{"id":40181,"date":"2025-04-17T05:02:42","date_gmt":"2025-04-16T21:02:42","guid":{"rendered":"https:\/\/17aitech.com\/?p=40181"},"modified":"2025-04-17T05:02:42","modified_gmt":"2025-04-16T21:02:42","slug":"%e8%bf%88%e5%90%91%e6%8e%a8%e7%90%86%e6%97%b6%e4%bb%a3%ef%bc%9a900%e7%af%87%e5%8f%82%e8%80%83%e6%96%87%e7%8c%ae%e6%8f%ad%e7%a4%ba%e9%95%bf%e9%93%be%e6%80%9d%e7%bb%b4%e7%9a%84%e5%89%8d%e4%b8%96","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=40181","title":{"rendered":"\u8fc8\u5411\u63a8\u7406\u65f6\u4ee3\uff1a900+\u7bc7\u53c2\u8003\u6587\u732e\u63ed\u793a\u957f\u94fe\u601d\u7ef4\u7684\u524d\u4e16\u4eca\u751f\uff0c\u6700\u5168\u7efc\u8ff0\u6765\u4e86"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-04-16\" target=\"_blank\">\u8fc8\u5411\u63a8\u7406\u65f6\u4ee3\uff1a900+\u7bc7\u53c2\u8003\u6587\u732e\u63ed\u793a\u957f\u94fe\u601d\u7ef4\u7684\u524d\u4e16\u4eca\u751f\uff0c\u6700\u5168\u7efc\u8ff0\u6765\u4e86<\/a><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-fa364dae7f3411d036bb9f46c3f4284d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-fa364dae7f3411d036bb9f46c3f4284d.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-95304eb296d0f5cd0e4b215f1e1f2851.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-95304eb296d0f5cd0e4b215f1e1f2851.png\"><\/a><\/section>\n<ul>\n<li>\n<p>\u8bba\u6587\u6807\u9898\uff1aTowards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Models<\/p>\n<\/li>\n<li>\n<p>\u8bba\u6587\u4e3b\u9875\uff1ahttps:\/\/long-cot.github.io\/<\/p>\n<\/li>\n<li>\n<p>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2503.09567<\/p>\n<\/li>\n<li>\n<p>Github \u4ed3\u5e93\uff1ahttps:\/\/github.com\/LightChen233\/Awesome-Long-Chain-of-Thought-Reasoning<\/p>\n<\/li>\n<\/ul>\n<p><strong>1. \u8fc8\u5411\u63a8\u7406\u65f6\u4ee3 \u00a0900 + \u7bc7\u6587\u732e\u63ed\u793a\u957f\u94fe\u601d\u7ef4\uff08Long CoT\uff09\u7684\u524d\u4e16\u4eca\u751f<\/strong><\/p>\n<p>\u8fd1\u5e74\u6765\uff0c\u5927\u6a21\u578b\uff08Large Language Models, LLMs\uff09\u5728\u6570\u5b66\u3001\u7f16\u7a0b\u7b49\u590d\u6742\u4efb\u52a1\u4e0a\u53d6\u5f97\u7a81\u7834\uff0cOpenAI-o1\u3001DeepSeek-R1 \u7b49\u63a8\u7406\u5927\u6a21\u578b\uff08Reasoning Large Language Models\uff0cRLLMs\uff09\u8868\u73b0\u5c24\u4e3a\u4eae\u773c\u3002<strong>\u4f46\u5b83\u4eec\u4e3a\u4f55\u5982\u6b64\u5f3a\u5927\u5462\uff1f<\/strong> \u6838\u5fc3\u5728\u4e8e<strong>\u957f\u94fe\u601d\u7ef4\uff08Long Chain-of-Thought, Long CoT\uff09<\/strong>\u2014\u2014 \u901a\u8fc7\u66f4\u6df1\u5165\u7684\u63a8\u7406\u3001\u66f4\u5e7f\u6cdb\u7684\u63a2\u7d22\uff0c\u4ee5\u53ca\u66f4\u5f3a\u7684\u81ea\u6211\u53cd\u601d\uff0c\u4f7f\u5927\u6a21\u578b\u80fd\u591f\u9ad8\u6548\u89e3\u51b3\u590d\u6742\u95ee\u9898\u3002<\/p>\n<p><strong>\u7136\u800c\uff0c\u957f\u601d\u7ef4\u94fe\u4ecd\u7136\u7f3a\u4e4f\u7cfb\u7edf\u6027\u7814\u7a76\uff0c\u751a\u81f3\u9762\u4e34\u8bf8\u591a\u672a\u89e3\u4e4b\u8c1c\uff1a<\/strong><\/p>\n<ul>\n<li>\n<p>\u957f\u94fe\u601d\u7ef4\u7a76\u7adf\u5982\u4f55\u63d0\u5347\u6a21\u578b\u80fd\u529b\uff1f\u4e0e\u77ed\u94fe\u601d\u7ef4\uff08Short CoT\uff09\u7684\u6838\u5fc3\u533a\u522b\u662f\u4ec0\u4e48\uff1f<\/p>\n<\/li>\n<li>\n<p>\u00a0\u201c\u63a8\u7406\u6269\u5c55\u6027\u201d\uff08Test-Time Scaling\uff09\u662f\u5173\u952e\u7a81\u7834\uff0c\u8fd8\u662f\u5e26\u6765\u4e86 \u201c\u8fc7\u5ea6\u601d\u8003\u201d\uff08Overthinking\uff09\uff1f<\/p>\n<\/li>\n<li>\n<p>\u957f\u601d\u7ef4\u94fe\u6709\u54ea\u4e9b\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\uff1f<\/p>\n<\/li>\n<\/ul>\n<p><strong>\u586b\u8865\u7814\u7a76\u7a7a\u767d\uff0c\u5efa\u7acb\u7edf\u4e00\u89c6\u89d2\uff1a\u6700\u5168\u957f\u601d\u7ef4\u94fe\u7efc\u8ff0\uff0c\u6765\u4e86\uff01<\/strong><\/p>\n<ul>\n<li>\n<p>\u5728\u8fd9\u7bc7\u7efc\u8ff0\u4e2d\uff0c\u8be5\u8bba\u6587\u9996\u5148\u5398\u6e05\u957f\u601d\u7ef4\u94fe\uff08Long CoT\uff09\u4e0e\u77ed\u601d\u7ef4\u94fe\uff08Short CoT\uff09\u7684\u672c\u8d28\u533a\u522b\uff0c\u63d0\u51fa\u65b0\u7684\u63a8\u7406\u8303\u5f0f\u5206\u7c7b\u6846\u67b6\uff0c\u660e\u786e\u5b83\u4eec\u5728\u63a8\u7406\u6df1\u5ea6\u3001\u63a2\u7d22\u8303\u56f4\u548c\u53cd\u9988\u673a\u5236\u4e0a\u7684\u5dee\u5f02\uff0c\u5e76\u4e3a\u4e0d\u540c\u4efb\u52a1\u7684\u63a8\u7406\u9700\u6c42\u63d0\u4f9b\u7cfb\u7edf\u6027\u8ba4\u77e5\u3002<\/p>\n<\/li>\n<li>\n<p>\u968f\u540e\uff0c\u8be5\u8bba\u6587\u6df1\u5165\u62c6\u89e3 \u957f\u601d\u7ef4\u94fe\u7684\u4e09\u5927\u5173\u952e\u7279\u6027\uff0c\u5373<strong>\u6df1\u5ea6\u63a8\u7406\u3001\u5e7f\u6cdb\u63a2\u7d22\u548c\u53ef\u884c\u6027\u53cd\u601d<\/strong>\uff0c\u63ed\u793a\u5176\u5982\u4f55\u5728\u63d0\u5347\u63a8\u7406\u80fd\u529b\u7684\u540c\u65f6\uff0c\u4e5f\u5e26\u6765\u4e86\u8ba1\u7b97\u5f00\u9500\u3001\u5197\u4f59\u63a8\u7406\u7b49\u6311\u6218\u3002<\/p>\n<\/li>\n<li>\n<p>\u8be5\u8bba\u6587\u8fdb\u4e00\u6b65\u63a2\u8ba8 \u957f\u601d\u7ef4\u94fe\u76f8\u5173\u7684\u6838\u5fc3\u63a8\u7406\u73b0\u8c61\uff0c\u5982<strong>\u8fc7\u5ea6\u601d\u8003\uff08Overthinking\uff09\u3001\u63a8\u7406\u6269\u5c55\u6027\uff08Test-Time Scaling\uff09\u4ee5\u53ca AI \u7684 \u201c\u987f\u609f\u65f6\u523b\u201d\uff08Aha Moment\uff09<\/strong>\uff0c\u5206\u6790\u8fd9\u4e9b\u73b0\u8c61\u5982\u4f55\u5f71\u54cd\u6a21\u578b\u7684\u63a8\u7406\u6548\u7387\u548c\u7b54\u6848\u8d28\u91cf\uff0c\u5e76\u8ba8\u8bba\u5176\u53ef\u80fd\u7684\u4f18\u5316\u65b9\u6848\u3002<\/p>\n<\/li>\n<li>\n<p>\u6700\u540e\uff0c\u8be5\u8bba\u6587\u5c55\u671b\u957f\u601d\u7ef4\u94fe\u7684\u672a\u6765\u53d1\u5c55\u65b9\u5411\uff0c\u5305\u62ec<strong>\u591a\u6a21\u6001\u63a8\u7406\u7684\u878d\u5408\u3001\u63a8\u7406\u6548\u7387\u7684\u4f18\u5316\u4ee5\u53ca\u77e5\u8bc6\u589e\u5f3a\u6280\u672f\u7b49\u591a\u4e2a\u672a\u6765\u7814\u7a76\u65b9\u5411\u5c55\u671b<\/strong>\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63d0\u5347\u63a8\u7406\u5927\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u548c\u5b9e\u9645\u5e94\u7528\u4ef7\u503c\u3002\u901a\u8fc7\u8fd9\u7bc7\u7efc\u8ff0\uff0c\u8be5\u8bba\u6587\u5e0c\u671b\u5efa\u7acb\u7edf\u4e00\u7684\u7814\u7a76\u89c6\u89d2\uff0c\u63a8\u52a8\u957f\u601d\u7ef4\u94fe\u7814\u7a76\u8d70\u5411\u66f4\u6df1\u5165\u7684\u7406\u8bba\u63a2\u7d22\u4e0e\u66f4\u9ad8\u6548\u7684\u5b9e\u8df5\u843d\u5730\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>2. \u957f\u601d\u7ef4\u94fe vs. \u77ed\u601d\u7ef4\u94fe<\/strong><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dcd528c43e1ca5697243b5a7f8b67e21.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-dcd528c43e1ca5697243b5a7f8b67e21.png\"><\/a><\/section>\n<p><sup>\u56fe 2\uff1a\u957f\u601d\u7ef4\u94fe\u548c\u601d\u7ef4\u94fe\u4e4b\u95f4\u7684\u5dee\u5f02\u5177\u6709\u4e09\u4e2a\u5173\u952e\u7279\u5f81\uff1a\u6df1\u5ea6\u7684\u63a8\u7406\uff0c\u9002\u5f53\u7684\u53cd\u601d\u548c\u5e7f\u6cdb\u7684\u63a2\u7d22\u3002\u6b64\u5916\uff0c\u957f\u601d\u7ef4\u94fe\u6574\u5408\u4e86\u6240\u6709\u8fd9\u4e9b\u7279\u5f81\u4ee5\u5b9e\u73b0\u66f4\u597d\u7684\u903b\u8f91\u6548\u679c\u3002<\/sup><\/p>\n<p>\u5728\u5927\u6a21\u578b\u63a8\u7406\u6f14\u5316\u7684\u8fc7\u7a0b\u4e2d\uff0c\u5982\u56fe 2 \u6240\u793a\uff0c<strong>\u957f\u601d\u7ef4\u94fe\uff08Long CoT\uff09<\/strong>\u4e0e<strong>\u77ed\u601d\u7ef4\u94fe\uff08Short CoT\uff09<\/strong>\u4ee3\u8868\u4e86\u4e24\u79cd\u622a\u7136\u4e0d\u540c\u7684\u63a8\u7406\u8303\u5f0f\u3002\u77ed\u601d\u7ef4\u94fe\u4ee5<strong>\u6d45\u5c42\u3001\u7ebf\u6027<\/strong>\u7684\u63a8\u7406\u65b9\u5f0f\u5feb\u901f\u5f97\u51fa\u7ed3\u8bba\uff0c\u903b\u8f91\u8def\u5f84\u77ed\uff0c\u63a2\u7d22\u6027\u4f4e\uff0c\u9002\u7528\u4e8e\u7ed3\u6784\u6e05\u6670\u3001\u89e3\u7b54\u660e\u786e\u7684\u95ee\u9898\u3002\u800c\u957f\u601d\u7ef4\u94fe\u5219\u5f3a\u8c03<strong>\u6df1\u5ea6\u63a8\u7406\u3001\u5e7f\u6cdb\u63a2\u7d22\u548c\u53ef\u884c\u6027\u53cd\u601d<\/strong>\uff0c\u5141\u8bb8\u6a21\u578b\u5728\u66f4\u590d\u6742\u7684\u903b\u8f91\u7f51\u7edc\u4e2d\u5c55\u5f00\u6df1\u5165\u5206\u6790\uff0c\u53d1\u73b0\u9690\u85cf\u5173\u7cfb\uff0c\u5e76\u4f18\u5316\u63a8\u7406\u8def\u5f84\u3002<\/p>\n<p>\u77ed\u601d\u7ef4\u94fe\u7684\u63a8\u7406\u65b9\u5f0f\u903b\u8f91\u8282\u70b9\u6570\u91cf\u6709\u9650\uff0c\u4e3b\u8981\u4f9d\u8d56\u987a\u5e8f\u63a8\u7406\uff0c\u4e14\u51e0\u4e4e\u4e0d\u4f1a\u56de\u6eaf\u6216\u68c0\u9a8c\u5148\u524d\u7684\u63a8\u7406\u8def\u5f84\u3002\u5728\u6570\u5b66\u3001\u7f16\u7a0b\u7b49\u9700\u8981\u903b\u8f91\u63a8\u5bfc\u7684\u4efb\u52a1\u4e2d\uff0c\u77ed\u601d\u7ef4\u94fe\u5f80\u5f80\u96be\u4ee5\u5e94\u5bf9\u590d\u6742\u95ee\u9898\uff0c\u800c\u957f\u601d\u7ef4\u94fe\u5219\u53ef\u4ee5\u901a\u8fc7\u6269\u5c55\u63a8\u7406\u6df1\u5ea6\u548c\u589e\u52a0\u63a2\u7d22\u7a7a\u95f4\uff0c\u63d0\u9ad8\u89e3\u9898\u80fd\u529b\u3002\u5f62\u5f0f\u4e0a\uff0c\u957f\u601d\u7ef4\u94fe\u653e\u5bbd\u4e86\u4f20\u7edf \u77ed\u601d\u7ef4\u94fe\u5728\u63a8\u7406\u8282\u70b9\u6570\u91cf\u3001\u5e76\u884c\u63a2\u7d22\u80fd\u529b\u4ee5\u53ca\u56de\u6eaf\u8c03\u6574\u65b9\u9762\u7684\u9650\u5236\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u5728\u66f4\u5927\u7684\u95ee\u9898\u7a7a\u95f4\u4e2d\u5bfb\u627e\u6700\u4f73\u89e3\u3002<\/p>\n<p><strong>3. \u957f\u601d\u7ef4\u94fe\u5177\u5907\u4e09\u5927\u6838\u5fc3\u7279\u6027 <\/strong><strong> \u4f7f\u5176\u63a8\u7406\u80fd\u529b\u8fdc\u8d85\u77ed\u601d\u7ef4\u94fe<\/strong><\/p>\n<p><strong>\u6df1\u5ea6\u63a8\u7406\uff08Deep Reasoning\uff09<\/strong>\uff1a\u77ed\u601d\u7ef4\u94fe\u53ea\u80fd\u5904\u7406\u6709\u9650\u6570\u91cf\u7684\u903b\u8f91\u8282\u70b9\uff0c\u9002\u7528\u4e8e\u8f83\u7b80\u5355\u7684\u95ee\u9898\uff0c\u800c \u957f\u601d\u7ef4\u94fe\u5141\u8bb8\u63a8\u7406\u8def\u5f84\u5927\u5e45\u6269\u5c55\uff0c\u6db5\u76d6\u66f4\u591a\u5c42\u6b21\u7684\u903b\u8f91\u5173\u7cfb\uff0c\u4ece\u800c\u89e3\u51b3\u590d\u6742\u7684\u591a\u6b65\u63a8\u7406\u95ee\u9898\u3002\u6570\u5b66\u63a8\u7406\u3001\u7b26\u53f7\u903b\u8f91\u63a8\u5bfc\u7b49\u9886\u57df\u5bf9\u8fd9\u4e00\u80fd\u529b\u6709\u6781\u9ad8\u9700\u6c42\u3002<\/p>\n<p><strong>\u5e7f\u6cdb\u63a2\u7d22\uff08Extensive Exploration\uff09<\/strong>\uff1a\u77ed\u601d\u7ef4\u94fe\u4ec5\u9650\u4e8e\u56fa\u5b9a\u7684\u903b\u8f91\u8def\u5f84\uff0c\u5f80\u5f80\u5bfc\u81f4\u63a8\u7406\u8fc7\u4e8e\u5355\u4e00\uff0c\u800c \u957f\u601d\u7ef4\u94fe\u901a\u8fc7\u62d3\u5c55\u903b\u8f91\u5206\u652f\u3001\u5e76\u884c\u63a2\u7d22\u591a\u4e2a\u53ef\u80fd\u89e3\uff0c\u80fd\u591f\u5927\u5e45\u63d0\u9ad8\u63a8\u7406\u7684\u7a33\u5065\u6027\u548c\u51c6\u786e\u5ea6\u3002\u8fd9\u79cd\u80fd\u529b\u5728\u5177\u6709\u591a\u79cd\u53ef\u80fd\u89e3\u6216\u9700\u8981\u8de8\u9886\u57df\u63a8\u7406\u7684\u4efb\u52a1\u4e2d\u5c24\u4e3a\u91cd\u8981\u3002<\/p>\n<p><strong>\u53ef\u884c\u6027\u53cd\u601d\uff08Feasible Reflection\uff09<\/strong>\uff1a\u77ed\u601d\u7ef4\u94fe\u7684\u63a8\u7406\u8fc7\u7a0b\u4e0d\u53ef\u56de\u6eaf\uff0c\u4e00\u65e6\u51fa\u9519\u5f88\u96be\u4fee\u6b63\uff0c\u800c \u957f\u601d\u7ef4\u94fe\u80fd\u591f\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u53cd\u601d\u3001\u6821\u9a8c\u5148\u524d\u7684\u903b\u8f91\u6b65\u9aa4\uff0c\u7ed3\u5408\u53cd\u9988\u673a\u5236\u4f18\u5316\u63a8\u7406\u8def\u5f84\uff0c\u51cf\u5c11\u63a8\u7406\u9519\u8bef\uff0c\u63d0\u5347\u6574\u4f53\u63a8\u7406\u8d28\u91cf\u3002<\/p>\n<p><strong>4. \u957f\u601d\u7ef4\u94fe\u73b0\u8c61\u5206\u6790 \u00a0\u516d\u5927\u63a8\u7406\u73b0\u8c61<\/strong><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-8d9e25b5fd9dbc287bfa1aa889bcbf24.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-8d9e25b5fd9dbc287bfa1aa889bcbf24.png\"><\/a><\/section>\n<p><sup>\u56fe 3\uff1a\u5bf9\u957f\u601d\u7ef4\u94fe\u5916\u90e8\u884c\u4e3a\u516d\u4e2a\u7ecf\u5178\u73b0\u8c61\u7684\u5206\u6790\uff1a\uff08a\uff09\u5f53\u524d\u63a8\u7406\u5927\u6a21\u578b\u4e3a\u4f55\u4f1a\u4ea7\u751f\u957f\u601d\u7ef4\u94fe\uff1b\uff08b\uff09\u5f53\u524d\u957f\u601d\u7ef4\u94fe\u7cfb\u7edf\u7684\u63a8\u7406\u8fb9\u754c\u53ca\u5176\u5c40\u9650\u6027\uff1b\uff08c\uff09\u5f53\u8d85\u51fa\u63a8\u7406\u5927\u6a21\u578b\u7684\u63a8\u7406\u8fb9\u754c\u65f6\uff0c\u8fc7\u5ea6\u601d\u8003\u53ef\u80fd\u5bfc\u81f4\u7684\u6027\u80fd\u8870\u9000\uff1b\uff08d\uff09\u6d4b\u8bd5\u9636\u6bb5\u6269\u5c55\u65b9\u6cd5\u7684\u5e94\u7528\u53ca\u5176\u6269\u5c55\u89c4\u5f8b\u4e0e\u5c40\u9650\u6027\uff1b\uff08e\uff09\u4f7f\u7528\u8fc7\u7a0b\u4e2d\u7684\u5956\u52b1\u6a21\u578b\u4e0e\u7ed3\u679c\u5956\u52b1\u6a21\u578b\u6548\u679c\u5bf9\u6bd4\uff1b\uff08f\uff09\u201c\u987f\u609f\u201d \u65f6\u523b\u7684\u63a2\u7d22\u53ca\u5176\u6f5c\u5728\u539f\u56e0\u3002<\/sup><\/p>\n<p>\u7814\u7a76\u8868\u660e\uff0c\u957f\u601d\u7ef4\u94fe\u901a\u8fc7\u591a\u79cd\u65b9\u5f0f\u589e\u5f3a\u4e86\u63a8\u7406\u5927\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4e3b\u8981\u8868\u73b0\u4e3a<strong>\u516d\u5927\u6838\u5fc3\u63a8\u7406\u73b0\u8c61<\/strong>\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u957f\u601d\u7ef4\u94fe\u6d8c\u73b0\u73b0\u8c61<\/strong>\uff1a\u7814\u7a76\u53d1\u73b0\uff0c\u957f\u601d\u7ef4\u94fe\u5728\u6a21\u578b\u9884\u8bad\u7ec3\u9636\u6bb5\u5df2\u88ab\u6f5c\u5728\u7f16\u7801\uff0c\u53ef\u901a\u8fc7<strong>\u4e0a\u4e0b\u6587\u793a\u4f8b\uff08contextual examples\uff09\u6216\u7279\u5b9a\u63d0\u793a\u8bcd<\/strong>\u8fdb\u884c\u6fc0\u6d3b\u3002\u8fdb\u4e00\u6b65\u5730\uff0c\u4e00\u4e9b\u5de5\u4f5c\u53d1\u73b0\u5927\u6a21\u578b\u80fd\u591f\u76f4\u63a5\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u76f4\u63a5\u9ad8\u6548\u7684\u6fc0\u53d1\u51fa\u957f\u601d\u7ef4\u94fe\u7684\u80fd\u529b\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u63a8\u7406\u8fb9\u754c\uff08Reasoning Boundary\uff09\u73b0\u8c61<\/strong>\uff1a\u63a8\u7406\u5927\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u5e76\u975e\u65e0\u9650\u6269\u5c55\uff0c\u800c\u662f\u5728\u8d85\u51fa\u4e00\u5b9a\u590d\u6742\u5ea6\u9608\u503c\u540e\u8fc5\u901f\u4e0b\u964d\uff0c\u79f0\u4e4b\u4e3a \u201c\u63a8\u7406\u8fb9\u754c\u201d\uff0c<strong>\u5e76\u53d1\u73b0<\/strong>\u7a81\u7834\u63a8\u7406\u8fb9\u754c\u4f1a\u5bfc\u81f4\u63a8\u7406\u9519\u8bef\u7d2f\u79ef\uff0c\u4ece\u800c\u964d\u4f4e\u6574\u4f53\u6027\u80fd\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u8fc7\u5ea6\u601d\u8003\uff08Overthinking\uff09\u73b0\u8c61<\/strong>\uff1a\u968f\u7740\u63a8\u7406\u94fe\u7684\u589e\u52a0\uff0c\u6a21\u578b\u6027\u80fd\u5728\u67d0\u4e2a\u70b9\u4e4b\u524d\u6301\u7eed\u63d0\u9ad8\uff0c\u4f46\u8d85\u8fc7\u8be5\u70b9\u540e\uff0c\u9519\u8bef\u7387\u53cd\u800c\u4e0a\u5347\u3002\u4f8b\u5982\uff0c\u8fc7\u957f\u7684\u63a8\u7406\u94fe\u53ef\u80fd\u5bfc\u81f4 \u201c\u96ea\u7403\u6548\u5e94\u201d\uff08snowball errors\uff09\uff0c\u5373\u524d\u671f\u63a8\u7406\u9519\u8bef\u4e0d\u65ad\u7d2f\u79ef\uff0c\u6700\u7ec8\u5f71\u54cd\u7ed3\u679c\u53ef\u9760\u6027\u3002\u800c\u4e00\u4e9b\u7814\u7a76\u8005\u8ba4\u4e3a\u8fc7\u5ea6\u601d\u8003\u662f\u7531\u4e8e\u63a8\u7406\u957f\u5ea6\u8d85\u8fc7\u4e86\u5176\u63a8\u7406\u8fb9\u754c\u5bfc\u81f4\u7684\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u63a8\u7406\u6269\u5c55\u6027\uff08Inference Test-Time Scaling\uff09\u73b0\u8c61<\/strong>\uff1a\u63a8\u7406\u8def\u5f84\u7684\u6269\u5c55\u53ef\u5206\u4e3a<strong>\u5782\u76f4\u6269\u5c55\uff08Vertical Scaling\uff09\u548c\u5e76\u884c\u6269\u5c55\uff08Parallel Scaling\uff09<\/strong>\uff0c\u4f46\u4e24\u8005\u5747\u5b58\u5728\u4e0a\u9650\u3002\u4f8b\u5982\uff0c\u5e76\u884c\u6269\u5c55\u867d\u7136\u53ef\u4ee5\u589e\u52a0\u7b54\u6848\u7684\u591a\u6837\u6027\uff0c\u4f46\u5e76\u4e0d\u80fd\u8d85\u8fc7\u6700\u4f18\u9a8c\u8bc1\u7b56\u7565\u7684\u9650\u5236\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u8fc7\u7a0b\u5956\u52b1 