{"id":751035,"date":"2026-03-20T20:30:40","date_gmt":"2026-03-20T12:30:40","guid":{"rendered":"https:\/\/www.alphaseek.cn\/?p=751035"},"modified":"2026-03-20T20:30:40","modified_gmt":"2026-03-20T12:30:40","slug":"%e9%b2%81%e6%a3%92rl%e8%b5%8b%e8%83%bdai%e7%bc%96%e7%a8%8b%ef%bc%81%e7%a0%b4%e5%b1%80%e4%bc%81%e4%b8%9a%e6%95%b0%e6%8d%ae%e5%99%aa%e5%a3%b0%e9%9a%be%e9%a2%98%ef%bc%8c%e5%90%8c%e7%ad%89%e7%ae%97","status":"publish","type":"post","link":"https:\/\/www.alphaseek.cn\/?p=751035","title":{"rendered":"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &#8211; \u91cf\u5b50\u4f4d"},"content":{"rendered":"<h1>\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b<\/h1>\n<div>              <span><img src='http:\/\/www.qbitai.com\/wp-content\/uploads\/2019\/06\/200-100x100.jpg' height='100' width='100' alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><em><a href=\"https:\/\/www.qbitai.com\/author\/qbitai\" title=\"\u7531 \u91cf\u5b50\u4f4d\u7684\u670b\u53cb\u4eec \u53d1\u5e03\" rel=\"author\">\u91cf\u5b50\u4f4d\u7684\u670b\u53cb\u4eec<\/a><\/em><\/span>                           <span>2026-02-16<\/span>              <span>19:30:47<\/span>           <span>           \u6765\u6e90\uff1a<a href='https:\/\/www.qbitai.com' target='blank' rel=\"nofllow noopener\">\u91cf\u5b50\u4f4d<\/a>            <\/span><\/div>\n<div>\n<p>\u8ba9\u566a\u58f0\u4ece\u300c\u5305\u88b1\u300d\u53d8\u300c\u71c3\u6599\u300d<\/p>\n<\/p><\/div>\n<p data-track=\"2\">\u7a0b\u5e8f\u5458\u4eec\u53c8\u80fd\u5c11\u6389\u5934\u53d1\u4e86\uff01<\/p>\n<p data-track=\"3\">\u65b0\u7814\u7a76\u901a\u8fc7\u8fc7\u6ee4\u6389\u8bad\u7ec3\u4e2d\u7684\u566a\u58f0\u548c\u5f02\u5e38\u503c\uff0c\u663e\u8457\u63d0\u5347\u4ee3\u7801\u5927\u6a21\u578b\u5728\u5b9e\u9645\u7f16\u8f91\u4efb\u52a1\u4e2d\u7684\u51c6\u786e\u6027\u548c\u6548\u7387\u3002<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/d99d60eaf79aa5ab9ccf8db23be73290.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<\/div>\n<p data-track=\"4\">\u5728AI\u8f85\u52a9\u7f16\u7a0b\u6210\u4e3a\u8f6f\u4ef6\u5f00\u53d1\u6838\u5fc3\u751f\u4ea7\u529b\u7684\u4eca\u5929\uff0c\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5df2\u6df1\u5ea6\u878d\u5165\u4ee3\u7801\u7f16\u8f91\u3001\u8c03\u8bd5\u4e0e\u4f18\u5316\u5168\u6d41\u7a0b\u3002<\/p>\n<p data-track=\"5\">\u7136\u800c\uff0c\u5f53\u4f01\u4e1a\u8bd5\u56fe\u7528<strong>\u771f\u5b9e\u590d\u6742\u7528\u6237\u73af\u5883\u4e2d\u91c7\u96c6\u7684\u6570\u636e<\/strong>\u5f00\u5c55\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u8bad\u7ec3\u65f6\uff0c\u4e00\u4e2a\u68d8\u624b\u7684\u5b9e\u9645\u95ee\u9898\u6d6e\u51fa\u6c34\u9762\uff1a\u590d\u6742\u4e0a\u4e0b\u6587\uff08context\uff09\u5bfc\u81f4\u5927\u6a21\u578b\u7684\u8f93\u51fa\u7b54\u6848\u9891\u7e41\u51fa\u73b0\u5f02\u5e38\u5185\u5bb9\uff0c\u5373rollout\u566a\u58f0\u66f4\u666e\u904d\uff0c\u4f7f\u5f97reward\u51fa\u73b0\u5f02\u5e38\u503c\uff08outliers\uff09\uff0c\u76f4\u63a5\u9020\u6210\u4f18\u52bf\u503c\uff08advantage\uff09\u4f30\u8ba1\u4e0d\u51c6\u786e\uff0c\u4e25\u91cd\u62d6\u7d2f\u5f3a\u5316\u5b66\u4e60\u6548\u679c\u3002<\/p>\n<p