{"id":6181,"date":"2025-09-18T11:55:32","date_gmt":"2025-09-18T03:55:32","guid":{"rendered":"https:\/\/nullthought.net\/?p=6181"},"modified":"2025-09-18T11:57:47","modified_gmt":"2025-09-18T03:57:47","slug":"deepseek-r1%e9%80%9a%e8%bf%87%e5%bc%ba%e5%8c%96%e5%ad%a6%e4%b9%a0%ef%bc%88rl%ef%bc%89%e6%bf%80%e5%8a%b1%e6%8e%a8%e7%90%86","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=6181","title":{"rendered":"DeepSeek-R1\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u6fc0\u52b1\u63a8\u7406"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/www.nature.com\/articles\/s41586-025-09422-z\" target=\"_blank\" rel=\"noreferrer noopener\">DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning<\/a><\/strong>\u63d0\u51fa\u4e86\u4e00\u4e2a\u4ee5\u201c\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u76f4\u63a5\u6fc0\u52b1\u63a8\u7406\u201d\u4e3a\u4e2d\u5fc3\u7684\u65b0\u8303\u5f0f\uff1a\u5728\u6ca1\u6709\u4eba\u5de5\u6807\u6ce8\u63a8\u7406\u8f68\u8ff9\u7684\u524d\u63d0\u4e0b\uff0c\u4ec5\u7528\u201c\u53ef\u9a8c\u8bc1\u201d\u7684\u6700\u7ec8\u7b54\u6848\u4f5c\u4e3a\u5956\u52b1\u4fe1\u53f7\uff0c\u5c31\u80fd\u8ba9\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5728\u8bad\u7ec3\u4e2d\u81ea\u53d1\u6d8c\u73b0\u51fa\u66f4\u957f\u3001\u66f4\u7cfb\u7edf\u7684\u601d\u8003\u94fe\u3001\u53cd\u601d\u4e0e\u81ea\u68c0\u7b49\u9ad8\u7ea7\u63a8\u7406\u884c\u4e3a\u3002\u4f5c\u8005\u5148\u5728 <strong><a href=\"https:\/\/nullthought.net\/?p=5463\" target=\"_blank\" rel=\"noreferrer noopener\">DeepSeek-V3<\/a><\/strong> Base \u4e0a\u7528 GRPO \u7b97\u6cd5\u8bad\u7ec3\u51fa\u7eaf RL \u7684 DeepSeek-R1-Zero\uff0c\u518d\u901a\u8fc7\u201c\u62d2\u7edd\u91c7\u6837 + \u6709\u9650 SFT + \u7b2c\u4e8c\u9636\u6bb5 RL\u201d\u7684\u591a\u9636\u6bb5\u6d41\u7a0b\u5f97\u5230\u66f4\u5747\u8861\u7684 <strong><a href=\"https:\/\/nullthought.net\/?p=5614\" target=\"_blank\" rel=\"noreferrer noopener\">DeepSeek-R1<\/a><\/strong>\u3002\u6a21\u578b\u5728\u6570\u5b66\u3001\u7f16\u7a0b\u7ade\u8d5b\u4e0e STEM \u7b49\u53ef\u9a8c\u8bc1\u4efb\u52a1\u4e0a\u4f18\u4e8e\u5e38\u89c4\u201c\u4eba\u7c7b\u793a\u4f8b\u76d1\u7763\u201d\u8def\u7ebf\uff0c\u5e76\u53ef\u5c06\u5f3a\u63a8\u7406\u80fd\u529b\u84b8\u998f\u5230\u66f4\u5c0f\u6a21\u578b\u4ee5\u964d\u4f4e\u80fd\u8017\u4e0e\u4f7f\u7528\u95e8\u69db\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u5e94\u4e3b\u8981\u6765\u81eaDeepSeek\u56e2\u961f\uff0c\u6881\u6587\u950b\u662f\u8bba\u6587\u901a\u8baf\u4f5c\u8005\u3002<\/p>\n\n\n\n<p>\u4e00\u3001\u7814\u7a76\u80cc\u666f\u4e0e\u95ee\u9898\u754c\u5b9a<br>\u4ee5 