{"id":26025,"date":"2025-05-26T09:09:43","date_gmt":"2025-05-26T01:09:43","guid":{"rendered":"https:\/\/aif.amtbbs.org\/?p=26025"},"modified":"2025-05-26T09:09:43","modified_gmt":"2025-05-26T01:09:43","slug":"deepseek%e7%94%a8%e7%9a%84grpo%e6%9c%89%e9%82%a3%e4%b9%88%e7%89%b9%e5%88%ab%e5%90%97%ef%bc%9f%e4%b8%87%e5%ad%97%e9%95%bf%e6%96%87%e5%88%86%e6%9e%90%e5%9b%9b%e7%af%87%e7%b2%be%e5%93%81%e8%ae%ba","status":"publish","type":"post","link":"https:\/\/aif.amtbbs.org\/index.php\/2025\/05\/26\/26025\/","title":{"rendered":"DeepSeek\u7528\u7684GRPO\u6709\u90a3\u4e48\u7279\u522b\u5417\uff1f\u4e07\u5b57\u957f\u6587\u5206\u6790\u56db\u7bc7\u7cbe\u54c1\u8bba\u6587"},"content":{"rendered":"<div class=\"article-desc\">\u672c\u6587\u8be6\u7ec6\u89e3\u8bfb\u4e86 Kimi k1.5\u3001OpenReasonerZero\u3001DAPO \u548c Dr. GRPO \u56db\u7bc7\u8bba\u6587\u4e2d\u7684\u521b\u65b0\u70b9\uff0c\u8bfb\u5b8c\u4f1a\u5bf9 GRPO \u53ca\u5176\u6539\u8fdb\u7b97\u6cd5\u6709\u66f4\u6df1\u7684\u7406\u89e3\uff0c\u8fdb\u800c\u542f\u53d1\u6784\u5efa\u63a8\u7406\u6a21\u578b\u7684\u65b0\u601d\u8def\u3002<\/div>\n<div id=\"postspictures\" class=\"article-content\">\n<div id=\"container\" class=\"container am-engine\" data-v-1d7a5742=\"\" data-element=\"root\">\n<p><img data-dominant-color=\"b5a2ad\" data-has-transparency=\"false\" style=\"--dominant-color: #b5a2ad;\" loading=\"lazy\" decoding=\"async\" class=\"not-transparent alignnone size-full wp-image-26027\" src=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/05\/81776e915125c8c304b151f683fa1813e3fe46-300x167-1.jpg\" width=\"300\" height=\"167\" alt=\"\" srcset=\"https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/05\/81776e915125c8c304b151f683fa1813e3fe46-300x167-1.jpg 300w, https:\/\/aiforumimage.oss-cn-shanghai.aliyuncs.com\/wp-content\/uploads\/2025\/05\/81776e915125c8c304b151f683fa1813e3fe46-300x167-1-150x84.jpg 150w\" sizes=\"auto, (max-width: 300px) 100vw, 300px\" \/><\/p>\n<p>\u968f\u7740 DeepSeek R1 \u7684\u6301\u7eed\u7206\u706b\uff0c\u63a8\u7406\u548c\u5f3a\u5316\u5b66\u4e60\u5df2\u7ecf\u6210\u4e3a AI \u9886\u57df\u7684\u70ed\u95e8\u8bcd\u6c47\u3002<\/p>\n<p>\u77ed\u77ed\u51e0\u4e2a\u6708\u7684\u65f6\u95f4\uff0c\u6211\u4eec\u5df2\u7ecf\u89c1\u8bc1\u4e86\u592a\u591a\u7684\u63a8\u7406\u5927\u6a21\u578b\uff0cAI \u66f4\u65b0\u8fed\u4ee3\u901f\u5ea6\u4f3c\u4e4e\u5df2\u7ecf\u5feb\u8fdb\u5230\u4e86\u4ee5\u5929\u4e3a\u5355\u4f4d\u3002<\/p>\n<p>\u4f46\u5728\u4f17\u591a\u7814\u7a76\u6210\u679c\u4e2d\u627e\u5230\u503c\u5f97\u5173\u6ce8\u7684\u5185\u5bb9\u5e76\u4e0d\u5bb9\u6613\u3002<\/p>\n<p>\u8fd9\u6709\u4e00\u7bc7\u4ef7\u503c\u975e\u5e38\u9ad8\u7684\u535a\u5ba2\uff0c\u53ef\u4ee5\u5e2e\u4f60\u68b3\u7406\u6700\u8fd1\u5173\u4e8e\u63a8\u7406\u6a21\u578b\u7684\u7814\u7a76\uff0c\u91cd\u70b9\u5173\u6ce8 DeepSeek R1 \u91cc\u7528\u5230\u7684 GRPO \u53ca\u540e\u7eed\u7684\u6539\u8fdb\u7b97\u6cd5\uff0c\u975e\u5e38\u503c\u5f97\u4e00\u8bfb\u3002\u4f5c\u8005\u662f\u6765\u81ea AI2 \u7684 ML \u79d1\u5b66\u5bb6 Nathan Lambert\uff0c\u4ed6\u535a\u58eb\u6bd5\u4e1a\u4e8e UC \u4f2f\u514b\u5229\uff0c\u66fe\u5728 HuggingFace \u9886\u5bfc RLHF \u56e2\u961f\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s7.51cto.com\/oss\/202505\/25\/c647c1871e6ba02bad3712b2c8f0fe316ba8d7.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u535a\u5ba2\u5730\u5740\uff1ahttps:\/\/www.interconnects.ai\/p\/papers-im-reading-base-model-rl-grpo<\/p>\n<p>\u6587\u7ae0\u5217\u4e3e\u4e86\u6700\u8fd1\u6bd4\u8f83\u706b\u7684\u8bba\u6587\u548c\u5927\u6a21\u578b\uff0c\u5305\u62ec\uff1a<\/p>\n<ul data-id=\"ua73dd4b-jWJb6TVS\">\n<li data-id=\"ld70c578-PYqdPadf\">Kimi k1.5\uff1a\u8fd9\u662f\u4e00\u4e2a\u4e0e DeepSeek R1 \u540c\u4e00\u5929\u53d1\u5e03\u7684\u63a8\u7406\u6a21\u578b\uff0c\u5176\u6280\u672f\u62a5\u544a\u975e\u5e38\u8be6\u7ec6\uff1b<\/li>\n<li data-id=\"ld70c578-28GaVB7D\">OpenReasonerZero\uff1a\u8fd9\u662f\u9996\u6b21\u5bf9\u57fa\u7840\u6a21\u578b\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u7684\u5b8c\u6574\u590d\u73b0\uff1b<\/li>\n<li data-id=\"ld70c578-PTbI6jOi\">DAPO\uff1a\u8fd9\u662f\u9996\u7bc7\u63a2\u8ba8\u5bf9 GRPO \u8fdb\u884c\u6539\u8fdb\u4ee5\u66f4\u597d\u5730\u9002\u5e94\u63a8\u7406\u8bad\u7ec3\u7684\u8bba\u6587\uff08\u53c2\u89c1\u300a\u8d85\u8d8a DeepSeek GRPO \u7684\u5173\u952e RL \u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534e AIR \u5f00\u6e90 DAPO\u300b\uff09\uff1b<\/li>\n<li data-id=\"ld70c578-EciKVrKa\">Dr. GRPO\uff1a\u8fd9\u7bc7\u8bba\u6587\u6df1\u5165\u7814\u7a76\u4e86\u4ece\u57fa\u7840\u6a21\u578b\u5f00\u59cb\u7684\u5f3a\u5316\u5b66\u4e60\uff0c\u5e76\u63d0\u51fa\u4e86\u5bf9 GRPO \u7684\u6539\u8fdb\u4ee5\u63d0\u5347\u5b66\u4e60\u6548\u679c\uff08\u53c2\u89c1\u300a\u63ed\u79d8 DeepSeek R1-Zero \u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO \u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848\u300b\uff09\u3002<\/li>\n<\/ul>\n<p>\u6b64\u5916\uff0c\u4f5c\u8005\u8fd8\u7ed9\u51fa\u4e86\u53c2\u8003\u8bba\u6587\u4e2d\u91cd\u590d\u770b\u5230\u7684\u635f\u5931\u51fd\u6570\uff0c\u6211\u4eec\u4e0d\u96be\u53d1\u73b0\u8fd9\u5e94\u8be5\u662f\u6bd4\u8f83\u91cd\u8981\u7684\u635f\u5931\u51fd\u6570\uff1a<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s3.51cto.com\/oss\/202505\/25\/87a07dd67d7339833df5968a2e7fcf3a6a0bc2.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<h4>\u662f\u65f6\u5019\u7ed9 GRPO \u964d\u964d\u6e29\u4e86<\/h4>\n<p>\u73b0\u5728\u5f88\u591a\u4eba\u88ab RL \u5728\u8bed\u8a00\u5efa\u6a21\u9886\u57df\u7684\u8868\u73b0\u6240\u5438\u5f15\uff0c\u8fd9\u7ed9\u4eba\u4e00\u79cd\u9519\u89c9\uff0c\u5f77\u4f5b GRPO \u548c DeepSeek R1\uff08\u4ee5\u53ca\u4e4b\u524d\u7684\u6a21\u578b\uff09\u7684\u5de5\u4f5c\u5df2\u7ecf\u5f00\u542f\u4e86 RL \u8bad\u7ec3\u7684\u5168\u65b0\u65f6\u4ee3\u3002<\/p>\n<p>\u4f46\u4e8b\u5b9e\u8fdc\u975e\u5982\u6b64\u3002<\/p>\n<p>\u5176\u5b9e GRPO \u5e76\u4e0d\u662f\u4e00\u79cd\u7279\u6b8a\u7684 RL \u7b97\u6cd5\u3002<\/p>\n<p>\u76ee\u524d\u8bb8\u591a\u9886\u5148\u7684\u7814\u7a76\u5de5\u4f5c\u548c\u5b9e\u9a8c\u5ba4\u5e76\u6ca1\u6709\u4f7f\u7528 GRPO \u8fdb\u884c\u7814\u7a76\u3002<\/p>\n<p>\u5b9e\u9645\u4e0a\uff0cGRPO \u4e0e\u5176\u4ed6 RL \u7b97\u6cd5\u5173\u7cfb\u6781\u4e3a\u5bc6\u5207 \u2014\u2014 \u5b83\u6e90\u81ea PPO\uff08\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff09\uff0c\u5e76\u4e14\u5177\u6709\u4e0e RLOO \uff08REINFORCE Leave One Out\uff09\u8d85\u7ea7\u76f8\u4f3c\u7684\u8ba1\u7b97\u4f18\u52bf\u3002<\/p>\n<p>GRPO \u786e\u5b9e\u5305\u542b\u4e86\u5de7\u5999\u7684\u6539\u8fdb\uff0c\u5c24\u5176\u662f\u5728\u63a8\u7406\u8bad\u7ec3\uff08reasoning training\uff09\u800c\u975e\u4f20\u7edf\u7684 RLHF \u573a\u666f\u4e0b\u3002<\/p>\n<p>\u4f20\u7edf RLHF \u5b9e\u8df5\u6cbf\u88ad\u4e86\u65e9\u671f RL \u6587\u732e\u7684\u505a\u6cd5\uff0c\u901a\u5e38\u6bcf\u4e2a\u6279\u6b21\u4e2d\u6bcf\u4e2a\u63d0\u793a\u8bcd\u4ec5\u91c7\u6837\u4e00\u4e2a\u751f\u6210\u7ed3\u679c\u8fdb\u884c\u8bad\u7ec3\u3002\u800c\u5728\u63a8\u7406\u4efb\u52a1\u4e2d\uff0c\u6211\u4eec\u73b0\u5728\u4f1a\u751f\u6210\u591a\u4e2a\u7b54\u6848\u3002<\/p>\n<p>\u82e5\u4e0d\u6df1\u5165\u6280\u672f\u7ec6\u8282\uff0c\u73b0\u4ee3\u5b9e\u73b0\u4e2d GRPO \u548c RLOO \u7684\u4f18\u52bf\u503c\u8ba1\u7b97\u51e0\u4e4e\u5982\u51fa\u4e00\u8f99 \u2014\u2014 \u8fd9\u4e0e PPO \u5f62\u6210\u9c9c\u660e\u5bf9\u6bd4\uff08PPO \u7684\u4f18\u52bf\u503c\u6765\u6e90\u4e8e\u4ef7\u503c\u51fd\u6570\uff0c\u901a\u5e38\u91c7\u7528 GAE \u65b9\u6cd5\u8ba1\u7b97\uff09\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s9.51cto.com\/oss\/202505\/25\/1590e752364a518c8609427b10634c0a5d7059.