data-track=\"6\">\u4e0a\u6d77\u4ea4\u901a\u5927\u5b66\u3001\u817e\u8bafCodeBuddy\u7b49\u56e2\u961f\u8054\u5408\u63d0\u51fa\u7684<strong>Group Adaptive Policy Optimization\uff08GAPO\uff09<\/strong>\u65b9\u6cd5\uff0c\u7cbe\u51c6\u76f4\u51fb\u8fd9\u4e00\u4ea7\u4e1a\u843d\u5730\u5173\u952e\u74f6\u9888\uff0c\u4e3a\u4ee3\u7801LLM\u7684\u5de5\u4e1a\u5316\u8bad\u7ec3\u63d0\u4f9b\u4e86\u517c\u5177\u79d1\u7814\u521b\u65b0\u6027\u4e0e\u5de5\u7a0b\u5b9e\u7528\u6027\u7684\u7a81\u7834\u65b9\u6848\uff0c\u5f15\u53d1AI\u79d1\u7814\u754c\u4e0e\u4ea7\u4e1a\u754c\u5e7f\u6cdb\u5173\u6ce8\u3002<\/p>\n<h1 data-track=\"7\">\u771f\u5b9e\u573a\u666f\u7684\u6838\u5fc3\u6897\u963b\uff1a\u590d\u6742\u4e0a\u4e0b\u6587\u2192rollout\u566a\u58f0\u2192\u4f18\u52bf\u4f30\u8ba1\u5931\u771f<\/h1>\n<p data-track=\"8\">\u4ee3\u7801\u7f16\u8f91\u7684\u6838\u5fc3\u96be\u70b9\u5728\u4e8e\uff0c\u771f\u5b9e\u7528\u6237\u573a\u666f\u7684\u8f93\u5165\u63d0\u793a\u7edd\u975e\u7b80\u5355\u7684\u4ee3\u7801\u7247\u6bb5\uff0c\u800c\u662f\u5305\u542b<strong>\u590d\u6742\u4e0a\u4e0b\u6587\u4fe1\u606f<\/strong>\u7684\u7efc\u5408\u6307\u4ee4\u2014\u2014\u65e2\u6db5\u76d6\u4ee3\u7801\u7684\u6a21\u5757\u8c03\u7528\u5173\u7cfb\u3001\u5386\u53f2\u7f16\u8f91\u8bb0\u5f55\u3001\u6307\u5b9a\u4fee\u6539\u8303\u56f4\uff0c\u8fd8\u5305\u542b\u7528\u6237\u6a21\u7cca\u7684\u9700\u6c42\u6697\u793a\u3002<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/011aefb9e56178160f4e0a6215ce0a6a.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<p>\u8f93\u5165\u7684Prompt\u7ed3\u6784<\/p>\n<\/p><\/div>\n<p data-track=\"10\">\u8fd9\u79cd\u590d\u6742\u4e0a\u4e0b\u6587\u5e26\u6765\u7684\u76f4\u63a5\u540e\u679c\u662f\uff1a\u5927\u6a21\u578b\u7684\u8f93\u51fa\uff08rollout\uff09\u6781\u6613\u51fa\u73b0\u5f02\u5e38\u503c\uff08outliers\uff09\uff0c\u4e14\u6b64\u7c7b\u566a\u58f0\u5728\u771f\u5b9e\u6570\u636e\u4e2d\u5e76\u975e\u4e2a\u4f8b\uff0c\u800c\u662f\u5e38\u6001\u3002<\/p>\n<p data-track=\"59\"><strong>1. \u590d\u6742\u4e0a\u4e0b\u6587\u4e3a\u4f55\u50ac\u751f\u5927\u91cfrollout\u566a\u58f0\uff1f<\/strong><\/p>\n<p data-track=\"12\">\u771f\u5b9e\u573a\u666f\u7684\u4ee3\u7801\u7f16\u8f91\u4efb\u52a1\u663e\u793a\uff0c\u8f93\u5165\u63d0\u793a\u957f\u5ea6\u8de8\u5ea6\u4ece1925\u523024883\u5b57\u7b26\uff0c\u8f93\u51fa\u7f16\u8f91\u957f\u5ea6\u4ece36\u5230833\u5b57\u7b26\uff0c\u4e14\u9700\u517c\u5bb9Go\u3001Python\u3001Java\u7b4910\u79cd\u4e3b\u6d41\u7f16\u7a0b\u8bed\u8a00\u3002<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/e51bdfdca27cc9d648aac305d8a3fbfb.