CoT \u63d0\u793a\u548c\u4eba\u7c7b\u793a\u4f8b\u76d1\u7763\u4e3a\u4ee3\u8868\u7684\u4f20\u7edf\u505a\u6cd5\uff0c\u5728\u590d\u6742\u63a8\u7406\u4e0a\u867d\u6709\u6548\u4f46\u4f9d\u8d56\u5927\u91cf\u4eba\u529b\u4e0e\u201c\u4eba\u7c7b\u601d\u7ef4\u8303\u5f0f\u201d\uff0c\u5bb9\u6613\u5f15\u5165\u504f\u89c1\u4e0e\u5929\u82b1\u677f\u6548\u5e94\u3002\u8bba\u6587\u5173\u6ce8\u7684\u95ee\u9898\u662f\uff1a\u5728\u4e0d\u63d0\u4f9b\u4eba\u7c7b\u63a8\u7406\u8f68\u8ff9\u7684\u6761\u4ef6\u4e0b\uff0c\u662f\u5426\u53ef\u4ee5\u53ea\u9760\u201c\u7ed3\u679c\u662f\u5426\u6b63\u786e\u201d\u7684\u5956\u52b1\uff0c\u8ba9\u6a21\u578b\u81ea\u5df1\u5b66\u4f1a\u66f4\u5f3a\u7684\u601d\u8003\u7b56\u7565\uff1f\u4f5c\u8005\u7ed9\u51fa\u7684\u7b54\u6848\u662f\u80af\u5b9a\u7684\u2014\u2014<strong>\u53ef\u9a8c\u8bc1\u4efb\u52a1\u3001\u53ef\u9760\u5224\u9898\u5668\/\u7f16\u8bd1\u5668\u4e0e\u89c4\u6a21\u5316 RL \u662f\u5173\u952e<\/strong>\u3002<\/p>\n\n\n\n<p>\u4e8c\u3001\u603b\u4f53\u6280\u672f\u8def\u7ebf\u4e0e\u8bad\u7ec3\u8303\u5f0f<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>R1-Zero\uff08\u7eaf RL \u8d77\u70b9\uff09\uff1a\u4ee5 DeepSeek-V3 Base \u4e3a\u7b56\u7565\u6a21\u578b\uff0c\u91c7\u7528 GRPO \u8fdb\u884c\u5927\u89c4\u6a21 RL\u3002\u8bad\u7ec3\u65f6\u4ec5\u5f3a\u5236\u8f93\u51fa\u7ed3\u6784\u6ee1\u8db3 &lt;think&gt;\u2026&lt;\/think&gt;&lt;answer&gt;\u2026&lt;\/answer&gt;\uff0c\u5956\u52b1\u53ea\u770b\u201c\u7b54\u6848\u6b63\u786e\u6027 + \u8f93\u51fa\u683c\u5f0f\u201d\uff0c\u4e0d\u7ea6\u675f\u201c\u5982\u4f55\u601d\u8003\u201d\uff0c\u4e14\u4e0d\u505a\u9884\u5148 SFT\uff0c\u907f\u514d\u628a\u63a2\u7d22\u675f\u7f1a\u5728\u201c\u4eba\u7c7b\u793a\u4f8b\u201d\u7684\u8f68\u9053\u4e0a\u3002<\/li>\n\n\n\n<li>R1\uff08\u591a\u9636\u6bb5\u6574\u5408\uff09\uff1a\u5728 R1-Zero \u7684\u57fa\u7840\u4e0a\uff0c\u5148\u5c11\u91cf\u6536\u96c6\u201c\u51b7\u542f\u52a8\u201d\u7684\u5bf9\u8bdd\u5f0f\u957f\u601d\u8003\u6570\u636e\uff0c\u518d\u505a\u7b2c\u4e00\u9636\u6bb5 RL\uff1b\u968f\u540e\u5f15\u5165\u201c\u62d2\u7edd\u91c7\u6837 + SFT\u201d\uff08\u6df7\u5408\u63a8\u7406\u4e0e\u901a\u7528\u8bed\u6599\uff09\u589e\u5f3a\u53ef\u8bfb\u6027\u4e0e\u901a\u7528\u4efb\u52a1\uff1b\u6700\u540e\u8fdb\u884c\u7b2c\u4e8c\u9636\u6bb5 