{"id":1172,"date":"2025-07-26T16:57:00","date_gmt":"2025-07-26T08:57:00","guid":{"rendered":"http:\/\/www.applicationsofllm.com\/?p=1172"},"modified":"2025-09-17T17:00:13","modified_gmt":"2025-09-17T09:00:13","slug":"%e4%b8%80%e6%96%87%e8%a7%a3%e6%9e%84%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%90%8e%e8%ae%ad%e7%bb%83%ef%bc%8cgrpo%e5%92%8c%e5%ae%83%e7%9a%84%e7%bb%a7%e4%bb%bb%e8%80%85%e4%bb%ac%e7%9a%84%e5%89%8d%e4%b8%96","status":"publish","type":"post","link":"http:\/\/www.applicationsofllm.com\/index.php\/2025\/07\/26\/%e4%b8%80%e6%96%87%e8%a7%a3%e6%9e%84%e5%a4%a7%e6%a8%a1%e5%9e%8b%e5%90%8e%e8%ae%ad%e7%bb%83%ef%bc%8cgrpo%e5%92%8c%e5%ae%83%e7%9a%84%e7%bb%a7%e4%bb%bb%e8%80%85%e4%bb%ac%e7%9a%84%e5%89%8d%e4%b8%96\/","title":{"rendered":"\u4e00\u6587\u89e3\u6784\u5927\u6a21\u578b\u540e\u8bad\u7ec3\uff0cGRPO\u548c\u5b83\u7684\u7ee7\u4efb\u8005\u4eec\u7684\u524d\u4e16\u4eca\u751f"},"content":{"rendered":"\n<p>\u5927\u8bed\u8a00\u6a21\u578b\u7684\u53d1\u5c55\u771f\u662f\u65e5\u65b0\u6708\u5f02\u3002<\/p>\n\n\n\n<p>\u4ece DeepSeek \u6a2a\u7a7a\u51fa\u4e16\u4ee5\u6765\uff0c\u5176\u5728\u5927\u6a21\u578b\u540e\u8bad\u7ec3\u7684\u521b\u65b0 GRPO \u4e00\u8dc3\u6210\u4e3a\u5f3a\u5316\u5b66\u4e60\u9ec4\u91d1\u8303\u5f0f\u3002<\/p>\n\n\n\n<p>GRPO \u5df2\u7ecf\u6210\u4e3a\u4e00\u79cd\u5927\u6a21\u578b\u901a\u7528\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\uff0c\u80fd\u591f\u7528\u5728\u5e7f\u6cdb\u7684\u540e\u8bad\u7ec3\u4efb\u52a1\u4e2d\uff0c\u751a\u81f3\u5305\u62ec\u8ba9\u5927\u6a21\u578b\u73a9 2048\uff1a<\/p>\n\n\n\n<p>\u800c\u5c31\u5728\u4eca\u5e74\uff0c\u5927\u6a21\u578b\u540e\u8bad\u7ec3\u7684\u7814\u7a76\u51fa\u73b0\u4e86\u51e0\u4e2a\u91cd\u78c5\u7ed3\u679c\uff0c\u5305\u62ec Seed \u56e2\u961f\u7684 DAPO\uff0cQwen \u56e2\u961f\u7684 GSPO\uff0c\u5fae\u8f6f\u56e2\u961f\u7684 GFPO \u7b49\u7b49\uff0c\u800c\u4ed6\u4eec\u65e0\u4e00\u4f8b\u5916\u90fd\u662f\u5bf9 GRPO \u8303\u5f0f\u7684\u6539\u8fdb\u3002<\/p>\n\n\n\n<p>\u770b\u8fd9\u4e9b\u540d\u5b57\u90fd\u7ed5\u6655\u4e86\uff0cGRPO \u5230\u5e95\u6709\u4ec0\u4e48\u9b54\u529b\uff0c\u80fd\u8ba9\u5404\u5927\u7814\u7a76\u56e2\u961f\u7ed5\u7740\u5b83\u56e2\u56e2\u8f6c\uff1bGRPO \u53c8\u6709\u4ec0\u4e48\u7f3a\u9677\uff0c\u5404\u5927\u56e2\u961f\u90fd\u8981\u5728\u5b83\u8eab\u4e0a\u52a8\u5200\uff1f<\/p>\n\n\n\n<p>\u901a\u8fc7\u8fd9\u7bc7\u6587\u7ae0\uff0c\u6211\u4eec\u5e0c\u671b\u80fd\u591f\u6df1\u5165\u6d45\u51fa\u7684\u89e3\u91ca\u5927\u6a21\u578b\u540e\u8bad\u7ec3\u7684\u539f\u7406\uff0c\u8fd1\u671f\u7684\u6280\u672f\u8fdb\u5316\u8def\u7ebf\uff0c\u4ee5\u671f\u4e3a\u8bfb\u8005\u6784\u5efa\u4e00\u4e2a\u5b8c\u6574\u7684\u77e5\u8bc6\u4f53\u7cfb\u3002<\/p>\n\n\n\n<p><strong>\u540e\u8bad\u7ec3\u4e0e\u5f3a\u5316\u5b66\u4e60<\/strong><\/p>\n\n\n\n<p>\u5f88\u591a\u4eba\u4f1a\u89c9\u5f97\uff0c\u5f3a\u5316\u5b66\u4e60\u662f\u4e00\u4e2a\u975e\u5e38\u53e4\u8001\u7684\u6982\u5ff5\uff0c\u548c\u5168\u65b0\u7684\u5927\u6a21\u578b\u597d\u4f3c\u683c\u683c\u4e0d\u5165\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u5148\u4ece\u5927\u6a21\u578b\u8bf4\u8d77\u3002<\/p>\n\n\n\n<p>\u5927\u4f17\u7406\u89e3\u7684\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6982\u5ff5\u4f3c\u4e4e\u5f88\u7b80\u5355\uff0c\u4ece\u6d77\u91cf\u6570\u636e\u4e2d\u81ea\u76d1\u7763\u5b66\u4e60\u51fa\u6765\u7684\u4e00\u4e2a\u6a21\u578b\uff0c\u80fd\u591f\u9884\u6d4b\u6587\u672c\u4e2d\u4e0b\u4e00\u4e2a\u51fa\u73b0\u7684\u8bcd\uff0c\u4ece\u800c\u8f93\u51fa\u8bed\u8a00\u6587\u672c\u3002<\/p>\n\n\n\n<p>\u4f46\u8fd9\u5e76\u4e0d\u5b8c\u5584\uff0c\u8fd9\u79cd\u7406\u89e3\u53ea\u7a81\u51fa\u4e86\u5927\u6a21\u578b\u300c\u9884\u8bad\u7ec3\u300d\u7684\u8fc7\u7a0b\uff0c\u800c\u5b8c\u5168\u5ffd\u7565\u4e86\u300c\u540e\u8bad\u7ec3\u300d\u8fd9\u4e00\u91cd\u8981\u8fc7\u7a0b\u3002<\/p>\n\n\n\n<p>\u7b80\u5355\u6765\u8bf4\uff0c\u4ece\u6d77\u91cf\u6570\u636e\u4e2d\u5b66\u4e60\u7684\u8fc7\u7a0b\u79f0\u4e3a\u300c\u9884\u8bad\u7ec3\u300d\uff0c\u9884\u8bad\u7ec3\u7684\u7ed3\u679c\u662f\u8ba9\u6a21\u578b\u638c\u63e1\u4e86\u901a\u7528\u8bed\u8a00\u80fd\u529b\uff0c\u4f46\u4ec5\u4ec5\u5982\u6b64\uff0c\u6a21\u578b\u751f\u6210\u7684\u5185\u5e76\u4e0d\u4e00\u5b9a\u7b26\u5408\u504f\u597d\uff1b\u53ef\u80fd\u751f\u6210\u5197\u957f\u3001\u4e0d\u51c6\u786e\u7684\u5185\u5bb9\uff1b\u53ef\u80fd\u4e0d\u7b26\u5408\u5e94\u7528\u4efb\u52a1\u7684\u9700\u6c42\u3002<\/p>\n\n\n\n<p>\u6362\u53e5\u8bdd\u8bf4\uff0c<strong>\u9884\u8bad\u7ec3\u540e\u7684\u5927\u6a21\u578b\u4f1a\u8bf4\u8bdd\uff0c\u4f46\u4e0d\u4e00\u5b9a\u4f1a\u300c\u8bf4\u5bf9\u8bdd\u300d\u3002<\/strong><\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u300c\u540e\u8bad\u7ec3\u300d\u8fc7\u7a0b\u5c31\u6781\u4e3a\u91cd\u8981\u3002\u540e\u8bad\u7ec3\u7684\u4e3b\u8981\u76ee\u6807\u662f\u5f3a\u5316\u6a21\u578b\u5728\u7279\u5b9a\u9886\u57df\u7684\u77e5\u8bc6\u548c\u5e94\u7528\u80fd\u529b\uff0c\u589e\u5f3a\u4e86\u6a21\u578b\u7684\u9002\u5e94\u6027\u548c\u7075\u6d3b\u6027\uff0c\u4f7f\u5176\u80fd\u591f\u66f4\u597d\u5730\u6ee1\u8db3\u5b9e\u9645\u5e94\u7528\u573a\u666f\u4e2d\u7684\u591a\u6837\u5316\u9700\u6c42\u3002<\/p>\n\n\n\n<p>\u800c\u5f3a\u5316\u5b66\u4e60\u5219\u662f\u5728\u540e\u8bad\u7ec3\u4e2d\u4e0d\u53ef\u6216\u7f3a\u7684\u6838\u5fc3\u90e8\u5206\u3002\u5173\u4e8e\u5f3a\u5316\u5b66\u4e60\u7684\u7406\u89e3\uff0c\u6211\u4eec\u53ef\u4ee5\u53c2\u8003\u3002<\/p>\n\n\n\n<p><strong>\u5f3a\u5316\u5b66\u4e60\u7684\u6838\u5fc3\u662f\u300c\u53cd\u9988\u300d<\/strong>\uff0c\u76ee\u6807\u662f\u589e\u52a0\u597d\u7ed3\u679c\u7684\u51fa\u73b0\u6982\u7387\uff0c\u964d\u4f4e\u574f\u7ed3\u679c\u7684\u51fa\u73b0\u6982\u7387\u3002<\/p>\n\n\n\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u5728\u5403\u8c46\u4eba\uff08Pacman\uff09\u6e38\u620f\u4e2d\uff1a\u5982\u679c\u5403\u6389\u4e00\u5757\u997c\u5e72\uff0c\u53cd\u9988\u662f\u52a0\u5206\uff1b\u5982\u679c\u4f60\u78b0\u5230\u654c\u4eba\uff0c\u53cd\u9988\u662f\u6263\u5206\u3002<\/p>\n\n\n\n<p>\u8fd9\u662f\u6700\u6734\u7d20\u7684\u5f3a\u5316\u5b66\u4e60\u65b9\u5f0f\u4e86\u3002\u6211\u4eec\u653e\u5230\u5927\u6a21\u578b\u8bad\u7ec3\u5f53\u4e2d\uff0c\u53c8\u8be5\u7528\u4ec0\u4e48\u65b9\u5f0f\u7ed9\u5927\u6a21\u578b\u52a0\u51cf\u5206\u7684\u53cd\u9988\u5462\uff1f<\/p>\n\n\n\n<p>\u6211\u4eec\u7684\u6838\u5fc3\u76ee\u6807\u662f\u8ba9\u5927\u6a21\u578b\u8f93\u51fa\u7b26\u5408\u6211\u4eec\u4efb\u52a1\u504f\u597d\u7684\u5185\u5bb9\uff0c\u90a3\u6700\u7b80\u5355\u7684\u65b9\u5f0f\u5c31\u662f\u4eba\u7c7b\u7684\u53cd\u9988\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u4f60\u4e5f\u8fd9\u4e48\u60f3\uff0c\u90a3\u4f60\u7684\u60f3\u6cd5\u548c OpenAI \u4e0d\u8c0b\u800c\u5408\u3002<\/p>\n\n\n\n<p>\u5728\u8bad\u7ec3 GPT \u7684\u65f6\u5019\uff0cOpenAI \u5c31<strong>\u91c7\u7528\u4e86 RLHF\uff08\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff09<\/strong>\u7684\u65b9\u6cd5\u3002\u5728\u8be5\u65b9\u6cd5\u4e2d\uff0c\u9700\u8981\u8bad\u7ec3\u4e00\u4e2a agent \u6765\u9488\u5bf9\u67d0\u4e2a\u95ee\u9898\uff08\u72b6\u6001\uff09\u751f\u6210\u4eba\u7c7b\u8ba4\u4e3a\u66f4\u6709\u7528\u7684\u8f93\u51fa\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"576\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-419-1024x576.jpg\" alt=\"\" class=\"wp-image-1178\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-419-1024x576.