{"id":5614,"date":"2025-02-09T12:21:07","date_gmt":"2025-02-09T04:21:07","guid":{"rendered":"https:\/\/nullthought.net\/?p=5614"},"modified":"2025-03-03T12:25:31","modified_gmt":"2025-03-03T04:25:31","slug":"deepseek-r1%e6%8a%80%e6%9c%af%e6%8a%a5%e5%91%8a%e6%a6%82%e8%bf%b0","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=5614","title":{"rendered":"DeepSeek-R1\u6280\u672f\u62a5\u544a\u6982\u8ff0"},"content":{"rendered":"\n<p>\u8fd1\u5e74\u6765\uff0c\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u5feb\u901f\u53d1\u5c55\u4f7f\u5176\u5728\u63a8\u7406\u3001\u4ee3\u7801\u751f\u6210\u3001\u79d1\u5b66\u8ba1\u7b97\u7b49\u4efb\u52a1\u4e0a\u5c55\u73b0\u51fa\u8d8a\u6765\u8d8a\u5f3a\u7684\u80fd\u529b\uff0c\u9010\u6b65\u7f29\u5c0f\u4e86\u4e0e\u4eba\u5de5\u901a\u7528\u667a\u80fd\uff08AGI\uff09\u4e4b\u95f4\u7684\u5dee\u8ddd\u3002\u7136\u800c\uff0c\u5373\u4f7f\u662f\u6700\u5148\u8fdb\u7684 LLM\uff0c\u5982 OpenAI \u7684 GPT-4o \u548c Claude-3.5-Sonnet\uff0c\u4ecd\u7136\u5728\u4e00\u4e9b\u9ad8\u9636\u63a8\u7406\u4efb\u52a1\u4e0a\u5b58\u5728\u4e00\u5b9a\u5c40\u9650\u6027\uff0c\u7279\u522b\u662f\u5728\u957f\u94fe\u63a8\u7406\uff08Chain-of-Thought, CoT\uff09\u3001\u6570\u5b66\u63a8\u7406\u3001\u590d\u6742\u7f16\u7a0b\u95ee\u9898\u89e3\u51b3\u7b49\u65b9\u9762\u3002<\/p>\n\n\n\n<p>\u540e\u8bad\u7ec3\uff08Post-training\uff09\u5df2\u6210\u4e3a LLM \u53d1\u5c55\u8fc7\u7a0b\u4e2d\u4e0d\u53ef\u6216\u7f3a\u7684\u90e8\u5206\u3002\u8fd1\u5e74\u6765\u7684\u7814\u7a76\u8868\u660e\uff0c\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning, RL\uff09\u548c\u76d1\u7763\u5fae\u8c03\uff08Supervised Fine-tuning, SFT\uff09\uff0c\u53ef\u4ee5\u6709\u6548\u63d0\u5347\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4f7f\u5176\u66f4\u597d\u5730\u5bf9\u9f50\u7528\u6237\u504f\u597d\uff0c\u5e76\u5728\u63a8\u7406\u4efb\u52a1\u4e0a\u53d6\u5f97\u66f4\u4f18\u8868\u73b0\u3002\u4f8b\u5982\uff0cOpenAI \u7684 o1 \u7cfb\u5217\u6a21\u578b\u901a\u8fc7\u63a8\u7406\u65f6\u52a8\u6001\u6269\u5c55 CoT \u957f\u5ea6\uff0c\u6781\u5927\u5730\u63d0\u5347\u4e86\u5728\u6570\u5b66\u3001\u7f16\u7a0b\u548c\u79d1\u5b66\u63a8\u7406\u65b9\u9762\u7684\u80fd\u529b\u3002\u7136\u800c\uff0c\u76ee\u524d\u7684\u7814\u7a76\u4ecd\u7136\u5b58\u5728\u4e00\u4e9b\u5173\u952e\u6311\u6218\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u63a8\u7406\u80fd\u529b\u7684\u5f3a\u5316\u5b66\u4e60\u4f18\u5316\u4ecd\u7136\u662f\u4e00\u4e2a\u5f00\u653e\u95ee\u9898<\/strong>\u3002\u5c3d\u7ba1\u5df2\u6709\u7814\u7a76\u63a2\u7d22\u4e86\u57fa\u4e8e\u8fc7\u7a0b\u5956\u52b1\u5efa\u6a21\uff08Process-based Reward Modeling, PRM\uff09\u3001\u5f3a\u5316\u5b66\u4e60\u3001\u641c\u7d22\u7b97\u6cd5\uff08\u5982\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22 Monte Carlo Tree Search, MCTS\uff09\u7b49\u65b9\u6cd5\u6765\u63d0\u5347\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u5c1a\u672a\u51fa\u73b0\u80fd\u591f\u5168\u9762\u8d85\u8d8a OpenAI o1 \u7cfb\u5217\u7684\u65b9\u6848\u3002<\/li>\n\n\n\n<li><strong>\u6d4b\u8bd5\u65f6\u63a8\u7406\u80fd\u529b\u7684\u53ef\u6269\u5c55\u6027\u4ecd\u7136\u6709\u9650<\/strong>\u3002\u5728\u5f53\u524d LLM \u7814\u7a76\u4e2d\uff0c\u5982\u4f55\u8ba9\u6a21\u578b\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u52a8\u6001\u5730\u4f18\u5316\u81ea\u8eab\u601d\u7ef4\u94fe\uff0c\u5e76\u4e14\u6709\u6548\u6269\u5c55\u5176\u63a8\u7406\u6df1\u5ea6\uff0c\u4ecd\u7136\u662f\u4e00\u4e2a\u5c1a\u672a\u5b8c\u5168\u89e3\u51b3\u7684\u95ee\u9898\u3002<\/li>\n\n\n\n<li><strong>\u76d1\u7763\u5fae\u8c03\u7684\u4f9d\u8d56<\/strong>\u3002\u5927\u591a\u6570\u63d0\u5347\u63a8\u7406\u80fd\u529b\u7684\u5de5\u4f5c\u90fd\u4f9d\u8d56\u4e8e\u5927\u91cf\u9ad8\u8d28\u91cf\u76d1\u7763\u6570\u636e\u8fdb\u884c\u5fae\u8c03\uff0c\u800c\u83b7\u53d6\u8fd9\u4e9b\u6570\u636e\u5f80\u5f80\u6210\u672c\u9ad8\u6602\uff0c\u5e76\u4e14\u53ef\u80fd\u9650\u5236\u6a21\u578b\u7684\u81ea\u9002\u5e94\u63a8\u7406\u80fd\u529b\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u5728\u6b64\u80cc\u666f\u4e0b\uff0c\u672c\u7814\u7a76\u63d0\u51fa\u4e86 <strong><a href=\"https:\/\/arxiv.org\/abs\/2501.12948\" target=\"_blank\" rel=\"noreferrer noopener\">DeepSeek-R1<\/a><\/strong>\uff0c\u8fd9\u662f\u4e00\u4e2a\u57fa\u4e8e<strong>\u7eaf\u5f3a\u5316\u5b66\u4e60<\/strong>\uff08RL\uff09\u8bad\u7ec3\u7684\u63a8\u7406\u4f18\u5316 LLM\uff0c\u65e8\u5728\u63a2\u7d22\u5982\u4f55\u5728<strong>\u4e0d\u4f9d\u8d56\u76d1\u7763\u6570\u636e<\/strong>\u7684\u60c5\u51b5\u4e0b\uff0c\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u6fc0\u52b1\u6a21\u578b\u81ea\u53d1\u5b66\u4e60\u63a8\u7406\u80fd\u529b\u3002\u6211\u4eec\u9996\u5148\u8bad\u7ec3\u4e86<strong>DeepSeek-R1-Zero<\/strong>\uff0c\u8fd9\u662f\u4e00\u4e2a<strong>\u5b8c\u5168\u7531 RL \u8bad\u7ec3\u7684\u6a21\u578b<\/strong>\uff0c\u5728\u65e0 SFT \u9884\u8bad\u7ec3\u7684\u60c5\u51b5\u4e0b\u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\u3002\u7136\u800c\uff0cDeepSeek-R1-Zero \u4ecd\u7136\u5b58\u5728<strong>\u53ef\u8bfb\u6027\u5dee\u3001\u8bed\u8a00\u6df7\u6742<\/strong>\u7b49\u95ee\u9898\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5f15\u5165\u4e86<strong>\u51b7\u542f\u52a8\u6570\u636e\uff08Cold Start Data\uff09<\/strong>\uff0c\u5e76\u91c7\u7528<strong>\u591a\u9636\u6bb5\u5f3a\u5316\u5b66\u4e60<\/strong>\u8bad\u7ec3\u7b56\u7565\uff0c\u6700\u7ec8\u5f97\u5230\u66f4\u7a33\u5b9a\u4e14\u63a8\u7406\u80fd\u529b\u66f4\u5f3a\u7684 <strong>DeepSeek-R1<\/strong>\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u8fd8\u63a2\u7d22\u4e86<strong><a href=\"https:\/\/nullthought.net\/?p=5098\" target=\"_blank\" rel=\"noreferrer noopener\">\u84b8\u998f\uff08Distillation\uff09\u6280\u672f<\/a><\/strong>\uff0c\u4ee5\u4fbf\u5c06 DeepSeek-R1 \u7684\u63a8\u7406\u80fd\u529b\u8fc1\u79fb\u5230\u66f4\u5c0f\u7684\u6a21\u578b\uff0c\u5982 <strong>Qwen \u548c Llama<\/strong>\uff0c\u4f7f\u5f97\u5c0f\u578b\u6a21\u578b\u4e5f\u80fd\u83b7\u5f97\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\u3002\u7814\u7a76\u8868\u660e\uff0c\u84b8\u998f\u7684\u5c0f\u6a21\u578b\u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8d85\u8fc7\u4e86\u5176\u4ed6\u540c\u7c7b\u5f00\u6e90\u6a21\u578b\uff0c\u5982 <strong>QwQ-32B-Preview<\/strong>\uff0c\u751a\u81f3\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\u4e0e OpenAI-o1-mini \u76f8\u5ab2\u7f8e\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"990\" height=\"624\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/02\/image-4.png\" alt=\"\" class=\"wp-image-5615\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/02\/image-4.png 990w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/02\/image-4-300x189.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/02\/image-4-768x484.png 768w\" sizes=\"auto, (max-width: 990px) 100vw, 990px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2501.12948\" target=\"_blank\" rel=\"noreferrer noopener\">DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning<\/a><\/strong><br>Credit: <strong><a href=\"https:\/\/www.deepseek.com\/\" target=\"_blank\" rel=\"noreferrer noopener\">Deepseek<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u672c\u7814\u7a76\u7684\u4e3b\u8981\u8d21\u732e\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u7eaf\u5f3a\u5316\u5b66\u4e60\u9a71\u52a8\u7684\u63a8\u7406\u80fd\u529b\u4f18\u5316<\/strong>\uff1aDeepSeek-R1-Zero \u662f\u9996\u4e2a\u5b8c\u5168\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u7684 