{"id":50551,"date":"2025-08-10T10:05:52","date_gmt":"2025-08-10T02:05:52","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/50551.html"},"modified":"2025-08-10T10:05:52","modified_gmt":"2025-08-10T02:05:52","slug":"%e8%ae%ba%e6%96%87%e9%98%85%e8%af%bb-%e4%ba%ba%e5%b7%a5%e6%99%ba%e8%83%bd-%e8%bd%af%e4%bb%b6%e5%b7%a5%e7%a8%8b-posterior-grpo%ef%bc%9a%e4%bc%98%e5%8c%96%e4%bb%a3%e7%a0%81%e7%94%9f%e6%88%90","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/50551.html","title":{"rendered":"[\u8bba\u6587\u9605\u8bfb] \u4eba\u5de5\u667a\u80fd + \u8f6f\u4ef6\u5de5\u7a0b | Posterior-GRPO\uff1a\u4f18\u5316\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u7684\u65b0\u6846\u67b6"},"content":{"rendered":"<h2>Posterior-GRPO&#xff1a;\u4f18\u5316\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u7684\u65b0\u6846\u67b6<\/h2>\n<h3>\u8bba\u6587&#xff1a;Posterior-GRPO: Rewarding Reasoning Processes in Code Generation<\/h3>\n<p>arXiv:2508.05170<br \/>\nPosterior-GRPO: Rewarding Reasoning Processes in Code Generation<br \/>\nLishui Fan, Yu Zhang, Mouxiang Chen, Zhongxin Liu<br \/>\nSubjects: Software Engineering (cs.SE); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG)<\/p>\n<h4>\u4e00\u6bb5\u8bdd\u603b\u7ed3&#xff1a;<\/h4>\n<p>\u672c\u6587\u63d0\u51faPosterior-GRPO&#xff08;P-GRPO&#xff09; \u6846\u67b6&#xff0c;\u65e8\u5728\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60&#xff08;RL&#xff09;\u4f18\u5316\u4ee3\u7801\u751f\u6210\u4e2d\u7684\u63a8\u7406\u8fc7\u7a0b\u8d28\u91cf&#xff0c;\u89e3\u51b3\u73b0\u6709\u65b9\u6cd5\u4f9d\u8d56\u7ed3\u679c\u5956\u52b1\u5bfc\u81f4\u7684\u63a8\u7406\u8fc7\u7a0b\u88ab\u5ffd\u89c6\u53ca\u5956\u52b1\u9ed1\u5ba2\u95ee\u9898\u3002\u8be5\u6846\u67b6\u5305\u542b\u4e09\u90e8\u5206&#xff1a;\u6784\u5efaLCB-RB\u57fa\u51c6&#xff08;\u542b187\u5bf9\u4f18\u52a3\u63a8\u7406\u8fc7\u7a0b\u504f\u597d\u5bf9&#xff09;\u7528\u4e8e\u63a8\u7406\u8bc4\u4f30&#xff1b;\u63d0\u51faOD-based\u5956\u52b1\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5&#xff0c;\u901a\u8fc7\u4f18\u5316\u548c\u964d\u7ea7\u63a8\u7406\u8def\u5f84&#xff08;\u57fa\u4e8e\u4e8b\u5b9e\u51c6\u786e\u6027\u3001\u903b\u8f91\u4e25\u8c28\u6027\u3001\u8fde\u8d2f\u6027\u7ef4\u5ea6&#xff09;\u751f\u6210\u9ad8\u8d28\u91cf\u504f\u597d\u5bf9&#xff0c;\u8bad\u7ec3\u76847B\u6a21\u578b\u5728LCB-RB\u4e0a\u8fbeSOTA&#xff1b;\u8bbe\u8ba1P-GRPO\u7b97\u6cd5&#xff0c;\u4ec5\u5bf9\u6210\u529f\u7ed3\u679c\u7684\u63a8\u7406\u8fc7\u7a0b\u5e94\u7528\u5956\u52b1&#xff0c;\u7f13\u89e3\u5956\u52b1\u9ed1\u5ba2\u3002\u5b9e\u9a8c\u663e\u793a&#xff0c;7B\u6a21\u578b\u5728\u4ee3\u7801\u751f\u6210\u4efb\u52a1\u4e0a\u8d85\u4ec5\u7ed3\u679c\u5956\u52b1\u57fa\u7ebf4.5%&#xff0c;\u6027\u80fd\u63a5\u8fd1GPT-4-Turbo&#xff0c;\u4e14\u5728\u6570\u5b66\u4efb\u52a1\u4e0a\u6cdb\u5316\u6027\u826f\u597d&#xff08;\u76f8\u5bf9\u63d0\u53477.3%&#xff09;\u3002<\/p>\n<h3>\u7814\u7a76\u80cc\u666f<\/h3>\n<p>\u5728\u5927\u8bed\u8a00\u6a21\u578b&#xff08;LLMs&#xff09;\u7684\u4ee3\u7801\u751f\u6210\u9886\u57df&#xff0c;\u5f3a\u5316\u5b66\u4e60&#xff08;RL&#xff09;\u5df2\u6210\u4e3a\u91cd\u8981\u7684\u540e\u8bad\u7ec3\u8303\u5f0f\u3002\u7136\u800c&#xff0c;\u73b0\u6709\u65b9\u6cd5\u5b58\u5728\u660e\u663e\u5c40\u9650&#xff1a;\u5b83\u4eec\u4ec5\u4f9d\u8d56\u6d4b\u8bd5\u7528\u4f8b\u7ed3\u679c&#xff08;\u5982\u4ee3\u7801\u901a\u8fc7\u7387&#xff09;\u6765\u7ed9\u4e88\u6a21\u578b\u5956\u52b1&#xff0c;\u5374\u5ffd\u89c6\u4e86\u6a21\u578b\u751f\u6210\u4ee3\u7801\u65f6\u7684\u4e2d\u95f4\u63a8\u7406\u8fc7\u7a0b\u8d28\u91cf\u3002<\/p>\n<p>\u8fd9\u5c31\u597d\u6bd4\u8001\u5e08\u6279\u6539\u4f5c\u4e1a\u53ea\u770b\u7b54\u6848\u5bf9\u9519&#xff0c;\u4e0d\u5173\u6ce8\u5b66\u751f\u