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u56e0\u6b64\uff0cREINFORCE \u4e0e GRPO \u7684\u552f\u4e00\u533a\u522b\u4ec5\u5728\u4e8e PPO \u7684 clipping logic \u673a\u5236 \u2014\u2014 \u5b83\u4eec\u672c\u8d28\u4e0a\u90fd\u662f\u540c\u5b97\u540c\u6e90\u7684\u7b56\u7565\u68af\u5ea6\u7b97\u6cd5\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u524d LLM \u65f6\u4ee3\u6d41\u884c\u7684\u53e6\u4e00\u4e2a RL \u7b97\u6cd5 A2C\uff0c\u6839\u636e\u8d85\u53c2\u6570\u8bbe\u7f6e\u7684\u4e0d\u540c\uff0c\u4e5f\u53ef\u4ee5\u89c6\u4e3a PPO \u7684\u7279\u6b8a\u53d8\u4f53\u3002<\/p>\n<p>\u8fd9\u91cc\u9700\u8981\u628a\u63e1\u7684\u6838\u5fc3\u8ba4\u77e5\u662f\uff1a\u5f53\u524d\u4f7f\u7528\u7684\u6240\u6709 RL \u7b97\u6cd5\u5728\u5b9e\u73b0\u5c42\u9762\u4e0a\u662f\u9ad8\u5ea6\u76f8\u4f3c\u7684\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u5c3d\u7ba1 GRPO \u662f\u5f53\u524d\u6700\u6d41\u884c\u7684\u7b97\u6cd5\uff0c\u4f46\u5982\u4eca RL \u7b97\u6cd5\u7684\u53d8\u9769\u5176\u5b9e\u53ea\u805a\u7126\u5728\u51e0\u4e2a\u6838\u5fc3\u7ef4\u5ea6\uff1a<\/p>\n<ul data-id=\"ua73dd4b-3DejDomM\">\n<li data-id=\"ld70c578-cOQr42hM\">\u4ef7\u503c\u51fd\u6570\u7684\u53d6\u820d\uff1a\u4e1a\u754c\u6b63\u9010\u6e10\u8f6c\u5411\u76f4\u63a5\u4f30\u8ba1\u4f18\u52bf\u503c\uff08advantage\uff09\uff0c\u56e0\u4e3a\u4ef7\u503c\u51fd\u6570\uff08value function\uff09\u7684\u5efa\u6a21\u5f80\u5f80\u8f83\u4e3a\u56f0\u96be\uff1b<\/li>\n<li data-id=\"ld70c578-paWcDCBq\">DeepSeek \u7684\u7814\u7a76\u6210\u679c\u5f15\u7206\u4e86\u8fd9\u573a\u53d8\u9769\uff0c\u56e0\u6b64\u4eba\u4eec\u81ea\u7136\u4ece\u5176 GRPO \u7b97\u6cd5\u5207\u5165\u63a2\u7d22\u3002\u4f46\u672c\u8d28\u4e0a\uff0c\u8fd9\u4e00\u6d6a\u6f6e\u7684\u63a8\u52a8\u529b\u662f\u5f3a\u5316\u5b66\u4e60\u8303\u5f0f\u7684\u8fdb\u5316\uff0c\u800c\u975e\u67d0\u4e2a\u7279\u5b9a\u7b97\u6cd5\u7684\u7a81\u7834\u3002<\/li>\n<\/ul>\n<h4>Kimi k1.5<\/h4>\n<p>\u300aKimi k1.5\u300b\u7684\u62a5\u544a\u5185\u5bb9\u975e\u5e38\u4e30\u5bcc\uff0c\u8bba\u6587\u957f\u8fbe 25 \u9875\u3002\u4e0d\u8fc7\uff0c\u5176\u5e76\u672a\u5f00\u653e\u6a21\u578b\u6743\u91cd\u3002<\/p>\n<p>\u8fd9\u7bc7\u8bba\u6587\u62a5\u544a\u4e86 Kimi k1.5 \u7684\u8bad\u7ec3\u5b9e\u8df5\uff0c\u8fd9\u662f Kimi \u56e2\u961f\u6700\u65b0\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\uff0c\u91c7\u7528 RL \u8fdb\u884c\u8bad\u7ec3\uff0c\u5305\u62ec\u5176\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u6280\u672f\u3001\u591a\u6a21\u6001\u6570\u636e\u914d\u65b9\u4ee5\u53ca\u57fa\u7840\u8bbe\u65bd\u4f18\u5316\u3002\u957f\u4e0a\u4e0b\u6587\u6269\u5c55\u548c\u6539\u8fdb\u7684\u7b56\u7565\u4f18\u5316\u65b9\u6cd5\u662f Kimi \u56e2\u961f\u65b9\u6cd5\u7684\u6838\u5fc3\u8981\u7d20\uff0c\u4ed6\u4eec\u5efa\u7acb\u4e86\u4e00\u4e2a\u7b80\u5355\u800c\u6709\u6548\u7684\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\uff0c\u65e0\u9700\u4f9d\u8d56\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22\u3001\u4ef7\u503c\u51fd\u6570\u548c\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\u7b49\u66f4\u590d\u6742\u7684\u6280\u672f\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s2.51cto.com\/oss\/202505\/25\/e243bfb368b28bb959e09726a1b5f4d2c4a36e.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<ul data-id=\"ua73dd4b-BDUQc8Md\">\n<li data-id=\"ld70c578-NDJuCQ5n\">\u8bba\u6587\u6807\u9898\uff1a KIMI K1.5: SCALING REINFORCEMENT LEARNING WITH LLMS<\/li>\n<li data-id=\"ld70c578-Y2EIwlje\">\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2501.12599<\/li>\n<\/ul>\n<p>\u8be5\u6a21\u578b\u5728 o3-mini \u53d1\u5e03\u4e4b\u524d\u5c31\u5df2\u7ecf\u63a8\u51fa\uff0c\u5176\u8bc4\u4f30\u7ed3\u679c\u975e\u5e38\u51fa\u8272\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s7.51cto.com\/oss\/202505\/25\/397c0b6654dfbfdb94544016c6e06b69b4607f.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p><strong>\u6570\u636e\u5206\u5e03<\/strong><\/p>\n<p>\u8fd9\u7bc7\u8bba\u6587\uff08\u4ee5\u53ca\u672c\u6587\u540e\u9762\u63d0\u5230\u7684\u300aOpen Reasoner Zero\u300b\uff09\u90fd\u5305\u542b\u4e86 01 \u548c R1 \u7248\u672c\u6240\u6ca1\u6709\u7684\u5bf9\u6570\u636e\u7684\u8ba8\u8bba\u3002Kimi \u56e2\u961f\u5f3a\u8c03\u4e86\u4e3a RL \u8fdb\u884c\u63d0\u793a\u7b56\u5212\uff08prompt curation\uff09\u7684\u91cd\u8981\u6027\u3002\u8fd9\u542c\u8d77\u6765\u5f88\u7b80\u5355\uff0c\u4f46\u5f3a\u5316\u5b66\u4e60\u63d0\u793a\u96c6\u7684\u8d28\u91cf\u548c\u591a\u6837\u6027\u5728\u786e\u4fdd\u5f3a\u5316\u5b66\u4e60\u7684\u6709\u6548\u6027\u65b9\u9762\u8d77\u7740\u5173\u952e\u4f5c\u7528\u3002\u7531\u6b64\uff0c\u56e2\u961f\u4eba\u5458\u6307\u51fa\u4e86\u4e24\u70b9\u4e0e\u6211\u4eec\u76ee\u524d\u770b\u5230\u7684\u5927\u591a\u6570\u4ec5\u6570\u5b66\u6a21\u578b\u4e0d\u540c\u7684\u5730\u65b9\uff1a<\/p>\n<ul data-id=\"ua73dd4b-WRQHbpTU\">\n<li data-id=\"ld70c578-v8w5UEjl\">\u591a\u6837\u5316\u8986\u76d6\uff1a\u63d0\u793a\u5e94\u6db5\u76d6\u5e7f\u6cdb\u7684\u5b66\u79d1\uff0c\u4f8b\u5982 STEM\u3001\u7f16\u7801\u548c\u4e00\u822c\u63a8\u7406\uff0c\u4ee5\u589e\u5f3a\u6a21\u578b\u7684\u9002\u5e94\u6027\u5e76\u786e\u4fdd\u5728\u4e0d\u540c\u9886\u57df\u7684\u5e7f\u6cdb\u9002\u7528\u6027\u3002<\/li>\n<li data-id=\"ld70c578-ieL0dwAp\">\u5e73\u8861\u96be\u5ea6\uff1a\u63d0\u793a\u96c6\u5e94\u5305\u62ec\u5206\u5e03\u5747\u5300\u7684\u6613\u3001\u4e2d\u3001\u96be\u95ee\u9898\uff0c\u4ee5\u4fc3\u8fdb\u6e10\u8fdb\u5b66\u4e60\u5e76\u9632\u6b62\u8fc7\u5ea6\u62df\u5408\u5230\u7279\u5b9a\u7684\u590d\u6742\u7a0b\u5ea6\u3002<\/li>\n<\/ul>\n<p>\u5728\u4efb\u52a1\u96be\u5ea6\u65b9\u9762\uff0cKimi \u56e2\u961f\u91c7\u7528\u4e86\u4e00\u79cd\u4e0e\u63a8\u7406\u6a21\u578b\u76f8\u5173\u7684\u8f83\u65b0\u7684\u65b9\u6cd5\uff1a \u4ed6\u4eec\u91c7\u7528\u57fa\u4e8e\u6a21\u578b\u7684\u65b9\u6cd5\uff0c\u5229\u7528\u6a21\u578b\u81ea\u8eab\u7684\u80fd\u529b\u6765\u9002\u5e94\u6027\u5730\u8bc4\u4f30\u6bcf\u4e2a\u63d0\u793a\u7684\u96be\u5ea6\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5bf9\u4e8e\u6bcf\u4e2a\u63d0\u793a\uff0c\u4e00\u4e2a\u7ecf\u8fc7\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u7684\u6a21\u578b\u4f7f\u7528\u76f8\u5bf9\u8f83\u9ad8\u7684\u91c7\u6837\u6e29\u5ea6\u751f\u6210\u7b54\u6848\u5341\u6b21\u3002\u7136\u540e\u8ba1\u7b97\u901a\u8fc7\u7387\uff0c\u5e76\u5c06\u5176\u4f5c\u4e3a\u63d0\u793a\u96be\u5ea6\u7684\u4ee3\u7406\uff08proxy\uff09\u2014\u2014 \u901a\u8fc7\u7387\u8d8a\u4f4e\uff0c\u96be\u5ea6\u8d8a\u9ad8\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u4ed6\u4eec\u8fd8\u79fb\u9664\u4e86\u4e00\u4e9b\u53ef\u80fd\u4fc3\u4f7f\u6a21\u578b\u731c\u6d4b\u800c\u4e0d\u662f\u8fdb\u884c\u63a8\u7406\u7684\u95ee\u9898\uff1a \u7ecf\u9a8c\u89c2\u5bdf\u8868\u660e\uff0c\u4e00\u4e9b\u590d\u6742\u7684\u63a8\u7406\u95ee\u9898\u53ef\u80fd\u6709\u76f8\u5bf9\u7b80\u5355\u4e14\u5bb9\u6613\u731c\u6d4b\u7684\u7b54\u6848\uff0c\u8fd9\u4f1a\u5bfc\u81f4\u5047\u9633\u6027\u9a8c\u8bc1 \u2014\u2014 \u6a21\u578b\u901a\u8fc7\u4e0d\u6b63\u786e\u7684\u63a8\u7406\u8fc7\u7a0b\u5f97\u51fa\u4e86\u6b63\u786e\u7b54\u6848\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u4ed6\u4eec\u6392\u9664\u4e86\u5bb9\u6613\u51fa\u73b0\u8fd9\u79cd\u9519\u8bef\u7684\u95ee\u9898\uff0c\u4f8b\u5982\u9009\u62e9\u9898\u3001\u57fa\u4e8e\u8bc1\u660e\u7684\u95ee\u9898\u3002<\/p>\n<p><strong>\u8bad\u7ec3\u65b9\u6cd5<\/strong><\/p>\n<p>Kimi K1.5 \u7684\u8bad\u7ec3\u65b9\u6848\u5305\u542b\u4e86\u8bb8\u591a\u6709\u8da3\u7684\u7ec6\u8282\uff0c\u4f46\u968f\u7740\u8bad\u7ec3\u6280\u672f\u7684\u6210\u719f\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u53ef\u80fd\u4e0d\u4f1a\u6210\u4e3a\u957f\u671f\u63a8\u8350\u7684\u6700\u4f73\u5b9e\u8df5\u3002<\/p>\n<p>\u4f8b\u5982\uff0c\u4ed6\u4eec\u7684\u521d\u59cb\u9636\u6bb5\u4e0e DeepSeek R1 \u8bba\u6587\u975e\u5e38\u76f8\u4f3c\uff1a\u91c7\u7528 SFT\uff08\u76d1\u7763\u5fae\u8c03\uff09\u9884\u70ed\uff0c\u7ed3\u5408\u957f\u601d\u7ef4\u94fe\uff08CoT\uff09\u548c\u62d2\u7edd\u91c7\u6837\uff08rejection sampling\uff09\u3002<\/p>\n<p>\u53c8\u6bd4\u5982\uff0c\u4ed6\u4eec\u91cd\u70b9\u5173\u6ce8\u6570\u636e\u4e2d\u7684\u884c\u4e3a\u6a21\u5f0f\uff0c\u5305\u62ec\u89c4\u5212\uff08planning\uff09\u3001\u8bc4\u4f30\uff08evaluation\uff09\u3001\u53cd\u601d\uff08reflection\uff09\u548c\u63a2\u7d22\uff08exploration\uff09\uff0c\u8fd9\u4e9b\u5bf9\u6700\u7ec8\u6027\u80fd\u63d0\u5347\u81f3\u5173\u91cd\u8981\u3002<\/p>\n<p>\u8fdb\u5165\u540e\u7eed\u8bad\u7ec3\u9636\u6bb5\u540e\uff0c\u4ed6\u4eec\u7684\u65b9\u6cd5\u53d8\u5f97\u66f4\u52a0\u6709\u8da3\uff1a\u672a\u91c7\u7528 GRPO\uff0c\u800c\u662f\u4f7f\u7528\u4e86\u4e00\u79cd\u5728\u7ebf\u7b56\u7565\u955c\u50cf\u4e0b\u964d\uff08online policy mirror descent\uff09 \u7684\u53d8\u4f53\uff08\u4ecd\u5c5e\u4e8e\u7b56\u7565\u68af\u5ea6\u7b97\u6cd5\u5bb6\u65cf\uff09\u3002<\/p>\n<p>\u9664\u6b64\u4e4b\u5916\uff0c\u4ed6\u4eec\u672a\u4f7f\u7528\u4ef7\u503c\u51fd\u6570\uff0c\u800c\u662f\u91c7\u7528\u8499\u7279\u5361\u6d1b\u5956\u52b1\u57fa\u7ebf\uff08Monte Carlo reward baseline\uff09\uff0c\u5176\u6838\u5fc3\u601d\u60f3\u4e0e GRPO \u7c7b\u4f3c\uff0c\u4f46\u5e76\u975e\u76f4\u63a5\u7528\u4e8e\u4f18\u52bf\u4f30\u8ba1\uff08advantage\uff09\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s7.51cto.com\/oss\/202505\/25\/c3c0f7e851aed32b0a0179661758384c7ff629.