b65\u9aa4 {R}\u3002<\/section>\n<section><\/section>\n<section>\u751f\u6210 {R} \u4e4b\u540e\uff0c\u5c31\u8981\u9a8c\u8bc1\u5176\u6b63\u786e\u6027\u3002\u8fd9\u91cc\u53ef\u4ee5\u518d\u6b21\u4f7f\u7528\u8fd9\u4e2a LLM \u7b56\u7565\u3002<\/section>\n<section><\/section>\n<section>\u4e4b\u540e\uff0c\u6536\u96c6\u5230\u7684 {Q, {R}, A} \u5c31\u53ef\u8fdb\u4e00\u6b65\u7528\u4e8e\u8bad\u7ec3\u7b56\u7565 \u03c0_LLM\uff0c\u63d0\u5347\u6709\u6548\u63a8\u7406\u6b65\u9aa4\u7684\u751f\u6210\u8fc7\u7a0b\u3002<\/section>\n<section><\/section>\n<section>\u5f53\u63a8\u7406\u5e8f\u5217\u8f83\u957f\u65f6\uff0c\u8fd8\u4f1a\u7528\u5230\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22\uff08MCTS\uff09\u3002<\/section>\n<section><\/section>\n<section><strong>\u81ea\u6211\u589e\u5f3a\u5f0f\u8bad\u7ec3<\/strong><\/section>\n<section><strong><br \/><\/strong><\/section>\n<section>\u5982\u56fe 4 \u6240\u793a\uff0cPRM v (s) \u548c LLM \u7b56\u7565 \u03c0_LLM \u53ef\u4ee5\u76f8\u4e92\u589e\u5f3a\u4ee5\u5b9e\u73b0\u81ea\u6211\u63d0\u5347\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-feb17b12ee8fb753c3431a24d080bf40.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-feb17b12ee8fb753c3431a24d080bf40.png\"><\/a><\/p>\n<section><em><sup>\u56fe 4\uff1a\u5c06 PRM \u7684\u4ef7\u503c\u51fd\u6570\u4e0e LLM \u7684\u7b56\u7565\u751f\u6210\u76f8\u7ed3\u5408\uff0c\u53ef\u786e\u4fdd\u5f97\u5230\u7684\u7ed3\u679c\u662f\u7ecf\u8fc7\u5f15\u5bfc\u4e14\u53ef\u63a7\u7684\u3002\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0cLLM \u7b56\u7565\u5f97\u5230\u7684\u751f\u6210\u7ed3\u679c\u548c PRM \u63d0\u4f9b\u7684\u8bc4\u4f30\u76f8\u4e92\u589e\u5f3a\uff0c\u4ece\u800c\u53ef\u8ba9\u8fd9\u4e24\u4e2a\u7ec4\u4ef6\u4e0d\u65ad\u81ea\u6211\u6539\u8fdb\u548c\u4f18\u5316\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<ul>\n<li>\n<section>PRM \u7684\u4ef7\u503c\u8fed\u4ee3<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u6709\u4e86\u63a8\u7406\u6570\u636e\u4e4b\u540e\uff0c\u4e0b\u4e00\u6b65\u5c31\u662f\u8bad\u7ec3\u4e16\u754c\u6a21\u578b\u4e86\uff0c\u4e5f\u88ab\u79f0\u4e3a\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff08PRM\uff09\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u7531\u4e8e\u72b6\u6001\u8f6c\u53d8\u662f\u786e\u5b9a\u548c\u5df2\u77e5\u7684\uff0c\u56e0\u6b64\u91cd\u70b9\u5c31\u53d8\u6210\u4e86\u5b66\u4e60\u4e00\u4e2a\u4e4b\u540e\u53ef\u7528\u4e8e\u5f15\u5bfc\u641c\u7d22\u3001\u63a8\u7406\u548c\u89e3\u7801\u8fc7\u7a0b\u7684\u901a\u7528\u5956\u52b1\u6a21\u578b\u3002<\/section>\n<section><\/section>\n<section>\u8be5\u5956\u52b1\u6a21\u578b\u901a\u5e38\u88ab\u79f0\u4e3a\u9a8c\u8bc1\u5668\uff0c\u8bb0\u4e3a vPRM (s)\uff0c\u53ef\u4ee5\u4f7f\u7528\u6709\u6807\u6ce8\u7684\u63a8\u7406\u6b65\u9aa4\u6570\u636e\u96c6\u8fdb\u884c\u8bad\u7ec3\u3002\u5176\u8bad\u7ec3\u901a\u5e38\u6d89\u53ca\u6839\u636e\u63a8\u7406\u6b65\u9aa4\u7684\u6b63\u786e\u6027\u4f18\u5316\u4e00\u4e2a\u5206\u7c7b\u635f\u5931\u51fd\u6570\uff1a<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-05a1dc5da09168ce5e52545f1cee5be6.