{"id":39605,"date":"2025-03-24T01:02:23","date_gmt":"2025-03-23T17:02:23","guid":{"rendered":"https:\/\/17aitech.com\/?p=39605"},"modified":"2025-03-24T01:02:23","modified_gmt":"2025-03-23T17:02:23","slug":"%e6%8f%ad%e7%a7%98deepseek-r1-zero%e8%ae%ad%e7%bb%83%e6%96%b9%e5%bc%8f%ef%bc%8cgrpo%e8%bf%98%e6%9c%89%e6%9e%81%e7%ae%80%e6%94%b9%e8%bf%9b%e6%96%b9%e6%a1%88","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=39605","title":{"rendered":"\u63ed\u79d8DeepSeek R1-Zero\u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO\u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-22-5\" target=\"_blank\">\u63ed\u79d8DeepSeek R1-Zero\u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO\u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848<\/a><\/p>\n<section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"29\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\u5176\u5b9e\u5927\u6a21\u578b\u5728DeepSeek-V3\u65f6\u671f\u5c31\u5df2\u7ecf\u300c\u987f\u609f\u300d\u4e86\uff1f<\/section>\n<\/section>\n<\/blockquote>\n<\/section>\n<p>DeepSeek-V3-Base \u5728\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u8c03\u4f18\u4e4b\u524d\u5c31\u5df2\u7ecf\u5c55\u73b0\u51fa\u300c\u987f\u609f\u65f6\u523b\u300d\uff1fRL \u8c03\u6574\u4e2d\u4e0d\u65ad\u589e\u52a0\u7684\u8f93\u51fa\u957f\u5ea6\u53ef\u80fd\u662f\u7531\u4e8e GRPO \u4e2d\u7684 BIAS \u9020\u6210\u7684\uff1f<\/p>\n<p>\u8fd1\u65e5\uff0c\u6765\u81ea Sea AI Lab\u3001\u65b0\u52a0\u5761\u56fd\u7acb\u5927\u5b66\u3001\u65b0\u52a0\u5761\u7ba1\u7406\u5927\u5b66\u7684\u7814\u7a76\u4eba\u5458\u7814\u7a76\u4e86\u9884\u8bad\u7ec3\u7279\u6027\u5f71\u54cd RL \u6027\u80fd\u7684\u539f\u7406\uff0c\u53d1\u8868\u4e86\u4e00\u7bc7\u540d\u4e3a\u300aUnderstanding R1-Zero-Like Training: A Critical Perspective\u300b\u7684\u7814\u7a76\u62a5\u544a\u3002<\/p>\n<p><strong>\u5206\u6790\u8868\u660e\uff0cDeepSeek-V3-Base \u5df2\u7ecf\u5c55\u73b0\u51fa\u300c\u987f\u609f\u65f6\u523b\u300d\uff0c\u800c Qwen2.5 \u57fa\u7840\u6a21\u578b\u5373\u4f7f\u6ca1\u6709\u63d0\u793a\u6a21\u677f\u4e5f\u8868\u73b0\u51fa\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\uff0c\u8fd9\u8868\u660e\u5b58\u5728\u6f5c\u5728\u7684\u9884\u8bad\u7ec3\u504f\u5dee\u3002<\/strong>\u6b64\u5916\uff0c\u4f5c\u8005\u8fd8\u5728\u7fa4\u4f53\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff08GRPO\uff09\u4e2d\u53d1\u73b0\u4e86\u4f18\u5316\u504f\u5dee\uff0c\u5b83\u5728\u8bad\u7ec3\u671f\u95f4\u4eba\u4e3a\u5730\u589e\u52a0\u4e86\u54cd\u5e94\u957f\u5ea6\uff08\u5c24\u5176\u662f\u5bf9\u4e8e\u9519\u8bef\u8f93\u51fa\uff09\u3002<\/p>\n<p>\u4e3a\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u7814\u7a76\u4eba\u5458\u5f15\u5165\u4e86 Dr. GRPO\uff0c\u8fd9\u662f\u4e00\u79cd\u65e0\u504f\u4f18\u5316\u65b9\u6cd5\uff0c\u53ef\u5728\u4fdd\u6301\u63a8\u7406\u6027\u80fd\u7684\u540c\u65f6\u63d0\u9ad8 token \u6548\u7387\u3002\u5229\u7528\u8fd9\u4e9b\u89c1\u89e3\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86\u4e00\u79cd\u7b80\u5316\u7684 R1-Zero \u65b9\u6848\uff0c\u4f7f\u7528 7B \u57fa\u7840\u6a21\u578b\u5728 AIME 2024 \u4e0a\u5b9e\u73b0\u4e86 43.3% \u7684\u51c6\u786e\u7387\uff0c\u5efa\u7acb\u4e86\u65b0\u7684 SOTA \u6c34\u51c6\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-809960926c9870a6df09192262a995cc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-809960926c9870a6df09192262a995cc.png\"><\/a><\/p>\n<p>\u9879\u76ee\u94fe\u63a5\uff1ahttps:\/\/github.com\/sail-sg\/understand-r1-zero<\/p>\n<p><strong>\u4ecb\u7ecd<\/strong><\/p>\n<p>\u6700\u8fd1\u4e00\u6bb5\u65f6\u95f4\uff0c\u4eba\u5de5\u667a\u80fd\u9886\u57df\u6b63\u5728\u8fce\u6765\u53d8\u9769\u3002DeepSeek-R1-Zero \u901a\u8fc7\u5f15\u5165\u7c7b\u4f3c R1-Zero \u7684\u8bad\u7ec3\u8303\u5f0f\u5f7b\u5e95\u6539\u53d8\u4e86\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM) \u540e\u8bad\u7ec3\u7684\u6d41\u7a0b\uff1a\u76f4\u63a5\u5c06 RL \u5e94\u7528\u4e8e\u57fa\u7840 LLM\uff0c\u800c\u4e0d\u4f9d\u8d56\u4efb\u4f55\u76d1\u7763\u5fae\u8c03\uff08SFT) \u4f5c\u4e3a\u521d\u6b65\u6b65\u9aa4\u3002<\/p>\n<p>\u8fd9\u79cd\u65b0\u8303\u5f0f\u56e0\u5176\u7b80\u5355\u6027\u548c\u5df2\u8bc1\u660e\u7684 RL \u6269\u5c55\u73b0\u8c61\u800c\u5177\u6709\u5438\u5f15\u529b\uff1a\u6a21\u578b\u63a8\u7406\u80fd\u529b\u968f\u7740\u6a21\u578b\u54cd\u5e94\u957f\u5ea6\u7684\u4e0d\u65ad\u589e\u52a0\u800c\u63d0\u9ad8\u3002\u8fd9\u79cd\u73b0\u8c61\u8fd8\u4f34\u968f\u7740\u300c\u987f\u609f\u65f6\u523b\u300d\uff08Aha Moment\uff09\uff0c\u6b64\u65f6\u6a21\u578b\u5c55\u73b0\u51fa\u4e86\u4eba\u7c7b\u4e00\u6837\u7684\u81ea\u6211\u53cd\u601d\u7b49\u65b0\u5174\u6280\u80fd\uff0c\u8ba9\u4eba\u4eec\u89c1\u8bc1\u4e86\u5f3a\u5316\u5b66\u4e60\u7684\u529b\u91cf\u548c\u7f8e\u611f\u3002<\/p>\n<p>\u5728\u672c\u6587\u4e2d\uff0c\u7814\u7a76\u4eba\u5458\u65e8\u5728\u901a\u8fc7\u7814\u7a76\u4e24\u4e2a\u57fa\u672c\u7ec4\u6210\u90e8\u5206\u6765\u7406\u89e3\u7c7b\u4f3c