{"id":38180,"date":"2025-02-11T14:00:52","date_gmt":"2025-02-11T06:00:52","guid":{"rendered":"https:\/\/17aitech.com\/?p=38180"},"modified":"2025-02-11T14:00:52","modified_gmt":"2025-02-11T06:00:52","slug":"sebastian-raschka%ef%bc%9a%e5%85%b3%e4%ba%8edeepseek-r1%e5%92%8c%e6%8e%a8%e7%90%86%e6%a8%a1%e5%9e%8b%ef%bc%8c%e6%88%91%e6%9c%89%e5%87%a0%e7%82%b9%e7%9c%8b%e6%b3%95","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=38180","title":{"rendered":"Sebastian Raschka\uff1a\u5173\u4e8eDeepSeek R1\u548c\u63a8\u7406\u6a21\u578b\uff0c\u6211\u6709\u51e0\u70b9\u770b\u6cd5"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-02-09-4\" target=\"_blank\">Sebastian Raschka\uff1a\u5173\u4e8eDeepSeek R1\u548c\u63a8\u7406\u6a21\u578b\uff0c\u6211\u6709\u51e0\u70b9\u770b\u6cd5<\/a><\/p>\n<section>\u8457\u540d AI \u7814\u7a76\u8005\u548c\u535a\u4e3b Sebastian Raschka \u53c8\u66f4\u65b0\u535a\u5ba2\u4e86\u3002\u8fd9\u4e00\u6b21\uff0c\u4ed6\u5c06\u7acb\u8db3\u4e8e DeepSeek \u6280\u672f\u62a5\u544a\uff0c\u4ecb\u7ecd\u7528\u4e8e\u6784\u5efa\u63a8\u7406\u6a21\u578b\u7684\u56db\u79cd\u4e3b\u8981\u65b9\u6cd5\uff0c\u4e5f\u5c31\u662f\u5982\u4f55\u901a\u8fc7\u63a8\u7406\u80fd\u529b\u6765\u589e\u5f3a LLM\u3002Sebastian Raschka \u8868\u793a\uff1a\u300c\u6211\u5e0c\u671b\u8fd9\u80fd\u63d0\u4f9b\u6709\u4ef7\u503c\u7684\u89c1\u89e3\uff0c\u5e76\u5e2e\u52a9\u4f60\u4e86\u89e3\u56f4\u7ed5\u8fd9\u4e00\u4e3b\u9898\u7684\u5feb\u901f\u6f14\u53d8\u7684\u6587\u732e\u548c\u8bdd\u9898\u7092\u4f5c\u3002\u300d<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-15d4c381c2263f771daf2ecd7558e848.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-15d4c381c2263f771daf2ecd7558e848.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u539f\u6587\u5730\u5740\uff1ahttps:\/\/sebastianraschka.com\/blog\/2025\/understanding-reasoning-llms.html<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>2024 \u5e74\uff0cLLM \u9886\u57df\u7684\u4e13\u4e1a\u5316\u7a0b\u5ea6\u4e0d\u65ad\u63d0\u9ad8\u3002\u9664\u4e86\u9884\u8bad\u7ec3\u548c\u5fae\u8c03\u4e4b\u5916\uff0c\u6211\u4eec\u8fd8\u89c1\u8bc1\u4e86\u4ece RAG \u5230\u4ee3\u7801\u52a9\u624b\u7b49\u4e13\u4e1a\u5e94\u7528\u7684\u5174\u8d77\u3002\u6211\u9884\u8ba1\u8fd9\u4e00\u8d8b\u52bf\u5c06\u5728 2025 \u5e74\u52a0\u901f\uff0c\u4e5f\u5c31\u662f\u66f4\u52a0\u91cd\u89c6\u7279\u5b9a\u4e8e\u5177\u4f53\u9886\u57df\u548c\u5e94\u7528\u7684\u4f18\u5316\uff08\u5373\u4e13\u4e1a\u5316 \/specialization\uff09\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-9adb959b77715e3b9741bd9402c11159.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-9adb959b77715e3b9741bd9402c11159.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u7b2c 1-3 \u9636\u6bb5\u662f\u5f00\u53d1 LLM \u7684\u5e38\u89c1\u6b65\u9aa4\u3002\u7b2c 4 \u9636\u6bb5\u662f\u9488\u5bf9\u5177\u4f53\u7528\u4f8b\u5bf9 LLM \u8fdb\u884c\u4e13\u95e8\u5316\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5f00\u53d1\u63a8\u7406\u6a21\u578b\u5c31\u662f\u4e00\u79cd\u4e13\u4e1a\u5316\u3002\u8fd9\u80fd\u8ba9 LLM \u66f4\u64c5\u957f\u5904\u7406\u590d\u6742\u4efb\u52a1 \u2014\u2014 \u89e3\u51b3\u8fd9\u4e9b\u4efb\u52a1\u65f6\u6700\u597d\u4f7f\u7528\u4e2d\u95f4\u6b65\u9aa4\uff0c\u4f8b\u5b50\u5305\u62ec\u89e3\u7b54\u8c1c\u9898\u3001\u9ad8\u7ea7\u6570\u5b66\u548c\u7f16\u7a0b\u96be\u9898\u3002\u4f46\u662f\uff0c\u8fd9\u79cd\u4e13\u4e1a\u5316\u5e76\u4e0d\u80fd\u53d6\u4ee3\u5176\u4ed6 LLM \u5e94\u7528\u3002\u56e0\u4e3a\u5c06 LLM \u8f6c\u6362\u4e3a\u63a8\u7406\u6a21\u578b\u4e5f\u4f1a\u5e26\u6765\u67d0\u4e9b\u7f3a\u70b9\uff0c\u6211\u5c06\u5728\u540e\u9762\u8ba8\u8bba\u3002<\/section>\n<section><\/section>\n<section>\u4ee5\u4e0b\u662f\u672c\u6587\u7684\u7b80\u8981\u7eb2\u76ee\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u5982\u4f55\u5b9a\u4e49\u300c\u63a8\u7406\u6a21\u578b\u300d\uff1f<\/section>\n<\/li>\n<li>\n<section>\u5e94\u8be5\u4f55\u65f6\u4f7f\u7528\u63a8\u7406\u6a21\u578b\uff1f<\/section>\n<\/li>\n<li>\n<section>\u7b80\u8981\u4e86\u89e3 DeepSeek \u8bad\u7ec3\u6d41\u7a0b<\/section>\n<\/li>\n<li>\n<section>\u6784\u5efa\u548c\u6539\u8fdb\u63a8\u7406\u6a21\u578b\u7684 4 \u79cd\u4e3b\u8981\u65b9\u6cd5<\/section>\n<\/li>\n<\/ul>\n<section>1) \u63a8\u7406\u65f6\u95f4\u6269\u5c55<\/section>\n<section>2) \u7eaf\u5f3a\u5316\u5b66\u4e60 (RL)<\/section>\n<section>3) \u76d1\u7763\u5fae\u8c03\u52a0\u5f3a\u5316\u5b66\u4e60 (SFT + RL)<\/section>\n<section>4) \u7eaf\u76d1\u7763\u5fae\u8c03 (SFT) \u548c\u84b8\u998f<\/section>\n<section>\u7ed3\u8bba<\/section>\n<ul>\n<li>\n<section>\u5bf9 DeepSeek R1 \u7684\u4e00\u4e9b\u601d\u8003<\/section>\n<\/li>\n<li>\n<section>\u5728\u6709\u9650\u7684\u9884\u7b97\u4e0b\u5f00\u53d1\u63a8\u7406\u6a21\u578b<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u5982\u4f55\u5b9a\u4e49\u300c\u63a8\u7406\u6a21\u578b\u300d\uff1f<\/strong><\/section>\n<section><\/section>\n<section>\u5982\u679c\u4f60\u5728 AI\uff08\u6216\u5e7f\u4e49\u4e0a\u7684\u673a\u5668\u5b66\u4e60\uff09\u9886\u57df\u5185\u5de5\u4f5c\uff0c\u4f60\u53ef\u80fd\u5df2\u7ecf\u5f88\u4e86\u89e3\u76f8\u5173\u5b9a\u4e49\u7684\u6a21\u7cca\u6027\u4e86\u3002\u300c<strong>\u63a8\u7406\u6a21\u578b\uff08reasoning model\uff09<\/strong>\u300d\u8fd9\u4e2a\u672f\u8bed\u4e5f\u4e0d\u4f8b\u5916\u3002\u6700\u7ec8\uff0c\u6709\u4eba\u4f1a\u5728\u4e00\u7bc7\u8bba\u6587\u4e2d\u6b63\u5f0f\u5b9a\u4e49\u5b83\uff0c\u4f46\u5374\u53c8\u4f1a\u5728\u4e0b\u4e00\u7bc7\u8bba\u6587\u4e2d\u91cd\u65b0\u5b9a\u4e49\u5b83\uff0c\u5982\u6b64\u8fde\u7ef5\u4e0d\u65ad\u3002<\/section>\n<section><\/section>\n<section>\u672c\u6587\u5c06\u300c\u63a8\u7406\u300d\u5b9a\u4e49\u4e3a\u56de\u7b54\u9700\u8981\u590d\u6742\u3001\u591a\u6b65\u9aa4\u751f\u6210\u548c\u4e2d\u95f4\u6b65\u9aa4\u7684\u95ee\u9898\u7684\u8fc7\u7a0b\u3002\u4f8b\u5982\uff0c\u50cf\u300c\u6cd5\u56fd\u9996\u90fd\u662f\u54ea\u91cc\uff1f\u300d\u8fd9\u6837\u7684\u4e8b\u5b9e\u6027\u95ee\u7b54\u4e0d\u9700\u8981\u63a8\u7406\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u50cf\u300c\u5982\u679c\u4e00\u5217\u706b\u8f66\u4ee5 60 \u82f1\u91cc \/ \u5c0f\u65f6\u7684\u901f\u5ea6\u884c\u9a76 3 \u5c0f\u65f6\uff0c\u5b83\u4f1a\u884c\u9a76\u591a\u8fdc\uff1f\u300d\u8fd9\u6837\u7684\u95ee\u9898\u9700\u8981\u4e00\u4e9b\u7b80\u5355\u7684\u63a8\u7406 \u2014\u2014 \u5b83\u9700\u8981\u8bc6\u522b\u8ddd\u79bb\u3001\u901f\u5ea6\u548c\u65f6\u95f4\u4e4b\u95f4\u7684\u5173\u7cfb\uff0c\u65b9\u80fd\u5f97\u51fa\u7b54\u6848\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-f41aa57cc59f5e9aab3e7c988b5f08f2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-f41aa57cc59f5e9aab3e7c988b5f08f2.png\"><\/a><\/section>\n<section><em><sup>\u5e38\u89c4\u7684 LLM \u53ef\u80fd\u53ea\u63d0\u4f9b\u7b80\u77ed\u7684\u7b54\u6848\uff08\u5982\u5de6\u56fe\u6240\u793a\uff09\uff0c\u800c\u63a8\u7406\u6a21\u578b\u901a\u5e38\u5305\u62ec\u63ed\u793a\u90e8\u5206\u601d\u7ef4\u8fc7\u7a0b\u7684\u4e2d\u95f4\u6b65\u9aa4\u3002\uff08\u8bf7\u6ce8\u610f\uff0c\u8bb8\u591a\u6ca1\u6709\u4e13\u95e8\u4e3a\u63a8\u7406\u4efb\u52a1\u5f00\u53d1\u7684 LLM \u4e5f\u80fd\u5728\u5176\u7b54\u6848\u4e2d\u63d0\u4f9b\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\u3002\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5927\u591a\u6570\u73b0\u4ee3 LLM \u90fd\u5177\u6709\u57fa\u672c\u7684\u63a8\u7406\u80fd\u529b\uff0c\u53ef\u4ee5\u56de\u7b54\u8bf8\u5982\u300c\u5982\u679c\u4e00\u5217\u706b\u8f66\u4ee5 60 \u82f1\u91cc \/ \u5c0f\u65f6\u7684\u901f\u5ea6\u884c\u9a76 3 \u5c0f\u65f6\uff0c\u5b83\u4f1a\u884c\u9a76\u591a\u8fdc\uff1f\u300d\u8fd9\u6837\u7684\u95ee\u9898\u3002\u56e0\u6b64\uff0c\u4eca\u5929\u5f53\u6211\u4eec\u63d0\u5230\u63a8\u7406\u6a21\u578b\u65f6\uff0c\u6211\u4eec\u901a\u5e38\u6307\u7684\u662f\u64c5\u957f\u66f4\u590d\u6742\u63a8\u7406\u4efb\u52a1\u7684 LLM\uff0c\u4f8b\u5982\u89e3\u51b3\u8c1c\u9898\u3001\u8c1c\u8bed\u548c\u6570\u5b66\u8bc1\u660e\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c\u5982\u4eca\u5927\u591a\u6570\u88ab\u79f0\u4e3a\u63a8\u7406\u6a21\u578b\u7684 LLM \u90fd\u5c06\u300c\u601d\u7ef4\u300d\u6216\u300c\u601d\u8003\u300d\u8fc7\u7a0b\u4f5c\u4e3a\u5176\u54cd\u5e94\u7684\u4e00\u90e8\u5206\u3002LLM \u662f\u5426\u4ee5\u53ca\u5982\u4f55\u771f\u6b63\u300c\u601d\u8003\u300d\u5219\u662f\u53e6\u4e00\u4e2a\u8bdd\u9898\u4e86\u3002<\/section>\n<section><\/section>\n<section>\u63a8\u7406\u6a21\u578b\u4e2d\u7684\u4e2d\u95f4\u6b65\u9aa4\u53ef\u4ee5\u4e24\u79cd\u65b9\u5f0f\u51fa\u73b0\u3002\u7b2c\u4e00\u79cd\uff0c\u4e2d\u95f4\u6b65\u9aa4\u53ef\u80fd\u663e\u5f0f\u5730\u51fa\u73b0\u5728\u54cd\u5e94\u4e2d\uff0c\u5982\u4e0a\u56fe\u6240\u793a\u3002\u7b2c\u4e8c\u79cd\uff0c\u4e00\u4e9b\u63a8\u7406 LLM\uff08\u4f8b\u5982 OpenAI o1\uff09\u5219\u4f1a\u8fd0\u884c\u591a\u4e2a\u8fed\u4ee3\uff0c\u4e2d\u95f4\u6b65\u9aa4\u5bf9\u7528\u6237\u6765\u8bf4\u662f\u4e0d\u53ef\u89c1\u7684\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-5d92a42f009254badad34a70311cd68a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-5d92a42f009254badad34a70311cd68a.