{"id":35144,"date":"2024-12-19T11:59:53","date_gmt":"2024-12-19T03:59:53","guid":{"rendered":"https:\/\/17aitech.com\/?p=35144"},"modified":"2024-12-19T11:59:53","modified_gmt":"2024-12-19T03:59:53","slug":"%e4%b8%a4%e4%b8%aa%e5%b0%8f%e6%a8%a1%e5%9e%8b%e4%ba%92%e7%9b%b8%e9%aa%8c%e8%af%81%ef%bc%8c%e7%9b%b4%e6%8e%a5%e6%af%94%e8%82%a9%e5%a4%a7%e6%a8%a1%e5%9e%8b%ef%bc%9f%e5%be%ae%e8%bd%af%e7%9a%84rstar","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=35144","title":{"rendered":"\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4\u63a5\u6bd4\u80a9\u5927\u6a21\u578b\uff1f\u5fae\u8f6f\u7684rStar\u751a\u81f3\u6ca1\u7528CoT\u548c\u5fae\u8c03"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-08-16\" target=\"_blank\">\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4\u63a5\u6bd4\u80a9\u5927\u6a21\u578b\uff1f\u5fae\u8f6f\u7684rStar\u751a\u81f3\u6ca1\u7528CoT\u548c\u5fae\u8c03<\/a><\/p>\n<section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"17\" data-source-title=\"\" data-type=\"2\" data-url=\"\"><p>\u4e92\u76f8\u68c0\u67e5\uff0c\u8ba9\u5c0f\u6a21\u578b\u4e5f\u80fd\u89e3\u51b3\u5927\u95ee\u9898\u3002<\/p><\/blockquote>\n<\/section>\n<p>\u4f17\u6240\u5468\u77e5\uff0cLLM \u5f88\u5f3a\u5927\uff0c\u4f46\u6267\u884c\u590d\u6742\u63a8\u7406\u7684\u80fd\u529b\u8fd8\u4e0d\u591f\u5f3a\u3002<\/p>\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5728 GSM8K \u6570\u636e\u96c6\u4e0a\uff0cMistral-7B \u5373\u4f7f\u4f7f\u7528\u601d\u7ef4\u94fe\uff08CoT\uff09\u7b49\u6280\u672f\uff0c\u4e5f\u53ea\u80fd\u8fbe\u5230 36.5% \u7684\u51c6\u786e\u5ea6\u3002\u5c3d\u7ba1\u5fae\u8c03\u786e\u5b9e\u4e5f\u80fd\u6709\u6548\u5730\u63d0\u5347\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u5927\u591a\u6570 LLM \u4f9d\u9760\u7684\u5fae\u8c03\u6570\u636e\u90fd\u662f\u7ecf\u8fc7 GPT-4 \u7b49\u66f4\u5f3a\u5927\u6a21\u578b\u84b8\u998f\u8fc7\u7684\uff0c\u751a\u81f3\u53ef\u80fd\u539f\u672c\u5c31\u662f\u8fd9\u4e9b\u5f3a\u5927\u6a21\u578b\u5408\u6210\u7684\u3002<\/p>\n<p>\u540c\u65f6\uff0c\u7814\u7a76\u8005\u4eec\u4e5f\u5728\u79ef\u6781\u5f00\u53d1\u4e00\u79cd\u80fd\u63d0\u4f9b\u8f85\u52a9\u4f46\u4e5f\u66f4\u56f0\u96be\u7684\u65b9\u6cd5\uff1a\u4f7f\u7528\u4e00\u4e2a\u66f4\u4f18\u7684\u6559\u5e08 LLM \u6765\u63d0\u5347\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<p>\u4e3a\u4e86\u5728\u6ca1\u6709\u66f4\u4f18\u6a21\u578b\u7684\u524d\u63d0\u4e0b\u63d0\u5347\u63a8\u7406\u80fd\u529b\uff0c\u4e00\u79cd\u9887\u6709\u5e0c\u671b\u7684\u8303\u5f0f\u662f\u5229\u7528 LLM \u81ea\u8eab\u4e4b\u4e2d\u7684\u77e5\u8bc6\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u4e00\u79cd\u540d\u4e3a RAP \u7684\u65b9\u6cd5\u91c7\u7528\u4e86\u4e00\u79cd\u81ea\u6211\u63a2\u7d22\u5f0f\u7684\u89e3\u51b3\u65b9\u6cd5\uff0c\u5373\u901a\u8fc7\u81ea\u6211\u5956\u52b1\u7684\u53cd\u9988\u6765\u8fed\u4ee3\u5f0f\u5730\u63d0\u5347 LLM \u7684\u63a8\u7406\u6027\u80fd\u3002\u4e0d\u5e78\u7684\u662f\uff0c\u7814\u7a76\u8868\u660e\u8fd9\u4e00\u8303\u5f0f\u5177\u6709\u4e24\u5927\u6839\u672c\u6027\u95ee\u9898\u3002<\/p>\n<p>\u7b2c\u4e00\uff0c\u5728\u6267\u884c\u63a8\u7406\u65f6\uff0cLLM \u5f80\u5f80\u96be\u4ee5\u6709\u6548\u5730\u63a2\u7d22\u89e3\u7b54\u7a7a\u95f4\u3002\u8fd9\u79cd\u81ea\u6211\u63a2\u7d22\u5f0f\u65b9\u6cd5\u5f80\u5f80\u4f1a\u56e0\u63a8\u7406\u6b65\u9aa4\u8d28\u91cf\u4e0d\u4f73\u800c\u53d7\u56f0\u4e8e\u67d0\u4e2a\u89e3\u7b54\u7a7a\u95f4\uff0c\u5373\u4f7f\u591a\u6b21\u5c1d\u8bd5\u4e5f\u662f\u5982\u6b64\u3002<\/p>\n<p>\u7b2c\u4e8c\uff0c\u5373\u4f7f\u81ea\u6211\u63a2\u7d22\u627e\u5230\u4e86\u9ad8\u8d28\u91cf\u7684\u63a8\u7406\u6b65\u9aa4\uff0c\u5c0f\u7248\u672c\u7684\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08SLM\uff09\u4e5f\u96be\u4ee5\u8fa8\u522b\u54ea\u4e9b\u63a8\u7406\u6b65\u9aa4\u7684\u8d28\u91cf\u66f4\u9ad8\uff0c\u4e5f\u96be\u4ee5\u786e\u5b9a\u6700\u7ec8\u7b54\u6848\u662f\u5426\u6b63\u786e\uff0c\u7531\u6b64\u96be\u4ee5\u6709\u6548\u5730\u5f15\u5bfc\u81ea\u6211\u63a2\u7d22\u3002\u7814\u7a76\u8868\u660e\uff0c\u57fa\u4e8e\u57fa\u672c\u7684\u5e38\u89c4\u5956\u52b1\u7684\u81ea\u6211\u63a2\u7d22\u5f15\u5bfc\u5f97\u5230\u7684\u7ed3\u679c\u5e76\u4e0d\u6bd4\u968f\u673a\u731c\u6d4b\u66f4\u597d\u3002<\/p>\n<p>\u66f4\u9ebb\u70e6\u7684\u662f\uff0c\u5c0f\u7248\u672c\u7684\u5927\u578b<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff08SLM\uff09\u66f4\u5bb9\u6613\u51fa\u73b0\u4e0a\u8ff0\u4e24\u4e2a\u95ee\u9898\uff0c\u56e0\u4e3a\u5b83\u4eec\u7684\u80fd\u529b\u66f4\u5dee\u4e00\u4e9b\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0cGPT-4 \u80fd\u901a\u8fc7\u81ea\u6211\u4f18\u5316\u6765\u63d0\u5347\u8f93\u51fa\u7ed3\u679c\uff0c\u4f46 SLM \u5374\u5f88\u96be\u505a\u5230\u8fd9\u4e00\u70b9\uff0c\u751a\u81f3\u53ef\u80fd\u5bfc\u81f4\u8f93\u51fa\u7ed3\u679c\u8d28\u91cf\u4e0b\u964d\u3002\u8fd9\u4f1a\u4e25\u91cd\u59a8\u788d\u795e\u7ecf<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u7684\u63a8\u5e7f\u5e94\u7528\u3002<\/p>\n<p>\u9488\u5bf9\u8fd9\u4e9b\u95ee\u9898\uff0c<mark data-type=\"institutions\" data-id=\"2c9b618a-41dc-4938-9d96-c5bd8967f08d\">\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662<\/mark>\u548c\u54c8\u4f5b\u5927\u5b66\u7684\u4e00\u4e2a\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86 Self-play muTuAl Reasoning\uff0c\u5373\u81ea\u535a\u5f08\u76f8\u4e92\u63a8\u7406\uff0c\u7b80\u79f0 rStar\u3002\u7b80\u5355\u6765\u8bf4\uff0c\u8be5\u65b9\u6cd5\u5c31\u7c7b\u4f3c\u4e8e\u8ba9\u4e24\u4e2a\u5b66\u4e60\u5e73\u5e73\u7684\u4eba\u4e92\u76f8\u68c0\u67e5\u8003\u5377\u7b54\u6848\uff0c\u6700\u7ec8\u63d0\u5347\u5f97\u5206\uff0c\u751a\u81f3\u8fbe\u5230\u6bd4\u80a9\u5b66\u9738\u7684\u7a0b\u5ea6\u3002\u8be5\u56e2\u961f\u5ba3\u79f0 rStar \u300c\u65e0\u9700\u5fae\u8c03\u6216\u66f4\u4f18\u6a21\u578b\u5c31\u80fd\u63d0\u5347 SLM \u7684\u63a8\u7406\u80fd\u529b\u300d\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-57b19af6a3fa257886fae56448547749.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-57b19af6a3fa257886fae56448547749.png\"><\/a><\/p>\n<ul>\n<li>\n<p>\u8bba\u6587\u6807\u9898\uff1aMutual Reasoning Makes Smaller LLMs Stronger Problem-Solvers<\/p>\n<\/li>\n<li>\n<p>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2408.06195<\/p>\n<\/li>\n<li>\n<p>\u4ee3\u7801\u5730\u5740\uff1ahttps:\/\/github.com\/zhentingqi\/rStar \uff08\u5c1a\u5f85\u53d1\u5e03\uff09<\/p>\n<\/li>\n<\/ul>\n<p><strong>\u65b9\u6cd5<\/strong><\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u96be\u9898\uff0crStar \u7684\u505a\u6cd5\u662f\u5c06\u63a8\u7406\u8fc7\u7a0b\u5206\u6210\u4e86\u89e3\u7b54\u751f\u6210\u548c\u76f8\u4e92\u9a8c\u8bc1\u4e24\u90e8\u5206\uff0c\u5982\u56fe 2 \u6240\u793a\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-68dede92d82332c3ecef2a911e70cd18.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-68dede92d82332c3ecef2a911e70cd18.png\"><\/a><\/p>\n<p>\u9488\u5bf9\u7b2c\u4e00\u4e2a\u96be\u9898\uff0c\u8be5\u56e2\u961f\u5f15\u5165\u4e86\u4e00\u4e2a\u96c6\u5408\uff0c\u5176\u4e2d\u5305\u542b\u4e30\u5bcc\u7684\u7c7b\u4f3c\u4eba\u7c7b\u7684\u63a8\u7406\u52a8\u4f5c\uff0c\u53ef\u900f\u5f7b\u5730\u63a2\u7d22\u591a\u79cd\u4e0d\u540c\u7684\u63a8\u7406\u4efb\u52a1\u7a7a\u95f4\u3002<\/p>\n<p>\u9488\u5bf9\u7b2c\u4e8c\u4e2a\u96be\u9898\uff0c\u4ed6\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u4e13\u95e8\u9488\u5bf9 SLM \u7684\u5956\u52b1\u51fd\u6570\uff0c\u8fd9\u80fd\u5bf9\u4e2d\u95f4\u6b65\u9aa4\u8fdb\u884c\u8bc4\u4f30\uff0c\u4ece\u800c\u907f\u514d\u4f9d\u8d56\u5b83\u4eec\u90a3\u5f80\u5f80\u5e76\u4e0d\u53ef\u9760\u7684\u81ea\u6211\u8bc4\u4f30\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u8be5\u56e2\u961f\u8fd8\u4f7f\u7528\u4e86\u53e6\u4e00\u4e2a SLM \u4f5c\u4e3a\u5224\u522b\u5668\u6765\u589e\u5f3a MCTS \u8fc7\u7a0b\uff0c\u4e0e\u5224\u522b\u5668 SLM \u4e92\u76f8\u9a8c\u8bc1\u6bcf\u6761\u8f68\u8ff9\u7684\u6b63\u786e\u6027\u3002<\/p>\n<p><strong>\u4f7f\u7528 MCTS Rollout \u81ea\u5df1\u751f\u6210\u63a8\u7406\u8f68\u8ff9<\/strong><\/p>\n<p>\u4e00\u4e2a\u5305\u542b\u4e30\u5bcc\u7684\u7c7b\u4eba\u63a8\u7406\u52a8\u4f5c\u7684\u96c6\u5408\u3002MCTS \u751f\u6210\u7684\u6838\u5fc3\u5728\u4e8e\u52a8\u4f5c\u7a7a\u95f4\uff0c\u5176\u5b9a\u4e49\u4e86\u6811\u63a2\u7d22\u7684\u8303\u56f4\u3002\u5927\u591a\u6570\u57fa\u4e8e MCTS \u7684\u65b9\u6cd5\u5728\u6784\u5efa\u6811\u65f6\u90fd\u4f7f\u7528\u4e86\u5355\u4e00\u52a8\u4f5c\u7c7b\u578b\u3002\u6bd4\u5982 RAP \u4e2d\u7684\u52a8\u4f5c\u662f\u63d0\u51fa\u4e0b\u4e00\u4e2a\u5b50\u95ee\u9898\uff0c\u800c