R1-Zero \u7684\u8bad\u7ec3\uff1a\u57fa\u7840\u6a21\u578b\u548c RL\u3002<\/p>\n<p>\u5728\u7b2c\u4e00\u90e8\u5206\u4e2d\uff0c\u4ed6\u4eec\u7814\u7a76\u4e86\u57fa\u7840\u6a21\u578b\u7684\u5404\u79cd\u5c5e\u6027\uff0c\u91cd\u70b9\u5173\u6ce8 Qwen2.5 \u6a21\u578b\u7cfb\u5217 \uff0c\u8be5\u6a21\u578b\u6700\u8fd1\u88ab\u7528\u4e8e\u91cd\u73b0 R1-Zero \uff0c\u4ee5\u53ca DeepSeek-V3-Base\uff0c\u771f\u6b63\u7684 R1-Zero \u6a21\u578b\u662f\u4ece\u8be5\u6a21\u578b\u4e2d\u7ecf\u8fc7 RL \u8c03\u6574\u7684\u3002<\/p>\n<p>\u5728\u7b2c\u4e8c\u90e8\u5206\u4e2d\uff0c\u4eba\u4eec\u786e\u5b9a\u4e86 GRPO \u4f18\u5316\u4e2d\u7684\u504f\u5dee\uff0c\u8fd9\u53ef\u80fd\u4f1a\u5bfc\u81f4\u9519\u8bef\u54cd\u5e94\u9010\u6e10\u53d8\u957f\u3002\u4e3a\u6b64\uff0c\u7814\u7a76\u4eba\u5458\u63d0\u51fa\u4e86\u4e00\u4e2a\u7b80\u5355\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u53ef\u4ee5\u5728\u4fdd\u6301\u63a8\u7406\u6027\u80fd\u7684\u540c\u65f6\u63d0\u9ad8 token \u6548\u7387\uff0c\u79f0\u4e3a Dr. GRPO\uff08GRPO Done Right\uff09\uff08\u56fe 1 \u4e2d\u663e\u793a\uff09\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-7e86f8edef74bab95f619b419f8416e9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-7e86f8edef74bab95f619b419f8416e9.png\"><\/a><\/p>\n<p>\u4e3a\u4e86\u7406\u89e3 R1-Zero \u4e0a\u7684\u8bad\u7ec3\u65b9\u5f0f\uff0c\u7814\u7a76\u4eba\u5458\u4ed4\u7ec6\u89c2\u5bdf\u4e86\u4e24\u4e2a\u6838\u5fc3\u90e8\u5206\uff1a\u57fa\u7840\u6a21\u578b\u548c\u5f3a\u5316\u5b66\u4e60\uff0c\u968f\u540e\u91cd\u70b9\u4ecb\u7ecd\u4e86\u5176\u53d1\u73b0\u3002<\/p>\n<p>\u4ed6\u4eec\u5bf9\u57fa\u7840\u6a21\u578b\u548c\u5f3a\u5316\u5b66\u4e60\u7684\u5206\u6790\u8868\u660e\uff0c<strong>\u5b9e\u73b0\u7c7b\u4f3c R1-Zero \u8bad\u7ec3\u53ef\u91c7\u7528\u6781\u7b80\u65b9\u6848\uff1a\u5229\u7528\u65e0\u504f\u7684 Dr. GRPO \u7b97\u6cd5\uff0c\u4ee5 Qwen-Math \u6a21\u677f\u5bf9 Qwen2.5-Math-7B \u6a21\u578b\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u8c03\u4f18\uff0c\u4f7f\u7528 MATH 3-5 \u7ea7\u95ee\u9898\u4f5c\u4e3a\u8bad\u7ec3\u6570\u636e\uff0c\u4ec5\u7528 8 \u5757 A100 GPU \u8bad\u7ec3 27 \u5c0f\u65f6\uff0c\u5c31\u5b9e\u73b0\u4e86\u6700\u5148\u8fdb\u7684\u6027\u80fd<\/strong>\uff08\u56fe 2 \u5c55\u793a\uff09\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-24d64ee7f112739c31ed3b4ec9a49835.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-24d64ee7f112739c31ed3b4ec9a49835.png\"><\/a><\/section>\n<p>\u8be5\u7814\u7a76\u7684\u4e3b\u8981\u89c2\u70b9\u603b\u7ed3\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>\n<p>\u6a21\u677f\u5bf9\u4e8e\u8ba9\u57fa\u7840\u6a21\u578b\u56de\u7b54\u95ee\u9898\uff08\u800c\u975e\u5b8c\u6210\u53e5\u5b50\uff09\u81f3\u5173\u91cd\u8981\u3002\u6b64\u5916\uff0c\u6240\u6709\u57fa\u7840\u6a21\u578b\u5728\u5f3a\u5316\u5b66\u4e60\u4e4b\u524d\u90fd\u5df2\u5177\u5907\u6570\u5b66\u89e3\u9898\u80fd\u529b\uff1b<\/p>\n<\/li>\n<li>\n<p>\u6709\u8da3\u7684\u662f\uff0cQwen-2.5 \u57fa\u7840\u6a21\u578b\u5728\u4e0d\u4f7f\u7528\u6a21\u677f\u65f6\u7acb\u5373\u83b7\u5f97\u7ea6 60% \u7684\u63d0\u5347\uff0c\u8fd9\u8ba9\u7814\u7a76\u4eba\u5458\u5047\u8bbe\u5b83\u4eec\u53ef\u80fd\u5728\u8bad\u7ec3\u6a21\u578b\u65f6\u5bf9\u62fc\u63a5\u7684\u95ee\u7b54\u6587\u672c\u8fdb\u884c\u4e86\u9884\u8bad\u7ec3\uff1b<\/p>\n<\/li>\n<li>\n<p>\u51e0\u4e4e\u6240\u6709\u57fa\u7840\u6a21\u578b\u90fd\u5df2\u8868\u73b0\u51fa\u300c\u987f\u609f\u65f6\u523b\u300d\uff0c\u5305\u62ec DeepSeek-V3-Base\uff1b<\/p>\n<\/li>\n<li>\n<p>\u65b0\u65b9\u6cd5 Dr. GRPO \u6709\u6548\u4fee\u590d\u4e86 GRPO \u5728\u4f18\u5316\u4e2d\u7684\u504f\u5dee\uff0c\u5b9e\u73b0\u4e86\u66f4\u597d\u7684\u6807\u8bb0\u6548\u7387\uff1b<\/p>\n<\/li>\n<li>\n<p>\u6a21\u578b\u4e0e\u6a21\u677f\u4e0d\u5339\u914d\u4f1a\u5728\u5f3a\u5316\u5b66\u4e60\u91cd\u5efa\u4e4b\u524d\u7834\u574f\u63a8\u7406\u80fd\u529b\uff1b<\/p>\n<\/li>\n<li>\n<p>\u5728 Llama-3.2-3B \u4e0a\u8fdb\u884c\u6570\u5b66\u9884\u8bad\u7ec3\u53ef\u63d0\u9ad8\u5176\u5f3a\u5316\u5b66\u4e60\u7684\u4e0a\u9650\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>\u57fa\u7840\u6a21\u578b\u5206\u6790<\/strong><\/p>\n<p>\u7814\u7a76\u4eba\u5458\u5bf9\u5e7f\u6cdb\u7684\u57fa\u7840\u6a21\u578b\u8fdb\u884c\u4e86\u5ba1\u67e5\uff0c\u5305\u62ec Qwen-2.5 \u7cfb\u5217\u3001Llama-3.1 \u548c DeepSeek \u7cfb\u5217\uff0c\u5411\u5b83\u4eec\u63d0\u51fa\u4ece MATH \u8bad\u7ec3\u96c6\u4e2d\u62bd\u53d6\u7684 500 \u4e2a\u95ee\u9898\uff0c\u5e76\u5206\u6790\u5b83\u4eec\u7684\u56de\u7b54\u3002<\/p>\n<p><strong>R1-Zero \u8bad\u7ec3\u80fd\u529b\uff1a\u6a21\u677f\u6784\u5efa\u63a2\u7d22\u6027\u57fa\u7840\u7b56\u7565<\/strong><\/p>\n<p>\u7531\u4e8e\u4ece\u57fa\u7840\u6a21\u578b\u8fdb\u884c\u8bad\u7ec3\u662f R1-Zero \u7c7b\u8303\u5f0f\u7684\u57fa\u672c\u8bbe\u7f6e\uff0c\u7814\u7a76\u4eba\u5458\u9996\u5148\u7814\u7a76\u5e7f\u6cdb\u4f7f\u7528\u7684\u5f00\u6e90\u57fa\u7840\u6a21\u578b\uff0c\u8fd9\u4e9b\u6a21\u578b\u901a\u5e38\u662f\u4e3a\u4e86\u53e5\u5b50\u8865\u5168\u800c\u8bad\u7ec3\u7684\u3002\u7814\u7a76\u4eba\u5458\u63a2\u7d22\u4e86\u662f\u5426\u53ef\u4ee5\u901a\u8fc7\u9002\u5f53\u7684\u6a21\u677f\u6709\u6548\u5730\u6fc0\u53d1\u5176\u95ee\u7b54\u80fd\u529b\uff0c\u4ece\u800c\u4f5c\u4e3a\u95ee\u7b54\u57fa\u7840\u7b56\u7565 \u3002<\/p>\n<p>\u5728\u5b9e\u9a8c\u8bbe\u7f6e\u4e2d\uff0c\u7814\u7a76\u4eba\u5458\u7eb3\u5165\u4e86 Qwen2.5-Math-1.5B\u3001Qwen2.5-Math-7B\u3001Qwen2.5-7B\u3001Llama-3.1-8B\u3001DeepSeek-Math-7B \u548c DeepSeek-V3-Base-685B \u8fdb\u884c\u5b9e\u9a8c\u3002<\/p>\n<p>\u5bf9\u4e8e\u6bcf\u4e2a\u6a21\u578b\uff0c\u4ed6\u4eec\u9996\u5148\u5e94\u7528\u65e0\u6a21\u677f\u6765\u83b7\u53d6\u6a21\u578b\u56de\u7b54\uff0c\u7136\u540e\u8ba9 GPT-4o-mini \u5224\u65ad\u6a21\u578b\u56de\u7b54\u662f\u5904\u4e8e\u95ee\u7b54\u683c\u5f0f\uff08\u65e0\u8bba\u8d28\u91cf\u5982\u4f55\uff09\u8fd8\u662f\u5904\u4e8e\u53e5\u5b50\u8865\u5168\u6a21\u5f0f\u3002\u4ed6\u4eec\u8bb0\u5f55\u503e\u5411\u4e8e\u56de\u7b54\u95ee\u9898\u7684\u56de\u7b54\u767e\u5206\u6bd4\u4f5c\u4e3a\u6307\u6807\u3002\u7136\u540e\uff0c\u4ed6\u4eec\u5206\u522b\u5e94\u7528 R1 \u6a21\u677f\u548c Qwen-Math \u6a21\u677f\u6765\u83b7\u53d6\u6a21\u578b\u56de\u7b54\uff0c\u5e76\u6839\u636e\u8be5\u6307\u6807\u786e\u5b9a\u6bcf\u4e2a\u6a21\u578b\u6700\u5408\u9002\u7684\u6a21\u677f\u3002\u6700\u540e\uff0c\u4f7f\u7528\u76f8\u5e94\u6a21\u677f\u8bc4\u4f30\u6bcf\u4e2a\u6a21\u578b\u7684 pass@8 \u51c6\u786e\u5ea6\uff0c\u4ee5\u5224\u65ad\u57fa\u7840\u7b56\u7565\u662f\u5426\u80fd\u591f\u63a2\u7d22\u51fa\u6709\u76ca\u4e8e\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6539\u8fdb\u7684\u8f68\u8ff9\u3002<\/p>\n<p>\u5728\u7ed3\u679c\u5448\u73b0\u4e2d\uff0c\u56fe 3 \u4e2d\u7684\u5de6\u56fe\u663e\u793a\u4e86\u57fa\u7840\u6a21\u578b\uff08\u65e0\u8bba\u662f\u5426\u4f7f\u7528\u6a21\u677f\uff09\u5bf9\u6240\u63d0\u4f9b\u95ee\u9898\u7684\u56de\u7b54\u80fd\u529b\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-86ea83a9b5cb9f561156d452c22e50df.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-86ea83a9b5cb9f561156d452c22e50df.png\"><\/a><\/p>\n<p>\u7814\u7a76\u4eba\u5458\u89c2\u5bdf\u5230\uff0cLlama \u548c DeepSeek \u6a21\u578b\u901a\u8fc7\u4f7f\u7528\u9002\u5f53\u7684\u6a21\u677f\uff08R1 \u6a21\u677f\uff09\u90fd\u63d0\u9ad8\u4e86\u56de\u7b54\u80fd\u529b\u3002\u7136\u800c\uff0cQwen2.5 \u6a21\u578b\u5728\u4e0d\u4f7f\u7528\u6a21\u677f\u65f6\u8868\u73b0\u6700\u4f73\uff08\u56de\u7b54\u7387\u4e3a 100%\uff09\u3002\u8fd9\u4e00\u6709\u8da3\u7684\u7279\u6027\u6fc0\u53d1\u4e86\u4ed6\u4eec\u8fdb\u4e00\u6b65\u7814\u7a76\u7684\u52a8\u529b\u3002\u540c\u65f6\uff0c\u65e0\u6a21\u677f\u65f6\u6700\u4f4e\u7684\u56de\u7b54\u7387\u8868\u660e DeepSeek-V3-Base \u662f\u4e00\u4e2a\u51e0\u4e4e\u7eaf\u7cb9\u7684\u57fa\u7840\u6a21\u578b\u3002\u8fd9\u4e00\u89c2\u5bdf\u7ed3\u679c\u4fc3\u4f7f\u4ed6\u4eec\u63a2\u8ba8\u50cf DeepSeekV3-Base \u8fd9\u6837\u7684\u7eaf\u7cb9\u57fa\u7840\u6a21\u578b\u662f\u5426\u8868\u73b0\u51fa \u300c\u987f\u609f\u65f6\u523b\u300d\u7684\u73b0\u8c61\u3002<\/p>\n<p>\u56fe 3 \u7684\u4e2d\u56fe\u663e\u793a\u4e86\u4e0d\u540c\u57fa\u7840\u6a21\u578b\uff08\u4f7f\u7528\u6a21\u677f\uff09\u5728\u4e0d\u540c\u91c7\u6837\u6e29\u5ea6\u4e0b\u7684 pass@8 \u51c6\u786e\u5ea6\u3002\u8be5\u6307\u6807\u53ef\u4ee5\u4f5c\u4e3a\u57fa\u7840\u7b56\u7565\u63a2\u7d22\u80fd\u529b\u7684\u6307\u6807\u3002\u4f8b\u5982\uff0c\u5982\u679c\u57fa\u7840\u7b56\u7565\u751a\u81f3\u65e0\u6cd5\u91c7\u6837\u51fa\u4e00\u6761\u901a\u5411\u6b63\u786e\u6700\u7ec8\u7b54\u6848\u7684\u8f68\u8ff9\uff0c\u90a3\u4e48\u7531\u4e8e\u6ca1\u6709\u5956\u52b1\u4fe1\u53f7\uff0c\u5f3a\u5316\u5b66\u4e60\u5c31\u4e0d\u53ef\u80fd\u6539\u8fdb\u8be5\u7b56\u7565\u3002\u4ed6\u4eec\u7684\u7ed3\u679c\u8868\u660e\uff0c\u6240\u6709\u6d4b\u8bd5\u7684\u6a21\u578b\u90fd\u5177\u6709\u63a2\u7d22\u6027\uff08\u56e0\u6b64\u5df2\u51c6\u5907\u597d\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\uff09\uff0c\u5176\u4e2d Qwen2.5 \u6a21\u578b\u8868\u73b0\u6700\u4f73\uff08\u751a\u81f3\u8d85\u8d8a\u4e86 DeekSeek-V3-Base\uff09\u3002<strong>\u8fd9\u53ef\u80fd\u90e8\u5206\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48\u5927\u591a\u6570 R1-Zero \u9879\u76ee\u90fd\u57fa\u4e8e Qwen2.5 \u6a21\u578b\u3002<\/strong><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-93f7d78ccec9e457bce24ec1b7b17649.