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u4e3a\u4e86\u63d0\u9ad8\u6a21\u578b\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u548c\u6548\u679c\uff0c\u7814\u7a76\u8005\u4eec\u91c7\u7528\u4e86\u591a\u79cd\u7b56\u7565\u3002\u5176\u4e2d\u4e00\u79cd\u7b56\u7565\u662f\u5f15\u5165\u957f\u5ea6\u60e9\u7f5a\uff0c\u5373\u9f13\u52b1\u751f\u6210\u8f83\u77ed\u7684\u56de\u7b54\uff0c\u5e76\u5728\u6b63\u786e\u56de\u7b54\u4e2d\u60e9\u7f5a\u8f83\u957f\u7684\u56de\u7b54\uff0c\u540c\u65f6\u660e\u786e\u60e9\u7f5a\u9519\u8bef\u7b54\u6848\u4e2d\u7684\u957f\u56de\u7b54\u3002\u8fd9\u79cd\u7b56\u7565\u6709\u52a9\u4e8e\u63a7\u5236\u6a21\u578b\u751f\u6210\u56de\u7b54\u7684\u957f\u5ea6\uff0c\u907f\u514d\u6a21\u578b\u8fc7\u5ea6\u751f\u6210\u5197\u957f\u4e14\u53ef\u80fd\u4e0d\u51c6\u786e\u7684\u5185\u5bb9\uff0c\u4ece\u800c\u63d0\u9ad8\u8bad\u7ec3\u7684\u7a33\u5b9a\u6027\u3002\u5c3d\u7ba1\u8fd9\u79cd\u65b9\u6cd5\u5728\u8bad\u7ec3\u521d\u671f\u53ef\u80fd\u4f1a\u51cf\u6162\u8bad\u7ec3\u901f\u5ea6\uff0c\u4f46\u7814\u7a76\u8005\u4eec\u4f1a\u9010\u6e10\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5f15\u5165\u8fd9\u79cd\u5956\u52b1\u673a\u5236\uff0c\u4ee5\u5b9e\u73b0\u66f4\u597d\u7684\u8bad\u7ec3\u6548\u679c\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u7814\u7a76\u8005\u4eec\u8fd8\u91c7\u7528\u4e86\u6570\u636e\u5e8f\u5217\u7b56\u7565\u6765\u8f85\u52a9\u6a21\u578b\u5b66\u4e60\u3002\u8fd9\u79cd\u65b9\u6cd5\u7c7b\u4f3c\u4e8e\u4e00\u79cd\u660e\u786e\u7684\u6559\u5b66\u5927\u7eb2\uff0c\u5373\u4ece\u8f83\u7b80\u5355\u7684\u4efb\u52a1\u5f00\u59cb\u8bad\u7ec3\uff0c\u5e76\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5bf9\u6a21\u578b\u8868\u73b0\u4e0d\u4f73\u7684\u4efb\u52a1\u8fdb\u884c\u91cd\u65b0\u91c7\u6837\uff0c\u589e\u52a0\u8fd9\u4e9b\u4efb\u52a1\u7684\u8bad\u7ec3\u9891\u7387\u3002\u8fd9\u79cd\u7b56\u7565\u7c7b\u4f3c\u4e8e\u9010\u6b65\u5f15\u5bfc\u6a21\u578b\u5b66\u4e60\uff0c\u7c7b\u4f3c\u4e8e\u4eba\u7c7b\u5b66\u4e60\u8fc7\u7a0b\u4e2d\u4ece\u6613\u5230\u96be\u7684\u9010\u6b65\u8fdb\u9636\u3002\u5c3d\u7ba1\u8fd9\u79cd\u65b9\u6cd5\u53ef\u80fd\u4f1a\u589e\u52a0\u8bad\u7ec3\u7684\u590d\u6742\u6027\uff0c\u4f46\u5b83\u88ab\u89c6\u4e3a\u4e00\u79cd\u6709\u6548\u7684\u6280\u5de7\uff0c\u53ef\u4ee5\u5e2e\u52a9\u6a21\u578b\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u9010\u6b65\u63d0\u5347\u6027\u80fd\u3002<\/p>\n<p>\u8fd9\u4e9b\u65b9\u6cd5\u867d\u7136\u5728\u77ed\u671f\u5185\u53ef\u80fd\u4f1a\u589e\u52a0\u8bad\u7ec3\u7684\u590d\u6742\u6027\uff0c\u4f46\u5b83\u4eec\u6709\u52a9\u4e8e\u6a21\u578b\u5728\u957f\u671f\u8bad\u7ec3\u4e2d\u4fdd\u6301\u7a33\u5b9a\u6027\u548c\u4e00\u81f4\u6027\uff0c\u4ece\u800c\u63d0\u9ad8\u6a21\u578b\u7684\u6574\u4f53\u6027\u80fd\u548c\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s5.51cto.com\/oss\/202505\/25\/9183a9f2436cce397fd212419e5ff19728ccb7.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u5728\u5173\u4e8e\u6a21\u578b\u5927\u5c0f\u7684\u6d88\u878d\u7814\u7a76\u4e2d\uff08\u5c3d\u7ba1\u6ca1\u6709\u660e\u786e\u63d0\u53ca\u6a21\u578b\u7684\u5177\u4f53\u5927\u5c0f\uff09\uff0c\u4ed6\u4eec\u53d1\u73b0\uff0c\u5c3d\u7ba1\u8f83\u5927\u7684\u6a21\u578b\u5728\u521d\u59cb\u9636\u6bb5\u8868\u73b0\u4f18\u4e8e\u8f83\u5c0f\u7684\u6a21\u578b\uff0c\u4f46\u8f83\u5c0f\u7684\u6a21\u578b\u901a\u8fc7\u5229\u7528\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u4f18\u5316\u7684\u66f4\u957f\u7684\u601d\u7ef4\u94fe\uff08CoTs\uff09\u4e5f\u80fd\u8fbe\u5230\u76f8\u5f53\u7684\u6027\u80fd\u3002\u7136\u800c\uff0c\u8f83\u5927\u7684\u6a21\u578b\u901a\u5e38\u5728 token \u6548\u7387\u65b9\u9762\u8868\u73b0\u5f97\u6bd4\u5c0f\u6a21\u578b\u66f4\u597d\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s8.51cto.com\/oss\/202505\/25\/d73e27d850065daac83231dd1bc57041237f5f.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u53e6\u5916\uff0c\u8fd9\u7bc7\u8bba\u6587\u5bf9\u6a21\u578b\u6700\u7ec8\u5b9e\u7528\u6027\u65b9\u9762\u7684\u603b\u7ed3\u975e\u5e38\u6709\u8da3\uff0c\u4e5f\u4e0e\u8fd1\u671f\u8bb8\u591a\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6587\u732e\u4e2d\u7684\u89c2\u70b9\u4e00\u81f4\uff1a \u5982\u679c\u76ee\u6807\u662f\u5c3d\u53ef\u80fd\u8fbe\u5230\u6700\u4f73\u6027\u80fd\uff0c\u90a3\u4e48\u6269\u5927\u8f83\u5927\u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u957f\u5ea6\u5177\u6709\u66f4\u9ad8\u7684\u4e0a\u9650\uff0c\u5e76\u4e14\u5728 token \u6548\u7387\u65b9\u9762\u66f4\u5177\u4f18\u52bf\u3002\u7136\u800c\uff0c\u5982\u679c\u6d4b\u8bd5\u65f6\u8ba1\u7b97\u8d44\u6e90\u6709\u9650\uff0c\u90a3\u4e48\u8bad\u7ec3\u5177\u6709\u8f83\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6\u7684\u5c0f\u578b\u6a21\u578b\u53ef\u80fd\u662f\u53ef\u884c\u7684\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n<p>\u8bba\u6587\u4e2d\u8fd8\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u4ed6\u4eec\u7684\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u6570\u636e\u96c6\u3001\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u57fa\u7840\u8bbe\u65bd\u3001\u957f\u601d\u7ef4\u94fe\u5230\u77ed\u601d\u7ef4\u94fe\u7684\u84b8\u998f\u8fc7\u7a0b\u3002\u611f\u5174\u8da3\u7684\u8bfb\u8005\u53ef\u4ee5\u67e5\u770b\u8bba\u6587\u6df1\u5165\u4e86\u89e3\u3002<\/p>\n<h4>Open- reasoner &#8211; zero<\/h4>\n<p>\u8fd9\u7bc7\u8bba\u6587\u7684\u4e3b\u8981\u8d21\u732e\u5728\u4e8e\uff0c\u5b83\u662f\u7b2c\u4e00\u7bc7\u5c55\u793a\u5728\u57fa\u7840\u6a21\u578b\u4e0a\u901a\u8fc7 RL \u53d6\u5f97\u975e\u5e38\u51fa\u8272\u7ed3\u679c\u7684\u7814\u7a76\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s3.51cto.com\/oss\/202505\/25\/193e9595398aa6de480240979b397a164ab3f0.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/github.com\/Open-Reasoner-Zero\/Open-Reasoner-Zero\/blob\/main\/ORZ_paper.pdf<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s2.51cto.com\/oss\/202505\/25\/32fb15830e111d0ace13293fd00b0c23e1a264.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u4e8b\u5b9e\u4e0a\uff0c\u4ed6\u4eec\u4f7f\u7528 Qwen-32B \u57fa\u5ea7\u6a21\u578b\u65f6\uff0c\u80fd\u591f\u8fbe\u5230 DeepSeek \u8bba\u6587\u5728\u5176\u84b8\u998f\u90e8\u5206\u63d0\u5230\u7684 Qwen-32B \u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6a21\u578b\u7684\u6027\u80fd\u3002\u4e0d\u8fc7\uff0cDeepSeek \u8bba\u6587\u4e2d\u63d0\u5230\u7684\u8fd9\u4e2a RL \u6a21\u578b\u5e76\u672a\u5f15\u8d77\u592a\u591a\u8ba8\u8bba\uff0c\u56e0\u4e3a\u540c\u4e00\u8868\u683c\u663e\u793a\uff0cDeepSeek \u53d1\u73b0\u5728\u8fd9\u79cd\u89c4\u6a21\u7684\u6a21\u578b\u4e2d\uff0c\u84b8\u998f\u63a8\u7406\u80fd\u529b\uff08\u800c\u975e\u4ece\u5f3a\u5316\u5b66\u4e60\u5f00\u59cb\uff09\u80fd\u5e26\u6765\u66f4\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s8.51cto.com\/oss\/202505\/25\/a69f25332c7e565d82b4443d7351775ef1c034.