AlphaMath \u548c MindStar \u4e2d\u7684\u52a8\u4f5c\u662f\u751f\u6210\u4e0b\u4e00\u63a8\u7406\u6b65\u9aa4\u3002\u4f46\u662f\uff0c\u4f9d\u8d56\u5355\u4e00\u52a8\u4f5c\u7c7b\u578b\u53ef\u80fd\u5bb9\u6613\u5bfc\u81f4\u7a7a\u95f4\u63a2\u7d22\u6548\u679c\u4e0d\u4f73\u3002<\/p>\n<p>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u8be5\u56e2\u961f\u56de\u987e\u4e86\u4eba\u7c7b\u6267\u884c\u63a8\u7406\u7684\u65b9\u6cd5\u3002\u4e0d\u540c\u7684\u4eba\u89e3\u51b3\u95ee\u9898\u7684\u65b9\u6cd5\u4e5f\u4e0d\u540c\uff1a\u67d0\u4e9b\u4eba\u4f1a\u5c06\u95ee\u9898\u5206\u89e3\u6210\u5b50\u95ee\u9898\uff0c\u53e6\u4e00\u4e9b\u5219\u4f1a\u76f4\u63a5\u89e3\u51b3\u95ee\u9898\uff0c\u8fd8\u6709\u4e9b\u4eba\u5219\u4f1a\u6362\u4e2a\u89c6\u89d2\u91cd\u65b0\u8868\u8ff0\u95ee\u9898\u3002\u6b64\u5916\uff0c\u4eba\u4eec\u8fd8\u4f1a\u6839\u636e\u5f53\u524d\u72b6\u6001\u8c03\u6574\u81ea\u5df1\u7684\u65b9\u6cd5\uff0c\u6309\u9700\u6c42\u9009\u62e9\u4e0d\u540c\u7684\u52a8\u4f5c\u3002<\/p>\n<p>\u53d7\u4eba\u7c7b\u63a8\u7406\u8fc7\u7a0b\u7684\u542f\u53d1\uff0c\u8be5\u56e2\u961f\u6784\u5efa\u4e86\u4e00\u4e2a\u66f4\u4e3a\u4e30\u5bcc\u7684\u6570\u636e\u96c6\uff0c\u5176\u4e2d\u5305\u542b 5 \u7c7b\u52a8\u4f5c\uff0c\u4ee5\u5c3d\u53ef\u80fd\u5730\u63d0\u5347 SLM \u6b63\u786e\u89e3\u51b3\u590d\u6742\u63a8\u7406\u95ee\u9898\u7684\u6f5c\u529b\u3002<\/p>\n<section>\u52a8\u4f5c 1\uff1a\u63d0\u8bae\u4e00\u6b65\u601d\u8def\u3002\u9488\u5bf9\u7ed9\u5b9a\u95ee\u9898\uff0c\u8be5\u52a8\u4f5c\u4f1a\u8ba9 LLM \u57fa\u4e8e\u5df2\u6709\u7684\u63a8\u7406\u6b65\u9aa4\u751f\u6210\u63a5\u4e0b\u6765\u7684\u4e00\u6b65\u601d\u8def\u3002<\/section>\n<section><\/section>\n<section>\u52a8\u4f5c 2\uff1a\u63d0\u8bae\u4f59\u4e0b\u7684\u601d\u8def\u6b65\u9aa4\u3002\u8be5\u52a8\u4f5c\u4e0e\u6807\u51c6 CoT \u4e00\u6837\uff0c\u80fd\u5b9e\u73b0\u300c\u5feb\u901f\u601d\u8003\u300d\uff0c\u4ece\u800c\u89e3\u51b3\u53ea\u9700\u5c11\u91cf\u6b65\u9aa4\u7684\u7b80\u5355\u95ee\u9898\u3002\u7ed9\u5b9a\u5df2\u7ecf\u751f\u6210\u7684\u63a8\u7406\u6b65\u9aa4\uff0c\u5b83\u4f1a\u8ba9 LLM \u76f4\u63a5\u751f\u6210\u5269\u4f59\u6b65\u9aa4\uff0c\u76f4\u5230\u5f97\u5230\u6700\u7ec8\u7b54\u6848\u3002<\/section>\n<section><\/section>\n<section>\u52a8\u4f5c 3\uff1a\u63d0\u8bae\u4e0b\u4e00\u4e2a\u5b50\u95ee\u9898\u53ca\u5176\u7b54\u6848\u3002<\/section>\n<section><\/section>\n<section>\u52a8\u4f5c 4\uff1a\u518d\u6b21\u56de\u7b54\u8fd9\u4e2a\u5b50\u95ee\u9898\u3002\u8003\u8651\u5230\u52a8\u4f5c 3 \u6709\u53ef\u80fd\u65e0\u6cd5\u6b63\u786e\u56de\u7b54\u5bf9\u5e94\u7684\u5b50\u95ee\u9898\uff0c\u56e0\u6b64\u8fd9\u4e2a\u52a8\u4f5c\u7684\u4f5c\u7528\u662f\u518d\u6b21\u56de\u7b54\u5b83\u3002<\/section>\n<section><\/section>\n<section>\u52a8\u4f5c 5\uff1a\u91cd\u65b0\u8868\u8ff0\u95ee\u9898 \/ \u5b50\u95ee\u9898\u3002\u8fd9\u4e2a\u65b0\u52a8\u4f5c\u662f\u4ee5\u66f4\u7b80\u5355\u7684\u65b9\u5f0f\u91cd\u65b0\u8868\u8ff0\u8be5\u95ee\u9898\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u8fd9\u91cc\u662f\u8ba9 LLM \u6e05\u6670\u5217\u51fa\u95ee\u9898\u9648\u8ff0\u4e2d\u7684\u6240\u6709\u6761\u4ef6\u3002<\/section>\n<p>\u4ee5\u4e0a\u4e94\u4e2a\u52a8\u4f5c\u5b9a\u4e49\u4e86\u4e00\u4e2a\u9ad8\u5ea6\u591a\u6837\u5316\u7684\u52a8\u4f5c\u7a7a\u95f4 {A1, A2, A3, A4, A5}\u3002<\/p>\n<p>\u5728\u6bcf\u4e2a\u6b65\u9aa4 i\uff0cMCTS \u4ece\u8be5\u7a7a\u95f4\u9009\u53d6\u4e00\u4e2a\u52a8\u4f5c a_i\u3002\u7136\u540e\u57fa\u4e8e\u5f53\u524d\u72b6\u6001\uff08\u5373\u4e4b\u524d\u751f\u6210\u7684\u8f68\u8ff9 x \u2295 s_1 \u2295 s_2 \u2295 &#8230; \u2295 s_{i\u22121}\uff09\uff0c\u4f7f\u7528\u8be5\u52a8\u4f5c a_i \u8ba9 LLM \u751f\u6210\u4e0b\u4e00\u63a8\u7406\u6b65\u9aa4 s_i\u3002\u8bf7\u6ce8\u610f\u67d0\u4e9b\u52a8\u4f5c\u9700\u8981\u6309\u987a\u5e8f\u6267\u884c\u3002\u56fe 3 \u7ed9\u51fa\u4e86\u4e00\u4e2a\u793a\u4f8b\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e113091491fd4f18c5e03a2b23078318.