vs. \u7ed3\u679c\u5956\u52b1\u7684\u8bad\u7ec3\u5dee\u5f02\u73b0\u8c61<\/strong>\uff1a\u5f3a\u5316\u5b66\u4e60\u4f18\u5316\u63a8\u7406\u8def\u5f84\u65f6\uff0c<strong>\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff08PRM\uff09<\/strong>\u80fd\u63d0\u4f9b\u66f4\u7cbe\u7ec6\u7684\u5956\u52b1\u53cd\u9988\uff0c\u800c<strong>\u7ed3\u679c\u5956\u52b1\u6a21\u578b\uff08ORM\uff09<\/strong>\u4f9d\u8d56\u6700\u7ec8\u7b54\u6848\u8fdb\u884c\u4f18\u5316\uff0c\u867d\u7136\u6536\u655b\u66f4\u5feb\uff0c\u4f46\u96be\u4ee5\u63d0\u4f9b\u4e2d\u95f4\u63a8\u7406\u7684\u7ec6\u7c92\u5ea6\u4f18\u5316\u4fe1\u606f\u3002\u4f46\u4e8c\u8005\u5728\u6570\u636e\u8d28\u91cf\u63a5\u8fd1\u7684\u60c5\u51b5\u4e0b\uff0c\u7406\u8bba\u4e0a\u6ca1\u6709\u8f83\u5927\u7684\u6027\u80fd\u5dee\u5f02\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u987f\u609f\u65f6\u523b\uff08Aha Moment\uff09\u73b0\u8c61<\/strong>\uff1a\u67d0\u4e9b\u60c5\u51b5\u4e0b\uff0c\u6a21\u578b\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u4f1a\u7ecf\u5386<strong>\u7a81\u5982\u5176\u6765\u7684\u8ba4\u77e5\u7a81\u7834<\/strong>\u3002\u7814\u7a76\u8868\u660e\uff0c\u57fa\u4e8e<strong>\u89c4\u5219\u7684\u5f3a\u5316\u5b66\u4e60\uff08rule-based RL\uff09<\/strong>\u53ef\u4ee5\u6709\u6548\u89e6\u53d1\u8fd9\u4e00\u73b0\u8c61\uff0c\u4f7f\u5f97\u5927\u6a21\u578b\u5728\u65e0\u76d1\u7763\u7684\u60c5\u51b5\u4e0b\u8fdb\u884c\u81ea\u6211\u53cd\u601d\u548c\u63a8\u7406\u4f18\u5316\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>5. \u6df1\u5ea6\u63a8\u7406 \u00a0\u6784\u5efa\u957f\u601d\u7ef4\u94fe\u903b\u8f91\u6df1\u5ea6\u7684\u5173\u952e\u8def\u5f84<\/strong><\/p>\n<p><strong>\u6df1\u5ea6\u63a8\u7406\uff08Deep Reasoning\uff09<\/strong>\u662f\u957f\u601d\u7ef4\u94fe\u7684\u6838\u5fc3\u80fd\u529b\uff0c\u5b83\u51b3\u5b9a\u4e86\u63a8\u7406\u5927\u6a21\u578b\u5728\u5904\u7406\u590d\u6742\u63a8\u7406\u4efb\u52a1\u65f6\u7684\u8868\u73b0\u3002\u7136\u800c\uff0c\u7f3a\u4e4f\u8db3\u591f\u7684\u63a8\u7406\u6df1\u5ea6\u4f1a\u5bfc\u81f4\u6a21\u578b\u5728\u9ad8\u96be\u5ea6\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u4e0b\u964d\u3002\u5f53\u524d\u7684\u7814\u7a76\u4e3b\u8981\u4ece<strong>\u63a8\u7406\u683c\u5f0f\uff08Deep Reasoning Format\uff09<\/strong>\u548c<strong>\u63a8\u7406\u5b66\u4e60\uff08Deep Reasoning Learning\uff09<\/strong>\u4e24\u65b9\u9762\u5165\u624b\uff0c\u901a\u8fc7\u4f18\u5316\u63a8\u7406\u7ed3\u6784\u548c\u8bad\u7ec3\u65b9\u6cd5\uff0c\u4f7f\u63a8\u7406\u5927\u6a21\u578b\u80fd\u591f\u66f4\u9ad8\u6548\u5730\u6267\u884c\u6df1\u5ea6\u63a8\u7406\u3002<\/p>\n<p><strong>5.1 \u6df1\u5ea6\u63a8\u7406\u683c\u5f0f\uff1a\u4f18\u5316\u63a8\u7406\u7ed3\u6784\u7684\u591a\u79cd\u8def\u5f84<\/strong><\/p>\n<p>\u957f\u601d\u7ef4\u94fe\u7684\u63a8\u7406\u80fd\u529b\u4f9d\u8d56\u4e8e\u4e09\u79cd\u4e3b\u8981\u7684\u63a8\u7406\u683c\u5f0f\uff1a<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-7066530b7e40315513a7faa06eaa218b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-7066530b7e40315513a7faa06eaa218b.png\"><\/a><\/section>\n<p><sup>\u56fe 4\uff1a\u6df1\u5ea6\u63a8\u7406\u683c\u5f0f\u4e3b\u8981\u5206\u4e3a\u4e09\u7c7b\uff1a\u81ea\u7136\u8bed\u8a00\u3001\u7ed3\u6784\u5316\u8bed\u8a00\u4e0e\u6f5c\u5728\u7a7a\u95f4\u63a8\u7406\uff08\u8fdb\u4e00\u6b65\u7ec6\u5206\u4e3a\u57fa\u4e8e token\u3001\u5411\u91cf\u548c\u7ba1\u7406\u5668\u9a71\u52a8\u7684\u6f5c\u5728\u63a8\u7406\uff09<\/sup><\/p>\n<ul>\n<li>\n<p><strong>\u81ea\u7136\u8bed\u8a00\u63a8\u7406\uff08Natural Language Reasoning\uff09<\/strong>\uff1a\u81ea\u7136\u8bed\u8a00\u63a8\u7406\u662f\u6700\u76f4\u89c2\u7684\u63a8\u7406\u65b9\u5f0f\uff0c\u5141\u8bb8\u6a21\u578b\u81ea\u7531\u7ec4\u7ec7\u903b\u8f91\u94fe\uff0c\u63d0\u9ad8\u63a8\u7406\u7684\u53ef\u89e3\u91ca\u6027\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u7ed3\u6784\u5316\u8bed\u8a00\u63a8\u7406\uff08Structured Language Reasoning\uff09<\/strong>\uff1a\u7ed3\u6784\u5316\u8bed\u8a00\u63a8\u7406\u4e3b\u8981\u4f9d\u8d56<strong>\u4ee3\u7801<\/strong>\u6216<strong>\u7b26\u53f7\u903b\u8f91<\/strong>\u8fdb\u884c\u66f4\u7cbe\u51c6\u7684\u903b\u8f91\u63a8\u7406\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6f5c\u5728\u7a7a\u95f4\u63a8\u7406\uff08Latent Space Reasoning\uff09<\/strong>\uff1a\u6f5c\u5728\u7a7a\u95f4\u63a8\u7406\u901a\u8fc7<strong>\u5728\u9690\u7a7a\u95f4\u6267\u884c\u63a8\u7406\u8ba1\u7b97<\/strong>\uff0c\u63d0\u9ad8\u63a8\u7406\u6548\u7387\u548c\u8fde\u8d2f\u6027\uff0c\u4e3b\u8981\u5305\u62ec\u4e09\u79cd\u65b9\u6cd5\uff1a\uff081\uff09<strong>\u57fa\u4e8e Token \u7684\u6f5c\u5728\u63a8\u7406<\/strong>\uff1a\u4f7f\u7528 \u201c\u63a8\u7406 Token\u201d \u6216 \u201c\u601d\u7ef4 Token\u201d \u5f15\u5bfc\u5927\u6a21\u578b\u8fdb\u884c\u9690\u5f0f\u63a8\u7406\u3002\uff082\uff09<strong>\u57fa\u4e8e\u5411\u91cf\u7684\u6f5c\u5728\u63a8\u7406<\/strong>\uff1a\u4f7f\u7528 \u201c\u601d\u7ef4\u5411\u91cf\uff08Thought Vectors\uff09\u201d \u5728\u4e0d\u540c\u5c42\u6b21\u95f4\u4f20\u64ad\u63a8\u7406\u72b6\u6001\uff0c\u4f7f\u63a8\u7406\u66f4\u52a0\u52a8\u6001\u548c\u7a33\u5065\u3002\uff083\uff09<strong>\u57fa\u4e8e\u7ba1\u7406\u5668\u7684\u6f5c\u5728\u63a8\u7406<\/strong>\uff1a\u901a\u8fc7 \u201c\u63a8\u7406\u7ba1\u7406\u5668\uff08Reasoning Manager\uff09\u201d \u4f18\u5316\u63a8\u7406\u94fe\u7684\u52a8\u6001\u63a7\u5236\uff0c\u63d0\u9ad8\u63a8\u7406\u6548\u7387\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>5.2 \u6df1\u5ea6\u63a8\u7406\u5b66\u4e60\uff1a\u4f18\u5316\u63a8\u7406\u80fd\u529b\u7684\u8bad\u7ec3\u65b9\u6cd5<\/strong><\/p>\n<p>\u63d0\u5347\u63a8\u7406\u5927\u6a21\u578b\u7684\u6df1\u5ea6\u63a8\u7406\u80fd\u529b\uff0c\u9700\u8981\u7ed3\u5408<strong>\u6a21\u4eff\u5b66\u4e60\uff08Imitation Learning\uff09<\/strong>\u548c<strong>\u81ea\u5b66\u4e60\uff08Self-Learning\uff09<\/strong>\u4e24\u79cd\u65b9\u6cd5\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u5728\u590d\u6742\u63a8\u7406\u4efb\u52a1\u4e2d\u66f4\u597d\u5730\u6cdb\u5316\u548c\u4f18\u5316\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-6537759a2378f82fe2795ea17da0379e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-6537759a2378f82fe2795ea17da0379e.png\"><\/a><\/section>\n<p><sup>\u56fe 5\uff1a\u6df1\u5ea6\u63a8\u7406\u5b66\u4e60\u7684\u4e0d\u540c\u7b56\u7565\uff0c\u5305\u62ec\uff1a(a) \u6df1\u5ea6\u63a8\u7406\u6a21\u4eff\u5b66\u4e60\uff0c\u5373\u6a21\u4eff\u6765\u81ea\u9ad8\u7ea7\u6df1\u5ea6\u63a8\u7406\u7cfb\u7edf\uff08\u5982\u9ad8\u7ea7\u63a8\u7406\u5927\u6a21\u578b\u3001MCTS \u7b49\uff09\u4ea7\u751f\u7684\u6570\u636e\uff0c\u901a\u8fc7\u76d1\u7763\u5fae\u8c03\u8bad\u7ec3\u63a8\u7406\u6a21\u578b\uff1b(b) \u6df1\u5ea6\u63a8\u7406\u81ea\u6211\u5b66\u4e60\uff0c\u5373\u901a\u8fc7\u9690\u5f0f\u5956\u52b1\u9a71\u52a8\u7684\u57fa\u4e8e\u504f\u597d\u7684\u5f3a\u5316\u5b66\u4e60\uff0c\u5b9e\u73b0\u63a8\u7406\u6a21\u578b\u7684\u81ea\u6211\u6539\u8fdb\u4e0e\u4f18\u5316\u3002<\/sup><\/p>\n<p>5.2.1 \u6df1\u5ea6\u63a8\u7406\u6a21\u4eff\uff08Imitation Learning\uff09<\/p>\n<p>\u6a21\u4eff\u5b66\u4e60\u901a\u8fc7<strong>\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09<\/strong>\uff0c\u8ba9\u6a21\u578b\u4ece<strong>\u4eba\u7c7b\u3001\u5148\u8fdb\u63a8\u7406\u5927\u6a21\u578b\u6216\u6269\u5c55\u91c7\u6837\u7b56\u7565<\/strong>\u4e2d\u5b66\u4e60\u63a8\u7406\u6a21\u5f0f\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u6a21\u4eff\u4eba\u7c7b\u63a8\u7406<\/strong>\uff1a\u4f7f\u7528\u9ad8\u8d28\u91cf\u7684\u4eba\u5de5\u6807\u6ce8\u6570\u636e\u8bad\u7ec3 \u63a8\u7406\u5927\u6a21\u578b\uff0c\u76f4\u63a5\u6a21\u4eff\u4eba\u7c7b\u884c\u4e3a\u8fdb\u884c\u6df1\u5ea6\u63a8\u7406\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6a21\u4eff\u5148\u8fdb\u63a8\u7406\u5927\u6a21\u578b\u63a8\u7406<\/strong>\uff1a\u5229\u7528\u5148\u8fdb\u7684\u63a8\u7406\u5927\u6a21\u578b\u4f5c\u4e3a\u6559\u5e08\u6a21\u578b\u8bad\u7ec3\u4e0d\u4f1a\u6df1\u5ea6\u63a8\u7406\u7684\u5b66\u751f\u5c0f\u6a21\u578b\uff0c\u66f4\u52a0\u5feb\u901f\u5730\u5347\u6570\u5b66\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6a21\u4eff\u6269\u5c55\u91c7\u6837<\/strong>\uff1a\u7ed3\u5408\u6269\u5c55\u641c\u7d22\u7b56\u7565\u4f18\u5316\u63a8\u7406\u8def\u5f84\uff0c\u4f8b\u5982\u91c7\u7528 MCTS \u6216\u8005\u5927\u91cf\u7ed3\u679c\u91c7\u6837\u641c\u7d22\u4f18\u5316\u63a8\u7406\uff0c\u4f7f\u6a21\u578b\u66f4\u597d\u5730\u9002\u5e94\u9762\u5411\u6df1\u5ea6\u63a8\u7406\u7684\u957f\u601d\u7ef4\u94fe\u4efb\u52a1\u3002<\/p>\n<\/li>\n<\/ul>\n<p>5.2.2 \u6df1\u5ea6\u63a8\u7406\u81ea\u5b66\u4e60\uff08Self-Learning\uff09<\/p>\n<p>\u81ea\u5b66\u4e60\u5141\u8bb8\u6a21\u578b<strong>\u81ea\u4e3b\u4f18\u5316\u63a8\u7406\u8fc7\u7a0b<\/strong>\uff0c\u4e3b\u8981\u91c7\u7528<strong>\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09<\/strong>\u548c<strong>\u6811\u641c\u7d22\uff08MCTS\uff09<\/strong>\u65b9\u6cd5\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u76f4\u63a5\u91c7\u6837\u81ea\u5b66\u4e60<\/strong>\uff1a\u91c7\u7528\u76f4\u63a5\u5bf9\u5927\u6a21\u578b\u81ea\u8eab\u8fdb\u884c\u91c7\u6837\uff0c\u4ece\u800c\u8fdb\u884c\u81ea\u6211\u5b66\u4e60\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6811\u641c\u7d22\u91c7\u6837\u81ea\u5b66\u4e60<\/strong>\uff1a\u901a\u8fc7 MCTS \u7b49\u65b9\u6cd5\u5bf9\u5927\u6a21\u578b\u8fdb\u884c\u6709\u6307\u5bfc\u7684\u6811\u641c\u7d22\u91c7\u6837\uff0c\u63d0\u9ad8\u63a8\u7406\u7684\u6df1\u5ea6\uff0c\u4f18\u5316 \u957f\u601d\u7ef4\u94fe\u751f\u6210\u8fc7\u7a0b\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>6. \u53ef\u884c\u6027\u53cd\u601d\uff08Feasible Reflection\uff09 \u00a0\u63d0\u5347\u957f\u601d\u7ef4\u94fe\u53ef\u9760\u6027\u7684\u5173\u952e\u673a\u5236<\/strong><\/p>\n<p>\u5728 \u957f\u601d\u7ef4\u94fe\u63a8\u7406\u4e2d\uff0c<strong>\u53ef\u884c\u6027\u53cd\u601d\uff08Feasible Reflection\uff09<\/strong>\u662f\u786e\u4fdd\u903b\u8f91\u94fe\u6761\u51c6\u786e\u6027\u7684\u91cd\u8981\u673a\u5236\uff0c\u5b83\u6d89\u53ca\u5bf9\u63a8\u7406\u8fc7\u7a0b\u7684\u53cd\u9988\uff08Feedback\uff09\u548c\u6539\u8fdb\uff08Refinement\uff09\u3002\u53cd\u9988\u673a\u5236\u63d0\u4f9b\u8bc4\u4f30\uff0c\u5e2e\u52a9\u6a21\u578b\u8bc6\u522b\u548c\u4fee\u6b63\u9519\u8bef\uff0c\u800c\u6539\u8fdb\u673a\u5236\u5219\u5141\u8bb8\u6a21\u578b\u57fa\u4e8e\u53cd\u9988\u8fdb\u884c\u8c03\u6574\uff0c\u4ee5\u63d0\u9ad8\u63a8\u7406\u8d28\u91cf\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-55e0fcd315eafc5f91a84a80545e5cb5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-55e0fcd315eafc5f91a84a80545e5cb5.png\"><\/a><\/section>\n<p><sup>\u56fe 6\uff1a\u53ef\u884c\u53cd\u601d\u7684\u53cd\u9988\u80fd\u529b\u6846\u67b6\u5305\u62ec\u603b\u4f53\u53cd\u9988\u4e0e\u8fc7\u7a0b\u53cd\u9988\u4e24\u79cd\u5f62\u5f0f\u3002\u603b\u4f53\u53cd\u9988\u5305\u62ec\uff1a\u7ed3\u679c\u5956\u52b1\u6a21\u578b\uff08\u4ee5\u6570\u503c\u5f62\u5f0f\u8fdb\u884c\u53cd\u9988\u3001\u57fa\u4e8e\u89c4\u5219\u7684\u6b63\u786e\u6027\u5224\u65ad\u673a\u5236\uff0c\u4ee5\u53ca\u57fa\u4e8e\u63a8\u7406\u5927\u6a21\u578b\u7684\u603b\u4f53\u8bc4\u4ef7\u6a21\u578b\u3002\u8fc7\u7a0b\u53cd\u9988\u5219\u5305\u542b\uff1a\u4ee5\u6570\u503c\u5f62\u5f0f\u63d0\u4f9b\u53cd\u9988\u7684\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff0c\u4ee5\u53ca\u57fa\u4e8e\u63a8\u7406\u5927\u6a21\u578b\u7684\u9010\u6b65\u8bc4\u4ef7\u6a21\u578b\u3002\u6b64\u5916\uff0c\u8fc7\u7a0b\u53cd\u9988\u8fd8\u53ef\u7ed3\u5408\u64cd\u4f5c\u7cfb\u7edf\u73af\u5883\u3001\u771f\u5b9e\u73af\u5883\u6216\u56fe\u5f62\u754c\u9762\u73af\u5883\u7b49\u8fdb\u884c\u4ea4\u4e92\u5f0f\u8bc4\u4f30\u3002<\/sup><\/p>\n<p>6.1 \u53cd\u9988\uff08Feedback\uff09\uff1a\u4ece\u6574\u4f53\u5230\u8fc7\u7a0b\u7684\u4f18\u5316<\/p>\n<p>\u53cd\u9988\u662f\u957f\u601d\u7ef4\u94fe\u63a8\u7406\u4f18\u5316\u7684\u5173\u952e\u73af\u8282\uff0c\u5b83\u53ef\u4ee5\u7528\u4e8e<strong>\u9a8c\u8bc1\u903b\u8f91\u6b63\u786e\u6027\u3001\u63d0\u9ad8\u6a21\u578b\u7a33\u5b9a\u6027\uff0c\u5e76\u51cf\u5c11\u9519\u8bef\u7d2f\u79ef\u3002<\/strong><\/p>\n<p>6.1.1 \u6574\u4f53\u53cd\u9988\uff08Overall Feedback\uff09\uff1a\u5168\u5c40\u4f18\u5316\u63a8\u7406\u8d28\u91cf<\/p>\n<p>\u6574\u4f53\u53cd\u9988\u5173\u6ce8\u63a8\u7406\u94fe\u6761\u7684\u6700\u7ec8\u7ed3\u679c\uff0c\u800c\u975e\u6bcf\u4e00\u6b65\u7684\u7ec6\u8282\u3002\u5b83\u5e38\u7528\u4e8e\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u4e2d\u7684\u5956\u52b1\u5efa\u6a21\uff0c\u4ee5\u63d0\u5347\u63a8\u7406\u5927\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4e3b\u8981\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u7ed3\u679c\u5956\u52b1\u6a21\u578b\uff08ORM\uff09<\/strong>\uff1a\u91c7\u7528\u5956\u52b1\u6a21\u578b\u8bc4\u4f30\u63a8\u7406\u7ed3\u679c\uff0c\u6216\u8005\u901a\u8fc7\u81ea\u4e00\u81f4\u6027\u4f18\u5316\u5956\u52b1\u53cd\u9988\uff0c\u83b7\u5f97\u6570\u503c\u5316\u7684\u7ed3\u679c\u5956\u52b1\u5206\u6570\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u89c4\u5219\u63d0\u53d6\uff08Rule Extraction\uff09<\/strong>\uff1a\u89c4\u5219\u53cd\u9988\u53ef\u63d0\u4f9b\u9ad8\u51c6\u786e\u5ea6\u7684\u7ed3\u679c\u6821\u6b63\uff0c\u4f8b\u5982 STaR\u3001ReST \u548c ReFT \u6846\u67b6\u8868\u660e\uff0c\u6570\u5b66\u95ee\u9898\u4e2d\u57fa\u4e8e\u89c4\u5219\u7684\u53cd\u9988\u6bd4 ORM \u66f4\u6709\u6548\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6279\u5224\u6027\u53cd\u9988\u6a21\u578b\uff08Critic Models\uff09<\/strong>\uff1a\u8fd9\u7c7b\u65b9\u6cd5\u5229\u7528\u5927\u6a21\u578b\u81ea\u6211\u53cd\u601d\uff08Self-Reflection\uff09\u673a\u5236\u8fdb\u884c\u81ea\u6211\u6279\u5224\uff0c\u63d0\u4f9b\u6587\u672c\u5316\u7ed3\u679c\u5956\u52b1\uff0c\u63d0\u9ad8\u63a8\u7406\u6548\u7387\u3002<\/p>\n<\/li>\n<\/ul>\n<p>6.1.2 \u8fc7\u7a0b\u53cd\u9988\uff08Process Feedback\uff09\uff1a\u7ec6\u5316\u63a8\u7406\u6b65\u9aa4<\/p>\n<p>\u8fc7\u7a0b\u53cd\u9988\u5173\u6ce8\u63a8\u7406\u7684\u4e2d\u95f4\u6b65\u9aa4\uff0c\u901a\u5e38\u53ef\u4ee5\u589e\u5f3a\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6216\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22\uff08MCTS\uff09\u7684\u4f18\u5316\uff0c\u4e3b\u8981\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff08PRM\uff09<\/strong>\uff1aPRM \u901a\u8fc7 SFT \u6216\u8005 DPO \u8bad\u7ec3\u4e00\u4e2a\u9762\u5411\u8fc7\u7a0b\u7684\u5956\u52b1\u6253\u5206\u5668\uff0c\u63d0\u4f9b\u57fa\u4e8e\u6b65\u9aa4\u7684\u6570\u503c\u5316\u5956\u52b1\u4fe1\u53f7\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6279\u5224\u6027\u53cd\u9988\u6a21\u578b\uff08Critic Models\uff09<\/strong>\uff1a\u5141\u8bb8 LLM \u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u5bf9\u63a8\u7406\u8fc7\u7a0b\u4e0e\u7ed3\u679c\u8fdb\u884c\u52a8\u6001\u8bc4\u4f30\u4e0e\u8c03\u6574\uff0c\u4f18\u5316\u63a8\u7406\u8def\u5f84\uff0c\u63d0\u4f9b\u66f4\u901a\u7528\u7684\u6b65\u9aa4\u7ea7\u6587\u672c\u5316\u53cd\u9988\u4fe1\u53f7\u3002<\/p>\n<\/li>\n<\/ul>\n<p>6.1.3 \u6df7\u5408\u53cd\u9988\uff08Hybrid Feedback\uff09<\/p>\n<p>\u7531\u4e8e\u6574\u4f53\u53cd\u9988\u548c\u8fc7\u7a0b\u53cd\u9988\u5404\u5177\u4f18\u52bf\uff0c\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u7ed3\u5408\u4e24\u8005\u7684\u6df7\u5408\u65b9\u6cd5\uff0c\u8fdb\u884c\u53cc\u91cd\u53cd\u9988\u4f18\u5316\uff0c\u63d0\u9ad8\u63a8\u7406\u8fde\u8d2f\u6027\u3002<\/p>\n<p><strong>6.2 \u6539\u8fdb\uff08Refinement\uff09\uff1a\u4f18\u5316\u63a8\u7406\u8def\u5f84\u7684\u5173\u952e\u624b\u6bb5<\/strong><\/p>\n<p>\u5728\u53cd\u9988\u673a\u5236\u7684\u57fa\u7840\u4e0a\uff0c<strong>\u6539\u8fdb\uff08Refinement\uff09<\/strong>\u8fdb\u4e00\u6b65\u5e2e\u52a9\u63a8\u7406\u5927\u6a21\u578b\u7ea0\u6b63\u9519\u8bef\u3001\u4f18\u5316\u63a8\u7406\u8def\u5f84\uff0c\u4ece\u800c\u63d0\u5347\u63a8\u7406\u80fd\u529b\u3002\u73b0\u6709\u7814\u7a76\u4e3b\u8981\u91c7\u7528\u4e09\u79cd\u65b9\u6cd5\uff1a<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-4612b931843467d0e948bf7b6b7f254c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-4612b931843467d0e948bf7b6b7f254c.png\"><\/a><\/section>\n<p><sup>\u56fe 