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>R1 \u8bba\u6587\uff1ahttps:\/\/arxiv.org\/pdf\/2501.12948<\/p>\n<p>\u8fd9\u91cc\uff0c\u95ee\u9898\u5728\u4e8e\u5e76\u975e\u6240\u6709\u7684 RL \u6b65\u9aa4\u90fd\u662f\u7b49\u4ef7\u7684\u3002\u5b83\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u53d6\u51b3\u4e8e\uff1a<\/p>\n<ul data-id=\"ua73dd4b-03fS6TjH\">\n<li data-id=\"ld70c578-EpE4HBWV\">\u7528\u4e8e RL \u7684\u6279\u6b21\u5927\u5c0f\uff0c\u66f4\u5927\u7684\u6279\u6b21\u5927\u5c0f\u610f\u5473\u7740\u66f4\u5c11\u7684\u6b65\u9aa4\uff1b<\/li>\n<li data-id=\"ld70c578-ueyxsWmU\">\u6bcf\u4e2a\u6b65\u9aa4\u4e2d\u4f7f\u7528\u7684\u68af\u5ea6\u6b65\u9aa4\u6570\uff0c\u66f4\u591a\u7684\u68af\u5ea6\u6b65\u9aa4\u53ef\u4ee5\u5bfc\u81f4\u66f4\u591a\u7684\u5b66\u4e60\uff0c\u4f46\u540c\u65f6\u4e5f\u53ef\u80fd\u5bfc\u81f4\u66f4\u591a\u7684\u4e0d\u7a33\u5b9a\u6027\uff1b<\/li>\n<li data-id=\"ld70c578-uDLlnixP\">\u6570\u636e\u96c6\u4e0e\u7279\u5b9a\u6a21\u578b\u7684\u5339\u914d\u7a0b\u5ea6\uff0cDeepSeek \u53ef\u80fd\u4f7f\u7528\u4e86\u4e0e DeepSeek \u57fa\u7840\u6a21\u578b\u76f8\u540c\u7684\u6570\u636e\u96c6\uff0c\u8fd9\u53ef\u80fd\u4e0d\u592a\u9002\u5408 Qwen \u6a21\u578b\uff08\u5b66\u4e60\u901f\u5ea6\u66f4\u6162\uff09\u3002<\/li>\n<\/ul>\n<p><strong>\u6570\u636e\u5206\u5e03<\/strong><\/p>\n<p>\u8fd9\u9879\u5de5\u4f5c\u7684\u6838\u5fc3\u6210\u529f\u4e4b\u5904\u5728\u4e8e\uff0c\u5b83\u975e\u5e38\u6e05\u6670\u5730\u5c55\u793a\u4e86\u6570\u636e\u5bf9\u4e8e\u5b66\u4e60\u7684\u91cd\u8981\u6027\u3002<\/p>\n<p>\u4ed6\u4eec\u4ece\u5404\u79cd\u6765\u6e90\u6536\u96c6\u516c\u5f00\u6570\u636e\uff0c\u5305\u62ec AIME\uff08\u622a\u81f3 2023 \u5e74\uff09\u3001MATH\u3001Numina-Math \u6570\u636e\u96c6\u3001Tulu3 MATH \u4ee5\u53ca\u5176\u4ed6\u5f00\u6e90\u6570\u636e\u96c6\u3002\u6839\u636e\u6570\u636e\u6765\u6e90\u548c\u95ee\u9898\u96be\u5ea6\uff0c\u4ed6\u4eec\u63d0\u53d6\u4e86 AMC\u3001AIME\u3001\u6570\u5b66\u3001\u5965\u6797\u5339\u514b\u7ade\u8d5b\u4ee5\u53ca AoPS \u8bba\u575b\u7684\u76f8\u5173\u90e8\u5206\uff0c\u4f5c\u4e3a\u96be\u5ea6\u8f83\u9ad8\u7684\u63d0\u793a\uff0c\u4ee5\u786e\u4fdd\u6570\u636e\u96c6\u7684\u96be\u5ea6\u6c34\u5e73\u9002\u5f53\u3002<\/p>\n<p>\u8be5\u7814\u7a76\u8fd8\u901a\u8fc7\u7a0b\u5e8f\u5316\u65b9\u6cd5\u5408\u6210\u989d\u5916\u7684\u63a8\u7406\u4efb\u52a1\uff0c\u4ee5\u6269\u5145\u6570\u636e\u96c6\u3002\u6b64\u5916\uff0c\u4ed6\u4eec\u8fd8\u5bf9\u6570\u636e\u96c6\u8fdb\u884c\u683c\u5f0f\u7b5b\u9009\u7b49\u4e00\u7cfb\u5217\u64cd\u4f5c\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u7814\u7a76\u8005\u6392\u9664\u4e86\u90a3\u4e9b\u96be\u4ee5\u7528\u57fa\u4e8e\u89c4\u5219\u5956\u52b1\u51fd\u6570\u8fdb\u884c\u8bc4\u4f30\u7684\u95ee\u9898\uff0c\u4f8b\u5982\u9009\u62e9\u9898\u548c\u8bc1\u660e\u9898\uff0c\u4ee5\u786e\u4fdd\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5956\u52b1\u8ba1\u7b97\u7684\u51c6\u786e\u6027\u548c\u4e00\u81f4\u6027\u3002<\/p>\n<p><strong>\u8bad\u7ec3\u6d88\u878d<\/strong><\/p>\n<p>\u300aOpenReasonerZero\u300b\u662f\u53e6\u4e00\u7bc7\u53d1\u73b0 GRPO \u5bf9\u4ed6\u4eec\u4e0d\u8d77\u4f5c\u7528\u7684\u8bba\u6587\u3002\u4f5c\u8005\u4f7f\u7528\u4e86\u5e26 GAE\uff08Generalized Advantage Estimation\uff09\u7684 PPO \u7b97\u6cd5\u6765\u5bf9\u4e00\u7ec4\u54cd\u5e94\u8fdb\u884c\u4f30\u8ba1\uff0c\u8fd9\u4e5f\u8fdb\u4e00\u6b65\u8bc1\u5b9e\u4e86\u6587\u7ae0\u4e4b\u524d\u63d0\u5230\u7684 GRPO \u5e76\u975e\u6709\u7279\u522b\u4e4b\u5904\u3002<\/p>\n<p>\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u4ed6\u4eec\u6ca1\u6709\u4f7f\u7528\u4efb\u4f55\u590d\u6742\u7684\u957f\u5ea6\u6216\u683c\u5f0f token\uff08\u4f8b\u5982 &lt; answer&gt;token\uff09\u6765\u6784\u5efa\u5956\u52b1\u51fd\u6570\uff0c\u800c\u662f\u53d1\u73b0\u4ec5\u6b63\u786e\u6027\u662f\u5fc5\u8981\u7684\u3002\u6b64\u5916\uff0c\u4ed6\u4eec\u8fd8\u79fb\u9664\u4e86\u6240\u6709\u7684 KL \u60e9\u7f5a\uff0c\u8fd9\u5bf9\u4e8e\u5141\u8bb8\u6a21\u578b\u5728\u54cd\u5e94\u957f\u5ea6\u4e0a\u8fdb\u884c\u663e\u8457\u53d8\u5316\u4ee5\u53ca\u5b66\u4e60\u65b0\u884c\u4e3a\u81f3\u5173\u91cd\u8981\uff0c\u8fd9\u4e9b\u884c\u4e3a\u6709\u52a9\u4e8e\u4e0b\u6e38\u6027\u80fd\u7684\u63d0\u5347\u3002<\/p>\n<p>\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u79fb\u9664 KL \u635f\u5931\u548c KL \u60e9\u7f5a\u80fd\u591f\u5b9e\u73b0\u6700\u4f18\u7684\u8bad\u7ec3\u7a33\u5b9a\u6027\u548c\u6700\u7ec8\u6027\u80fd\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s2.51cto.com\/oss\/202505\/25\/19c6a0770b67ac41417026987d8ba5eea8e2a7.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u8fd9\u7bc7\u8bba\u6587\u7684\u9644\u5f55\u91cc\u8fd8\u6709\u66f4\u591a\u6709\u8da3\u7684\u6d88\u878d\u5b9e\u9a8c\uff0c\u6bd4\u5982\u8c03\u6574 RL \u8bad\u7ec3\u65f6\u7684\u91c7\u6837\u6e29\u5ea6\u3001\u4fee\u6539 RL \u8d85\u53c2\u6570\uff0c\u6216\u662f\u8c03\u6574\u6279\u6b21\u5927\u5c0f\u548c\u66f4\u65b0\u89c4\u5219\u7b49\uff0c\u611f\u5174\u8da3\u7684\u8bfb\u8005\u53ef\u4ee5\u67e5\u770b\u539f\u6587\u7ae0\u3002<\/p>\n<h4>DAPO\uff1a\u4e00\u4e2a\u5927\u89c4\u6a21\u5f00\u6e90 LLM \u5f3a\u5316\u5b66\u4e60\u7cfb\u7edf<\/h4>\n<p>\u5728\u5f00\u59cb\u8ba8\u8bba\u63a5\u4e0b\u6765\u7684\u4e24\u7bc7\u8bba\u6587\u524d\uff0c\u6211\u4eec\u9700\u8981\u4e86\u89e3\u4e00\u4e9b\u80cc\u666f\uff0c\u4ee5\u4fbf\u7406\u89e3\u673a\u5668\u5b66\u4e60\u7279\u522b\u662f\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u7b97\u6cd5\u8fdb\u6b65\u3002\u6709\u4e00\u4e2a\u7ecf\u9a8c\u6cd5\u5219\u662f\uff0c\u5982\u679c\u4f60\u770b\u5230\u4e00\u7bc7\u8bba\u6587\u4e2d\u63d0\u51fa\u7684\u65b9\u6cd5\u6ca1\u6709\u63d0\u9ad8\u5230\u57fa\u51c6\u65b9\u6cd5\u7684 2 \u500d\u5de6\u53f3\uff0c\u90a3\u4e48\u8fd9\u4e2a\u89e3\u51b3\u65b9\u6848\u7684\u6210\u529f\u5f88\u53ef\u80fd\u4e3b\u8981\u5f52\u529f\u4e8e\u8d85\u53c2\u6570\u8c03\u6574\u6216\u5176\u4ed6\u6df7\u6dc6\u53d8\u91cf\u3002\u8fd9\u662f\u5bf9\u8bed\u8a00\u6a21\u578b\u65b0\u578b\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u5e94\u6301\u6709\u7684\u9002\u5f53\u6000\u7591\u6001\u5ea6\u3002<\/p>\n<p>\u8bfb\u8fd9\u4e9b\u8bba\u6587\u65f6\uff0c\u4f60\u5f88\u5bb9\u6613\u60f3\u5230\u300c\u54c7\uff0c\u6211\u7684\u9879\u76ee\u73b0\u5728\u5c31\u80fd\u987a\u5229\u8fd0\u884c\u4e86\u300d\u3002\u5b9e\u9645\u60c5\u51b5\u8fdc\u975e\u5982\u6b64\u3002\u8fd9\u4e9b\u8bba\u6587\u662f\u5b66\u4e60 GRPO \u635f\u5931\u51fd\u6570\u590d\u6742\u7ec6\u8282\u7684\u7edd\u4f73\u7ec3\u4e60\u3002\u591a\u5e74\u6765\uff0c\u8fd9\u7c7b\u8bba\u6587\u79ef\u7d2f\u8d77\u6765\u4f1a\u5e26\u6765\u5de8\u5927\u7684\u76f4\u89c9\u589e\u76ca\u3002\u4f46\u76ee\u524d\u5728\u8bad\u7ec3\u771f\u6b63\u7684 SOTA \u6a21\u578b\u65f6\uff0c\u5927\u591a\u6570\u6539\u53d8\u5728\u4ee3\u7801\u590d\u6742\u6027\u4e0a\u53ef\u80fd\u4f1a\u8fc7\u4e8e\u7e41\u91cd\uff0c\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u4e13\u6ce8\u4e8e\u8c03\u6574\u6570\u636e\u5206\u5e03\uff08\u5982\u4e0a\u8ff0\u8bba\u6587\u6240\u8ba8\u8bba\u7684\uff09\u66f4\u4e3a\u91cd\u8981\u3002<\/p>\n<p>\u8ba9\u6211\u4eec\u6765\u770b\u770b DAPO\u2014\u2014 \u8fd9\u662f\u5bf9\u4e4b\u524d Twitter \u4e0a\u7b80\u77ed\u4ecb\u7ecd\u7684\u6269\u5c55\u7248\u672c\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s7.51cto.com\/oss\/202505\/25\/0677e5055ba24b9d862879a8534fef6e1c34f9.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<ul data-id=\"ua73dd4b-EU7aXkIZ\">\n<li data-id=\"ld70c578-Yb9WdjTZ\">\u8bba\u6587\u6807\u9898\uff1aDAPO: An Open-Source LLM Reinforcement Learning System at Scale<\/li>\n<li data-id=\"ld70c578-KI5uBNwL\">\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2503.14476<\/li>\n<\/ul>\n<p>\u4ed6\u4eec\u5c55\u793a\u7684\u5b66\u4e60\u66f2\u7ebf\u4e0d\u9519\uff0c\u4f46\u6709\u4e9b\u6df7\u4e71\uff0c\u56e0\u4e3a\u300cDeepSeek R1 Zero Qwen 32B\u300d\u6a21\u578b\u7684\u8bad\u7ec3\u6b65\u9aa4\u7cbe\u786e\u6bd4\u8f83\u5b9e\u9645\u4e0a\u5e76\u4e0d\u5b58\u5728\uff08\u5982\u4e0a\u6240\u8ff0\uff09\uff0c\u5c24\u5176\u662f\uff0cx \u8f74\u662f\u6709\u8bef\u5bfc\u6027\u7684\u3002\u518d\u5f3a\u8c03\u4e00\u904d\uff0cDeepSeek \u6240\u505a\u7684\u5de5\u4f5c\u5e76\u975e\u4e0d\u53ef\u590d\u5236\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s9.51cto.com\/oss\/202505\/25\/012a57e421e860e6cb0236e2fbce5bdf1708a7.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u8fd9\u662f\u4e00\u7bc7\u975e\u5e38\u6574\u6d01\u7684\u5173\u4e8e\u63a8\u7406\u7684\u5f3a\u5316\u5b66\u4e60\u8bba\u6587\u3002\u6211\u4eec\u5c06\u4ecb\u7ecd\u7684 GRPO \u6539\u8fdb\u5305\u62ec\uff1a<\/p>\n<p>1. \u4e24\u4e2a\u4e0d\u540c\u7684\u88c1\u526a\u8d85\u53c2\u6570\uff0c\u4f7f\u6b63\u5411\u88c1\u526a\u80fd\u591f\u66f4\u591a\u5730\u63d0\u5347\u610f\u5916\u7684 token\u3002<\/p>\n<p>2. \u52a8\u6001\u91c7\u6837 \u2014\u2014 \u4ece\u6279\u6b21\u4e2d\u79fb\u9664\u5177\u6709\u5e73\u5766\u5956\u52b1\u7684\u6837\u672c\u4ee5\u63d0\u9ad8\u6548\u7387\u3002<\/p>\n<p>3. \u4f7f\u7528\u6bcf\u4e2a token \u7684\u635f\u5931\uff08\u800c\u975e\u6bcf\u4e2a\u56de\u5e94\uff08per-response\uff09\u7684\u635f\u5931\uff09\u6765\u6539\u5584\u5b66\u4e60\u52a8\u6001\u3002<\/p>\n<p>4. \u5728\u635f\u5931\u51fd\u6570\u4e2d\u7ba1\u7406\u8fc7\u957f\u751f\u6210\u4ee5\u83b7\u5f97\u66f4\u597d\u7684\u7a33\u5b9a\u6027\u3002<\/p>\n<p>\u6211\u5e0c\u671b\u8fd9\u7bc7\u8bba\u6587\uff0c\u6b63\u5982\u6211\u5c06\u5728\u4e0b\u9762\u8ba8\u8bba Dr. GRPO \u8bba\u6587\u65f6\u63d0\u5230\u7684\uff0c\u80fd\u505a\u66f4\u591a\u5173\u4e8e\u6700\u7ec8\u6027\u80fd\u7684\u6bd4\u8f83\u3002\u6211\u4eec\u5173\u5fc3\u7684\u662f\u8bc4\u4f30\u7ed3\u679c\uff0c\u6240\u4ee5\u5728\u7b97\u6cd5\u53d8\u5316\u5e26\u6765\u6027\u80fd\u63d0\u5347\u4e4b\u524d\uff0c\u6211\u5f88\u96be\u8bf4\u8fd9\u4e9b\u662f\u5173\u952e\u7684\u5b9e\u73b0\u51b3\u7b56\u3002<\/p>\n<p>\u603b\u4e4b\uff0cDAPO \u770b\u8d77\u6765\u5982\u4e0b\uff1a<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s8.51cto.com\/oss\/202505\/25\/138666f9609a78726f64924809d0b849358a66.