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-e113091491fd4f18c5e03a2b23078318.png\"><\/a><\/p>\n<p>\u5982\u8868 1 \u6240\u793a\uff0c\u5728\u63d0\u5347\u6700\u7ec8\u63a8\u7406\u51c6\u786e\u5ea6\u65b9\u9762\uff0c\u6bcf\u4e2a\u52a8\u4f5c\u90fd\u5177\u6709\u91cd\u8981\u4f5c\u7528\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-6eb0180eccc255fd43c65c4b65d3cab1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-6eb0180eccc255fd43c65c4b65d3cab1.png\"><\/a><\/p>\n<ul>\n<li>\n<p>\u5956\u52b1\u51fd\u6570<\/p>\n<\/li>\n<\/ul>\n<p>MCTS \u7684\u53e6\u4e00\u4e2a\u5173\u952e\u7ec4\u4ef6\u662f\u5956\u52b1\u51fd\u6570\uff0c\u5176\u4f5c\u7528\u662f\u8bc4\u4f30\u6bcf\u4e2a\u52a8\u4f5c\u7684\u4ef7\u503c\u5e76\u4e3a\u6811\u7684\u6269\u5c55\u63d0\u4f9b\u6307\u793a\u3002\u9488\u5bf9 SLM\uff0c\u8be5\u56e2\u961f\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u7b80\u5355\u5374\u6709\u6548\u7684\u5956\u52b1\u51fd\u6570\u3002\u4ed6\u4eec\u7684\u65b9\u6cd5\u7075\u611f\u6765\u81ea AlphaGo\uff0c\u5373\u57fa\u4e8e\u6bcf\u4e2a\u4e2d\u95f4\u8282\u70b9\u5bf9\u6700\u7ec8\u6b63\u786e\u7b54\u6848\u7684\u8d21\u732e\u5bf9\u5b83\u4eec\u8fdb\u884c\u8bc4\u5206\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u7ecf\u5e38\u5f97\u5230\u6b63\u786e\u7b54\u6848\u7684\u52a8\u4f5c\u5c31\u80fd\u83b7\u5f97\u66f4\u9ad8\u5956\u52b1\uff0c\u5b83\u4eec\u4e5f\u5c31\u66f4\u53ef\u80fd\u5728\u672a\u6765\u7684 MCTS \u6811\u6269\u5c55\u4e2d\u88ab\u9009\u53d6\u3002<\/p>\n<p>\u8fd9\u91cc\u5c06\u6267\u884c\u52a8\u4f5c a \u540e\u751f\u6210\u7684\u8282\u70b9 s \u7684\u5956\u52b1\u503c\u5b9a\u4e49\u4e3a Q (s, a)\u3002\u4e00\u5f00\u59cb\uff0c\u6240\u6709\u672a\u88ab\u63a2\u7d22\u8fc7\u7684\u8282\u70b9\u90fd\u88ab\u5206\u914d\u4e86 Q (s_i, a_i) = 0\uff0c\u4ece\u800c\u5b9e\u73b0\u968f\u673a\u7684\u6811\u6269\u5c55\u3002\u5728\u62b5\u8fbe\u9996\u4e2a\u7aef\u8282\u70b9 n_d \u65f6\uff0c\u6839\u636e\u5176\u662f\u5426\u5f97\u5230\u6b63\u786e\u7b54\u6848\u800c\u8ba1\u7b97\u4e00\u4e2a\u5956\u52b1\u5206\u6570 Q (s_d, a_d)\u3002<\/p>\n<p>\u7136\u540e\uff0c\u6cbf\u8f68\u8ff9 t = x \u2295 s_1 \u2295 s_2 \u2295 &#8230; \u2295 s_d \u5c06\u8be5\u5206\u6570\u53cd\u5411\u4f20\u64ad\u7ed9\u6bcf\u4e2a\u4e2d\u95f4\u8282\u70b9\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5bf9\u4e8e\u6bcf\u4e2a s_i\uff0c\u90fd\u4ee5\u5982\u4e0b\u65b9\u5f0f\u66f4\u65b0\u5176 Q \u503c\uff1aQ (s_i, a_i) = Q (s_i, a_i) + Q (s_d, a_d)\u3002\u4e3a\u4e86\u8ba1\u7b97\u7aef\u8282\u70b9\u7684 Q (s_d, a_d)\uff0c\u8fd9\u91cc\u4f7f\u7528\u7684\u5956\u52b1\u503c\u662f\u81ea\u6d3d\u591a\u6570\u6295\u7968\u7684\u4f3c\u7136\uff08\u7f6e\u4fe1\u5ea6\uff09\u3002<\/p>\n<ul>\n<li>\n<p>\u4f7f\u7528 MCTS Rollout \u751f\u6210\u89e3\u7b54<\/p>\n<\/li>\n<\/ul>\n<p>\u4e0b\u9762\u63cf\u8ff0 MCTS \u751f\u6210\u5019\u9009\u63a8\u7406\u8f68\u8ff9\u7684\u65b9\u5f0f\u3002\u4ece\u521d\u59cb\u7684\u6839\u8282\u70b9 s_0 \u5f00\u59cb\uff0c\u6267\u884c\u5305\u62ec\u9009\u53d6\u3001\u6269\u5c55\u3001\u6a21\u62df\u548c\u53cd\u5411\u4f20\u64ad\u5728\u5185\u7684\u591a\u79cd\u641c\u7d22\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6a21\u62df\u4f7f\u7528\u7684\u662f\u9ed8\u8ba4\u7684 Rollout \u7b56\u7565\u3002\u4e3a\u4e86\u5f97\u5230\u66f4\u51c6\u786e\u7684\u5956\u52b1\u4f30\u8ba1\uff0c\u8be5\u56e2\u961f\u4f1a\u6267\u884c\u591a\u6b21 Rollout\u3002\u4e3a\u4e86\u5e73\u8861\u63a2\u7d22\u4e0e\u5229\u7528\uff0c\u4ed6\u4eec\u4f7f\u7528\u4e86\u8457\u540d\u7684 UCT\uff08\u6811\u7684\u7f6e\u4fe1\u5ea6\u4e0a\u754c\uff09\u6765\u9009\u53d6\u6bcf\u4e2a\u8282\u70b9\u3002\u8fd9\u4e2a\u9009\u53d6\u8fc7\u7a0b\u7684\u6570\u5b66\u5f62\u5f0f\u4e3a\uff1a<\/p>\n<p>\u5176\u4e2d N (s, a) \u662f\u4e4b\u524d\u7684\u8fed\u4ee3\u4e2d\u8282\u70b9 s \u88ab\u8bbf\u95ee\u7684\u6b21\u6570\uff0cN_parent (s) \u8868\u793a\u5bf9 s \u7684\u7236\u8282\u70b9\u7684\u8bbf\u95ee\u6b21\u6570\u3002Q (s, a) \u662f\u4f30\u8ba1\u7684\u5956\u52b1\u503c\uff0c\u4f1a\u5728\u53cd\u5411\u4f20\u64ad\u8fc7\u7a0b\u4e2d\u5f97\u5230\u66f4\u65b0\u3002c \u662f\u5e73\u8861\u63a2\u7d22\u4e0e\u5229\u7528\u7684\u5e38\u91cf\u3002<\/p>\n<p>\u4e00\u65e6\u641c\u7d22\u5230\u8fbe\u67d0\u4e2a\u7aef\u8282\u70b9\uff08\u53ef\u80fd\u662f\u4e00\u4e2a\u7ec8\u7aef\u72b6\u6001\uff0c\u4e5f\u53ef\u80fd\u5230\u8fbe\u4e86\u9884\u5b9a\u4e49\u7684\u6700\u5927\u6811\u6df1\u5ea6 d\uff09\uff0c\u4fbf\u80fd\u5f97\u5230\u4e00\u6761\u4ece\u6839\u5230\u7aef\u8282\u70b9\u7684\u8f68\u8ff9\u3002\u5c06 Rollout \u8fed\u4ee3\u5f97\u5230\u7684\u6240\u6709\u8f68\u8ff9\u6536\u96c6\u8d77\u6765\u4f5c\u4e3a\u5019\u9009\u89e3\u7b54\u3002\u63a5\u4e0b\u6765\u5c31\u9700\u8981\u5bf9\u5b83\u4eec\u8fdb\u884c\u9a8c\u8bc1\u3002<\/p>\n<p><strong>\u4f7f\u7528\u4e92\u6070\u6027\u9009\u62e9\u63a8\u7406\u8f68\u8ff9<\/strong><\/p>\n<p>\u57fa\u4e8e\u6536\u96c6\u5230\u7684\u6240\u6709\u8f68\u8ff9\uff0c\u8be5\u56e2\u961f\u63d0\u51fa\u4f7f\u7528\u63a8\u7406\u4e92\u6070\u6027\u6765\u9009\u62e9\u7b54\u6848\u3002<\/p>\n<ul>\n<li>\n<p>\u901a\u8fc7\u5224\u522b\u5668 SLM \u5b9e\u73b0\u63a8\u7406\u4e92\u6070\u6027<\/p>\n<\/li>\n<\/ul>\n<p>\u5982\u56fe 2 \u6240\u793a\uff0c\u9664\u4e86\u76ee\u6807 SLM \u5916\uff0c\u8be5\u56e2\u961f\u8fd8\u5f15\u5165\u4e86\u4e00\u4e2a\u5224\u522b\u5668 SLM\uff0c\u5176\u4f5c\u7528\u662f\u4e3a\u6bcf\u4e2a\u5019\u9009\u8f68\u8ff9\u63d0\u4f9b\u5916\u90e8\u65e0\u76d1\u7763\u53cd\u9988\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u5bf9\u4e8e t = x \u2295 s_1 \u2295 s_2 \u2295 &#8230; \u2295 s_d\uff0c\u906e\u63a9\u4ece\u67d0\u4e2a\u968f\u673a\u91c7\u6837\u7684\u6b65\u9aa4 i \u5904\u5f00\u59cb\u7684\u63a8\u7406\u6b65\u9aa4\u3002\u7136\u540e\u5c06\u4e4b\u524d\u7684\u63a8\u7406\u8f68\u8ff9 t = x \u2295 s_1 \u2295 s_2 \u2295 &#8230; \u2295 s_{i-1} \u4f5c\u4e3a prompt \u63d0\u4f9b\u7ed9\u5224\u522b\u5668 SLM\uff0c\u8ba9\u5176\u8865\u5168\u5269\u4f59\u6b65\u9aa4\u3002\u7531\u4e8e\u5c06\u4e4b\u524d\u7684 i-1 \u4e2a\u63a8\u7406\u6b65\u9aa4\u4f5c\u4e3a\u4e86\u63d0\u793a\uff0c\u56e0\u6b64\u96be\u5ea6\u964d\u4f4e\u4e86\uff0c\u5224\u522b\u5668 SLM \u4fbf\u66f4\u6709\u53ef\u80fd\u7ed9\u51fa\u6b63\u786e\u7b54\u6848\u3002<\/p>\n<p>\u56fe 4 \u4e2d\u6bd4\u8f83\u4e86\u5224\u522b\u5668 SLM \u8865\u5168\u7684\u7b54\u6848\u662f\u5426\u4e0e\u539f\u59cb\u8f68\u8ff9 t \u5339\u914d\u3002\u5982\u679c\u4e24\u8005\u4e00\u81f4\uff0c\u5219\u8ba4\u4e3a t \u662f\u53ef\u4ee5\u6700\u7ec8\u9009\u62e9\u7684\u5df2\u9a8c\u8bc1\u8f68\u8ff9\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-88fe5425b8af0b1906726e4daa52c289.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-88fe5425b8af0b1906726e4daa52c289.png\"><\/a><\/p>\n<p>\u7531\u76ee\u6807 SLM \u9009\u53d6\u6700\u7ec8\u8f68\u8ff9\u3002\u5728\u5bf9\u6240\u6709\u5019\u9009\u8f68\u8ff9\u4f7f\u7528\u4e86\u63a8\u7406\u4e92\u6070\u6027\u4e4b\u540e\uff0c\u518d\u56de\u5230\u76ee\u6807 SLM\uff0c\u8ba9\u5176\u4ece\u5df2\u9a8c\u8bc1\u8f68\u8ff9\u4e2d\u9009\u51fa\u6700\u7ec8\u8f68\u8ff9\u3002\u4e3a\u4e86\u8ba1\u7b97\u6bcf\u6761\u8f68\u8ff9\u7684\u6700\u7ec8\u5206\u6570\uff0c\u8be5\u56e2\u961f\u7684\u505a\u6cd5\u662f\u7528\u5176\u5956\u52b1\u4e58\u4ee5\u901a\u8fc7 Rollout \u5f97\u5230\u7684\u5176\u7aef\u8282\u70b9\u7684\u7f6e\u4fe1\u5ea6\u5206\u6570\u3002\u6700\u7ec8\u5206\u6570\u6700\u9ad8\u7684\u8f68\u8ff9\u88ab\u9009\u4f5c\u89e3\u7b54\u3002<\/p>\n<p><strong>\u5b9e\u9a8c<\/strong><\/p>\n<p><strong>\u5b9e\u9a8c\u8bbe\u7f6e<\/strong><\/p>\n<p>rStar \u9002\u7528\u4e8e\u591a\u79cd LLM \u548c\u63a8\u7406\u4efb\u52a1\u3002\u8be5\u56e2\u961f\u8bc4\u4f30\u4e86 5 \u4e2a SLM\uff1aPhi3-mini\u3001LLaMA2-7B\u3001Mistral-7B\u3001LLaMA3-8B\u3001LLaMA3-8B-Instruct\u3002<\/p>\n<p>\u6d4b\u8bd5\u7684\u63a8\u7406\u4efb\u52a1\u6709 5 \u4e2a\uff0c\u5176\u4e2d\u5305\u62ec 4 \u4e2a\u6570\u5b66\u4efb\u52a1\uff08GSM8K\u3001GSM-Hard\u3001MATH\u3001SVAMP\uff09\u548c 1 \u4e2a\u5e38\u8bc6\u4efb\u52a1\uff08StrategyQA\uff09\u3002<\/p>\n<p>\u5b9e\u9a8c\u7ec6\u8282\u8bf7\u8bbf\u95ee\u539f\u8bba\u6587\u3002<\/p>\n<p><strong>\u4e3b\u8981\u7ed3\u679c<\/strong><\/p>\n<p>\u8be5\u56e2\u961f\u9996\u5148\u8bc4\u4f30\u4e86 rStar \u5728\u4e00\u822c\u63a8\u7406<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u4e0a\u7684\u6709\u6548\u6027\u3002\u8868 2 \u6bd4\u8f83\u4e86 rStar \u548c\u5176\u5b83\u5f53\u524d\u6700\u4f73\u65b9\u6cd5\u5728\u4e0d\u540c SLM \u548c\u63a8\u7406\u6570\u636e\u96c6\u4e0a\u7684\u51c6\u786e\u5ea6\u3002\u4e3a\u4e86\u6f14\u793a\u65b0\u751f\u6210\u5668\u7684\u6548\u679c\uff0c\u8be5\u56e2\u961f\u8fd8\u63d0\u4f9b\u4e86 rStar (generator @maj) \u7684\u51c6\u786e\u5ea6\uff0c\u5373\u4e0d\u4f7f\u7528\u5224\u522b\u5668\uff0c\u4ec5\u4f7f\u7528\u591a\u6570\u6295\u7968\u6765\u9a8c\u8bc1\u7b54\u6848\u800c\u5f97\u5230\u7684\u51c6\u786e\u5ea6\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-2eb5988cf3c69c5b78149a51f2ffcc85.