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<p>\u771f\u5b9e\u4e16\u754c\u4ee3\u7801\u7f16\u8f91\u4efb\u52a1\u7684\u8bad\u7ec3\u6570\u636e\u7edf\u8ba1\u3002\u201cInput\u201d\u548c\u201cOutput\u201d\u5206\u522b\u8868\u793a\u8f93\u5165\u63d0\u793a\u7b26\u548c\u8f93\u51fa\u771f\u5b9e\u6587\u672c\u957f\u5ea6\u8303\u56f4\u3002<\/p>\n<\/p><\/div>\n<p data-track=\"14\">\u66f4\u5173\u952e\u7684\u662f\uff0c\u6bcf\u4e2a\u4efb\u52a1\u90fd\u5305\u542b\u591a\u5c42\u7ea7\u4fe1\u606f\uff1a\u6bd4\u5982Go\u8bed\u8a00\u4efb\u52a1\u4e2d\uff0c\u53ef\u80fd\u6d89\u53ca\u8de8\u6587\u4ef6\u6a21\u5757\u8c03\u7528\u3001\u591a\u8f6e\u7f16\u8f91\u5386\u53f2\u7684\u5173\u8054\uff0c\u4ee5\u53ca\u7528\u6237\u672a\u660e\u786e\u8bf4\u660e\u4f46\u9690\u542b\u7684\u6027\u80fd\u4f18\u5316\u9700\u6c42\u3002<\/p>\n<p data-track=\"15\">\u8fd9\u79cd\u590d\u6742\u4e0a\u4e0b\u6587\u5bf9\u6a21\u578b\u7684\u7406\u89e3\u80fd\u529b\u63d0\u51fa\u6781\u9ad8\u8981\u6c42\uff0c\u8fd9\u662f\u6a21\u578b\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6ca1\u89c1\u8fc7\u7684\u6570\u636e\u533a\u57df\uff0c\u5bfc\u81f4\u6a21\u578b\u8f93\u51fa\u4e0d\u786e\u5b9a\u6027\u589e\u52a0\uff0c\u65f6\u5e38\u51fa\u73b0\u201c\u504f\u79bb\u9700\u6c42\u7684\u5f02\u5e38\u7f16\u8f91\u201d\u2014\u2014\u53ef\u80fd\u662f\u9057\u6f0f\u5173\u952e\u903b\u8f91\u3001\u4fee\u6539\u65e0\u5173\u4ee3\u7801\uff0c\u751a\u81f3\u751f\u6210\u8bed\u6cd5\u9519\u8bef\u7247\u6bb5\uff0c\u8fd9\u4e9b\u5747\u5c5e\u4e8erollout\u566a\u58f0\uff08outliers\uff09\u3002<\/p>\n<p data-track=\"16\"><strong>2. Reward\u566a\u58f0\u5bfc\u81f4\u4f18\u52bf\u503c\u4f30\u8ba1\u5931\u7075\uff0c\u4f7fRL\u4fe1\u53f7\u5931\u771f<\/strong><\/p>\n<p data-track=\"17\">\u4f20\u7edf\u5206\u7ec4\u76f8\u5bf9RL\u65b9\u6cd5\uff08\u5982GRPO\u3001DAPO\uff09\u7684\u4f18\u52bf\u503c\u8ba1\u7b97\u4f9d\u8d56\u7fa4\u4f53\u5747\u503c\uff1a<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/5315857efe60e8bd0942cfc1b9803269.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<\/div>\n<p data-track=\"18\">\u6b63\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u771f\u5b9e\u6570\u636e\u7684\u5956\u52b1\u5206\u5e03\u4e2d\uff0c14.4%\u5448\u73b0\u53f3\u504f\u300115.5%\u5448\u73b0\u5de6\u504f\u7279\u5f81\u2014\u2014\u8fd9\u79cd\u504f\u659c\u6b63\u662foutliers\u4e3b\u5bfc\u7684\u7ed3\u679c\uff1a\u5de6\u504f\u5206\u5e03\u4e2d\uff0c\u4f4e\u5956\u52b1\u533a\u57df\u7684outliers\u62c9\u4f4e\u5747\u503c\uff0c\u5bfc\u81f4\u4f18\u52bf\u503c\u88ab\u9ad8\u4f30\uff1b\u53f3\u504f\u5206\u5e03\u4e2d\uff0c\u9ad8\u5956\u52b1\u533a\u57df\u7684outliers\u62ac\u9ad8\u5747\u503c\uff0c\u5bfc\u81f4\u591a\u6570\u6709\u6548\u8f93\u51fa\u7684\u4f18\u52bf\u503c\u88ab\u4f4e\u4f30\u3002<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/58b040984d5d4fe50318e9ee32b8b5cc.