7684\u89e3\u9898\u601d\u8def\u3002\u957f\u6b64\u4ee5\u5f80&#xff0c;\u6a21\u578b\u53ef\u80fd\u5b66\u4f1a\u201c\u8d70\u6377\u5f84\u201d\u2014\u2014\u6bd4\u5982\u751f\u6210\u770b\u4f3c\u6b63\u786e\u4f46\u903b\u8f91\u6df7\u4e71\u7684\u4ee3\u7801&#xff0c;\u6216\u8005\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u5b58\u5728\u6f0f\u6d1e\u5374\u4fa5\u5e78\u901a\u8fc7\u6d4b\u8bd5&#xff0c;\u8fd9\u79cd\u73b0\u8c61\u88ab\u79f0\u4e3a\u201c\u5956\u52b1\u9ed1\u5ba2\u201d\u3002<\/p>\n<p>\u66f4\u5173\u952e\u7684\u662f&#xff0c;\u7814\u7a76\u53d1\u73b0\u63a8\u7406\u8fc7\u7a0b\u8d28\u91cf\u4e0e\u6700\u7ec8\u4ee3\u7801\u6b63\u786e\u6027\u5b58\u5728\u663e\u8457\u5173\u8054&#xff08;\u03c7\u00b2\u68c0\u9a8c\u663e\u793ap&#061;9.3\u00d710\u207b\u00b9\u2075\u226a0.001&#xff09;&#xff0c;\u5ffd\u89c6\u63a8\u7406\u8fc7\u7a0b\u53ef\u80fd\u5bfc\u81f4\u6a21\u578b\u6027\u80fd\u96be\u4ee5\u6301\u7eed\u63d0\u5347\u3002\u56e0\u6b64&#xff0c;\u5982\u4f55\u8ba9\u6a21\u578b\u5728\u751f\u6210\u6b63\u786e\u4ee3\u7801\u7684\u540c\u65f6&#xff0c;\u5177\u5907\u9ad8\u8d28\u91cf\u7684\u63a8\u7406\u8fc7\u7a0b&#xff0c;\u6210\u4e3a\u4e9f\u5f85\u89e3\u51b3\u7684\u95ee\u9898\u3002<\/p>\n<h3>\u4e3b\u8981\u4f5c\u8005\u53ca\u5355\u4f4d\u4fe1\u606f<\/h3>\n<ul>\n<li>\u4f5c\u8005&#xff1a;Lishui Fan\u3001Yu Zhang*\u3001Mouxiang Chen\u3001Zhongxin Liu\u2020<\/li>\n<li>\u5355\u4f4d&#xff1a;1. \u533a\u5757\u94fe\u4e0e\u6570\u636e\u5b89\u5168\u56fd\u5bb6\u91cd\u70b9\u5b9e\u9a8c\u5ba4&#xff1b;2. \u6d59\u6c5f\u5927\u5b66<\/li>\n<\/ul>\n<h3>\u521b\u65b0\u70b9<\/h3>\n<li>LCB-RB\u57fa\u51c6&#xff1a;\u9996\u4e2a\u4e13\u95e8\u7528\u4e8e\u8bc4\u4f30\u5956\u52b1\u6a21\u578b\u5bf9\u63a8\u7406\u8fc7\u7a0b\u533a\u5206\u80fd\u529b\u7684\u57fa\u51c6&#xff0c;\u5305\u542b187\u5bf9\u201c\u4f18\u8d28\u63a8\u7406&#043;\u6b63\u786e\u4ee3\u7801\u201d\u4e0e\u201c\u52a3\u8d28\u63a8\u7406&#043;\u9519\u8bef\u4ee3\u7801\u201d\u7684\u504f\u597d\u5bf9\u3002<\/li>\n<li>OD-based\u5956\u52b1\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5&#xff1a;\u901a\u8fc7\u7cfb\u7edf\u5730\u4f18\u5316\u548c\u964d\u7ea7\u521d\u59cb\u63a8\u7406\u8def\u5f84&#xff08;\u57fa\u4e8e\u4e8b\u5b9e\u51c6\u786e\u6027\u3001\u903b\u8f91\u4e25\u8c28\u6027\u3001\u8fde\u8d2f\u6027\u4e09\u4e2a\u7ef4\u5ea6&#xff09;&#xff0c;\u751f\u6210\u9ad8\u8d28\u91cf\u504f\u597d\u5bf9&#xff0c;\u8ba9\u5956\u52b1\u6a21\u578b\u66f4\u7cbe\u51c6\u5730\u8bc4\u4f30\u63a8\u7406\u8d28\u91cf\u3002<\/li>\n<li>Posterior-GRPO&#xff08;P-GRPO&#xff09;\u7b97\u6cd5&#xff1a;\u4e00\u79cd\u65b0\u578b\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5&#xff0c;\u4ec5\u5bf9\u201c\u6d4b\u8bd5\u901a\u8fc7\u7684\u4ee3\u7801\u201d\u5bf9\u5e94\u7684\u63a8\u7406\u8fc7\u7a0b\u7ed9\u4e88\u5956\u52b1&#xff0c;\u65e2\u907f\u514d\u5956\u52b1\u9ed1\u5ba2&#xff0c;\u53c8\u8ba9\u6a21\u578b\u7684\u63a8\u7406\u8fc7\u7a0b\u4e0e\u4ee3\u7801\u6b63\u786e\u6027\u5bf9\u9f50\u3002<\/li>\n<h3>\u7814\u7a76\u65b9\u6cd5\u548c\u601d\u8def<\/h3>\n<h4>1. LCB-RB\u57fa\u51c6\u6784\u5efa<\/h4>\n<ul>\n<li>\u4eceLiveCodeBench\u9009\u53d6880\u4e2a\u4ee3\u7801\u95ee\u9898&#xff0c;\u7528Qwen2.5-Coder-32B-Instruct\u751f\u621050\u4e2a\u5e26\u63a8\u7406\u8fc7\u7a0b\u7684\u89e3\u51b3\u65b9\u6848&#xff1b;<\/li>\n<li>\u7ecfGPT-4o\u8fc7\u6ee4&#xff0c;\u4fdd\u7559\u201c\u63a8\u7406\u4e0e\u4ee3\u7801\u4e00\u81f4\u201d\u7684\u6837\u672c&#xff0c;\u6700\u7ec8\u5f62\u6210187\u5bf9\u504f\u597d\u5bf9&#xff08;\u4f18\u8d28\u63a8\u7406&#043;\u6b63\u786e\u4ee3\u7801 vs \u52a3\u8d28\u63a8\u7406&#043;\u9519\u8bef\u4ee3\u7801&#xff09;\u3002<\/li>\n<\/ul>\n<h4>2. OD-based\u5956\u52b1\u6a21\u578b\u8bad\u7ec3<\/h4>\n<ul>\n<li>\u6b65\u9aa41&#xff1a;\u7528\u5f3a\u5927\u7684LLM\u751f\u6210\u521d\u59cb\u63a8\u7406\u8def\u5f84&#xff1b;<\/li>\n<li>\u6b65\u9aa42&#xff1a;\u57fa\u4e8e\u201c\u4e8b\u5b9e\u51c6\u786e\u6027\u3001\u903b\u8f91\u4e25\u8c28\u6027\u3001\u8fde\u8d2f\u6027\u201d\u4e09\u4e2a\u7ef4\u5ea6&#xff0c;\u751f\u6210\u4f18\u5316\u7248&#xff08;y\u207a&#xff09;\u548c\u964d\u7ea7\u7248&#xff08;y\u207b&#xff09;\u63a8\u7406\u8def\u5f84&#xff1b;<\/li>\n<li>\u6b65\u9aa43&#xff1a;\u7528\u4e09\u79cd\u504f\u597d\u5bf9&#xff08;(x,y\u207a,y\u207b)\u3001(x,y,y\u207b)\u3001(x,y\u207a,y)&#xff09;\u8bad\u7ec3Bradley-Terry\u5956\u52b1\u6a21\u578b&#xff0c;\u4f7f\u5176\u80fd\u7ed9\u63a8\u7406\u8d28\u91cf\u6253\u5206&#xff08;0-1\u4e4b\u95f4&#xff09;\u3002<\/li>\n<\/ul>\n<h4>3. P-GRPO\u7b97\u6cd5\u8bbe\u8ba1<\/h4>\n<ul>\n<li>\u5956\u52b1\u7ec4\u6210&#xff1a;\u683c\u5f0f\u5956\u52b1&#xff08;R<sup>f&#xff0c;\u68c0\u67e5\u8f93\u51fa\u7ed3\u6784\u662f\u5426\u5408\u89c4&#xff09;\u3001\u89c4\u5219\u5956\u52b1&#xff08;R<\/sup>o&#xff0c;\u6d4b\u8bd5\u7528\u4f8b\u901a\u8fc7\u7387&#xff09;\u3001\u601d\u7ef4\u5956\u52b1&#xff08;R^t&#xff0c;\u63a8\u7406\u8d28\u91cf\u8bc4\u5206&#xff09;&#xff1b;<\/li>\n<li>\u540e\u9a8c\u7b56\u7565&#xff1a;\u4ec5\u5f53R<sup>o&#061;1&#xff08;\u4ee3\u7801\u901a\u8fc7\u6240\u6709\u6d4b\u8bd5&#xff09;\u65f6&#xff0c;\u624d\u4fdd\u7559R<\/sup>t&#xff0c;\u5426\u5219R^t&#061;0&#xff1b;<\/li>\n<li>\u603b\u5956\u52b1\u516c\u5f0f&#xff1a;R_i &#061; R_i^f &#043; R_i^o &#043; R_i<sup>o\u00b7R_i<\/sup>t&#xff0c;\u89e3\u51b3\u4e86\u4f20\u7edfGRPO\u4e2d\u201c\u6210\u529f\u6837\u672c\u5956\u52b1\u65e0\u5dee\u5f02\u201d\u7684\u95ee\u9898\u3002<\/li>\n<\/ul>\n<h4>\u5b9e\u9a8c\u65b9\u6cd5<\/h4>\n<ul>\n<li>\u5956\u52b1\u6a21\u578b&#xff1a;\u57fa\u4e8eQwen2.5-Coder-3B\/7B-Base\u8bad\u7ec3&#xff0c;\u5728LCB-RB\u548cRewardBench\u4e0a\u8bc4\u4f30&#xff1b;<\/li>\n<li>\u5f3a\u5316\u5b66\u4e60&#xff1a;\u4ee5Qwen2.5-Coder-7B-Instruct\u4e3a\u7b56\u7565\u6a21\u578b&#xff0c;\u5728HumanEval(&#043;)\u3001MBPP(&#043;)\u7b494\u4e2a\u4ee3\u7801\u57fa\u51c6\u4e0a\u6d4b\u8bd5&#xff0c;\u7528Pass&#064;1\u6307\u6807\u8bc4\u4f30&#xff1b;<\/li>\n<li>\u6570\u5b66\u4efb\u52a1\u6269\u5c55&#xff1a;\u7528Qwen2.5Math-7B\u6a21\u578b&#xff0c;\u5728MATH500\u7b493\u4e2a\u6570\u5b66\u57fa\u51c6\u4e0a\u9a8c\u8bc1\u6cdb\u5316\u6027\u3002<\/li>\n<\/ul>\n<h3>\u4e3b\u8981\u8d21\u732e<\/h3>\n<li>\u6027\u80fd\u63d0\u5347\u663e\u8457&#xff1a;7B\u53c2\u6570\u6a21\u578b\u7ecfP-GRPO\u8bad\u7ec3\u540e&#xff0c;\u5728\u4ee3\u7801\u751f\u6210\u4efb\u52a1\u4e0a\u5e73\u5747\u8d85\u201c\u4ec5\u7ed3\u679c\u5956\u52b1\u201d\u57fa\u7ebf4.5%&#xff0c;\u5728LiveCodeBench\u4e0a\u76f8\u5bf9\u63d0\u534718.1%&#xff0c;\u6027\u80fd\u63a5\u8fd1GPT-4-Turbo\u3002<\/li>\n<li>\u5956\u52b1\u6a21\u578b\u8868\u73b0\u4f18\u5f02&#xff1a;OD-based\u65b9\u6cd5\u8bad\u7ec3\u76847B\u6a21\u578b\u5728LCB-RB\u4e0a\u51c6\u786e\u7387\u8fbe58.28%&#xff0c;\u8d85GPT-4-Turbo&#xff0c;\u5728RewardBench\u63a8\u7406\u5b50\u96c6\u4e0a\u5e73\u5747\u51c6\u786e\u738782.22%&#xff0c;\u4e3aSOTA\u3002<\/li>\n<li>\u8de8\u9886\u57df\u6cdb\u5316\u6027\u5f3a&#xff1a;\u6269\u5c55\u5230\u6570\u5b66\u4efb\u52a1\u65f6&#xff0c;Qwen2.5Math-7B\u76f8\u5bf9\u201c\u4ec5\u7ed3\u679c\u5956\u52b1\u201d\u57fa\u7ebf\u63d0\u53477.3%&#xff0c;\u9a8c\u8bc1\u4e86\u65b9\u6cd5\u7684\u901a\u7528\u6027\u3002<\/li>\n<li>\u89e3\u51b3\u6838\u5fc3\u75db\u70b9&#xff1a;\u6709\u6548\u7f13\u89e3\u5956\u52b1\u9ed1\u5ba2\u95ee\u9898&#xff0c;\u8ba9\u6a21\u578b\u540c\u65f6\u5173\u6ce8\u201c\u63a8\u7406\u8d28\u91cf\u201d\u548c\u201c\u7ed3\u679c\u6b63\u786e\u6027\u201d&#xff0c;\u4e3a\u4ee3\u7801\u751f\u6210\u4e0e\u590d\u6742\u63a8\u7406\u4efb\u52a1\u63d0\u4f9b\u65b0\u8303\u5f0f\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250810020549-6897fe7ddaa3f.