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u4f7f\u7528\u300c\u63a8\u7406\u300d\u7684\u4e24\u4e2a\u4e0d\u540c\u5c42\u7ea7\uff1a1\uff09\u901a\u8fc7\u591a\u4e2a\u4e2d\u95f4\u6b65\u9aa4\u5904\u7406\u8f93\u5165\u548c\u751f\u6210\u7ed3\u679c\uff0c2\uff09\u5c06\u67d0\u79cd\u5f62\u5f0f\u7684\u63a8\u7406\u653e\u5728\u54cd\u5e94\u4e2d\u63d0\u4f9b\u7ed9\u7528\u6237\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u4ec0\u4e48\u65f6\u5019\u5e94\u8be5\u4f7f\u7528\u63a8\u7406\u6a21\u578b\uff1f<\/strong><\/section>\n<section><\/section>\n<section>\u73b0\u5728\u6211\u4eec\u5df2\u7ecf\u5b9a\u4e49\u4e86\u63a8\u7406\u6a21\u578b\uff0c\u63a5\u4e0b\u6765\u8ba8\u8bba\u66f4\u6709\u8da3\u7684\u90e8\u5206\uff1a\u5982\u4f55\u6784\u5efa\u548c\u6539\u8fdb\u7528\u4e8e\u63a8\u7406\u4efb\u52a1\u7684 LLM\u3002\u7136\u800c\uff0c\u5728\u6df1\u5165\u7814\u7a76\u6280\u672f\u7ec6\u8282\u4e4b\u524d\uff0c\u91cd\u8981\u7684\u662f\u601d\u8003\u4f55\u65f6\u624d\u771f\u6b63\u9700\u8981\u63a8\u7406\u6a21\u578b\u3002<\/section>\n<section><\/section>\n<section>\u6211\u4eec\u4ec0\u4e48\u65f6\u5019\u9700\u8981\u63a8\u7406\u6a21\u578b\uff1f\u63a8\u7406\u6a21\u578b\u64c5\u957f\u89e3\u51b3\u590d\u6742\u4efb\u52a1\uff0c\u4f8b\u5982\u89e3\u51b3\u96be\u9898\u3001\u9ad8\u7ea7\u6570\u5b66\u95ee\u9898\u548c\u9ad8\u96be\u5ea6\u7f16\u7a0b\u4efb\u52a1\u3002\u4f46\u662f\uff0c\u5bf9\u4e8e\u603b\u7ed3\u3001\u7ffb\u8bd1\u6216\u57fa\u4e8e\u77e5\u8bc6\u7684\u95ee\u7b54\u7b49\u7b80\u5355\u4efb\u52a1\uff0c\u5b83\u4eec\u5e76\u4e0d\u662f\u5fc5\u9700\u7684\u3002\u4e8b\u5b9e\u4e0a\uff0c\u5982\u679c\u4f60\u5c06\u63a8\u7406\u6a21\u578b\u7528\u4e8e\u6240\u6709\u4e8b\u60c5\uff0c\u5219\u53ef\u80fd\u906d\u9047\u4f4e\u6548\u7387\u548c\u9ad8\u6210\u672c\u95ee\u9898\u3002\u4f8b\u5982\uff0c\u63a8\u7406\u6a21\u578b\u901a\u5e38\u4f7f\u7528\u8d77\u6765\u66f4\u6602\u8d35\u3001\u66f4\u5197\u957f\uff0c\u6709\u65f6\u7531\u4e8e\u300c\u8fc7\u5ea6\u601d\u8003\u300d\u800c\u66f4\u5bb9\u6613\u51fa\u9519\u3002\u5f88\u5bb9\u6613\u7406\u89e3\uff1a\u4e3a\u4e86\u5b8c\u6210\u4efb\u52a1\uff0c\u9700\u8981\u4f7f\u7528\u6b63\u786e\u7684\u5de5\u5177\uff08\u6216 LLM \u7c7b\u578b\uff09\u3002<\/section>\n<section><\/section>\n<section>\u4e0b\u56fe\u603b\u7ed3\u4e86\u63a8\u7406\u6a21\u578b\u7684\u4e3b\u8981\u4f18\u52bf\u548c\u5c40\u9650\u6027\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-4a173649f6818ef1c4ae1a83d79b85bf.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-4a173649f6818ef1c4ae1a83d79b85bf.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u63a8\u7406\u6a21\u578b\u7684\u4e3b\u8981\u4f18\u52bf\u548c\u52a3\u52bf\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u63a8\u7406\u6a21\u5f0f\u7684\u4f18\u52bf\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u6f14\u7ece\u6216\u5f52\u7eb3\u63a8\u7406\uff08\u4f8b\u5982\u89e3\u5bc6\u3001\u6570\u5b66\u8bc1\u660e\uff09<\/section>\n<\/li>\n<li>\n<section>\u601d\u7ef4\u94fe\u63a8\u7406\uff08\u5206\u89e3\u6210\u591a\u6b65\u9aa4\u95ee\u9898\uff09<\/section>\n<\/li>\n<li>\n<section>\u590d\u6742\u7684\u51b3\u7b56\u4efb\u52a1<\/section>\n<\/li>\n<li>\n<section>\u80fd\u66f4\u597d\u5730\u6cdb\u5316\u5230\u65b0\u95ee\u9898<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u63a8\u7406\u6a21\u5f0f\u7684\u52a3\u52bf\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u5feb\u901f\u4e14\u4f4e\u6210\u672c\u7684\u54cd\u5e94\uff08\u66f4\u591a\u63a8\u7406\u65f6\u95f4\uff09<\/section>\n<\/li>\n<li>\n<section>\u57fa\u4e8e\u77e5\u8bc6\u7684\u4efb\u52a1\uff08\u5e7b\u89c9\uff09<\/section>\n<\/li>\n<li>\n<section>\u7b80\u5355\u4efb\u52a1\uff08\u8fc7\u5ea6\u601d\u8003\uff09<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u7b80\u8981\u4e86\u89e3 DeepSeek \u8bad\u7ec3\u6d41\u7a0b<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u4e0b\u4e00\u8282\u8ba8\u8bba\u6784\u5efa\u548c\u6539\u8fdb\u63a8\u7406\u6a21\u578b\u7684\u56db\u79cd\u4e3b\u8981\u65b9\u6cd5\u4e4b\u524d\uff0c\u8fd9\u91cc\u7b80\u8981\u6982\u8ff0\u4e00\u4e0b DeepSeek R1 \u7684\u5de5\u4f5c\u6d41\u7a0b \u2014\u2014 \u4fe1\u606f\u6765\u6e90\u662f DeepSeek R1 \u6280\u672f\u62a5\u544a\u3002\u8be5\u62a5\u544a\u65e2\u662f\u4e00\u4e2a\u6709\u8da3\u7684\u6848\u4f8b\u7814\u7a76\uff0c\u4e5f\u53ef\u4f5c\u4e3a\u4e00\u4efd\u5f00\u53d1\u63a8\u7406 LLM \u7684\u84dd\u56fe\u3002<\/section>\n<section><\/section>\n<section>\u8bf7\u6ce8\u610f\uff0cDeepSeek \u5e76\u6ca1\u6709\u53d1\u5e03\u5355\u4e00\u7684 R1 \u63a8\u7406\u6a21\u578b\uff0c\u800c\u662f\u53d1\u5e03\u4e86\u4e09\u4e2a\u4e0d\u540c\u7684\u53d8\u4f53\uff1aDeepSeek-R1-Zero\u3001DeepSeek-R1 \u548c DeepSeek-R1-Distill\u3002<\/section>\n<section><\/section>\n<section>\u6839\u636e\u5176\u6280\u672f\u62a5\u544a\u4e2d\u7684\u63cf\u8ff0\uff0c\u6211\u5728\u4e0b\u56fe\u4e2d\u603b\u7ed3\u4e86\u8fd9\u4e9b\u6a21\u578b\u7684\u5f00\u53d1\u8fc7\u7a0b\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-235236949720e024045aae72fe34da1d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-235236949720e024045aae72fe34da1d.