{"id":39505,"date":"2025-03-19T04:02:22","date_gmt":"2025-03-18T20:02:22","guid":{"rendered":"https:\/\/17aitech.com\/?p=39505"},"modified":"2025-03-19T04:02:22","modified_gmt":"2025-03-18T20:02:22","slug":"%e8%b6%85%e8%b6%8adeepseek-grpo%e7%9a%84%e5%85%b3%e9%94%aerl%e7%ae%97%e6%b3%95%ef%bc%8c%e5%ad%97%e8%8a%82%e3%80%81%e6%b8%85%e5%8d%8eair%e5%bc%80%e6%ba%90dapo","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=39505","title":{"rendered":"\u8d85\u8d8aDeepSeek GRPO\u7684\u5173\u952eRL\u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534eAIR\u5f00\u6e90DAPO"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-18-9\" target=\"_blank\">\u8d85\u8d8aDeepSeek GRPO\u7684\u5173\u952eRL\u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534eAIR\u5f00\u6e90DAPO<\/a><\/p>\n<section>DeepSeek \u63d0\u51fa\u7684 GRPO \u53ef\u4ee5\u6781\u5927\u63d0\u5347 LLM \u7684\u5f3a\u5316\u5b66\u4e60\u6548\u7387\uff0c\u4e0d\u8fc7\u5176\u8bba\u6587\u4e2d\u4f3c\u4e4e\u8fd8\u7f3a\u5c11\u4e00\u4e9b\u5173\u952e\u7ec6\u8282\uff0c\u8ba9\u4eba\u96be\u4ee5\u590d\u73b0\u51fa\u5927\u89c4\u6a21\u548c\u5de5\u4e1a\u7ea7\u7684\u5f3a\u5316\u5b66\u4e60\u7cfb\u7edf\u3002<\/section>\n<section><\/section>\n<section>\u8fd1\u65e5\uff0c\u6e05\u534e AIR \u548c\u5b57\u8282\u8054\u5408 SIA Lab \u53d1\u5e03\u4e86\u4ed6\u4eec\u7684\u7b2c\u4e00\u9879\u7814\u7a76\u6210\u679c\uff1aDAPO\uff0c\u5373 Decoupled Clip and Dynamic sAmpling Policy Optimization\uff08\u89e3\u8026\u526a\u8f91\u548c\u52a8\u6001\u91c7\u6837\u7b56\u7565\u4f18\u5316\uff09\u3002\u8fd9\u662f\u4e00\u4e2a\u53ef\u5b9e\u73b0\u5927\u89c4\u6a21 LLM \u5f3a\u5316\u5b66\u4e60\u7684\u5f00\u6e90 SOTA \u7cfb\u7edf\u3002\u6b64\u5916\uff0c\u4f7f\u7528\u8be5\u7b97\u6cd5\u8bad\u7ec3\u7684\u6a21\u578b\u4e5f\u5c06\u5728\u8fd1\u671f\u5f00\u6e90\u53d1\u5e03\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ef623e220fdaceda3039115e21bd85bb.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ef623e220fdaceda3039115e21bd85bb.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u9879\u76ee\u9875\u9762\uff1ahttps:\/\/dapo-sia.github.io\/<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/dapo-sia.github.io\/static\/pdf\/dapo_paper.pdf<\/section>\n<\/li>\n<li>\n<section>\u4ee3\u7801\u5730\u5740\uff1ahttps:\/\/github.com\/volcengine\/verl\/tree\/gm-tyx\/puffin\/main\/recipe\/dapo<\/section>\n<\/li>\n<li>\n<section>\u6570\u636e\uff1ahttps:\/\/huggingface.co\/datasets\/BytedTsinghua-SIA\/DAPO-Math-17k<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u4f7f\u7528\u8be5\u7b97\u6cd5\uff0c\u8be5\u56e2\u961f\u6210\u529f\u8ba9 Qwen2.5-32B \u6a21\u578b\u5728 AIME 2024 \u57fa\u51c6\u4e0a\u83b7\u5f97\u4e86 50 \u5206\uff0c\u4f18\u4e8e\u540c\u7b49\u89c4\u6a21\u7684 DeepSeek-R1-Zero-Qwen-32B\uff0c\u540c\u65f6 DAPO \u7248 Qwen2.5-32B \u4f7f\u7528\u7684\u8bad\u7ec3\u6b65\u6570\u8fd8\u5c11 50%\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d493bc040697a98e6744db84b7b6b7ab.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d493bc040697a98e6744db84b7b6b7ab.png\"><\/a><\/section>\n<section>\u76f8\u8f83\u4e4b\u4e0b\uff0c\u5982\u679c\u4f7f\u7528 GRPO\uff0cQwen2.5-32B \u6a21\u578b\u5728 AIME 2024 \u57fa\u51c6\u4e0a\u53ea\u80fd\u83b7\u5f97 30 \u5206\u3002<\/section>\n<section><\/section>\n<section>30 \u5206\u7684\u6210\u7ee9\u8fdc\u4f4e\u4e8e DeepSeek \u7684\u5f3a\u5316\u5b66\u4e60\uff0847 \u5206\uff09\u3002\u8be5\u56e2\u961f\u5206\u6790\u53d1\u73b0\uff0c\u539f\u751f\u7248 GRPO \u9762\u4e34\u7740\u51e0\u5927\u5173\u952e\u95ee\u9898\uff0c\u6bd4\u5982\u71b5\u5d29\u6e83\u3001\u5956\u52b1\u566a\u97f3\u548c\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u3002\u4e8b\u5b9e\u4e0a\uff0c\u8be5\u56e2\u961f\u8868\u793a\uff0c\u5f88\u591a\u7814\u7a76\u56e2\u961f\u5728\u590d\u73b0 DeepSeek \u7684\u7ed3\u679c\u65f6\u4e5f\u9047\u5230\u4e86\u7c7b\u4f3c\u7684\u96be\u9898\u3002\u4ed6\u4eec\u8868\u793a\uff1a\u300c\u8fd9\u8868\u660e R1 \u8bba\u6587\u4e2d\u53ef\u80fd\u7701\u7565\u4e86\u5f00\u53d1\u5de5\u4e1a\u7ea7\u3001\u5927\u89c4\u6a21\u4e14\u53ef\u91cd\u73b0\u7684\u5f3a\u5316\u5b66\u4e60\u7cfb\u7edf\u6240\u9700\u7684\u5173\u952e\u8bad\u7ec3\u7ec6\u8282\u3002\u300d<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-02c528226f7e5ba70ab7b4b1857d364a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-02c528226f7e5ba70ab7b4b1857d364a.png\"><\/a><\/section>\n<section><em><sup>Allen AI \u7814\u7a76\u8005 Nathan Lambert \u603b\u7ed3\u4e86 DAPO \u5bf9 GRPO \u7684\u6539\u8fdb\uff0c\u5305\u62ec\u4e24\u4e2a\u4e0d\u540c\u7684\u88c1\u526a\u8d85\u53c2\u6570\u3001\u52a8\u6001\u91c7\u6837\u3001token \u5c42\u9762\u7684\u7b56\u7565\u68af\u5ea6\u635f\u5931\u3001\u8fc7\u957f\u5956\u52b1\u5851\u9020<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u4e0b\u9762\u5c06\u4ece PPO \u5230 GRPO \u518d\u5230 DAPO \u9010\u6b65\u4ecb\u7ecd\uff0c\u770b\u770b\u8fd9\u4e2a\u65b0\u7684\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u7a76\u7adf\u662f\u5982\u4f55\u70bc\u6210\u7684\u3002<\/section>\n<section><\/section>\n<section><strong>\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff08PPO\uff09<\/strong><\/section>\n<section><\/section>\n<section>PPO \u5f15\u5165\u4e86\u88c1\u526a\u5f0f\u66ff\u4ee3\u76ee\u6807\uff08clipped surrogate objective\uff09\u6765\u6267\u884c\u7b56\u7565\u4f18\u5316\u3002\u901a\u8fc7\u4f7f\u7528\u88c1\u526a\u5c06\u7b56\u7565\u66f4\u65b0\u9650\u5236\u5728\u5148\u524d\u7b56\u7565\u7684\u8fd1\u7aef\u533a\u57df\u5185\uff0cPPO \u53ef\u4ee5\u8ba9\u8bad\u7ec3\u7a33\u5b9a\u5e76\u63d0\u9ad8\u6837\u672c\u6548\u7387\u3002\u5177\u4f53\u800c\u8a00\uff0cPPO \u66f4\u65b0\u7b56\u7565\u7684\u65b9\u5f0f\u662f\u6700\u5927\u5316\u4ee5\u4e0b\u76ee\u6807\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-89f3f97cdbee913636c6ce9c1409e9f1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-89f3f97cdbee913636c6ce9c1409e9f1.png\"><\/a><\/section>\n<section><strong>\u7fa4\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff08GRPO\uff09<\/strong>\u00a0<\/section>\n<section><\/section>\n<section>\u4e0e PPO \u76f8\u6bd4\uff0cGRPO \u6d88\u9664\u4e86\u4ef7\u503c\u51fd\u6570\u5e76\u4ee5\u7fa4\u7ec4\u76f8\u5173\u7684\u65b9\u5f0f\u6765\u4f30\u8ba1\u4f18\u52bf\u3002\u5bf9\u4e8e\u7279\u5b9a\u7684\u95ee\u7b54\u5bf9 (q, a),\u884c\u4e3a\u7b56\u7565 \u03c0_\u03b8_old \u91c7\u6837\u4e00\u7ec4 G \u4e2a\u4e2a\u4f53\u54cd\u5e94<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f3c2cfe1c01a6bc97d5f23187966e53a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f3c2cfe1c01a6bc97d5f23187966e53a.png\"><\/a>\u3002\u7136\u540e\uff0c\u901a\u8fc7\u5bf9\u7fa4\u7ec4\u7ea7\u5956\u52b1<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-b76816c26a83d32a91cf3fe7eb223848.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-b76816c26a83d32a91cf3fe7eb223848.png\"><\/a>\u8fdb\u884c\u5f52\u4e00\u5316\u6765\u8ba1\u7b97\u7b2c i \u4e2a\u54cd\u5e94\u7684\u4f18\u52bf\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-eee8d3bfea0dc8ff9c5c22220ba84ca0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-eee8d3bfea0dc8ff9c5c22220ba84ca0.png\"><\/a><\/section>\n<section>\u4e0e PPO \u7c7b\u4f3c\uff0cGRPO \u4e5f\u91c7\u7528\u4e86\u88c1\u526a\u76ee\u6807\uff0c\u540c\u65f6\u8fd8\u6709\u4e00\u4e2a\u76f4\u63a5\u6dfb\u52a0\u7684 KL \u60e9\u7f5a\u9879\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d82a6235befb8c7679f1321e1318ab3f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-d82a6235befb8c7679f1321e1318ab3f.png\"><\/a><\/section>\n<section>\u8fd8\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cGRPO \u662f\u5728\u6837\u672c\u5c42\u7ea7\u8ba1\u7b97\u76ee\u6807\u3002\u786e\u5207\u5730\u8bf4\uff0cGRPO \u9996\u5148\u4f1a\u8ba1\u7b97\u6bcf\u4e2a\u751f\u6210\u5e8f\u5217\u4e2d\u7684\u5e73\u5747\u635f\u5931\uff0c\u7136\u540e\u518d\u5bf9\u4e0d\u540c\u6837\u672c\u7684\u635f\u5931\u8fdb\u884c\u5e73\u5747\u3002\u6b63\u5982\u540e\u9762\u4f1a\u8ba8\u8bba\u7684\u90a3\u6837\uff0c\u8fd9\u79cd\u5dee\u5f02\u53ef\u80fd\u4f1a\u5bf9\u7b97\u6cd5\u7684\u6027\u80fd\u4ea7\u751f\u5f71\u54cd\u3002<\/section>\n<section><\/section>\n<section><strong>\u53e6\u5916\u4e24\u9879\u6539\u8fdb<\/strong><\/section>\n<section><\/section>\n<section><strong>\u6d88\u9664 KL \u504f\u79bb<\/strong><\/section>\n<section><\/section>\n<section>KL \u60e9\u7f5a\u9879\u7684\u4f5c\u7528\u662f\u8c03\u8282\u5728\u7ebf\u7b56\u7565\u548c\u51bb\u7ed3\u53c2\u8003\u7b56\u7565\u4e4b\u95f4\u7684\u504f\u79bb\u60c5\u51b5\u3002\u5728 RLHF \u573a\u666f\u4e2d\uff0cRL \u7684\u76ee\u6807\u662f\u5728\u4e0d\u504f\u79bb\u521d\u59cb\u6a21\u578b\u592a\u8fdc\u7684\u60c5\u51b5\u4e0b\u8c03\u6574\u6a21\u578b\u884c\u4e3a\u3002\u7136\u800c\uff0c\u5728\u8bad\u7ec3\u957f CoT \u63a8\u7406\u6a21\u578b\u65f6\uff0c\u6a21\u578b\u5206\u5e03\u53ef\u80fd\u4f1a\u4e0e\u521d\u59cb\u6a21\u578b\u6709\u663e\u8457\u5dee\u5f02\uff0c\u56e0\u6b64\u8fd9\u79cd\u9650\u5236\u662f\u4e0d\u5fc5\u8981\u7684\u3002\u56e0\u6b64\uff0c\u5728 DAPO \u4e2d\uff0cKL \u9879\u88ab\u6392\u9664\u5728\u5916\u3002<\/section>\n<section><\/section>\n<section><strong>\u57fa\u4e8e\u89c4\u5219\u7684\u5956\u52b1\u5efa\u6a21<\/strong><\/section>\n<section><\/section>\n<section>\u5956\u52b1\u6a21\u578b\u7684\u4f7f\u7528\u901a\u5e38\u4f1a\u53d7\u5230\u5956\u52b1 hacking \u95ee\u9898\u7684\u5f71\u54cd\u3002\u4f5c\u4e3a\u66ff\u4ee3\uff0c\u8be5\u56e2\u961f\u76f4\u63a5\u4f7f\u7528\u53ef\u9a8c\u8bc1\u4efb\u52a1\u7684\u6700\u7ec8\u51c6\u786e\u7387\u4f5c\u4e3a\u7ed3\u679c\u5956\u52b1\uff0c\u8ba1\u7b97\u89c4\u5219\u5982\u4e0b\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f3ce47a4d23d5364d5cf2024b1cac9e1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-f3ce47a4d23d5364d5cf2024b1cac9e1.png\"><\/a><\/section>\n<section>\u4e8b\u5b9e\u8bc1\u660e\uff0c\u8fd9\u662f\u6fc0\u6d3b\u57fa\u7840\u6a21\u578b\u63a8\u7406\u80fd\u529b\u7684\u6709\u6548\u65b9\u6cd5\uff0c\u8fd9\u4e5f\u5728\u591a\u4e2a\u9886\u57df\u5f97\u5230\u4e86\u8bc1\u660e\uff0c\u5305\u62ec\u81ea\u52a8\u5b9a\u7406\u8bc1\u660e\u3001\u8ba1\u7b97\u673a\u7f16\u7a0b\u548c\u6570\u5b66\u7ade\u8d5b\u3002<\/section>\n<section><\/section>\n<section><strong>DAPO<\/strong><\/section>\n<section><\/section>\n<section>\u7814\u7a76\u8005\u63d0\u51fa\u4e86\u89e3\u8026\u526a\u8f91\uff08Decouple Clip\uff09\u548c\u52a8\u6001\u91c7\u6837\u7b56\u7565\u4f18\u5316\uff08DAPO\uff09\u7b97\u6cd5\u3002DAPO \u5bf9\u6bcf\u4e2a\u95ee\u9898 q \u548c\u7b54\u6848 a \u7684\u4e00\u7ec4\u8f93\u51fa<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-375a454e5154525bfd786c35f98f455e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-375a454e5154525bfd786c35f98f455e.png\"><\/a>\u8fdb\u884c\u91c7\u6837\uff0c\u5e76\u901a\u8fc7\u4ee5\u4e0b\u76ee\u6807\u4f18\u5316\u7b56\u7565\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ebe474fb9702f46d862009fbfe0b97ec.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ebe474fb9702f46d862009fbfe0b97ec.png\"><\/a><\/section>\n<section>\u6b64\u5904<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-9959a380c7dbe5757bb37d11d40d8697.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-9959a380c7dbe5757bb37d11d40d8697.png\"><\/a><\/section>\n<section>\u4e0b\u9762\u5c06\u4ecb\u7ecd\u4e0e DAPO \u76f8\u5173\u7684\u5173\u952e\u6280\u672f\u3002<\/section>\n<section><\/section>\n<section><strong>\u62ac\u9ad8\u5929\u82b1\u677f\uff1aClip-Higher<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u4f7f\u7528 PPO \u6216 GRPO \u8fdb\u884c\u7684\u521d\u6b65\u5b9e\u9a8c\u4e2d\uff0c\u7814\u7a76\u8005\u89c2\u5bdf\u5230\u4e86\u71b5\u5d29\u6e83\u73b0\u8c61\uff1a\u968f\u7740\u8bad\u7ec3\u7684\u8fdb\u884c\uff0c\u7b56\u7565\u7684\u71b5\u8fc5\u901f\u4e0b\u964d\uff08\u5982\u56fe 2b\uff09\u3002\u67d0\u4e9b\u7ec4\u7684\u91c7\u6837\u54cd\u5e94\u901a\u5e38\u51e0\u4e4e\u76f8\u540c\u3002\u8fd9\u8868\u660e\u6709\u9650\u7684\u63a2\u7d22\u548c\u65e9\u671f\u7684\u786e\u5b9a\u6027\u7b56\u7565\u4f1a\u963b\u788d\u6269\u5c55\u8fc7\u7a0b\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-99584130fce6a339afcba5b238a88735.