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/li>\n<hr \/>\n<h4>\u601d\u7ef4\u5bfc\u56fe&#xff1a;<\/h4>\n<p><img decoding=\"async\" src=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250810020550-6897fe7e420a6.png\" alt=\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\" \/><\/p>\n<hr \/>\n<h4>\u8be6\u7ec6\u603b\u7ed3&#xff1a;<\/h4>\n<h5>1. \u7814\u7a76\u80cc\u666f\u4e0e\u6311\u6218<\/h5>\n<ul>\n<li>\u73b0\u72b6&#xff1a;\u73b0\u6709RL\u65b9\u6cd5\u5728\u4ee3\u7801\u751f\u6210\u4e2d\u4ec5\u4f9d\u8d56\u7ed3\u679c\u5956\u52b1&#xff08;\u5982\u6d4b\u8bd5\u901a\u8fc7\u7387&#xff09;&#xff0c;\u5ffd\u89c6\u63a8\u7406\u8fc7\u7a0b\u8d28\u91cf&#xff0c;\u53ef\u80fd\u5bfc\u81f4\u63a8\u7406\u8fc7\u7a0b\u4e0d\u4f73&#xff0c;\u6700\u7ec8\u5f71\u54cd\u7ed3\u679c\u51c6\u786e\u6027\u3002<\/li>\n<li>\u6311\u6218&#xff1a;\n<ul>\n<li>\u7f3a\u4e4f\u8bc4\u4f30\u63a8\u7406\u8fc7\u7a0b\u7684\u57fa\u51c6&#xff08;\u73b0\u6709\u57fa\u51c6\u4fa7\u91cd\u7ed3\u679c&#xff09;&#xff1b;<\/li>\n<li>\u7f3a\u5c11\u9488\u5bf9\u63a8\u7406\u8bc4\u4f30\u7684\u53ef\u9760\u5956\u52b1\u6a21\u578b&#xff08;\u73b0\u6709\u6a21\u578b\u57fa\u4e8e\u7ed3\u679c\u8bad\u7ec3&#xff09;&#xff1b;<\/li>\n<li>\u73b0\u6709RL\u7b97\u6cd5\u6613\u53d7\u5956\u52b1\u9ed1\u5ba2\u5f71\u54cd&#xff08;\u6a21\u578b\u5229\u7528\u5956\u52b1\u4fe1\u53f7\u800c\u975e\u63d0\u5347\u7ed3\u679c&#xff09;\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h5>2. \u6838\u5fc3\u65b9\u6cd5<\/h5>\n<table>\n<tr>\u7ec4\u6210\u90e8\u5206\u7ec6\u8282\u63cf\u8ff0\u5173\u952e\u6210\u679c<\/tr>\n<tbody>\n<tr>\n<td>LCB-RB\u57fa\u51c6<\/td>\n<td>\u57fa\u4e8eLiveCodeBench\u7684880\u4e2a\u95ee\u9898&#xff0c;\u7528Qwen2.5-Coder-32B-Instruct\u751f\u621050\u4e2a\u5e26\u63a8\u7406\u7684\u89e3\u51b3\u65b9\u6848&#xff0c;\u7ecfGPT-4o\u8fc7\u6ee4&#xff0c;\u6700\u7ec8\u5f97\u5230187\u5bf9\u504f\u597d\u5bf9&#xff08;\u4f18\u63a8\u7406&#043;\u6b63\u786e\u4ee3\u7801 vs \u52a3\u63a8\u7406&#043;\u9519\u8bef\u4ee3\u7801&#xff09;<\/td>\n<td>\u9996\u4e2a\u9488\u5bf9\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u7684\u504f\u597d\u8bc4\u4f30\u57fa\u51c6<\/td>\n<\/tr>\n<tr>\n<td>OD-based\u5956\u52b1\u6a21\u578b\u8bad\u7ec3<\/td>\n<td>1. \u751f\u6210\u521d\u59cb\u63a8\u7406\u8def\u5f84&#xff1b;2. \u57fa\u4e8e\u4e8b\u5b9e\u51c6\u786e\u6027\u3001\u903b\u8f91\u4e25\u8c28\u6027\u3001\u8fde\u8d2f\u6027\u4f18\u5316&#xff08;y\u207a&#xff09;\u548c\u964d\u7ea7&#xff08;y\u207b&#xff09;&#xff1b;3. \u7528\u4e09\u79cd\u504f\u597d\u5bf9&#xff08;(x,y\u207a,y\u207b)\u3001(x,y,y\u207b)\u3001(x,y\u207a,y)&#xff09;\u8bad\u7ec3Bradley-Terry\u6a21\u578b<\/td>\n<td>7B\u6a21\u578b\u5728LCB-RB\u51c6\u786e\u738758.28%&#xff0c;\u8d85GPT-4-Turbo&#xff0c;\u5728RewardBench\u63a8\u7406\u5b50\u96c6\u8fbeSOTA<\/td>\n<\/tr>\n<tr>\n<td>P-GRPO\u7b97\u6cd5<\/td>\n<td>\u5956\u52b1\u516c\u5f0f&#xff1a;(R_i &#061; R_i^f &#043; R_i^o &#043; R_i^o \\\\cdot R_i<sup>t)&#xff0c;\u5176\u4e2d(R<\/sup>t)\u4ec5\u5728(R^o&#061;1)\u65f6\u6709\u6548&#xff1b;\u89e3\u51b3GRPO\u4e2d\u6210\u529f\u6837\u672c\u5956\u52b1\u65e0\u5dee\u5f02\u95ee\u9898<\/td>\n<td>\u63d0\u5347\u6570\u636e\u5229\u7528\u6548\u7387&#xff0c;\u4f7f\u6210\u529f\u6837\u672c\u56e0\u63a8\u7406\u8d28\u91cf\u5dee\u5f02\u4ea7\u751f\u5956\u52b1\u533a\u5206<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h5>3. \u5b9e\u9a8c\u7ed3\u679c<\/h5>\n<ul>\n<li>\u4ee3\u7801\u751f\u6210\u4efb\u52a1&#xff08;\u88681&#xff09;&#xff1a;\n<ul>\n<li>7B\u6a21\u578b&#xff08;Qwen2.5-Coder-Instruct&#xff09;\u7ecfP-GRPO\u8bad\u7ec3&#xff0c;\u5728HumanEval(&#043;)\u3001MBPP(&#043;)\u7b494\u4e2a\u57fa\u51c6\u4e0a\u5e73\u5747\u76f8\u5bf9\u63d0\u534713.9%&#xff