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<p>\u5728Qwen2.5-Coder-14B\u4e0a\u8bad\u7ec3\u7684reward\u5206\u5e03\u7edf\u8ba1<\/p>\n<\/p><\/div>\n<p data-track=\"20\">\u66f4\u4e25\u5cfb\u7684\u662f\uff0c\u590d\u6742\u4e0a\u4e0b\u6587\u7684\u591a\u6837\u6027\u8ba9\u566a\u58f0\u5206\u5e03\u6beb\u65e0\u89c4\u5f8b\uff1a<\/p>\n<p data-track=\"21\">\u4e0d\u540c\u7f16\u7a0b\u8bed\u8a00\u3001\u4e0d\u540c\u7f16\u8f91\u573a\u666f\uff08\u5982bug\u4fee\u590d\u3001\u529f\u80fd\u65b0\u589e\uff09\u7684outliers\u7279\u5f81\u6709\u5f88\u5927\u5dee\u5f02\uff0c\u4f20\u7edf\u65b9\u6cd5\u7684\u56fa\u5b9a\u5747\u503c\u57fa\u51c6\u65e0\u6cd5\u9002\u914d\u8fd9\u79cd\u52a8\u6001\u566a\u58f0\u3002<\/p>\n<p data-track=\"22\">\u5bf9\u4f01\u4e1a\u800c\u8a00\uff0c\u8fd9\u610f\u5473\u7740\u6295\u5165\u5927\u91cf\u8d44\u6e90\u91c7\u96c6\u7684\u771f\u5b9e\u6570\u636e\uff0c\u4e0d\u4ec5\u96be\u4ee5\u63d0\u5347\u6a21\u578b\u6548\u679c\uff0c\u53cd\u800c\u53ef\u80fd\u56e0\u4f18\u52bf\u4f30\u8ba1\u5931\u771f\u5bfc\u81f4\u6a21\u578b\u8bad\u7ec3\u201c\u8d8a\u8bad\u8d8a\u504f\u201d\uff0c\u5b9e\u9645\u90e8\u7f72\u4e2d\u4ecd\u9700\u5927\u91cf\u4eba\u5de5\u6821\u9a8c\uff0c\u8fdd\u80cc\u4e86AI\u8f85\u52a9\u7f16\u7a0b\u201c\u964d\u672c\u589e\u6548\u201d\u7684\u6838\u5fc3\u8bc9\u6c42\uff1b\u5bf9\u79d1\u7814\u9886\u57df\u800c\u8a00\uff0c\u8fd9\u79cd\u771f\u5b9e\u573a\u666f\u7279\u6709\u7684\u201c\u590d\u6742\u4e0a\u4e0b\u6587-\u566a\u58f0-\u4f18\u52bf\u4f30\u8ba1\u201d\u8fde\u9501\u95ee\u9898\uff0c\u957f\u671f\u7f3a\u4e4f\u9488\u5bf9\u6027\u7a81\u7834\u65b9\u6848\uff0c\u6210\u4e3a\u5236\u7ea6\u4ee3\u7801LLM\u843d\u5730\u7684\u5173\u952e\u74f6\u9888\u3002<\/p>\n<h1 data-track=\"23\">GAPO\u7684\u521b\u65b0\u7a81\u7834\uff1a\u7cbe\u51c6\u964d\u566a+\u7a33\u5065\u4f30\u8ba1\uff0c\u7a81\u7834\u771f\u5b9e\u6570\u636e\u8bad\u7ec3\u56f0\u5c40<\/h1>\n<p data-track=\"24\">\u9488\u5bf9\u201c\u590d\u6742\u4e0a\u4e0b\u6587\u2192rollout\u566a\u58f0\u2192\u4f18\u52bf\u503c\u4f30\u8ba1\u4e0d\u51c6\u201d\u7684\u6838\u5fc3\u75db\u70b9\uff0cGAPO\u5728\u4e0d\u6539\u53d8\u73b0\u6709\u5206\u7ec4\u76f8\u5bf9RL\u6846\u67b6\u6838\u5fc3\u76ee\u6807\u7684\u524d\u63d0\u4e0b\uff0c\u4ec5\u901a\u8fc7\u4f18\u5316\u4f18\u52bf\u8ba1\u7b97\u73af\u8282\uff0c\u5b9e\u73b0\u201c\u5373\u63d2\u5373\u7528\u201d\u7684\u9ad8\u6548\u7a81\u7834\uff0c\u5176\u539f\u7406\u53ef\u901a\u8fc7\u8bba\u6587\u5173\u952e\u56fe\u8868\u76f4\u89c2\u5448\u73b0\uff1a<\/p>\n<p data-track=\"25\"><strong>1. \u7b2c\u4e00\u6b65\uff1a\u9501\u5b9a\u9ad8\u4fe1\u566a\u6bd4\u533a\u57df\uff0c\u7cbe\u51c6\u8fc7\u6ee4outliers<\/strong><\/p>\n<p data-track=\"26\">\u5982\u4e0b\u56fe\u6240\u793a\uff0cGAPO\u7684\u6838\u5fc3\u521b\u65b0\u662f\u5148\u4ece\u542b\u566a\u5956\u52b1\u5206\u5e03\u4e2d\uff0c\u7b5b\u9009\u51fa\u4e0d\u53d7outliers\u5e72\u6270\u7684\u201c\u6709\u6548\u4fe1\u53f7\u533a\u201d\u3002\u5b83\u5c06\u8fd9\u4e00\u95ee\u9898\u8f6c\u5316\u4e3a\u7ecf\u5178\u7684\u67e5\u627e\u6700\u9ad8\u5bc6\u5ea6\u533a\u95f4\uff08HDI\uff09\u6c42\u89e3\uff0c\u901a\u8fc7\u6ed1\u52a8\u7a97\u53e3\u7b97\u6cd5\uff08\u8bba\u6587\u7b97\u6cd51\uff09\uff0c\u5728\u6bcf\u4e2a\u8f93\u5165\u63d0\u793a\u7684\u5956\u52b1\u96c6\u5408\u4e2d\uff0c\u627e\u5230\u8986\u76d6\u03c4\u6bd4\u4f8b\uff08\u9ed8\u8ba40.5\uff09\u5956\u52b1\u70b9\u7684\u6700\u7a84\u533a\u95f4\u2014\u2014\u8fd9\u4e00\u533a\u95f4\u6b63\u662f\u6700\u9ad8\u4fe1\u566a\u6bd4\uff08SNR\uff09\u533a\u57df\uff0c\u80fd\u6700\u5927\u7a0b\u5ea6\u6392\u9664\u590d\u6742\u4e0a\u4e0b\u6587\u50ac\u751f\u7684outliers\uff08\u5982\u5de6\u504f\u5206\u5e03\u4e2d\u7684\u4f4e\u5956\u52b1\u5f02\u5e38\u503c\u3001\u53f3\u504f\u5206\u5e03\u4e2d\u7684\u9ad8\u5956\u52b1\u5f02\u5e38\u503c\uff09\u3002<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/a1bd16de7536d9ae23c5c12ae85b5672.