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-93f7d78ccec9e457bce24ec1b7b17649.png\"><\/a><\/p>\n<p><strong>\u57fa\u7840\u6a21\u578b\u4e2d\u5df2\u51fa\u73b0\u300c\u987f\u609f\u65f6\u523b\u300d<\/strong><\/p>\n<p><strong>\u5305\u62ec DeepSeek-V3-Base \u6a21\u578b<\/strong><\/p>\n<p>DeepSeek-R1-Zero \u6700\u4ee4\u4eba\u632f\u594b\u7684\u6210\u679c\u4e4b\u4e00\uff0c\u662f\u901a\u8fc7\u7eaf\u7cb9\u7684\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\uff0c\u6a21\u578b\u6d8c\u73b0\u51fa\u4e86\u81ea\u6211\u53cd\u601d\u884c\u4e3a\uff0c\u4e5f\u5c31\u662f\u6240\u8c13\u7684 \u300c\u987f\u609f\u65f6\u523b\u300d\uff08Aha moment\uff09\u3002<\/p>\n<p>\u4e00\u4e9b\u65e9\u671f\u7814\u7a76\u6307\u51fa\uff0c\u5728\u5f00\u6e90\u7684 R1 \u590d\u73b0\u9879\u76ee\u4e2d\u53ef\u80fd\u4e0d\u5b58\u5728\u300c\u987f\u609f\u65f6\u523b\u300d\uff0c\u56e0\u4e3a\u8fd9\u4e9b\u9879\u76ee\u6240\u4f7f\u7528\u7684\u57fa\u7840\u6a21\u578b\u5df2\u7ecf\u8868\u73b0\u51fa\u81ea\u6211\u53cd\u601d\u76f8\u5173\u7684\u5173\u952e\u8bcd\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u7814\u7a76\u5e76\u672a\u6d4b\u8bd5\u8fc7\u4f5c\u4e3a\u771f\u5b9e R1-Zero \u6a21\u578b\u5f3a\u5316\u5b66\u4e60\u5fae\u8c03\u57fa\u7840\u7684 DeepSeek-V3-Base \u6a21\u578b\u3002\u7814\u7a76\u4eba\u5458\u901a\u8fc7\u81ea\u5df1\u90e8\u7f72 DeepSeek-V3-Base-685B \u6a21\u578b\uff0c\u5e76\u4f7f\u7528 R1 \u6a21\u677f\u5bf9\u5176\u56de\u7b54\u7684 500 \u9053 MATH \u9898\u76ee\u8fdb\u884c\u5206\u6790\uff0c\u586b\u8865\u4e86\u8fd9\u4e00\u7a7a\u767d\u3002<\/p>\n<p>\u4ece\u56fe 3 \u7684\u53f3\u56fe\u53ef\u4ee5\u770b\u51fa\uff0cDeepSeek-V3-Base \u6a21\u578b\u4e5f\u4ea7\u751f\u4e86\u76f8\u5f53\u6570\u91cf\u7684\u81ea\u6211\u53cd\u601d\uff0c\u8fdb\u4e00\u6b65\u9a8c\u8bc1\u4e86\u65e9\u671f\u7814\u7a76\u7684\u4e3b\u5f20\u3002\u6b64\u5916\uff0c\u8be5\u7814\u7a76\u4eba\u5458\u5728\u56fe 4 \u4e2d\u5c55\u793a\u4e86 DeepSeek-V3-Base \u6a21\u578b\u751f\u6210\u8bf8\u5982 \u300cAha\u300d\u300c\u7b49\u7b49\u300d\u300c\u91cd\u65b0\u68c0\u67e5\u95ee\u9898\u300d\u7b49\u5173\u952e\u8bcd\u7684\u4f8b\u5b50\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-40d0629c7ddd49f78cc69f39edf0b413.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-40d0629c7ddd49f78cc69f39edf0b413.png\"><\/a><\/p>\n<p><em>DeepSeek-V3-Base \u5df2\u7ecf\u5c55\u73b0\u51fa\u4e86\u987f\u609f\u65f6\u523b\u3002<\/em><\/p>\n<p>\u53e6\u4e00\u4e2a\u91cd\u8981\u7684\u95ee\u9898\u662f\uff0c\u81ea\u6211\u53cd\u601d\u884c\u4e3a\u662f\u5426\u5728\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u540e\u63d0\u5347\u4e86\u6a21\u578b\u6027\u80fd\u3002\u4e3a\u4e86\u63a2\u7a76\u8fd9\u4e00\u70b9\uff0c\u7814\u7a76\u4eba\u5458\u90e8\u7f72\u4e86 DeepSeek-R1-Zero\uff0c\u5e76\u5206\u6790\u4e86\u5b83\u5bf9 MATH \u6570\u636e\u96c6\u4e2d\u76f8\u540c\u95ee\u9898\u7684\u56de\u7b54\u3002\u4ed6\u4eec\u53d1\u73b0\uff0c\u4e0e\u57fa\u7840\u6a21\u578b\u76f8\u6bd4\uff0cR1-Zero \u4e2d\u81ea\u6211\u53cd\u601d\u884c\u4e3a\u7684\u51fa\u73b0\u9891\u7387\u66f4\u9ad8\u3002\u7136\u800c\uff0c\u6ca1\u6709\u660e\u786e\u7684\u8bc1\u636e\u8868\u660e\u8fd9\u4e9b\u81ea\u6211\u53cd\u601d\u884c\u4e3a\u4e0e\u66f4\u9ad8\u7684\u51c6\u786e\u7387\u76f8\u5173\u3002<\/p>\n<p>\u770b\u8d77\u6765\uff0c\u5728\u57fa\u7840\u6a21\u578b\u9884\u8bad\u7ec3\u4e2d\u5b58\u5728\u7684\u504f\u5dee\uff0c\u81ea\u6211\u53cd\u601d\u884c\u4e3a\u3001\u6570\u5b66\u89e3\u51b3\u80fd\u529b\u5728 RL \u901a\u8fc7\u5956\u52b1\u4fe1\u53f7\u5f3a\u5316\u4e4b\u524d\u5c31\u5df2\u7ecf\u88ab\u6ce8\u5165\u4e86\u3002\u90a3\u4e48\u8d8a\u6765\u8d8a\u957f\u7684\u54cd\u5e94\u662f\u8fd9\u79cd RL \u8fc7\u7a0b\u7684\u7ed3\u679c\u5417\uff1f<\/p>\n<p><strong>\u5f3a\u5316\u5b66\u4e60\u5206\u6790<\/strong><\/p>\n<p>\u5728 Deepseek-R1-Zero \u7684\u8bad\u7ec3\u548c\u7814\u7a76\u8fc7\u7a0b\u4e2d\uff0c\u4e00\u4e2a\u663e\u8457\u7684\u8d8b\u52bf\u662f\u6a21\u578b\u5728\u6574\u4e2a\u8bad\u7ec3\u4e2d\u54cd\u5e94\u957f\u5ea6\u589e\u52a0\uff0c\u8fd9\u4e00\u8d8b\u52bf\u88ab\u5e7f\u6cdb\u89e3\u91ca\u4e3a\u6a21\u578b\u6b63\u5728\u53d1\u5c55\u300c\u81ea\u6211\u53cd\u601d\u300d\u7b49\u9ad8\u7ea7\u63a8\u7406\u80fd\u529b\u7684\u6807\u5fd7\uff0c\u540e\u7eed\u4e5f\u6709\u5f88\u591a\u7814\u7a76\u8005\u901a\u8fc7\u4e0d\u540c\u7684\u7b97\u6cd5\u548c\u5b9e\u73b0\u65b9\u5f0f\u590d\u73b0\u4e86\u8fd9\u4e00\u8d8b\u52bf\u3002<\/p>\n<p>\u7136\u800c\uff0c\u6709\u5b66\u8005\u63d0\u51fa\uff0c\u89c2\u5bdf\u5230\u7684\u54cd\u5e94\u957f\u5ea6\u589e\u52a0\u53ef\u80fd\u5e76\u975e\u5b8c\u5168\u7531\u9ad8\u7ea7\u63a8\u7406\u80fd\u529b\u7684\u53d1\u5c55\u6240\u9a71\u52a8\uff0c\u800c\u662f\u5f52\u56e0\u4e8e GRPO \u76ee\u6807\u51fd\u6570\u56fa\u6709\u7684\u504f\u5dee\uff1a<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-eee33ae251c49d5bdad95407d5ba897a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-eee33ae251c49d5bdad95407d5ba897a.png\"><\/a><\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0e4acff587e4fd22b2183791c3a7d9cb.