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u50cf\u73b0\u5728\u7684\u8bb8\u591a\u8bba\u6587\u4e00\u6837\uff0c\u4ed6\u4eec\u4e5f\u5efa\u8bae\u4ece GRPO \u4e2d\u79fb\u9664 KL \u6563\u5ea6\u60e9\u7f5a\u4ee5\u5e2e\u52a9\u5b66\u4e60\u3002\u8bb8\u591a\u4eba\u8868\u793a\uff0c\u5982\u679c\u6ca1\u6709\u8981\u8fc7\u5ea6\u4f18\u5316\u7684\u5956\u52b1\u6a21\u578b\uff0c\u8fd9\u4e2a\u60e9\u7f5a\u5c31\u4e0d\u662f\u5fc5\u8981\u7684\u3002\u5bf9\u4e8e\u57fa\u7840\u6a21\u578b\u7684\u5f3a\u5316\u5b66\u4e60\uff0c\u6211\u540c\u610f\u8fd9\u4e00\u70b9\uff0c\u56e0\u4e3a\u6a21\u578b\u901a\u5e38\u9700\u8981\u66f4\u5927\u7684\u53d8\u5316\u624d\u80fd\u6210\u4e3a\u5b8c\u6574\u7684\u63a8\u7406\u6a21\u578b\u3002\u4f46\u5982\u679c\u5bf9\u6307\u4ee4\u6a21\u578b\u8fdb\u884c\u53ef\u9a8c\u8bc1\u5956\u52b1\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLVR\uff09\uff0cKL \u60e9\u7f5a\u53ef\u80fd\u4ecd\u7136\u6709\u7528\u3002<\/p>\n<p><strong>GRPO \u6539\u8fdb\u70b9 1\uff1a\u66f4\u9ad8\u7684\u88c1\u526a \/ \u5206\u79bb\u88c1\u526a\u8d85\u53c2\u6570\u00a0<\/strong><\/p>\n<p>PPO \u548c GRPO \u6709\u4e00\u4e2a\u63a7\u5236\u66f4\u65b0\u6b65\u957f\u7684\u88c1\u526a\u8d85\u53c2\u6570\u3002\u8fd9\u662f PPO \u548c TRPO \u76f8\u6bd4 REINFORCE \u6216\u666e\u901a\u7b56\u7565\u68af\u5ea6\u7684\u6838\u5fc3\u601d\u60f3\u3002DAPO \u5c06\u5176\u6539\u4e3a\u4e24\u4e2a\u8d85\u53c2\u6570\uff0c\u8fd9\u6837\u4e0a\u9650 \/ \u6b63\u5411\u5bf9\u6570\u6bd4\u7387\u6b65\u957f\u53ef\u4ee5\u66f4\u5927\u3002\u8fd9\u662f\u4e3a\u4e86\u589e\u52a0 token \u7684\u6982\u7387\uff0c\u6bd4\u5982\u63a8\u7406\u94fe\u4e2d\u4ee4\u4eba\u60ca\u8bb6\u7684\u65b0 token\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s2.51cto.com\/oss\/202505\/25\/c4a9f923919a1a90cc7519efccbdf9dbd2eec5.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>PPO\/GRPO \u66f4\u65b0\u662f\u57fa\u4e8e\u5bf9\u6570\u6bd4\u7387\u7684\uff0c\u6240\u4ee5\u6982\u7387\u8f83\u5c0f\u4f46\u6b63\u5728\u589e\u52a0\u53ef\u80fd\u6027\u7684\u4e0d\u592a\u53ef\u80fd\u7684 token \u4f1a\u53d8\u6210\u66f4\u5927\u7684\u5bf9\u6570\u6bd4\u7387\uff0c\u6bd4\u53ef\u80fd\u6027\u5df2\u7ecf\u5f88\u9ad8\u7684 token \u66f4\u5bb9\u6613\u88ab\u88c1\u526a\u3002\u8fd9\u5bf9\u4e8e\u63d0\u5347\u6548\u679c\u6765\u8bf4\u662f\u4e00\u4e2a\u76f8\u5f53\u590d\u6742\u7684\u53d8\u5316\uff0c\u4f46\u5b83\u5f88\u597d\u5730\u8bf4\u660e\u4e86\u88c1\u526a\u5982\u4f55\u5f71\u54cd\u5b66\u4e60\u52a8\u6001\u3002<\/p>\n<p>\u8fd9\u7bc7\u8bba\u6587\u5bf9\u4ed6\u4eec\u7684\u6539\u8fdb\u505a\u4e86\u5f88\u597d\u7684\u6d88\u878d\u5b9e\u9a8c\uff01\u5982\u4e0b\u56fe\uff08\u53f3\uff09\u6240\u793a\uff0c\u4ed6\u4eec\u663e\u793a\u6a21\u578b\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4fdd\u6301\u4e86\u66f4\u9ad8\u7684\u71b5\uff08\u5373\u63a2\u7d22 \/ \u968f\u673a\u6027\uff09\u3002\u4e0d\u9519\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s5.51cto.com\/oss\/202505\/25\/a7a7cd175618dfda7b236522eb9fb15134450b.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p><strong>GRPO \u6539\u8fdb\u70b9 2\uff1a\u4ece\u6279\u6b21\u4e2d\u79fb\u9664\u4e0d\u5fc5\u8981\u7684\u6837\u672c\u00a0<\/strong><\/p>\n<p>\u672c\u8d28\u4e0a\uff0c\u5728 GRPO \u4e2d\uff0c\u5982\u679c\u6279\u6b21\u4e2d\u9488\u5bf9\u4e00\u4e2a\u63d0\u793a\u7684\u6240\u6709\u6837\u672c\u5177\u6709\u76f8\u540c\u7684\u5956\u52b1\uff0c\u5219\u6ca1\u6709\u5b66\u4e60\u4fe1\u53f7\uff0c\u56e0\u4e3a\u6bcf\u4e2a\u7b54\u6848\u7684\u4f18\u52bf\u662f\u8ba1\u7b97\u4e3a\u8be5\u7b54\u6848\u4e0e\u6279\u6b21\u4e2d\u7ec4\u5e73\u5747\u503c\u7684\u5dee\u5f02\u3002\u79fb\u9664\u5b83\u4eec\u53ef\u4ee5\u901a\u8fc7\u8ba1\u7b97\u66f4\u5c11\u7684\u68af\u5ea6\u6765\u63d0\u9ad8\u5b66\u4e60\u901f\u5ea6\u3002<\/p>\n<p>\u8fd9\u5b9e\u9645\u4e0a\u662f GRPO \u7684\u4e00\u4e2a\u7b80\u5355\u4e8b\u5b9e\u3002\u4ece\u7406\u8bba\u4e0a\u8bb2\uff0c\u5f53\u6279\u6b21\u4e2d\u7684\u7b54\u6848\u6ca1\u6709\u4fe1\u53f7\u65f6\uff0c\u5b83\u4eec\u4e0d\u4f1a\u5f71\u54cd\u6a21\u578b\uff0c\u4f46\u8fd9\u4e5f\u4e0e\u4e3a\u4ec0\u4e48\u66f4\u5927\u7684\u6a21\u578b\u53ef\u80fd\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u5b66\u4e60\u5f97\u66f4\u597d\u6709\u5173\u3002\u66f4\u5927\u6a21\u578b\u7684\u5f3a\u5316\u5b66\u4e60\u6b65\u9aa4\u4e0d\u592a\u53ef\u80fd\u65e0\u610f\u4e2d\u4f24\u5bb3\u6a21\u578b\u4e2d\u4e0d\u5728\u5b66\u4e60\u6279\u6b21\u4e2d\u7684\u5176\u4ed6\u533a\u57df\uff0c\u56e0\u4e3a\u5b83\u4eec\u7684\u80fd\u529b\u5206\u5e03\u5728\u66f4\u591a\u53c2\u6570\u4e0a\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s3.51cto.com\/oss\/202505\/25\/968634a40982a445ee7109dbd3b60fb256fe7d.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p><strong>GRPO \u6539\u8fdb\u70b9 3\uff1atoken \u7ea7\u7b56\u7565\u68af\u5ea6<\/strong><\/p>\n<p>\u8bba\u6587\u4f5c\u8005\u8868\u793a\uff0ctoken \u7ea7\u635f\u5931\u6709\u52a9\u4e8e\u51cf\u8f7b\u975e\u5e38\u957f\u7684\u63a8\u7406\u94fe\u4e2d\u7684\u91cd\u590d\u884c\u4e3a\uff0c\u540c\u65f6\u4ecd\u7136\u9f13\u52b1\u6a21\u578b\u4ece\u6b63\u5411\u7684\u957f\u4e0a\u4e0b\u6587\u793a\u4f8b\u4e2d\u5b66\u4e60\u3002\u8fd9\u4e0e\u6807\u51c6 RLHF \u5b9e\u73b0\u4e0d\u540c\uff08\u89c1\u4e0b\u6587\u5173\u4e8e Dr. GRPO \u7684\u8ba8\u8bba\uff09\u3002\u4ed6\u4eec\u7684\u6539\u53d8\u6bd4\u6211\u4eec\u7a0d\u540e\u8ba8\u8bba\u7684\u66f4\u4e3a\u6e29\u548c\u3002<\/p>\n<p>\u4f5c\u8005\u7684\u76f4\u89c9\u5f88\u597d\uff0c\u6211\u4eec\u5c06\u5728\u63a5\u4e0b\u6765\u7684\u8bba\u6587\u4e2d\u770b\u5230\u5f88\u591a\u5173\u4e8e\u8fd9\u79cd\u6743\u8861\u7684\u8ba8\u8bba\uff1a<\/p>\n<blockquote><p>\u7531\u4e8e\u6240\u6709\u6837\u672c\u5728\u635f\u5931\u8ba1\u7b97\u4e2d\u88ab\u8d4b\u4e88\u76f8\u540c\u7684\u6743\u91cd\uff0c\u8f83\u957f\u56de\u5e94\u4e2d\u7684 token\uff08\u5305\u542b\u66f4\u591a token\uff09\u5bf9\u6574\u4f53\u635f\u5931\u7684\u8d21\u732e\u53ef\u80fd\u4e0d\u6210\u6bd4\u4f8b\u5730\u4f4e\uff0c\u8fd9\u53ef\u80fd\u5bfc\u81f4\u4e24\u79cd\u4e0d\u826f\u5f71\u54cd\u3002\u9996\u5148\uff0c\u5bf9\u4e8e\u9ad8\u8d28\u91cf\u7684\u957f\u6837\u672c\uff0c\u8fd9\u79cd\u6548\u679c\u53ef\u80fd\u963b\u788d\u6a21\u578b\u5b66\u4e60\u5176\u4e2d\u4e0e\u63a8\u7406\u76f8\u5173\u7684\u6a21\u5f0f\u7684\u80fd\u529b\u3002\u5176\u6b21\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\u8fc7\u957f\u7684\u6837\u672c\u901a\u5e38\u8868\u73b0\u51fa\u4f4e\u8d28\u91cf\u7684\u6a21\u5f0f\uff0c\u5982\u80e1\u8a00\u4e71\u8bed\u548c\u91cd\u590d\u8bcd\u6c47\u3002\u56e0\u6b64\uff0c\u6837\u672c\u7ea7\u635f\u5931\u8ba1\u7b97\u7531\u4e8e\u65e0\u6cd5\u6709\u6548\u60e9\u7f5a\u957f\u6837\u672c\u4e2d\u90a3\u4e9b\u4e0d\u53d7\u6b22\u8fce\u7684\u6a21\u5f0f\uff0c\u5bfc\u81f4\u71b5\u548c\u56de\u5e94\u957f\u5ea6\u7684\u4e0d\u5065\u5eb7\u589e\u52a0\u3002<\/p><\/blockquote>\n<p>\u5c06\u957f\u5ea6\u5f52\u4e00\u5316 1\/|o | \u79fb\u5230\u7ec4\u603b\u548c\u4e4b\u5916\uff0c\u4f7f\u5f97\u635f\u5931\u8ba1\u7b97\u4ec5\u901a\u8fc7\u5bf9\u4e24\u4e2a\u603b\u548c\u5185\u90e8\u7684 token \u6c42\u548c\u6765\u5b8c\u6210\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s4.51cto.com\/oss\/202505\/25\/d8c8b4f574ac0c6000e64166cf440ab88a4fa7.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u5728\u8fd9\u91cc\uff0c\u7b54\u6848\u7ec4\u662f\u6309\u7167\u8be5\u63d0\u793a\u7684\u603b\u4f53 token \u8ba1\u6570\u5f52\u4e00\u5316\u7684\u3002\u9ed8\u8ba4 GRPO \u53ea\u5bf9 token \u6240\u5bf9\u5e94\u7684\u56de\u5e94\u8fdb\u884c\u6bcf token \u635f\u5931\u5f52\u4e00\u5316\u3002<\/p>\n<p>\u8fd9\u5f88\u9177\uff0c\u4f46\u9700\u8981\u66f4\u591a\u7684\u590d\u5236\uff01\u89c1\u4e0b\u9762\u7684\u8ba8\u8bba\u3002\u5728\u4ed6\u4eec\u7684\u8bbe\u7f6e\u4e2d\uff0c\u4ed6\u4eec\u770b\u5230\u4e86\u76f8\u5f53\u4e0d\u540c\u7684\u884c\u4e3a\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s8.51cto.com\/oss\/202505\/25\/39d3155527f3929b30d199dbfce1ef5e7b0fd5.