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<p>GRPO\/DAPO\u4e0e\u56e2\u961f\u63d0\u51fa\u7684GAPO\u7b97\u6cd5\u793a\u610f\u56fe\u3002<\/p>\n<\/p><\/div>\n<p data-track=\"28\">\u4e0a\u56fe\u4e2dr\u1d62\u548c\u00c2(i,t)\u5206\u522b\u8868\u793a\u7b2ci\u4e2a\u5956\u52b1\uff0c\u4ee5\u53ca\u7b2ct\u6b65\u65f6\u4e00\u4e2a\u7ec4\u5185\u7684\u4f18\u52bf\u503c\u3002Q\u209c\u88ab\u5b9a\u4e49\u4e3a\u81ea\u9002\u5e94\u6700\u9ad8\u5bc6\u5ea6\u533a\u95f4\uff08Highest-Density Interval\uff0cHDI\uff09\u7684\u4e2d\u4f4d\u6570\uff0c\u8be5\u533a\u95f4\u7531\u6bcf\u4e2a\u63d0\u793a\u5728rollout\u8fc7\u7a0b\u4e2d\u7684\u5956\u52b1\u5206\u5e03\u63a8\u5bfc\u5f97\u5230\u3002<\/p>\n<p data-track=\"29\"><strong>2. \u7b2c\u4e8c\u6b65\uff1a\u7528\u4e2d\u4f4d\u6570\u66ff\u4ee3\u5747\u503c\uff0c\u7a33\u5065\u4f30\u8ba1\u4f18\u52bf\u503c<\/strong><\/p>\n<p data-track=\"30\">\u5728\u7b5b\u9009\u51fa\u7684HDI\u533a\u95f4\u5185\uff0cGAPO\u91c7\u7528\u4e2d\u4f4d\u6570\u800c\u975e\u5747\u503c\u4f5c\u4e3a\u81ea\u9002\u5e94Q\u503c\uff1a<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/5d9e2c656395ce5a6f67a1e65e0bade8.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<\/div>\n<p data-track=\"31\">\u76f8\u8f83\u4e8e\u5747\u503c\u5bf9outliers\u7684\u654f\u611f\u6027\uff0c\u4e2d\u4f4d\u6570\u80fd\u66f4\u7a33\u5b9a\u5730\u53cd\u6620\u201c\u6709\u6548\u8f93\u51fa\u201d\u7684\u5956\u52b1\u57fa\u51c6\uff1a\u5de6\u504f\u5206\u5e03\uff08\u6613\u9898\uff09\u4e2d\uff0cQ\u503c\u9ad8\u4e8e\u5168\u91cf\u5747\u503c\uff0c\u76f8\u5bf9\u4ea7\u751f\u66f4\u591a\u8d1fadvantage\u6837\u672c\uff0c\u8ba9\u8fd9\u4e9b\u6570\u636e\u8d21\u732e\u6cdb\u5316\u6027\uff1b\u53f3\u504f\u5206\u5e03\uff08\u96be\u9898\uff09\u4e2d\uff0cQ\u503c\u4f4e\u4e8e\u5168\u91cf\u5747\u503c\uff0c\u76f8\u5bf9\u4ea7\u751f\u66f4\u591a\u6b63advantage\u6837\u672c\uff0c\u8ba9\u8fd9\u4e9b\u6570\u636e\u8d21\u732e\u63d0\u70b9\u9700\u6c42\u3002<\/p>\n<p data-track=\"32\">Rollout\u566a\u58f0\u867d\u7136\u4e0d\u5229\u4e8e\u771f\u5b9eadvantage\u4f30\u8ba1\uff0c\u4f46\u8fd9\u4e9bnoise\u4e5f\u662f\u4e00\u79cd\u6a21\u578b\u80fd\u529b\u8fb9\u754c\u7684\u8868\u8fbe\uff0c\u6240\u4ee5\u7814\u7a76\u56e2\u961f\u4e5f\u5c06\u5176\u7eb3\u5165RL\u8fc7\u7a0b\uff0c\u4ee5\u8ba9\u6a21\u578b\u83b7\u5f97\u66f4\u6e05\u6670\u7684\u80fd\u529b\u8fb9\u754c\uff0c\u4ece\u800c\u64c5\u4e8e\u5904\u7406\u590d\u6742\u8f93\u5165\u3002<\/p>\n<p data-track=\"33\"><strong>3. \u5de5\u7a0b\u9002\u914d\uff1a\u4f4e\u590d\u6742\u5ea6+\u9ad8\u517c\u5bb9\u6027\uff0c\u5bf9\u63a5\u4f01\u4e1a\u771f\u5b9e\u6570\u636e\u8bad\u7ec3\u9700\u6c42<\/strong><\/p>\n<p