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0e4acff587e4fd22b2183791c3a7d9cb.png\"><\/a><\/p>\n<p>\u5176\u4e2d\uff0c<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ad2ef434c66a22dfd6739bb266b7af4f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ad2ef434c66a22dfd6739bb266b7af4f.png\"><\/a><\/p>\n<p>R (q, oi) \u8868\u793a Deepseek-R1-Zero \u4e2d\u7ed9\u5b9a\u95ee\u9898 q \u548c\u91c7\u6837\u54cd\u5e94 oi \u7684\u7ed3\u679c\u5956\u52b1\uff08\u8be5\u7ed3\u8bba\u4e5f\u9002\u7528\u4e8e\u8fc7\u7a0b\u5956\u52b1\u6848\u4f8b\uff09\u3002<\/p>\n<p>\u4e0e\u516c\u5f0f\uff082\uff09\u4e2d\u7684\u76ee\u6807\u51fd\u6570\u76f8\u6bd4\uff0cGRPO \u5f15\u5165\u4e86\u4e24\u79cd\u504f\u5dee\uff1a<\/p>\n<ul>\n<li>\n<p>\u54cd\u5e94\u5c42\u9762\u7684\u957f\u5ea6\u504f\u5dee\uff1a\u8fd9\u79cd\u504f\u5dee\u6e90\u4e8e\u5bf9\u54cd\u5e94\u957f\u5ea6\u8fdb\u884c\u5f52\u4e00\u5316\u5904\u7406\u3002\u5f53\u6a21\u578b\u751f\u6210\u6b63\u786e\u54cd\u5e94\u65f6\uff08\u5373\u4f18\u52bf\u503c\u4e3a\u6b63\uff09\uff0c\u8f83\u77ed\u7684\u54cd\u5e94\u4f1a\u83b7\u5f97\u66f4\u5927\u7684\u68af\u5ea6\u66f4\u65b0\uff0c\u4ece\u800c\u4fc3\u4f7f\u7b56\u7565\u503e\u5411\u4e8e\u751f\u6210\u7b80\u6d01\u7684\u6b63\u786e\u7b54\u6848\u3002\u7136\u800c\uff0c\u5f53\u6a21\u578b\u751f\u6210\u9519\u8bef\u54cd\u5e94\u65f6\uff08\u5373\u4f18\u52bf\u503c\u4e3a\u8d1f\uff09\uff0c\u8f83\u957f\u7684\u54cd\u5e94\u7531\u4e8e\u957f\u5ea6\u8f83\u5927\u800c\u53d7\u5230\u7684\u60e9\u7f5a\u8f83\u8f7b\uff0c\u5bfc\u81f4\u7b56\u7565\u66f4\u503e\u5411\u4e8e\u751f\u6210\u5197\u957f\u7684\u9519\u8bef\u7b54\u6848\u3002\u8fd9\u79cd\u957f\u5ea6\u89c4\u8303\u5316\u673a\u5236\u4f7f\u5f97\u6a21\u578b\u5728\u6b63\u786e\u65f6\u504f\u597d\u7b80\u77ed\uff0c\u800c\u5728\u9519\u8bef\u65f6\u504f\u597d\u5197\u957f\uff0c\u5f62\u6210\u4e86\u4e00\u79cd\u4e0d\u5bf9\u79f0\u7684\u4f18\u5316\u503e\u5411\u3002<\/p>\n<\/li>\n<li>\n<p>\u95ee\u9898\u5c42\u9762\u7684\u96be\u5ea6\u504f\u5dee\uff1a\u8fd9\u79cd\u504f\u5dee\u6e90\u4e8e\u5bf9\u95ee\u9898\u7684\u5956\u52b1\u8fdb\u884c\u6807\u51c6\u5dee\u5f52\u4e00\u5316\u5904\u7406\u3002\u5bf9\u4e8e\u6807\u51c6\u5dee\u8f83\u4f4e\u7684\u95ee\u9898\uff08\u5373\u95ee\u9898\u8fc7\u4e8e\u7b80\u5355\u6216\u8fc7\u4e8e\u56f0\u96be\uff09\uff0c\u7b56\u7565\u66f4\u65b0\u65f6\u4f1a\u8d4b\u4e88\u66f4\u9ad8\u7684\u6743\u91cd\u3002\u5c3d\u7ba1\u4f18\u52bf\u503c\u5f52\u4e00\u5316\u662f\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u5e38\u89c1\u6280\u5de7\uff0c\u4f46 GRPO \u5c06\u5176\u5e94\u7528\u4e8e\u5355\u4e2a\u95ee\u9898\u5c42\u9762\uff0c\u800c\u975e\u6574\u4e2a\u6279\u6b21\uff0c\u8fd9\u5bfc\u81f4\u4e0d\u540c\u95ee\u9898\u5728\u76ee\u6807\u51fd\u6570\u4e2d\u7684\u6743\u91cd\u5206\u5e03\u4e0d\u5747\u3002\u8fd9\u79cd\u673a\u5236\u4f7f\u5f97\u6a21\u578b\u66f4\u504f\u597d\u6781\u7aef\u7b80\u5355\u6216\u6781\u7aef\u56f0\u96be\u7684\u95ee\u9898\uff0c\u800c\u5ffd\u89c6\u4e86\u4e2d\u7b49\u96be\u5ea6\u7684\u95ee\u9898\uff0c\u8fdb\u4e00\u6b65\u5f71\u54cd\u4e86\u7b56\u7565\u7684\u4f18\u5316\u65b9\u5411\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u957f\u5ea6\u504f\u5dee\u5e76\u4e0d\u4ec5\u9650\u4e8e GRPO\uff0c\u5728\u5f00\u6e90\u7684\u6807\u51c6 PPO \u7b97\u6cd5\u4e2d\u4e5f\u666e\u904d\u5b58\u5728\u3002<\/p>\n<p>\u968f\u540e\u4f5c\u8005\u5bf9\u591a\u4e2a\u6d41\u884c\u7684\u5f00\u6e90 PPO \u5b9e\u73b0\u8fdb\u884c\u4e86\u5206\u6790\uff0c\u53d1\u73b0\u8fd9\u4e9b\u5b9e\u73b0\u65e0\u4e00\u4f8b\u5916\u5730\u5728\u635f\u5931\u51fd\u6570\u4e2d\u8868\u73b0\u51fa\u4e86\u57fa\u4e8e\u54cd\u5e94\u957f\u5ea6\u7684\u504f\u5dee\uff08\u8be6\u89c1\u9644\u5f55 1 \u548c\u8868 2\uff09\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-2394f95e1690bca373c1e1e1775697cc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-2394f95e1690bca373c1e1e1775697cc.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0e3750d852c4624d04a3944db5ab99fc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0e3750d852c4624d04a3944db5ab99fc.png\"><\/a><\/section>\n<p>\u7814\u7a76\u8005\u63a8\u6d4b\uff0c\u8fd9\u79cd\u9010\u8bcd\u5f52\u4e00\u5316\u7684\u8bbe\u8ba1\u53ef\u80fd\u6e90\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u9010\u8bcd\u9884\u8bad\u7ec3\u673a\u5236\uff0c\u5176\u521d\u8877\u662f\u8ba9\u6bcf\u4e2a\u8bcd\u5bf9\u76ee\u6807\u51fd\u6570\u7684\u5f71\u54cd\u5747\u8861\u3002\u7136\u800c\uff0c\u5728\u5f3a\u5316\u5b66\u4e60\u7684\u8bed\u5883\u4e0b\uff0c\u901a\u8fc7\u9664\u4ee5\u54cd\u5e94\u957f\u5ea6 | oi | \u8fdb\u884c\u5f52\u4e00\u5316\uff0c\u5374\u610f\u5916\u5f15\u5165\u4e86\u8fd9\u79cd\u504f\u5dee\u3002\u56e0\u6b64\u8fd9\u4e00\u53d1\u73b0\u4e3a\u4f18\u5316\u63d0\u4f9b\u4e86\u65b0\u7684\u601d\u8def\u3002<\/p>\n<p>\u4e3a\u4e86\u4f18\u5316 GRPO \u7b97\u6cd5\u5e76\u89e3\u51b3\u5176\u5b58\u5728\u7684\u504f\u5dee\u95ee\u9898\uff0c\u7814\u7a76\u9488\u5bf9 GRPO \u4e24\u5904\u4fee\u6539\uff1a<strong>\u4e00\u662f\u5220\u9664\u4e86\u957f\u5ea6\u5f52\u4e00\u5316\u9879 \u2223oi\u2223\uff0c\u4e8c\u662f\u79fb\u9664\u4e86\u6807\u51c6\u5dee\uff08std\uff09\u6807\u51c6\u5316\u9879\u3002<\/strong><\/p>\n<p>\u540c\u65f6\uff0c\u4e3a\u786e\u4fdd\u4f18\u5316\u76ee\u6807\u7684\u65e0\u504f\u6027\uff0c\u7814\u7a76\u5728 Listing 1 \u7684\u63a9\u7801\u5747\u503c\u51fd\u6570\u4e2d\u5c06 mask.sum (axis=dim) \u66ff\u6362\u4e3a\u4e00\u4e2a\u56fa\u5b9a\u503c MAX_TOKENS\uff0c\u5177\u4f53\u5982\u7eff\u8272\u9ad8\u4eae\u90e8\u5206\u6240\u793a\u3002<\/p>\n<p>\u57fa\u4e8e\u4e0a\u8ff0\u6539\u8fdb\uff0c\u7814\u7a76\u8005\u5c06\u8fd9\u4e00\u4f18\u5316\u7b97\u6cd5\u547d\u540d\u4e3a Dr. GRPO\uff0c\u5e76\u901a\u8fc7\u5b9e\u9a8c\u9a8c\u8bc1\u4e86\u5176\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u7684\u6709\u6548\u6027\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d0f8d7c2f9f9125d782422ddcc5ad626.