{"id":53777,"date":"2025-02-16T09:42:13","date_gmt":"2025-02-16T01:42:13","guid":{"rendered":"https:\/\/fwq.ai\/blog\/53777\/"},"modified":"2025-02-16T09:42:13","modified_gmt":"2025-02-16T01:42:13","slug":"deepseek-grpo%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%e7%ae%97%e6%b3%95","status":"publish","type":"post","link":"https:\/\/fwq.ai\/blog\/53777\/","title":{"rendered":"DeepSeek GRPO\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5"},"content":{"rendered":"<p>\u5982\u679c\u4f60\u5728\u5730\u7403\u4e0a\uff0c\u4f60\u73b0\u5728\u4e00\u5b9a\u542c\u8bf4\u8fc7 DeepSeek \u4ee5\u53ca\u5b83\u662f\u5982\u4f55\u5bfc\u81f4\u7f8e\u56fd\u80a1\u5e02\u5d29\u76d8\u7684\u3002\u8fd9\u4e2a\u9769\u547d\u6027\u6a21\u578b\u7684\u6839\u6e90\u662f\u201c\u6211\u7684\u521d\u604b\u201d\uff0c\u5373\u5f3a\u5316\u5b66\u4e60\u3002\u56e0\u6b64\uff0c\u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u5c06\u5c1d\u8bd5\u4e86\u89e3 GRPO\uff0c\u8fd9\u662f\u4e00\u79cd\u4f7f\u8fd9\u4e00\u5207\u6210\u4e3a\u53ef\u80fd\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u3002<\/p>\n<h2>1\u3001\u9884\u5907\u77e5\u8bc6<\/h2>\n<h3>1.1 \u4ec0\u4e48\u662f\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u7b56\u7565\uff1f<\/h3>\n<p>\u5728\u5f3a\u5316\u5b66\u4e60 (RL) \u4e2d\uff0c\u7b56\u7565\uff08policy\uff09\u662f\u4ee3\u7406\u7528\u6765\u51b3\u5b9a\u5728\u7ed9\u5b9a\u72b6\u6001\u4e0b\u91c7\u53d6\u54ea\u79cd\u884c\u52a8\u7684\u7b56\u7565\u3002<\/p>\n<ul>\n<li>\u786e\u5b9a\u6027\u7b56\u7565 (\u03c0(s)) \u2192 \u59cb\u7ec8\u4e3a\u7ed9\u5b9a\u72b6\u6001\u9009\u62e9\u76f8\u540c\u7684\u52a8\u4f5c\u3002<\/li>\n<li>\u968f\u673a\u7b56\u7565 (\u03c0(a|s)) \u2192 \u5b9a\u4e49\u7ed9\u5b9a\u72b6\u6001\u4e0b\u52a8\u4f5c\u7684\u6982\u7387\u5206\u5e03\u3002<\/li>\n<\/ul>\n<p>\u5f3a\u5316\u5b66\u4e60\u7684\u76ee\u6807\u662f\u5b66\u4e60\u6700\u5927\u5316\u7d2f\u79ef\u5956\u52b1\u7684\u6700\u4f73\u7b56\u7565 (\u03c0*)\u3002<\/p>\n<h3>1.2 \u4ec0\u4e48\u662f\u4ef7\u503c\u51fd\u6570\uff1f<\/h3>\n<p>\u4ef7\u503c\u51fd\u6570\uff08value function\uff09\u544a\u8bc9\u6211\u4eec\u4e00\u4e2a\u72b6\u6001\u6216\u884c\u52a8\u5728\u672a\u6765\u5956\u52b1\u65b9\u9762\u6709\u591a\u597d\u3002<\/p>\n<ul>\n<li>\u72b6\u6001\u503c (V) \u2192 \u5904\u4e8e\u67d0\u79cd\u72b6\u6001\u6709\u591a\u597d\u3002<\/li>\n<li>\u52a8\u4f5c\u503c (Q) \u2192 \u5728\u67d0\u4e2a\u72b6\u6001\u4e0b\u91c7\u53d6\u7279\u5b9a\u52a8\u4f5c\u6709\u591a\u597d\u3002<\/li>\n<\/ul>\n<p>\u5b83\u5982\u4f55\u4e0e\u7b56\u7565\u76f8\u5173\u8054\uff1a<\/p>\n<ul>\n<li>\u7b56\u7565\u51b3\u5b9a\u91c7\u53d6\u4ec0\u4e48\u52a8\u4f5c\u3002<\/li>\n<li>\u4ef7\u503c\u51fd\u6570\u901a\u8fc7\u4f30\u8ba1\u54ea\u4e9b\u72b6\u6001\/\u52a8\u4f5c\u4f1a\u5e26\u6765\u66f4\u9ad8\u7684\u56de\u62a5\u6765\u5e2e\u52a9\u6539\u8fdb\u7b56\u7565\u3002<\/li>\n<li>\u4ef7\u503c\u4f30\u8ba1\u8d8a\u597d\uff0c\u7b56\u7565\u5c31\u4f1a\u53d8\u5f97\u8d8a\u597d\uff01<\/li>\n<\/ul>\n<h3>1.3 RL \u4e2d\u7684 Critic-Actor \u6a21\u578b\u662f\u4ec0\u4e48\uff1f<\/h3>\n<p>\u4e0e GAN \u975e\u5e38\u76f8\u4f3c\uff0c\u5728 GAN \u4e2d\uff0c\u751f\u6210\u5668\u548c\u9274\u522b\u5668\u6a21\u578b\u5728\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u90fd\u5728\u76f8\u4e92\u7ade\u4e89\uff0cRL Actor-Critic-Models \u4e2d\u4e5f\u5b58\u5728\u7c7b\u4f3c\u7684\u6982\u5ff5\uff0c\u5176\u4e2dActor-Critic (A2C\u3001A3C) \u662f\u4e00\u79cd\u6d41\u884c\u7684 RL \u67b6\u6784\uff0c\u5b83\u7ed3\u5408\u4e86\u4e24\u4e2a\u7ec4\u4ef6\uff1a<\/p>\n<ul>\n<li>Actor \u2014 \u5b66\u4e60\u7b56\u7565 (\u03c0\u03b8) \u5e76\u51b3\u5b9a\u52a8\u4f5c\u3002<\/li>\n<li>Critic \u2014 \u8bc4\u4f30\u4ef7\u503c\u51fd\u6570 (V(s)) \u4ee5\u6307\u5bfc Actor\u3002<\/li>\n<\/ul>\n<p>\u5de5\u4f5c\u539f\u7406\uff1a<\/p>\n<ul>\n<li>Actor \u6839\u636eCritic\u7684\u53cd\u9988\u66f4\u65b0\u7b56\u7565\u3002<\/li>\n<li>Critic\u4f30\u8ba1\u9884\u671f\u5956\u52b1\u5e76\u51cf\u5c11\u5b66\u4e60\u4e2d\u7684\u5dee\u5f02\u3002<\/li>\n<\/ul>\n<h2>2\u3001GRPO \u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5<\/h2>\n<p>\u5728\u8fdb\u5165\u6570\u5b66\u4e4b\u524d\uff0c\u6211\u4eec\u5c06\u4ece\u4e00\u4e2a\u4f8b\u5b50\u5f00\u59cb\u3002<\/p>\n<p>GRPO\uff08\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff09\u662f\u5f3a\u5316\u5b66\u4e60 (RL) \u4e2d\u4f7f\u7528\u7684\u4e00\u79cd\u65b9\u6cd5\uff0c\u901a\u8fc7\u6bd4\u8f83\u4e0d\u540c\u7684\u52a8\u4f5c\u5e76\u4f7f\u7528\u4e00\u7ec4\u89c2\u5bdf\u7ed3\u679c\u8fdb\u884c\u5c0f\u7684\u53d7\u63a7\u66f4\u65b0\u6765\u5e2e\u52a9\u6a21\u578b\u66f4\u597d\u5730\u5b66\u4e60\u3002\u8fd9\u5c31\u50cf\u662f\u4e00\u79cd\u4ece\u7ecf\u9a8c\u4e2d\u5b66\u4e60\u7684\u806a\u660e\u65b9\u6cd5\uff0c\u800c\u4e0d\u4f1a\u505a\u51fa\u53ef\u80fd\u641e\u7838\u4e8b\u60c5\u7684\u91cd\u5927\u6539\u53d8\u3002<\/p>\n<p>\u60f3\u8c61\u4e00\u4e0b\uff0c\u4f60\u6b63\u5728\u6559\u673a\u5668\u4eba\u73a9\u4e00\u4e2a\u7b80\u5355\u7684\u6e38\u620f\uff0c\u5b83\u5fc5\u987b\u5728\u4e0d\u540c\u7684\u8def\u5f84\u4e4b\u95f4\u505a\u51fa\u9009\u62e9\u624d\u80fd\u8fbe\u5230\u76ee\u6807\u3002\u673a\u5668\u4eba\u9700\u8981\u5b66\u4e60\u54ea\u4e9b\u8def\u5f84\u662f\u597d\u7684\uff0c\u54ea\u4e9b\u4e0d\u662f\u3002<\/p>\n<p>GRPO \u901a\u8fc7\u4ee5\u4e0b\u65b9\u5f0f\u5e2e\u52a9\u673a\u5668\u4eba\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff1a<\/p>\n<ul>\n<li>\u5c1d\u8bd5\u4e0d\u540c\u7684\u8def\u5f84\uff1a\u673a\u5668\u4eba\u5c1d\u8bd5\u5f53\u524d\u7b56\u7565\uff08\u7b56\u7565\uff09\u4e2d\u7684\u51e0\u6761\u4e0d\u540c\u8def\u5f84\uff08\u64cd\u4f5c\uff09\u3002<\/li>\n<li>\u6bd4\u8f83\u6027\u80fd\uff1a\u6bd4\u8f83\u6bcf\u6761\u8def\u5f84\u7684\u6548\u679c\u3002<\/li>\n<li>\u8fdb\u884c\u5c0f\u5e45\u8c03\u6574\uff1a\u6839\u636e\u6bd4\u8f83\u7ed3\u679c\uff0c\u673a\u5668\u4eba\u5bf9\u5176\u7b56\u7565\u8fdb\u884c\u5c0f\u5e45\u8c03\u6574\u4ee5\u63d0\u9ad8\u6027\u80fd\u3002<\/li>\n<\/ul>\n<h3>2.1 \u793a\u4f8b\uff1a\u673a\u5668\u4eba\u9009\u62e9\u8def\u5f84<\/h3>\n<p>\u5047\u8bbe\u673a\u5668\u4eba\u5728\u8ff7\u5bab\u4e2d\uff0c\u5fc5\u987b\u5728\u4e09\u6761\u8def\u5f84\uff08A\u3001B \u548c C\uff09\u4e4b\u95f4\u8fdb\u884c\u9009\u62e9\u624d\u80fd\u5230\u8fbe\u76ee\u6807\u3002\u4ee5\u4e0b\u662f GRPO \u7684\u5206\u6b65\u5de5\u4f5c\u65b9\u5f0f\uff1a<\/p>\n<blockquote><p>\n  \u793a\u4f8b\u8def\u5f84\uff1a\n<\/p><\/blockquote>\n<p>\u673a\u5668\u4eba\u5c1d\u8bd5\u6bcf\u6761\u8def\u5f84\u51e0\u6b21\u5e76\u8bb0\u5f55\u7ed3\u679c\u3002<\/p>\n<ul>\n<li>\u8def\u5f84 A\uff1a3 \u6b21\u4e2d\u6709 2 \u6b21\u6210\u529f\u3002<\/li>\n<li>\u8def\u5f84 B\uff1a3 \u6b21\u4e2d\u6709 1 \u6b21\u6210\u529f\u3002<\/li>\n<li>\u8def\u5f84 C\uff1a3 \u6b21\u4e2d\u6709 3 \u6b21\u6210\u529f\u3002<\/li>\n<\/ul>\n<blockquote><p>\n  \u8ba1\u7b97\u6307\u6807\uff1a\n<\/p><\/blockquote>\n<p>\u673a\u5668\u4eba\u8ba1\u7b97\u6bcf\u6761\u8def\u5f84\u7684\u6210\u529f\u7387\uff1a<\/p>\n<ul>\n<li>\u8def\u5f84 A\uff1a\u6210\u529f\u7387\u4e3a 66.67%\u3002<\/li>\n<li>\u8def\u5f84 B\uff1a\u6210\u529f\u7387\u4e3a 33.33%\u3002<\/li>\n<li>\u8def\u5f84 C\uff1a\u6210\u529f\u7387\u4e3a 100%\u3002<\/li>\n<\/ul>\n<p>\u8bf7\u6ce8\u610f\uff0c\u66f4\u65b0\u5c06\u5728\u89c2\u5bdf\u5230\u201cx\u201d\u6b21\u89c2\u5bdf\u540e\u8fdb\u884c\uff0c\u800c\u4e0d\u662f\u5728\u6bcf\u6b21\u8bd5\u9a8c\u540e\u8fdb\u884c\u3002\u4e0e\u5c0f\u6279\u91cf\u68af\u5ea6\u975e\u5e38\u76f8\u4f3c<\/p>\n<blockquote><p>\n  \u6bd4\u8f83\u8def\u5f84\uff1a\n<\/p><\/blockquote>\n<p>\u673a\u5668\u4eba\u6bd4\u8f83\u8fd9\u4e9b\u6210\u529f\u7387\uff0c\u4ee5\u786e\u5b9a\u54ea\u6761\u8def\u5f84\u6700\u4f73\u3002<\/p>\n<p>\u8def\u5f84 C \u663e\u7136\u662f\u6700\u597d\u7684\uff0c\u56e0\u4e3a\u5b83\u7684\u6210\u529f\u7387\u6700\u9ad8\u3002<\/p>\n<blockquote><p>\n  \u8c03\u6574\u7b56\u7565\uff1a\n<\/p><\/blockquote>\n<p>\u673a\u5668\u4eba\u66f4\u65b0\u5176\u7b56\u7565\uff0c\u4ee5\u4fbf\u5c06\u6765\u66f4\u9891\u7e41\u5730\u9009\u62e9\u8def\u5f84 C\u3002<\/p>\n<p>\u4f46\u5b83\u5e76\u6ca1\u6709\u5b8c\u5168\u5ffd\u7565\u8def\u5f84 A \u548c B\u3002\u5b83\u4ecd\u7136\u4f1a\u5076\u5c14\u5c1d\u8bd5\u5b83\u4eec\uff0c\u770b\u770b\u5b83\u4eec\u662f\u5426\u6709\u6240\u6539\u5584\u3002<\/p>\n<blockquote><p>\n  \u53d7\u63a7\u66f4\u65b0\uff1a\n<\/p><\/blockquote>\n<p>\u673a\u5668\u4eba\u786e\u4fdd\u4e0d\u4f1a\u8fc7\u5ea6\u6539\u53d8\u5176\u7b56\u7565\u3002\u4f8b\u5982\uff0c\u5b83\u53ef\u80fd\u4f1a\u5c06\u9009\u62e9\u8def\u5f84 C \u7684\u6982\u7387\u4ece 30% \u63d0\u9ad8\u5230 50%\uff0c\u4f46\u4e0d\u4f1a\u63d0\u9ad8\u5230 100%\u3002\u8fd9\u6837\uff0c\u5b83\u4ecd\u7136\u53ef\u4ee5\u63a2\u7d22\u5176\u4ed6\u8def\u5f84\u5e76\u4e86\u89e3\u66f4\u591a\u4fe1\u606f\u3002<\/p>\n<p>\u73b0\u5728\u8ba9\u6211\u4eec\u8fdb\u5165\u6570\u5b66<\/p>\n<h3>2.2 GRPO \u7b97\u6cd5\u6570\u5b66<\/h3>\n<blockquote><p>\n  \u7b56\u7565\u548c\u884c\u52a8\n<\/p><\/blockquote>\n<ul>\n<li>\u8ba9\u7b56\u7565\u8868\u793a\u4e3a \u03c0\u03b8\uff0c\u5176\u4e2d \u03b8 \u8868\u793a\u7b56\u7565\u7684\u53c2\u6570\u3002<\/li>\n<li>\u5bf9\u4e8e\u7ed9\u5b9a\u72b6\u6001 s\uff0c\u7b56\u7565\u8f93\u51fa\u52a8\u4f5c\u7684\u6982\u7387\u5206\u5e03\uff1a\u03c0\u03b8(a\u2223s)\uff0c\u5373\uff08\u7ed9\u5b9a\u72b6\u6001\u9009\u62e9\u54ea\u4e2a\u52a8\u4f5c\uff09<\/li>\n<li>\u76ee\u6807\u662f\u6700\u5927\u5316\u9884\u671f\u7d2f\u79ef\u5956\u52b1 J(\u03b8)\uff1a<\/li>\n<\/ul>\n<p>\u5176\u4e2d \u03c4=(s0,a0,s1,a1,\u2026 ) \u662f\u8f68\u8ff9\uff08\u8fc7\u53bb\u91c7\u53d6\u7684\u72b6\u6001\u52a8\u4f5c\uff09\uff0cr(st, at) \u662f\u65f6\u95f4 t \u7684\u5956\u52b1\u3002<\/p>\n<blockquote><p>\n  \u7ec4\u62bd\u6837\n<\/p><\/blockquote>\n<ul>\n<li>\u5bf9\u4e8e\u7ed9\u5b9a\u72b6\u6001 s\uff0cGRPO \u4f7f\u7528\u5f53\u524d\u7b56\u7565 \u03c0\u03b8 \u751f\u6210\u4e00\u7ec4 N \u4e2a\u52a8\u4f5c {a1,a2,\u2026,aN}\u3002<\/li>\n<li>\u6bcf\u4e2a\u52a8\u4f5c ai \u90fd\u4ece\u7b56\u7565\u4e2d\u62bd\u6837\uff1aai\u223c\u03c0\u03b8(a\u2223s)\u3002<\/li>\n<\/ul>\n<blockquote><p>\n  \u5956\u52b1\u8bc4\u5206\n<\/p><\/blockquote>\n<ul>\n<li>\u6bcf\u4e2a\u52a8\u4f5c\u201cai\u201d\u90fd\u4f7f\u7528\u5956\u52b1\u51fd\u6570 R(ai) \u8fdb\u884c\u8bc4\u4f30\uff0c\u8be5\u51fd\u6570\u8861\u91cf\u52a8\u4f5c\u7684\u8d28\u91cf\u3002<\/li>\n<li>\u5956\u52b1\u51fd\u6570\u53ef\u4ee5\u57fa\u4e8e\u5373\u65f6\u5956\u52b1\u6216\u672a\u6765\u5956\u52b1\u7684\u6298\u6263\u603b\u548c\u3002<\/li>\n<\/ul>\n<blockquote><p>\n  \u4f18\u52bf\u8ba1\u7b97\n<\/p><\/blockquote>\n<p>\u4f18\u52bf A(ai) \u8861\u91cf\u52a8\u4f5c ai \u4e0e\u7ec4\u7684\u5e73\u5747\u8868\u73b0\u76f8\u6bd4\u597d\u591a\u5c11\u6216\u5dee\u591a\u5c11\u3002<\/p>\n<blockquote><p>\n  \u7b56\u7565\u66f4\u65b0\n<\/p><\/blockquote>\n<p>GRPO \u66f4\u65b0\u7b56\u7565\u53c2\u6570 \u03b8\uff0c\u4ee5\u589e\u52a0\u5177\u6709\u6b63\u4f18\u52bf\u7684\u52a8\u4f5c\u7684\u53ef\u80fd\u6027\uff0c\u5e76\u964d\u4f4e\u5177\u6709\u8d1f\u4f18\u52bf\u7684\u52a8\u4f5c\u7684\u53ef\u80fd\u6027\u3002<\/p>\n<blockquote><p>\n  KL \u6563\u5ea6\u7ea6\u675f\n<\/p><\/blockquote>\n<p>\u4e3a\u4e86\u786e\u4fdd\u66f4\u65b0\u540e\u7684\u7b56\u7565 \u03c0\u03b8\u200b \u4e0d\u4f1a\u4e0e\u65e7\u7b56\u7565\u504f\u79bb\u592a\u591a\uff0cGRPO \u5305\u542b\u4e00\u4e2a KL \u6563\u5ea6\u7ea6\u675f\uff1a<br \/>\u6b64\u7ea6\u675f\u53ef\u786e\u4fdd\u66f4\u65b0\u7a33\u5b9a\uff0c\u5e76\u9632\u6b62\u7b56\u7565\u53d1\u751f\u592a\u5927\u53d8\u5316\u3002<\/p>\n<blockquote><p>\n  GRPO \u76ee\u6807\n<\/p><\/blockquote>\n<p>GRPO \u7684\u603b\u4f53\u76ee\u6807\u662f\u5728\u4fdd\u6301\u7b56\u7565\u66f4\u65b0\u7a33\u5b9a\u7684\u540c\u65f6\u6700\u5927\u5316\u9884\u671f\u7d2f\u79ef\u5956\u52b1<\/p>\n<h3>2.3 GRPO \u4e3a\u4f55\u6709\u6548<\/h3>\n<ul>\n<li>\u7ec4\u6bd4\u8f83\uff1a\u901a\u8fc7\u6bd4\u8f83\u7ec4\u5185\u7684\u52a8\u4f5c\uff0cGRPO \u51cf\u5c11\u4e86\u7b56\u7565\u66f4\u65b0\u7684\u65b9\u5dee\u5e76\u786e\u4fdd\u66f4\u7a33\u5b9a\u7684\u5b66\u4e60\u3002<\/li>\n<li>\u53d7\u63a7\u66f4\u65b0\uff1aKL \u6563\u5ea6\u7ea6\u675f\u53ef\u9632\u6b62\u5bf9\u7b56\u7565\u8fdb\u884c\u5927\u7684\u3001\u4e0d\u7a33\u5b9a\u7684\u66f4\u6539\u3002<\/li>\n<li>\u6548\u7387\uff1aGRPO \u907f\u514d\u4e86\u8bc4\u4f30\u6bcf\u4e2a\u53ef\u80fd\u7684\u64cd\u4f5c\u7684\u9700\u8981\uff0c\u4f7f\u5176\u5177\u6709\u8ba1\u7b97\u6548\u7387\u3002<\/li>\n<\/ul>\n<p>3\u3001\u5982\u4f55\u4f7f\u7528 GRPO \u6765\u8bad\u7ec3 LLM\uff1f<\/p>\n<ul>\n<li>\u7ec4\u62bd\u6837\uff1a\u5bf9\u4e8e\u7ed9\u5b9a\u7684\u63d0\u793a\uff0cLLM \u4f1a\u751f\u6210\u591a\u4e2a\u54cd\u5e94\u3002<\/li>\n<li>\u5956\u52b1\u8bc4\u5206\uff1a\u5956\u52b1\u6a21\u578b\u8bc4\u4f30\u6bcf\u4e2a\u54cd\u5e94\u7684\u8d28\u91cf\u3002<\/li>\n<li>\u4f18\u52bf\u8ba1\u7b97\uff1a\u5c06\u54cd\u5e94\u4e0e\u7ec4\u7684\u5e73\u5747\u5956\u52b1\u8fdb\u884c\u6bd4\u8f83\uff0c\u4ee5\u786e\u5b9a\u54ea\u4e9b\u66f4\u597d\u6216\u66f4\u5dee\u3002<\/li>\n<li>\u7b56\u7565\u66f4\u65b0\uff1a\u4f7f\u7528 KL \u6563\u5ea6\u7ea6\u675f\u8c03\u6574 LLM \u7684\u7b56\u7565\u4ee5\u652f\u6301\u9ad8\u5956\u52b1\u54cd\u5e94\uff0c\u540c\u65f6\u907f\u514d\u5267\u70c8\u53d8\u5316\u3002<\/li>\n<li>\u8fed\u4ee3\u8bad\u7ec3\uff1a\u6b64\u8fc7\u7a0b\u91cd\u590d\u8fdb\u884c\uff0c\u9010\u6e10\u63d0\u9ad8 LLM \u751f\u6210\u9ad8\u8d28\u91cf\u5bf9\u9f50\u6587\u672c\u7684\u80fd\u529b\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u5c31\u662f\u7eaf\u5f3a\u5316\u5b66\u4e60\u7528\u4e8e\u8bad\u7ec3LLM\u7684\u65b9\u5f0f\u3002<\/p>\n<hr>\n","protected":false},"excerpt":{"rendered":"<p>\u5982\u679c\u4f60\u5728\u5730\u7403\u4e0a\uff0c\u4f60\u73b0\u5728\u4e00\u5b9a\u542c\u8bf4\u8fc7 DeepSeek \u4ee5\u53ca\u5b83\u662f\u5982\u4f55\u5bfc\u81f4\u7f8e\u56fd\u80a1\u5e02\u5d29\u76d8\u7684\u3002\u8fd9\u4e2a\u9769\u547d\u6027\u6a21\u578b\u7684\u6839\u6e90\u662f\u201c\u6211\u7684\u521d\u604b\u201d\uff0c\u5373\u5f3a\u5316\u5b66\u4e60\u3002\u56e0\u6b64\uff0c\u5728\u8fd9\u7bc7\u6587\u7ae0\u4e2d\uff0c\u6211\u4eec\u5c06\u5c1d\u8bd5\u4e86\u89e3 GRPO\uff0c\u8fd9\u662f\u4e00\u79cd\u4f7f\u8fd9\u4e00\u5207\u6210\u4e3a\u53ef\u80fd\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u3002 1\u3001\u9884\u5907\u77e5\u8bc6 1.1 \u4ec0\u4e48\u662f\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u7b56\u7565\uff1f \u5728\u5f3a\u5316\u5b66\u4e60 (RL) \u4e2d\uff0c\u7b56\u7565\uff08policy\uff09\u662f\u4ee3\u7406\u7528\u6765\u51b3\u5b9a\u5728\u7ed9\u5b9a\u72b6\u6001\u4e0b\u91c7\u53d6\u54ea\u79cd\u884c\u52a8\u7684\u7b56\u7565\u3002 \u786e\u5b9a\u6027\u7b56\u7565 (\u03c0(s)) \u2192 \u59cb\u7ec8\u4e3a\u7ed9\u5b9a\u72b6\u6001\u9009\u62e9\u76f8\u540c\u7684\u52a8\u4f5c\u3002 \u968f\u673a\u7b56\u7565 (\u03c0(a|s)) \u2192 \u5b9a\u4e49\u7ed9\u5b9a\u72b6\u6001\u4e0b\u52a8\u4f5c\u7684\u6982\u7387\u5206\u5e03\u3002 \u5f3a\u5316\u5b66\u4e60\u7684\u76ee\u6807\u662f\u5b66\u4e60\u6700\u5927\u5316\u7d2f\u79ef\u5956\u52b1\u7684\u6700\u4f73\u7b56\u7565 (\u03c0*)\u3002 1.2 \u4ec0\u4e48\u662f\u4ef7\u503c\u51fd\u6570\uff1f \u4ef7\u503c\u51fd\u6570\uff08value function\uff09\u544a\u8bc9\u6211\u4eec\u4e00\u4e2a\u72b6\u6001\u6216\u884c\u52a8\u5728\u672a\u6765\u5956\u52b1\u65b9\u9762\u6709\u591a\u597d\u3002 \u72b6\u6001\u503c (V) \u2192 \u5904\u4e8e\u67d0\u79cd\u72b6\u6001\u6709\u591a\u597d\u3002 \u52a8\u4f5c\u503c (Q) \u2192 \u5728\u67d0\u4e2a\u72b6\u6001\u4e0b\u91c7\u53d6\u7279\u5b9a\u52a8\u4f5c\u6709\u591a\u597d\u3002 \u5b83\u5982\u4f55\u4e0e\u7b56\u7565\u76f8\u5173\u8054\uff1a \u7b56\u7565\u51b3\u5b9a\u91c7\u53d6\u4ec0\u4e48\u52a8\u4f5c\u3002 \u4ef7\u503c\u51fd\u6570\u901a\u8fc7\u4f30\u8ba1\u54ea\u4e9b\u72b6\u6001\/\u52a8\u4f5c\u4f1a\u5e26\u6765\u66f4\u9ad8\u7684\u56de\u62a5\u6765\u5e2e\u52a9\u6539\u8fdb\u7b56\u7565\u3002 \u4ef7\u503c\u4f30\u8ba1\u8d8a\u597d\uff0c\u7b56\u7565\u5c31\u4f1a\u53d8\u5f97\u8d8a\u597d\uff01 1.3 RL \u4e2d\u7684 Critic-Actor \u6a21\u578b\u662f\u4ec0\u4e48\uff1f \u4e0e GAN \u975e\u5e38\u76f8\u4f3c\uff0c\u5728 GAN \u4e2d\uff0c\u751f\u6210\u5668\u548c\u9274\u522b\u5668\u6a21\u578b\u5728\u6574\u4e2a\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u90fd\u5728\u76f8\u4e92\u7ade\u4e89\uff0cRL Actor-Critic-Models \u4e2d\u4e5f\u5b58\u5728\u7c7b\u4f3c\u7684\u6982\u5ff5\uff0c\u5176\u4e2dActor-Critic (A2C\u3001A3C) \u662f\u4e00\u79cd\u6d41\u884c\u7684 RL \u67b6\u6784\uff0c\u5b83\u7ed3\u5408\u4e86\u4e24\u4e2a\u7ec4\u4ef6\uff1a Actor \u2014 \u5b66\u4e60\u7b56\u7565 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[13],"tags":[],"class_list":["post-53777","post","type-post","status-publish","format-standard","hentry","category-ai"],"_links":{"self":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53777","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/comments?post=53777"}],"version-history":[{"count":0,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/posts\/53777\/revisions"}],"wp:attachment":[{"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/media?parent=53777"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/categories?post=53777"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/fwq.ai\/blog\/wp-json\/wp\/v2\/tags?post=53777"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}