data-track=\"34\">GAPO\u7684\u8ba1\u7b97\u590d\u6742\u5ea6\u4ec5\u4e3aO(n log n)\uff08\u4e3b\u8981\u6765\u81ea\u5956\u52b1\u6392\u5e8f\uff09\uff0c\u6ed1\u52a8\u7a97\u53e3\u626b\u63cf\u4ec5\u9700O(n)\u65f6\u95f4\uff0c\u4e0d\u4f1a\u7ed9\u8bad\u7ec3\u5e26\u6765\u989d\u5916\u7b97\u529b\u8d1f\u62c5\u3002<\/p>\n<p data-track=\"35\">\u66f4\u91cd\u8981\u7684\u662f\uff0c\u5b83\u4ec5\u4fee\u6539\u4f18\u52bf\u8ba1\u7b97\u51fd\u6570\uff0c\u65e0\u9700\u8c03\u6574RL\u6846\u67b6\u5176\u4ed6\u6a21\u5757\uff0c\u4ec5\u9700\u5c11\u91cf\u4ee3\u7801\u5373\u53ef\u96c6\u6210\u5230verl\u7b49\u4e3b\u6d41\u6846\u67b6\uff0c\u9002\u914d\u4f01\u4e1a\u57fa\u4e8e\u771f\u5b9e\u590d\u6742\u6570\u636e\u7684\u8bad\u7ec3\u6d41\u7a0b\u2014\u2014\u65e0\u9700\u5bf9\u6570\u636e\u8fdb\u884c\u989d\u5916\u964d\u566a\u9884\u5904\u7406\uff0c\u5373\u53ef\u76f4\u63a5\u5e94\u5bf9\u590d\u6742\u4e0a\u4e0b\u6587\u50ac\u751f\u7684rollout\u566a\u58f0\u3002<\/p>\n<h1 data-track=\"36\">\u5b9e\u8bc1\u9a8c\u8bc1\uff1aGAPO\u8ba9\u771f\u5b9e\u6570\u636e\u53d1\u6325\u5176\u4ef7\u503c<\/h1>\n<p data-track=\"37\">GAPO\u7684\u6838\u5fc3\u4ef7\u503c\u5728\u4e8e\uff0c\u5b83\u8ba9\u4f01\u4e1a\u91c7\u96c6\u7684\u771f\u5b9e\u590d\u6742\u6570\u636e\u771f\u6b63\u53d1\u6325\u6548\u7528\uff0c\u5176\u6027\u80fd\u4f18\u52bf\u901a\u8fc7\u8bba\u6587\u4e30\u5bcc\u7684\u5b9e\u9a8c\u56fe\u8868\u5f97\u5230\u5145\u5206\u9a8c\u8bc1\uff1a<\/p>\n<p data-track=\"38\"><strong>1. \u4f18\u52bf\u503c\u4f30\u8ba1\u66f4\u51c6\uff0c\u6a21\u578b\u7cbe\u5ea6\u663e\u8457\u63d0\u5347<\/strong><\/p>\n<p data-track=\"39\">\u5982\u4e0b\u8868\u6240\u793a\uff0cGAPO\u57289\u4e2a\u4e3b\u6d41LLM\uff083B-14B\u53c2\u6570\uff09\u4e0a\u5747\u5b9e\u73b0\u7a33\u5b9a\u63d0\u5347\uff1a\u4ee3\u7801\u4e13\u7528\u578b\u6a21\u578b\u53d7\u76ca\u6700\u663e\u8457\uff0cQwen2.5-Coder-14B\u57df\u5185\uff08ID\uff09\u7cbe\u786e\u5339\u914d\u51c6\u786e\u7387\u8fbe46.25%\uff0c\u8f83GRPO\u63d0\u53474.35\u4e2a\u767e\u5206\u70b9\u3002<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/60d91b5aa3a1acf23736421e72c5f2e9.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<p>\u5728\u6d4b\u8bd5\u96c6\u4e0a\u4e5d\u79cdLLM\u7684\u7cbe\u786e\u5339\u914d\u51c6\u786e\u7387<\/p>\n<\/p><\/div>\n<p data-track=\"41\">\u8de8\u57df\uff08OOD\uff09\u573a\u666f\u4e0b\uff0cQwen2.5-Coder-7B\u5728zeta\u6570\u636e\u96c6\u4e0a\u51c6\u786e\u7387\u63d0\u53475.30\u4e2a\u767e\u5206\u70b9\uff0c\u76f8\u5bf9\u63d0\u5347\u8fbe38.89%\u3002\u8fd9\u80cc\u540e\u6b63\u662f\u4f18\u52bf\u503c\u4f30\u8ba1\u5931\u771f\u95ee\u9898\u7684\u6709\u6548\u5904\u7406\u2014\u2014\u6a21\u578b\u80fd\u7cbe\u51c6\u5b66\u4e60\u771f\u5b9e\u6570\u636e\u4e2d\u7684\u6709\u6548\u7f16\u8f91\u6a21\u5f0f\uff0c\u800c\u975e\u88aboutliers\u8bef\u5bfc\u3002<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/5186d34406fc890327bf0addb87092ed.