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-05a1dc5da09168ce5e52545f1cee5be6.png\"><\/a><\/p>\n<section>\u8fd8\u6709\u53e6\u4e00\u79cd\u65b9\u6cd5\uff0c\u662f\u5c06 PRM \u89c6\u4e3a\u4e00\u4e2a\u53ef\u4ee5\u901a\u8fc7\u4ef7\u503c\u8fed\u4ee3\u65b9\u6cd5\u8bad\u7ec3\u7684\u4ef7\u503c\u51fd\u6570\uff0c\u4f7f\u5176\u80fd\u591f\u9884\u6d4b\u7d2f\u79ef\u5956\u52b1\u5e76\u901a\u8fc7\u6700\u4f73\u52a8\u4f5c\u9009\u62e9\u6307\u5bfc\u63a8\u7406\u8fc7\u7a0b\u3002<\/section>\n<section><\/section>\n<section>\u5047\u8bbe\u6709\u4e00\u4e2a\u63a8\u7406\u8fc7\u7a0b\uff0c\u5176\u4e2d\u72b6\u6001 s \u8868\u793a\u5f53\u524d\u72b6\u6001\uff0c\u5e76\u4e14\u5176\u6574\u5408\u4e86\u4e4b\u524d\u7684\u6240\u6709\u72b6\u6001\u3002\u8be5\u4ef7\u503c\u8fed\u4ee3\u65b9\u6cd5\u7684\u76ee\u6807\u662f\u5b66\u4e60\u4e00\u4e2a\u7531 \u03b8 \u53c2\u6570\u5316\u7684\u4ef7\u503c\u51fd\u6570 V_\u03b8 (s)\uff0c\u5176\u53ef\u9884\u6d4b\u4ece\u72b6\u6001 s \u5f00\u59cb\u7684\u9884\u671f\u7d2f\u79ef\u5956\u52b1\u3002\u8be5\u4ef7\u503c\u51fd\u6570\u53ef\u901a\u8fc7\u8bc4\u4f30\u4e0d\u540c\u52a8\u4f5c\u7684\u6f5c\u5728\u7ed3\u679c\u6765\u6307\u5bfc\u63a8\u7406\u8fc7\u7a0b\u3002r_\u03c6 (s) \u662f\u5956\u52b1\u51fd\u6570\uff0c\u5176\u4f1a\u6839\u636e\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u6216\u6700\u7ec8\u7b54\u6848\u7684\u6b63\u786e\u6027\u4e3a\u72b6\u6001 s \u5206\u914d\u4e00\u4e2a\u6807\u91cf\u5956\u52b1\u3002\u03b3 \u662f\u6298\u6263\u56e0\u5b50\uff0c\u51b3\u5b9a\u4e86\u672a\u6765\u5956\u52b1\u7684\u76f8\u5bf9\u91cd\u8981\u6027\u3002\u8be5 PRM \u7684\u8d1d\u5c14\u66fc\u65b9\u7a0b\u4e3a\uff1a<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0c1b7960d9b88b5c9a9d7c585e79bfbb.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-0c1b7960d9b88b5c9a9d7c585e79bfbb.png\"><\/a><\/p>\n<section>\u4e3a\u4e86\u5b66\u4e60\u8be5\u4ef7\u503c\u51fd\u6570\u7684 \u03b8\uff0c\u8fd9\u91cc\u5c06 TD \u635f\u5931\u51fd\u6570\u5b9a\u4e49\u6210\u5f53\u524d\u503c\u4e0e\u8d1d\u5c14\u66fc\u76ee\u6807\u4e4b\u95f4\u7684\u5e73\u65b9\u8bef\u5dee\uff1a<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-5c056e1b8aed3f5a7231784caf1c0e31.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-5c056e1b8aed3f5a7231784caf1c0e31.png\"><\/a><\/p>\n<ul>\n<li>\n<section>LLM \u7b56\u7565\u7684\u7b56\u7565\u8fed\u4ee3<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u5f97\u5230\u4e86 PRM \u4e4b\u540e\uff0c\u5c31\u53ef\u4ee5\u8bad\u7ec3 LLM \u7b56\u7565\u4ee5\u63d0\u5347\u5176\u63a8\u7406\u80fd\u529b\u4e86\u3002\u8fd9\u9700\u8981\u8d85\u8d8a\u4f20\u7edf\u76d1\u7763\u5b66\u4e60\u6846\u67b6\u7684\u65b9\u6cd5\u3002\u901a\u8fc7\u6574\u5408\u5728\u7ebf\u5f3a\u5316\u5b66\u4e60\u6765\u4f18\u5316\u63a8\u7406\u4efb\u52a1\uff0cPRM \u5728\u6b64\u8fc7\u7a0b\u4e2d\u53d1\u6325\u7740\u91cd\u8981\u4f5c\u7528\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u91cc\u6765\u770b\u770b Group Relative Policy Optimisation\uff08GRPO\uff09\uff0c\u5373\u5206\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\u3002<\/section>\n<section><\/section>\n<section>\u5047\u8bbe\u5bf9\u4e8e\u6bcf\u4e2a\u95ee\u9898 Q = q\uff0c\u7b56\u7565\u90fd\u4f1a\u751f\u6210\u63a8\u7406\u6b65\u9aa4 {o_1, o_2, . . . , o_G}\uff0c\u6bcf\u4e2a\u8f93\u51fa o_i \u7531\u591a\u4e2a\u6b65\u9aa4 {a_{i,1}, a_{i,2}, . . . , a_{i,Ki} } \u7ec4\u6210\uff0c\u5176\u4e2d K_i \u662f\u8f93\u51fa o_i \u4e2d\u7684\u63a8\u7406\u6b65\u9aa4\uff08\u6216 token\uff09\u603b\u6570\u3002\u73b0\u5728\uff0c\u53ef\u4ee5\u6784\u5efa\u901a\u8fc7 PRM \u5b66\u4e60 LLM \u7b56\u7565\u7684 GRPO \u4e86\uff0c\u5982\u4e0b\u6240\u793a\u3002<\/section>\n<section><\/section>\n<section>\u5bf9\u4e8e\u6bcf\u4e2a\u95ee\u9898 