LLM\uff0c\u5728\u63a8\u7406\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u663e\u8457\u63d0\u5347\uff0c\u800c\u65e0\u9700 SFT \u9884\u8bad\u7ec3\u3002<\/li>\n\n\n\n<li><strong>\u591a\u9636\u6bb5\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u6846\u67b6<\/strong>\uff1aDeepSeek-R1 \u91c7\u7528\u51b7\u542f\u52a8\u6570\u636e\u548c\u591a\u8f6e RL \u8fed\u4ee3\uff0c\u6700\u7ec8\u5728\u63a8\u7406\u4efb\u52a1\u4e0a\u8fbe\u5230\u4e86 OpenAI-o1-1217 \u7684\u6c34\u5e73\u3002<\/li>\n\n\n\n<li><strong>\u9ad8\u6548\u7684\u84b8\u998f\u65b9\u6cd5<\/strong>\uff1a\u6211\u4eec\u5c06 DeepSeek-R1 \u7684\u63a8\u7406\u80fd\u529b\u84b8\u998f\u81f3\u5c0f\u578b\u6a21\u578b\uff081.5B-70B \u53c2\u6570\u89c4\u6a21\uff09\uff0c\u5e76\u53d1\u73b0\u84b8\u998f\u7684\u5c0f\u6a21\u578b\u80fd\u591f\u663e\u8457\u8d85\u8d8a\u5e38\u89c4 LLM \u8bad\u7ec3\u65b9\u6cd5\u6240\u80fd\u8fbe\u5230\u7684\u63a8\u7406\u6c34\u5e73\u3002<\/li>\n<\/ol>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>\u4e00\u3001DeepSeek-R1-Zero\uff1a\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u7684\u63a8\u7406\u80fd\u529b\u81ea\u6211\u8fdb\u5316<\/strong><\/h5>\n\n\n\n<p>DeepSeek-R1-Zero \u662f\u5728 <strong><a href=\"https:\/\/nullthought.net\/?p=5463\" target=\"_blank\" rel=\"noreferrer noopener\">DeepSeek-V3-Base<\/a><\/strong>\uff08\u57fa\u7840\u6a21\u578b\uff09\u4e0a\u76f4\u63a5\u5e94\u7528\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u8bad\u7ec3\u51fa\u7684\u63a8\u7406\u6a21\u578b\uff0c\u5176<strong>\u6838\u5fc3\u7279\u70b9\u662f\u5b8c\u5168\u4e0d\u4f9d\u8d56\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09<\/strong>\u3002\u8fd9\u610f\u5473\u7740\u8be5\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u5b8c\u5168\u901a\u8fc7 RL \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u81ea\u6211\u8fdb\u5316\u4ea7\u751f\uff0c\u800c\u4e0d\u662f\u4f9d\u8d56\u4eba\u7c7b\u63d0\u4f9b\u7684\u9ad8\u8d28\u91cf\u8bad\u7ec3\u6570\u636e\u3002<\/p>\n\n\n\n<p>\u8bad\u7ec3 DeepSeek-R1-Zero \u7684\u6838\u5fc3\u65b9\u6cd5\u5305\u62ec\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u91c7\u7528<strong>\u7ec4\u76f8\u5bf9\u7b56\u7565\u4f18\u5316\uff08Group Relative Policy Optimization, GRPO\uff09<\/strong>\uff0c\u8be5\u65b9\u6cd5\u4e0d\u4f7f\u7528\u4f20\u7edf\u7684\u4ef7\u503c\u7f51\u7edc\uff0c\u800c\u662f\u901a\u8fc7\u5bf9\u6bd4\u591a\u4e2a\u7b56\u7565\u8f93\u51fa\u7684\u5f97\u5206\u6765\u8fdb\u884c\u4f18\u5316\u3002\u8fd9\u79cd\u65b9\u6cd5\u51cf\u5c11\u4e86\u8ba1\u7b97\u6210\u672c\uff0c\u4f7f RL \u8bad\u7ec3\u66f4\u52a0\u9ad8\u6548\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u5956\u52b1\u5efa\u6a21\uff08Reward Modeling\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li><strong>\u51c6\u786e\u6027\u5956\u52b1\uff08Accuracy Rewards\uff09<\/strong>\uff1a\u5728\u6570\u5b66\u3001\u7f16\u7a0b\u7b49\u4efb\u52a1\u4e0a\uff0c\u4f7f\u7528\u89c4\u5219\u68c0\u6d4b\u7b54\u6848\u662f\u5426\u6b63\u786e\u3002\u4f8b\u5982\uff0c\u6570\u5b66\u95ee\u9898\u8981\u6c42\u6a21\u578b\u63d0\u4f9b\u6700\u7ec8\u7b54\u6848\uff0c\u5e76\u4f7f\u7528\u6807\u51c6\u683c\u5f0f\u8fdb\u884c\u68c0\u67e5\uff0c\u800c\u7f16\u7a0b\u4efb\u52a1\u5219\u5229\u7528\u7f16\u8bd1\u5668\u81ea\u52a8\u9a8c\u8bc1\u4ee3\u7801\u8f93\u51fa\u7684\u6b63\u786e\u6027\u3002<\/li>\n\n\n\n<li><strong>\u683c\u5f0f\u5956\u52b1\uff08Format Rewards\uff09<\/strong>\uff1a\u8bad\u7ec3\u65f6\u8981\u6c42\u6a21\u578b\u5c06\u63a8\u7406\u8fc7\u7a0b\u5305\u88f9\u5728 <code>&lt;think&gt;...&lt;\/think&gt;<\/code> \u6807\u7b7e\u4e2d\uff0c\u5e76\u5c06\u6700\u7ec8\u7b54\u6848\u5305\u88f9\u5728 <code>&lt;answer&gt;...