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-99584130fce6a339afcba5b238a88735.png\"><\/a><\/section>\n<section>\u9488\u5bf9\u8fd9\u4e00\u95ee\u9898\uff0c\u7814\u7a76\u8005\u63d0\u51fa\u4e86 Clip-Higher \u7b56\u7565\u3002\u5bf9\u91cd\u8981\u5ea6\u91c7\u6837\u7387\u8fdb\u884c\u526a\u8f91\u662f PPO-Clip \u4e2d\u7684\u4e00\u79cd\u7b56\u7565\uff0c\u7528\u4e8e\u9650\u5236\u4fe1\u4efb\u533a\u57df\u5e76\u589e\u5f3a RL \u7684\u7a33\u5b9a\u6027\u3002\u4e0a\u526a\u8f91\u53ef\u4ee5\u9650\u5236\u7b56\u7565\u7684\u63a2\u7d22\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u63d0\u9ad8\u300c\u5229\u7528 token\u300d\u7684\u6982\u7387\u8981\u6bd4\u63d0\u9ad8\u4e0d\u53ef\u80fd\u7684\u300c\u63a2\u7d22 token\u300d\u7684\u6982\u7387\u5bb9\u6613\u5f97\u591a\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-c2e88550d376f5d4cf1ddcc83387bec3.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-c2e88550d376f5d4cf1ddcc83387bec3.png\"><\/a><\/section>\n<section>\u4ed6\u4eec\u8fd8\u89c2\u5bdf\u5230\uff0c\u88ab\u526a\u8f91 token \u7684\u6700\u5927\u6982\u7387\u7ea6\u4e3a<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1b53662e228420b65002b28ceb699f28.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1b53662e228420b65002b28ceb699f28.png\"><\/a>\uff08\u56fe 3a\uff09\u3002\u8fd9\u4e00\u53d1\u73b0\u652f\u6301\u4e86\u4ed6\u4eec\u7684\u5206\u6790\uff0c\u5373\u4e0a\u9650\u526a\u8f91\u9608\u503c\u786e\u5b9e\u9650\u5236\u4e86\u4f4e\u6982\u7387 token \u7684\u6982\u7387\u589e\u957f\uff0c\u4ece\u800c\u53ef\u80fd\u9650\u5236\u4e86\u7cfb\u7edf\u7684\u591a\u6837\u6027\u3002<\/section>\n<section><\/section>\n<section>\u5982\u516c\u5f0f 10 \u6240\u793a\uff0c\u6839\u636e Clip-Higher \u7b56\u7565\uff0c\u7814\u7a76\u8005\u5c06\u8f83\u4f4e\u548c\u8f83\u9ad8\u7684\u526a\u8f91\u8303\u56f4\u89e3\u8026\u4e3a \u03b5_low \u548c \u03b5_high\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-a231fb1630f9777def5b093abd8a9734.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-a231fb1630f9777def5b093abd8a9734.png\"><\/a><\/section>\n<section>\u7814\u7a76\u8005\u589e\u52a0\u4e86 \u03b5_high \u7684\u503c\uff0c\u4ee5\u4fbf\u4e3a\u4f4e\u6982\u7387 token \u7684\u589e\u52a0\u7559\u51fa\u66f4\u591a\u7a7a\u95f4\u3002\u5982\u56fe 2 \u6240\u793a\uff0c\u8fd9\u4e00\u8c03\u6574\u6709\u6548\u5730\u63d0\u9ad8\u4e86\u7b56\u7565\u7684\u71b5\uff0c\u6709\u5229\u4e8e\u751f\u6210\u66f4\u591a\u6837\u5316\u7684\u6837\u672c\u3002\u7814\u7a76\u8005\u9009\u62e9\u5c06 \u03b5_low \u4fdd\u6301\u5728\u76f8\u5bf9\u8f83\u5c0f\u7684\u8303\u56f4\u5185\uff0c\u56e0\u4e3a\u589e\u5927 \u03b5_low \u4f1a\u5c06\u8fd9\u4e9b token \u7684\u6982\u7387\u538b\u5236\u4e3a 0\uff0c\u4ece\u800c\u5bfc\u81f4\u91c7\u6837\u7a7a\u95f4\u7684\u5d29\u6e83\u3002<\/section>\n<section><\/section>\n<section><strong>\u8d8a\u591a\u8d8a\u597d\uff1a\u52a8\u6001\u91c7\u6837<\/strong><\/section>\n<section><\/section>\n<section>\u5f53\u67d0\u4e9b\u63d0\u793a\u7684\u51c6\u786e\u5ea6\u7b49\u4e8e 1 \u65f6\uff0c\u73b0\u6709\u7684 RL \u7b97\u6cd5\u5c31\u4f1a\u51fa\u73b0\u68af\u5ea6\u9012\u51cf\u95ee\u9898\u3002\u6839\u636e\u7ecf\u9a8c\uff0c\u51c6\u786e\u7387\u7b49\u4e8e 1 \u7684\u6837\u672c\u6570\u91cf\u4f1a\u7ee7\u7eed\u589e\u52a0\uff0c\u5982\u56fe 3b \u6240\u793a\u3002\u8fd9\u610f\u5473\u7740\u6bcf\u6279\u6837\u672c\u4e2d\u7684\u6709\u6548\u63d0\u793a\u6b21\u6570\u4f1a\u4e0d\u65ad\u51cf\u5c11\uff0c\u4ece\u800c\u5bfc\u81f4\u68af\u5ea6\u65b9\u5dee\u589e\u5927\uff0c\u6291\u5236\u4e86\u6a21\u578b\u8bad\u7ec3\u7684\u68af\u5ea6\u4fe1\u53f7\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u6b64\uff0c\u7814\u7a76\u8005\u5efa\u8bae\u8fdb\u884c\u8fc7\u5ea6\u91c7\u6837\uff0c\u8fc7\u6ee4\u6389\u7b49\u5f0f 11 \u4e2d\u6240\u793a\u7cbe\u5ea6\u7b49\u4e8e 1 \u548c 0 \u7684\u63d0\u793a\u8bed\uff0c\u4fdd\u7559\u6279\u6b21\u4e2d\u6240\u6709\u5177\u6709\u6709\u6548\u68af\u5ea6\u7684\u63d0\u793a\u8bed\uff0c\u5e76\u4fdd\u6301\u4e00\u81f4\u7684\u63d0\u793a\u8bed\u6570\u91cf\u3002\u5728\u8bad\u7ec3\u4e4b\u524d\u4e0d\u65ad\u91c7\u6837\uff0c\u76f4\u5230\u6279\u6b21\u4e2d\u5168\u90e8\u90fd\u662f\u51c6\u786e\u7387\u65e2\u4e0d\u7b49\u4e8e 0 \u4e5f\u4e0d\u7b49\u4e8e 1 \u7684\u6837\u672c\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-5d17937af18e738a48fa2e4d5f66faf1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-5d17937af18e738a48fa2e4d5f66faf1.png\"><\/a><\/section>\n<section>\u53e6\u5916\u4e00\u70b9\u53d1\u73b0\u5982\u56fe 6 \u6240\u793a\uff0c\u5728\u52a8\u6001\u91c7\u6837\u7684\u60c5\u51b5\u4e0b\uff0c\u5b9e\u9a8c\u80fd\u66f4\u5feb\u5730\u5b9e\u73b0\u76f8\u540c\u7684\u6027\u80fd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-568e69bdee401c89eac1eb8ad0a8ee3c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-568e69bdee401c89eac1eb8ad0a8ee3c.png\"><\/a><\/section>\n<section><strong>Rebalancing Act\uff1aToken \u7ea7\u7b56\u7565\u68af\u5ea6\u635f\u5931<\/strong><\/section>\n<section><\/section>\n<section>\u7814\u7a76\u8005\u89c2\u5bdf\u5230\uff0c\u7531\u4e8e\u6240\u6709\u6837\u672c\u5728\u635f\u5931\u8ba1\u7b97\u4e2d\u7684\u6743\u91cd\u76f8\u540c\uff0c\u56e0\u6b64\u957f\u56de\u590d\u4e2d\u7684 token \u5bf9\u603b\u4f53\u635f\u5931\u7684\u8d21\u732e\u53ef\u80fd\u4f1a\u4e0d\u6210\u6bd4\u4f8b\u5730\u964d\u4f4e\uff0c\u8fd9\u53ef\u80fd\u4f1a\u5bfc\u81f4\u4e24\u79cd\u4e0d\u5229\u5f71\u54cd\u3002<\/section>\n<section><\/section>\n<section>\u9996\u5148\uff0c\u5bf9\u4e8e\u9ad8\u8d28\u91cf\u7684\u957f\u6837\u672c\u6765\u8bf4\uff0c\u8fd9\u79cd\u5f71\u54cd\u4f1a\u963b\u788d\u6a21\u578b\u5b66\u4e60\u5176\u4e2d\u4e0e\u63a8\u7406\u76f8\u5173\u7684\u6a21\u5f0f\u7684\u80fd\u529b\u3002\u5176\u6b21\uff0c\u8fc7\u957f\u7684\u6837\u672c\u5f80\u5f80\u8868\u73b0\u51fa\u4f4e\u8d28\u91cf\u7684\u6a21\u5f0f\uff0c\u5982\u80e1\u8a00\u4e71\u8bed\u548c\u91cd\u590d\u8bcd\u8bed\u3002<\/section>\n<section><\/section>\n<section>\u5982\u56fe 4a \u548c\u56fe 4b \u6240\u793a\uff0c\u6837\u672c\u7ea7\u635f\u5931\u8ba1\u7b97\u7531\u4e8e\u65e0\u6cd5\u6709\u6548\u60e9\u7f5a\u957f\u6837\u672c\u4e2d\u7684\u4e0d\u826f\u6a21\u5f0f\uff0c\u4f1a\u5bfc\u81f4\u71b5\u548c\u54cd\u5e94\u957f\u5ea6\u7684\u4e0d\u5065\u5eb7\u589e\u957f\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-4f3ae81bd47d80f013f60669d569cb6f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-4f3ae81bd47d80f013f60669d569cb6f.png\"><\/a><\/section>\n<section><strong>\u6349\u8ff7\u85cf\uff1a\u8fc7\u957f\u7684\u5956\u52b1\u5851\u9020<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u7814\u7a76\u5956\u52b1\u566a\u58f0\u7684\u5f71\u54cd\uff0c\u7814\u7a76\u8005\u9996\u5148\u5e94\u7528\u4e86\u8d85\u957f\u8fc7\u6ee4\u7b56\u7565\uff0c\u4ee5\u63a9\u76d6\u622a\u65ad\u6837\u672c\u7684\u635f\u5931\u3002\u5982\u56fe 5 \u6240\u793a\uff0c\u8fd9\u79cd\u65b9\u6cd5\u5927\u5927\u7a33\u5b9a\u4e86\u8bad\u7ec3\u5e76\u63d0\u9ad8\u4e86\u6027\u80fd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-7cde722f3328b0929ed02ed501195fb9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-7cde722f3328b0929ed02ed501195fb9.png\"><\/a><\/section>\n<section>\u6b64\u5916\uff0c\u4ed6\u4eec\u8fd8\u63d0\u51fa\u4e86\u300cSoft Overlong Punishment\u300d\uff08\u7b49\u5f0f 13\uff09\uff0c\u8fd9\u662f\u4e00\u79cd\u957f\u5ea6\u611f\u77e5\u60e9\u7f5a\u673a\u5236\uff0c\u65e8\u5728\u5851\u9020\u622a\u65ad\u6837\u672c\u7684\u5956\u52b1\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5f53\u54cd\u5e94\u957f\u5ea6\u8d85\u8fc7\u9884\u5b9a\u4e49\u7684\u6700\u5927\u503c\u65f6\uff0c\u7814\u7a76\u8005\u4f1a\u5b9a\u4e49\u4e00\u4e2a\u60e9\u7f5a\u533a\u95f4\u3002\u5728\u8fd9\u4e2a\u533a\u95f4\u5185\uff0c\u54cd\u5e94\u8d8a\u957f\uff0c\u53d7\u5230\u7684\u60e9\u7f5a\u5c31\u8d8a\u5927\u3002\u8fd9\u79cd\u60e9\u7f5a\u4f1a\u6dfb\u52a0\u5230\u57fa\u4e8e\u89c4\u5219\u7684\u539f\u59cb\u6b63\u786e\u6027\u5956\u52b1\u4e2d\uff0c\u4ece\u800c\u5411\u6a21\u578b\u53d1\u51fa\u4fe1\u53f7\uff0c\u907f\u514d\u8fc7\u957f\u7684\u54cd\u5e94\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-274ed664f7b84ebd767e661f8a9d5469.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-274ed664f7b84ebd767e661f8a9d5469.png\"><\/a><\/section>\n<section><strong>DAPO \u7684\u5b9e\u9a8c\u8868\u73b0<\/strong><\/section>\n<section><\/section>\n<section>\u57fa\u4e8e Qwen-32B \u57fa\u7840\u6a21\u578b\uff0c\u8be5\u56e2\u961f\u8fdb\u884c\u4e86\u4e00\u7cfb\u5217\u5b9e\u9a8c\uff0c\u9a8c\u8bc1\u4e86\u65b0\u63d0\u51fa\u7684 DAPO \u7b97\u6cd5\u7684\u6709\u6548\u6027\u548c\u4f18\u52bf\u3002\u8fd9\u91cc\u6211\u4eec\u7565\u8fc7\u5b9e\u9a8c\u7ec6\u8282\uff0c\u91cd\u70b9\u6765\u770b\u770b\u5b9e\u9a8c\u7ed3\u679c\u3002<\/section>\n<section><\/section>\n<section>\u6574\u4f53\u6765\u770b\uff0c\u5728 AIME 