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d0f8d7c2f9f9125d782422ddcc5ad626.png\"><\/a><\/p>\n<p>\u5982\u56fe 5 \u6240\u5c55\u793a\u7684\u591a\u9879\u5b9e\u9a8c\u5bf9\u6bd4\u7ed3\u679c\uff0c\u8bc1\u660e Dr. GRPO \u80fd\u591f\u6709\u6548\u7f13\u89e3\u4f18\u5316\u504f\u5dee\u5e76\u663e\u8457\u51cf\u5c11\u4e86\u9519\u8bef\u56de\u7b54\u7684\u957f\u5ea6\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ab7f7c11c0f76d69658b0a6d5b534bf9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ab7f7c11c0f76d69658b0a6d5b534bf9.png\"><\/a><\/p>\n<p>\u5982\u679c\u5bf9\u8be6\u7ec6\u63a8\u5bfc\u8fc7\u7a0b\u611f\u5174\u8da3\uff0c\u53ef\u4ee5\u53c2\u8003\u9644\u5f55 A\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u7531\u4e8e Qwen2.5-Math \u57fa\u7840\u6a21\u578b\u53ef\u4ee5\u5728\u6ca1\u6709\u4efb\u4f55\u63d0\u793a\u6a21\u677f\u7684\u60c5\u51b5\u4e0b\u8f7b\u677e\u56de\u7b54\u95ee\u9898\u5e76\u8fbe\u5230\u9ad8\u7cbe\u5ea6\uff0c\u4f5c\u8005\u63a2\u8ba8\u4e86\u4e0d\u540c\u6a21\u677f\u5bf9 RL \u8bad\u7ec3\u7684\u5f71\u54cd\u3002\u9488\u5bf9\u300c\u66f4\u5927\u95ee\u9898\u8986\u76d6\u7387\u63d0\u5347\u6027\u80fd\u300d\u7684\u666e\u904d\u89c2\u70b9\uff0c\u4f5c\u8005\u8fd8\u7814\u7a76\u4e86\u4e0d\u540c\u6a21\u677f\u4e0e\u95ee\u9898\u8986\u76d6\u7387\u4e4b\u95f4\u7684\u4ea4\u4e92\u4f5c\u7528\u3002\uff08\u5b9e\u9a8c\u8bbe\u7f6e\u8be6\u89c1\u8868 3\uff09<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-3967f1b9cfcdcb72e907dcccc90d92aa.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-3967f1b9cfcdcb72e907dcccc90d92aa.png\"><\/a><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-5ff3105deee0b23dc5b8fa4849325fe5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-5ff3105deee0b23dc5b8fa4849325fe5.png\"><\/a><\/p>\n<p>\u7ed3\u679c\u5982\u56fe 6 \u5c55\u793a\u4e86\u4e0d\u540c\u5b9e\u9a8c\u7684 RL \u8bad\u7ec3\u66f2\u7ebf\uff0c\u63ed\u793a\u4e86\u4ee5\u4e0b\u7ed3\u8bba\uff1a<\/p>\n<ul>\n<li>\n<p>\u6a21\u677f\u51b3\u5b9a\u4e86\u521d\u59cb\u7b56\u7565\u7684\u6027\u80fd\uff0c\u4f46 RL \u53ef\u4ee5\u5c06\u6240\u6709\u7b56\u7565\u63d0\u5347\u81f3\u7ea6 40% \u7684\u76f8\u4f3c\u6c34\u5e73\uff08\u7ed9\u5b9a\u5408\u9002\u7684\u95ee\u9898\u96c6\uff09\uff1b<\/p>\n<\/li>\n<li>\n<p>\u4f7f\u7528 R1 \u6a21\u677f\u65f6\uff0c\u95ee\u9898\u96c6\u8986\u76d6\u7387\u663e\u8457\u5f71\u54cd RL \u52a8\u6001\uff0c\u8986\u76d6\u7387\u4e0d\u8db3\u4f1a\u5bfc\u81f4\u6027\u80fd\u4e0a\u9650\u964d\u4f4e\u3002\u800c\u4f7f\u7528 Qwen-Math \u6a21\u677f\u65f6\uff0cRL \u5728 GSM-8K \u4e0a\u53d6\u5f97\u6700\u4f73\u8868\u73b0\uff0c\u8868\u660e\u5728\u66f4\u7b80\u5355\uff08\u4e14\u5206\u5e03\u5916\uff09\u95ee\u9898\u4e0a\u8bad\u7ec3\u53ef\u5927\u5e45\u63d0\u5347\uff08\u8fd1\u4e4e\u7ffb\u500d\uff09\u5728\u66f4\u96be\u9898\u4e0a\u7684\u6d4b\u8bd5\u51c6\u786e\u7387\u3002<\/p>\n<\/li>\n<\/ul>\n<p>\u8fdb\u4e00\u6b65\u5206\u6790\u8868\u660e\uff1a<\/p>\n<ul>\n<li>\n<p>Qwen2.5-Math-1.5B \u57fa\u7840\u6a21\u578b\u5df2\u5177\u5907\u8f83\u5f3a\u7684\u6570\u5b66\u89e3\u9898\u80fd\u529b\uff08\u89c1\u56fe 6 \u53f3\u56fe\u7684\u8d77\u70b9\uff09\uff0c\u6a21\u677f\u7684\u4f7f\u7528\u53cd\u800c\u4f1a\u7834\u574f\u8fd9\u4e00\u80fd\u529b\uff0c\u76f4\u5230 RL \u91cd\u5efa\u3002\u56e0\u6b64\u8fd9\u8bf4\u660e\u5728\u5f3a\u8c03\u7eaf RL \u5e26\u6765\u7684\u5de8\u5927\u6536\u76ca\u65f6\u5e94\u66f4\u52a0\u8c28\u614e\u3002<\/p>\n<\/li>\n<li>\n<p>\u5f53\u57fa\u7840\u6a21\u578b\u4e0e\u6a21\u677f\u5b58\u5728\u8f83\u5927\u4e0d\u5339\u914d\u65f6\uff08\u5982 R1 \u6a21\u677f\u4e0e Qwen2.5-Math-1.5B\uff09\uff0c\u7b56\u7565\u6539\u8fdb\u4e3b\u8981\u4f9d\u8d56 RL \u8c03\u4f18\uff0c\u9700\u95ee\u9898\u96c6\u5177\u6709\u826f\u597d\u7684\u8986\u76d6\u7387\uff08\u89c1\u56fe 6 \u5de6\u56fe\uff09\u3002\u53cd\u4e4b\uff0c\u5373\u4f7f\u662f\u4e00\u4e2a\u5c0f\u578b\u4e14\u5b8c\u5168\u5206\u5e03\u5916\u7684\u95ee\u9898\u96c6\uff0c\u4e5f\u80fd\u901a\u8fc7\u5f3a\u5316\u6b63\u786e\u63a8\u7406\u884c\u4e3a\u800c\u975e\u6ce8\u5165\u65b0\u77e5\u8bc6\uff0c\u6709\u6548\u63d0\u5347\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<\/li>\n<\/ul>\n<p><strong>\u9886\u57df\u7279\u5b9a\u9884\u8bad\u7ec3\u53ef\u63d0\u5347 RL \u4e0a\u9650<\/strong><\/p>\n<p>\u6700\u8fd1\u6210\u529f\u7684 R1-Zero \u7c7b\u6570\u5b66\u63a8\u7406\u5668\u590d\u5236\u5927\u591a\u4ee5 Qwen2.5 \u57fa\u7840\u6a21\u578b\u4f5c\u4e3a\u521d\u59cb\u7b56\u7565\uff0c\u8fd9\u4e9b\u6a21\u578b\u672c\u8eab\u5df2\u662f\u5f3a\u5927\u7684\u6570\u5b66\u89e3\u9898\u5668\uff0c\u5e76\u5c55\u73b0\u51fa\u81ea\u6211\u53cd\u601d\u7684\u6a21\u5f0f\u3002\u4e0d\u8fc7\uff0c\u8be5\u7814\u7a76\u4eba\u5458\u5e0c\u671b\u63a2\u8ba8\u53e6\u4e00\u65b9\u9762\uff1aR1-Zero \u7c7b\u8bad\u7ec3\u80fd\u5426\u5728\u6700\u521d\u6570\u5b66\u63a8\u7406\u80fd\u529b\u8f83\u5f31\u7684\u57fa\u7840\u6a21\u578b\u4e0a\u53d6\u5f97\u6210\u529f\uff1f\u4ed6\u4eec\u7684\u56de\u7b54\u662f\u80af\u5b9a\u7684\uff0c\u5e76\u4e14\u89c2\u5bdf\u5230\u6570\u5b66\u9884\u8bad\u7ec3\u80fd\u591f\u63d0\u5347\u5f3a\u5316\u5b66\u4e60\u7684\u4e0a\u9650\u3002<\/p>\n<p>\u5728\u5b9e\u9a8c\u8bbe\u7f6e\u4e2d\uff0c\u4ed6\u4eec\u4ee5 Llama-3.2-3B \u57fa\u7840\u6a21\u578b\u4e3a\u8d77\u70b9\uff0c\u91c7\u7528\u65e0\u504f\u7684 Dr. GRPO \u7b97\u6cd5\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u5fae\u8c03\uff0c\u5e76\u4f7f\u7528 R1 \u6a21\u677f\u3002\u4ed6\u4eec\u5047\u8bbe\u9886\u57df\u7279\u5b9a\u7684\u9884\u8bad\u7ec3\u6709\u52a9\u4e8e\u5f3a\u5316\u5b66\u4e60\uff0c\u56e0\u6b64\u9009\u7528\u4e86\u5728 FineMath \u6570\u636e\u96c6\u4e0a\u6301\u7eed\u9884\u8bad\u7ec3\u7684 Llama-3.2-3B-FineMath4 \u6a21\u578b\u3002\u6b64\u5916\uff0c\u6b63\u5982\u4e4b\u524d\u5047\u8bbe\u7684\u90a3\u6837\uff0cQwen2.5 \u6a21\u578b\u5f88\u53ef\u80fd\u662f\u5728\u62fc\u63a5\u7684\u95ee\u7b54\u6587\u672c\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\u7684\uff0c\u56e0\u6b64\u4ed6\u4eec\u540c\u6837\u4ece NuminaMath1.5 \u4e2d\u51c6\u5907\u4e86\u4e00\u4e2a\u62fc\u63a5\u7684\u6570\u636e\u96c6\uff0c\u5e76\u7528 1e-5 \u7684\u5b66\u4e60\u7387\u5bf9 Llama-3.2-3B-FineMath \u8fdb\u884c\u4e86 2 \u8f6e\u6301\u7eed\u9884\u8bad\u7ec3\u3002\u4ed6\u4eec\u5c06\u8fd9\u79cd\u62fc\u63a5\u6301\u7eed\u9884\u8bad\u7ec3\u7684\u6a21\u578b\u79f0\u4e3a Llama-3.2-3BNuminaQA\u3002<\/p>\n<p>\u4ee5\u4e0b\u662f\u7ed3\u679c\u5448\u73b0\u3002\u56fe 7 \u7684\u5de6\u56fe\u4e2d\u5c55\u793a\u4e86\u4e0d\u540c\u57fa\u7840\u6a21\u578b\u7684\u5f3a\u5316\u5b66\u4e60\u66f2\u7ebf\u3002\u4ed6\u4eec\u89c2\u5bdf\u5230\uff0c\u5373\u4f7f\u662f\u5bf9\u57fa\u7840\u7684 Llama \u6a21\u578b\uff0c\u5f3a\u5316\u5b66\u4e60\u4e5f\u80fd\u63d0\u5347\u5176\u6027\u80fd\uff0c\u4f46\u63d0\u5347\u5e45\u5ea6\u5f88\u5c0f\u3002\u5728\u7ecf\u8fc7\u6301\u7eed\u9884\u8bad\u7ec3\uff08\u4ee5\u53ca\u62fc\u63a5\u6301\u7eed\u9884\u8bad\u7ec3\uff09\u4ee5\u878d\u5165\u6570\u5b66\u9886\u57df\u77e5\u8bc6\u540e\uff0cLlama \u6a21\u578b\u5c55\u73b0\u51fa\u4e86\u66f4\u5f3a\u5927\u7684\u5f3a\u5316\u5b66\u4e60\u6027\u80fd\uff0c\u9a8c\u8bc1\u4e86\u4ed6\u4eec\u7684\u5047\u8bbe\u3002<\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0c8ed42715a34ae63c26940228a3bccf.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-0c8ed42715a34ae63c26940228a3bccf.png\"><\/a><\/section>\n<p>\u4f5c\u8005\u8fd8\u4f7f\u7528 Llama \u57fa\u7840\u6a21\u578b\u91cd\u65b0\u5ba1\u89c6\u4e86 GRPO \u7684\u4f18\u5316\u504f\u5dee\u3002\u56fe 7 \u7684\u53f3\u56fe\u6bd4\u8f83\u4e86\u4f7f\u7528 GRPO \u548c Dr. GRPO \u8bad\u7ec3\u7684\u6a21\u578b\u6027\u80fd\u548c\u56de\u7b54\u957f\u5ea6\uff0c\u53ef\u4ee5\u6e05\u695a\u5730\u770b\u5230\uff0cGRPO \u53ef\u4ee5\u4ea7\u751f\u300c\u53cc\u91cd\u589e\u52a0\u300d\u73b0\u8c61\uff0c\u53ef\u80fd\u5bfc\u81f4\u8bef\u89e3\uff0c\u5373\u5728\u6570\u5b66\u9884\u8bad\u7ec3\u540e\uff0c\u957f\u601d\u7ef4\u94fe\uff08long-CoT\uff09\u4e5f\u80fd\u5728 Llama \u6a21\u578b\u4e0a\u51fa\u73b0\u3002\u7136\u800c\uff0c\u957f\u5ea6\u7684\u589e\u52a0\u53ef\u80fd\u662f\u7531\u4e8e\u4f18\u5316\u504f\u5dee\uff0c\u800c\u8fd9\u79cd\u504f\u5dee\u53ef\u4ee5\u901a\u8fc7\u4ed6\u4eec\u63d0\u51fa\u7684 Dr. GRPO \u6709\u6548\u7f13\u89e3\uff08\u56fe 7 \u7684\u53f3\u56fe\uff09\u3002<\/p>\n<p><strong>\u7ed3\u8bed<\/strong><\/p>\n<p>\u7814\u7a76\u4eba\u5458\u5bf9\u7528\u4e8e R1-Zero \u7c7b\u4f3c\u8bad\u7ec3\u7684\u57fa\u7840\u6a21\u578b\u4ee5\u53ca\u5f3a\u5316\u5b66\u4e60\u4e2d\u6240\u7528\u7b97\u6cd5\u8fdb\u884c\u4e86\u6279\u5224\u6027\u5ba1\u89c6\u3002\u901a\u8fc7\u5206\u6790\uff0c\u4f5c\u8005\u63ed\u793a\u4e86\u9884\u8bad\u7ec3\u504f\u5dee\u5982\u4f55\u5f71\u54cd\u5f3a\u5316\u5b66\u4e60\u7684\u7ed3\u679c\uff0c\u4ee5\u53ca\u8bf8\u5982 GRPO \u8fd9\u6837\u7684\u4f18\u5316\u9009\u62e9\u5982\u4f55\u65e0\u610f\u4e2d\u6539\u53d8\u6a21\u578b\u7684\u884c\u4e3a\u3002\u4ed6\u4eec\u63d0\u51fa\u7684 Dr. GRPO \u7b97\u6cd5\u63d0\u4f9b\u4e86\u4e00\u4e2a\u7b80\u5355\u7684\u4fee\u6b63\u65b9\u6848\uff0c\u5728\u4fdd\u6301\u63a8\u7406\u6027\u80fd\u7684\u540c\u65f6\u63d0\u9ad8 token \u6548\u7387\u3002<\/p>\n<p>\u6700\u7ec8\u7684\u7814\u7a76\u7ed3\u679c\u8868\u660e\uff0c\u6269\u5927\u5f3a\u5316\u5b66\u4e60\u7684\u89c4\u6a21\u65e2\u80fd\u63d0\u5347\u6548\u679c\uff0c\u53c8\u80fd\u63d0\u9ad8\u6548\u7387 \u2014\u2014 \u6709\u65f6\uff0c\u7b80\u5355\u5316\u53cd\u800c\u80fd\u5b9e\u73b0\u66f4\u4f18\u7684\u6027\u80fd\u3002<\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-22-5\" target=\"_blank\">\u63ed\u79d8DeepSeek