&lt;\/answer&gt;<\/code> \u4e2d\uff0c\u4ee5\u786e\u4fdd\u8f93\u51fa\u7ed3\u6784\u7684\u89c4\u8303\u6027\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u81ea\u6211\u8fdb\u5316\u8fc7\u7a0b<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0cDeepSeek-R1-Zero \u5728 AIME 2024 \u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684 <strong>pass@1 \u51c6\u786e\u7387\u4ece 15.6% \u63d0\u5347\u81f3 71.0%<\/strong>\uff0c\u5e76\u5728\u591a\u6570\u6295\u7968\uff08majority voting\uff09\u6a21\u5f0f\u4e0b\u8fdb\u4e00\u6b65\u63d0\u5347\u81f3 <strong>86.7%<\/strong>\uff0c\u8fbe\u5230 OpenAI-o1-0912 \u7684\u6c34\u5e73\u3002<\/li>\n\n\n\n<li>\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u81ea\u7136\u6d8c\u73b0\u4e86<strong>\u53cd\u601d\uff08Reflection\uff09<\/strong>\u3001<strong>\u81ea\u6211\u9a8c\u8bc1\uff08Self-verification\uff09<\/strong> \u7b49\u9ad8\u7ea7\u63a8\u7406\u80fd\u529b\uff0c\u8868\u73b0\u51fa\u7c7b\u4f3c\u201c\u987f\u609f\u201d\uff08Aha moment\uff09\u7684\u884c\u4e3a\uff0c\u5373\u6a21\u578b\u80fd\u591f\u81ea\u53d1\u5730\u91cd\u65b0\u5ba1\u89c6\u81ea\u5df1\u7684\u63a8\u7406\u8fc7\u7a0b\uff0c\u5e76\u5728\u5fc5\u8981\u65f6\u8fdb\u884c\u7ea0\u6b63\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p>\u5c3d\u7ba1 DeepSeek-R1-Zero \u5c55\u73b0\u51fa\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u5b83\u4ecd\u7136\u5b58\u5728\u4e00\u4e9b\u660e\u663e\u7684\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u53ef\u8bfb\u6027\u8f83\u5dee<\/strong>\uff1a\u63a8\u7406\u8fc7\u7a0b\u5f80\u5f80\u6df7\u6742\u591a\u79cd\u8bed\u8a00\uff0c\u5f71\u54cd\u7406\u89e3\u3002<\/li>\n\n\n\n<li><strong>\u8bed\u8a00\u6df7\u6742<\/strong>\uff1a\u4e0d\u540c\u8bed\u8a00\u7684\u6df7\u5408\u8f93\u51fa\u5bfc\u81f4\u6587\u672c\u96be\u4ee5\u9605\u8bfb\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>\u4e8c\u3001DeepSeek-R1\uff1a\u7ed3\u5408\u51b7\u542f\u52a8\u6570\u636e\u4e0e\u5f3a\u5316\u5b66\u4e60\u7684\u589e\u5f3a\u7248\u6a21\u578b<\/strong><\/h5>\n\n\n\n<p>\u4e3a\u4e86\u89e3\u51b3 DeepSeek-R1-Zero \u7684\u53ef\u8bfb\u6027\u95ee\u9898\uff0c\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86 <strong>DeepSeek-R1<\/strong>\uff0c\u5e76\u5f15\u5165\u4e86<strong>\u51b7\u542f\u52a8\u6570\u636e\uff08Cold Start Data\uff09\u548c\u591a\u9636\u6bb5\u8bad\u7ec3\u7b56\u7565<\/strong>\u3002DeepSeek-R1 \u7684\u8bad\u7ec3\u6d41\u7a0b\u5305\u62ec\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u51b7\u542f\u52a8\u6570\u636e\u7684\u5f15\u5165<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u91c7\u96c6\u5305\u542b<strong>\u8be6\u7ec6 Chain-of-Thought\uff08CoT\uff09<\/strong> \u7684\u9ad8\u8d28\u91cf\u6570\u636e\uff0c\u5bf9 DeepSeek-V3-Base \u8fdb\u884c\u5fae\u8c03\uff0c\u4ee5\u63d0\u4f9b\u4e00\u4e2a\u66f4\u7a33\u5b9a\u7684 RL \u8bad\u7ec3\u8d77\u70b9\u3002<\/li>\n\n\n\n<li>\u901a\u8fc7<strong>\u4eba\u5de5\u7b5b\u9009<\/strong>\u4e0e<strong>\u540e\u5904\u7406<\/strong>\u786e\u4fdd\u6570\u636e\u53ef\u8bfb\u6027\uff0c\u907f\u514d\u8bed\u8a00\u6df7\u6742\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u63a8\u7406\u5f3a\u5316\u5b66\u4e60\uff08Reasoning-Oriented RL\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u91c7\u7528\u7c7b\u4f3c DeepSeek-R1-Zero \u7684 RL \u8bad\u7ec3\u65b9\u6cd5\uff0c\u5e76\u989d\u5916\u5f15\u5165<strong>\u8bed\u8a00\u4e00\u81f4\u6027\u5956\u52b1<\/strong>\uff08\u9632\u6b62\u4e0d\u540c\u8bed\u8a00\u6df7\u6742\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u62d2\u7edd\u91c7\u6837\u4e0e\u76d1\u7763\u5fae\u8c03\uff08Rejection Sampling &amp; SFT\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u4ece RL \u751f\u6210\u7684\u6837\u672c\u4e2d\u7b5b\u9009\u9ad8\u8d28\u91cf\u63a8\u7406\u6570\u636e\uff0c\u8fdb\u884c\u989d\u5916\u7684 SFT \u8bad\u7ec3\u3002<\/li>\n\n\n\n<li>\u989d\u5916\u52a0\u5165\u975e\u63a8\u7406\u4efb\u52a1\uff08\u5982\u5199\u4f5c\u3001\u95ee\u7b54\u7b49\uff09\u7684\u6570\u636e\uff0c\u4ee5\u63d0\u5347\u6a21\u578b\u7684\u901a\u7528\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u6700\u7ec8\u5f3a\u5316\u5b66\u4e60\u4f18\u5316\uff08Final