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u8fd9\u4e2a\u60f3\u6cd5\u662f\u4e3a\u4e86\u80fd\u66f4\u597d\u5730\u4ece\u957f\u7b54\u6848\u4e2d\u5b66\u4e60\u3002\u597d\u7684\u957f\u7b54\u6848\u5e94\u8be5\u5f97\u5230\u5145\u5206\u5956\u52b1\uff0c\u800c\u91cd\u590d\u7684\u3001\u7cdf\u7cd5\u7684\u957f\u7b54\u6848\u9700\u8981\u88ab\u60e9\u7f5a\u3002<\/p>\n<p><strong>GRPO \u6539\u8fdb\u70b9 4\uff1a\u907f\u514d\u622a\u65ad\u7684\u5956\u52b1\u5851\u9020<\/strong><\/p>\n<p>\u8fd9\u662f DAPO \u4e2d\u6700\u5fae\u5c0f\u7684\u6539\u53d8\u3002\u672c\u8d28\u4e0a\uff0c\u4ed6\u4eec\u6dfb\u52a0\u4e86\u4e00\u4e2a\u67d4\u6027\u7136\u540e\u662f\u4e00\u4e2a\u4e25\u683c\u7684\u60e9\u7f5a\uff0c\u5f53\u6a21\u578b\u751f\u6210\u957f\u5ea6\u8d85\u8fc7\u9650\u5236\u65f6\u3002\u5bf9\u4e8e\u6700\u5927\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e3a 16k token \u7684\u6a21\u578b\uff0c\u4ed6\u4eec\u5728 12k token \u5f00\u59cb\u5e94\u7528\u60e9\u7f5a\uff0c\u5e76\u7ebf\u6027\u589e\u52a0\u5230 16k\u3002\u8fd9\u79cd\u957f\u5ea6\u63a7\u5236\u673a\u5236\u611f\u89c9\u5c06\u4f1a\u8fc7\u65f6\uff0c\u6216\u8005\u5728\u672a\u6765\u53ea\u662f\u4e00\u4e2a\u975e\u5e38\u5c0f\u7684\u6280\u5de7\u3002<\/p>\n<p>\u53c2\u8003\u4e00\u4e0b\uff0c\u8bb8\u591a\u5f3a\u5316\u5b66\u4e60\u5b9e\u73b0\u5df2\u7ecf\u5305\u542b\u4e86\u4e00\u4e9b\u66f4\u5927\u7684\u60e9\u7f5a\uff0c\u5982\u679c\u6a21\u578b\u622a\u65ad\u81ea\u8eab\uff08\u5373\u4ece\u4e0d\u751f\u6210 EOS token \u6216\u7b54\u6848\uff09\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s8.51cto.com\/oss\/202505\/25\/c114a7663bbc943666c861317b7e27def4b7ba.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u4ed6\u4eec\u79f0\u4e4b\u4e3a\u89e3\u8026\u88c1\u526a\u548c\u52a8\u6001\u91c7\u6837\u7b56\u7565\u4f18\u5316\uff08DAPO\uff09\u7b97\u6cd5\u3002\u79f0\u5176\u4e3a\u65b0\u7b97\u6cd5\u4f3c\u4e4e\u8db3\u591f\u516c\u5e73\uff0c\u8fd9\u5728\u5f3a\u5316\u5b66\u4e60\u4e2d\u4e00\u76f4\u5982\u6b64\uff0c\u4f46\u5b9e\u9645\u4e0a\u8fd9\u57fa\u672c\u4e0a\u662f GRPO++\u3002<\/p>\n<p>\u8bad\u7ec3\u4e13\u6ce8\u4e8e AIME\uff0c\u6240\u4ee5\u6700\u7ec8\u6a21\u578b\u5e76\u4e0d\u8d85\u7ea7\u6709\u8da3\uff0c\u4f46\u5b83\u4eec\u9648\u8ff0\u4e86\u4e00\u4e2a\u6c38\u6052\u771f\u7406\uff1a \u5373\u4f7f\u662f\u521d\u59cb\u6761\u4ef6\u7684\u5fae\u5c0f\u53d8\u5316\uff0c\u5982\u6570\u636e\u548c\u8d85\u53c2\u6570\u7684\u53d8\u5316\uff0c\u4e5f\u53ef\u4ee5\u901a\u8fc7\u8fed\u4ee3\u5f3a\u5316\u5b66\u4e60\u8fc7\u7a0b\u653e\u5927\uff0c\u4ea7\u751f\u5b9e\u8d28\u6027\u7684\u7ed3\u679c\u504f\u5dee\u3002<\/p>\n<h4>Dr. GRPO<\/h4>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s3.51cto.com\/oss\/202505\/25\/c91595193d406175ceb577e853c3e8a39a457f.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<ul data-id=\"ua73dd4b-IajN7l7A\">\n<li data-id=\"ld70c578-l5fo71er\">\u8bba\u6587\u6807\u9898\uff1aUnderstanding R1-Zero-Like Training: A Critical Perspective<\/li>\n<li data-id=\"ld70c578-Im8SOOXG\">\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2503.20783<\/li>\n<\/ul>\n<p>\u8fd9\u662f\u6211\u4eec\u83b7\u5f97\u7684\u7b2c\u4e8c\u7bc7\u5173\u4e8e\u4fee\u6539 GRPO \u4ee5\u4f7f\u5176\u66f4\u9002\u5408\u63a8\u7406\u8bad\u7ec3\uff08\u5b9e\u9645\u4e0a\uff0c\u5c31\u662f\u8ba9\u5b83\u66f4\u6709\u6548\uff09\u7684\u8bba\u6587\u3002\u8be5\u8bba\u6587\u8fd8\u5305\u62ec\u4e00\u4e9b\u4f18\u79c0\u7684\u5b9e\u9a8c\uff0c\u5e2e\u52a9\u7406\u89e3\u4e0d\u540c\u57fa\u7840\u6a21\u578b\u5982\u4f55\u5f71\u54cd\u5b66\u4e60\u5230\u7684\u63a8\u7406\u884c\u4e3a\u3002<\/p>\n<p>\u8bba\u6587\u7684\u6838\u5fc3\u56fe\u8868\u5982\u4e0b\uff1a<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s3.51cto.com\/oss\/202505\/25\/5377f03996494c71ec39977075c197b4389e31.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u8bba\u6587\u7684\u6838\u5fc3\u601d\u60f3\u662f\uff0c\u901a\u8fc7\u4fee\u6539 GRPO\uff0c\u4ed6\u4eec\u53ef\u4ee5\u6539\u5584\u5b66\u4e60\u52a8\u6001\uff0c\u4f7f\u5f97\u5728\u751f\u6210\u957f\u5ea6\u589e\u52a0\u8f83\u5c11\u7684\u60c5\u51b5\u4e0b\u5b9e\u73b0\u66f4\u5f3a\u7684\u6027\u80fd\u3002\u8fd9\u662f\u6bcf\u4e2a\u4eba\u90fd\u5e94\u8be5\u60f3\u8981\u7684\uff01<\/p>\n<p>\u4ed6\u4eec\u5047\u8bbe\u9ed8\u8ba4\u7684 GRPO \u5b9e\u73b0\u5b9e\u9645\u4e0a\u8bbe\u7f6e\u4e86\u4e00\u4e2a\u504f\u7f6e\uff0c\u4f7f\u751f\u6210\u957f\u5ea6\u589e\u52a0\u7684\u7a0b\u5ea6\u8d85\u8fc7\u4e86\u5b9e\u9645\u6709\u7528\u7684\u8303\u56f4\u3002<\/p>\n<p><strong>\u5173\u4e8e GRPO \u7684\u4fee\u6539<\/strong><\/p>\n<p>\u4ed6\u4eec\u63d0\u51fa\u7684\u6838\u5fc3\u4fee\u6539\u6709\u4e9b\u5fae\u5999\uff0c\u4e0e GRPO \u5b9e\u73b0\u7684\u5e38\u89c1\u505a\u6cd5\u5bc6\u5207\u76f8\u5173\u3002GRPO \u5b9e\u73b0\u7684\u4e00\u4e2a\u6838\u5fc3\u6b65\u9aa4\u5728 DeepSeekMath \u8bba\u6587\u4e2d\u6709\u8be6\u7ec6\u8bf4\u660e\uff1a<\/p>\n<p>\u300c\u7ed3\u679c\u76d1\u7763\u5728\u6bcf\u4e2a\u8f93\u51fa\ud835\udc5c_\ud835\udc56\u7684\u672b\u5c3e\u63d0\u4f9b\u5f52\u4e00\u5316\u7684\u5956\u52b1\uff0c\u5e76\u5c06\u8f93\u51fa\u4e2d\u6240\u6709 token \u7684\u4f18\u52bf\ud835\udc34\u02c6_(i,t)\u8bbe\u7f6e\u4e3a\u5f52\u4e00\u5316\u5956\u52b1&#8230;\u300d<\/p>\n<p>\u672c\u8d28\u4e0a\uff0c\u4e0d\u662f\u53ea\u5728\u9a8c\u8bc1\u7ed3\u679c\u7684\u6700\u7ec8 token \u4e0a\u5206\u914d\u4f18\u52bf\uff0c\u800c\u662f\u6279\u6b21\u4e2d\u7684\u6bcf\u4e2a token \u90fd\u88ab\u5206\u914d\u4e86\u4f18\u52bf\u3002\u7136\u540e\u4f7f\u7528\u8fd9\u4e9b\u7b56\u7565\u68af\u5ea6\u7b97\u6cd5\u8ba1\u7b97\u6bcf\u4e2a token \u7684\u635f\u5931\u3002<\/p>\n<p>\u8981\u4e86\u89e3\u8fd9\u662f\u5982\u4f55\u5de5\u4f5c\u7684\uff0c\u8ba9\u6211\u4eec\u91cd\u65b0\u56de\u987e\u8bba\u6587\u4e2d\u7684\u635f\u5931\u51fd\u6570\uff1a<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s9.51cto.com\/oss\/202505\/25\/53302de00bc52a907dc729febc46e35d980108.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u8fd9\u91cc\u53d1\u751f\u7684\u4e8b\u60c5\u662f\uff0c\u7b2c\u4e00\u4e2a\u6c42\u548c\u7ba1\u7406\u95ee\u9898\u7684\u56de\u5e94\u7ec4 G\uff0c\u5185\u90e8\u6c42\u548c\u7ba1\u7406\u6bcf\u4e2a token \u7684\u635f\u5931\u3002\u4f5c\u8005\u6b63\u5728\u7ea0\u6b63\u5b66\u4e60\u4e2d\u7684\u4e24\u79cd\u884c\u4e3a\uff1a<\/p>\n<ul data-id=\"ua73dd4b-YpTNXI1f\">\n<li data-id=\"ld70c578-seqsyW4t\">\u504f\u7231\u77ed\u800c\u6b63\u786e\u7684\u7b54\u6848\u3002\u5bf9\u4e8e\u6b63\u5411\u4f18\u52bf\uff0c\u5373\u4e00\u7ec4\u4e2d\u6b63\u786e\u4e14\u66f4\u597d\u7684\u56de\u5e94\uff0c\u539f\u59cb\u635f\u5931\u4f1a\u5956\u52b1\u77ed\u56de\u5e94\u3002\u8981\u7406\u89e3\u8fd9\u4e00\u70b9\uff0c\u8003\u8651\u4e00\u7ec4\u4e2d\u4e24\u4e2a\u90fd\u662f\u6b63\u786e\u7684\u7b54\u6848\uff0c\u4e00\u4e2a\u6709 10 \u4e2a token\uff0c\u4e00\u4e2a\u6709 1000 \u4e2a\u3002\u8fd9\u4e24\u4e2a\u90fd\u6709\u76f8\u540c\u7684\u4f18\u52bf\uff0c\u4f46\u8f83\u77ed\u7b54\u6848\u7684\u6982\u7387\u4f1a\u589e\u52a0\u66f4\u591a\uff0c\u56e0\u4e3a\u635f\u5931\u8fd1\u4f3c\u4e8e\u4f18\u52bf \/ \u957f\u5ea6\u3002\u5982\u679c\u8fd9\u4ecd\u7136\u4ee4\u4eba\u56f0\u60d1\uff0c\u53ef\u4ee5\u8003\u8651 EOS token\uff0c\u5b83\u51fa\u73b0\u5728\u6bcf\u4e2a\u56de\u5e94\u7684\u6700\u540e\u4e00\u4e2a token \u2014\u2014 \u5bf9\u4e8e\u77ed\u56de\u5e94\uff0c\u8fd9\u4e2a token \u5c06\u88ab\u589e\u52a0\u5f97\u66f4\u591a\u3002<\/li>\n<li data-id=\"ld70c578-eh6h7Nku\">\u4e0d\u60e9\u7f5a\u957f\u3001\u91cd\u590d\u3001\u9519\u8bef\u7684\u7b54\u6848\u3002\u8fd9\u4e00\u70b9\u66f4\u7b80\u5355\uff0c\u5982\u679c\u4f18\u52bf\u662f\u8d1f\u9762\u7684\uff0c\u5f52\u4e00\u5316\u56e0\u5b50 1\/|o | \u901a\u8fc7\u51cf\u5c11\u6bcf\u4e2a token \u4e0a\u4f18\u52bf\u7684\u6709\u6548\u5927\u5c0f\u6765\u51cf\u5c11\u5bf9\u8f83\u957f\u7b54\u6848\u7684\u5b66\u4e60\u4fe1\u53f7\u3002\u8f83\u5c0f\u7684\u60e9\u7f5a\u610f\u5473\u7740\u5b83\u6bd4\u9519\u8bef\u3001\u77ed\u7b54\u6848\u7684\u53ef\u80fd\u6027\u66f4\u5927\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u5408\u5728\u4e00\u8d77\uff0c\u6709\u70b9\u4e0e\u6211\u4eec\u60f3\u8981\u7684\u76f8\u53cd\uff08\u4e0e DAPO \u7684\u60f3\u6cd5\u975e\u5e38\u76f8\u5173\uff09\u3002\u6211\u4eec\u5e0c\u671b\u5728\u63a8\u7406\u65f6\u6709\u66f4\u957f\u7684\u6b63\u786e\u7b54\u6848\u4ee5\u63d0\u9ad8\u6269\u5c55\u6027\uff0c\u5e76\u4e14\u4e0d\u60f3\u6d6a\u8d39 token\u3002\u4e2a\u4eba\u800c\u8a00\uff0c\u6211\u66f4\u559c\u6b22 DAPO \u7684\u89e3\u51b3\u65b9\u6848\uff0c\u5c06\u957f\u5ea6\u5f52\u4e00\u5316\u79fb\u5230\u7ec4\u5916\uff0c\u800c\u4e0d\u662f\u5b8c\u5168\u53bb\u9664\u5b83\u3002<\/p>\n<p>\u4ed6\u4eec\u63d0\u51fa\u7684\u7b2c\u4e8c\u4e2a\u4fee\u6539\u975e\u5e38\u806a\u660e\uff08\u5df2\u5728 TRL \u4e2d\u5b9e\u73b0\uff09\u2014\u2014 \u79fb\u9664\u95ee\u9898\u7ea7\u96be\u5ea6\u504f\u7f6e\u3002\u5f53\u6267\u884c\u50cf GRPO \u8fd9\u6837\u7684\u66f4\u65b0\uff08\u4f8b\u5982\u4e5f\u7528 PPO\uff09\u65f6\uff0c\u4f18\u52bf\u7684\u5927\u5c0f\u5f71\u54cd\u68af\u5ea6\u66f4\u65b0\u7684\u5927\u5c0f\u3002\u5728\u8fd9\u91cc\uff0c\u76f8\u5bf9\u4e8e\u66f4\u5bb9\u6613\u89e3\u51b3\uff08\u6216\u5931\u8d25\uff09\u7684\u95ee\u9898\uff0c\u5177\u6709\u9ad8\u65b9\u5dee\u7684\u95ee\u9898\u4f1a\u53d7\u5230\u60e9\u7f5a \u2014\u2014 \u4ece\u76f4\u89c9\u4e0a\u8bb2\uff0c\u8fd9\u751a\u81f3\u53ef\u80fd\u4e0e\u6211\u4eec\u60f3\u8981\u7684\u76f8\u53cd\uff01\u8f83\u96be\u7684\u95ee\u9898\uff0c\u7279\u522b\u662f\u5728\u5b66\u4e60\u7684\u5173\u952e\u9636\u6bb5\uff0c\u5c06\u6709\u66f4\u9ad8\u7684\u65b9\u5dee\u3002\u4f5c\u8005\u4e5f\u89e3\u91ca\u4e86\u4e0e\u4e4b\u524d\u7684\u5f3a\u5316\u5b66\u4e60\u5b9e\u8df5\u7684\u5173\u7cfb\uff1a<\/p>\n<blockquote><p>\u867d\u7136\u4f18\u52bf\u5f52\u4e00\u5316\u5728\u5f3a\u5316\u5b66\u4e60\u4e2d\u662f\u4e00\u79cd\u5e38\u89c1\u6280\u5de7\uff0c\u4f46\u5b83\u901a\u5e38\u662f\u5728\u6574\u4e2a\u6279\u6b21\u4e2d\u8ba1\u7b97\u7684\u3002<\/p><\/blockquote>\n<p>\u5b9e\u9645\u4e0a\uff0c\u8fd9\u79cd\u53d8\u5316\u7684\u5f71\u54cd\u53ef\u4ee5\u5b8c\u5168\u88ab\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u5de5\u7a0b\u6240\u5438\u6536\uff0c\u6b63\u5982\u4e0a\u9762\u5176\u4ed6\u8bba\u6587\u4e2d\u8ba8\u8bba\u7684\u90a3\u6837\uff0c\u6279\u6b21\u4e2d\u95ee\u9898\u96be\u5ea6\u7684\u5206\u5e03\u662f\u5747\u5300\u7684\u3002<\/p>\n<p>\u4e0d\u8fc7\uff0c\u8fd9\u4e9b\u53d8\u5316\u4e5f\u6ca1\u6709\u514d\u8d39\u7684\u5348\u9910 \u2014\u2014 \u6211\u7684\u540c\u4e8b Costa Huang \u63d0\u9192\u6211\uff0c\u4f4e\u65b9\u5dee\u95ee\u9898\u53ef\u80fd\u662f\u6211\u4eec\u6a21\u578b\u5b66\u4e60\u7684\u5173\u952e\u3002\u5728\u6a21\u578b 9 \u6b21\u56de\u7b54\u9519\u8bef\u3001\u53ea\u6709 1 \u6b21\u6b63\u786e\u7684\u60c5\u51b5\u4e0b\uff0c\u79fb\u9664\u6807\u51c6\u5dee\u4f1a\u964d\u4f4e\u90a3\u4e00\u4e2a\u6b63\u786e\u7b54\u6848\u7684\u6743\u91cd\u3002\u8fd9\u53ef\u80fd\u662f\u6a21\u578b\u9700\u8981\u5b66\u4e60\u7684\u5173\u952e\uff01<\/p>\n<p>\u4ed6\u4eec\u5c06\u8fd9\u4e24\u4e2a\u53d8\u5316\u79f0\u4e3a\u300cGRPO Done Right\u300d\uff0c\u5373 Dr. GRPO\u3002\u5f53\u4ed6\u4eec\u5c06\u8fd9\u4e9b\u7ed3\u5408\u5728\u4e00\u8d77\u65f6\uff0c\u6a21\u578b\u663e\u793a\u4e86\u4ed6\u4eec\u9884\u671f\u7684\u8f93\u51fa\u957f\u5ea6\u53d8\u5316 \u2014\u2014 \u603b\u4f53\u4e0a\u66f4\u77ed\uff0c\u7279\u522b\u662f\u9519\u8bef\u7b54\u6848\u66f4\u77ed\u3002\u8fd9\u91cc\u7684\u7f3a\u70b9\u662f\uff0c\u4ed6\u4eec\u5b9e\u9645\u4e0a\u6ca1\u6709\u5c55\u793a\u66f4\u597d\u7684\u4e0b\u6e38\u6700\u7ec8\u6027\u80fd\u3002\u6700\u7ec8\u6027\u80fd\u662f\u76ee\u6807\uff0c\u5f88\u53ef\u80fd\u66f4\u591a\u5730\u6765\u81ea\u6570\u636e\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s6.51cto.com\/oss\/202505\/25\/f4b0b8d819011ac9c8e5246083a4d4912b2117.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p><strong>\u5173\u4e8e\u4e3a\u5f3a\u5316\u5b66\u4e60\u66f4\u6362\u57fa\u7840\u6a21\u578b<\/strong><\/p>\n<p>\u8bba\u6587\u4f5c\u8005\u6bd4\u8f83\u4e86\u6765\u81ea Qwen 2.5\u3001Llama 3.1 \u548c DeepSeek \u7684\u57fa\u7840\u6a21\u578b\uff0c\u770b\u5b83\u4eec\u5982\u4f55\u56de\u7b54 MATH \u95ee\u9898\u3002\u5bf9\u4e8e\u8fd9\u4e9b\u6a21\u578b\uff0c\u5b83\u4eec\u4f7f\u7528 R1 \u6a21\u677f\u3001Qwen MATH \u6a21\u677f\u548c\u65e0\u6a21\u677f\u3002\u5f53\u6a21\u578b\u5df2\u7ecf\u901a\u8fc7\u5e26\u6709\u63a8\u7406 \/ CoT \u8f68\u8ff9\u7684\u4e2d\u95f4\u8bad\u7ec3\u65f6\uff0c\u8fd9\u4e9b\u6a21\u677f\u8d77\u7740\u81f3\u5173\u91cd\u8981\u7684\u4f5c\u7528\uff1a<\/p>\n<blockquote><p>\u6a21\u677f 1\uff08R1 \u6a21\u677f\uff09\uff1aA conversation between User and Assistant. The User asks a question, and the Assistant solves it. The Assistant first thinks about the reasoning process in the mind and then provides the User with the answer.<\/p>\n<p>The reasoning process is enclosed within &lt;think&gt; &lt;\/think&gt; and answer is enclosed within &lt;answer&gt; &lt;\/answer&gt; tags, respectively, i.e., &lt;think&gt; reasoning process here &lt;\/think&gt; &lt;answer&gt; answer here &lt;\/answer&gt;.<\/p>\n<p>\u6a21\u677f 2\uff08Qwen-Math \u6a21\u677f\uff09\uff1a &lt;|im_start|&gt;system Please reason step by step, and put your final answer within \\boxed {}. &lt;|im_end|&gt; &lt;|im_start|&gt;user {question} &lt;|im_end|&gt; &lt;|im_start|&gt;assistant<\/p>\n<p>\u6a21\u677f 3\uff08\u65e0\u6a21\u677f\uff09\uff1a{question}<\/p><\/blockquote>\n<p>\u4ed6\u4eec\u53d1\u73b0 Llama \u548c DeepSeek \u4f7f\u7528 R1 \u6a21\u677f\u80fd\u591f\u6700\u597d\u5730\u9075\u5faa\u6307\u4ee4\uff0c\u800c Qwen \u5728\u6ca1\u6709\u6a21\u677f\u7684\u60c5\u51b5\u4e0b\u6548\u679c\u6700\u597d\u3002\u8bb0\u4f4f\u8fd9\u4e9b\u57fa\u7840\u6a21\u578b\u5728\u4e0d\u540c\u7684\u300c\u5fae\u8c03\u300d\u673a\u5236\u4e0b\u6709\u591a\u5927\u5dee\u5f02\u662f\u81f3\u5173\u91cd\u8981\u7684\u3002\u683c\u5f0f\u5408\u89c4\u6027\u7684\u6bd4\u8f83\u5982\u4e0b\u6240\u793a\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s4.51cto.com\/oss\/202505\/25\/647ecb7062de7778cdb11158169f6124987e8e.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u4e2d\u95f4\u56fe\u663e\u793a\u4e86\u6570\u5b66\u95ee\u9898\u7684 pass@8 \u7387\uff0c\u53ef\u4ee5\u7406\u89e3\u4e3a\u300c\u5982\u679c\u6211\u5bf9 1 \u4e2a\u95ee\u9898\u91c7\u6837 8 \u4e2a\u7b54\u6848\uff0c\u6b63\u786e\u7b54\u6848\u51fa\u73b0\u7684\u9891\u7387\u662f\u591a\u5c11\uff1f\u300d\u8fd9\u662f\u8861\u91cf\u6a21\u578b\u5728\u8bad\u7ec3\u5f00\u59cb\u65f6\u5b66\u4e60\u96be\u6613\u7a0b\u5ea6\u7684\u6307\u6807\u3002Qwen \u518d\u6b21\u8868\u73b0\u6700\u4f73\uff0c\u6e29\u5ea6\u7684\u5f71\u54cd\u6bd4\u6211\u60f3\u8c61\u7684\u8981\u5c0f\u3002<\/p>\n<p>\u6700\u53f3\u8fb9\u4ed6\u4eec\u663e\u793a\uff0c\u66f4\u5927\u7684 Qwen \u6a21\u578b\u5728\u4efb\u4f55\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u4e4b\u524d\u5c31\u5df2\u7ecf\u6709\u53cd\u601d\u884c\u4e3a\uff01\u5982\u679c\u4f60\u6b63\u5728\u4f7f\u7528\u8fd9\u4e9b\u6a21\u578b\uff0c\u8fd9\u5e76\u4e0d\u4ee4\u4eba\u60ca\u8bb6\uff0c\u4f46\u8fd9\u662f\u4e00\u4e2a\u5f88\u597d\u7684\u6570\u636e\u70b9\uff0c\u53ef\u4ee5\u6de1\u5316\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u4e2d\u300c\u554a\u54c8\u65f6\u523b\u300d\u7684\u4f5c\u7528\u3002\u8fd9\u4e9b\u6a21\u578b\u4e3b\u8981\u662f\u5728\u653e\u5927\uff0c\u800c\u4e0d\u662f\u5b66\u4e60\u65b0\u4e1c\u897f\u3002<\/p>\n<p>\u8bba\u6587\u8fd8\u6709\u5176\u4ed6\u4e0d\u9519\u7684\u7ed3\u679c\uff0c\u4f8b\u5982\u5728\u66f4\u591a\u9886\u57df\u7279\u5b9a\u7684\u6570\u5b66\u6570\u636e\u4e0a\u7ee7\u7eed\u9884\u8bad\u7ec3\u53ef\u4ee5\u63d0\u9ad8\u5f3a\u5316\u5b66\u4e60\u6027\u80fd\uff0c\u5f88\u591a\u4eba\u901a\u8fc7\u5728 Qwen-MATH \u6a21\u578b\u4e0a\u8bad\u7ec3\u7684\u5bb9\u6613\u7a0b\u5ea6\u4e5f\u80fd\u770b\u5230\u8fd9\u4e00\u70b9\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s4.51cto.com\/oss\/202505\/25\/c661c6232662df61c1a800e22a971eb937e60e.