q\uff0cGRPO \u4ece\u65e7\u7b56\u7565 \u03c0_\u03b8_old \u91c7\u6837\u4e00\u7ec4\u8f93\u51fa {o_1, o_2, . . . , o_G}\uff0c\u76ee\u6807\u662f\u901a\u8fc7\u6700\u5927\u5316\u4ee5\u4e0b\u76ee\u6807\u6765\u4f18\u5316\u7b56\u7565\uff1a<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-aca675a6e0b9526ebbf581b18194abf0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-aca675a6e0b9526ebbf581b18194abf0.png\"><\/a><\/p>\n<section>GRPO \u4e0d\u4f1a\u5c06 KL \u60e9\u7f5a\u76f4\u63a5\u7eb3\u5165\u5956\u52b1\uff0c\u5176\u89c4\u8303\u7b56\u7565\u7684\u65b9\u5f0f\u662f\u5c06\u5f53\u524d\u7b56\u7565 \u03c0_\u03b8 \u548c\u53c2\u8003\u7b56\u7565 \u03c0_\u03b8_ref \u4e4b\u95f4\u7684 KL \u6563\u5ea6\u76f4\u63a5\u6dfb\u52a0\u5230\u635f\u5931\u51fd\u6570\u4e2d\u3002\u8fd9\u53ef\u786e\u4fdd\u66f4\u65b0\u540e\u7684\u7b56\u7565\u5728\u8bad\u7ec3\u671f\u95f4\u4e0d\u4f1a\u8fc7\u5ea6\u504f\u79bb\u53c2\u8003\u7b56\u7565\uff0c\u4ece\u800c\u6709\u52a9\u4e8e\u4fdd\u6301\u7a33\u5b9a\u6027\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u79cd GRPO \u5f62\u5f0f\u662f\u901a\u8fc7\u5229\u7528\u63a8\u7406\u6b65\u9aa4\u548c\u6700\u7ec8\u6b65\u9aa4\u4e2d\u7684\u5206\u7ec4\u76f8\u5bf9\u5956\u52b1\u6765\u4f18\u5316 LLM \u7b56\u7565\uff0c\u4e13\u95e8\u9002\u7528\u4e8e\u901a\u8fc7\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\u7684\u63a8\u7406\u4efb\u52a1\u3002\u5f52\u4e00\u5316\u7684\u4f18\u52bf\u51fd\u6570\uff08advantage function\uff09\u662f\u6839\u636e\u76f8\u5bf9\u6027\u80fd\u8ba1\u7b97\u7684\uff0c\u9f13\u52b1\u7b56\u7565\u504f\u5411\u5728\u4e00\u7ec4\u91c7\u6837\u8f93\u51fa\u4e2d\u8868\u73b0\u66f4\u597d\u7684\u8f93\u51fa\u3002\u6b64\u5916\uff0cKL \u6b63\u5219\u5316\u53ef\u786e\u4fdd\u66f4\u65b0\u540e\u7684\u7b56\u7565\u4e0e\u53c2\u8003\u7b56\u7565\u4fdd\u6301\u63a5\u8fd1\uff0c\u4ece\u800c\u63d0\u9ad8\u8bad\u7ec3\u7a33\u5b9a\u6027\u548c\u6548\u7387\u3002\u8be5\u6846\u67b6\u63d0\u4f9b\u4e86\u4e00\u79cd\u7a33\u5065\u7684\u65b9\u6cd5\uff0c\u53ef\u901a\u8fc7\u57fa\u4e8e PRM \u7684\u4f18\u5316\u6765\u6307\u5bfc LLM \u63a8\u7406\u3002<\/section>\n<section><\/section>\n<section>\u53e6\u5916\uff0c\u8fd8\u6709 token \u7ea7 DPO \u7b49\u4e0d\u4f7f\u7528 PRM \u7684\u66f4\u9ad8\u6548\u7684\u79bb\u7ebf\u65b9\u6cd5\uff0c\u8be6\u89c1\u76f8\u5173\u8bba\u6587\u300aToken-level direct preference optimization\u300b\u3002<\/section>\n<section><\/section>\n<section><strong>\u63a8\u7406\u65f6\u95f4\u8ba1\u7b97<\/strong><\/section>\n<section><\/section>\n<section>\u8bad\u7ec3\u5b8c\u6210\u540e\uff0cLLM \u7b56\u7565\u5fc5\u987b\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u9ad8\u6548\u5730\u751f\u6210\u8f93\u51fa\u3002LLM \u5e38\u7528\u7684\u65b9\u6cd5\u662f\u81ea\u56de\u5f52\uff0c\u5373\u6839\u636e\u4e4b\u524d\u7684 token \u9010\u4e00\u751f\u6210\u65b0 \u00a0token\u3002\u4f46\u662f\uff0c\u5bf9\u4e8e\u63a8\u7406\u4efb\u52a1\uff0c\u8fd8\u5fc5\u9700\u66f4\u590d\u6742\u7684\u89e3\u7801\u6280\u672f\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u5728\u6548\u7387\u548c\u6548\u679c\u4e4b\u95f4\u53d6\u5f97\u5e73\u8861\uff0c\u6709\u7814\u7a76\u53d1\u73b0\uff0c\u6ce2\u675f\u641c\u7d22\u7b49\u66f4\u7075\u6d3b\u7684\u65b9\u6cd5\u6709\u5229\u4e8e\u63a8\u7406\u4efb\u52a1\u3002\u5bf9\u4e8e\u66f4\u590d\u6742\u7684\u63a8\u7406\u4efb\u52a1\uff0c\u53ef\u4ee5\u4f7f\u7528 MCTS \u7b49\u5411\u524d\u770b\u7684\u6a21\u578b\u3002<\/section>\n<section><\/section>\n<section>MCTS \u53ef\u6a21\u62df\u591a\u6761\u63a8\u7406\u8def\u5f84\uff0c\u5e76\u6839\u636e\u5956\u52b1\u7cfb\u7edf\u5bf9\u5176\u8fdb\u884c\u8bc4\u4f30\uff0c\u9009\u62e9\u9884\u671f\u5956\u52b1\u6700\u9ad8\u7684\u8def\u5f84\u3002\u8fd9\u5141\u8bb8\u6a21\u578b\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u63a2\u7d22\u66f4\u5927\u8303\u56f4\u7684\u53ef\u80fd\u6027\uff0c\u4ece\u800c\u589e\u52a0\u5176\u83b7\u5f97\u6700\u4f18\u89e3\u7684\u673a\u4f1a\u3002\u4f7f\u7528 MDP\uff0c\u53ef\u4ee5\u4ece\u5f62\u5f0f\u4e0a\u5b9a\u4e49\u5176\u63a8\u7406\u8fc7\u7a0b\u7ed3\u6784\u3002<\/section>\n<section><\/section>\n<section>\u5b9a\u4e49 2\uff1a<strong>\u539f\u751f\u601d\u7ef4\u94fe\uff08NCoT\uff09<\/strong>\u662f\u6307\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u56fa\u6709\u7684\u63a8\u7406\u80fd\u529b\uff0c\u8fd9\u8ba9\u5176\u65e0\u9700\u5916\u90e8\u63d0\u793a\u8bcd\u4fbf\u53ef\u4ee5\u81ea\u52a8\u6267\u884c\u9010\u6b65\u5f0f\u7684\u7ed3\u6784\u5316\u63a8\u7406\u3002\u8be5\u80fd\u529b\u53ef\u4ee5\u8868\u8ff0\u4e3a\u4e00\u4e2a\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b\uff08MDP\uff09(S, A, \u03c0, R)\uff0c\u5176\u4e2d<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>S \u662f\u72b6\u6001\u7a7a\u95f4\uff0c\u8868\u793a\u751f\u6210\u5230\u7ed9\u5b9a\u4f4d\u7f6e\u5904\u7684 token \u5e8f\u5217\u6216\u63a8\u7406\u6b65\u9aa4\uff1b<\/section>\n<\/li>\n<li>\n<section>A \u662f\u52a8\u4f5c\u7a7a\u95f4\uff0c\u7531\u6f5c\u5728\u63a8\u7406\u6b65\u9aa4 R_t \u6216\u6700\u7ec8\u7b54\u6848 A \u7ec4\u6210\uff1b<\/section>\n<\/li>\n<li>\n<section>\u03c0_LLM (a_t | s_t) \u662f\u63a7\u5236\u52a8\u4f5c\u9009\u62e9\u7684\u7b56\u7565\uff08\u4e5f\u662f LLM\uff09\uff0c\u5176\u53ef\u6839\u636e\u5f53\u524d\u72b6\u6001 s_t \u786e\u5b9a\u4e0b\u4e00\u4e2a\u63a8\u7406\u6b65\u9aa4\u6216\u6700\u7ec8\u7b54\u6848\uff1b<\/section>\n<\/li>\n<li>\n<section>R (s_t a_t) \u662f\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff08PRM\uff0c\u5176\u4f5c\u7528\u662f\u6839\u636e\u6240\u9009\u52a8\u4f5c a_t \u7684\u8d28\u91cf\u548c\u76f8\u5173\u6027\u5206\u914d\u5956\u52b1 r_t\uff0c\u4ee5\u5f15\u5bfc\u63a8\u7406\u8fc7\u7a0b\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u8be5\u6a21\u578b\u65e2\u53ef\u4ee5\u901a\u8fc7\u5c55\u5f00 MDP \u6765\u9075\u5faa\u987a\u5e8f\u63a8\u7406\u8def\u5f84\uff0c\u4e5f\u53ef\u4ee5\u901a\u8fc7\u5728\u6bcf\u4e2a\u72b6\u6001\u4e0b\u91c7\u6837\u4e0d\u540c\u7684\u63a8\u7406\u6b65\u9aa4\u6765\u63a2\u7d22\u591a\u4e2a\u8f68\u8ff9\uff0c\u5f62\u6210\u6811\u72b6\u7ed3\u6784\uff08\u56fe 5\uff09\u3002\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b R \u63d0\u4f9b\u4e86\u5bf9\u8be5\u7a7a\u95f4\u7684\u5f15\u5bfc\u641c\u7d22\uff0c\u5176\u63a7\u5236\u63a8\u7406\u8f68\u8ff9\u7684\u65b9\u5f0f\u662f\u652f\u6301\u80fd\u5f97\u5230\u66f4\u6709\u610f\u4e49\u6216\u66f4\u6b63\u786e\u7684\u63a8\u7406\u6b65\u9aa4\u7684\u52a8\u4f5c\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-95abdddb0de17e671ab9fc469cc9fad7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-95abdddb0de17e671ab9fc469cc9fad7.png\"><\/a><\/p>\n<section><em><sup>\u56fe 5\uff1a\u501f\u52a9 PRM\uff0cLLM \u53ef\u4ee5\u901a\u8fc7\u4e09\u79cd\u65b9\u6cd5\u6267\u884c\u975e\u81ea\u56de\u5f52\u63a8\u7406\uff1a1) \u91c7\u6837\u591a\u6761\u63a8\u7406\u8f68\u8ff9\uff0c2) \u5bf9\u6f5c\u5728\u63a8\u7406\u8def\u5f84\u7684\u6811\u7ed3\u6784\u8fdb\u884c\u8499\u7279\u5361\u6d1b\u641c\u7d22\uff0c3) \u7ed3\u5408\u524d\u4e24\u79cd\u65b9\u6cd5\u6765\u589e\u5f3a\u63a8\u7406\u7684\u7075\u6d3b\u6027\u548c\u7a33\u5065\u6027\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u76f8\u5173\u6587\u732e<\/strong><\/section>\n<section><\/section>\n<section>\u4e0b\u9762\u5c06\u56de\u987e\u5e76\u8ba8\u8bba\u76f8\u5173\u9886\u57df\u7684\u51e0\u7bc7\u5173\u952e\u8bba\u6587\uff0c\u4ecb\u7ecd\u5b83\u4eec\u7684\u8d21\u732e\u548c\u5c40\u9650\u6027\u3002\u56fe 6 \u63cf\u8ff0\u4e86\u8fd9\u4e9b\u7814\u7a76\u4e0e\u66f4\u5e7f\u6cdb\u7684\u7814\u7a76\u56fe\u666f\u4e4b\u95f4\u7684\u8054\u7cfb\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-c915af0478d1e4d5dac60a2c6f4c4d4d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-c915af0478d1e4d5dac60a2c6f4c4d4d.png\"><\/a><\/p>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 6\uff1aLLM \u539f\u751f\u601d\u7ef4\u94fe\u7814\u7a76\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u63a8\u7406\u65f6\u95f4\u8ba1\u7b97<\/strong><\/section>\n<section><\/section>\n<section>\u6709\u51e0\u7bc7\u8bba\u6587\u91cd\u70b9\u5173\u6ce8\u4f18\u5316 LLM \u5728\u63a8\u7406\u65f6\u95f4\u7684\u63a8\u7406\uff0c\u6bd4\u5982\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u300aAlphazero-like tree-search can guide large language model decoding and training\u300b\u63d0\u51fa\u4e86\u4e00\u79cd\u5c06\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22\uff08MCTS\uff09\u4e0e LLM \u89e3\u7801\u6574\u5408\u8d77\u6765\u7684\u65b9\u6cd5\uff0c\u7814\u7a76\u8bc1\u660e\u8fd9\u79cd\u7ec4\u5408\u80fd\u591f\u6709\u6548\u5730\u5f15\u5bfc\u63a8\u7406\uff0c\u5c24\u5176\u662f\u590d\u6742\u7684\u591a\u6b65\u9aa4\u4efb\u52a1\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aScaling llm test-time compute optimally can be more effective than scaling model parameters\u300b\u5f3a\u8c03\u4e86\u4f18\u5316\u6d4b\u8bd5\u65f6\u95f4\u8ba1\u7b97\u7684\u91cd\u8981\u6027\uff0c\u5176\u901a\u8fc7\u5b9e\u8bc1\u7814\u7a76\u8868\u660e\uff0c\u63a8\u7406\u65f6\u95f4\u63a8\u7406\u589e\u5f3a\u901a\u5e38\u53ef\u4ee5\u6bd4\u7b80\u5355\u5730\u6269\u5c55\u6a21\u578b\u53c2\u6570\u4ea7\u751f\u66f4\u5b9e\u8d28\u6027\u7684\u6539\u8fdb\u3002\u8fd9\u53cd\u6620\u4e86\u4eba\u4eec\u65e5\u76ca\u589e\u957f\u7684\u7406\u89e3\uff0c\u5373\u53ef\u4ee5\u5229\u7528\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u66f4\u591a\u8ba1\u7b97\u6765\u5b9e\u73b0\u66f4\u9ad8\u8d28\u91cf\u7684\u63a8\u7406\uff0c\u800c\u4e0d\u5fc5\u589e\u52a0\u6a21\u578b\u7684\u5927\u5c0f\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aThink before you speak: Training language models with pause tokens\u300b\u63d0\u51fa\u4e86\u53e6\u4e00\u79cd\u65b9\u6cd5\uff1a\u5728\u63a8\u7406\u9636\u6bb5\u4f7f\u7528\u6682\u505c token \u5f3a\u8feb\u6a21\u578b\u6682\u505c\u5e76\u300c\u601d\u8003\u300d\u3002\u8be5\u65b9\u6cd5\u4f1a\u5f15\u5165\u4e00\u4e2a\u9690\u5f0f\u7684\u63a8\u7406\u6a21\u578b\uff0c\u4ece\u800c\u9f13\u52b1 LLM \u5bf9\u4fe1\u606f\u8fdb\u884c\u5206\u5757\uff0c\u6a21\u4eff\u4eba\u7c7b\u7684\u601d\u8003\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u9a8c\u8bc1\u5668\u6a21\u578b<\/strong><\/section>\n<section><\/section>\n<section>\u9a8c\u8bc1\u5668\u6a21\u578b\uff08\u7ed3\u679c\u5956\u52b1\u6a21\u578b\u548c\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff09\u5df2\u6210\u4e3a\u63d0\u9ad8 