RL Fine-tuning\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728\u5168\u9886\u57df\u4efb\u52a1\uff08\u63a8\u7406\u3001\u5199\u4f5c\u3001\u7f16\u7a0b\u7b49\uff09\u4e0a\u5e94\u7528 RL\uff0c\u4f7f\u6a21\u578b\u66f4\u597d\u5730\u5bf9\u9f50\u4eba\u7c7b\u504f\u597d\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p>\u6700\u7ec8\uff0cDeepSeek-R1 \u5728\u591a\u4e2a\u63a8\u7406\u4efb\u52a1\u4e0a\u8fbe\u5230\u4e86 OpenAI-o1-1217 \u7684\u6c34\u5e73\uff0c\u5e76\u5728\u90e8\u5206\u4efb\u52a1\u4e0a\u8d85\u8d8a\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>\u4e09\u3001\u84b8\u998f\u5c0f\u6a21\u578b<\/strong><\/h5>\n\n\n\n<p>\u7814\u7a76\u56e2\u961f\u8fd8\u5c06 DeepSeek-R1 \u7684\u63a8\u7406\u80fd\u529b\u84b8\u998f\u5230\u66f4\u5c0f\u7684\u6a21\u578b\uff0c\u5982 Qwen \u548c Llama\u3002\u7ed3\u679c\u8868\u660e\uff0c\u84b8\u998f\u7684\u5c0f\u6a21\u578b\u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8d85\u8fc7\u4e86\u5176\u4ed6\u540c\u7c7b\u5f00\u6e90\u6a21\u578b\uff0c\u5982 QwQ-32B-Preview\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>\u56db\u3001\u5b9e\u9a8c\u4e0e\u8bc4\u4f30<\/strong><\/h5>\n\n\n\n<p>\u7814\u7a76\u56e2\u961f\u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u8bc4\u4f30\u4e86 DeepSeek-R1 \u53ca\u5176\u84b8\u998f\u6a21\u578b\uff0c\u4e3b\u8981\u5305\u62ec\u6570\u5b66\u63a8\u7406\u3001\u4ee3\u7801\u63a8\u7406\u3001\u77e5\u8bc6\u95ee\u7b54\u4ee5\u53ca\u901a\u7528\u6587\u672c\u751f\u6210\u4efb\u52a1\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cDeepSeek-R1 \u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8fbe\u5230\u4e86\u6216\u8d85\u8d8a\u4e86 OpenAI-o1-1217 \u7684\u6c34\u5e73\uff0c\u800c\u5176\u84b8\u998f\u6a21\u578b\u5728\u5c0f\u578b\u53c2\u6570\u89c4\u6a21\u4e0b\u4e5f\u5c55\u73b0\u51fa\u5f3a\u52b2\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n\n\n\n<h6 class=\"wp-block-heading\">1.<strong>DeepSeek-R1 \u8bc4\u4f30<\/strong><\/h6>\n\n\n\n<p>DeepSeek-R1 \u5728\u591a\u4e2a\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u7684\u8868\u73b0\u5747\u4f18\u4e8e DeepSeek-R1-Zero\uff0c\u5e76\u5728\u90e8\u5206\u4efb\u52a1\u4e0a\u8d85\u8d8a OpenAI-o1-1217\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6570\u5b66\u63a8\u7406\uff08AIME 2024, MATH-500\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728 AIME 2024\uff08\u7f8e\u56fd\u6570\u5b66\u9080\u8bf7\u8d5b\uff09\u4e0a\uff0cDeepSeek-R1 \u7684 <strong>pass@1 \u5f97\u5206\u8fbe\u5230 79.8%<\/strong>\uff0c\u8d85\u8d8a OpenAI-o1-1217\uff0879.2%\uff09\u3002<\/li>\n\n\n\n<li>\u5728 MATH-500 \u6570\u636e\u96c6\u4e0a\uff0cDeepSeek-R1 \u7684 <strong>pass@1 \u5f97\u5206\u9ad8\u8fbe 97.3%<\/strong>\uff0c\u7565\u9ad8\u4e8e OpenAI-o1-1217\uff0896.4%\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u4ee3\u7801\u63a8\u7406\uff08Codeforces, LiveCodeBench\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728 Codeforces \u7f16\u7a0b\u7ade\u8d5b\u4efb\u52a1\u4e0a\uff0cDeepSeek-R1 \u53d6\u5f97 <strong>96.3% \u7684\u4eba\u7c7b\u6392\u540d\u767e\u5206\u6bd4\uff08percentile\uff09\uff0cElo \u8bc4\u5206\u4e3a 2029<\/strong>\uff0c\u4e0e OpenAI-o1-1217\uff08Elo 2061\uff09\u63a5\u8fd1\u3002<\/li>\n\n\n\n<li>\u5728 LiveCodeBench \u4ee3\u7801\u751f\u6210\u4efb\u52a1\u4e0a\uff0cDeepSeek-R1 \u7684 <strong>pass@1 \u5f97\u5206\u4e3a 65.9%<\/strong>\uff0c\u4f18\u4e8e OpenAI-o1-mini\uff0853.8%\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u77e5\u8bc6\u95ee\u7b54\uff08MMLU, GPQA Diamond\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728 MMLU\uff08\u591a\u4efb\u52a1\u8bed\u8a00\u7406\u89e3\uff09\u6d4b\u8bd5\u4e0a\uff0cDeepSeek-R1 \u53d6\u5f97 <strong>90.8% pass@1<\/strong>\uff0c\u7565\u4f4e\u4e8e OpenAI-o1-1217\uff0891.8%\uff09\uff0c\u4f46\u663e\u8457\u4f18\u4e8e DeepSeek-V3\uff0888.5%\uff09\u3002<\/li>\n\n\n\n<li>\u5728 GPQA Diamond\uff08\u9ad8\u96be\u5ea6\u77e5\u8bc6\u95ee\u7b54\uff09\u6d4b\u8bd5\u4e2d\uff0cDeepSeek-R1 \u53d6\u5f97 <strong>71.5% pass@1<\/strong>\uff0c\u8d85\u8d8a DeepSeek-V3\uff0859.1%\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u5f00\u653e\u5f0f\u6587\u672c\u751f\u6210<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728 AlpacaEval 2.0 \u6d4b\u8bd5\u4e2d\uff0cDeepSeek-R1 \u4ee5 <strong>87.6% \u7684\u80dc\u7387<\/strong> \u8d85\u8d8a Claude-3.5 \u548c GPT-4o\u3002<\/li>\n\n\n\n<li>\u5728 Arena-Hard \u6d4b\u8bd5\u4e2d\uff0cDeepSeek-R1 \u4ee5 <strong>92.3% \u7684\u80dc\u7387<\/strong> \u8d85\u8d8a GPT-4o \u548c OpenAI-o1-mini\uff0c\u5c55\u73b0\u51fa\u5728\u901a\u7528\u4efb\u52a1\u4e0a\u7684\u5f3a\u5927\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<h6 class=\"wp-block-heading\">2.<strong>\u84b8\u998f\u5c0f\u6a21\u578b\u8bc4\u4f30<\/strong><\/h6>\n\n\n\n<p>\u7814\u7a76\u56e2\u961f\u8fd8\u5c06 DeepSeek-R1 \u7684\u63a8\u7406\u80fd\u529b\u84b8\u998f\u5230\u66f4\u5c0f\u7684\u6a21\u578b\uff0c\u5982 <strong>Qwen \u548c Llama<\/strong>\uff0c\u53d1\u73b0\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>DeepSeek-R1-Distill-Qwen-7B \u8d85\u8d8a QwQ-32B-Preview<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728 AIME 2024 \u4e0a\uff0cDeepSeek-R1-Distill-Qwen-7B \u53d6\u5f97 <strong>55.5% pass@1<\/strong>\uff0c\u663e\u8457\u8d85\u8fc7 QwQ-32B-Preview\uff0850.0%\uff09\u3002<\/li>\n\n\n\n<li>\u5728 MATH-500 \u4e0a\uff0cDeepSeek-R1-Distill-Qwen-7B \u53d6\u5f97 <strong>92.8% pass@1<\/strong>\uff0c\u8d85\u8fc7 OpenAI-o1-mini\uff0890.0%\uff09\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>DeepSeek-R1-Distill-Qwen-32B \u548c DeepSeek-R1-Distill-Llama-70B \u8d85\u8d8a OpenAI-o1-mini<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u5728 GPQA Diamond\u3001LiveCodeBench \u4ee5\u53ca Codeforces \u7b49\u63a8\u7406\u4efb\u52a1\u4e0a\uff0cDeepSeek-R1 \u7684\u84b8\u998f\u6a21\u578b\u5747\u8d85\u8d8a\u4e86 OpenAI-o1-mini\uff0c\u5e76\u4e0e OpenAI-o1-1217 \u63a5\u8fd1\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c<strong>\u84b8\u998f\u540e\u7684\u5c0f\u6a21\u578b\u5728\u63a8\u7406\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u5927\u5e45\u8d85\u8d8a\u540c\u7b49\u53c2\u6570\u89c4\u6a21\u7684 LLM<\/strong>\uff0c\u8868\u660e DeepSeek-R1 \u7684\u63a8\u7406\u80fd\u529b\u53ef\u4ee5\u9ad8\u6548\u5730\u8fc1\u79fb\u5230\u5c0f\u6a21\u578b\u4e0a\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>\u4e94\u3001\u8ba8\u8bba<\/strong><\/h5>\n\n\n\n<h6 class=\"wp-block-heading\">1.<strong>\u84b8\u998f vs \u5f3a\u5316\u5b66\u4e60<\/strong><\/h6>\n\n\n\n<p>\u7814\u7a76\u56e2\u961f\u53d1\u73b0\uff0c<strong>\u76f4\u63a5\u5bf9\u5c0f\u6a21\u578b\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u7684\u6548\u679c\u4e0d\u5982\u84b8\u998f<\/strong>\u3002\u5b9e\u9a8c\u8868\u660e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u76f4\u63a5\u4f7f\u7528 RL \u8bad\u7ec3\u7684 32B \u6a21\u578b\uff08DeepSeek-R1-Zero-Qwen-32B\uff09\uff0c\u5176\u63a8\u7406\u80fd\u529b\u4ec5\u4e0e QwQ-32B-Preview \u6301\u5e73<\/strong>\u3002<\/li>\n\n\n\n<li><strong>\u800c\u84b8\u998f\u540e\u7684 DeepSeek-R1-Distill-Qwen-32B \u5219\u663e\u8457\u4f18\u4e8e\u524d\u8005\uff0c\u5728\u591a\u4e2a\u63a8\u7406\u4efb\u52a1\u4e0a\u90fd\u8d85\u8fc7\u4e86 QwQ-32B-Preview<\/strong>\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e00\u53d1\u73b0\u8868\u660e\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u84b8\u998f\u53ef\u4ee5\u9ad8\u6548\u5730\u5c06\u5927\u578b\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u8fc1\u79fb\u5230\u5c0f\u578b\u6a21\u578b\u4e0a<\/strong>\uff0c\u5e76\u4e14\u6bd4\u76f4\u63a5\u5bf9\u5c0f\u6a21\u578b\u8fdb\u884c