2024 \u4e0a\uff0c\u4f7f\u7528 DAPO \u8bad\u7ec3\u7684 Qwen-32B \u6a21\u578b\u6210\u957f\u4e3a\u4e86\u4e00\u4e2a\u5f3a\u5927\u7684\u63a8\u7406\u6a21\u578b\uff0c\u6027\u80fd\u4f18\u4e8e\u4f7f\u7528 R1 \u65b9\u6cd5\u8bad\u7ec3\u7684 Qwen2.5-32B\u3002<\/section>\n<section><\/section>\n<section>\u5982\u56fe 1 \u6240\u793a\uff0c\u53ef\u4ee5\u770b\u5230 DAPO \u8bad\u7ec3\u7684 Qwen2.5-32B \u5728 AIME 2024 \u57fa\u51c6\u4e0a\u7684\u6027\u80fd\u63d0\u5347\u60c5\u51b5\u3002\u968f\u7740\u8bad\u7ec3\u6b65\u6570\u589e\u957f\uff0c\u6a21\u578b\u51c6\u786e\u5ea6\u4ece 0% \u7a33\u6b65\u5347\u81f3\u4e86 50%\u3002\u9700\u8981\u91cd\u70b9\u6307\u51fa\uff1a\u8fbe\u6210\u8fd9\u4e00\u6027\u80fd\u6240\u4f7f\u7528\u7684\u6b65\u6570\u4ec5\u4e3a DeepSeek-R1-Zero-Qwen-32B \u6240\u9700\u6b65\u6570\u7684\u4e00\u534a\u3002<\/section>\n<section><\/section>\n<section>\u8868 1 \u5c55\u793a\u4e86\u65b0\u65b9\u6cd5\u4e2d\u6bcf\u79cd\u8bad\u7ec3\u6280\u672f\u7684\u8d21\u732e\u3002\u770b\u5f97\u51fa\u6765\uff0c\u6bcf\u79cd\u6280\u672f\u90fd\u5bf9\u51c6\u786e\u5ea6\u7684\u589e\u957f\u6709\u6240\u8d21\u732e\u3002\u53ef\u4ee5\u770b\u5230\uff0c\u539f\u751f GRPO \u53ea\u80fd\u8ba9 Qwen2.5-32B \u57fa\u7840\u6a21\u578b\u7684\u51c6\u786e\u5ea6\u8fbe\u5230 30%\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-79391d383b66f4a0e3e2ff2e1a3c185b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-79391d383b66f4a0e3e2ff2e1a3c185b.png\"><\/a><\/section>\n<section>\u81f3\u4e8e token \u7ea7\u635f\u5931\uff0c\u867d\u7136\u5b83\u5e26\u6765\u7684\u6027\u80fd\u63d0\u5347\u8f83\u5c11\uff0c\u4f46\u8be5\u56e2\u961f\u53d1\u73b0\u5b83\u53ef\u589e\u5f3a\u8bad\u7ec3\u7a33\u5b9a\u6027\u5e76\u4f7f\u957f\u5ea6\u589e\u52a0\u5f97\u66f4\u5065\u5eb7\u3002<\/section>\n<section><\/section>\n<section><strong>\u8bad\u7ec3\u52a8\u6001<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u83b7\u5f97\u66f4\u900f\u5f7b\u7684\u5206\u6790\uff0c\u8be5\u56e2\u961f\u4e5f\u5206\u6790\u4e86\u8bad\u7ec3\u52a8\u6001\u548c\u4e2d\u95f4\u7ed3\u679c\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-8890e6a44f8077d1dde53b32f386ed81.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-8890e6a44f8077d1dde53b32f386ed81.png\"><\/a><\/section>\n<section>\u751f\u6210\u54cd\u5e94\u7684\u957f\u5ea6\uff1a\u8be5\u6307\u6807\u4e0e\u8bad\u7ec3\u7a33\u5b9a\u6027\u548c\u6027\u80fd\u5bc6\u5207\u76f8\u5173\u3002\u5982\u56fe 7a \u6240\u793a\u3002\u957f\u5ea6\u7684\u589e\u52a0\u53ef\u4e3a\u6a21\u578b\u63d0\u4f9b\u66f4\u5927\u7684\u63a2\u7d22\u7a7a\u95f4\uff0c\u5141\u8bb8\u91c7\u6837\u66f4\u590d\u6742\u7684\u63a8\u7406\u884c\u4e3a\u5e76\u901a\u8fc7\u8bad\u7ec3\u9010\u6e10\u5f3a\u5316\u3002\u4f46\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u957f\u5ea6\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5e76\u4e0d\u603b\u662f\u4fdd\u6301\u6301\u7eed\u7684\u4e0a\u5347\u8d8b\u52bf\u3002\u5728\u4e00\u4e9b\u76f8\u5f53\u957f\u7684\u65f6\u671f\u5185\uff0c\u5b83\u53ef\u4ee5\u505c\u6ede\u751a\u81f3\u4e0b\u964d\u3002\u901a\u5e38\u7684\u505a\u6cd5\u662f\u5c06\u957f\u5ea6\u4e0e\u9a8c\u8bc1\u51c6\u786e\u5ea6\u7ed3\u5408\u8d77\u6765\u4f5c\u4e3a\u8bc4\u4f30\u5b9e\u9a8c\u662f\u5426\u6b63\u5728\u6076\u5316\u7684\u6307\u6807\u3002<\/section>\n<section><\/section>\n<section>\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u5956\u52b1\u52a8\u6001\uff1a\u8fd9\u4e00\u76f4\u662f\u5f3a\u5316\u5b66\u4e60\u4e2d\u81f3\u5173\u91cd\u8981\u7684\u76d1\u6d4b\u6307\u6807\u4e4b\u4e00\uff0c\u5982\u56fe 7b \u6240\u793a\u3002\u5728\u8fd9\u91cc\u7684\u5927\u591a\u6570\u5b9e\u9a8c\u4e2d\uff0c\u5956\u52b1\u589e\u52a0\u7684\u8d8b\u52bf\u76f8\u5bf9\u7a33\u5b9a\uff0c\u4e0d\u4f1a\u56e0\u4e3a\u5b9e\u9a8c\u8bbe\u7f6e\u7684\u8c03\u6574\u800c\u51fa\u73b0\u5927\u5e45\u6ce2\u52a8\u6216\u4e0b\u964d\u3002\u8fd9\u8868\u660e\uff0c\u5728\u7ed9\u5b9a\u53ef\u9760\u5956\u52b1\u4fe1\u53f7\u7684\u60c5\u51b5\u4e0b\uff0c\u8bed\u8a00\u6a21\u578b\u53ef\u4ee5\u7a33\u5065\u5730\u62df\u5408\u8bad\u7ec3\u96c6\u7684\u5206\u5e03\u3002\u7136\u800c\uff0c\u8be5\u56e2\u961f\u53d1\u73b0\uff0c\u5728\u8bad\u7ec3\u96c6\u4e0a\u7684\u6700\u7ec8\u5956\u52b1\u5f80\u5f80\u4e0e\u5728\u9a8c\u8bc1\u96c6\u4e0a\u7684\u51c6\u786e\u5ea6\u76f8\u5173\u6027\u4e0d\u5927\uff0c\u8fd9\u8868\u660e\u5bf9\u8bad\u7ec3\u96c6\u5b58\u5728\u8fc7\u62df\u5408\u73b0\u8c61\u3002<\/section>\n<section>\u00a0<\/section>\n<section>Actor \u6a21\u578b\u7684\u71b5\u548c\u751f\u6210\u6982\u7387\uff1a\u8fd9\u4e0e\u6a21\u578b\u7684\u63a2\u7d22\u80fd\u529b\u6709\u5173\uff0c\u540c\u65f6\u4e5f\u662f\u5b9e\u9a8c\u4e2d\u5bc6\u5207\u76d1\u63a7\u7684\u5173\u952e\u6307\u6807\u3002\u76f4\u89c2\u5730\u8bb2\uff0c\u6a21\u578b\u7684\u71b5\u9700\u8981\u4fdd\u6301\u5728\u9002\u5f53\u7684\u8303\u56f4\u5185\u3002\u8fc7\u4f4e\u7684\u71b5\u8868\u793a\u6982\u7387\u5206\u5e03\u8fc7\u4e8e\u5c16\u9510\uff0c\u8fd9\u4f1a\u5bfc\u81f4\u63a2\u7d22\u80fd\u529b\u4e27\u5931\u3002\u76f8\u53cd\uff0c\u8fc7\u9ad8\u7684\u71b5\u5f80\u5f80\u4e0e\u8fc7\u5ea6\u63a2\u7d22\u7684\u95ee\u9898\u6709\u5173\uff0c\u4f8b\u5982\u4e71\u7801\u548c\u91cd\u590d\u751f\u6210\u3002\u5bf9\u4e8e\u751f\u6210\u6982\u7387\uff0c\u60c5\u51b5\u6070\u6070\u76f8\u53cd\u3002\u5982\u524d\u6240\u793a\uff0c\u901a\u8fc7\u5e94\u7528 Clip-Higher \u7b56\u7565\uff0c\u53ef\u4ee5\u6709\u6548\u5730\u89e3\u51b3\u71b5\u5d29\u6e83\u7684\u95ee\u9898\u3002\u5728\u540e\u7eed\u7684\u5b9e\u9a8c\u4e2d\uff0c\u8be5\u56e2\u961f\u8fd8\u53d1\u73b0\u4fdd\u6301\u71b5\u7f13\u6162\u4e0a\u5347\u7684\u8d8b\u52bf\u6709\u5229\u4e8e\u6a21\u578b\u63d0\u5347\u6027\u80fd\uff0c\u5982\u56fe 7c \u548c\u56fe 7d \u6240\u793a\u3002<\/section>\n<section><\/section>\n<section><strong>\u6848\u4f8b\u7814\u7a76<\/strong><\/section>\n<section><\/section>\n<section>\u5728 RL \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u7814\u7a76\u8005\u89c2\u5bdf\u5230\u4e00\u4e2a\u6709\u8da3\u7684\u73b0\u8c61\uff1aActor \u6a21\u578b\u7684\u63a8\u7406\u6a21\u5f0f\u4f1a\u968f\u7740\u65f6\u95f4\u7684\u63a8\u79fb\u800c\u52a8\u6001\u6f14\u5316\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u7b97\u6cd5\u4e0d\u4ec5\u5f3a\u5316\u4e86\u6709\u52a9\u4e8e\u6b63\u786e\u89e3\u51b3\u95ee\u9898\u7684\u73b0\u6709\u63a8\u7406\u6a21\u5f0f\uff0c\u8fd8\u9010\u6e10\u4ea7\u751f\u4e86\u6700\u521d\u4e0d\u5b58\u5728\u7684\u5168\u65b0\u63a8\u7406\u6a21\u5f0f\u3002\u8fd9\u4e00\u53d1\u73b0\u63ed\u793a\u4e86 RL \u7b97\u6cd5\u7684\u9002\u5e94\u6027\u548c\u63a2\u7d22\u80fd\u529b\uff0c\u5e76\u4e3a\u6a21\u578b\u7684\u5b66\u4e60\u673a\u5236\u63d0\u4f9b\u4e86\u65b0\u7684\u89c1\u89e3\u3002<\/section>\n<section><\/section>\n<section>\u4f8b\u5982\uff0c\u5728\u6a21\u578b\u8bad\u7ec3\u7684\u65e9\u671f\u9636\u6bb5\uff0c\u51e0\u4e4e\u4e0d\u5b58\u5728\u5bf9\u4e4b\u524d\u63a8\u7406\u6b65\u9aa4\u7684\u68c0\u67e5\u548c\u53cd\u601d\u3002\u7136\u800c\uff0c\u968f\u7740\u8bad\u7ec3\u7684\u8fdb\u884c\uff0c\u6a21\u578b\u8868\u73b0\u51fa\u660e\u663e\u7684\u53cd\u601d\u548c\u56de\u6eaf\u884c\u4e3a\uff0c\u5982\u8868 2 \u6240\u793a\u3002\u8fd9\u4e00\u89c2\u5bdf\u7ed3\u679c\u4e3a\u8fdb\u4e00\u6b65\u63a2\u7d22\u89e3\u91ca\u63a8\u7406\u80fd\u529b\u5728 RL \u8fc7\u7a0b\u4e2d\u7684\u51fa\u73b0\u63d0\u4f9b\u4e86\u542f\u793a\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-046653e647157e926991a70fe7a73be2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-046653e647157e926991a70fe7a73be2.png\"><\/a><\/section>\n<section>\u66f4\u591a\u7814\u7a76\u7ec6\u8282\uff0c\u53ef\u53c2\u8003\u539f\u8bba\u6587\u3002<\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-18-9\" target=\"_blank\">\u8d85\u8d8aDeepSeek GRPO\u7684\u5173\u952eRL\u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534eAIR\u5f00\u6e90DAPO<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u8d85\u8d8aDeepSeek  [&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":[81],"class_list":["post-39505","post","type-post","status-publish","format-standard","hentry","category-news","tag-81"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>\u8d85\u8d8aDeepSeek GRPO\u7684\u5173\u952eRL\u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534eAIR\u5f00\u6e90DAPO - \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=39505\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39505#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39505\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"\u8d85\u8d8aDeepSeek