7\uff1a\u7ec6\u5316\u65b9\u6cd5\u4e3b\u8981\u5206\u4e3a\u4e09\u7c7b\uff1a(a) \u57fa\u4e8e\u63d0\u793a\u7684\u6539\u8fdb\u751f\u6210\uff0c\u901a\u8fc7\u6784\u9020\u5408\u9002\u7684\u63d0\u793a\u7b56\u7565\uff0c\u5f15\u5bfc\u6a21\u578b\u8fdb\u884c\u53ef\u884c\u7684\u53cd\u601d\u8fc7\u7a0b\u5e76\u4ea7\u51fa\u4fee\u6b63\u7b54\u6848\uff1b(b) \u57fa\u4e8e\u76d1\u7763\u5fae\u8c03\u7684\u6539\u8fdb\u6a21\u4eff\uff0c\u5229\u7528\u6765\u81ea\u9ad8\u7ea7\u63a8\u7406\u5927\u6a21\u578b\u7684\u7ec6\u5316\u793a\u4f8b\uff0c\u901a\u8fc7\u76d1\u7763\u5b66\u4e60\u5bf9\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff1b(c) \u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u7684\u6539\u8fdb\u5b66\u4e60\uff0c\u6a21\u578b\u6839\u636e\u53cd\u9988\u4fe1\u53f7\uff08\u5982\u5956\u52b1\uff09\u5bf9\u81ea\u8eab\u751f\u6210\u7684\u4fee\u6b63\u8fdb\u884c\u4f18\u5316\uff0c\u9010\u6b65\u5b66\u4f1a\u81ea\u6211\u6539\u8fdb\u63a8\u7406\u4e0e\u8f93\u51fa\u3002<\/sup><\/p>\n<ul>\n<li>\n<p><strong>\u57fa\u4e8e\u63d0\u793a\u7684\u6539\u8fdb\uff08Prompt-Based Refinement\uff09<\/strong>\uff1a\u8be5\u65b9\u6cd5\u901a\u8fc7 LLM \u751f\u6210\u521d\u59cb\u63a8\u7406\uff0c\u518d\u5229\u7528\u81ea\u8eab\u7684\u53cd\u9988\u8fed\u4ee3\u4f18\u5316\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u57fa\u4e8e\u76d1\u7763\u5fae\u8c03\u7684\u6539\u8fdb\uff08SFT-Based Refinement\uff09<\/strong>\uff1aSFT \u65b9\u6cd5\u901a\u8fc7\u4ece<strong>\u5148\u8fdb LLMs \u6216\u4eba\u5de5\u6807\u6ce8\u6570\u636e\u4e2d\u5b66\u4e60\u63a8\u7406\u6539\u8fdb\u65b9\u5f0f<\/strong>\uff0c\u63d0\u5347\u63a8\u7406\u5927\u6a21\u578b\u53cd\u601d\u80fd\u529b\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u7684\u6539\u8fdb\uff08RL-Based Refinement\uff09<\/strong>\uff1aRL \u65b9\u6cd5\u5141\u8bb8\u63a8\u7406\u5927\u6a21\u578b\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u81ea\u9002\u5e94\u8c03\u6574\u63a8\u7406\u8def\u5f84\uff0c\u51cf\u5c11\u5bf9\u4eba\u5de5\u6807\u6ce8\u6570\u636e\u7684\u4f9d\u8d56\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>7. \u5e7f\u6cdb\u63a2\u7d22\uff08Extensive Exploration\uff09 \u62d3\u5c55 \u957f\u601d\u7ef4\u94fe\u63a8\u7406\u7684\u8fb9\u754c<\/strong><\/p>\n<p>\u63a2\u7d22\u80fd\u529b\u662f \u957f\u601d\u7ef4\u94fe\u63a8\u7406\u7684\u5173\u952e\uff0c\u4f7f\u63a8\u7406\u5927\u6a21\u578b\u80fd\u591f\u5728\u590d\u6742\u95ee\u9898\u7a7a\u95f4\u4e2d\u8fdb\u884c\u7b56\u7565\u6027\u5206\u652f\u548c\u8fed\u4ee3\u4f18\u5316\u3002\u7814\u7a76\u8868\u660e\uff0c\u5047\u8bbe\u5206\u652f\u548c\u57fa\u4e8e\u53cd\u601d\u7684\u9519\u8bef\u56de\u6eaf\u662f\u7a81\u7834\u7ebf\u6027\u63a8\u7406\u8def\u5f84\u5c40\u9650\u7684\u91cd\u8981\u65b9\u6cd5\u3002\u5f53\u524d\u7814\u7a76\u4e3b\u8981\u5173\u6ce8\u63a2\u7d22\u89c4\u6a21\u5316\u3001\u5185\u90e8\u63a2\u7d22\u3001\u5916\u90e8\u63a2\u7d22\uff0c\u4ee5\u63d0\u9ad8\u6a21\u578b\u7684\u63a8\u7406\u8d28\u91cf\u548c\u9002\u5e94\u6027\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-d4a806b6fd67669d8eb0ff1f8562b18f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-d4a806b6fd67669d8eb0ff1f8562b18f.png\"><\/a><\/section>\n<p><sup>\u56fe 8\uff1a\u63a8\u7406\u6d4b\u8bd5\u9636\u6bb5\u7684\u4e24\u79cd\u5e38\u89c1\u6269\u5c55\u7b56\u7565\u793a\u610f\u56fe\uff0c\u5305\u62ec\uff1a(a) \u7eb5\u5411\u6269\u5c55\uff1a\u901a\u8fc7\u5ef6\u957f\u6a21\u578b\u7684\u957f\u94fe\u5f0f\u601d\u7ef4\u8fc7\u7a0b\u4ee5\u589e\u5f3a\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u53d7\u9650\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u8fb9\u754c\uff0c\u6700\u7ec8\u53ef\u80fd\u96be\u4ee5\u7a81\u7834\u6027\u80fd\u74f6\u9888\u3002(b) \u6a2a\u5411\u6269\u5c55\uff1a\u901a\u8fc7\u589e\u52a0\u91c7\u6837\u6b21\u6570\u751f\u6210\u591a\u4e2a\u63a8\u7406\u7ed3\u679c\uff0c\u5e76\u7ed3\u5408\u81ea\u4e00\u81f4\u6027\u3001\u81ea\u9a8c\u8bc1\u7b49\u673a\u5236\u8fdb\u884c\u7ed3\u679c\u9a8c\u8bc1\u4e0e\u6574\u5408\uff0c\u4ece\u800c\u63d0\u5347\u8f93\u51fa\u7684\u6574\u4f53\u8d28\u91cf\uff0c\u4f46\u5176\u6027\u80fd\u4e0a\u9650\u4ecd\u4e0d\u8d85\u8fc7 Pass@k \u6307\u6807\u6240\u4ee3\u8868\u7684\u7406\u8bba\u6700\u4f18\u6c34\u5e73\u3002<\/sup><\/p>\n<p><strong>7.1 \u63a2\u7d22\u89c4\u6a21\u5316\uff08Exploration Scaling\uff09\uff1a\u4f18\u5316\u63a8\u7406\u957f\u5ea6\u4e0e\u5e7f\u5ea6<\/strong><\/p>\n<p>\u63a2\u7d22\u89c4\u6a21\u5316\u65e8\u5728\u6269\u5c55\u63a8\u7406\u5927\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u63d0\u9ad8\u4efb\u52a1\u6c42\u89e3\u6548\u7387\u3002<\/p>\n<p>7.1.1 \u5782\u76f4\u6269\u5c55\uff08Vertical Scaling\uff09<\/p>\n<p>\u901a\u8fc7\u5ef6\u957f\u63a8\u7406\u8def\u5f84\u589e\u52a0\u63a8\u7406\u6df1\u5ea6\uff0c\u5982\u63a8\u7406\u65f6\u6269\u5c55\uff08Test-time Scaling\uff09\u4f18\u5316\u8ba1\u7b97\u8d44\u6e90\u4ee5\u63d0\u5347\u63a8\u7406\u6548\u679c\u3002\u6b64\u5916\uff0c\u9690\u7a7a\u95f4\u6269\u5c55\uff08Latent Space Scaling\uff09\u5728\u6ce8\u610f\u529b\u8de8\u5ea6\u9650\u5236\u4e0b\uff0c\u5229\u7528\u9012\u5f52\u8ba1\u7b97\u589e\u5f3a\u63a8\u7406\u6df1\u5ea6\u3002<\/p>\n<p>7.1.2 \u5e76\u884c\u6269\u5c55\uff08Parallel Scaling\uff09<\/p>\n<p>\u91c7\u7528\u591a\u6b21\u63a8\u7406\u91c7\u6837\u7ed3\u5408\u9a8c\u8bc1\u673a\u5236\uff0c\u4ee5\u7b5b\u9009\u6700\u4f18\u63a8\u7406\u8def\u5f84\u3002<\/p>\n<p><strong>\u63a2\u7d22\u9a8c\u8bc1\u4f18\u5316\u65b9\u9762<\/strong>\uff0c\u6700\u65e9\u5730\uff0c\u81ea\u4e00\u81f4\u6027\uff08Self-Consistency\uff09\u65b9\u6cd5\u5c31\u53ef\u901a\u8fc7\u591a\u8f6e\u63a8\u7406\u91c7\u6837\u53ca\u591a\u6570\u6295\u7968\u673a\u5236\u4f18\u5316\u7b54\u6848\u9009\u62e9\uff0c\u800c\u4e00\u4e9b\u5de5\u4f5c\u901a\u8fc7\u8fdb\u4e00\u6b65\u5f15\u5165\u7ec6\u7c92\u5ea6\u81ea\u4e00\u81f4\u6027\uff08Fine-grained Self-Consistency\uff09\u7ed3\u5408\u9010\u6b65\u9a8c\u8bc1\u673a\u5236\uff08Step-wise Verification\uff09\uff0c\u6709\u6548\u51cf\u5c11\u63a8\u7406\u9a8c\u8bc1\u9519\u8bef\u7684\u7d2f\u79ef\u3002<\/p>\n<p><strong>\u63a2\u7d22\u8def\u5f84\u4f18\u5316\u65b9\u9762<\/strong>\uff0c\u77ed\u8def\u5f84\u591a\u6837\u5316\u91c7\u6837\uff08Diverse Short-path Sampling\uff09\u65b9\u6cd5\u901a\u8fc7\u4f18\u5316\u91c7\u6837\u6e29\u5ea6\u548c\u8de8\u8bed\u8a00\u63a8\u7406\u62d3\u5c55\uff0c\u63d0\u9ad8\u63a2\u7d22\u7684\u591a\u6837\u6027\u548c\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-04312f787e8d9a0b9f0855ad7e5e8a9b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-04312f787e8d9a0b9f0855ad7e5e8a9b.png\"><\/a><\/section>\n<p><sup>\u56fe 9\uff1a\u7528\u4e8e\u4f18\u5316\u5185\u90e8\u63a2\u7d22\u7684\u4e24\u7c7b\u4e3b\u8981\u65b9\u6cd5\uff1a(a) \u5f3a\u5316\u5b66\u4e60\u7b56\u7565\uff1a\u901a\u8fc7\u53c2\u8003\u6a21\u578b\u4e0e\u4ef7\u503c\u6a21\u578b\u8f85\u52a9\uff0c\u7ed3\u5408\u5956\u52b1\u6a21\u578b\u8fdb\u884c\u4f18\u52bf\u805a\u5408\uff0c\u4ece\u800c\u63d0\u5347\u7b56\u7565\u6a21\u578b\u7684\u8f93\u51fa\u8d28\u91cf\uff0c\u5b9e\u73b0\u7b56\u7565\u4f18\u5316\u3002(b) \u5956\u52b1\u7b56\u7565\uff1a\u6839\u636e\u6a21\u578b\u8f93\u51fa\u8bbe\u8ba1\u5956\u52b1\u673a\u5236\uff0c\u5305\u62ec\u89c4\u5219\u9a71\u52a8\u5956\u52b1\uff0c\u4f8b\u5982\u4f7f\u7528\u6b63\u5219\u5339\u914d\u6216\u6d4b\u8bd5\u7528\u4f8b\u6253\u5206\uff1b\u4ee5\u53ca\u6a21\u578b\u9a71\u52a8\u5956\u52b1\uff0c\u57fa\u4e8e\u5982\u548c PRM \u7b49\u8bc4\u4ef7\u6a21\u578b\u751f\u6210\u5956\u52b1\uff0c\u4ee5\u63d0\u5347\u5f3a\u5316\u5b66\u4e60\u6027\u80fd\u3002<\/sup><\/p>\n<p><strong>7.2 \u5185\u90e8\u81ea\u4e3b\u63a2\u7d22\uff08Internal Exploration\uff09\uff1a\u5f3a\u5316\u5b66\u4e60\u9a71\u52a8\u7684\u63a8\u7406\u4f18\u5316<\/strong><\/p>\n<p>\u5185\u90e8\u81ea\u4e3b\u63a2\u7d22\u5f3a\u8c03\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning, RL\uff09\u548c\u5956\u52b1\u7b56\u7565\uff08Reward Strategies\uff09\uff0c\u4f7f\u63a8\u7406\u5927\u6a21\u578b\u80fd\u591f\u81ea\u4e3b\u4f18\u5316\u63a8\u7406\u7b56\u7565\uff0c\u63d0\u9ad8\u6cdb\u5316\u80fd\u529b\u548c\u63a2\u7d22\u6df1\u5ea6\u3002<\/p>\n<p>7.2.1 \u5f3a\u5316\u5b66\u4e60\u7b56\u7565<\/p>\n<p>\u4e3b\u8981\u5206\u4e3a\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u57fa\u4e8e\u5956\u52b1\u6a21\u578b\u7684 RL<\/strong>\uff1a\u5982 Proximal Policy Optimization\uff08PPO\uff09\u548c Group Relative Policy Optimization\uff08GRPO\uff09\u57fa\u4e8e\u5956\u52b1\u6a21\u578b\u6216\u5956\u52b1\u7b56\u7565\u4ece\u800c\u5b8c\u6210\u7b56\u7565\u4f18\u5316\uff0c\u63d0\u5347\u63a8\u7406\u80fd\u529b\uff1b<\/p>\n<\/li>\n<li>\n<p><strong>\u65e0\u5956\u52b1\u6a21\u578b RL<\/strong>\uff1a\u91c7\u7528 DPO \u98ce\u683c\u6216\u8005\u89c4\u5219\u62bd\u53d6\u7684\u5956\u52b1\u8ba1\u7b97\u65b9\u6848\uff0c\u901a\u8fc7\u660e\u786e\u7684\u89c4\u5219\u5956\u52b1\u53cd\u9988\uff0c\u63d0\u9ad8\u7b56\u7565\u5b66\u4e60\u7684\u7a33\u5b9a\u6027\u3002<\/p>\n<\/li>\n<\/ul>\n<p>7.2.2 \u5956\u52b1\u7b56\u7565<\/p>\n<p>\u5956\u52b1\u7b56\u7565\u7684\u4f18\u5316\u662f\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u5173\u952e\uff0c\u5305\u62ec\u6b63\u786e\u6027\u5956\u52b1\uff08Correctness Rewarding\uff09\u3001\u683c\u5f0f\u5956\u52b1\uff08Format Rewarding\uff09\u548c\u89c4\u6a21\u5956\u52b1\uff08Scaling Rewarding\uff09\u3002\u6b63\u786e\u6027\u5956\u52b1\u901a\u8fc7\u63d0\u9ad8\u7b54\u6848\u51c6\u786e\u6027\u5f15\u5bfc\u6a21\u578b\u4f18\u5316\u63a8\u7406\u8def\u5f84\uff0c\u683c\u5f0f\u5956\u52b1\u786e\u4fdd\u63a8\u7406\u8fc7\u7a0b\u7b26\u5408\u89c4\u8303\uff0c\u800c\u89c4\u6a21\u5956\u52b1\u5219\u9f13\u52b1\u66f4\u957f\u7684\u63a8\u7406\u94fe\u6761\uff0c\u4f46\u7814\u7a76\u8868\u660e\u8fc7\u5ea6\u6269\u5c55\u53ef\u80fd\u5bfc\u81f4\u4f4e\u6548\u63a8\u7406\u3002<\/p>\n<p>\u4e3a\u4e86\u5c55\u793a\u5f53\u524d\u4e3b\u6d41\u5956\u52b1\u7b56\u7565\u5728\u5404\u7c7b\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u8868\u73b0\uff0c\u8868 5 \u603b\u7ed3\u4e86\u4e0d\u540c\u65b9\u6cd5\u5728 GSM8K\u3001AIME 2024\u3001MATH500 \u7b49\u6570\u636e\u96c6\u4e0a\u7684\u6027\u80fd\u5bf9\u6bd4\uff0c\u80fd\u591f\u76f4\u89c2\u4f53\u73b0\u7b56\u7565\u8bbe\u8ba1\u5bf9\u63a8\u7406\u80fd\u529b\u7684\u5f71\u54cd\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-64e517fd7cd8a8a8366a4fee1d1d6bb9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-64e517fd7cd8a8a8366a4fee1d1d6bb9.png\"><\/a><\/section>\n<p><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u8868 5\uff1a\u4e0d\u540c\u5185\u90e8\u63a2\u7d22\u65b9\u6cd5\u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u6027\u80fd\u8868\u73b0\uff0c\u4e3b\u8981\u6309 AIME 2024 \u6392\u5e8f\u3002\u201c-\u201d \u8868\u793a\u8bba\u6587\u672a\u62a5\u544a\u8be5\u5f97\u5206\u3002<\/sup><\/p>\n<p><strong>7.3 \u5916\u90e8\u5f15\u5bfc\u63a2\u7d22\uff08External Exploration\uff09\uff1a\u7ed3\u5408\u5916\u90e8\u7cfb\u7edf\u589e\u5f3a\u63a8\u7406\u80fd\u529b<\/strong><\/p>\n<p>\u5916\u90e8\u5f15\u5bfc\u63a2\u7d22\u7ed3\u5408\u4eba\u7c7b\u6216\u5916\u90e8\u7cfb\u7edf\uff0c\u63d0\u9ad8\u63a8\u7406\u5927\u6a21\u578b\u7684\u63a2\u7d22\u80fd\u529b\u548c\u63a8\u7406\u8d28\u91cf\u3002<\/p>\n<p>7.3.1 \u4eba\u7c7b\u9a71\u52a8\u63a2\u7d22\uff08Human-driven Exploration\uff09<\/p>\n<p>\u4f9d\u8d56\u4e8e\u9884\u8bbe\u7684\u63a8\u7406\u7ed3\u6784\uff0c\u5982 Tree-of-Thought\uff08ToT\uff09\u91c7\u7528\u6811\u72b6\u63a8\u7406\u7ed3\u6784\u589e\u5f3a\u590d\u6742\u4efb\u52a1\u6c42\u89e3\u80fd\u529b\uff0c\u800c Forest-of-Thought\uff08FoT\uff09\u7ed3\u5408\u591a\u68f5\u63a8\u7406\u6811\u63d0\u9ad8\u63a2\u7d22\u7684\u5168\u9762\u6027\u3002<\/p>\n<p>7.3.2 \u6a21\u578b\u9a71\u52a8\u63a2\u7d22\uff08Model-driven Exploration\uff09<\/p>\n<p>\u901a\u8fc7\u81ea\u9002\u5e94\u641c\u7d22\u7ed3\u6784\u548c\u53cd\u9988\u673a\u5236\u8fdb\u884c\u4f18\u5316\uff0c\u63a2\u7d22\u8def\u5f84\u4f18\u5316\u65b9\u6cd5\u5305\u62ec<strong>\u675f\u641c\u7d22\uff08Beam Search\uff09\u3001A* \u641c\u7d22\uff08A* Search\uff09<\/strong>\u548c<strong>\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22\uff08Monte Carlo Tree Search, MCTS\uff09<\/strong>\uff0c\u63d0\u5347\u641c\u7d22\u6548\u7387\u548c\u63a8\u7406\u8d28\u91cf\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-acb6088b170a38306bae1c237706f7bd.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-acb6088b170a38306bae1c237706f7bd.png\"><\/a><\/section>\n<p><sup>\u56fe 11\uff1a\u5916\u90e8\u63a2\u7d22\u7b56\u7565\u6839\u636e\u8fc7\u7a0b\u7ba1\u7406\u89d2\u8272\u7684\u4e0d\u540c\u53ef\u5206\u4e3a\u4e24\u7c7b\uff1a(a) \u4eba\u7c7b\u9a71\u52a8\u7684\u63a2\u7d22\uff1a\u7531\u4eba\u5de5\u8bbe\u5b9a\u7684\u63d0\u793a\u8bcd\u4e0e\u56fa\u5b9a\u6d41\u7a0b\u63a7\u5236\u63a8\u7406\u8fc7\u7a0b\uff0c\u63a2\u7d22\u7ed3\u6784\u901a\u5e38\u4e3a\u7ebf\u6027\u3001\u6811\u72b6\u6216\u56fe\u72b6\uff0c\u4f46\u8def\u5f84\u56fa\u5b9a\u3001\u7f3a\u4e4f\u9002\u5e94\u6027\u3002(b) \u6a21\u578b\u9a71\u52a8\u7684\u63a2\u7d22\uff1a\u7531\u6a21\u578b\u4e3b\u5bfc\u8fdb\u884c\u63a2\u7d22\uff0c\u5177\u5907\u53cd\u601d\u4e0e\u81ea\u9002\u5e94\u641c\u7d22\u80fd\u529b\uff0c\u91c7\u7528\u52a8\u6001\u7ed3\u6784\u5e76\u7ed3\u5408 Beam\u3001A*\u3001MCTS \u7b49\u641c\u7d22\u903b\u8f91\u4e0e\u591a\u79cd\u53cd\u9988\u673a\u5236\uff08\u5982 PRM\u3001Critic\u3001Advantage\u3001Q-Value\u3001Verifier\uff09\u6765\u5b9e\u73b0\u66f4\u7075\u6d3b\u4e14\u9ad8\u6548\u7684\u6df1\u5ea6\u63a8\u7406\u3002<\/sup><\/p>\n<p><strong>8. \u524d\u6cbf\u7814\u7a76\u4e0e\u672a\u6765\u65b9\u5411<\/strong><\/p>\n<p>\u957f\u601d\u7ef4\u94fe\u6b63\u5728\u4e0d\u65ad\u62d3\u5c55\u5176\u5e94\u7528\u8fb9\u754c\uff0c\u63a8\u52a8\u591a\u6a21\u6001\u3001\u8de8\u8bed\u8a00\u3001\u667a\u80fd\u4f53\u3001\u6548\u7387\u4f18\u5316\u3001\u77e5\u8bc6\u589e\u5f3a\u53ca\u5b89\u5168\u6027\u65b9\u9762\u7684\u7814\u7a76\u53d1\u5c55\u3002\u672a\u6765\u7684\u63a2\u7d22\u65b9\u5411\u4e3b\u8981\u6db5\u76d6\u4ee5\u4e0b\u516d\u4e2a\u5173\u952e\u9886\u57df\uff1a<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-193f7b1d9cc1cdad079ca9e15ca7721e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-193f7b1d9cc1cdad079ca9e15ca7721e.png\"><\/a><\/section>\n<p><sup>\u56fe 11\uff1a\u957f\u601d\u7ef4\u94fe\u7684\u672a\u6765\u53d1\u5c55\u65b9\u5411\u5305\u62ec\uff1a(a) \u591a\u6a21\u6001\u957f\u601d\u7ef4\u94fe\uff1a\u878d\u5408\u591a\u79cd\u6a21\u6001\u7684\u8f93\u5165\u4e0e\u8f93\u51fa\uff08\u5982\u56fe\u50cf\u3001\u6587\u5b57\u3001\u516c\u5f0f\u7b49\uff09\uff0c\u589e\u5f3a\u63a8\u7406\u8868\u73b0\u529b\u4e0e\u9002\u5e94\u6027\uff1b(b) \u591a\u8bed\u8a00\u957f\u601d\u7ef4\u94fe\uff1a\u652f\u6301\u8de8\u8bed\u8a00\u63a8\u7406\u4e0e\u5e94\u7528\uff0c\u63d0\u5347\u6a21\u578b\u7684\u591a\u8bed\u79cd\u901a\u7528\u6027\uff1b(c) \u5177\u8eab\u667a\u80fd\u4e0e\u4ee3\u7406\u578b\u957f\u601d\u7ef4\u94fe\uff1a\u901a\u8fc7\u5f15\u5165\u5177\u8eab\u4ea4\u4e92\u673a\u5236\uff0c\u63d0\u5347\u6a21\u578b\u5728\u771f\u5b9e\u73af\u5883\u4e2d\u7684\u63a8\u7406\u4e0e\u884c\u52a8\u80fd\u529b\uff1b(d) \u9ad8\u6548\u957f\u601d\u7ef4\u94fe\uff1a\u4f18\u5316\u63a8\u7406\u8def\u5f84\u4e0e\u6548\u7387\uff0c\u52a0\u5feb\u601d\u7ef4\u94fe\u751f\u6210\u901f\u5ea6\uff1b(e) \u77e5\u8bc6\u589e\u5f3a\u578b\u957f\u601d\u7ef4\u94fe\uff1a\u5f15\u5165\u5916\u90e8\u77e5\u8bc6\u5e93\u4e30\u5bcc\u63a8\u7406\u8fc7\u7a0b\uff0c\u63d0\u9ad8\u51c6\u786e\u6027\u4e0e\u89e3\u91ca\u6027\uff1b(f) \u5b89\u5168\u6027\u4fdd\u969c\u7684\u957f\u601d\u7ef4\u94fe\uff1a\u5f3a\u5316\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u53ef\u9760\u6027\uff0c\u907f\u514d\u4ea7\u751f\u6709\u5bb3\u6216\u8bef\u5bfc\u6027\u7684\u8f93\u51fa\uff0c\u786e\u4fdd\u4f26\u7406\u5408\u89c4\u3002<\/sup><\/p>\n<p><strong>8.1 \u591a\u6a21\u6001\u957f\u601d\u7ef4\u94fe\uff08Multimodal Long CoT\uff09<\/strong><\/p>\n<p>\u591a\u6a21\u6001\u63a8\u7406\u7684\u7814\u7a76\u805a\u7126\u4e8e\u5c06\u957f\u601d\u7ef4\u94fe\u6269\u5c55\u81f3\u56fe\u50cf\u3001\u89c6\u9891\u3001\u97f3\u9891\u7b49\u591a\u79cd\u6570\u636e\u6a21\u6001\uff0c\u4ee5\u63d0\u5347\u6a21\u578b\u5bf9\u590d\u6742\u4efb\u52a1\u7684\u7406\u89e3\u548c\u63a8\u7406\u80fd\u529b\u3002\u7814\u7a76\u65b9\u5411\u4e3b\u8981\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u591a\u6a21\u6001\u63d0\u793a\uff08Prompting\uff09<\/strong>\uff1a\u6539\u8fdb\u89c6\u89c9 LLMs \u7684\u4e0a\u4e0b\u6587\u63cf\u8ff0\u80fd\u529b\uff0c\u5e76\u7ed3\u5408\u81ea\u53cd\u9988\u4f18\u5316\u63a8\u7406\u957f\u5ea6\u4e0e\u51c6\u786e\u6027\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u591a\u6a21\u6001\u84b8\u998f\uff08Imitation\uff09<\/strong>\uff1a\u5229\u7528\u6570\u636e\u84b8\u998f\u7b56\u7565\u8bad\u7ec3\u6a21\u578b\uff0c\u4f7f\u5176\u80fd\u591f\u6a21\u4eff\u957f\u601d\u7ef4\u94fe\u8fc7\u7a0b\uff0c\u63d0\u9ad8\u590d\u6742\u95ee\u9898\u6c42\u89e3\u80fd\u529b\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u57fa\u4e8e\u5956\u52b1\u6a21\u578b\u7684\u63a8\u7406\u63a2\u7d22\uff08Reward