RL \u8bad\u7ec3\u66f4\u9ad8\u6548\u3002<\/li>\n\n\n\n<li><strong>\u5c0f\u6a21\u578b\u4ecd\u7136\u53d7\u5230\u8ba1\u7b97\u8d44\u6e90\u7684\u9650\u5236<\/strong>\uff0c\u5373\u4f7f\u5e94\u7528\u5927\u89c4\u6a21 RL \u8bad\u7ec3\uff0c\u6548\u679c\u4ecd\u4e0d\u53ca\u4ece\u5927\u578b\u6a21\u578b\u84b8\u998f\u800c\u6765\u7684\u7248\u672c\u3002<\/li>\n<\/ol>\n\n\n\n<h6 class=\"wp-block-heading\">2.<strong>\u5931\u8d25\u5c1d\u8bd5<\/strong><\/h6>\n\n\n\n<p>\u5728\u7814\u7a76\u8fc7\u7a0b\u4e2d\uff0c\u7814\u7a76\u56e2\u961f\u8fd8\u5c1d\u8bd5\u4e86\u4e00\u4e9b\u65b9\u6cd5\uff0c\u4f46\u53d1\u73b0\u5b83\u4eec\u5728\u5927\u89c4\u6a21\u63a8\u7406\u4efb\u52a1\u4e2d\u7684\u6548\u679c\u6709\u9650\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u8fc7\u7a0b\u5956\u52b1\u6a21\u578b\uff08PRM\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u8be5\u65b9\u6cd5\u8bd5\u56fe\u901a\u8fc7\u5956\u52b1\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u6bcf\u4e2a\u4e2d\u95f4\u6b65\u9aa4\u6765\u5f15\u5bfc\u6a21\u578b\u3002\u7136\u800c\uff0c\u5728\u901a\u7528\u63a8\u7406\u4efb\u52a1\u4e0a\uff0c\u7cbe\u786e\u5b9a\u4e49\u6bcf\u4e00\u6b65\u662f\u5426\u6b63\u786e\u975e\u5e38\u56f0\u96be\uff0c\u5e76\u4e14\u5bb9\u6613\u5bfc\u81f4<strong>\u5956\u52b1\u4f5c\u5f0a\uff08reward hacking\uff09<\/strong>\uff0c\u5373\u6a21\u578b\u901a\u8fc7\u67d0\u4e9b\u6377\u5f84\u83b7\u53d6\u9ad8\u5206\uff0c\u800c\u975e\u771f\u6b63\u4f18\u5316\u63a8\u7406\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22\uff08MCTS\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u53d7 AlphaGo \u542f\u53d1\uff0c\u7814\u7a76\u56e2\u961f\u5c1d\u8bd5\u4f7f\u7528 MCTS \u8fdb\u884c\u63a8\u7406\u4f18\u5316\u3002\u7136\u800c\uff0c\u8bed\u8a00\u4efb\u52a1\u7684\u641c\u7d22\u7a7a\u95f4\u8fdc\u5927\u4e8e\u56f4\u68cb\u6216\u56fd\u9645\u8c61\u68cb\uff0c\u5bfc\u81f4 MCTS \u96be\u4ee5\u6269\u5c55\u3002\u6b64\u5916\uff0c\u7531\u4e8e MCTS \u9700\u8981\u4f9d\u8d56\u7cbe\u786e\u7684\u4ef7\u503c\u7f51\u7edc\uff0c\u800c\u5f53\u524d\u8bed\u8a00\u6a21\u578b\u7684\u4ef7\u503c\u4f30\u8ba1\u4ecd\u7136\u5b58\u5728\u4e0d\u7a33\u5b9a\u6027\uff0c\u56e0\u6b64\u6700\u7ec8\u672a\u80fd\u6210\u529f\u63d0\u5347\u63a8\u7406\u80fd\u529b\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<h5 class=\"wp-block-heading\"><strong>\u516d\u3001\u7ed3\u8bba\u4e0e\u672a\u6765\u65b9\u5411<\/strong><\/h5>\n\n\n\n<p>\u672c\u7814\u7a76\u8bc1\u660e\u4e86<strong>\u7eaf\u5f3a\u5316\u5b66\u4e60\u53ef\u4ee5\u6709\u6548\u63d0\u5347 LLM \u7684\u63a8\u7406\u80fd\u529b<\/strong>\uff0c\u5e76\u63d0\u51fa\u4e86\u4e00\u79cd\u7ed3\u5408\u51b7\u542f\u52a8\u6570\u636e\u548c\u591a\u9636\u6bb5 RL \u8bad\u7ec3\u7684\u65b9\u6cd5\uff0c\u6700\u7ec8\u8bad\u7ec3\u51fa\u6027\u80fd\u63a5\u8fd1 OpenAI-o1-1217 \u7684 DeepSeek-R1\u3002\u6b64\u5916\uff0c\u7814\u7a76\u56e2\u961f\u8fd8\u63a2\u7d22\u4e86<strong>\u84b8\u998f\u7b56\u7565<\/strong>\uff0c\u5e76\u6210\u529f\u5c06 DeepSeek-R1 \u7684\u63a8\u7406\u80fd\u529b\u8fc1\u79fb\u5230\u5c0f\u578b\u6a21\u578b\uff0c\u663e\u8457\u63d0\u5347\u4e86\u5c0f\u6a21\u578b\u7684\u63a8\u7406\u6c34\u5e73\u3002<\/p>\n\n\n\n<p>\u672a\u6765\uff0c\u7814\u7a76\u56e2\u961f\u8ba1\u5212\u7ee7\u7eed\u4f18\u5316\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u901a\u7528\u80fd\u529b<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u76ee\u524d DeepSeek-R1 \u5728\u51fd\u6570\u8c03\u7528\uff08function calling\uff09\u3001\u591a\u8f6e\u5bf9\u8bdd\uff08multi-turn dialogue\uff09\u3001\u590d\u6742\u89d2\u8272\u626e\u6f14\uff08complex