GRPO\u7684\u5173\u952eRL\u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534eAIR\u5f00\u6e90DAPO\",\"datePublished\":\"2025-03-18T20:02:22+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39505\"},\"wordCount\":194,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39505#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/03\\\/frc-ef623e220fdaceda3039115e21bd85bb.png\",\"keywords\":[\"\u6a21\u578b\u8bad\u7ec3\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=39505#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39505\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=39505\",\"name\":\"\u8d85\u8d8aDeepSeek GRPO\u7684\u5173\u952eRL\u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534eAIR\u5f00\u6e90DAPO - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39505#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39505#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/03\\\/frc-ef623e220fdaceda3039115e21bd85bb.png\",\"datePublished\":\"2025-03-18T20:02:22+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39505#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=39505\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39505#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/03\\\/frc-ef623e220fdaceda3039115e21bd85bb.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/03\\\/frc-ef623e220fdaceda3039115e21bd85bb.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=39505#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u8d85\u8d8aDeepSeek GRPO\u7684\u5173\u952eRL\u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534eAIR\u5f00\u6e90DAPO\"}]},{\"@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":"\u8d85\u8d8aDeepSeek GRPO\u7684\u5173\u952eRL\u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534eAIR\u5f00\u6e90DAPO - \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=39505","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=39505#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=39505"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"\u8d85\u8d8aDeepSeek GRPO\u7684\u5173\u952eRL\u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534eAIR\u5f00\u6e90DAPO","datePublished":"2025-03-18T20:02:22+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=39505"},"wordCount":194,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=39505#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ef623e220fdaceda3039115e21bd85bb.png","keywords":["\u6a21\u578b\u8bad\u7ec3"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=39505#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=39505","url":"https:\/\/17aitech.com\/?p=39505","name":"\u8d85\u8d8aDeepSeek GRPO\u7684\u5173\u952eRL\u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534eAIR\u5f00\u6e90DAPO - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=39505#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=39505#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ef623e220fdaceda3039115e21bd85bb.png","datePublished":"2025-03-18T20:02:22+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=39505#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=39505"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=39505#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ef623e220fdaceda3039115e21bd85bb.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-ef623e220fdaceda3039115e21bd85bb.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=39505#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u8d85\u8d8aDeepSeek GRPO\u7684\u5173\u952eRL\u7b97\u6cd5\uff0c\u5b57\u8282\u3001\u6e05\u534eAIR\u5f00\u6e90DAPO"}]},{"@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\/39505","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=39505"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39505\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=39505"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=39505"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=39505"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}