Model-Based Exploration\uff09<\/strong>\uff1a\u901a\u8fc7\u5956\u52b1\u6a21\u578b\u4f18\u5316\u63a8\u7406\u8fc7\u7a0b\uff0c\u5728\u63a2\u7d22\u548c\u8bad\u7ec3\u9636\u6bb5\u63d0\u5347\u591a\u6a21\u6001\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u4e3b\u8981\u6311\u6218\u5305\u62ec<strong>\u5982\u4f55\u6709\u6548\u878d\u5408\u89c6\u89c9\u4fe1\u606f\u8f85\u52a9\u903b\u8f91\u63a8\u7406\uff0c\u4ee5\u53ca\u5982\u4f55\u4f18\u5316\u957f\u601d\u7ef4\u94fe\u7684\u63a8\u7406\u957f\u5ea6\u548c\u6d4b\u8bd5\u65f6\u6269\u5c55\u80fd\u529b\u3002<\/strong><\/p>\n<p><strong>8.2 \u591a\u8bed\u8a00\u957f\u601d\u7ef4\u94fe\uff08Multilingual Long CoT\uff09<\/strong><\/p>\n<p>\u5c3d\u7ba1\u63a8\u7406\u5927\u6a21\u578b\u5728\u82f1\u8bed\u9886\u57df\u53d6\u5f97\u4e86\u663e\u8457\u8fdb\u5c55\uff0c\u4f46\u8981\u5b9e\u73b0\u771f\u6b63\u7684\u901a\u7528\u667a\u80fd\uff0c\u591a\u8bed\u8a00\u63a8\u7406\u80fd\u529b\u4ecd\u9700\u8fdb\u4e00\u6b65\u589e\u5f3a\u3002\u5f53\u524d\u7814\u7a76\u4e3b\u8981\u96c6\u4e2d\u5728\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u8de8\u8bed\u8a00\u63d0\u793a\uff08Multilingual Prompting\uff09<\/strong>\uff1a\u4f7f\u7528\u901a\u7528\u6a21\u677f\u63d0\u793a\uff0c\u4fc3\u8fdb\u4e0d\u540c\u8bed\u8a00\u95f4\u7684\u903b\u8f91\u63a8\u7406\u80fd\u529b\u5bf9\u9f50\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u591a\u8bed\u8a00\u8bad\u7ec3\uff08Multilingual Training\uff09<\/strong>\uff1a\u91c7\u7528\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u6216\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u8bad\u7ec3\uff0c\u63d0\u5347\u4f4e\u8d44\u6e90\u8bed\u8a00\u7684\u63a8\u7406\u4e00\u81f4\u6027\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6d4b\u8bd5\u65f6\u6269\u5c55\uff08Test-Time Scaling\uff09<\/strong>\uff1a\u5229\u7528\u591a\u8bed\u8a00\u6811\u641c\u7d22\uff08Tree Search\uff09\u7b49\u65b9\u6cd5\uff0c\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u52a8\u6001\u9009\u62e9\u6700\u4f73\u8bed\u8a00\uff0c\u63d0\u5347\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u672a\u6765\u6311\u6218\u5305\u62ec<strong>\u8de8\u8bed\u8a00\u77e5\u8bc6\u8fc1\u79fb\u53ca\u4f4e\u8d44\u6e90\u8bed\u8a00\u7684\u63a8\u7406\u80fd\u529b\u4f18\u5316\uff0c\u4ee5\u786e\u4fdd\u591a\u8bed\u8a00\u63a8\u7406\u5728\u4e0d\u540c\u6587\u5316\u80cc\u666f\u4e0b\u7684\u4e00\u81f4\u6027\u3002<\/strong><\/p>\n<p><strong>8.3 \u957f\u601d\u7ef4\u94fe\u52a9\u529b\u667a\u80fd\u4f53\u4e0e\u5177\u8eab\u667a\u80fd\uff08Long CoT for Agentic &amp; Embodied AI\uff09<\/strong><\/p>\n<p>\u667a\u80fd\u4f53\u4e0e\u5177\u8eab\u667a\u80fd\u7cfb\u7edf\u4e0d\u4ec5\u4ec5\u9700\u8981\u901a\u8fc7\u957f\u601d\u7ef4\u94fe\u8fdb\u884c\u4efb\u52a1\u63a8\u7406\uff0c\u8fd8\u9700\u8981\u5728\u52a8\u6001\u73af\u5883\u4e2d\u901a\u8fc7\u4e0d\u65ad\u4e0e\u73af\u5883\u4ea4\u4e92\uff0c\u9010\u6b65\u8c03\u6574\u548c\u4f18\u5316\u957f\u601d\u7ef4\u94fe\uff0c\u4fee\u6b63\u63a8\u7406\u8def\u5f84\u5e76\u5b9e\u73b0\u6709\u6548\u7684\u76ee\u6807\u5bfc\u5411\u884c\u4e3a\u3002\u5f53\u524d\u7684\u7814\u7a76\u65b9\u5411\u4e3b\u8981\u96c6\u4e2d\u5728\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u57fa\u4e8e\u6811\u641c\u7d22\u7684\u63a8\u7406\u589e\u5f3a<\/strong>\uff1a\u901a\u8fc7\u5f15\u5165\u6811\u641c\u7d22\u7b97\u6cd5\uff08\u5982 MCTS\uff09\uff0c\u5bf9\u63a8\u7406\u8fc7\u7a0b\u8fdb\u884c\u4f18\u5316\uff0c\u667a\u80fd\u4f53\u548c\u5177\u8eab\u673a\u5668\u4eba\u80fd\u591f\u7ed3\u5408\u957f\u601d\u7ef4\u94fe\uff0c\u66f4\u52a0\u9ad8\u6548\u5730\u63a2\u7d22\u72b6\u6001\u7a7a\u95f4\uff0c\u9010\u6b65\u903c\u8fd1\u6700\u4f18\u89e3\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u73af\u5883\u4ea4\u4e92\u4f18\u5316<\/strong>\uff1a\u667a\u80fd\u4f53\u548c\u5177\u8eab\u673a\u5668\u4eba\u901a\u8fc7\u4e0e\u73af\u5883\u7684\u4e92\u52a8\uff0c\u4e0d\u65ad\u66f4\u65b0\u5386\u53f2\u8bb0\u5fc6\uff0c\u5e76\u6839\u636e\u52a8\u6001\u53d8\u5316\u7684\u73af\u5883\u8fdb\u884c\u81ea\u9002\u5e94\u8c03\u6574\u957f\u601d\u7ef4\u94fe\u7684\u63a8\u7406\u5185\u5bb9\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u591a\u667a\u80fd\u4f53\u534f\u4f5c<\/strong>\uff1a\u4e3a\u4e86\u5e94\u5bf9\u66f4\u590d\u6742\u7684\u4efb\u52a1\uff0c\u4e00\u4e9b\u7814\u7a76\u6b63\u5728\u63a2\u7d22\u591a\u4e2a\u80fd\u591f\u8fdb\u884c\u957f\u601d\u7ef4\u94fe\u63a8\u7406\u7684\u9ad8\u7ea7\u667a\u80fd\u4f53\u548c\u5177\u8eab\u673a\u5668\u4eba\u534f\u540c\u5de5\u4f5c\u4ee5\u63d0\u9ad8\u5de5\u4f5c\u6548\u7387\u3002\u901a\u8fc7\u591a\u667a\u80fd\u4f53\u4e4b\u95f4\u7684\u4fe1\u606f\u5171\u4eab\u4e0e\u957f\u601d\u7ef4\u94fe\u63a8\u7406\u534f\u4f5c\uff0c\u5171\u540c\u89e3\u51b3\u5355\u4e00\u667a\u80fd\u4f53\u96be\u4ee5\u5e94\u5bf9\u7684\u6311\u6218\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u4e3b\u8981\u6311\u6218\u5305\u62ec<strong>\u5728\u4e0d\u786e\u5b9a\u73af\u5883\u4e2d\u4fdd\u6301\u51b3\u7b56\u7a33\u5065\u6027<\/strong>\uff0c\u4ee5\u53ca<strong>\u5728\u591a\u667a\u80fd\u4f53\u534f\u4f5c\u8fc7\u7a0b\u4e2d\u4f18\u5316\u63a8\u7406\u6548\u7387\u3002<\/strong><\/p>\n<p><strong>8.4 \u9ad8\u6548\u957f\u601d\u7ef4\u94fe\uff08Efficient Long CoT\uff09<\/strong><\/p>\n<p>\u957f\u601d\u7ef4\u94fe\u7684\u6df1\u5ea6\u3001\u63a2\u7d22\u6027\u548c\u53cd\u601d\u673a\u5236\u5f80\u5f80\u5bfc\u81f4\u63a8\u7406\u8fc7\u7a0b\u5197\u957f\uff0c\u56e0\u6b64\u4f18\u5316\u63a8\u7406\u901f\u5ea6\u6210\u4e3a\u5173\u952e\u95ee\u9898\u3002\u5f53\u524d\u7814\u7a76\u65b9\u5411\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u63a8\u7406\u94fe\u538b\u7f29\uff08Direct Compression\uff09<\/strong>\uff1a\u901a\u8fc7\u526a\u679d\u3001\u8df3\u6b65\u6216\u76f4\u63a5\u9884\u6d4b\u51cf\u5c11\u5197\u4f59\u63a8\u7406\u8fc7\u7a0b\uff0c\u63d0\u9ad8\u63a8\u7406\u6548\u7387\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u9690\u85cf\u7a7a\u95f4\u63a8\u7406\uff08Hidden Space Reasoning\uff09<\/strong>\uff1a\u5728\u6f5c\u7a7a\u95f4\u4e2d\u6267\u884c\u63a8\u7406\uff0c\u907f\u514d\u663e\u5f0f\u89e3\u7801\uff0c\u4ee5\u51cf\u5c11\u8ba1\u7b97\u8d44\u6e90\u6d88\u8017\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u672a\u6765\u6311\u6218\u5305\u62ec\u5982\u4f55\u6784\u5efa\u81ea\u9002\u5e94\u63a8\u7406\u7b56\u7565\uff0c\u4f7f<strong>\u6a21\u578b\u80fd\u591f\u6839\u636e\u4efb\u52a1\u590d\u6742\u6027\u52a8\u6001\u8c03\u6574\u63a8\u7406\u6df1\u5ea6<\/strong>\uff0c\u5e76<strong>\u63a2<\/strong><strong>\u7d22\u66f4\u9ad8\u6548\u7684\u63a8\u7406\u683c\u5f0f\uff0c\u5982\u591a\u6a21\u6001\u6216\u9690\u53d8\u91cf\u63a8\u7406\u3002<\/strong><\/p>\n<p><strong>8.5 \u77e5\u8bc6\u589e\u5f3a\u957f\u601d\u7ef4\u94fe\uff08Knowledge-Augmented Long CoT\uff09<\/strong><\/p>\n<p>\u5c3d\u7ba1\u63a8\u7406\u5927\u6a21\u578b\u5177\u6709\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u5728\u67d0\u4e9b\u4e13\u4e1a\u9886\u57df\u4ecd\u7136\u5b58\u5728\u77e5\u8bc6\u7a7a\u767d\uff0c\u56e0\u6b64\u7ed3\u5408\u5916\u90e8\u77e5\u8bc6\u5e93\u6210\u4e3a\u5173\u952e\u7814\u7a76\u65b9\u5411\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08Retrieval-Augmented Generation, RAG\uff09<\/strong>\uff1a\u52a8\u6001\u7ed3\u5408\u77e5\u8bc6\u68c0\u7d22\u4e0e\u6587\u6863\u63a8\u7406\uff0c\u4ee5\u63d0\u5347\u590d\u6742\u4efb\u52a1\u7684\u6c42\u89e3\u80fd\u529b\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u6a21\u578b\u77e5\u8bc6\u6ce8\u5165\uff08Model Knowledge Injection\uff09<\/strong>\uff1a\u5728 SFT \u6216 RL \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5f15\u5165\u5916\u90e8\u77e5\u8bc6\uff0c\u4ee5\u63d0\u5347\u4e13\u4e1a\u9886\u57df\u7684\u63a8\u7406\u8868\u73b0\uff0c\u5982\u533b\u5b66\u63a8\u7406\u4efb\u52a1\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u4e3b\u8981\u6311\u6218\u5728\u4e8e\u5982\u4f55\u6709\u6548\u5730\u5c06\u66f4\u591a\u5916\u90e8\u77e5\u8bc6\u878d\u5165\u5230\u957f\u601d\u7ef4\u94fe\u7684\u63a8\u7406\u8fc7\u7a0b\uff0c\u5e76<strong>\u5f00\u53d1\u53ef\u6269\u5c55\u7684\u77e5\u8bc6\u5b58\u50a8\u4e0e\u68c0\u7d22\u673a\u5236<\/strong>\uff0c\u4ee5<strong>\u786e\u4fdd\u5b9e\u65f6\u66f4\u65b0\u548c\u77e5\u8bc6\u4e00\u81f4\u6027\u3002<\/strong><\/p>\n<p><strong>8.6 \u957f\u601d\u7ef4\u94fe\u7684\u5b89\u5168\u6027\uff08Safety in Long CoT\uff09<\/strong><\/p>\n<p>\u957f\u601d\u7ef4\u94fe\u867d\u7136\u63d0\u5347\u4e86\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u4e5f\u589e\u52a0\u4e86\u6f5c\u5728\u7684\u5b89\u5168\u98ce\u9669\uff0c\u4f8b\u5982\u8bef\u5bfc\u6027\u4fe1\u606f\u548c\u6709\u5bb3\u5185\u5bb9\u7684\u751f\u6210\u3002\u5f53\u524d\u7814\u7a76\u4e3b\u8981\u5173\u6ce8\uff1a<\/p>\n<ul>\n<li>\n<p><strong>\u63a8\u7406\u653b\u51fb\uff08\u957f\u601d\u7ef4\u94fe Attack\uff09<\/strong>\uff1a\u7814\u7a76\u53d1\u73b0\u957f\u601d\u7ef4\u94fe\u53ef\u80fd\u4f7f\u6a21\u578b\u66f4\u6613\u53d7\u5230\u9519\u8bef\u63a8\u7406\u6216\u4e0d\u5b89\u5168\u8f93\u51fa\u7684\u5f71\u54cd\uff0c\u4f8b\u5982 &#8220;OverThink \u653b\u51fb&#8221; \u548c &#8220;\u8fed\u4ee3\u6df7\u4e71\u653b\u51fb&#8221;\u3002<\/p>\n<\/li>\n<li>\n<p><strong>\u5b89\u5168\u4f18\u5316\uff08\u957f\u601d\u7ef4\u94fe Safety Improvement\uff09<\/strong>\uff1a\u5f15\u5165\u52a8\u6001\u5b89\u5168\u63d0\u793a\u3001\u504f\u89c1\u68c0\u6d4b\u3001\u5f3a\u5316\u5b66\u4e60\u7b49\u65b9\u6cd5\uff0c\u4ee5\u51cf\u5c11\u6709\u5bb3\u8f93\u51fa\u7684\u6982\u7387\uff0c\u63d0\u9ad8\u6a21\u578b\u7684\u53ef\u9760\u6027\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u672a\u6765\u6311\u6218\u5305\u62ec\u5982\u4f55<strong>\u5e73\u8861\u63a8\u7406\u80fd\u529b\u4e0e\u5b89\u5168\u6027<\/strong>\uff0c\u9632\u6b62\u957f\u601d\u7ef4\u94fe\u5bfc\u81f4\u7684\u8ba4\u77e5\u8fc7\u8f7d\uff0c\u540c\u65f6\u786e\u4fdd<strong>\u63a8\u7406\u7684\u53ef\u89e3\u91ca\u6027\u548c\u53ef\u63a7\u6027<\/strong>\uff0c\u4ee5\u5e94\u5bf9\u6f5c\u5728\u7684\u5bf9\u6297\u6027\u653b\u51fb\u3002<\/p>\n<p><strong>9. \u603b\u7ed3<\/strong><\/p>\n<p>\u672c\u6587\u7cfb\u7edf\u56de\u987e\u4e86\u957f\u601d\u7ef4\u94fe\u5728\u63a8\u7406\u5927\u6a21\u578b\u4e2d\u7684\u6838\u5fc3\u4f5c\u7528\uff0c\u68b3\u7406\u4e86\u5176\u4ece\u63d0\u51fa\u5230\u53d1\u5c55\u7684\u5173\u952e\u8109\u7edc\u3002\u8be5\u8bba\u6587\u901a\u8fc7 900 \u7bc7\u53c2\u8003\u6587\u732e\uff0c\u6784\u5efa\u4e86\u4e00\u4e2a\u8d2f\u7a7f\u7406\u8bba\u4e0e\u5b9e\u8df5\u7684\u7cfb\u7edf\u6027\u6846\u67b6\uff0c\u9996\u6b21\u6e05\u6670\u5730\u533a\u5206\u957f\u601d\u7ef4\u94fe\u4e0e\u77ed\u601d\u7ef4\u94fe\u5728\u63a8\u7406\u7ed3\u6784\u3001\u601d\u7ef4\u80fd\u529b\u4e0e\u53cd\u9988\u673a\u5236\u4e0a\u7684\u672c\u8d28\u5dee\u5f02\uff0c\u63d0\u51fa \u201c\u6df1\u5ea6\u63a8\u7406\u3001\u5e7f\u6cdb\u63a2\u7d22\u3001\u53ef\u884c\u53cd\u601d\u201d \u4e09\u5927\u6838\u5fc3\u80fd\u529b\u6a21\u578b\uff0c\u5e76\u636e\u6b64\u89e3\u6790\u4e86\u516d\u5927\u5173\u952e\u63a8\u7406\u73b0\u8c61\uff08\u5982\u63a8\u7406\u8fb9\u754c\u3001\u8fc7\u5ea6\u601d\u8003\u3001\u987f\u609f\u65f6\u523b\u7b49\uff09\uff0c\u9610\u91ca\u5176\u5e95\u5c42\u673a\u5236\u4e0e\u884c\u4e3a\u8868\u73b0\u3002<\/p>\n<p>\u66f4\u91cd\u8981\u7684\u662f\uff0c\u8be5\u8bba\u6587\u4e0d\u4ec5\u4ec5\u505c\u7559\u5728\u73b0\u8c61\u63cf\u8ff0\uff0c\u800c\u662f\u8fdb\u4e00\u6b65\u5c55\u5f00\u4e86\u65b9\u6cd5\u8bba\u603b\u7ed3\uff0c\u5168\u9762\u6574\u7406\u4e86\u5f53\u524d\u4e3b\u6d41\u7684\u957f\u601d\u7ef4\u94fe\u4f18\u5316\u7b56\u7565\uff0c\u5305\u62ec\u57fa\u4e8e\u683c\u5f0f\u8bbe\u8ba1\u7684\u6df1\u5ea6\u63a8\u7406\u3001\u57fa\u4e8e\u6a21\u4eff\u4e0e\u81ea\u5b66\u4e60\u7684\u80fd\u529b\u63d0\u5347\u3001\u57fa\u4e8e\u53cd\u9988\u4e0e\u5956\u52b1\u7684\u8fc7\u7a0b\u7ec6\u5316\u3001\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u4e0e\u5916\u90e8\u5f15\u5bfc\u7684\u63a2\u7d22\u673a\u5236\u7b49\uff0c\u5e76\u5c55\u671b\u4e86\u591a\u6a21\u6001\u3001\u8de8\u8bed\u8a00\u3001\u667a\u80fd\u4f53\u4ea4\u4e92\u3001\u6548\u7387\u4f18\u5316\u3001\u77e5\u8bc6\u589e\u5f3a\u4e0e\u5b89\u5168\u4fdd\u969c\u7b49\u672a\u6765\u7814\u7a76\u65b9\u5411\u3002<\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-04-16\" target=\"_blank\">\u8fc8\u5411\u63a8\u7406\u65f6\u4ee3\uff1a900+\u7bc7\u53c2\u8003\u6587\u732e\u63ed\u793a\u957f\u94fe\u601d\u7ef4\u7684\u524d\u4e16\u4eca\u751f\uff0c\u6700\u5168\u7efc\u8ff0\u6765\u4e86<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u8fc8\u5411\u63a8\u7406\u65f6\u4ee3\uff1a900+ [&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":[70,71,73,68,81,82],"class_list":["post-40181","post","type-post","status-publish","format-standard","hentry","category-news","tag-agent","tag-rag","tag-73","tag-68","tag-81","tag-82"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u8fc8\u5411\u63a8\u7406\u65f6\u4ee3\uff1a900+\u7bc7\u53c2\u8003\u6587\u732e\u63ed\u793a\u957f\u94fe\u601d\u7ef4\u7684\u524d\u4e16\u4eca\u751f\uff0c\u6700\u5168\u7efc\u8ff0\u6765\u4e86 - \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=40181\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=40181\",\"url\":\"https:\/\/17aitech.com\/?p=40181\",\"name\":\"\u8fc8\u5411\u63a8\u7406\u65f6\u4ee3\uff1a900+\u7bc7\u53c2\u8003\u6587\u732e\u63ed\u793a\u957f\u94fe\u601d\u7ef4\u7684\u524d\u4e16\u4eca\u751f\uff0c\u6700\u5168\u7efc\u8ff0\u6765\u4e86 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=40181#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=40181#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-fa364dae7f3411d036bb9f46c3f4284d.png\",\"datePublished\":\"2025-04-16T21:02:42+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=40181#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=40181\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=40181#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-fa364dae7f3411d036bb9f46c3f4284d.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-fa364dae7f3411d036bb9f46c3f4284d.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=40181#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u8fc8\u5411\u63a8\u7406\u65f6\u4ee3\uff1a900+\u7bc7\u53c2\u8003\u6587\u732e\u63ed\u793a\u957f\u94fe\u601d\u7ef4\u7684\u524d\u4e16\u4eca\u751f\uff0c\u6700\u5168\u7efc\u8ff0\u6765\u4e86\"}]},{\"@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":"\u8fc8\u5411\u63a8\u7406\u65f6\u4ee3\uff1a900+\u7bc7\u53c2\u8003\u6587\u732e\u63ed\u793a\u957f\u94fe\u601d\u7ef4\u7684\u524d\u4e16\u4eca\u751f\uff0c\u6700\u5168\u7efc\u8ff0\u6765\u4e86 - \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=40181","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=40181","url":"https:\/\/17aitech.com\/?p=40181","name":"\u8fc8\u5411\u63a8\u7406\u65f6\u4ee3\uff1a900+\u7bc7\u53c2\u8003\u6587\u732e\u63ed\u793a\u957f\u94fe\u601d\u7ef4\u7684\u524d\u4e16\u4eca\u751f\uff0c\u6700\u5168\u7efc\u8ff0\u6765\u4e86 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=40181#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=40181#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-fa364dae7f3411d036bb9f46c3f4284d.png","datePublished":"2025-04-16T21:02:42+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=40181#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=40181"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=40181#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-fa364dae7f3411d036bb9f46c3f4284d.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/04\/frc-fa364dae7f3411d036bb9f46c3f4284d.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=40181#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u8fc8\u5411\u63a8\u7406\u65f6\u4ee3\uff1a900+\u7bc7\u53c2\u8003\u6587\u732e\u63ed\u793a\u957f\u94fe\u601d\u7ef4\u7684\u524d\u4e16\u4eca\u751f\uff0c\u6700\u5168\u7efc\u8ff0\u6765\u4e86"}]},{"@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\/40181","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=40181"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/40181\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=40181"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=40181"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=40181"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}