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<p>Qwen2.5-Coder\u548cQwen3\u5728OOD zeta\u6570\u636e\u96c6\u4e0a\u7684\u7cbe\u786e\u5339\u914d\u51c6\u786e\u7387\u3002<\/p>\n<\/p><\/div>\n<p data-track=\"43\">\u6b64\u5916\uff0c\u56e2\u961f\u7684GAPO\u65b9\u6cd5\u5bf9baseline\u7684\u63d0\u5347\u4e5f\u6bd4\u8f83\u7a33\u5b9a\u3002<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/e00904d725d08b488c58c7ee211cac45.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<p>\u57fa\u4e8eQwen2.5-Coder-3B\/7B\u6a21\u578b\u7684DAPO\u7b97\u6cd5\u5728\u6d4b\u8bd5\u96c6\u4e0a\u7684\u6027\u80fd\u66f2\u7ebf\uff08W&amp;B\u8bb0\u5f55\uff09<\/p>\n<\/p><\/div>\n<p data-track=\"45\"><strong>2. \u8bad\u7ec3\u66f4\u7a33\u5b9a\uff0c\u7b97\u529b\u5229\u7528\u7387\u4f18\u5316<\/strong><\/p>\n<p data-track=\"46\">\u4e0b\u56fe\u88c1\u526a\u6bd4\u4f8b\u66f2\u7ebf\u663e\u793a\uff0cGAPO\u7684pg_clipfrac\uff08\u68af\u5ea6\u66f4\u65b0\u4e2d\u6982\u7387\u6bd4\u88ab\u88c1\u526a\u7684\u6bd4\u4f8b\uff09\u663e\u8457\u4f4e\u4e8eGRPO\/DAPO\uff0c\u8bf4\u660e\u4f18\u52bf\u503c\u4f30\u8ba1\u66f4\u5408\u7406\uff0c\u653f\u7b56\u66f4\u65b0\u66f4\u51c6\u786e\uff0c\u80fd\u5145\u5206\u5229\u7528\u771f\u5b9e\u6570\u636e\u4e2d\u7684\u5956\u52b1\u4fe1\u606f\uff0c\u51cf\u5c11\u65e0\u6548\u8fed\u4ee3\u3002<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/be41e1a003096304c615810233e9bcc9.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<p>\u57fa\u4e8eQwen2.5-Coder-3B\/7B\u6a21\u578b\u7684DAPO\u7b97\u6cd5\u5728\u8bad\u7ec3\u96c6\u4e0a\u7684\u88c1\u526a\u6bd4\u4f8b\uff08Clip fraction\uff09\u66f2\u7ebf<\/p>\n<\/p><\/div>\n<p data-track=\"48\">\u540c\u65f6\uff0cGPU\u541e\u5410\u91cf\u66f2\u7ebf\u663e\u793a\uff0cGAPO\u8ba93B\u6a21\u578b\u7684\u5e73\u5747\u541e\u5410\u91cf\u63d0\u53474.96%\uff0c\u8bad\u7ec3\u540e\u671f\u4f18\u52bf\u66f4\u660e\u663e\u2014\u2014\u5bf9\u4f01\u4e1a\u800c\u8a00\uff0c\u8fd9\u610f\u5473\u7740\u7528\u540c\u6837\u7684\u7b97\u529b\uff0c\u80fd\u4ece\u771f\u5b9e\u590d\u6742\u6570\u636e\u4e2d\u83b7\u5f97\u66f4\u597d\u7684\u8bad\u7ec3\u6548\u679c\uff0c\u4ece\u800c\u964d\u4f4eAI\u6a21\u578b\u8bad\u7ec3\u6210\u672c\u3002<\/p>\n<div>\n<p><img decoding=\"async\" src=\"https:\/\/i.qbitai.com\/wp-content\/uploads\/2026\/02\/bad3cae88312c81a7a590440716bcab6.webp\" alt=\"\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b &amp;#8211; \u91cf\u5b50\u4f4d\" \/><\/p>\n<p>\u57fa\u4e8eQwen2.5-Coder-3B\u6a21\u578b\u7684DAPO\u7b97\u6cd5\u5728\u8bad\u7ec3\u96c6\u4e0a\u7684GPU\u541e\u5410\u91cf\u66f2\u7ebf\u3002<\/p>\n<\/p><\/div>\n<h1 data-track=\"50\">\u7ed3\u8bed\uff1a\u8ba9\u771f\u5b9e\u6570\u636e\u6210\u4e3aAI\u8f85\u52a9\u7f16\u7a0b\u7684\u201c\u71c3\u6599\u201d\uff0c\u800c\u975e\u201c\u5305\u88b1\u201d<\/h1>\n<p data-track=\"51\">GAPO\u7814\u7a76\u5de5\u4f5c\u7cbe\u51c6\u7a81\u7834\u4e86\u771f\u5b9e\u590d\u6742\u7528\u6237\u73af\u5883\u4e2d\u201c\u590d\u6742\u4e0a\u4e0b\u6587\u2192rollout\u566a\u58f0\u2192\u4f18\u52bf\u503c\u4f30\u8ba1\u4e0d\u51c6\u201d\u7684\u8fde\u9501\u96be\u9898\uff0c\u8ba9\u4f01\u4e1a\u91c7\u96c6\u7684\u771f\u5b9e\u6570\u636e\u4ece\u8bad\u7ec3\u201c\u5305\u88b1\u201d\u53d8\u6210\u4e86\u63d0\u5347\u6a21\u578b\u6548\u679c\u7684\u201c\u71c3\u6599\u201d\u3002<\/p>\n<p