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0DeepSeek R1 \u6280\u672f\u62a5\u544a\u4e2d\u8ba8\u8bba\u7684\u4e09\u79cd\u4e0d\u540c\u63a8\u7406\u6a21\u578b\u7684\u5f00\u53d1\u8fc7\u7a0b\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u7b80\u5355\u8fc7\u4e00\u904d\u4e0a\u56fe\u6240\u793a\u7684\u6d41\u7a0b\u3002\u4e0b\u4e00\u8282\u5c06\u4ecb\u7ecd\u66f4\u591a\u7ec6\u8282 \u2014\u2014 \u5c06\u8ba8\u8bba\u6784\u5efa\u548c\u6539\u8fdb\u63a8\u7406\u6a21\u578b\u7684\u56db\u79cd\u4e3b\u8981\u65b9\u6cd5\u3002<\/section>\n<section><\/section>\n<section>(1)\u00a0<strong>DeepSeek-R1-Zero<\/strong>\uff1a\u8be5\u6a21\u578b\u57fa\u4e8e 2024 \u5e74 12 \u6708\u53d1\u5e03\u7684 671B \u9884\u8bad\u7ec3\u7248 DeepSeek-V3 \u57fa\u7840\u6a21\u578b\u3002\u8be5\u7814\u7a76\u56e2\u961f\u4f7f\u7528\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u5bf9\u5176\u8fdb\u884c\u4e86\u8bad\u7ec3\uff0c\u5e76\u63d0\u4f9b\u4e86\u4e24\u79cd\u7c7b\u578b\u7684\u5956\u52b1\u3002\u8fd9\u79cd\u65b9\u6cd5\u88ab\u79f0\u4e3a\u300c\u51b7\u542f\u52a8\uff08cold start\uff09\u300d\u8bad\u7ec3\uff0c\u56e0\u4e3a\u5b83\u4e0d\u5305\u62ec\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u6b65\u9aa4\uff0c\u800c\u8fd9\u901a\u5e38\u662f\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09\u7684\u4e00\u90e8\u5206\u3002<\/section>\n<section><\/section>\n<section>(2)\u00a0<strong>DeepSeek-R1<\/strong>\uff1a\u8fd9\u662f DeepSeek \u7684\u65d7\u8230\u63a8\u7406\u6a21\u578b\uff0c\u57fa\u4e8e DeepSeek-R1-Zero \u800c\u6784\u5efa\u3002\u8be5\u56e2\u961f\u4f7f\u7528\u4e86\u989d\u5916\u7684 SFT \u9636\u6bb5\u548c\u8fdb\u4e00\u6b65\u7684 RL \u8bad\u7ec3\u5bf9\u5176\u8fdb\u884c\u4e86\u8fdb\u4e00\u6b65\u5fae\u8c03\uff0c\u4ece\u800c\u5728\u300c\u51b7\u542f\u52a8\u300d\u7684 R1-Zero \u6a21\u578b\u57fa\u7840\u4e0a\u5b9e\u73b0\u4e86\u63d0\u5347\u3002<\/section>\n<section><\/section>\n<section>(3)\u00a0<strong>D<\/strong><strong>eepSeek-R1-Distill<\/strong>\uff1aDeepSeek \u56e2\u961f\u5229\u7528\u524d\u9762\u6b65\u9aa4\u751f\u6210\u7684 SFT \u6570\u636e\u5bf9 Qwen \u548c Llama \u6a21\u578b\u8fdb\u884c\u4e86\u5fae\u8c03\uff0c\u4ee5\u589e\u5f3a\u5176\u63a8\u7406\u80fd\u529b\u3002\u867d\u7136\u8fd9\u4e0d\u662f\u4f20\u7edf\u610f\u4e49\u4e0a\u7684\u84b8\u998f\uff0c\u4f46\u8fd9\u4e2a\u8fc7\u7a0b\u4e5f\u786e\u5b9e\u662f\u5728\u8f83\u5927\u7684 DeepSeek-R1 671B \u6a21\u578b\u7684\u8f93\u51fa\u4e0a\u8bad\u7ec3\u8f83\u5c0f\u7684\u6a21\u578b\uff08Llama 8B \u548c 70B\uff0c\u4ee5\u53ca Qwen 1.5B-30B\uff09\u3002<\/section>\n<section><\/section>\n<section><strong>\u6784\u5efa\u548c\u6539\u8fdb\u63a8\u7406\u6a21\u578b\u7684 4 \u79cd\u4e3b\u8981\u65b9\u6cd5<\/strong><\/section>\n<section><\/section>\n<section>\u4e0b\u9762\u5c06\u6982\u8ff0\u5f53\u524d\u7528\u4e8e\u589e\u5f3a LLM \u63a8\u7406\u80fd\u529b\u548c\u6784\u5efa\u4e13\u95e8\u7684\u63a8\u7406\u6a21\u578b\uff08\u5982 DeepSeek-R1\u3001OpenAI \u7684 o1 \u548c o3 \u7b49\uff09\u7684\u5173\u952e\u6280\u672f\u3002<\/section>\n<section><\/section>\n<section>\u6ce8\u610f\uff1aOpenAI \u5e76\u672a\u8be6\u7ec6\u8bf4\u660e o1 \u548c o3 \u7684\u5177\u4f53\u5de5\u4f5c\u539f\u7406\u3002\u7136\u800c\uff0c\u636e\u4f20\u5b83\u4eec\u7ec4\u5408\u5f0f\u5730\u5229\u7528\u4e86\u63a8\u7406\u548c\u8bad\u7ec3\u6280\u672f\u3002<\/section>\n<section><\/section>\n<section><strong>1. \u63a8\u7406\u65f6\u95f4\u6269\u5c55<\/strong><\/section>\n<section><\/section>\n<section>\u63d0\u9ad8 LLM \u63a8\u7406\u80fd\u529b\uff08\u6216\u4efb\u4f55\u80fd\u529b\uff09\u7684\u4e00\u79cd\u65b9\u6cd5\u662f\u63a8\u7406\u65f6\u95f4\u6269\u5c55\uff08inference-time scaling\uff09\u3002\u8fd9\u4e2a\u672f\u8bed\u53ef\u4ee5\u6709\u591a\u79cd\u542b\u4e49\uff0c\u4f46\u5728\u8fd9\u91cc\uff0c\u5b83\u6307\u7684\u662f\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u589e\u52a0\u8ba1\u7b97\u8d44\u6e90\u4ee5\u63d0\u9ad8\u8f93\u51fa\u8d28\u91cf\u3002<\/section>\n<section><\/section>\n<section>\u505a\u4e2a\u7c97\u7565\u7684\u7c7b\u6bd4\uff1a\u4eba\u7c7b\u5728\u89e3\u51b3\u590d\u6742\u95ee\u9898\u65f6\uff0c\u5982\u679c\u7ed9\u4ed6\u66f4\u591a\u65f6\u95f4\uff0c\u5f97\u5230\u7684\u7b54\u6848\u5f80\u5f80\u4f1a\u66f4\u597d\u3002\u7c7b\u4f3c\u5730\uff0c\u6211\u4eec\u53ef\u4ee5\u5e94\u7528\u4e00\u4e9b\u6280\u672f\u6765\u9f13\u52b1 LLM \u5728\u751f\u6210\u7b54\u6848\u65f6\u66f4\u591a\u5730\u300c\u601d\u8003\u300d\u3002\uff08\u4e0d\u8fc7 LLM \u7a76\u7adf\u662f\u5426\u771f\u7684\u4f1a\u300c\u601d\u8003\u300d\u8fd8\u6709\u5f85\u8ba8\u8bba\u3002\uff09<\/section>\n<section><\/section>\n<section>\u63a8\u7406\u65f6\u95f4\u6269\u5c55\u7684\u4e00\u79cd\u76f4\u63a5\u7b80\u5355\u7684\u65b9\u6cd5\u662f\u5de7\u5999\u7684\u63d0\u793a\u8bcd\u5de5\u7a0b\u3002\u4e00\u4e2a\u5178\u578b\u7684\u4f8b\u5b50\u662f\u601d\u7ef4\u94fe\uff08CoT\uff09\u63d0\u793a\u65b9\u6cd5\uff0c\u5373\u5728\u8f93\u5165\u7684\u63d0\u793a\u8bcd\u4e2d\u5305\u542b\u300c<strong>think