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-2eb5988cf3c69c5b78149a51f2ffcc85.png\"><\/a><\/p>\n<p>\u8be5\u56e2\u961f\u6307\u51fa\u4e86\u5176\u4e2d\u7684\u4e09\u9879\u5173\u952e\u7ed3\u679c\uff1a<\/p>\n<p>1. \u5f97\u5230 rStar \u52a9\u529b\u7684 SLM \u89e3\u51b3\u95ee\u9898\u7684\u80fd\u529b\u66f4\u5f3a\u3002\u6bd4\u5982\uff0c\u5728 GSM8K \u6570\u636e\u96c6\u4e0a\uff0c\u4f7f\u7528\u5c11\u6837\u672c CoT \u7684 LLaMA2-7B \u7684\u51c6\u786e\u5ea6\u53ea\u6709 12.51%\u3002\u4f46\u6709\u4e86 rStar \u7684\u5e2e\u52a9\uff0c\u5176\u51c6\u786e\u5ea6\u63d0\u5347\u5230\u4e86 63.91%\uff0c\u8fd9\u4e00\u6210\u7ee9\u63a5\u8fd1\u4f7f\u7528\u5fae\u8c03\u5f97\u5230\u7684\u51c6\u786e\u5ea6\uff0c\u5982\u56fe 1 \u6240\u793a\u3002\u7c7b\u4f3c\u5730\uff0c\u4f7f\u7528 rStar \u7684 Mistral \u7684\u6027\u80fd\u751a\u81f3\u6bd4\u5fae\u8c03\u7248\u7684 MetaMath \u8fd8\u9ad8 4.18%\u3002\u8fd9\u6837\u7684\u63d0\u5347\u8868\u660e\uff0cSLM \u672c\u8eab\u5df2\u7ecf\u5177\u5907\u5f88\u5f3a\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u9700\u8981\u5f15\u5bfc\u624d\u80fd\u751f\u6210\u548c\u9009\u51fa\u6b63\u786e\u89e3\u7b54\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-3cfacb12da7b5f09de08c7c4b00dfc46.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-3cfacb12da7b5f09de08c7c4b00dfc46.png\"><\/a><\/p>\n<p>2.rStar \u53ef\u4ee5\u7a33\u5b9a\u5730\u5c06\u88ab\u8bc4\u4f30\u7684\u591a\u79cd SLM \u5728\u4e0d\u540c\u4efb\u52a1\u4e0a\u7684\u63a8\u7406\u51c6\u786e\u5ea6\u63d0\u5347\u81f3\u5f53\u524d\u6700\u4f73\u6c34\u5e73\u3002\u76f8\u8f83\u4e4b\u4e0b\uff0c\u5176\u5b83\u5bf9\u6bd4\u65b9\u6cd5\u90fd\u65e0\u6cd5\u7a33\u5b9a\u5730\u5728\u6240\u6709\u56db\u4e2a<mark data-type=\"concepts\" data-id=\"308c3a45-0fee-4ec6-858e-85b15f440fc0\">\u57fa\u51c6<\/mark>\u4e0a\u53d6\u5f97\u4f18\u826f\u8868\u73b0\u3002\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5c3d\u7ba1 SC\uff08\u81ea\u6211\u4e00\u81f4\u6027\uff09\u64c5\u957f\u4e09\u4e2a\u6570\u5b66\u4efb\u52a1\uff0c\u4f46\u5374\u65e0\u6cd5\u6709\u6548\u89e3\u51b3 StrategyQA \u7684<mark data-type=\"concepts\" data-id=\"95a97f4b-79d2-4bbc-91ae-300f074dff9f\">\u903b\u8f91<\/mark>\u63a8\u7406\u4efb\u52a1\u3002<\/p>\n<p>3. \u5373\u4f7f\u6ca1\u6709\u65b0\u63d0\u51fa\u7684\u7528\u4e8e\u9a8c\u8bc1\u63a8\u7406\u8f68\u8ff9\u7684\u5224\u522b\u5668\uff0c\u65b0\u63d0\u51fa\u7684 MCTS \u751f\u6210\u5668\u5728\u63d0\u5347 SLM \u7684\u63a8\u7406\u51c6\u786e\u5ea6\u65b9\u9762\u4f9d\u7136\u6548\u679c\u5f88\u597d\u3002\u6bd4\u5982\uff0c\u5728 GSM8K \u6570\u636e\u96c6\u4e0a\uff0crStar (generator @maj) \u7684\u51c6\u786e\u5ea6\u6bd4 RAP \u9ad8 2.88%-16.39%\u3001\u6bd4 ToT \u9ad8 10.60%- 38.37%\u3001\u6bd4 SC \u9ad8 1.69% &#8211; 7.34%\u3002<\/p>\n<ul>\n<li>\n<p>\u5728\u9ad8\u96be\u5ea6\u6570\u5b66\u6570\u636e\u96c6\u4e0a\u7684\u7ed3\u679c<\/p>\n<\/li>\n<\/ul>\n<p>\u8be5\u56e2\u961f\u8fd8\u5728\u4e00\u4e2a\u66f4\u9ad8\u96be\u5ea6\u7684\u6570\u5b66\u6570\u636e\u96c6\u4e0a\u8bc4\u4f30\u4e86 rStar\u3002\u4e3a\u6b64\u4ed6\u4eec\u9009\u62e9\u4e86 GSM-Hard \u548c MATH \u6570\u636e\u96c6\u3002\u9075\u7167\u540c\u7c7b\u7814\u7a76\u7684\u60ef\u4f8b\uff0c\u4ed6\u4eec\u4f7f\u7528\u4e86 MATH-500\uff0c\u8fd9\u662f\u6765\u81ea MATH \u6570\u636e\u96c6\u7684\u4e00\u4e2a\u5305\u542b\u4ee3\u8868\u6027\u95ee\u9898\u7684\u5b50\u96c6\u3002\u8fd9\u6837\u505a\u662f\u4e3a\u4e86\u63d0\u5347\u8bc4\u4f30\u901f\u5ea6\u3002\u5982\u8868 2 \u548c 3 \u6240\u793a\uff0crStar \u80fd\u591f\u663e\u8457\u63d0\u9ad8 SLM \u5728\u8fd9\u4e9b\u9ad8\u96be\u5ea6\u6570\u5b66\u6570\u636e\u96c6\u4e0a\u7684\u63a8\u7406\u51c6\u786e\u5ea6\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-eb8c166fbd40a8ed28ec8122b2c69c04.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-eb8c166fbd40a8ed28ec8122b2c69c04.png\"><\/a><\/p>\n<p><strong>\u6d88\u878d\u7814\u7a76<\/strong><\/p>\n<ul>\n<li>\n<p>\u4e0d\u540c Rollout \u7684\u6709\u6548\u6027<\/p>\n<\/li>\n<\/ul>\n<p>rStar \u4f7f\u7528\u4e86 Rollout \u7b56\u7565\u6765\u6267\u884c MCTS \u6811\u6269\u5c55\u3002\u66f4\u591a Rollout \u4f1a\u751f\u6210\u66f4\u591a\u5019\u9009\u89e3\u7b54\u8f68\u8ff9\uff0c\u4f46\u4e5f\u4f1a\u62ac\u9ad8\u63a8\u7406\u6210\u672c\u3002\u56fe 5 \u6bd4\u8f83\u4e86\u5728 GSM8K \u4e0a\uff0cSC\u3001RAP \u548c rStar \u4f7f\u7528\u4e0d\u540c Rollout \u65f6\u7684\u51c6\u786e\u5ea6\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-0edc08b3fa1ec4d98564043fb641e1ad.