0c;\u8d85\u4ec5\u7ed3\u679c\u5956\u52b1\u57fa\u7ebf4.5%&#xff0c;LiveCodeBench\u4e0a\u76f8\u5bf9\u63d0\u534718.1%&#xff0c;\u6027\u80fd\u63a5\u8fd1GPT-4-Turbo\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u6570\u5b66\u4efb\u52a1&#xff08;\u88683&#xff09;&#xff1a;\n<ul>\n<li>Qwen2.5Math-7B\u7ecfP-GRPO\u8bad\u7ec3&#xff0c;\u5728MATH500\u3001AIME2024\u7b493\u4e2a\u57fa\u51c6\u4e0a\u5e73\u5747\u76f8\u5bf9\u63d0\u53477.3%&#xff0c;\u8d85\u4ec5\u7ed3\u679c\u5956\u52b1\u57fa\u7ebf&#xff0c;\u6027\u80fd\u63a5\u8fd1Eurus-2-PRIME\u7b49SOTA\u6a21\u578b\u3002<\/li>\n<\/ul>\n<\/li>\n<li>\u5956\u52b1\u6a21\u578b\u5bf9\u6bd4&#xff1a;\n<ul>\n<li>OD-based\u65b9\u6cd5\u8bad\u7ec3\u76847B\u6a21\u578b\u5728LCB-RB\u3001RewardBench&#xff08;\u4ee3\u7801&#043;\u6570\u5b66&#xff09;\u4e0a\u5e73\u5747\u51c6\u786e\u738782.22%&#xff0c;\u8d85Starling-RM&#xff08;75.71%&#xff09;\u3001EURS-RM&#xff08;76.44%&#xff09;\u7b49\u57fa\u7ebf\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<h5>4. \u7ed3\u8bba\u4e0e\u5c55\u671b<\/h5>\n<ul>\n<li>\u6838\u5fc3\u8d21\u732e&#xff1a;\u63d0\u51faLCB-RB\u57fa\u51c6\u3001OD-based\u5956\u52b1\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5\u3001P-GRPO\u7b97\u6cd5&#xff0c;\u6709\u6548\u63d0\u5347\u63a8\u7406\u8d28\u91cf\u548c\u4efb\u52a1\u6027\u80fd\u3002<\/li>\n<li>\u5c40\u9650\u4e0e\u672a\u6765&#xff1a;\u53d7\u8ba1\u7b97\u8d44\u6e90\u9650\u5236&#xff0c;\u8ba1\u5212\u6269\u5c55\u81f3\u66f4\u5927\u6a21\u578b&#xff08;\u5982DeepSeek-R1-Distill-Qwen-7B&#xff09;&#xff0c;\u5f00\u53d1\u81ea\u8fed\u4ee3\u5b66\u4e60\u6846\u67b6\u3002<\/li>\n<\/ul>\n<hr \/>\n<h4>\u5173\u952e\u95ee\u9898&#xff1a;<\/h4>\n<li>\n<p>\u95ee\u9898&#xff1a;P-GRPO\u5982\u4f55\u7f13\u89e3\u5956\u52b1\u9ed1\u5ba2\u95ee\u9898&#xff1f;\u5176\u5956\u52b1\u673a\u5236\u6709\u4f55\u7279\u70b9&#xff1f;<br \/>\n\u7b54\u6848&#xff1a;P-GRPO\u901a\u8fc7\u201c\u540e\u9a8c\u5956\u52b1\u5206\u914d\u7b56\u7565\u201d\u7f13\u89e3\u5956\u52b1\u9ed1\u5ba2&#xff1a;\u4ec5\u5f53\u89c4\u5219\u5956\u52b1(R<sup>o&#061;1)&#xff08;\u4ee3\u7801\u901a\u8fc7\u6240\u6709\u6d4b\u8bd5&#xff09;\u65f6&#xff0c;\u624d\u4fdd\u7559\u601d\u7ef4\u5956\u52b1(R<\/sup>t)&#xff1b;\u82e5(R<sup>o\u22601)&#xff0c;\u5219(R<\/sup>t&#061;0)\u3002\u8fd9\u786e\u4fdd\u6a21\u578b\u4ec5\u56e0\u6210\u529f\u7ed3\u679c\u7684\u9ad8\u8d28\u91cf\u63a8\u7406\u53d7\u5956\u52b1&#xff0c;\u907f\u514d\u5229\u7528\u9519\u8bef\u7ed3\u679c\u7684\u63a8\u7406\u83b7\u53d6\u5956\u52b1\u3002\u5956\u52b1\u673a\u5236\u5305\u542b\u4e09\u90e8\u5206&#xff1a;\u683c\u5f0f\u5956\u52b1&#xff08;\u786e\u4fdd\u8f93\u51fa\u7ed3\u6784\u5408\u89c4&#xff09;\u3001\u89c4\u5219\u5956\u52b1&#xff08;\u6d4b\u8bd5\u901a\u8fc7\u7387&#xff09;\u3001\u601d\u7ef4\u5956\u52b1&#xff08;\u63a8\u7406\u8d28\u91cf\u8bc4\u5206&#xff09;&#xff0c;\u6700\u7ec8\u5956\u52b1\u516c\u5f0f\u4e3a(R_i &#061; R_i^f &#043; R_i^o &#043; R_i^o \\\\cdot R_i^t)\u3002<\/p>\n<\/li>\n<li>\n<p>\u95ee\u9898&#xff1a;OD-based\u5956\u52b1\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5\u4e0e\u73b0\u6709\u65b9\u6cd5\u76f8\u6bd4&#xff0c;\u4f18\u52bf\u4f55\u5728&#xff1f;<br \/>\n\u7b54\u6848&#xff1a;OD-based\u65b9\u6cd5\u901a\u8fc7\u7cfb\u7edf\u4f18\u5316\u548c\u964d\u7ea7\u521d\u59cb\u63a8\u7406\u8def\u5f84&#xff08;\u57fa\u4e8e\u4e8b\u5b9e\u51c6\u786e\u6027\u3001\u903b\u8f91\u4e25\u8c28\u6027\u3001\u8fde\u8d2f\u6027&#xff09;\u751f\u6210\u5bf9\u6bd4\u9c9c\u660e\u7684\u504f\u597d\u5bf9&#xff0c;\u800c\u975e\u4f9d\u8d56\u76f4\u63a5\u6570\u503c\u8bc4\u5206\u3002\u4f18\u52bf\u5728\u4e8e&#xff1a;1. \u89e3\u51b3LLM\u5bf9\u7ec6\u7c92\u5ea6\u6570\u503c\u4e0d\u654f\u611f\u7684\u95ee\u9898&#xff0c;\u63d0\u4f9b\u66f4\u6e05\u6670\u7684\u5b66\u4e60\u4fe1\u53f7&