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-419-300x169.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-419-768x432.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-419.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u53cd\u9988\u6709\u4e86\uff0c\u5e76\u975e\u4e00\u5207\u4e07\u4e8b\u5927\u5409\u4e86\u3002\u6211\u4eec\u901a\u8fc7 RLHF \u83b7\u5f97\u4e86\u53cd\u9988\uff0c\u901a\u8fc7\u8fd9\u4e2a\u53cd\u9988\u7684 Reward \u4f5c\u4e3a\u4e00\u4e2a\u7edd\u5bf9\u7684\u6807\u51c6\u53bb\u76f4\u63a5\u8bad\u7ec3\u6a21\u578b\uff0c\u4f1a\u51fa\u73b0\u663e\u8457\u7684\u6fc0\u52b1\u4e0d\u5145\u5206\u548c\u65b9\u5dee\u8fc7\u5927\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"574\" height=\"110\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-169.png\" alt=\"\" class=\"wp-image-1174\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-169.png 574w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-169-300x57.png 300w\" sizes=\"auto, (max-width: 574px) 100vw, 574px\" \/><\/figure>\n\n\n\n<p>\u5047\u5982\u6709\u4e24\u4e2a\u6a21\u578b\uff0cA \u7684\u521d\u59cb\u80fd\u529b\u663e\u8457\u6bd4 B \u5f3a\uff0c\u901a\u8fc7\u76f4\u63a5\u53cd\u9988\u4f1a\u51fa\u73b0\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6a21\u578b B \u5373\u4f7f\u4ece 30 \u63d0\u5347\u5230 60\uff0c\u4f46\u548c\u6a21\u578b A \u7684 80 \u76f8\u6bd4\uff0c\u4ecd\u7136\u663e\u5f97\u5f88\u5dee\uff0c\u4f18\u5316\u65f6\u5b83\u5f97\u5230\u7684\u6fc0\u52b1\u4ecd\u7136\u6709\u9650\u3002<\/li>\n\n\n\n<li>\u6a21\u578b A \u5728\u8ffd\u6c42\u66f4\u9ad8\u5206\u65f6\uff0c\u53ef\u80fd\u51fa\u73b0\u4e00\u4e9b\u6fc0\u8fdb\u7684\u53d8\u5316\uff0c\u5bfc\u81f4 reward \u6709\u65f6\u98d9\u5347\uff0c\u6709\u65f6\u8fc5\u901f\u56de\u843d\uff0c\u8bad\u7ec3\u8fc7\u7a0b\u4e0d\u7a33\u5b9a\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>PPO \u7684\u7a33\u5b9a\u7b56\u7565<\/strong><\/p>\n\n\n\n<p>\u4e3a\u4e86\u5728\u6b64\u57fa\u7840\u4e0a\u7a33\u5b9a\u7684\u5b9e\u73b0 RLHF\uff0cOpenAI \u6784\u5efa\u4e86&nbsp;<strong>PPO<\/strong><strong>\uff08Proximal Policy Optimization\uff0c\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff09<\/strong>\u673a\u5236\uff0c\u52a0\u5165\u4e86&nbsp;<strong>Critic<\/strong><strong>\u3001CLIP \u64cd\u4f5c\u548c Reference Model<\/strong>\uff0c\u5728\u4fdd\u8bc1 \u7b56\u7565\u66f4\u65b0\u4e0d\u8fc7\u5ea6 \u7684\u540c\u65f6\uff0c\u4f9d\u65e7\u80fd \u9ad8\u6548\u63d0\u5347\u6027\u80fd\u3002\u73b0\u5728\u5df2\u7ecf\u6210\u4e3a\u5f3a\u5316\u5b66\u4e60\u9886\u57df\u7684 \u6807\u51c6\u65b9\u6cd5\u4e4b\u4e00\uff0c\u51e0\u4e4e\u662f RLHF \u7684\u9ed8\u8ba4\u9009\u62e9\u3002<\/p>\n\n\n\n<p>\u9488\u5bf9\u7b2c\u4e00\u6761\u95ee\u9898\uff0cPPO \u5f15\u5165\u4e86 Critic\uff1a<\/p>\n\n\n\n<p>\u901a\u4fd7\u6765\u8bf4\uff0c\u6211\u4eec\u4e0d\u518d\u53ea\u4f7f\u7528\u7eaf\u7cb9\u7684 Reward \u6765\u53cd\u9988\uff0c\u800c\u662f\u8bbe\u7f6e\u4e00\u4e2a\u300c\u4ef7\u503c\u51fd\u6570<img loading=\"lazy\" decoding=\"async\" width=\"100\" height=\"66\" src=\"blob:http:\/\/www.applicationsofllm.com\/edd4e74a-ad7e-4381-a153-8a139bbcff6c\" alt=\"\u56fe\u7247\">\u300d\u4f5c\u4e3a\u53c2\u8003\uff0c\u8bad\u7ec3\u76ee\u6807\u4ece\u300cReward\u300d\u8fdb\u5316\u6210\u300cAdvantage\u300d\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"316\" height=\"88\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-168.png\" alt=\"\" class=\"wp-image-1173\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-168.png 316w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-168-300x84.png 300w\" sizes=\"auto, (max-width: 316px) 100vw, 316px\" \/><\/figure>\n\n\n\n<p>\u5bf9\u67d0\u4e2a\u52a8\u4f5c\uff0c\u5982\u679c\u5b9e\u9645 Reward \u8d85\u8fc7\u4e86 Critic \u7684\u9884\u671f\uff0c\u5c31\u4f5c\u4e3a\u5956\u52b1\uff0c\u82e5\u4f4e\u4e8e\u9884\u671f\u5219\u4e3a\u8d1f\u53cd\u9988\u3002\u4f18\u5316\u76ee\u6807\u5c31\u53d8\u6210:<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"116\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-417-1024x116.jpg\" alt=\"\" class=\"wp-image-1176\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-417-1024x116.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-417-300x34.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-417-768x87.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-417.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u4e5f\u5c31\u662f\u8bf4\uff0c\u6211\u4eec\u62e5\u6709\u4e86\u4e00\u4e2a<strong>\u76f8\u5bf9\u8bc4\u4f30\u6a21\u578b\u8fdb\u6b65\u7a0b\u5ea6<\/strong>\u7684\u65b0\u8303\u5f0f\uff0c\u800c\u975e\u91c7\u7528\u7edd\u5bf9 Reward \u53cd\u9988\u3002\u5f15\u5165 Critic \u53ef\u4ee5\u663e\u8457\u964d\u4f4e\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u65b9\u5dee\uff0c\u76f8\u5bf9\u4e8e Reward \u53cd\u9988\uff0c\u6a21\u578b\u8fdb\u6b65\u80fd\u83b7\u5f97\u7684\u68af\u5ea6\u66f4\u663e\u8457\u3002<\/p>\n\n\n\n<p>\u9488\u5bf9\u7b2c\u4e8c\u6761\u95ee\u9898\uff0cPPO \u91c7\u7528\u4e86 Clip \u7b56\u7565\uff1a<\/p>\n\n\n\n<p>\u4e3a\u4e86\u907f\u514d\u6a21\u578b\u53d8\u5316\u8fc7\u5927\u5bfc\u81f4\u7684\u4e0d\u7a33\u5b9a\uff0cClip \u7b56\u7565\u52a0\u5165\u4e86\u9650\u5236\u6761\u4ef6\uff0c\u5728\u76ee\u6807\u51fd\u6570\u4e2d\u53ef\u4ee5\u4f53\u73b0\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"167\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-416-1024x167.jpg\" alt=\"\" class=\"wp-image-1175\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-416-1024x167.