RL\uff0c\u5c06\u201c\u53ef\u9a8c\u8bc1\u63a8\u7406\u5956\u52b1 + \u6709\u5bb3\u6027\/\u6709\u7528\u6027\u504f\u597d\u5956\u52b1 + \u8bed\u8a00\u4e00\u81f4\u6027\u5956\u52b1\u201d\u6574\u5408\uff0c\u5f97\u5230\u5728\u63a8\u7406\u4e0e\u901a\u7528\u5bf9\u9f50\u4e0a\u66f4\u5747\u8861\u7684 DeepSeek-R1\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u4e09\u3001GRPO \u7b97\u6cd5\u8981\u70b9\u4e0e\u76f8\u5bf9 PPO \u7684\u5dee\u5f02<br>GRPO \u7684\u6838\u5fc3\u662f\u201c\u7ec4\u76f8\u5bf9\u4f18\u52bf\u201d\uff1a\u5bf9\u6bcf\u4e2a\u95ee\u9898\u4e00\u6b21\u91c7\u6837\u4e00\u7ec4 G \u4e2a\u8f93\u51fa\uff0c\u7528\u8be5\u7ec4\u5185\u7684\u76f8\u5bf9\u5956\u52b1\u5206\u5e03\u76f4\u63a5\u5f52\u4e00\u5316\u5f97\u5230\u4f18\u52bf\uff0c\u4ece\u800c\u65e0\u9700\u5355\u72ec\u8bad\u7ec3\u4ef7\u503c\u7f51\u7edc\uff1b\u518d\u914d\u5408\u5bf9\u53c2\u8003\u7b56\u7565\u7684 KL \u6b63\u5219\u4e0e\u88c1\u526a\uff0c\u7b80\u5316\u4e86 PPO \u7684\u8bad\u7ec3\u6808\u5e76\u63d0\u5347\u53ef\u6269\u5c55\u6027\u3002\u76f4\u89c2\u4e0a\uff0c\u5b83\u662f\u5728\u201c\u540c\u9898\u591a\u89e3\u201d\u7684\u5c0f\u578b\u201c\u9526\u6807\u8d5b\u201d\u91cc\uff0c\u628a\u66f4\u4f18\u89e3\u7684\u76f8\u5bf9\u4f18\u52bf\u76f4\u63a5\u53d8\u6210\u66f4\u65b0\u4fe1\u53f7\u3002<\/p>\n\n\n\n<p>\u56db\u3001\u5956\u52b1\u8bbe\u8ba1\u4e0e\u201c\u53ef\u9a8c\u8bc1\u201d\u95ed\u73af<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u89c4\u5219\u5f0f\u5956\u52b1\uff1a<br>\u2022 \u6b63\u786e\u6027\u5956\u52b1\uff1a\u6570\u5b66\u9898\u7528\u6807\u51c6\u7b54\u6848\u6bd4\u5bf9\uff08\u6309\u6307\u5b9a\u683c\u5f0f\u8f93\u51fa\uff09\uff0c\u4ee3\u7801\u9898\u7528\u7f16\u8bd1\/\u6d4b\u8bd5\u7528\u4f8b\u5224\u5206\uff0c\u903b\u8f91\u9898\u7528\u89c4\u5219\u9a8c\u8bc1\u3002<br>\u2022 \u683c\u5f0f\u5956\u52b1\uff1a\u5f3a\u5236 &lt;think&gt;\/&lt;answer&gt; \u7ed3\u6784\uff0c\u4fbf\u4e8e\u5206\u79bb\u601d\u8003\u8fc7\u7a0b\u4e0e\u7ed3\u8bba\u3002<br>\u4e24\u8005\u540c\u6743\u76f8\u52a0\uff0c\u63a8\u7406\u57df\u4e0d\u5f15\u5165\u201c\u795e\u7ecf\u5956\u52b1\u6a21\u578b\u201d\uff0c\u4ee5\u89c4\u907f\u5927\u89c4\u6a21 RL \u4e2d\u7684\u201c\u5956\u52b1\u9ed1\u5ba2\u201d\u3002<\/li>\n\n\n\n<li>\u6a21\u578b\u5f0f\u5956\u52b1\uff1a<br>\u2022 \u6709\u7528\u6027\u504f\u597d\u6a21\u578b\uff1a\u7528\u6210\u5bf9\u504f\u597d\u6570\u636e\u8bad\u7ec3\uff0c\u5173\u6ce8\u6700\u7ec8\u603b\u7ed3\u7684\u5b9e\u7528\u4ef7\u503c\u4e0e\u76f8\u5173\u6027\u3002<br>\u2022 \u5b89\u5168\u6027\u6a21\u578b\uff1a\u7528\u5e26\u201c\u5b89\u5168\/\u4e0d\u5b89\u5168\u201d\u6807\u7b7e\u7684\u6570\u636e\u8fdb\u884c\u70b9\u5f0f\u8bad\u7ec3\uff0c\u8bc6\u522b\u6574\u4f53\u54cd\u5e94\u98ce\u9669\u3002<br>\u901a\u7528\u6570\u636e\u4e0a\u4f7f\u7528\u504f\u597d\u5956\u52b1\uff1b\u63a8\u7406\u6570\u636e\u4ecd\u575a\u6301\u89c4\u5219\u5f0f\u5956\u52b1\u3002<\/li>\n\n\n\n<li>\u8bed\u8a00\u4e00\u81f4\u6027\u5956\u52b1\uff1a\u6309\u76ee\u6807\u8bed\u8bcd\u5360\u6bd4\u8ba1\u5206\uff0c\u7f13\u89e3\u4e2d\u82f1\u5939\u6742\u4ee5\u63d0\u5347\u53ef\u8bfb\u6027\uff08\u4f1a\u7565\u6709\u63a8\u7406\u6027\u80fd\u4ee3\u4ef7\uff09\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u4e94\u3001\u89c4\u6a21\u5316\u8bad\u7ec3\u4e0e\u5173\u952e\u8d85\u53c2\uff08\u5de5\u7a0b\u7ec6\u8282\uff09<br>R1-Zero \u91c7\u7528\u5b66\u4e60\u7387 3e-6\u3001KL \u7cfb\u6570\u7ea6 0.001\u3001\u6bcf\u9898\u91c7\u6837 16 \u4e2a\u8f93\u51fa\uff1b\u524d 8.2k \u6b65\u6700\u5927\u957f\u5ea6 32,768 \u4ee4\u724c\uff0c\u4e4b\u540e\u5347\u81f3 65,536\uff0c\u6027\u80fd\u4e0e\u8f93\u51fa\u957f\u5ea6\u5728\u8be5\u6b65\u51fa\u73b0\u8dc3\u8fc1\uff1b\u6279\u5927\u5c0f\u7ea6 512\/\u6b65\uff1b\u53c2\u8003\u6a21\u578b\u5b9a\u671f\u5237\u65b0\uff1b\u5355\u8f6e rollout \u751f\u6210\u6570\u5343\u5230\u4e0a\u4e07\u8f93\u51fa\u518d\u5206\u5c0f\u6279\u66f4\u65b0\u3002\u7b2c\u4e00\u9636\u6bb5\u4e0e\u7b2c\u4e8c\u9636\u6bb5 RL\u5728\u6b64\u57fa\u7840\u4e0a\u5fae\u8c03\u4e86\u6e29\u5ea6\u3001\u88c1\u526a\u7cfb\u6570\u3001\u8bed\u8a00\u4e00\u81f4\u6027\u6743\u91cd\u4e0e\u4efb\u52a1\u6df7\u5408\u6bd4\u4f8b\uff0c\u4e14\u5728\u540e\u671f\u624d\u5f15\u5165\u901a\u7528\u6307\u4ee4\u6570\u636e\u4e0e\u504f\u597d\u5956\u52b1\uff0c\u4ee5\u964d\u4f4e\u201c\u5956\u52b1\u9ed1\u5ba2\u201d\u98ce\u9669\u3002<\/p>\n\n\n\n<p>\u516d\u3001\u6d8c\u73b0\u884c\u4e3a\u4e0e\u201caha moment\u201d<br>\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u81ea\u52a8\u5ef6\u957f\u201c\u601d\u8003\u65f6\u957f\u201d\uff0c\u5e73\u5747\u8f93\u51fa\u957f\u5ea6\u968f\u6b65\u6570\u7a33\u6b65\u4e0a\u5347\uff0c\u5e76\u9010\u6b65\u5f62\u6210\u201c\u81ea\u6211\u6821\u9a8c\u3001\u53cd\u601d\u91cd\u4f30\u3001\u5c1d\u8bd5\u5907\u9009\u8def\u5f84\u201d\u7b49\u7b56\u7565\u3002\u4f5c\u8005\u89c2\u5bdf\u5230\u4e0e\u201c\u53cd\u601d\/\u6821\u9a8c\u201d\u76f8\u5173\u8bcd\u9891\u663e\u8457\u4e0a\u5347\uff0c\u5e76\u5