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p><strong>\u4e0e GRPO \u8bad\u7ec3\u5931\u8d25\u7684\u5173\u7cfb<\/strong><\/p>\n<p>\u5728\u6211\u4eec\u65e9\u671f\u7684\u7814\u7a76\u4e2d\u770b\u5230\u7684\u4e00\u4e9b\u5931\u8d25\uff0c\u6211\u4e5f\u4ece\u5176\u4ed6\u5b9e\u9a8c\u5ba4\u542c\u5230\u8fc7\uff0c\u662f GRPO \u53ef\u80fd\u4f1a\u5931\u8d25\uff0c\u5f00\u59cb\u751f\u6210\u975e\u5e38\u91cd\u590d\u548c\u957f\u7684\u56de\u5e94\u3002\u7b54\u6848\u7684\u5e8f\u5217\u957f\u5ea6\u81a8\u80c0\u5230\u8bad\u7ec3\u8bbe\u7f6e\u4e2d\u5141\u8bb8\u7684\u6700\u5927\u503c\uff0c\u4e0b\u6e38\u8bc4\u4f30\u7684\u6027\u80fd\u964d\u81f3 0\u3002\u8fd9\u4e9b\u6f5c\u5728\u7684\u5e72\u9884\u63aa\u65bd\uff0c\u5bf9\u4e8e Dr. GRPO \u548c DAPO \u6765\u8bf4\uff0c\u90fd\u6709\u52a9\u4e8e\u7f13\u89e3\u8fd9\u79cd\u60c5\u51b5\u3002\u6211\u4eec\u5f88\u5feb\u5c31\u4f1a\u6709\u72ec\u7acb\u7684\u590d\u5236\u5b9e\u9a8c\u4e86\uff01<\/p>\n<p><strong>\u4e0e\u73b0\u6709 RLHF \u57fa\u7840\u8bbe\u65bd\u7684\u5173\u7cfb<\/strong><\/p>\n<p>\u5728\u8fc7\u53bb\u51e0\u5468\u91cc\uff0c\u6211\u53c2\u4e0e\u4e86\u8bb8\u591a\u5173\u4e8e RLHF \u57fa\u7840\u8bbe\u65bd\u4e2d\u6bcf\u6279\u6b21\u4f7f\u7528\u603b\u548c\u635f\u5931\u8fd8\u662f\u5e73\u5747\u635f\u5931\u7684\u8ba8\u8bba\u3002\u5176\u4e2d\u7684\u6838\u5fc3\u95ee\u9898\u662f\uff1a\u5f3a\u5316\u5b66\u4e60\u5e94\u8be5\u5e73\u7b49\u5730\u5bf9\u5f85\u6bcf\u4e2a token\uff08\u5373\u4ece\u6279\u6b21\u4e2d\u8f83\u957f\u7684\u7b54\u6848\u5b66\u4e60\u66f4\u591a\uff09\u8fd8\u662f\u5e94\u8be5\u76f8\u5bf9\u4e8e\u95ee\u9898\u5bf9\u5b83\u4eec\u8fdb\u884c\u5f52\u4e00\u5316\uff1f\u6211\u7684\u57fa\u672c\u89c2\u70b9\u662f\uff0c\u57fa\u4e8e\u6bcf\u4e2a\u95ee\u9898\u7684\u5f52\u4e00\u5316\u66f4\u6709\u610f\u4e49\uff0c\u56e0\u4e3a\u6a21\u578b\u9700\u8981\u9488\u5bf9\u4e0d\u540c\u95ee\u9898\u5b66\u4e60\u4e0d\u540c\u7684\u884c\u4e3a\uff0c\u4f46\u8fd9\u79cd\u5b66\u4e60\u52a8\u6001\u5f88\u5fae\u5999\u3002<\/p>\n<p>\u8fd9\u7bc7\u8bba\u6587\u5bf9\u76ee\u524d\u6240\u6709\u6d41\u884c\u7684\u5f00\u6e90\u5f3a\u5316\u5b66\u4e60\u5de5\u5177\u63d0\u51fa\u4e86\u6279\u8bc4\uff0c\u79f0 per-response \u7684\u65b9\u5f0f\u504f\u5411\u4e8e\u66f4\u7cdf\u7cd5\u7684\u60c5\u51b5\u3002\u5b9e\u9645\u60c5\u51b5\u6bd4\u8fd9\u590d\u6742\u5f97\u591a \u2014\u2014per-response \u6c42\u548c\u80af\u5b9a\u662f\u6807\u51c6\u505a\u6cd5\u3002<\/p>\n<p>\u533a\u522b\u5728\u4e8e\u4ece\u50cf TRL \u8fd9\u6837\u7684\u5e93\u4e2d\u7684 masked_mean \u53d8\u4e3a\u4f5c\u8005\u4f7f\u7528\u7684 masked_sum\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s2.51cto.com\/oss\/202505\/25\/166a00445980f59b941836e643cbcdd4144d7b.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u4e3a\u4ec0\u4e48 per-response \u53ef\u80fd\u662f\u597d\u7684\u4e00\u4e2a\u76f4\u89c9\u662f\uff0c\u5982\u679c\u4f60\u6709\u4e00\u4e2a KL \u60e9\u7f5a\uff0c\u4f60\u4e0d\u5e0c\u671b\u4e00\u4e2a\u975e\u5e38\u5947\u602a\u7684 token\uff08KL \u7206\u70b8\uff09\u5f71\u54cd\u6279\u6b21\u4e2d\u7684\u6bcf\u4e2a token\u3002\u8fd9\u4e9b\u5b9e\u73b0\u5dee\u5f02\u975e\u5e38\u4f9d\u8d56\u4e8e\u7279\u5b9a\u7684\u8bad\u7ec3\u8bbe\u7f6e\u3002<\/p>\n<p>\u4e0d\u8fc7\uff0c\u5173\u4e8e\u63a8\u7406 vs.RLHF \u7684\u4e0a\u4e0b\u6587\u4e5f\u5f88\u9177\u3002\u662f\u7684\uff0c\u5728\u5927\u591a\u6570\u65b9\u9762\uff0c\u5b9e\u73b0\u90fd\u5982\u4f5c\u8005\u6240\u8bf4\uff0c\u4f46\u8fd9\u53ea\u662f\u56e0\u4e3a\u793e\u533a\u4e4b\u524d\u4e0d\u50cf\u6211\u4eec\u73b0\u5728\u5bf9\u63a8\u7406\u6a21\u578b\u90a3\u6837\u5173\u5fc3\u4e0a\u4e0b\u6587\u957f\u5ea6\u5b66\u4e60\u52a8\u6001\u3002<\/p>\n<p>\u4ee5\u524d\uff0c\u91cd\u70b9\u662f\u5728\u5956\u52b1\u4e0a\u3002\u73b0\u5728\uff0c\u91cd\u70b9\u662f\u5956\u52b1\u6b63\u786e\u7684\u957f\u4e0a\u4e0b\u6587\u884c\u4e3a\u5e76\u60e9\u7f5a\u91cd\u590d\u7684\u957f\u884c\u4e3a\u3002\u6240\u4ee5\uff0c\u662f\u7684\uff0c\u56de\u7b54\u503e\u5411\u4e8e\u53cd\u5bf9\u8f83\u957f\u7684\u3001\u9ad8\u5956\u52b1\u7684\u5e8f\u5217\uff0c\u4f46\u6211\u4eec\u4e0d\u5728\u4e4e\uff01<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s4.51cto.com\/oss\/202505\/25\/528faf662b9d960a0b5758454f5227022af319.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<p>\u5728\u539f\u6587\u7684\u300cFurther reading\u300d\u4f5c\u8005\u8fd8\u63a8\u8350\u4e86\u5176\u4ed6\u8bba\u6587\uff0c\u611f\u5174\u8da3\u7684\u8bfb\u8005\u53ef\u4ee5\u53c2\u8003\u4e0b\u56fe\u3002<\/p>\n<p><img decoding=\"async\" title=\"image.png\" src=\"https:\/\/s6.51cto.com\/oss\/202505\/25\/f72502183440b6094b2294f482d308b07a168f.webp\" alt=\"image.png\" data-type=\"inline\" \/><\/p>\n<\/div>\n<p>\u6587\u7ae0\u6765\u81ea\uff1a51CTO<\/p>\n<\/div>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_26025\" class=\"pvc_stats total_only  \" data-element-id=\"26025\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u6587\u8be6\u7ec6\u89e3\u8bfb\u4e86 Kimi k1.5\u3001OpenReasonerZero\u3001DAPO \u548c Dr. GRPO \u56db\u7bc7\u8bba\u6587 [&hellip;]<\/p>\n<div class=\"pvc_clear\"><\/div>\n<p id=\"pvc_stats_26025\" class=\"pvc_stats total_only  \" data-element-id=\"26025\" style=\"\"><i class=\"pvc-stats-icon medium\" aria-hidden=\"true\"><svg xmlns=\"http:\/\/www.w3.org\/2000\/svg\" version=\"1.0\" viewBox=\"0 0 502 315\" preserveAspectRatio=\"xMidYMid meet\"><g transform=\"translate(0,332) scale(0.1,-0.1)\" fill=\"\" stroke=\"none\"><path d=\"M2394 3279 l-29 -30 -3 -207 c-2 -182 0 -211 15 -242 39 -76 157 -76 196 0 15 31 17 60 15 243 l-3 209 -33 29 c-26 23 -41 29 -80 29 -41 0 -53 -5 -78 -31z\"\/><path d=\"M3085 3251 c-45 -19 -58 -50 -96 -229 -47 -217 -49 -260 -13 -295 52 -53 146 -42 177 20 16 31 87 366 87 410 0 70 -86 122 -155 94z\"\/><path d=\"M1751 3234 c-13 -9 -29 -31 -37 -50 -12 -29 -10 -49 21 -204 19 -94 39 -189 45 -210 14 -50 54 -80 110 -80 34 0 48 6 76 34 21 21 34 44 34 59 0 14 -18 113 -40 219 -37 178 -43 195 -70 221 -36 32 -101 37 -139 11z\"\/><path d=\"M1163 3073 c-36 -7 -73 -59 -73 -102 0 -56 133 -378 171 -413 34 -32 83 -37 129 -13 70 36 67 87 -16 290 -86 209 -89 214 -129 231 -35 14 -42 15 -82 7z\"\/><path d=\"M3689 3066 c-15 -9 -33 -30 -42 -48 -48 -103 -147 -355 -147 -375 0 -98 131 -148 192 -74 13 15 57 108 97 206 80 196 84 226 37 273 -30 30 -99 39 -137 18z\"\/><path d=\"M583 2784 c-38 -19 -67 -74 -58 -113 9 -42 211 -354 242 -373 16 -10 45 -18 66 -18 51 0 107 52 107 100 0 39 -1 41 -124 234 -80 126 -108 162 -133 173 -41 17 -61 16 -100 -3z\"\/><path d=\"M4250 2784 c-14 -9 -74 -91 -133 -183 -95 -150 -107 -173 -107 -213 0 -55 33 -94 87 -104 67 -13 90 8 211 198 130 202 137 225 78 284 -27 27 -42 34 -72 34 -22 0 -50 -8 -64 -16z\"\/><path d=\"M2275 2693 c-553 -48 -1095 -270 -1585 -649 -135 -104 -459 -423 -483 -476 -23 -49 -22 -139 2 -186 73 -142 361 -457 571 -626 285 -228 642 -407 990 -497 242 -63 336 -73 660 -74 310 0 370 5 595 52 535 111 1045 392 1455 803 122 121 250 273 275 326 19 41 19 137 0 174 -41 79 -309 363 -465 492 -447 370 -946 591 -1479 653 -113 14 -422 18 -536 8z m395 -428 c171 -34 330 -124 456 -258 112 -119 167 -219 211 -378 27 -96 24 -300 -5 -401 -72 -255 -236 -447 -474 -557 -132 -62 -201 -76 -368 -76 -167 0 -236 14 -368 76 -213 98 -373 271 -451 485 -162 444 86 934 547 1084 153 49 292 57 452 25z m909 -232 c222 -123 408 -262 593 -441 76 -74 138 -139 138 -144 0 -16 -233 -242 -330 -319 -155 -123 -309 -223 -461 -299 l-81 -41 32 46 c18 26 49 83 70 128 143 306 141 649 -6 957 -25 52 -61 116 -79 142 l-34 47 45 -20 c26 -10 76 -36 113 -56z m-2057 25 c-40 -58 -105 -190 -130 -263 -110 -324 -59 -707 132 -981 25 -35 42 -64 37 -64 -19 0 -241 119 -326 174 -188 122 -406 314 -532 468 l-58 71 108 103 c185 178 428 349 672 473 66 33 121 60 123 61 2 0 -10 -19 -26 -42z\"\/><path d=\"M2375 1950 c-198 -44 -350 -190 -395 -379 -18 -76 -8 -221 19 -290 114 -284 457 -406 731 -260 98 52 188 154 231 260 27 69 37 214 19 290 -38 163 -166 304 -326 360 -67 23 -215 33 -279 19z\"\/><\/g><\/svg><\/i> <img loading=\"lazy\" decoding=\"async\" width=\"16\" height=\"16\" alt=\"Loading\" src=\"https:\/\/aif.amtbbs.org\/wp-content\/plugins\/page-views-count\/ajax-loader-2x.gif\" border=0 \/><\/p>\n<div class=\"pvc_clear\"><\/div>\n","protected":false},"author":56,"featured_media":26027,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[23,17,20,80],"tags":[1712],"class_list":["post-26025","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-23","category-17","category-20","category-80","tag-deepseekgrpo"],"_links":{"self":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/26025","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/users\/56"}],"replies":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/comments?post=26025"}],"version-history":[{"count":1,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/26025\/revisions"}],"predecessor-version":[{"id":26028,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/posts\/26025\/revisions\/26028"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media\/26027"}],"wp:attachment":[{"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/media?parent=26025"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/categories?post=26025"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aif.amtbbs.org\/index.php\/wp-json\/wp\/v2\/tags?post=26025"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}