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-416-300x49.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-416-768x126.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-416.jpg 1107w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u5176\u4e2d\uff0c<img loading=\"lazy\" decoding=\"async\" width=\"360\" height=\"132\" src=\"blob:http:\/\/www.applicationsofllm.com\/6f569f99-c312-441d-b3f9-c97cf3d7a895\" alt=\"\u56fe\u7247\"><\/p>\n\n\n\n<p>\u5b83\u8868\u793a\u65b0\u7b56\u7565\u76f8\u5bf9\u4e8e\u65e7\u7b56\u7565\uff0c\u5728\u52a8\u4f5c\u4e0a\u7684\u6982\u7387\u53d8\u5316\u5e45\u5ea6\u3002\u5982\u679c\u8fd9\u4e2a\u6bd4\u503c\u504f\u79bb 1 \u592a\u591a\uff0c\u5c31\u4f1a\u88ab<strong>\u9650\u5236\u5728\u4e00\u5b9a\u7684\u8303\u56f4\u5185<\/strong>\uff0c\u907f\u514d\u6a21\u578b\u4e00\u6b21\u66f4\u65b0\u7684\u5e45\u5ea6\u8fc7\u5927\u3002<\/p>\n\n\n\n<p>\u9664\u6b64\u4ee5\u5916\uff0cPPO \u7b56\u7565\u91c7\u7528 Reference Model \u4e0a\u4e86\u53cc\u4fdd\u9669\uff0c\u5728\u635f\u5931\u51fd\u6570\u4e2d\u52a0\u5165\u76f8\u5bf9<strong>\u521d\u59cb\u6a21\u578b\u7684 KL \u6563\u5ea6<\/strong>\uff0c\u540c\u6837\u53ef\u4ee5\u907f\u514d\u4e3a\u4e86\u77ed\u671f\u53cd\u9988\u800c\u8131\u79bb\u5408\u7406\u7684\u7b56\u7565\u3002<\/p>\n\n\n\n<p>\u4e8e\u662f\uff0cPPO \u7684\u635f\u5931\u51fd\u6570\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"96\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-420-1024x96.jpg\" alt=\"\" class=\"wp-image-1179\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-420-1024x96.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-420-300x28.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-420-768x72.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-420.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>\u4ece PPO \u5230 GRPO<\/strong><\/p>\n\n\n\n<p>\u4e0a\u9762\u7684\u63cf\u8ff0\u5e94\u8be5\u5f88\u597d\u7406\u89e3 PPO \u5728\u505a\u4ec0\u4e48\u4e8b\u60c5\u3002<\/p>\n\n\n\n<p>\u4f46\u662f PPO \u6709\u4e00\u4e2a\u4e25\u91cd\u7684\u95ee\u9898\uff0c\u7531\u4e8e\u4ef7\u503c\u51fd\u6570\u662f\u968f\u7740\u6a21\u578b\u8bad\u7ec3\u540c\u65f6\u53d8\u52a8\u7684\uff0c\u4e5f\u5c31\u610f\u5473\u7740\u7b56\u7565\u6a21\u578b\u672c\u8eab\u548c Critic \u6a21\u578b\uff08\u4ef7\u503c\u51fd\u6570\uff09\u90fd\u9700\u8981\u8fdb\u884c\u8bad\u7ec3\uff0c\u5e76\u4e14 Critic \u6a21\u578b\u7684\u5927\u5c0f\u548c\u7b56\u7565\u6a21\u578b\u76f8\u540c\u3002\u56e0\u6b64\u4f1a\u5e26\u6765<strong>\u989d\u5916\u7684\u5185\u5b58\u548c\u8ba1\u7b97\u8d1f\u62c5<\/strong>\uff0c\u8bad\u7ec3\u6210\u672c\u975e\u5e38\u9ad8\uff0c\u5f88\u96be\u8fdb\u884c scale up\u3002\u8fd9\u53ef\u80fd\u662f\u8d22\u5927\u6c14\u7c97\u7684 OpenAI \u73a9\u5f97\u8f6c\u5e76\u4e14\u53d6\u5f97\u9886\u5148\u7684\u539f\u56e0\u4e4b\u4e00\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u6539\u5584\u8fd9\u4e2a\u53cc\u6a21\u578b\u7b56\u7565\u7684\u9ad8\u6210\u672c\u95ee\u9898\uff0c\u6700\u7b80\u5355\u7684\u65b9\u6cd5\u5c31\u662f\uff1a<strong>\u53bb\u6389\u4e00\u4e2a\u7f51\u7edc<\/strong>\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u4f60\u4e00\u6837\u8fd9\u4e48\u60f3\uff0c\u90a3\u4f60\u548c DeepSeek \u53c8\u4e0d\u8c0b\u800c\u5408\u3002<\/p>\n\n\n\n<p>DeepSeek \u5728\u5927\u6a21\u578b\u8bad\u7ec3\u6539\u8fdb\u7684\u4e3b\u8981\u52a8\u673a\u662f\u60f3\u529e\u6cd5\u53bb\u6389 Critic \u6a21\u578b\uff0c\u4e3a\u6b64\u63d0\u51fa\u4e86\u4e00\u79cd\u66ff\u4ee3\u65b9\u6cd5\uff0c\u4e5f\u5c31\u662f\u5927\u540d\u9f0e\u9f0e\u7684<strong>\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff08Group Relative Policy Optimization, GRPO\uff09<\/strong>\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"445\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-421-1024x445.jpg\" alt=\"\" class=\"wp-image-1180\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-421-1024x445.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-421-300x130.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-421-768x334.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-421.jpg 1107w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>&nbsp; &nbsp; &nbsp; PPO vs GRPO \u6d41\u7a0b\u7b97\u6cd5\u5bf9\u6bd4<\/sup><\/p>\n\n\n\n<p>\u4ece\u6d41\u7a0b\u7b97\u6cd5\u5bf9\u6bd4\u4e2d\u53ef\u4ee5\u770b\u51fa\u6765\uff0c\u4ef7\u503c\u51fd\u6570\u76f4\u63a5\u6d88\u5931\u4e86\u3002\u90a3\u4e0d\u7528\u4ef7\u503c\u51fd\u6570\uff0c\u6211\u4eec\u5982\u4f55\u786e\u5b9a\u6a21\u578b\u7684 Advantage \u5462\uff1f<\/p>\n\n\n\n<p>GRPO \u91c7\u7528\u4e86\u4e00\u4e2a\u975e\u5e38\u5408\u7406\u7684\u65b9\u6cd5\uff0c\u4e0d\u7528\u300c\u5b66\u4e60\u300d\u4e00\u4e2a\u5355\u72ec\u7684\u4ef7\u503c\u7f51\u7edc\u5f53 Critic\uff0c\u800c\u662f\u7528\u8fd9\u4e2a\u6a21\u578b\u8fc7\u53bb\u591a\u6b21\u7684\u300c\u8003\u8bd5\u6210\u7ee9\u300d\u6765\u786e\u5b9a\u4e00\u4e2a\u57fa\u51c6\u7ebf\u3002<\/p>\n\n\n\n<p>\u5bf9\u540c\u4e00\u9053\u9898\u76ee\u3001\u540c\u4e00\u4e2a\u72b6\u6001\uff0c\u5148\u7528\u65e7\u7b56\u7565<strong>\u91c7\u6837\u591a\u6761\u8f93\u51fa<\/strong>\uff0c\u7136\u540e\u628a\u8fd9\u4e9b\u8f93\u51fa\u7684\u5e73\u5747 Reward \u5f53\u4f5c baseline\uff1b \u8d85\u8fc7\u5e73\u5747\u503c\u5c31\u76f8\u5f53\u4e8e\u300c\u6b63\u5411 Advantage\u300d\uff0c\u4f4e\u4e8e\u5e73\u5747\u503c\u5c31\u662f\u300c\u8d1f\u5411 Advantage\u300d\u3002<\/p>\n\n\n\n<p>\u5728 GRPO \u91cc\uff0c\u9664\u4e86\u8fd9\u4e00\u6b65\uff0c\u8fd8\u4fdd\u7559\u4e86 PPO \u4e2d\u7684 Clip \u548c\u5bf9 Reference Model \u7684 KL \u6b63\u5219\uff0c\u8fd9\u4e9b\u90fd\u53ef\u4ee5\u4fdd\u969c\u66f4\u65b0\u7684\u7a33\u5b9a\u6027\u3002\u4e0d\u8fc7\uff0cKL \u6563\u5ea6\u5728 GRPO \u7684\u76ee\u6807\u51fd\u6570\u76f4\u63a5\u653e\u5728\u4e86\u635f\u5931\u51fd\u6570\uff0c\u8fd9\u964d\u4f4e\u4e86\u5956\u52b1\u51fd\u6570\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u5e76\u4e14\u5b83\u7684\u8ba1\u7b97\u65b9\u6848\u80fd\u591f\u4fdd\u8bc1\u8fdb\u884c\u5f52\u4e00\u5316\u7684 KL \u503c\u6bcf\u6b21\u90fd\u662f\u6b63\u503c\u3002\u800c\u5728 PPO \u4e2d\uff0cKL \u6563\u5ea6\u653e\u5728\u5956\u52b1\u51fd\u6570\u4e2d\u3002<\/p>\n\n\n\n<p>GRPO \u8ddf PPO \u7684\u91cd\u8981\u533a\u522b\uff0c\u4e3b\u8981\u662f<strong>\u53bb\u6389\u4e86\u4ef7\u503c\u51fd\u6570\uff0c\u540c\u65f6\u4f7f\u7528\u7b56\u7565\u6a21\u578b\u7684\u591a\u4e2a\u8f93\u51fa\u91c7\u6837\u7684\u5956\u52b1\u6a21\u578b\u8f93\u51fa\u7684\u591a\u4e2a\u5956\u52b1\u7684\u5e73\u5747\u503c\u4f5c\u4e3a Advantage \u51fd\u6570<\/strong>\u3002<\/p>\n\n\n\n<p>\u4e8e\u662f\uff0c\u6211\u4eec\u5f97\u5230\u4e86 GRPO \u7684\u635f\u5931\u51fd\u6570\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"105\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-422-1024x105.jpg\" alt=\"\" class=\"wp-image-1181\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-422-1024x105.