step by step<\/strong>\u300d\u7b49\u77ed\u8bed\u3002\u8fd9\u80fd\u9f13\u52b1\u6a21\u578b\u751f\u6210\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\uff0c\u800c\u4e0d\u662f\u76f4\u63a5\u8df3\u5230\u6700\u7ec8\u7b54\u6848\uff0c\u8fd9\u901a\u5e38\uff08\u4f46\u5e76\u4e0d\u4e00\u5b9a\uff09\u53ef\u4ee5\u5728\u66f4\u590d\u6742\u7684\u95ee\u9898\u4e0a\u5f97\u5230\u66f4\u51c6\u786e\u7684\u7ed3\u679c\u3002\uff08\u8bf7\u6ce8\u610f\uff0c\u5bf9\u4e8e\u66f4\u7b80\u5355\u7684\u57fa\u4e8e\u77e5\u8bc6\u7684\u95ee\u9898\uff0c\u4f8b\u5982\u300c\u6cd5\u56fd\u7684\u9996\u90fd\u662f\u4ec0\u4e48\u300d\uff0c\u91c7\u7528\u8fd9\u79cd\u7b56\u7565\u662f\u6ca1\u6709\u610f\u4e49\u7684\uff0c\u8fd9\u53c8\u662f\u4e00\u4e2a\u5f88\u597d\u7684\u7ecf\u9a8c\u6cd5\u5219\uff0c\u53ef\u4ee5\u627e\u51fa\u63a8\u7406\u6a21\u578b\u662f\u5426\u5bf9\u7ed9\u5b9a\u7684\u8f93\u5165\u67e5\u8be2\u6709\u7528\u3002\uff09<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-0624a54d31af901da5f76f2231e2796c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-0624a54d31af901da5f76f2231e2796c.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u4e00\u4e2a\u7ecf\u5178 CoT \u63d0\u793a\u793a\u4f8b\uff0c\u6765\u81ea\u8bba\u6587\u300aLarge Language Models are Zero-Shot Reasoners\u300b<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u4e0a\u8ff0 CoT \u65b9\u6cd5\u53ef\u88ab\u89c6\u4e3a\u4e00\u79cd\u63a8\u7406\u65f6\u95f4\u6269\u5c55\uff0c\u56e0\u4e3a\u5b83\u4f1a\u751f\u6210\u66f4\u591a\u8f93\u51fa token\uff0c\u4f7f\u63a8\u7406\u7684\u6210\u672c\u4e5f\u4f1a\u66f4\u9ad8\u3002<\/section>\n<section><\/section>\n<section>\u63a8\u7406\u65f6\u95f4\u6269\u5c55\u7684\u53e6\u4e00\u79cd\u65b9\u6cd5\u662f\u4f7f\u7528\u6295\u7968\u548c\u641c\u7d22\u7b56\u7565\u3002\u4e00\u4e2a\u7b80\u5355\u7684\u4f8b\u5b50\u662f\u591a\u6570\u6295\u7968 \u2014\u2014 \u8ba9 LLM \u751f\u6210\u591a\u4e2a\u7b54\u6848\uff0c\u7136\u540e\u6211\u4eec\u901a\u8fc7\u591a\u6570\u6295\u7968\u9009\u62e9\u6b63\u786e\u7684\u7b54\u6848\u3002\u540c\u6837\uff0c\u6211\u4eec\u53ef\u4ee5\u4f7f\u7528\u96c6\u675f\u641c\u7d22\u548c\u5176\u4ed6\u641c\u7d22\u7b97\u6cd5\u6765\u751f\u6210\u66f4\u597d\u7684\u54cd\u5e94\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u91cc\u5f3a\u70c8\u63a8\u8350\u4e00\u7bc7\u8bba\u6587\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-d6b558e3f427e68c4bdae700164be97f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-d6b558e3f427e68c4bdae700164be97f.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u8bba\u6587\u6807\u9898\uff1aScaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2408.03314<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-506805b7e02197d1a0e3457aeb47f83b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-506805b7e02197d1a0e3457aeb47f83b.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u4e0d\u540c\u7684\u57fa\u4e8e\u641c\u7d22\u7684\u65b9\u6cd5\u4f9d\u8d56\u4e8e\u57fa\u4e8e\u8fc7\u7a0b\u5956\u52b1\u7684\u6a21\u578b\u6765\u9009\u62e9\u6700\u4f73\u7b54\u6848\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>DeepSeek R1 \u6280\u672f\u62a5\u544a\u6307\u51fa\u5176\u6a21\u578b\u6ca1\u6709\u4f7f\u7528\u63a8\u7406\u65f6\u95f4\u6269\u5c55\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u6280\u672f\u901a\u5e38\u5728 LLM \u4e4b\u4e0a\u7684\u5e94\u7528\u5c42\u5b9e\u73b0\uff0c\u56e0\u6b64 DeepSeek \u6709\u53ef\u80fd\u4f1a\u5728\u5176\u5e94\u7528\u4e2d\u4f7f\u7528\u5b83\u3002<\/section>\n<section><\/section>\n<section>\u6211\u731c\u60f3 OpenAI \u7684 o1 \u548c o3 \u6a21\u578b\u4f7f\u7528\u4e86\u63a8\u7406\u65f6\u95f4\u6269\u5c55\uff0c\u8fd9\u53ef\u4ee5\u89e3\u91ca\u4e3a\u4ec0\u4e48\u5b83\u4eec\u4e0e GPT-4o \u7b49\u6a21\u578b\u76f8\u6bd4\u76f8\u5bf9\u6602\u8d35\u3002\u9664\u4e86\u63a8\u7406\u65f6\u95f4\u6269\u5c55\u4e4b\u5916\uff0co1 \u548c o3 \u53ef\u80fd\u4f7f\u7528\u7c7b\u4f3c\u4e8e DeepSeek R1 \u4f7f\u7528\u7684 RL \u6d41\u7a0b\u8fdb\u884c\u8bad\u7ec3\u3002\u4e0b\u9762\u4e24\u8282\u5c06\u8be6\u7ec6\u4ecb\u7ecd\u5f3a\u5316\u5b66\u4e60\u3002<\/section>\n<section><\/section>\n<section><strong>2. \u7eaf RL<\/strong><\/section>\n<section><\/section>\n<section>DeepSeek R1 \u8bba\u6587\u7684\u4eae\u70b9\u4e4b\u4e00\u662f\u4ed6\u4eec\u53d1\u73b0\u63a8\u7406\u662f\u7eaf RL \u6d8c\u73b0\u51fa\u7684\u4e00\u79cd\u884c\u4e3a\u3002\u8ba9\u6211\u4eec\u66f4\u8be6\u7ec6\u5730\u63a2\u8ba8\u4e00\u4e0b\u8fd9\u610f\u5473\u7740\u4ec0\u4e48\u3002<\/section>\n<section><\/section>\n<section>\u5982\u524d\u6240\u8ff0\uff0cDeepSeek \u5f00\u53d1\u4e86\u4e09\u79cd\u7c7b\u578b\u7684 R1 \u6a21\u578b\u3002\u7b2c\u4e00\u4e2a DeepSeek-R1-Zero \u5efa\u7acb\u5728 DeepSeek-V3 \u57fa\u7840\u6a21\u578b\u4e4b\u4e0a\uff0c\u8fd9\u662f\u4ed6\u4eec\u4e8e 2024 \u5e74 12 \u6708\u53d1\u5e03\u7684\u6807\u51c6\u9884\u8bad\u7ec3 LLM\u3002\u4e0e\u5178\u578b\u7684 RL \u6d41\u7a0b\u4e0d\u540c\uff08\u5373\u5728 RL \u4e4b\u524d\u5e94\u7528\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\uff09\uff0cDeepSeek-R1-Zero \u4ec5\u4f7f\u7528 RL \u8fdb\u884c\u8bad\u7ec3\uff0c\u6ca1\u6709\u521d\u59cb SFT \u9636\u6bb5\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c1c225d27294863b84002322826393b1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-c1c225d27294863b84002322826393b1.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 DeepSeek-R1-Zero \u6a21\u578b\u7684\u5f00\u53d1\u8fc7\u7a0b\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5c3d\u7ba1\u5982\u6b64\uff0c\u8fd9\u79cd RL \u8fc7\u7a0b\u7c7b\u4f3c\u4e8e\u5e38\u7528\u7684 RLHF \u65b9\u6cd5\uff0c\u540e\u8005\u901a\u5e38\u5e94\u7528\u4e8e\u504f\u597d\u5fae\u8c03 LLM\u3002\u66f4\u591a\u8be6\u60c5\u53ef\u53c2\u9605\u300a<a data-itemshowtype=\"0\" data-linktype=\"2\" href=\"https:\/\/mp.weixin.qq.com\/s?__biz=MzA3MzI4MjgzMw==&amp;mid=2650892359&amp;idx=3&amp;sn=b67d23ce2c2453d730ae7d926471d814&amp;scene=21#wechat_redirect\" target=\"_blank\">LLM \u6210\u529f\u4e0d\u53ef\u6216\u7f3a\u7684\u57fa\u77f3\uff1aRLHF \u53ca\u5176\u66ff\u4ee3\u6280\u672f<\/a>\u300b\u3002\u4f46\u662f\uff0c\u5982\u4e0a\u6240\u8ff0\uff0cDeepSeek-R1-Zero \u7684\u4e3b\u8981\u533a\u522b\u5728\u4e8e\u4ed6\u4eec\u8df3\u8fc7\u4e86\u7528\u4e8e\u6307\u4ee4\u8c03\u6574\u7684\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u9636\u6bb5\u3002\u8fd9\u5c31\u662f\u4ed6\u4eec\u5c06\u5176\u79f0\u4e3a\u300c\u7eaf\u300dRL \u7684\u539f\u56e0\u3002\uff08\u4e0d\u8fc7\u9700\u8981\u6307\u51fa\uff0cLLM \u80cc\u666f\u4e0b\u7684 RL \u4e0e\u4f20\u7edf RL \u6709\u5f88\u5927\u4e0d\u540c\u3002\uff09<\/section>\n<section><\/section>\n<section>\u5bf9\u4e8e\u5956\u52b1\uff0c\u4ed6\u4eec\u6ca1\u6709\u4f7f\u7528\u6839\u636e\u4eba\u7c7b\u504f\u597d\u8bad\u7ec3\u7684\u5956\u52b1\u6a21\u578b\uff0c\u800c\u662f\u91c7\u7528\u4e86\u4e24\u79cd\u7c7b\u578b\u7684\u5956\u52b1\uff1a\u51c6\u786e\u5ea6\u5956\u52b1\u548c\u683c\u5f0f\u5956\u52b1\u3002<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u51c6\u786e\u5ea6\u5956\u52b1\u4f7f\u7528 LeetCode \u7f16\u8bd1\u5668\u6765\u9a8c\u8bc1\u7f16\u7801\u7b54\u6848\uff0c\u5e76\u4f7f\u7528\u786e\u5b9a\u6027\u7cfb\u7edf\u6765\u8bc4\u4f30\u6570\u5b66\u54cd\u5e94\u3002<\/section>\n<\/li>\n<li>\n<section>\u683c\u5f0f\u5956\u52b1\u4f9d\u9760 LLM \u8bc4\u5224\u5458\u6765\u786e\u4fdd\u54cd\u5e94\u9075\u5faa\u9884\u671f\u683c\u5f0f\uff0c\u4f8b\u5982\u5c06\u63a8\u7406\u6b65\u9aa4\u653e\u5728\u6807\u7b7e\u5185\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u4ee4\u4eba\u60ca\u8bb6\u7684\u662f\uff0c\u8fd9\u79cd\u65b9\u6cd5\u8db3\u4ee5\u8ba9 LLM \u53d1\u5c55\u51fa\u57fa\u672c\u7684\u63a8\u7406\u6280\u80fd\u3002\u7814\u7a76\u4eba\u5458\u89c2\u5bdf\u5230\u4e86\u4e00\u4e2a\u300c<strong>\u554a\u54c8\uff01<\/strong>\u300d\u65f6\u523b\uff0c\u5c3d\u7ba1\u6ca1\u6709\u7ecf\u8fc7\u660e\u786e\u7684\u8bad\u7ec3\uff0c\u4f46\u6a21\u578b\u5f00\u59cb\u5728\u5176\u54cd\u5e94\u4e2d\u751f\u6210\u63a8\u7406\u75d5\u8ff9\uff0c\u5982\u4e0b\u56fe\u6240\u793a\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-1910a159e4ebd78614e8c64297c5bd74.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-1910a159e4ebd78614e8c64297c5bd74.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0DeepSeek R1 \u6280\u672f\u62a5\u544a\u5c55\u793a\u7684\u300c\u554a\u54c8\u300d\u65f6\u523b\uff0chttps:\/\/arxiv.org\/abs\/2501.12948<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u867d\u7136 R1-Zero \u5e76\u4e0d\u662f\u8868\u73b0\u6700\u597d\u7684\u63a8\u7406\u6a21\u578b\uff0c\u4f46\u5b83\u786e\u5b9e\u901a\u8fc7\u751f\u6210\u4e2d\u95f4\u7684\u300c\u601d\u8003\u300d\u6b65\u9aa4\u5c55\u793a\u4e86\u63a8\u7406\u80fd\u529b\uff0c\u5982\u4e0a\u56fe\u6240\u793a\u3002\u8fd9\u8bc1\u5b9e\u4e86\u4f7f\u7528\u7eaf RL \u5f00\u53d1\u63a8\u7406\u6a21\u578b\u662f\u53ef\u80fd\u7684\uff0cDeepSeek \u56e2\u961f\u662f\u7b2c\u4e00\u4e2a\u5c55\u793a\uff08\u6216\u81f3\u5c11\u53d1\u5e03\uff09\u8fd9\u79cd\u65b9\u6cd5\u7684\u56e2\u961f\u3002<\/section>\n<section><\/section>\n<section><strong>3. \u76d1\u7763\u5fae\u8c03\u52a0\u5f3a\u5316\u5b66\u4e60\uff08SFT + RL\uff09<\/strong><\/section>\n<section><\/section>\n<section>\u63a5\u4e0b\u6765\uff0c\u8ba9\u6211\u4eec\u770b\u770b DeepSeek \u7684\u65d7\u8230\u63a8\u7406\u6a21\u578b DeepSeek-R1 \u7684\u5f00\u53d1\u8fc7\u7a0b\uff0c\u8fd9\u53ef\u4f5c\u4e3a\u6784\u5efa\u63a8\u7406\u6a21\u578b\u7684\u84dd\u56fe\u3002\u5176\u65b9\u6cd5\u662f\u901a\u8fc7\u7ed3\u5408\u989d\u5916\u7684 SFT \u548c RL \u6765\u6539\u8fdb DeepSeek-R1-Zero\uff0c\u4ee5\u63d0\u9ad8\u5176\u63a8\u7406\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section>\u8bf7\u6ce8\u610f\uff0c\u5728 RL \u4e4b\u524d\u5305\u542b SFT \u9636\u6bb5\u5b9e\u9645\u4e0a\u5f88\u5e38\u89c1\uff0c\u5982\u6807\u51c6 RLHF \u6d41\u7a0b\u90a3\u6837\u3002OpenAI o1 \u5f88\u53ef\u80fd\u5c31\u4f7f\u7528\u4e86\u7c7b\u4f3c\u7684\u65b9\u6cd5\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-8c348ec24c8475d4953b839ee226973b.