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-0edc08b3fa1ec4d98564043fb641e1ad.png\"><\/a><\/p>\n<p>\u8fd9\u91cc\u5f97\u5230\u4e24\u4e2a\u5173\u952e\u89c2\u5bdf\u7ed3\u679c\uff1a<\/p>\n<section>1. \u5373\u4f7f\u4ec5 2 \u6b21 Rollout\uff0crStar \u4e5f\u80fd\u5927\u5e45\u63d0\u5347 SLM \u7684\u63a8\u7406\u51c6\u786e\u5ea6\uff0c\u8fd9\u8868\u660e\u4e86\u5176\u6709\u6548\u6027\uff1b<\/section>\n<section>2.Rollout \u66f4\u591a\u65f6\u5bf9 rStar \u548c SC \u90fd\u6709\u5229\uff0c\u800c RAP \u5728 4 \u6b21 Rollout \u4e4b\u540e\u5f80\u5f80\u4f1a\u9971\u548c\u751a\u81f3\u4e0b\u964d\u3002\u4e00\u4e2a\u539f\u56e0\u662f RAP \u7684\u5355\u7c7b\u578b\u52a8\u4f5c\u7a7a\u95f4\u4f1a\u9650\u5236 MCTS \u63a2\u7d22\u7684\u6548\u679c\u3002<\/section>\n<ul>\n<li>\n<p>MCTS \u751f\u6210\u5668\u7684\u6709\u6548\u6027<\/p>\n<\/li>\n<\/ul>\n<p>\u8be5\u56e2\u961f\u6bd4\u8f83\u4e86 MCTS \u751f\u6210\u5668\u4e0e\u5176\u5b83\u4e09\u79cd\u751f\u6210\u5668\u7684\u6548\u679c\u3002\u5982\u8868 4 \u6240\u793a\uff0c\u65b0\u63d0\u51fa\u7684 MCTS \u751f\u6210\u5668\u5168\u9762\u80dc\u8fc7\u5176\u5b83\u751f\u6210\u5668\u3002\u6b64\u5916\uff0c\u9488\u5bf9 SLM \u8c03\u6574\u8fc7\u7684\u5956\u52b1\u51fd\u6570\u7684\u6709\u6548\u6027\u4e5f\u5f97\u5230\u4e86\u8bc1\u660e\uff0c\u56e0\u4e3a\u81ea\u6211\u8bc4\u4f30\u4f1a\u964d\u4f4e\u65b0\u751f\u6210\u5668\u7684\u51c6\u786e\u5ea6\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-6f26a06d1a85a2be917e45cd415227bc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-6f26a06d1a85a2be917e45cd415227bc.png\"><\/a><\/p>\n<ul>\n<li>\n<p>\u5224\u522b\u5668\u7684\u6709\u6548\u6027<\/p>\n<\/li>\n<\/ul>\n<p>\u8be5\u56e2\u961f\u8bbe\u7f6e\u4e86\u4e24\u4e2a\u8bc4\u4f30\u5b9e\u9a8c\u3002<\/p>\n<p>\u7b2c\u4e00\u4e2a\u5b9e\u9a8c\u662f\u5c06\u5224\u522b\u65b9\u6cd5\u4e0e\u591a\u6570\u6295\u7968\u548c\u81ea\u6211\u9a8c\u8bc1\u65b9\u6cd5\u8fdb\u884c\u6bd4\u8f83\u3002\u7ed3\u679c\u89c1\u8868 5\uff08\u5de6\uff09\uff0c\u53ef\u4ee5\u770b\u5230\u5224\u522b\u65b9\u6cd5\u7684\u4f18\u52bf\u975e\u5e38\u663e\u8457\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-c66b1a365178aa8c1ff222f82983de25.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-c66b1a365178aa8c1ff222f82983de25.png\"><\/a><\/p>\n<p>\u7b2c\u4e8c\u4e2a\u5b9e\u9a8c\u5219\u662f\u7814\u7a76\u4e0d\u540c\u7684\u5224\u522b\u5668\u6a21\u578b\u7684\u5f71\u54cd\u3002\u7ed3\u679c\u89c1\u8868 5\uff08\u53f3\uff09\uff0c\u53ef\u4ee5\u770b\u5230\u9009\u62e9\u4e0d\u540c\u7684\u5224\u522b\u5668\u6a21\u578b\u901a\u5e38\u4e0d\u4f1a\u5f71\u54cd\u63a8\u7406\u4e92\u6070\u6027\u65b9\u6cd5\u9a8c\u8bc1\u7b54\u6848\u7684\u6548\u679c\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5373\u4f7f\u4f7f\u7528\u5f3a\u5927\u7684 GPT-4 \u4f5c\u4e3a\u5224\u522b\u5668\uff0c\u6027\u80fd\u4e5f\u53ea\u6709\u7565\u5fae\u63d0\u5347\uff08\u4ece 91.13% \u63d0\u5347\u5230 92.57%\uff09\u3002\u8fd9\u8868\u660e\u63a8\u7406\u4e92\u6070\u6027\u65b9\u6cd5\u53ef\u4ee5\u6709\u6548\u5730\u4f7f\u7528 SLM \u6765\u9a8c\u8bc1\u7b54\u6848\u3002<\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-08-16\" target=\"_blank\">\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4\u63a5\u6bd4\u80a9\u5927\u6a21\u578b\uff1f\u5fae\u8f6f\u7684rStar\u751a\u81f3\u6ca1\u7528CoT\u548c\u5fae\u8c03<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4 [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68],"class_list":["post-35144","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4\u63a5\u6bd4\u80a9\u5927\u6a21\u578b\uff1f\u5fae\u8f6f\u7684rStar\u751a\u81f3\u6ca1\u7528CoT\u548c\u5fae\u8c03 - \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=35144\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35144#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35144\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4\u63a5\u6bd4\u80a9\u5927\u6a21\u578b\uff1f\u5fae\u8f6f\u7684rStar\u751a\u81f3\u6ca1\u7528CoT\u548c\u5fae\u8c03\",\"datePublished\":\"2024-12-19T03:59:53+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35144\"},\"wordCount\":270,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35144#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/08\\\/frc-57b19af6a3fa257886fae56448547749.png\",\"keywords\":[\"\u5927\u6a21\u578b\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=35144#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35144\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=35144\",\"name\":\"\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4\u63a5\u6bd4\u80a9\u5927\u6a21\u578b\uff1f\u5fae\u8f6f\u7684rStar\u751a\u81f3\u6ca1\u7528CoT\u548c\u5fae\u8c03 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35144#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35144#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/08\\\/frc-57b19af6a3fa257886fae56448547749.png\",\"datePublished\":\"2024-12-19T03:59:53+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35144#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=35144\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35144#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/08\\\/frc-57b19af6a3fa257886fae56448547749.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/08\\\/frc-57b19af6a3fa257886fae56448547749.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=35144#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4\u63a5\u6bd4\u80a9\u5927\u6a21\u578b\uff1f\u5fae\u8f6f\u7684rStar\u751a\u81f3\u6ca1\u7528CoT\u548c\u5fae\u8c03\"}]},{\"@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\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"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":"\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4\u63a5\u6bd4\u80a9\u5927\u6a21\u578b\uff1f\u5fae\u8f6f\u7684rStar\u751a\u81f3\u6ca1\u7528CoT\u548c\u5fae\u8c03 - \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=35144","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=35144#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=35144"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4\u63a5\u6bd4\u80a9\u5927\u6a21\u578b\uff1f\u5fae\u8f6f\u7684rStar\u751a\u81f3\u6ca1\u7528CoT\u548c\u5fae\u8c03","datePublished":"2024-12-19T03:59:53+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=35144"},"wordCount":270,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=35144#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-57b19af6a3fa257886fae56448547749.png","keywords":["\u5927\u6a21\u578b"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=35144#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=35144","url":"https:\/\/17aitech.com\/?p=35144","name":"\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4\u63a5\u6bd4\u80a9\u5927\u6a21\u578b\uff1f\u5fae\u8f6f\u7684rStar\u751a\u81f3\u6ca1\u7528CoT\u548c\u5fae\u8c03 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=35144#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=35144#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-57b19af6a3fa257886fae56448547749.png","datePublished":"2024-12-19T03:59:53+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=35144#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=35144"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=35144#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-57b19af6a3fa257886fae56448547749.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/08\/frc-57b19af6a3fa257886fae56448547749.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=35144#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u4e24\u4e2a\u5c0f\u6a21\u578b\u4e92\u76f8\u9a8c\u8bc1\uff0c\u76f4\u63a5\u6bd4\u80a9\u5927\u6a21\u578b\uff1f\u5fae\u8f6f\u7684rStar\u751a\u81f3\u6ca1\u7528CoT\u548c\u5fae\u8c03"}]},{"@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":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","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\/35144","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=35144"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/35144\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=35144"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=35144"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=35144"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}