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-422-300x31.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-422-768x78.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-422.jpg 1107w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u5bf9\u4e8e PPO \u5230 GRPO\uff0c\u77e5\u4e4e\u7f51\u53cb\u5c06\u4e24\u8005\u5728\u591a\u4e2a\u7ef4\u5ea6\u4e0a\u8fdb\u884c\u4e86\u6bd4\u8f83\uff0c\u5982\u56fe\u8868\u6240\u793a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"389\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-423-1024x389.jpg\" alt=\"\" class=\"wp-image-1182\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-423-1024x389.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-423-300x114.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-423-768x292.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-423.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><sup>&nbsp; &nbsp; &nbsp; \u77e5\u4e4e\u7f51\u53cb@\u675e\u92c2 \u5206\u4eabPPO\u4e0eGRPO\u7684\u5bf9\u6bd4<\/sup><\/p>\n\n\n\n<p><strong>\u4ece GRPO \u5f00\u679d\u6563\u53f6<\/strong><\/p>\n\n\n\n<p>GRPO \u5728\u51fa\u73b0\u540e\u8fc5\u901f\u6210\u4e3a\u4e00\u4e2a\u540e\u8bad\u7ec3\u8303\u5f0f\u7684\u91cd\u8981\u8282\u70b9\uff0cDeepSeek \u7684\u6a21\u578b\u6210\u529f\u5145\u5206\u8bc1\u660e\u4e86 GRPO \u8303\u5f0f\u7684\u6709\u6548\u6027\u548c\u4f18\u8d8a\u6027\u3002\u4e5f\u56e0\u6b64\uff0c\u540e\u7eed\u7684\u6539\u8fdb\u5de5\u4f5c\u5927\u591a\u90fd\u662f\u5728 GRPO \u7684\u65b9\u6cd5\u57fa\u7840\u4e0a\u8fdb\u884c\u3002<\/p>\n\n\n\n<p>\u90a3\u4e48 GRPO \u5230\u5e95\u6709\u5565\u95ee\u9898\uff0c\u5404\u4e2a\u65b0\u5de5\u4f5c\u90fd\u8981\u5728\u5b83\u8eab\u4e0a\u52a8\u5200\u5462\uff1f<\/p>\n\n\n\n<p>\u6700\u81f4\u547d\u7684\u95ee\u9898\uff0c\u54ea\u6015 GRPO \u5728 PPO \u7684\u57fa\u7840\u4e0a\u8fdb\u884c\u4e86\u6539\u8fdb\uff0c\u4f46\u5728\u7a33\u5b9a\u6027\u4e0a\u4e0e PPO \u65b9\u6cd5\u4ecd\u7136\u534a\u65a4\u516b\u4e24\u3002\u4e5f\u5c31\u662f\u8bf4&nbsp;<strong>GRPO <\/strong><strong>\u4ecd\u7136\u5b58\u5728\u4e25\u91cd\u7684\u7a33\u5b9a\u6027\u95ee\u9898<\/strong>\uff0c\u5f88\u5bb9\u6613\u5bfc\u81f4\u8bad\u7ec3\u5d29\u6e83\u3002<\/p>\n\n\n\n<p>\u6839\u636e\u6570\u5b66\u4e2d\u56fd\u7684\u8bf4\u6cd5\uff0c DeepSeek \u7684\u6570\u636e\u8db3\u591f\u591a\uff0c\u591a\u5230\u53ef\u4ee5\u5b8c\u7f8e\u5730\u907f\u5f00 GRPO \u7684\u7a33\u5b9a\u6027\u7f3a\u9677\u3002\u6bcf\u6b21\u7684\u7b56\u7565\u68af\u5ea6\u8ba1\u7b97\uff0c<strong>\u53ea\u8981 Batch \u6570\u636e\u8db3\u591f\u591a\uff0c\u5c31\u80fd\u6709\u6548\u964d\u4f4e\u7b56\u7565\u68af\u5ea6\u7684\u65b9\u5dee\uff0c\u5c31\u80fd\u83b7\u5f97\u6bd4\u8f83\u7a33\u5b9a\u7684\u8fed\u4ee3\u4e86<\/strong>\u3002\u5bf9\u4e8e\u4e2d\u5c0f\u89c4\u6a21\u7684 RL \u8bad\u7ec3\uff0cGRPO \u5e76\u975e\u4e00\u4e2a\u597d\u7684\u9009\u62e9\uff0c\u5c24\u5176\u662f\u5f53\u6bcf\u6b21\u4f7f\u7528\u7684\u6570\u636e\u6279\u91cf\u6bd4\u8f83\u5c0f\u7684\u65f6\u5019\uff0c\u5b83\u7684\u7a33\u5b9a\u6027\u7f3a\u9677\u5c06\u662f\u81f4\u547d\u7684\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u6700\u65b0\u7684\u4e00\u4e9b\u65b9\u6cd5\u9488\u5bf9 GPRO \u7684\u4e0d\u540c\u90e8\u5206\u8fdb\u884c\u4e86\u8fed\u4ee3\uff0c\u5177\u4f53\u7f3a\u9677\u548c\u4f18\u5316\u65b9\u5f0f\u5728\u4ecb\u7ecd\u65b0\u5de5\u4f5c\u65f6\u7ec6\u8bb2\u3002<\/p>\n\n\n\n<p><strong>DAPO<\/strong><\/p>\n\n\n\n<p>\u9996\u5148\u8981\u8bb2\u7684\u4f18\u5316\u8303\u5f0f\u662f\uff0c\u8fd9\u662f\u5b57\u8282\u3001\u6e05\u534e AIR \u5728\u4eca\u5e74\u4e09\u6708\u5f00\u6e90\u7684\u7b97\u6cd5\u3002<\/p>\n\n\n\n<p>\u4f7f\u7528\u8be5\u7b97\u6cd5\uff0c\u8be5\u56e2\u961f\u6210\u529f\u8ba9 Qwen2.5-32B \u6a21\u578b\u5728 AIME 2024 \u57fa\u51c6\u4e0a\u83b7\u5f97\u4e86 50 \u5206\uff0c\u4f18\u4e8e\u540c\u7b49\u89c4\u6a21\u7684 DeepSeek-R1-Zero-Qwen-32B\uff0c\u540c\u65f6 DAPO \u7248 Qwen2.5-32B \u4f7f\u7528\u7684\u8bad\u7ec3\u6b65\u6570\u8fd8\u5c11 50%\u3002<\/p>\n\n\n\n<p>\u4f46\u662f\u503c\u5f97\u4e00\u63d0\u7684\u662f\uff0cDAPO \u65b9\u6cd5\u5e76\u6ca1\u6709<strong>\u5728\u6570\u5b66\u539f\u7406\u4e0a\u6709\u4ec0\u4e48\u672c\u8d28\u4e0a\u7684\u6539\u53d8<\/strong>\uff0c\u57fa\u672c\u4f18\u5316\u76ee\u6807\u4ecd\u7136\u6cbf\u7528\u4e86 GRPO \u7684\u5f62\u5f0f\uff0c\u53ea\u662f\u5bf9 Clip \u7b49\u53c2\u6570\u548c\u91c7\u6837\u673a\u5236\u505a\u51fa\u4e86\u6539\u8fdb\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u628a DAPO \u653e\u5728\u6700\u65e9\u8ba8\u8bba\u7684\u987a\u4f4d\u3002<\/p>\n\n\n\n<p>\u5728\u5b9e\u8df5\u8fc7\u7a0b\u4e2d\uff0cGRPO \u5b58\u5728\u4ee5\u4e0b\u51e0\u4e2a\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Token \u7ea7\u522b\u7684 Clip \u5bb9\u6613\u5bfc\u81f4<strong>\u71b5\u5d29\u6e83<\/strong>\uff1a\u6a21\u578b\u5f88\u5feb\u6536\u655b\u5230\u5c11\u91cf\u56fa\u5b9a\u7b54\u6848\uff0c\u5bfc\u81f4\u591a\u6837\u6027\u548c\u63a2\u7d22\u80fd\u529b\u4e0d\u8db3\uff08\u71b5\u5d29\u6e83\uff09\u3002<\/li>\n\n\n\n<li>Batch \u91c7\u6837\u4e2d\u51fa\u73b0<strong>\u5956\u52b1\u6781\u7aef\u5316<\/strong>\uff1a\u90e8\u5206\u6837\u672c\u7684\u5956\u52b1\u53ef\u80fd\u5168\u90e8\u4e3a 1 \u6216 0\uff0c\u4ece\u800c\u4ea7\u751f\u300c\u96f6\u68af\u5ea6\u300d\u95ee\u9898\uff0c\u524a\u5f31\u8bad\u7ec3\u4fe1\u53f7\u3002<\/li>\n\n\n\n<li>\u957f\u5e8f\u5217\u8bad\u7ec3\u7684<strong>\u68af\u5ea6\u5206\u5e03\u5931\u8861<\/strong>\uff1a\u6743\u91cd\u5206\u5e03\u8ba9\u6781\u5c11\u6570 token \u7684\u68af\u5ea6\u5360\u636e\u4e3b\u5bfc\uff0c\u5bfc\u81f4\u8bb8\u591a\u9ad8\u8d28\u91cf\u7684\u957f\u5e8f\u5217\u6837\u672c\u88ab\u5ffd\u89c6\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4e3a\u6b64\uff0cDAPO \u6839\u636e\u5b9e\u8df5\u4e2d\u51fa\u73b0\u7684\u95ee\u9898\u63d0\u51fa\u4e86\u9488\u5bf9\u6027\u7684\u4f18\u5316\uff1a<\/p>\n\n\n\n<p>1.