c55\u793a\u4e86\u201c\u7b49\u7b49\uff0c\u6211\u9700\u8981\u91cd\u60f3\u4e00\u4e0b\u201d\u7684\u5178\u578b\u7247\u6bb5\uff0c\u6807\u5fd7\u63a8\u7406\u6a21\u5f0f\u53d1\u751f\u9636\u6bb5\u6027\u8dc3\u8fc1\u3002<\/p>\n\n\n\n<p>\u4e03\u3001\u5b9e\u9a8c\u7ed3\u679c\u4e0e\u9636\u6bb5\u6027\u5bf9\u6bd4<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u8bad\u7ec3\u8f68\u8ff9\u4e0e\u53ef\u9a8c\u8bc1\u8868\u73b0\uff1aR1-Zero \u5728 AIME \u7b49\u7ade\u8d5b\u9898\u4e0a\u4ece\u8f83\u4f4e\u8d77\u70b9\u5feb\u901f\u63d0\u5347\u5230\u63a5\u8fd1\u6216\u8d85\u8fc7\u4eba\u7c7b\u5e73\u5747\u6c34\u5e73\uff1b\u5728\u7f16\u7a0b\u7ade\u8d5b\u4e0e\u7406\u5316\u751f\u7b49\u53ef\u9a8c\u8bc1\u9898\u578b\u4e0a\u4ea6\u663e\u8457\u8fdb\u6b65\u3002<\/li>\n\n\n\n<li>\u591a\u9636\u6bb5\u6f14\u8fdb\uff1a<br>\u2022 \u63a8\u7406\u5411\u6307\u6807\uff08\u5982 MMLU-Pro\u3001MATH-500\u3001Codeforces\u3001SWE-bench Verified\u3001AIME \u7b49\uff09\u6301\u7eed\u4e0a\u5347\uff0cR1-Zero \u2192 \u7b2c\u4e00\u9636\u6bb5 RL \u2192 \u589e\u52a0 SFT \u2192 \u6700\u7ec8 R1 \u5448\u73b0\u201c\u63a8\u7406\u66f4\u5f3a\u3001\u5bf9\u9f50\u66f4\u597d\u201d\u7684\u6298\u4e2d\u6700\u4f18\u3002<br>\u2022 \u901a\u7528\u5411\u6307\u6807\uff08\u5982 Arena-Hard\u3001AlpacaEval\u3001IF-Eval \u7b49\uff09\u5728\u5f15\u5165\u6df7\u5408 SFT \u4e0e\u504f\u597d\u5956\u52b1\u540e\u663e\u8457\u63d0\u5347\u3002<\/li>\n\n\n\n<li>\u77e5\u8bc6\u84b8\u998f\uff1a\u5c06\u957f CoT \u80fd\u529b\u8fc1\u79fb\u81f3\u66f4\u5c0f\u6a21\u578b\uff0c\u663e\u8457\u4f18\u4e8e\u540c\u4f53\u91cf\u5e38\u89c4\u6307\u4ee4\u5fae\u8c03\u7248\u672c\uff0c\u964d\u4f4e\u63a8\u7406\u80fd\u8017\u4e0e\u7b97\u529b\u95e8\u69db\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u516b\u3001\u4e0e\u4e3b\u6d41\u8def\u7ebf\u7684\u7cfb\u7edf\u5bf9\u6bd4<br>\u2022 \u76f8\u5bf9\u201c\u4eba\u7c7b\u793a\u4f8b\u76d1\u7763 + CoT\u201d\uff1a\u672c\u7814\u7a76\u628a\u201c\u5982\u4f55\u601d\u8003\u201d\u7684\u81ea\u7531\u4ea4\u8fd8\u7ed9\u6a21\u578b\uff0c\u8ba9\u201c\u53ef\u9a8c\u8bc1\u7ed3\u679c\u201d\u9a71\u52a8\u7b56\u7565\u81ea\u53d1\u73b0\uff0c\u907f\u514d\u4eba\u7c7b\u8303\u5f0f\u8bbe\u9650\u3002<br>\u2022 \u76f8\u5bf9 RLHF\uff1aR1 \u5728\u63a8\u7406\u57df\u5c3d\u91cf\u907f\u514d\u201c\u795e\u7ecf\u504f\u597d\u5956\u52b1\u201d\uff0c\u6539\u4ee5\u53ef\u9a8c\u89c4\u5219\uff0c\u51cf\u5c11\u5956\u52b1\u9ed1\u5ba2\uff1b\u504f\u597d\u6a21\u578b\u4ec5\u7528\u4e8e\u901a\u7528\u4efb\u52a1\u5bf9\u9f50\u3002<br>\u2022 \u76f8\u5bf9\u6d4b\u8bd5\u65f6\u6269\u5c55\uff08\u591a\u6570\u6295\u7968\/MCTS \u7b49\uff09\uff1aR1 \u80fd\u6309\u9898\u76ee\u96be\u5ea6\u81ea\u9002\u5e94\u5206\u914d\u201c\u601d\u8003\u4ee4\u724c\u201d\uff0c\u590d\u6742\u9898\u201c\u591a\u60f3\u201d\u3001\u7b80\u5355\u9898\u201c\u5c11\u60f3\u201d\uff0c\u5b9e\u73b0\u201c\u8bad\u7ec3\u65f6\u5851\u9020 + \u6d4b\u8bd5\u65f6\u81ea\u9002\u5e94\u201d\u3002<\/p>\n\n\n\n<p>\u4e5d\u3001\u80fd\u529b\u8fb9\u754c\u4e0e\u98ce\u9669\u6e05\u5355<br>\u2022 \u5de5\u5177\u4e0e\u7ed3\u6784\u5316\u8f93\u51fa\uff1a\u5c1a\u672a\u5728 RL \u73af\u5883\u4e2d\u7cfb\u7edf\u7eb3\u5165\u68c0\u7d22\/\u5de5\u5177\/\u8ba1\u7b97\u5668\u4ea4\u4e92\uff1b\u662f\u540e\u7eed\u53ef\u8865\u5f3a\u65b9\u5411\u3002<br>\u2022 \u4ee4\u724c\u6548\u7387\uff1a\u867d\u80fd\u81ea\u8c03\u6df1\u601d\uff0c\u4f46\u4ecd\u53ef\u80fd\u201c\u8fc7\u5ea6\u601d\u8003\u201d\u3002<br>\u2022 \u8bed\u8a00\u6df7\u7528\uff1a\u975e\u4e3b\u8bed\u8a00\u573a\u666f\u53ef\u80fd\u51fa\u73b0\u8de8\u8bed\u79cd\u6df7\u6742\u3002<br>\u2022 \u63d0\u793a\u654f\u611f\uff1afew-shot \u53ef\u80fd\u964d\u5206\uff0c\u5efa\u8bae\u96f6\u6837\u672c\u5e76\u660e\u786e\u8f93\u51fa\u683c\u5f0f\u3002<br>\u2022 \u5de5\u7a0b\u7c7b\u957f\u5468\u671f\u8bc4\u6d4b\uff1a\u8bc4\u6d4b\u8017\u65f6\u957f\u5bfc\u81f4 RL \u8fed\u4ee3\u6162\uff0c\u63d0\u5347\u4e0d\u5982\u53ef\u9a8c\u8bc1\u63a8\u7406\u57df\u663e\u8457\u3002<br>\u2022 \u5956\u52b1\u9ed1\u5ba2\uff1a\u5f53\u5956\u52b1\u4f9d\u8d56\u6a21\u578b\u6253\u5206\u4e14\u96be\u4ee5\u53ef\u9760\u8bc4\u6d4b\u65f6\uff0c\u7b56\u7565\u53ef\u80fd\u201c\u627e\u6377\u5f84\u201d\uff1b\u8bba\u6587\u901a\u8fc7\u9650\u5236\u504f\u597d\u5956\u52b1\u7684\u4f7f\u7528\u9636\u6bb5\u4e0e\u6b65\u6570\u7f13\u89e3\u3002<\/p>\n\n\n\n<p>\u5341\u3001\u5b9e\u8df5\u542f\u793a\uff08\u5982\u4f55\u7528\u3001\u5982\u4f55\u8bad\uff09<br>\u2022 \u628a\u4efb\u52a1\u8f6c\u5316\u4e3a\u201c\u53ef\u9a8c\u8bc1\u201d\u7684\u95ed\u73af\uff1a\u6807\u51c6\u7b54\u6848\u3001\u5224\u9898\u5668\u3001\u6d4b\u8bd5\u96c6\u3001\u7ea6\u675f\u683c\u5f0f\uff0c\u8ba9\u201c\u6b63\u786e\u6027\u201d\u76f4\u63a5\u53d8\u6210\u53ef\u4f18\u5316\u7684 RL \u4fe1\u53f7\u3002<br>\u2022 \u63d0\u793a\u7b56\u7565\uff1a\u96f6\u6837\u672c + \u660e\u786e\u8f93\u51fa\u7ed3\u6784\u6700\u7a33\uff1b\u9700\u8981\u53ef\u8bfb\u6027\u65f6\u52a0\u5165\u8f7b\u91cf\u201c\u8bed\u8a00\u4e00\u81f4\u6027\u7ea6\u675f\u201d\u3002<br>\u2022 \u5de5\u7a0b\u843d\u5730\uff1a\u5bf9\u4ee3\u7801\/\u6570\u7406\u9898\uff0c\u5148\u5efa\u9ad8\u53ef\u4fe1\u5ea6\u7f16\u8bd1\/\u5224\u9898\u6d41\u6c34\u7ebf\uff0c\u518d\u7528 GRPO \u505a\u7ec4\u91c7\u6837\u4e0e\u76f8\u5bf9\u4f18\u52bf\u66f4\u65b0\uff1b\u8bc4\u6d4b\u6602\u8d35\u9886\u57df\u53ef\u7528\u201c\u62d2\u7edd\u91c7\u6837 + \u5f02\u6b65\u8bc4\u6d4b\/\u4ee3\u7406\u6307\u6807\u201d\u4ee5\u964d\u4f4e RL \u5faa\u73af\u5ef6\u8fdf\u3002<\/p>\n\n\n\n<p>\u5341\u4e00\u3001\u5bf9\u672a\u6765\u7814\u7a76\u7684\u5efa\u8bae<br>\u2022 \u201c\u5de5\u5177\u589e\u5f3a\u63a8\u7406\u201d\u7684\u7aef\u5230\u7aef RL\uff1a\u5c06\u68c0\u7d22\u5668\u3001\u8ba1\u7b97\u5668\u3001\u7f16\u8bd1\u5668\u3001\u5916\u90e8\u73af\u5883\u7b49\u7eb3\u5165\u5956\u60e9\u95ed\u73af\u3002<br>\u2022 \u66f4\u7ec6\u7c92\u5ea6\u7684\u5bc6\u96c6\u5956\u52b1\uff1a\u4ece\u201c\u53ea\u770b\u7ec8\u70b9\u6b63\u786e\u201d\u8d70\u5411\u201c\u8fc7\u7a0b\u53ef\u68c0\u201d\uff08\u6b65\u9aa4\u5224\u5b9a\u3001\u5355\u6d4b\u8986\u76d6\u3001\u8349\u7a3f\u81ea\u68c0\u4fe1\u53f7\uff09\u3002<br>\u2022 \u53ef\u89e3\u91ca\u6027\u4e0e\u884c\u4e3a\u8ddf\u8e2a\uff1a\u91cf\u5316\u201c\u53cd\u601d\u8bcd\/\u81ea\u68c0\u884c\u4e3a\u201d\uff0c\u7814\u7a76\u201caha \u65f6\u523b\u201d\u7684\u89e6\u53d1\u6761\u4ef6\u4e0e\u8fc1\u79fb\u6027\u3002<br>\u2022 \u4ee4\u724c\u9884\u7b97\u8c03\u5ea6\uff1a\u56f4\u7ed5\u957f CoT \u7684\u5197\u4f59\uff0c\u7814\u7a76\u201c\u505c\u6b62\u51c6\u5219\/\u81ea\u9002\u5e94\u6df1\u601d\u7b56\u7565\u201d\u7684 RL \u7ea6\u675f\u3002<\/p>\n\n\n\n<p>\u5341\u4e8c\u3001\u5173\u952e\u672f\u8bed\u901f\u89c8<br>\u2022 GRPO\uff1a\u57fa\u4e8e\u7ec4\u76f8\u5bf9\u4f18\u52bf\u7684\u7b56\u7565\u4f18\u5316\uff0c\u65e0\u9700\u4ef7\u503c\u7f51\u7edc\uff0c\u914d\u5408 KL \u6b63\u5219\u4e0e\u88c1\u526a\u5b9e\u73b0\u7a33\u5b9a\u66f4\u65b0\u3002<br>\u2022 pass@1 \/ cons@K\uff1a\u5355\u6837\u672c\u4e00\u6b21\u547d\u4e2d\u7387 \/ \u81ea\u4e00\u81f4\u6027\uff08\u5bf9\u540c\u9898\u91c7\u6837 K \u6b21\u6295\u7968\uff09\u3002<br>\u2022 \u51b7\u542f\u52a8\u957f CoT\uff1a\u5c11\u91cf\u4eba\u5de5\u7b5b\u9009\u7684\u5bf9\u8bdd\u5f0f\u957f\u63a8\u7406\u6837\u672c\uff0c\u7528\u4e8e\u628a\u201c\u53ef\u8bfb\u3001\u5bf9\u9f50\u201d\u7684\u601d\u7ef4\u98ce\u683c\u5f15\u5165\u7b56\u7565\uff0c\u518d\u4ea4\u7ed9 RL \u653e\u5927\u3002<\/p>\n\n\n\n<p>\u5341\u4e09\u3001\u6279\u5224\u6027\u8ba8\u8bba\u4e0e\u5f00\u653e\u95ee\u9898<br>\u8fd9\u9879\u5de5\u4f5c\u7684\u4ef7\u503c\u5728\u4e8e\u628a\u201c\u9ad8\u8d28\u91cf\u63a8\u7406\u201d\u7684\u6838\u5fc3\u9a71\u52a8\u4ece\u201c\u6a21\u4eff\u4eba\u7c7b\u601d\u7ef4\u8303\u5f0f\u201d\u8f6c\u5411\u201c\u8ba9\u53ef\u9a8c\u8bc1\u4efb\u52a1\u4e0e\u89c4\u6a21\u5316 RL \u9a71\u52a8\u81ea\u6f14\u5316\u201d\u3002\u5176\u8fb9\u754c\u4e5f\u6e05\u6670\uff1a\u5f53\u4efb\u52a1\u96be\u4ee5\u5b9a\u4e49\u7a33\u5b9a\u3001\u6297\u9ed1\u5ba2\u7684\u5956\u52b1\uff08\u5982\u5f00\u653e\u5199\u4f5c\u3001\u4ef7\u503c\u53d6\u5411\u95ee\u9898\uff09\uff0c\u7eaf RL \u4ecd\u6b65\u5c65\u7ef4\u8270\uff1b\u5f53\u8bc4\u6d4b\u5468\u671f\u8fc7\u957f\uff0cRL \u7684\u201c\u4fe1\u606f\u65f6\u5ef6\u201d\u4f1a\u62d6\u6162\u6539\u8fdb\u3002\u4e0b\u4e00\u6b65\u7684\u7ade\u4e89\u7126\u70b9\uff0c\u53ef\u80fd\u5728\u201c\u5982\u4f55\u628a\u66f4\u591a\u73b0\u5b9e\u95ee\u9898\u8f6c\u5316\u4e3a\u53ef\u9a8c\u8bc1\u5b50\u76ee\u6807\u94fe\u201d\uff0c\u4ee5\u53ca\u201c\u5982\u4f55\u628a\u5de5\u5177\u4e0e\u7ed3\u6784\u5316\u8f93\u51fa\u7eb3\u5165\u7aef\u5230\u7aef RL \u95ed\u73af\u201d\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587DeepSeek-R1 incentivizes reasoning in LLMs through re [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","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":"","ast-disable-related-posts":"","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":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"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":""},"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":[8],"tags":[39,96,86],"class_list":["post-6181","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-deepseek","tag-llm"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587DeepSeek-R1 incentivizes reasoning in LLMs through re&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6181","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6181"}],"version-history":[{"count":2,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6181\/revisions"}],"predecessor-version":[{"id":6184,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6181\/revisions\/6184"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6181"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6181"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6181"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}