#xff1b;2. \u8bad\u7ec3\u76847B\u6a21\u578b\u5728LCB-RB\u4e0a\u51c6\u786e\u738758.28%&#xff0c;\u8d85\u5206\u6570\u57fa\u7ebf23.5%&#xff0c;\u5728RewardBench\u63a8\u7406\u5b50\u96c6\u8fbeSOTA&#xff08;\u5e73\u574782.22%&#xff09;&#xff1b;3. \u6cdb\u5316\u80fd\u529b\u5f3a&#xff0c;\u53ef\u8fc1\u79fb\u81f3\u5176\u4ed6\u63a8\u7406\u8bc4\u4f30\u57fa\u51c6\u3002<\/p>\n<\/li>\n<li>\n<p>\u95ee\u9898&#xff1a;P-GRPO\u5728\u4ee3\u7801\u751f\u6210\u548c\u6570\u5b66\u4efb\u52a1\u4e0a\u7684\u6027\u80fd\u8868\u73b0\u5982\u4f55&#xff1f;\u4f53\u73b0\u4e86\u5176\u4ec0\u4e48\u7279\u6027&#xff1f;<br \/>\n\u7b54\u6848&#xff1a;\u5728\u4ee3\u7801\u751f\u6210\u4efb\u52a1\u4e0a&#xff0c;7B\u6a21\u578b\u7ecfP-GRPO\u8bad\u7ec3\u540e&#xff0c;\u5e73\u5747\u8d85\u4ec5\u7ed3\u679c\u5956\u52b1\u57fa\u7ebf4.5%&#xff0c;\u5728LiveCodeBench\u4e0a\u76f8\u5bf9\u63d0\u534718.1%&#xff0c;\u6027\u80fd\u63a5\u8fd1GPT-4-Turbo&#xff1b;\u5728\u6570\u5b66\u4efb\u52a1\u4e0a&#xff0c;Qwen2.5Math-7B\u76f8\u5bf9\u63d0\u53477.3%&#xff0c;\u5728AIME2024\u7b49\u57fa\u51c6\u4e0a\u8868\u73b0\u4f18\u4e8e\u57fa\u7ebf\u3002\u8fd9\u4f53\u73b0\u4e86P-GRPO\u4e0d\u4ec5\u80fd\u6709\u6548\u63d0\u5347\u4ee3\u7801\u751f\u6210\u4e2d\u7684\u63a8\u7406\u8d28\u91cf\u548c\u7ed3\u679c\u6b63\u786e\u6027&#xff0c;\u8fd8\u5177\u6709\u8de8\u9886\u57df\u6cdb\u5316\u80fd\u529b&#xff0c;\u53ef\u8fc1\u79fb\u81f3\u4f9d\u8d56\u9ad8\u8d28\u91cf\u63a8\u7406\u7684\u6570\u5b66\u4efb\u52a1\u3002<\/p>\n<\/li>\n<h3>\u603b\u7ed3<\/h3>\n<p>\u672c\u6587\u63d0\u51fa\u7684Posterior-GRPO\u6846\u67b6&#xff0c;\u901a\u8fc7\u6784\u5efaLCB-RB\u57fa\u51c6\u3001OD-based\u5956\u52b1\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5\u548cP-GRPO\u7b97\u6cd5&#xff0c;\u7cfb\u7edf\u6027\u5730\u89e3\u51b3\u4e86\u73b0\u6709\u5f3a\u5316\u5b66\u4e60\u5728\u4ee3\u7801\u751f\u6210\u4e2d\u5ffd\u89c6\u63a8\u7406\u8fc7\u7a0b\u7684\u95ee\u9898\u3002<\/p>\n<p>\u89e3\u51b3\u7684\u4e3b\u8981\u95ee\u9898&#xff1a;<\/p>\n<ul>\n<li>\u7f3a\u4e4f\u8bc4\u4f30\u63a8\u7406\u8fc7\u7a0b\u7684\u57fa\u51c6&#xff1b;<\/li>\n<li>\u7f3a\u5c11\u9488\u5bf9\u63a8\u7406\u7684\u53ef\u9760\u5956\u52b1\u6a21\u578b&#xff1b;<\/li>\n<li>\u5f3a\u5316\u5b66\u4e60\u6613\u53d7\u5956\u52b1\u9ed1\u5ba2\u5f71\u54cd\u3002<\/li>\n<\/ul>\n<p>\u4e3b\u8981\u6210\u679c&#xff1a;<\/p>\n<ul>\n<li>7B\u6a21\u578b\u5728\u4ee3\u7801\u751f\u6210\u4efb\u52a1\u4e0a\u6027\u80fd\u63a5\u8fd1GPT-4-Turbo&#xff1b;<\/li>\n<li>\u5956\u52b1\u6a21\u578b\u5728\u63a8\u7406\u8bc4\u4f30\u4e0a\u8fbeSOTA&#xff1b;<\/li>\n<li>\u65b9\u6cd5\u5728\u6570\u5b66\u4efb\u52a1\u4e0a\u5c55\u73b0\u5f3a\u6cdb\u5316\u6027\u3002<\/li>\n<\/ul>\n<p>\u8be5\u7814\u7a76\u4e0d\u4ec5\u63d0\u5347\u4e86\u6a21\u578b\u7684\u4ee3\u7801\u751f\u6210\u80fd\u529b&#xff0c;\u66f4\u63a8\u52a8\u4e86\u201c\u63a8\u7406\u8fc7\u7a0b\u4f18\u5316\u201d\u5728AI\u4efb\u52a1\u4e2d\u7684\u91cd\u89c6&#xff0c;\u76f8\u5173\u6a21\u578b\u3001\u6570\u636e\u96c6\u548c\u4ee3\u7801\u5df2\u516c\u5f00&#xff0c;\u4e3a\u540e\u7eed\u7814\u7a76\u63d0\u4f9b\u4fbf\u5229\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb942\u6b21\uff0c\u70b9\u8d5e19\u6b21\uff0c\u6536\u85cf20\u6b21\u3002\u672c\u6587\u4ecb\u7ecd\u4e86Posterior-GRPO\uff08P-GRPO\uff09\u6846\u67b6\uff0c\u8be5\u6846\u67b6\u9488\u5bf9\u73b0\u6709\u5f3a\u5316\u5b66\u4e60\u5728\u4ee3\u7801\u751f\u6210\u4e2d\u5ffd\u89c6\u63a8\u7406\u8fc7\u7a0b\u7684\u95ee\u9898\uff0c\u63d0\u51fa\u4e09\u5927\u521b\u65b0\uff1aLCB-RB\u57fa\u51c6\uff08\u8bc4\u4f30\u63a8\u7406\u8fc7\u7a0b\uff09\u3001OD-based\u5956\u52b1\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5\uff08\u7cbe\u51c6\u8bc4\u5206\u63a8\u7406\u8d28\u91cf\uff09\u3001P-GRPO\u7b97\u6cd5\uff08\u7f13\u89e3\u5956\u52