R1-Zero\u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO\u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u63ed\u79d8DeepSeek  [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[27],"tags":[68],"class_list":["post-39605","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>\u63ed\u79d8DeepSeek R1-Zero\u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO\u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=39605\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39605#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39605\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"\u63ed\u79d8DeepSeek R1-Zero\u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO\u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848\",\"datePublished\":\"2025-03-23T17:02:23+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39605\"},\"wordCount\":273,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39605#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/03\\\/frc-809960926c9870a6df09192262a995cc.png\",\"keywords\":[\"\u5927\u6a21\u578b\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=39605#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39605\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=39605\",\"name\":\"\u63ed\u79d8DeepSeek R1-Zero\u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO\u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39605#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39605#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/03\\\/frc-809960926c9870a6df09192262a995cc.png\",\"datePublished\":\"2025-03-23T17:02:23+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39605#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=39605\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39605#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/03\\\/frc-809960926c9870a6df09192262a995cc.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/03\\\/frc-809960926c9870a6df09192262a995cc.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39605#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u63ed\u79d8DeepSeek R1-Zero\u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO\u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\",\"url\":\"https:\\\/\\\/17aitech.com\\\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/17aitech.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"url\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"contentUrl\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u63ed\u79d8DeepSeek R1-Zero\u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO\u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=39605","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=39605#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=39605"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"\u63ed\u79d8DeepSeek R1-Zero\u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO\u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848","datePublished":"2025-03-23T17:02:23+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=39605"},"wordCount":273,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=39605#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-809960926c9870a6df09192262a995cc.png","keywords":["\u5927\u6a21\u578b"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=39605#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=39605","url":"https:\/\/17aitech.com\/?p=39605","name":"\u63ed\u79d8DeepSeek R1-Zero\u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO\u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=39605#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=39605#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-809960926c9870a6df09192262a995cc.png","datePublished":"2025-03-23T17:02:23+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=39605#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=39605"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=39605#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-809960926c9870a6df09192262a995cc.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-809960926c9870a6df09192262a995cc.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=39605#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u63ed\u79d8DeepSeek R1-Zero\u8bad\u7ec3\u65b9\u5f0f\uff0cGRPO\u8fd8\u6709\u6781\u7b80\u6539\u8fdb\u65b9\u6848"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39605","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=39605"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39605\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=39605"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=39605"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=39605"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}