role-playing\uff09\u7b49\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u4ecd\u6709\u63d0\u5347\u7a7a\u95f4\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u8bed\u8a00\u4e00\u81f4\u6027<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u76ee\u524d DeepSeek-R1 \u4e3b\u8981\u9488\u5bf9\u4e2d\u82f1\u6587\u4f18\u5316\uff0c\u5728\u5904\u7406\u5176\u4ed6\u8bed\u8a00\u65f6\u53ef\u80fd\u4f1a\u51fa\u73b0\u8bed\u8a00\u6df7\u6742\u7684\u60c5\u51b5\u3002\u672a\u6765\u8ba1\u5212\u6269\u5c55\u81f3\u591a\u8bed\u8a00\u652f\u6301\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u63d0\u793a\u5de5\u7a0b\uff08Prompt Engineering\uff09<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u7814\u7a76\u53d1\u73b0\uff0cDeepSeek-R1 \u5bf9\u63d0\u793a\u8bcd\uff08prompt\uff09\u8f83\u4e3a\u654f\u611f\uff0c\u7279\u522b\u662f\u5c11\u6837\u672c\uff08few-shot\uff09\u63d0\u793a\u5f80\u5f80\u4f1a\u964d\u4f4e\u63a8\u7406\u8868\u73b0\u3002\u56e0\u6b64\uff0c\u672a\u6765\u5c06\u4f18\u5316\u63d0\u793a\u8bbe\u8ba1\uff0c\u4f7f\u6a21\u578b\u5728\u4e0d\u540c\u63d0\u793a\u4e0b\u4fdd\u6301\u7a33\u5b9a\u8868\u73b0\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u8f6f\u4ef6\u5de5\u7a0b\u4efb\u52a1<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li>\u7531\u4e8e\u4ee3\u7801\u8bc4\u4f30\u65f6\u95f4\u8f83\u957f\uff0c\u5bfc\u81f4\u5927\u89c4\u6a21 RL \u8bad\u7ec3\u5728\u8f6f\u4ef6\u5de5\u7a0b\u4efb\u52a1\u4e0a\u7684\u4f18\u5316\u4ecd\u7136\u6709\u9650\u3002\u672a\u6765\u8ba1\u5212\u901a\u8fc7<strong>\u62d2\u7edd\u91c7\u6837\uff08Rejection Sampling\uff09\u6216\u5f02\u6b65\u8bc4\u4f30<\/strong>\u6765\u63d0\u5347\u6a21\u578b\u5728\u8f6f\u4ef6\u5de5\u7a0b\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<p>DeepSeek-R1 \u53ca\u5176\u84b8\u998f\u6a21\u578b\u7684\u5f00\u6e90\u9884\u8ba1\u5c06\u8fdb\u4e00\u6b65\u63a8\u52a8 LLM \u5728\u63a8\u7406\u4efb\u52a1\u4e0a\u7684\u53d1\u5c55\uff0c\u5e76\u4e3a\u7814\u7a76\u793e\u533a\u63d0\u4f9b\u66f4\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\u5de5\u5177\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<p>\u53c2\u8003\u9605\u8bfb\uff1a<strong><a href=\"https:\/\/nullthought.net\/?p=5627\" target=\"_blank\" rel=\"noreferrer noopener\">OpenAI o3-mini \u7cfb\u7edf\u8bf4\u660e\uff08OpenAI o3-mini System Card\uff09<\/a><\/strong><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8fd1\u5e74\u6765\uff0c\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u5feb\u901f\u53d1\u5c55\u4f7f\u5176\u5728\u63a8\u7406\u3001\u4ee3\u7801\u751f\u6210\u3001\u79d1\u5b66\u8ba1\u7b97\u7b49\u4efb\u52a1\u4e0a\u5c55\u73b0\u51fa\u8d8a\u6765\u8d8a\u5f3a\u7684\u80fd\u529b\uff0c\u9010\u6b65\u7f29 [&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":[39,96,86,97],"class_list":["post-5614","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-deepseek","tag-llm","tag-distillation"],"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":"\u8fd1\u5e74\u6765\uff0c\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u5feb\u901f\u53d1\u5c55\u4f7f\u5176\u5728\u63a8\u7406\u3001\u4ee3\u7801\u751f\u6210\u3001\u79d1\u5b66\u8ba1\u7b97\u7b49\u4efb\u52a1\u4e0a\u5c55\u73b0\u51fa\u8d8a\u6765\u8d8a\u5f3a\u7684\u80fd\u529b\uff0c\u9010\u6b65\u7f29&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5614","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=5614"}],"version-history":[{"count":2,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5614\/revisions"}],"predecessor-version":[{"id":5630,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5614\/revisions\/5630"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5614"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5614"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5614"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}