b1\u9ed1\u5ba2\uff09\u3002\u5b9e\u9a8c\u663e\u793a\uff0c7B\u53c2\u6570\u6a21\u578b\u5728\u4ee3\u7801\u751f\u6210\u4efb\u52a1\u4e0a\u8d85\u57fa\u7ebf4.5%\uff0c\u6027\u80fd\u63a5\u8fd1GPT-4-Turbo\uff0c\u4e14\u5728\u6570\u5b66\u4efb\u52a1\u4e0a\u6cdb\u5316\u6027\u826f\u597d\u3002\u8be5\u7814\u7a76\u6709\u6548\u63d0\u5347\u4e86\u6a21\u578b\u7684\u63a8\u7406\u8d28\u91cf\u4e0e\u4efb\u52a1\u6027\u80fd\uff0c\u4e3aAI\u63a8\u7406\u4f18\u5316\u63d0\u4f9b\u4e86\u65b0\u8303\u5f0f\u3002<\/p>\n","protected":false},"author":2,"featured_media":50549,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[50,3528,3546,383],"topic":[],"class_list":["post-50551","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-50","tag-3528","tag-3546","tag-383"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v20.3 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>[\u8bba\u6587\u9605\u8bfb] \u4eba\u5de5\u667a\u80fd + \u8f6f\u4ef6\u5de5\u7a0b | Posterior-GRPO\uff1a\u4f18\u5316\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u7684\u65b0\u6846\u67b6 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/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:\/\/www.wsisp.com\/helps\/50551.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"[\u8bba\u6587\u9605\u8bfb] \u4eba\u5de5\u667a\u80fd + \u8f6f\u4ef6\u5de5\u7a0b | Posterior-GRPO\uff1a\u4f18\u5316\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u7684\u65b0\u6846\u67b6 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb942\u6b21\uff0c\u70b9\u8d5e19\u6b21\uff0c\u6536\u85cf20\u6b21\u3002\u672c\u6587\u4ecb\u7ecd\u4e86Posterior-GRPO\uff08P-GRPO\uff09\u6846\u67b6\uff0c\u8be5\u6846\u67b6\u9488\u5bf9\u73b0\u6709\u5f3a\u5316\u5b66\u4e60\u5728\u4ee3\u7801\u751f\u6210\u4e2d\u5ffd\u89c6\u63a8\u7406\u8fc7\u7a0b\u7684\u95ee\u9898\uff0c\u63d0\u51fa\u4e09\u5927\u521b\u65b0\uff1aLCB-RB\u57fa\u51c6\uff08\u8bc4\u4f30\u63a8\u7406\u8fc7\u7a0b\uff09\u3001OD-based\u5956\u52b1\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5\uff08\u7cbe\u51c6\u8bc4\u5206\u63a8\u7406\u8d28\u91cf\uff09\u3001P-GRPO\u7b97\u6cd5\uff08\u7f13\u89e3\u5956\u52b1\u9ed1\u5ba2\uff09\u3002\u5b9e\u9a8c\u663e\u793a\uff0c7B\u53c2\u6570\u6a21\u578b\u5728\u4ee3\u7801\u751f\u6210\u4efb\u52a1\u4e0a\u8d85\u57fa\u7ebf4.5%\uff0c\u6027\u80fd\u63a5\u8fd1GPT-4-Turbo\uff0c\u4e14\u5728\u6570\u5b66\u4efb\u52a1\u4e0a\u6cdb\u5316\u6027\u826f\u597d\u3002\u8be5\u7814\u7a76\u6709\u6548\u63d0\u5347\u4e86\u6a21\u578b\u7684\u63a8\u7406\u8d28\u91cf\u4e0e\u4efb\u52a1\u6027\u80fd\uff0c\u4e3aAI\u63a8\u7406\u4f18\u5316\u63d0\u4f9b\u4e86\u65b0\u8303\u5f0f\u3002\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/50551.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2025-08-10T02:05:52+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250810020549-6897fe7ddaa3f.png\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"3 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/50551.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/50551.html\",\"name\":\"[\u8bba\u6587\u9605\u8bfb] \u4eba\u5de5\u667a\u80fd + \u8f6f\u4ef6\u5de5\u7a0b | Posterior-GRPO\uff1a\u4f18\u5316\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u7684\u65b0\u6846\u67b6 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2025-08-10T02:05:52+00:00\",\"dateModified\":\"2025-08-10T02:05:52+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/50551.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/50551.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/50551.