LLM \u63a8\u7406\u53ef\u9760\u6027\u7684\u91cd\u8981\u7814\u7a76\u9886\u57df\u3002<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u300aTraining verifiers to solve math word problems\u300b\u6700\u65e9\u5c1d\u8bd5\u5728\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u4e2d\u4f7f\u7528\u9a8c\u8bc1\u5668\uff08\u4ec5\u7ed3\u679c\u5956\u52b1\uff09\uff0c\u4e3a\u540e\u7eed\u7814\u7a76\u5960\u5b9a\u4e86\u57fa\u7840\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aSolving math word problems with process-and outcome-based feedback\u300b\u6269\u5c55\u4e86\u9a8c\u8bc1\u5668\u7684\u6982\u5ff5\uff0c\u6574\u5408\u4e86\u57fa\u4e8e\u8fc7\u7a0b\u7684\u63a8\u7406\u673a\u5236\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aLet\u2019s verify step by step\u300b \u7814\u7a76\u4e86\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff0c\u4ea6\u53ef\u53c2\u9605\u673a\u5668\u4e4b\u5fc3\u62a5\u9053\u300a<a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"http:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650879157&amp;idx=1&amp;sn=eac1b887a62133667d5987ea995b49fe&amp;chksm=84e4f0cbb39379dd724f6e21344eb101980458c3dbcca5ce87f5794e24810731c31883fe31de&amp;scene=21#wechat_redirect\" target=\"_blank\">OpenAI \u8981\u4e3a GPT-4 \u89e3\u51b3\u6570\u5b66\u95ee\u9898\u4e86\uff1a\u5956\u52b1\u6a21\u578b\u6307\u9519\uff0c\u89e3\u9898\u6c34\u5e73\u8fbe\u5230\u65b0\u9ad8\u5ea6<\/a>\u300b\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aMaking large language models better reasoners with step-aware verifier\u300b\u5c06\u9a8c\u8bc1\u5668\u6a21\u578b\u4e0e\u5927\u591a\u6570\u6295\u7968\u673a\u5236\u7ec4\u5408\u5230\u4e86\u4e00\u8d77\uff0c\u4ee5\u5728\u63a8\u7406\u4efb\u52a1\u4e2d\u5f97\u5230\u66f4\u53ef\u9760\u7684\u8f93\u51fa\u3002\u4e3a\u4e86\u589e\u5f3a\u9a8c\u8bc1\u8fc7\u7a0b\u7684\u7a33\u5065\u6027\uff0c\u8be5\u65b9\u6cd5\u4f1a\u4ea4\u53c9\u68c0\u67e5\u591a\u6761\u63a8\u7406\u8def\u5f84\u5e76\u8fc7\u6ee4\u6389\u4e0d\u6b63\u786e\u7684\u6b65\u9aa4\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u83b7\u53d6\u7528\u4e8e\u63a8\u7406\u4efb\u52a1\u7684\u6570\u636e<\/strong><\/section>\n<section><\/section>\n<section>\u83b7\u53d6\u63a8\u7406\u6570\u636e\u4e00\u76f4\u90fd\u662f\u4e00\u5927\u7814\u7a76\u91cd\u5fc3\uff0c\u7814\u7a76\u5305\u62ec\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u300aStar: Bootstrapping reasoning with reasoning\u300b\u63a2\u7d22\u4e86\u81ea\u52a8\u83b7\u53d6\u4e0e\u63a8\u7406\u6b65\u9aa4\u76f8\u5173\u7684\u6570\u636e\u7684\u65b9\u6cd5\u3002STaR \u63d0\u51fa\u4e86\u4e00\u79cd\u81ea\u5b66\u4e60\u8303\u5f0f\uff0c\u8ba9\u6a21\u578b\u53ef\u901a\u8fc7\u751f\u6210\u548c\u6279\u8bc4\u81ea\u5df1\u7684\u6b65\u9aa4\u6765\u63d0\u9ad8\u5176\u63a8\u7406\u80fd\u529b\uff0c\u4ece\u800c\u4ea7\u751f\u66f4\u53ef\u9760\u7684\u4e2d\u95f4\u6b65\u9aa4\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aMath-shepherd: Verify and reinforce llms step-by-step without human annotations\u300b\u8fdb\u4e00\u6b65\u63a8\u8fdb\u4e86\u8be5\u65b9\u6cd5\uff0c\u8868\u660e\u65e0\u9700\u6210\u672c\u9ad8\u6602\u7684\u6807\u6ce8\u4e5f\u80fd\u9010\u6b65\u8bad\u7ec3 