&nbsp;<strong>Clip-Higher <\/strong><strong>\u673a\u5236<\/strong>\uff1a\u5c06 Clip \u7684\u4e0a\u4e0b\u9650\u5206\u5f00 &nbsp;\uff0c\u7814\u7a76\u8005\u5c06\u8f83\u4f4e\u548c\u8f83\u9ad8\u7684\u526a\u8f91\u8303\u56f4\u89e3\u8026\u4e3a \u03b5_low \u548c \u03b5_high\uff0c\u7814\u7a76\u8005\u589e\u52a0\u4e86 \u03b5_high \u7684\u503c\uff0c\u4ee5\u4fbf\u4e3a\u4f4e\u6982\u7387 token \u7684\u589e\u52a0\u7559\u51fa\u66f4\u591a\u7a7a\u95f4\uff0c\u80fd\u591f\u663e\u8457\u63d0\u5347\u6a21\u578b\u8bad\u7ec3\u65e9\u671f\u7684\u71b5\u3002<\/p>\n\n\n\n<p>2.&nbsp;<strong>\u52a8\u6001\u91c7\u6837<\/strong>\uff1a\u8fdb\u884c\u8fc7\u5ea6\u91c7\u6837\uff0c\u8fc7\u6ee4\u6389\u5956\u52b1\u7b49\u4e8e 1 \u548c 0 \u7684\u63d0\u793a\u8bed\uff0c\u53ea\u4fdd\u7559\u6709\u6548\u68af\u5ea6\u7684\u6837\u672c\uff0c\u63d0\u9ad8\u8bad\u7ec3\u6548\u7387\u3002<\/p>\n\n\n\n<p>3.&nbsp;<strong>Token <\/strong><strong>\u7ea7\u7b56\u7565\u68af\u5ea6\u635f\u5931<\/strong>\uff1a\u5bf9\u6240\u6709 token \u4e00\u8d77\u6c42\u5e73\u5747\uff0c\u4fdd\u8bc1\u957f\u5e8f\u5217\u7684\u6240\u6709 token \u90fd\u516c\u5e73\u5730\u4e3a batch loss \u505a\u8d21\u732e\uff0c\u5e76\u9632\u6b62\u957f\u5e8f\u5217\u7684\u4f18\u5316\u68af\u5ea6\u88ab\u8fc7\u5ea6\u7f29\u5c0f\u3002<\/p>\n\n\n\n<p>4.&nbsp;<strong>\u8d85\u957f\u5956\u52b1\u8c03\u6574<\/strong>\uff1a\u9488\u5bf9\u8d85\u957f\u6837\u672c\uff0c\u5f53\u54cd\u5e94\u957f\u5ea6\u8d85\u8fc7\u9884\u5b9a\u4e49\u7684\u6700\u5927\u503c\u65f6\uff0c\u7814\u7a76\u8005\u5b9a\u4e49\u4e00\u4e2a\u300csoft \u7f5a\u5206\u300d\u3002\u5728\u8fd9\u4e2a\u533a\u95f4\u5185\uff0c\u54cd\u5e94\u8d8a\u957f\uff0c\u53d7\u5230\u7684\u60e9\u7f5a\u5c31\u8d8a\u5927\uff0c\u4ee5\u6b64\u907f\u514d\u8fc7\u957f\u7684\u54cd\u5e94\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0cDAPO \u7684\u4f18\u5316\u635f\u5931\u51fd\u6570\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"201\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-424-1024x201.jpg\" alt=\"\" class=\"wp-image-1183\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-424-1024x201.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-424-300x59.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-424-768x151.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-424.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u867d\u7136 DAPO \u4f9d\u7136\u662f&nbsp;<strong>token <\/strong><strong>\u7ea7\u522b<\/strong>\u7684\u91cd\u8981\u6027\u91c7\u6837\uff0c\u4f46\u8bad\u7ec3\u66f2\u7ebf\u548c\u6700\u7ec8\u6027\u80fd\u63d0\u5347\u975e\u5e38\u660e\u663e\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"390\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-426-1024x390.jpg\" alt=\"\" class=\"wp-image-1185\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-426-1024x390.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-426-300x114.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-426-768x293.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-426.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u9879\u76ee\u9875\u9762\uff1ahttps:\/\/dapo-sia.github.io\/<\/li>\n\n\n\n<li>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/dapo-sia.github.io\/static\/pdf\/dapo_paper.pdf<\/li>\n<\/ul>\n\n\n\n<p><strong>GSPO<\/strong><\/p>\n\n\n\n<p>\u5927\u7684\u6765\u4e86\u3002\u540e\u8bad\u7ec3\u9886\u57df\u91cc\u91cd\u8981\u7684\u7a81\u7834\u662f Qwen3 \u4f7f\u7528\u7684\u3002<\/p>\n\n\n\n<p>\u4e0a\u6587\u90a3\u4e48\u591a\u6587\u5b57\u4e00\u76f4\u5728\u63d0\u53ca PPO \u7c7b\u4f3c\u65b9\u6cd5\u7684\u91cd\u8981\u7ea7\u91c7\u6837\u5747\u4e3a token \u7ea7\uff0c\u8fed\u4ee3\u65b9\u6cd5\u4e00\u76f4\u6ca1\u6709\u7a81\u7834 token \u91c7\u6837\u7684\u9650\u5236\uff0c\u800c&nbsp;<strong>GSPO <\/strong><strong>\u771f\u6b63\u5728\u539f\u7406\u4e0a\u505a\u51fa\u4e86\u6539\u8fdb<\/strong>\u3002<\/p>\n\n\n\n<p>\u6700\u8fd1 Qwen \u7684\u7814\u7a76\u8868\u660e\uff0c\u4f7f\u7528 GRPO \u8bad\u7ec3\u5927\u8bed\u8a00\u6a21\u578b\u65f6\u5b58\u5728\u4e25\u91cd\u7684\u7a33\u5b9a\u6027\u95ee\u9898\uff0c\u5f80\u5f80\u4f1a\u5bfc\u81f4\u6a21\u578b\u4e0d\u53ef\u9006\u5730\u5d29\u6e83\u3002\u5728 Qwen \u56e2\u961f\u7684\u7814\u7a76\u4e2d\uff0c\u63ed\u793a\u4e86 GPRO \u65b9\u6cd5\u7684\u4e25\u91cd\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728\u6bcf\u4e2a token \u7ea7\u522b\u5e94\u7528\u91cd\u8981\u6027\u91c7\u6837\uff0c\u4f1a\u5728<strong>\u957f\u5e8f\u5217\u4e2d\u79ef\u7d2f\u9ad8\u65b9\u5dee\uff0c\u5bfc\u81f4\u8bad\u7ec3\u4e0d\u7a33\u5b9a<\/strong>\u3002<\/li>\n\n\n\n<li>\u8fd9\u4e00\u95ee\u9898\u5728\u00a0<strong>\u4e13\u5bb6\u6df7\u5408\u6a21\u578b\uff08Mixture-of-Experts, MoE\uff09 \u4e2d\u5c24\u4e3a\u4e25\u91cd<\/strong>\uff0c\u56e0\u4e3a token \u7ea7\u522b\u7684\u8def\u7531\u53d8\u5316\u4f1a\u52a0\u5267\u4e0d\u7a33\u5b9a\u6027\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5982\u679c\u8bf4 DAPO \u662f\u5728 GRPO \u6846\u67b6\u5185\u505a\u5fae\u8c03\uff0c\u90a3\u4e48 GSPO \u5219\u662f\u76f4\u63a5\u8c03\u6574\u4e86\u4f18\u5316\u76ee\u6807\u7684\u9897\u7c92\u5ea6 \u2014\u2014&nbsp;<strong>\u4ece token \u7ea7\u8df3\u5230\u5e8f\u5217\u7ea7<\/strong>\u3002<\/p>\n\n\n\n<p>\u91cd\u8981\u6027\u91c7\u6837\u7684\u4f5c\u7528\u662f\uff1a\u6765\u7f13\u89e3 off-policy \u5e26\u6765\u7684\u5206\u5e03\u5dee\u5f02\u60c5\u51b5\uff0c\u4e5f\u5c31\u662f\u8bf4\uff1a<\/p>\n\n\n\n<p>\u6211\u4eec\u60f3\u8981\u4f30\u8ba1\u4e00\u4e2a\u9884\u671f\u7684\u5206\u5e03\uff0c\u4f46\u662f\u6211\u4eec\u624b\u4e0a\u53ea\u6709\u53e6\u884c\u4e3a\u6a21\u578b\u7684\u5206\u5e03\uff0c\u6211\u4eec\u5c31\u53ea\u80fd\u5728\u884c\u4e3a\u7b56\u7565\u4e0b\u8fdb\u884c\u91c7\u6837\uff0c\u901a\u8fc7\u8fd9\u4e2a\u6837\u672c\uff0c\u8d4b\u4e88\u91cd\u8981\u6027\u6743\u91cd\uff0c\u6765\u4f30\u8ba1\u51fa\u76ee\u6807\u7b56\u7565\u4e0b\u51fd\u6570\u7684\u503c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"334\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-425-1024x334.jpg\" alt=\"\" class=\"wp-image-1184\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-425-1024x334.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-425-300x98.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-425-768x250.