data-track=\"52\">\u8bba\u6587\u4e2d\u7684\u56fe\u8868\u6570\u636e\u5145\u5206\u8bc1\u660e\uff0c\u8be5\u65b9\u6cd5\u5728\u7cbe\u5ea6\u3001\u6cdb\u5316\u80fd\u529b\u3001\u8bad\u7ec3\u6548\u7387\u4e0e\u786c\u4ef6\u5229\u7528\u7387\u4e0a\u5747\u5b9e\u73b0\u7a81\u7834\uff0c\u65e2\u4e3aAI\u79d1\u7814\u4eba\u5458\u63d0\u4f9b\u4e86\u9488\u5bf9\u771f\u5b9e\u573a\u666f\u566a\u58f0\u95ee\u9898\u7684\u65b0\u6280\u672f\u601d\u8def\uff0c\u4e5f\u4e3a\u4f01\u4e1a\u964d\u4f4eAI\u8f85\u52a9\u7f16\u7a0b\u843d\u5730\u95e8\u69db\u3001\u63d0\u5347\u7814\u53d1\u6548\u7387\u63d0\u4f9b\u4e86\u5207\u5b9e\u53ef\u884c\u7684\u65b9\u6848\u3002<\/p>\n<p data-track=\"53\">\u968f\u7740GAPO\u4ee3\u7801\u7684\u5f00\u6e90\uff0c\u7814\u7a76\u56e2\u961f\u4e5f\u671f\u5f85\u66f4\u591a\u79d1\u7814\u4eba\u5458\u4e0e\u4f01\u4e1a\u5f00\u53d1\u8005\u53c2\u4e0e\u63a2\u7d22\uff0c\u8ba9AI\u8f85\u52a9\u7f16\u7a0b\u66f4\u6df1\u5165\u5730\u878d\u5165\u8f6f\u4ef6\u5f00\u53d1\u5168\u6d41\u7a0b\uff0c\u63a8\u52a8\u8f6f\u4ef6\u4ea7\u4e1a\u5411\u66f4\u9ad8\u6548\u3001\u66f4\u667a\u80fd\u7684\u65b9\u5411\u53d1\u5c55\u3002<\/p>\n<p data-track=\"54\"><strong>\u8bba\u6587\u6807\u9898\uff1a<\/strong><\/p>\n<p data-track=\"55\">GAPO:Robust Advantage Estimation for Real-World Code LLMs<br \/><strong>\u8bba\u6587\u94fe\u63a5\uff1a<\/strong><\/p>\n<p data-track=\"56\">https:\/\/arxiv.org\/pdf\/2510.21830<br \/><strong>\u5f00\u6e90\u4ee3\u7801\uff1a<\/strong><\/p>\n<p data-track=\"57\">https:\/\/github.com\/TsingZ0\/verl-GAPO<\/p>\n<p>(\u6765\u6e90:\u91cf\u5b50\u4f4d)                                                   <\/p>\n<div><span><\/span><em>\u7248\u6743\u6240\u6709\uff0c\u672a\u7ecf\u6388\u6743\u4e0d\u5f97\u4ee5\u4efb\u4f55\u5f62\u5f0f\u8f6c\u8f7d\u53ca\u4f7f\u7528\uff0c\u8fdd\u8005\u5fc5\u7a76\u3002<\/em><span><\/span><\/div>\n","protected":false},"excerpt":{"rendered":"<p>\u9c81\u68d2RL\u8d4b\u80fdAI\u7f16\u7a0b\uff01\u7834\u5c40\u4f01\u4e1a\u6570\u636e\u566a\u58f0\u96be\u9898\uff0c\u540c\u7b49\u7b97\u529b\u8bad\u51fa\u66f4\u597d\u6a21\u578b \u91cf\u5b50\u4f4d\u7684\u670b\u53cb\u4eec 2026-02-16 19: [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":[],"categories":[380],"tags":[],"views":25,"_links":{"self":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/751035"}],"collection":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=751035"}],"version-history":[{"count":0,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=\/wp\/v2\/posts\/751035\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=751035"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=751035"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.alphaseek.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=751035"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}