LLM\uff0c\u5176\u4e3a\u63a8\u7406\u6570\u636e\u95ee\u9898\u63d0\u4f9b\u66f4\u5177\u53ef\u6269\u5c55\u6027\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aMulti-step problem solving through a verifier: An empirical analysis on model-induced process supervision\u300b\u5f3a\u8c03\u4e86\u5b9e\u9645\u6570\u636e\u91c7\u96c6\u5bf9\u4e8e\u63a8\u7406\u4efb\u52a1\u7684\u91cd\u8981\u6027\uff0c\u7279\u522b\u662f\u5bf9\u4e8e\u7f16\u7a0b\u95ee\u9898\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aAlphazero-like tree-search can guide large language model decoding and training\u300b\u4f7f\u7528\u4e86 MCTS \u6765\u83b7\u53d6\u6570\u636e\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aImprove mathematical reasoning in language models by automated process supervision\u300b\u5219\u5728\u6b64\u57fa\u7840\u4e0a\u4f7f\u7528\u4e86\u7ebf\u6027\u641c\u7d22\u6765\u63d0\u5347\u6548\u7387\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u7406\u89e3\u548c\u7cfb\u7edf\u7ea7\u63d0\u5347<\/strong><\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u4e5f\u6709\u4e0d\u5c11\u7814\u7a76\u8005\u81f4\u529b\u4e8e\u7406\u89e3 LLM \u9010\u6b65\u63a8\u7406\u80cc\u540e\u7684\u673a\u5236\uff0c\u5982\u8bba\u6587\u300aWhy can large language models generate correct chain-of-thoughts?\u300b\u548c\u300aWhy think step by step? reasoning emerges from the locality of experience\u300b\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aLlama: Open and efficient foundation language models\u300b\u5219\u662f\u4ece\u56fe\u6a21\u578b\u89d2\u5ea6\u6765\u5206\u6790\u601d\u7ef4\u94fe\u673a\u5236\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aWhy think step by step? reasoning emerges from the locality of experience\u300b\u63a2\u7d22\u4e86\u63a8\u7406\u4f5c\u4e3a LLM \u7684\u4e00\u79cd\u81ea\u7136\u80fd\u529b\u7684\u5185\u5728\u539f\u56e0\u3002\u5176\u8ba4\u4e3a\u63a8\u7406\u662f\u8bed\u8a00\u6a21\u578b\u5904\u7406\u672c\u5730\u5316\u7ecf\u9a8c\u548c\u77e5\u8bc6\u7684\u4e00\u4e2a\u526f\u4ea7\u7269\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aCritique ability of large language models\u300b\u5bf9 LLM \u6279\u8bc4\u81ea\u5df1\u7684\u80fd\u529b\u8fdb\u884c\u5b9e\u8bc1\u8bc4\u4f30\uff0c\u7ed3\u679c\u8868\u660e\u81ea\u6211\u6279\u8bc4\u5f80\u5f80\u5f88\u6709\u9650\uff0c\u5e76\u4e14\u901a\u5e38\u53ea\u6709\u5f53\u6a21\u578b\u8db3\u591f\u5927\u65f6\u624d\u4f1a\u6d8c\u73b0\u8fd9\u79cd\u80fd\u529b\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u300aPangu-agent: A fine-tunable generalist agent with structured reasoning\u300b\u4ece\u7cfb\u7edf\u89d2\u5ea6\u63d0\u51fa\u4e86\u8d85\u8d8a\u4f20\u7edf\u6a21\u578b\u7684\u7ed3\u6784\u5316\u63a8\u7406\u673a\u5236\uff0c\u7c7b\u4f3c\u4e8e OpenAI \u03bf1 \u6a21\u578b\u3002\u8fd9\u9879\u7814\u7a76\u53cd\u6620\u4e86\u5411\u66f4\u901a\u7528\u7684\u63a8\u7406\u667a\u80fd\u4f53\u7684\u8f6c\u53d8\uff0c\u8fd9\u4e9b\u667a\u80fd\u4f53\u80fd\u4ee5\u66f4\u9ad8\u7684\u7cbe\u5ea6\u548c\u7075\u6d3b\u6027\u5904\u7406\u66f4\u5e7f\u6cdb\u7684\u4efb\u52a1\uff0c\u63cf\u7ed8\u4e86\u4e0b\u4e00\u4ee3\u63a8\u7406\u6a21\u578b\u7684\u613f\u666f\u3002<\/section>\n<\/li>\n<\/ul>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-10-11-4\" target=\"_blank\">\u4e00\u6587\u770b\u61c2LLM\u63a8\u7406\uff0cUCL\u6c6a\u519b\u6559\u6388\u89e3\u8bfbOpenAI \u03bf1\u7684\u76f8\u5173\u65b9\u6cd5<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u4e00\u6587\u770b\u61c2LLM\u63a8\u7406\uff0cU [&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":[70],"class_list":["post-36024","post","type-post","status-publish","format-standard","hentry","category-news","tag-agent"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u4e00\u6587\u770b\u61c2LLM\u63a8\u7406\uff0cUCL\u6c6a\u519b\u6559\u6388\u89e3\u8bfbOpenAI \u03bf1\u7684\u76f8\u5173\u65b9\u6cd5 - \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=36024\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=36024\",\"url\":\"https:\/\/17aitech.com\/?p=36024\",\"name\":\"\u4e00\u6587\u770b\u61c2LLM\u63a8\u7406\uff0cUCL\u6c6a\u519b\u6559\u6388\u89e3\u8bfbOpenAI \u03bf1\u7684\u76f8\u5173\u65b9\u6cd5 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=36024#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=36024#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ca83ca16a797821b782c07f24651568a.png\",\"datePublished\":\"2025-01-04T09:00:20+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=36024#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=36024\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=36024#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ca83ca16a797821b782c07f24651568a.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ca83ca16a797821b782c07f24651568a.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=36024#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4e00\u6587\u770b\u61c2LLM\u63a8\u7406\uff0cUCL\u6c6a\u519b\u6559\u6388\u89e3\u8bfbOpenAI \u03bf1\u7684\u76f8\u5173\u65b9\u6cd5\"}]},{\"@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":"\u4e00\u6587\u770b\u61c2LLM\u63a8\u7406\uff0cUCL\u6c6a\u519b\u6559\u6388\u89e3\u8bfbOpenAI \u03bf1\u7684\u76f8\u5173\u65b9\u6cd5 - \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=36024","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=36024","url":"https:\/\/17aitech.com\/?p=36024","name":"\u4e00\u6587\u770b\u61c2LLM\u63a8\u7406\uff0cUCL\u6c6a\u519b\u6559\u6388\u89e3\u8bfbOpenAI \u03bf1\u7684\u76f8\u5173\u65b9\u6cd5 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=36024#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=36024#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ca83ca16a797821b782c07f24651568a.png","datePublished":"2025-01-04T09:00:20+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=36024#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=36024"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=36024#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ca83ca16a797821b782c07f24651568a.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/10\/frc-ca83ca16a797821b782c07f24651568a.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=36024#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u4e00\u6587\u770b\u61c2LLM\u63a8\u7406\uff0cUCL\u6c6a\u519b\u6559\u6388\u89e3\u8bfbOpenAI \u03bf1\u7684\u76f8\u5173\u65b9\u6cd5"}]},{"@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\/36024","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=36024"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/36024\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=36024"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=36024"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=36024"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}