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-425.jpg 1107w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u4f46\u662f\u8fd9\u79cd\u91c7\u6837\u7684\u524d\u63d0\u5728\u4e8e\u591a\u6b21\u91c7\u6837\uff0c\u5982\u679c\u53ea\u6709\u4e00\u6b21\u91c7\u6837\uff0c\u5e76\u4e0d\u80fd\u8d77\u5230\u5206\u5e03\u77eb\u6b63\u7684\u4f5c\u7528\u3002\u95ee\u9898\u5728\u4e8e\u5927\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u91cd\u8981\u6027\u91c7\u6837\u90fd\u662f \u5728 token \u7ea7\u522b\u8fdb\u884c\u7684\uff0c\u5355\u4e2a token \u8fdb\u884c\u7684\u91cd\u8981\u6027\u91c7\u6837\u662f\u65e0\u6cd5\u8d77\u5230\u5206\u5e03\u77eb\u6b63\u7684\u4f5c\u7528\u7684\uff0c\u76f8\u53cd\uff0c\u8fd9\u79cd\u91c7\u6837\u624b\u6bb5\u53cd\u800c\u4f1a\u5e26\u6765\u5f88\u5927\u65b9\u5dee\u7684\u566a\u58f0\u3002<\/p>\n\n\n\n<p>\u5728\u8bad\u7ec3\u65f6\uff0c\u5956\u52b1\u5176\u5b9e\u662f\u9488\u5bf9\u6574\u6bb5\u56de\u7b54\u6253\u7684\u5206\uff0c\u6bd4\u5982\u4e00\u53e5\u8bdd\u3001\u4e00\u4e2a\u5b8c\u6574\u56de\u590d\u90fd\u4f1a\u5f97\u5230\u4e00\u4e2a\u6574\u4f53\u8bc4\u4ef7\u3002<\/p>\n\n\n\n<p>\u4f46\u662f\u5728\u6a21\u578b\u4f18\u5316\u65f6\uff0c\u6211\u4eec\u901a\u5e38\u662f\u5728 token \u5c42\u9762\u8fdb\u884c\u91c7\u6837\u548c\u66f4\u65b0\u3002\u4e8e\u662f\u5e38\u89c1\u7684\u505a\u6cd5\u662f\uff1a\u628a\u5956\u52b1\u76f4\u63a5\u5206\u644a\u5230\u6bcf\u4e00\u4e2a token \u4e0a\uff0c\u518d\u9010\u4e2a\u53bb\u8c03\u6574\u3002<\/p>\n\n\n\n<p>\u8fd9\u5c31\u5bfc\u81f4\u4e86&nbsp;<strong>\u4f18\u5316\u76ee\u6807\u548c\u5956\u52b1\u76ee\u6807\u7684\u9897\u7c92\u5ea6\u4e0d\u5339\u914d<\/strong>\uff1a\u6a21\u578b\u53ef\u80fd\u5728\u5355\u4e2a token \u4e0a\u5b66\u5f97\u5f88\u7528\u529b\uff0c\u4f46\u8fd9\u5e76\u4e0d\u80fd\u5b8c\u5168\u5bf9\u5e94\u6574\u6bb5\u56de\u7b54\u7684\u8d28\u91cf\u3002<\/p>\n\n\n\n<p>\u4e3a\u6b64\uff0cQwen \u56e2\u961f\u5c06 GRPO \u8fdb\u5316\u4e3a<strong>\u7ec4\u5e8f\u5217\u7b56\u7565\u4f18\u5316\uff08Group Sequence Policy Optimization, GSPO\uff09<\/strong>\u3002<\/p>\n\n\n\n<p>\u6b63\u5982\u5176\u540d\u79f0\u6240\u6697\u793a\u7684\uff0cGSPO \u7684\u6838\u5fc3\u5728\u4e8e\u5c06\u91cd\u8981\u6027\u91c7\u6837<strong>\u4ece token \u7ea7\u8f6c\u79fb\u81f3\u5e8f\u5217\u7ea7<\/strong>\uff0c\u5176\u91cd\u8981\u6027\u6bd4\u503c\u57fa\u4e8e\u6574\u4e2a\u5e8f\u5217\u7684\u4f3c\u7136\u5ea6\u8ba1\u7b97\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"143\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-427-1024x143.jpg\" alt=\"\" class=\"wp-image-1186\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-427-1024x143.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-427-300x42.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-427-768x107.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-427.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u8fd9\u79cd\u91c7\u6837\u6743\u91cd\u7684\u8bbe\u8ba1\u81ea\u7136\u5730\u7f13\u89e3\u4e86\u9010 token \u65b9\u5dee\u7684\u7d2f\u79ef\u95ee\u9898\uff0c\u4ece\u800c\u663e\u8457\u63d0\u5347\u4e86\u8bad\u7ec3\u8fc7\u7a0b\u7684\u7a33\u5b9a\u6027\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0cGSPO \u7684\u635f\u5931\u51fd\u6570\u4e3a\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"92\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-428-1024x92.jpg\" alt=\"\" class=\"wp-image-1187\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-428-1024x92.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-428-300x27.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-428-768x69.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-428.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>GRPO\uff1a\u91cd\u8981\u6027\u6743\u91cd\u5728 token \u7ea7\uff0c\u6bcf\u4e2a token \u90fd\u53ef\u80fd\u88ab\u5355\u72ec\u88c1\u526a\u3002<\/li>\n\n\n\n<li>GSPO\uff1a\u91cd\u8981\u6027\u6743\u91cd\u5728 \u5e8f\u5217\u7ea7\uff0c\u88c1\u526a\u65f6\u76f4\u63a5\u4f5c\u7528\u4e8e\u6574\u4e2a\u56de\u7b54\uff0c\u66f4\u7b26\u5408\u5956\u52b1\u4fe1\u53f7\u7684\u6574\u4f53\u6027\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6b64\u5916\uff0cGSPO \u5bf9 \u5e8f\u5217\u7ea7\u7684\u91cd\u8981\u6027\u8fd8\u505a\u4e86 \u957f\u5ea6\u5f52\u4e00\u5316\uff0c\u4e0d\u540c\u95ee\u9898\u7684\u56de\u7b54\u957f\u5ea6\u5dee\u522b\u5f88\u5927\uff0c\u5982\u679c\u4e0d\u5f52\u4e00\u5316\uff0cimportance ratio \u4f1a\u5bf9\u957f\u5ea6\u975e\u5e38\u654f\u611f\uff0c\u9020\u6210\u4e0d\u7a33\u5b9a\u3002<\/p>\n\n\n\n<p>\u6700\u540e\uff0c\u56e0\u4e3a\u540c\u4e00\u4e2a\u5e8f\u5217\u4e2d\u7684\u6240\u6709 token \u5171\u7528\u540c\u4e00\u4e2a\u91cd\u8981\u6027\u6743\u91cd\uff0c\u4e00\u65e6\u53d1\u751f clipping\uff0c\u88ab\u88c1\u526a\u6389\u7684\u5c31\u662f \u6574\u4e2a\u5e8f\u5217\uff0c\u800c\u4e0d\u662f\u50cf GRPO \u90a3\u6837\u53ea\u5f71\u54cd\u90e8\u5206 token\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0cGSPO \u63d0\u51fa\u7684\u300c\u5e8f\u5217\u7ea7\u91cd\u8981\u6027\u91c7\u6837\u300d\u663e\u8457\u63d0\u9ad8\u4e86\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\uff0c\u5f88\u53ef\u80fd\u4f1a\u6210\u4e3a\u672a\u6765\u540e\u8bad\u7ec3\u5f3a\u5316\u5b66\u4e60\u7684\u65b0\u6807\u51c6\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"325\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-429-1024x325.jpg\" alt=\"\" class=\"wp-image-1188\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-429-1024x325.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-429-300x95.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-429-768x244.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-429.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bba\u6587\u6807\u9898\uff1aGroup Sequence Policy Optimization<\/li>\n\n\n\n<li>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/huggingface.co\/papers\/2507.18071<\/li>\n\n\n\n<li>\u535a\u5ba2\u94fe\u63a5\uff1ahttps:\/\/qwenlm.github.io\/blog\/gspo\/<\/li>\n<\/ul>\n\n\n\n<p><strong>GFPO<\/strong><\/p>\n\n\n\n<p>\u5728 GSPO \u4e4b\u540e\u4e0d\u4e45\uff0c\u5fae\u8f6f\u7814\u7a76\u5458\u66dd\u51fa\u4e00\u4e2a\u65b0\u6210\u679c\uff1a<strong>\u7ec4\u8fc7\u6ee4\u7b56\u7565\u4f18\u5316\uff08Group Filtered Policy Optimization\uff0cGFPO\uff09<\/strong>\uff0c\u53e6\u4e00\u79cd\u98a0\u8986\u6027\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u3002<\/p>\n\n\n\n<p>\u5728 GFPO \u5de5\u4f5c\u4e2d\uff0c\u5fae\u8f6f\u7814\u7a76\u56e2\u961f\u6307\u51fa\u4e86 GRPO \u7684\u4e00\u4e2a\u5173\u952e\u9650\u5236\uff1a<\/p>\n\n\n\n<p>GRPO \u4f9d\u8d56\u4e8e\u5355\u4e00\u7684\u6807\u91cf\u5956\u52b1\u4fe1\u53f7\uff0c\u8fd9\u4f7f\u5f97\u5b83<strong>\u96be\u4ee5\u8054\u5408\u4f18\u5316\u591a\u4e2a\u5c5e\u6027<\/strong>\uff0c\u4f8b\u5982\u540c\u65f6\u4f18\u5316\u7b80\u6d01\u6027\u548c\u51c6\u786e\u5ea6\u3002<\/p>\n\n\n\n<p>\u7ed3\u679c\u5c31\u662f\uff0cGRPO \u786e\u5b9e\u80fd\u63d0\u9ad8\u51c6\u786e\u5ea6\uff0c\u4f46\u4e5f\u4f1a\u8ba9\u54cd\u5e94\u957f\u5ea6\u5927\u5e45\u589e\u52a0\u3002\u8fd9\u4e5f\u5bfc\u81f4\u4e86\u5927\u6a21\u578b\u9047\u5230\u4e00\u4e2a\u7a0d\u5fae\u68d8\u624b\u7684\u95ee\u9898\uff0c\u5c31\u4f1a\u50cf\u9677\u5165\u6c89\u601d\u4e00\u6837\u957f\u7bc7\u5927\u8bba\u5730\u63a8\u4e0b\u53bb\uff0c\u8017\u65f6\u8017\u7b97\u529b\uff0c\u7ed3\u679c\u5374\u672a\u5fc5\u9760\u8c31\u3002<\/p>\n\n\n\n<p>GFPO \u6b63\u662f\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u800c\u751f\u7684\uff0c<strong>\u5b83\u53ef\u4ee5\u540c\u65f6\u4f18\u5316\u591a\u4e2a\u54cd\u5e94\u5c5e\u6027<\/strong>\u3002<\/p>\n\n\n\n<p>GFPO \u662f\u4e00\u79cd\u7b80\u5355\u800c\u6709\u6548\u7684\u65b9\u6cd5\uff0c\u53ef\u4ee5\u9488\u5bf9\u60f3\u8981\u7684\u54cd\u5e94\u5c5e\u6027\u8fdb\u884c\u6709\u9488\u5bf9\u6027\u7684\u7b56\u7565\u4f18\u5316\u3002<\/p>\n\n\n\n<p>GFPO \u4f1a\u4e3a\u6bcf\u4e2a\u95ee\u9898\u91c7\u6837\u66f4\u5927\u7684\u5019\u9009\u54cd\u5e94\u7ec4\uff0c\u4ece\u800c<strong>\u6269\u5927\u54cd\u5e94\u6c60<\/strong>\u4ee5\u5305\u542b\u66f4\u591a\u5177\u6709\u6240\u9700\u7279\u6027\u7684\u5019\u9009\u54cd\u5e94\uff0c\u7136\u540e\u5728\u8ba1\u7b97\u7b56\u7565\u68af\u5ea6\u65f6\u663e\u5f0f\u5730\u8fc7\u6ee4\u8fd9\u4e9b\u7279\u6027\uff0c\u4e0d\u7b26\u5408\u76ee\u6807\u5c5e\u6027\u7684\u54cd\u5e94\u4e0d\u8fdb\u5165\u4f18\u5316\u3002<\/p>\n\n\n\n<p>\u6570\u636e\u8fc7\u6ee4\u662f\u4e00\u79cd\u9690\u5f0f\u3001\u7075\u6d3b\u7684\u5956\u52b1\u5851\u9020\u5f62\u5f0f \u2014\u2014 \u7c7b\u4f3c\u4e8e<strong>\u4f7f\u7528\u9009\u62e9\u6027\u91c7\u6837\u6765\u653e\u5927\u7279\u5b9a\u6a21\u578b\u884c\u4e3a\u7684\u8fed\u4ee3\u5f0f\u81ea\u6211\u6539\u8fdb\u65b9\u6cd5\u3002<\/strong>\u8fc7\u6ee4\u673a\u5236\u4f1a\u8fed\u4ee3\u5730\u653e\u5927\u6a21\u578b\u5728\u76ee\u6807\u5c5e\u6027\u4e0a\u7684\u8868\u73b0\uff0c\u5c31\u50cf\u5f3a\u5316\u5b66\u4e60\u91cc\u7684\u300c\u504f\u597d\u653e\u5927\u5668\u300d\u3002<\/p>\n\n\n\n<p>\u5728\u6b64\u663e\u5f0f\u8fc7\u6ee4\u6b65\u9aa4\u5206\u79bb\u51fa\u6240\u9700\u7684\u54cd\u5e94\u540e\uff0c\u5c06\u5728\u6240\u9009\u7ec4\u5185\u4f7f\u7528\u6807\u51c6\u5956\u52b1\u6765\u8ba1\u7b97\u76f8\u5bf9\u4f18\u52bf\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0cGFPO \u65e0\u9700\u590d\u6742\u7684\u5956\u52b1\u5de5\u7a0b\uff0c\u5373\u53ef\u540c\u65f6\u4f18\u5316\u591a\u4e2a\u6240\u9700\u5c5e\u6027\uff08\u4f8b\u5982\u957f\u5ea6\u548c\u51c6\u786e\u5ea6\uff09\u3002<\/p>\n\n\n\n<p>GFPO \u7684\u5f62\u5f0f\u5316\u5b9a\u4e49\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"415\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-430-1024x415.jpg\" alt=\"\" class=\"wp-image-1189\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-430-1024x415.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-430-300x122.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-430-768x311.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-430.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>GFPO \u7684\u4e3b\u8981\u5e72\u9884\u63aa\u65bd\u662f\u5728 Advantage \u4f30\u8ba1\u5c42\u9762\uff0c\u4f7f\u5176\u53ef\u4e0e\u4efb\u4f55 GRPO \u7c7b\u4f3c\u7684\u65b9\u6cd5\u517c\u5bb9\uff0c\u4f8b\u5982 DAPO\u3001Dr. GRPO \u6216\u5e26\u6709 Dual-Clip PPO \u635f\u5931\u7684 GRPO\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"315\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-431-1024x315.jpg\" alt=\"\" class=\"wp-image-1190\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-431-1024x315.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-431-300x92.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-431-768x236.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-431.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u8bba\u6587\u6807\u9898\uff1aSample More to Think Less: Group Filtered Policy Optimization for Concise Reasoning<\/li>\n\n\n\n<li>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2508.09726<\/li>\n<\/ul>\n\n\n\n<p><strong>GRPO \u7684\u4e00\u4e9b\u5176\u4ed6\u95ee\u9898<\/strong><\/p>\n\n\n\n<p>\u9664\u6b64\u4ee5\u5916\uff0c\u4e5f\u6709\u4e9b\u7814\u7a76\u8005\u53d1\u73b0\u4e86 GRPO \u7684\u4e00\u4e9b\u5176\u4ed6\u7f3a\u9677\uff0c\u6216\u8bb8\u53ef\u4ee5\u4e3a\u672a\u6765\u7684\u7814\u7a76\u5de5\u4f5c\u63d0\u4f9b\u4e00\u4e9b\u65b0\u601d\u8def\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"331\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-432-1024x331.jpg\" alt=\"\" class=\"wp-image-1191\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-432-1024x331.jpg 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-432-300x97.jpg 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-432-768x248.jpg 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-432.jpg 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u535a\u5ba2\u94fe\u63a5\uff1ahttps:\/\/aryagxr.com\/blogs\/grpo-limitations.html<\/li>\n<\/ul>\n\n\n\n<p><strong>\u7f3a\u9677 1\uff1a\u5956\u52b1\u7684\u6b67\u4e49\u6027<\/strong><\/p>\n\n\n\n<p>\u590d\u6742\u7684\u63a8\u7406\u95ee\u9898\u901a\u5e38\u9700\u8981\u591a\u4e2a\u5956\u52b1\u4fe1\u53f7\uff0c\u56e0\u6b64\u6211\u4eec\u4f1a\u4e3a\u6bcf\u4e2a\u8bc4\u5224\u6807\u51c6\u5355\u72ec\u8bbe\u8ba1\u5956\u52b1\u51fd\u6570\u3002\u7136\u540e\uff0c\u628a\u6240\u6709\u5956\u52b1\u51fd\u6570\u7684\u5206\u6570\u52a0\u5728\u4e00\u8d77\uff0c\u6298\u53e0\u6210\u4e00\u4e2a\u5355\u4e00\u7684\u5956\u52b1\u4fe1\u53f7\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"816\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-170-1024x816.png\" alt=\"\" class=\"wp-image-1192\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-170-1024x816.png 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-170-300x239.png 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-170-768x612.png 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-170.png 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u95ee\u9898\u5728\u4e8e\uff0c\u6a21\u578b\u6839\u672c\u65e0\u6cd5\u77e5\u9053 \u81ea\u5df1\u5230\u5e95\u662f\u56e0\u4e3a\u4ec0\u4e48\u884c\u4e3a\u88ab\u5956\u52b1\u7684\u3002\u867d\u7136\u6211\u4eec\u5199\u4e86\u4e0d\u540c\u7684\u5956\u52b1\u51fd\u6570\uff0c\u4f46\u6700\u540e\u6240\u6709\u5956\u52b1\u4f9d\u7136\u88ab\u5408\u5e76\u4e3a\u4e00\u4e2a\u6807\u91cf\u4fe1\u53f7\u3002\u6a21\u578b\u53c8\u600e\u4e48\u77e5\u9053\u5956\u52b1\u662f\u6765\u81ea\u300c\u7b54\u6848\u6b63\u786e\u300d\uff0c\u8fd8\u662f\u300c\u63a8\u7406\u6e05\u6670\u300d\uff0c\u8fd8\u662f\u300c\u683c\u5f0f\u89c4\u8303\u300d\u5462\uff1f<\/p>\n\n\n\n<p>\u5373\u4f7f\u6211\u4eec\u8c03\u6574\u4e0d\u540c\u5956\u52b1\u7ec4\u4ef6\u7684\u6743\u91cd\uff0c\u6a21\u578b\u770b\u5230\u7684\u4ecd\u7136\u53ea\u662f\u4e00\u4e2a\u603b\u7684\u6807\u91cf\u5956\u52b1\u3002<\/p>\n\n\n\n<p>GFPO \u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u6539\u5584\u4e86\u4e0a\u8ff0\u95ee\u9898\u3002<\/p>\n\n\n\n<p><strong>\u7f3a\u9677 2\uff1a\u6807\u91cf\u53cd\u9988<\/strong><\/p>\n\n\n\n<p>\u5728\u63a8\u7406\u4efb\u52a1\u4e2d\uff0cGRPO \u4f1a\u4e22\u5f03\u6240\u6709\u4e2d\u95f4\u7684\u6587\u672c\u53cd\u9988\uff0c\u56e0\u4e3a\u4f20\u7ed9\u6a21\u578b\u7684\u53ea\u662f\u4e00\u4e2a\u6570\u503c\u5316\u7684\u5956\u52b1\u4fe1\u53f7\u3002<\/p>\n\n\n\n<p>\u4e3e\u4e2a\u4f8b\u5b50\uff0c\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4f1a\u6253\u5370\u4e00\u4e9b\u8f93\u51fa\uff0c\u6bcf\u6b21\u731c\u6d4b\u90fd\u6709\u6587\u5b57\u53cd\u9988\uff0c\u6bd4\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u300c\u5b57\u6bcd \u2018c\u2019 \u4e0d\u5e94\u8be5\u5728\u7b54\u6848\u91cc\u300d<\/li>\n\n\n\n<li>\u300c\u2018n\u2019 \u4e0d\u80fd\u51fa\u73b0\u5728\u4f4d\u7f6e 3\u300d<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e9b\u6587\u5b57\u53cd\u9988\u5bf9\u6a21\u578b\u5176\u5b9e\u5f88\u6709\u5e2e\u52a9\uff0c\u4f46\u5728 GRPO \u6846\u67b6\u4e0b\u5b8c\u5168\u7528\u4e0d\u4e0a\uff0c\u56e0\u4e3a\u5b83\u4eec\u6700\u7ec8\u90fd\u4f1a\u88ab\u62bd\u8c61\u6210\u4e00\u4e2a\u6807\u91cf\u5956\u52b1\u3002<\/p>\n\n\n\n<p><strong>\u7f3a\u9677 3\uff1a\u591a\u8f6e\u63a8\u7406<\/strong><\/p>\n\n\n\n<p>\u53e6\u4e00\u4e2a\u74f6\u9888\u662f \u591a\u8f6e\u63a8\u7406 \u4efb\u52a1\u5728 GRPO \u4e0b\u7684\u8868\u73b0\u3002\u95ee\u9898\u7684\u5173\u952e\u5728\u4e8e\uff1a<\/p>\n\n\n\n<p>\u5728\u591a\u8f6e\u5bf9\u8bdd\u4e2d\uff0c\u6bcf\u4e00\u8f6e\u7684\u53cd\u9988\u90fd\u4f1a\u88ab\u91cd\u65b0\u8f93\u5165\u5230\u57fa\u7840\u6a21\u578b\u7684 prompt \u4e2d\uff0c\u4ece\u800c\u5bfc\u81f4 \u6307\u6570\u7ea7\u5206\u652f\uff08exponential forking\uff09\uff0c\u4f7f\u5f97 GRPO \u5728\u591a\u8f6e\u4efb\u52a1\u4e2d\u7684\u8bad\u7ec3\u53d8\u5f97\u975e\u5e38\u75db\u82e6\u3002\u89c1\u4e0b\u56fe\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"290\" src=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-171-1024x290.png\" alt=\"\" class=\"wp-image-1193\" srcset=\"http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-171-1024x290.png 1024w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-171-300x85.png 300w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-171-768x218.png 768w, http:\/\/www.applicationsofllm.com\/wp-content\/uploads\/2025\/09\/image-171.png 1108w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><strong>\u5199\u5728\u6700\u540e<\/strong><\/p>\n\n\n\n<p>\u7b80\u5355\u603b\u7ed3\u4e00\u4e0b\uff0c\u540e\u8bad\u7ec3\u7684\u53d1\u5c55\u8109\u7edc\u5176\u5b9e\u5f88\u6e05\u6670\u3002\u4ece OpenAI \u63d0\u51fa PPO \u7684\u540e\u8bad\u7ec3\u65b9\u6cd5\u5f00\u59cb\uff0c\u90fd\u5728\u6b64\u57fa\u7840\u4e0a\u7f1d\u7f1d\u8865\u8865\u3002<\/p>\n\n\n\n<p>GRPO \u662f PPO \u57fa\u7840\u4e0a\u91cd\u8981\u7684\u66f4\u65b0\u8303\u5f0f\uff0c\u81ea GRPO \u8d77\uff0c\u540e\u8bad\u7ec3\u7b56\u7565\u4f18\u5316\u5c31\u4f5c\u4e3a\u5927\u6a21\u578b\u7684\u4e00\u4e2a\u91cd\u8981\u7814\u7a76\u65b9\u5411\u8fdb\u884c\uff0c\u5c31\u50cf\u6811\u8282\u70b9\u4e00\u6837\u5411\u5916\u5ef6\u4f38\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>PPO<\/strong>\uff1a\u4ee5 token \u4e3a\u6838\u5fc3\uff0c\u4f9d\u8d56\u4ef7\u503c\u51fd\u6570\u3002<\/li>\n\n\n\n<li><strong>GRPO<\/strong>\uff1a\u63d0\u51fa\u7ec4\u4f18\u5316\u601d\u8def\uff0c\u5728\u7ec4\u5185\u5bf9\u5956\u52b1\u505a\u5f52\u4e00\u5316\uff0c\u4ece\u800c\u6446\u8131\u4ef7\u503c\u51fd\u6570\u4f9d\u8d56\uff1b\u4f46\u4ecd\u505c\u7559\u5728 token \u7ea7\uff0c\u65b9\u5dee\u4f9d\u65e7\u8f83\u5927\u3002<\/li>\n\n\n\n<li><strong>DAPO<\/strong>\uff1a\u5728 GRPO \u57fa\u7840\u4e0a\u52a0\u5165\u5927\u91cf\u5de5\u7a0b\u6539\u8fdb\uff08\u5982 Clip-Higher\u3001Dynamic Sampling \u7b49\uff09\uff0c\u4e00\u5b9a\u7a0b\u5ea6\u7f13\u89e3\u5927\u6a21\u578b RL \u7684\u8bad\u7ec3\u74f6\u9888\uff0c\u4f46\u4ecd\u505c\u7559\u5728 token \u7ea7\u3002<\/li>\n\n\n\n<li><strong>GSPO<\/strong>\uff1a\u5b9e\u73b0\u8303\u5f0f\u8f6c\u53d8\uff0c\u5c06 off-policy \u4e0e clip \u5168\u90e8\u63d0\u5347\u5230 \u5e8f\u5217\u7ea7\uff0c\u663e\u8457\u964d\u4f4e\u65b9\u5dee\uff0c\u517c\u5177\u7b97\u6cd5\u7b80\u6d01\u6027\u4e0e\u6027\u80fd\u8868\u73b0\uff0c\u5df2\u6210\u4e3a Qwen3 RL \u7684\u6838\u5fc3\u5b9e\u8df5\u6846\u67b6\u3002<\/li>\n\n\n\n<li><strong>GFPO<\/strong>\uff1a\u9488\u5bf9\u540c\u65f6\u4f18\u5316\u591a\u4e2a\u6240\u9700\u5c5e\u6027\u7684\u76ee\u6807\u8fdb\u884c\u4f18\u5316\uff0c\u52a0\u5165\u6570\u636e\u8fc7\u6ee4\u64cd\u4f5c\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u5927\u8bed\u8a00\u6a21\u578b\u7684\u53d1\u5c55\u771f\u662f\u65e5\u65b0\u6708\u5f02\u3002 \u4ece DeepSeek \u6a2a\u7a7a\u51fa\u4e16\u4ee5\u6765\uff0c\u5176\u5728\u5927\u6a21\u578b\u540e\u8bad\u7ec3\u7684\u521b\u65b0 GRPO \u4e00\u8dc3\u6210\u4e3a [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":1177,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[10],"tags":[],"class_list":["post-1172","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-10"],"_links":{"self":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1172","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/comments?post=1172"}],"version-history":[{"count":1,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1172\/revisions"}],"predecessor-version":[{"id":1194,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/1172\/revisions\/1194"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media\/1177"}],"wp:attachment":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media?parent=1172"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/categories?post=1172"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/tags?post=1172"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}