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"[\u8bba\u6587\u9605\u8bfb] \u4eba\u5de5\u667a\u80fd + \u8f6f\u4ef6\u5de5\u7a0b | Posterior-GRPO\uff1a\u4f18\u5316\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u7684\u65b0\u6846\u67b6\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"[\u8bba\u6587\u9605\u8bfb] \u4eba\u5de5\u667a\u80fd + \u8f6f\u4ef6\u5de5\u7a0b | Posterior-GRPO\uff1a\u4f18\u5316\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u7684\u65b0\u6846\u67b6 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","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:\/\/www.wsisp.com\/helps\/50551.html","og_locale":"zh_CN","og_type":"article","og_title":"[\u8bba\u6587\u9605\u8bfb] \u4eba\u5de5\u667a\u80fd + \u8f6f\u4ef6\u5de5\u7a0b | Posterior-GRPO\uff1a\u4f18\u5316\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u7684\u65b0\u6846\u67b6 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb942\u6b21\uff0c\u70b9\u8d5e19\u6b21\uff0c\u6536\u85cf20\u6b21\u3002\u672c\u6587\u4ecb\u7ecd\u4e86Posterior-GRPO\uff08P-GRPO\uff09\u6846\u67b6\uff0c\u8be5\u6846\u67b6\u9488\u5bf9\u73b0\u6709\u5f3a\u5316\u5b66\u4e60\u5728\u4ee3\u7801\u751f\u6210\u4e2d\u5ffd\u89c6\u63a8\u7406\u8fc7\u7a0b\u7684\u95ee\u9898\uff0c\u63d0\u51fa\u4e09\u5927\u521b\u65b0\uff1aLCB-RB\u57fa\u51c6\uff08\u8bc4\u4f30\u63a8\u7406\u8fc7\u7a0b\uff09\u3001OD-based\u5956\u52b1\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5\uff08\u7cbe\u51c6\u8bc4\u5206\u63a8\u7406\u8d28\u91cf\uff09\u3001P-GRPO\u7b97\u6cd5\uff08\u7f13\u89e3\u5956\u52b1\u9ed1\u5ba2\uff09\u3002\u5b9e\u9a8c\u663e\u793a\uff0c7B\u53c2\u6570\u6a21\u578b\u5728\u4ee3\u7801\u751f\u6210\u4efb\u52a1\u4e0a\u8d85\u57fa\u7ebf4.5%\uff0c\u6027\u80fd\u63a5\u8fd1GPT-4-Turbo\uff0c\u4e14\u5728\u6570\u5b66\u4efb\u52a1\u4e0a\u6cdb\u5316\u6027\u826f\u597d\u3002\u8be5\u7814\u7a76\u6709\u6548\u63d0\u5347\u4e86\u6a21\u578b\u7684\u63a8\u7406\u8d28\u91cf\u4e0e\u4efb\u52a1\u6027\u80fd\uff0c\u4e3aAI\u63a8\u7406\u4f18\u5316\u63d0\u4f9b\u4e86\u65b0\u8303\u5f0f\u3002","og_url":"https:\/\/www.wsisp.com\/helps\/50551.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2025-08-10T02:05:52+00:00","og_image":[{"url":"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/08\/20250810020549-6897fe7ddaa3f.png"}],"author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"3 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/50551.html","url":"https:\/\/www.wsisp.com\/helps\/50551.html","name":"[\u8bba\u6587\u9605\u8bfb] \u4eba\u5de5\u667a\u80fd + \u8f6f\u4ef6\u5de5\u7a0b | Posterior-GRPO\uff1a\u4f18\u5316\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u7684\u65b0\u6846\u67b6 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2025-08-10T02:05:52+00:00","dateModified":"2025-08-10T02:05:52+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/50551.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/50551.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/50551.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"[\u8bba\u6587\u9605\u8bfb] \u4eba\u5de5\u667a\u80fd + \u8f6f\u4ef6\u5de5\u7a0b | Posterior-GRPO\uff1a\u4f18\u5316\u4ee3\u7801\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u7684\u65b0\u6846\u67b6"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/50551","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=50551"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/50551\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media\/50549"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=50551"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=50551"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=50551"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=50551"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}