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-8c348ec24c8475d4953b839ee226973b.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0DeepSeek-R1 \u6a21\u578b\u7684\u5f00\u53d1\u8fc7\u7a0b\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5982\u4e0a\u56fe\u6240\u793a\uff0cDeepSeek \u56e2\u961f\u4f7f\u7528 DeepSeek-R1-Zero \u751f\u6210\u4e86\u6240\u8c13\u7684\u300c\u51b7\u542f\u52a8\u300dSFT \u6570\u636e\u3002\u672f\u8bed\u300c\u51b7\u542f\u52a8\u300d\u6307\u7684\u662f\u8fd9\u4e9b\u6570\u636e\u662f\u7531 DeepSeek-R1-Zero \u751f\u6210\u7684\uff0c\u800c DeepSeek-R1-Zero \u672c\u8eab\u5e76\u672a\u63a5\u53d7\u8fc7\u4efb\u4f55 SFT \u6570\u636e\u7684\u8bad\u7ec3\u3002<\/section>\n<section><\/section>\n<section>DeepSeek \u968f\u540e\u4f7f\u7528\u6b64\u51b7\u542f\u52a8 SFT \u6570\u636e\u901a\u8fc7\u6307\u4ee4\u5fae\u8c03\u8bad\u7ec3\u6a21\u578b\uff0c\u7136\u540e\u8fdb\u884c\u53e6\u4e00\u4e2a RL \u9636\u6bb5\u3002\u6b64 RL \u9636\u6bb5\u4fdd\u7559\u4e86 DeepSeek-R1-Zero RL \u8fc7\u7a0b\u4e2d\u4f7f\u7528\u7684\u76f8\u540c\u51c6\u786e\u5ea6\u548c\u683c\u5f0f\u5956\u52b1\u3002\u4e0d\u8fc7\uff0c\u4ed6\u4eec\u8fd8\u6dfb\u52a0\u4e86\u4e00\u81f4\u6027\u5956\u52b1\u4ee5\u9632\u6b62\u8bed\u8a00\u6df7\u5408 \u2014\u2014 \u5f53\u6a21\u578b\u5728\u54cd\u5e94\u4e2d\u5728\u591a\u79cd\u8bed\u8a00\u4e4b\u95f4\u5207\u6362\u65f6\u5c31\u4f1a\u53d1\u751f\u8fd9\u79cd\u60c5\u51b5\u3002<\/section>\n<section><\/section>\n<section>RL \u9636\u6bb5\u4e4b\u540e\u662f\u53e6\u4e00\u8f6e SFT \u6570\u636e\u6536\u96c6\u3002\u5728\u6b64\u9636\u6bb5\uff0c\u4f7f\u7528\u6700\u65b0\u7684\u6a21\u578b\u68c0\u67e5\u70b9\u751f\u6210 600K \u601d\u7ef4\u94fe SFT \u6837\u672c\uff0c\u540c\u65f6\u4f7f\u7528 DeepSeek-V3 \u57fa\u7840\u6a21\u578b\u521b\u5efa\u53e6\u5916 200K \u57fa\u4e8e\u77e5\u8bc6\u7684 SFT \u6837\u672c\u3002<\/section>\n<section><\/section>\n<section>\u7136\u540e\u5c06\u8fd9\u4e9b 600K + 200K SFT \u6837\u672c\u7528\u4e8e\u53e6\u4e00\u8f6e RL\u3002\u5728\u6b64\u9636\u6bb5\uff0c\u4ed6\u4eec\u518d\u6b21\u4f7f\u7528\u57fa\u4e8e\u89c4\u5219\u7684\u65b9\u6cd5\u5bf9\u6570\u5b66\u548c\u7f16\u7a0b\u95ee\u9898\u8fdb\u884c\u51c6\u786e\u5ea6\u5956\u52b1\uff0c\u800c\u5bf9\u5176\u4ed6\u95ee\u9898\u7c7b\u578b\u4f7f\u7528\u4eba\u7c7b\u504f\u597d\u6807\u7b7e\u3002<\/section>\n<section><\/section>\n<section>\u6700\u7ec8\u6a21\u578b DeepSeek-R1 \u7531\u4e8e\u589e\u52a0\u4e86 SFT \u548c RL \u9636\u6bb5\uff0c\u6027\u80fd\u5728 DeepSeek-R1-Zero \u57fa\u7840\u4e0a\u6709\u4e86\u660e\u663e\u63d0\u5347\uff0c\u5982\u4e0b\u8868\u6240\u793a\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-cc077def7e63d35ef7f5f4736862f3c9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-cc077def7e63d35ef7f5f4736862f3c9.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0OpenAI o1 \u548c DeepSeek R1 \u6a21\u578b\u7684\u57fa\u51c6\u8bc4\u6d4b\u7ed3\u679c\u6bd4\u8f83\uff0c\u6765\u81ea DeepSeek-R1 \u6280\u672f\u62a5\u544a\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u7eaf\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u548c\u84b8\u998f<\/strong><\/section>\n<section><\/section>\n<section>\u5230\u76ee\u524d\u4e3a\u6b62\uff0c\u6211\u4eec\u5df2\u7ecf\u4ecb\u7ecd\u4e86\u6784\u5efa\u548c\u6539\u8fdb\u63a8\u7406\u6a21\u578b\u7684\u4e09\u79cd\u4e3b\u8981\u65b9\u6cd5\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section><strong>\u63a8\u7406\u65f6\u95f4\u6269\u5c55<\/strong>\uff0c\u8fd9\u662f\u4e00\u79cd\u65e0\u9700\u8bad\u7ec3\u6216\u4ee5\u5176\u4ed6\u65b9\u5f0f\u4fee\u6539\u5e95\u5c42\u6a21\u578b\u5373\u53ef\u63d0\u9ad8\u63a8\u7406\u80fd\u529b\u7684\u6280\u672f\u3002<\/section>\n<\/li>\n<li>\n<section><strong>\u7eaf RL<\/strong>\uff0c\u5982 DeepSeek-R1-Zero\uff0c\u5b83\u8868\u660e\u63a8\u7406\u53ef\u4ee5\u5728\u6ca1\u6709\u76d1\u7763\u5fae\u8c03\u7684\u60c5\u51b5\u4e0b\u6210\u4e3a\u4e00\u79cd\u5b66\u4e60\u884c\u4e3a\u3002<\/section>\n<\/li>\n<li>\n<section><strong>SFT + RL<\/strong>\uff0c\u8fd9\u5f97\u5230\u4e86 DeepSeek \u7684\u65d7\u8230\u63a8\u7406\u6a21\u578b DeepSeek-R1\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u90a3\u4e48\uff0c\u8fd8\u6709\u4ec0\u4e48\u65b9\u6cd5\uff1f\u6a21\u578b\u300c<strong>\u84b8\u998f<\/strong>\u300d\u3002<\/section>\n<section><\/section>\n<section>\u4ee4\u4eba\u60ca\u8bb6\u7684\u662f\uff0cDeepSeek \u8fd8\u53d1\u5e03\u4e86\u901a\u8fc7\u6240\u8c13\u300c\u84b8\u998f\u300d\u8fc7\u7a0b\u8bad\u7ec3\u7684\u8f83\u5c0f\u6a21\u578b\u3002\u7136\u800c\uff0c\u5728 LLM \u8bed\u5883\u4e2d\uff0c\u84b8\u998f\u5e76\u4e0d\u4e00\u5b9a\u9075\u5faa\u6df1\u5ea6\u5b66\u4e60\u4e2d\u4f7f\u7528\u7684\u7ecf\u5178\u77e5\u8bc6\u84b8\u998f\u65b9\u6cd5\u3002\u4f20\u7edf\u4e0a\uff0c\u5728\u77e5\u8bc6\u84b8\u998f\u4e2d\uff0c\u8f83\u5c0f\u7684\u5b66\u751f\u6a21\u578b\u5728\u8f83\u5927\u7684\u6559\u5e08\u6a21\u578b\u548c\u76ee\u6807\u6570\u636e\u96c6\u7684 logits \u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002<\/section>\n<section><\/section>\n<section>\u76f8\u53cd\uff0c\u8fd9\u91cc\u7684\u84b8\u998f\u662f\u6307\u5728\u8f83\u5927\u7684 LLM \u751f\u6210\u7684 SFT \u6570\u636e\u96c6\u4e0a\u5bf9\u8f83\u5c0f\u7684 LLM\uff08\u4f8b\u5982 Llama 8B \u548c 70B \u4ee5\u53ca Qwen 2.5 \u6a21\u578b\uff080.5B \u5230 32B\uff09\uff09\u8fdb\u884c\u6307\u4ee4\u5fae\u8c03\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u8fd9\u4e9b\u8f83\u5927\u7684 LLM \u662f DeepSeek-V3 \u548c DeepSeek-R1 \u7684\u4e2d\u95f4\u68c0\u67e5\u70b9\u3002\u4e8b\u5b9e\u4e0a\uff0c\u7528\u4e8e\u6b64\u84b8\u998f\u8fc7\u7a0b\u7684 SFT \u6570\u636e\u4e0e\u7528\u4e8e\u8bad\u7ec3 DeepSeek-R1 \u7684\u6570\u636e\u96c6\u76f8\u540c\uff0c\u5982\u4e0a\u4e00\u8282\u6240\u8ff0\u3002<\/section>\n<section><\/section>\n<section>\u4e0b\u56fe\u4e2d\u7a81\u51fa\u5c55\u793a\u4e86\u84b8\u998f\u90e8\u5206\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-eedc1d7afff81ce6cf32c26e53172897.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-eedc1d7afff81ce6cf32c26e53172897.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u84b8\u998f\u7248 DeepSeek R1 \u6a21\u578b\u7684\u5f00\u53d1\u8fc7\u7a0b\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u4ed6\u4eec\u4e3a\u4ec0\u4e48\u8981\u5f00\u53d1\u8fd9\u4e9b\u84b8\u998f\u6a21\u578b\uff1f\u6211\u8ba4\u4e3a\u6709\u4e24\u4e2a\u4e3b\u8981\u539f\u56e0\uff1a<\/section>\n<section><\/section>\n<ol>\n<li>\n<section>\u8f83\u5c0f\u7684\u6a21\u578b\u6548\u7387\u66f4\u9ad8\u3002\u8fd9\u610f\u5473\u7740\u5b83\u4eec\u8fd0\u884c\u8d77\u6765\u66f4\u4fbf\u5b9c\uff0c\u4f46\u5b83\u4eec\u4e5f\u53ef\u4ee5\u5728\u4f4e\u7aef\u786c\u4ef6\u4e0a\u8fd0\u884c\uff0c\u8fd9\u5bf9\u8bb8\u591a\u50cf\u6211\u4e00\u6837\u7684\u7814\u7a76\u4eba\u5458\u548c\u4fee\u8865\u5320\u6765\u8bf4\u5c24\u5176\u6709\u8da3\u3002<\/section>\n<\/li>\n<li>\n<section>\u7eaf SFT \u6848\u4f8b\u7814\u7a76\u3002\u8fd9\u4e9b\u84b8\u998f\u6a21\u578b\u53ef\u4f5c\u4e3a\u6709\u8da3\u7684\u57fa\u51c6\uff0c\u5c55\u793a\u7eaf\u76d1\u7763\u5fae\u8c03 (SFT) \u5728\u6ca1\u6709 RL \u7684\u60c5\u51b5\u4e0b\u53ef\u4ee5\u8ba9\u6a21\u578b\u8d70\u591a\u8fdc\u3002<\/section>\n<\/li>\n<\/ol>\n<section><\/section>\n<section>\u4e0b\u8868\u6bd4\u8f83\u4e86\u8fd9\u4e9b\u84b8\u998f\u6a21\u578b\u4e0e\u5176\u4ed6\u6d41\u884c\u6a21\u578b\u4ee5\u53ca DeepSeek-R1-Zero \u548c DeepSeek-R1 \u7684\u6027\u80fd\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-21cc4f4d59f3d049ccf2ee9a354fb491.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-21cc4f4d59f3d049ccf2ee9a354fb491.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u84b8\u998f\u6a21\u578b\u4e0e\u975e\u84b8\u998f\u6a21\u578b\u7684\u57fa\u51c6\u6bd4\u8f83\u3002\u6ce8\u91ca\u56fe\u6765\u81ea DeepSeek-R1 \u6280\u672f\u62a5\u544a\u3002 <\/sup><\/em><\/section>\n<section><\/section>\n<section>\u6211\u4eec\u53ef\u4ee5\u770b\u5230\uff0c\u84b8\u998f\u540e\u7684\u6a21\u578b\u660e\u663e\u5f31\u4e8e DeepSeek-R1\uff0c\u4f46\u4e0e DeepSeek-R1-Zero \u76f8\u6bd4\uff0c\u5b83\u4eec\u5374\u51fa\u5947\u5730\u5f3a\u5927\uff0c\u5c3d\u7ba1\u89c4\u6a21\u5c0f\u4e86\u51e0\u4e2a\u6570\u91cf\u7ea7\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u4e9b\u6a21\u578b\u4e0e o1 mini \u76f8\u6bd4\uff0c\u8868\u73b0\u975e\u5e38\u597d\uff08\u6211\u6000\u7591 o1-mini \u672c\u8eab\u53ef\u80fd\u662f o1 \u7684\u4e00\u4e2a\u7c7b\u4f3c\u7684\u84b8\u998f\u7248\u672c\uff09\u3002<\/section>\n<section><\/section>\n<section>\u5728\u7ed3\u675f\u672c\u8282\u4e4b\u524d\uff0c\u8fd8\u6709\u4e00\u4e2a\u6709\u8da3\u7684\u6bd4\u8f83\u503c\u5f97\u4e00\u63d0\u3002DeepSeek \u56e2\u961f\u6d4b\u8bd5\u4e86 DeepSeek-R1-Zero \u4e2d\u51fa\u73b0\u7684\u6d8c\u73b0\u63a8\u7406\u884c\u4e3a\u662f\u5426\u4e5f\u4f1a\u51fa\u73b0\u5728\u8f83\u5c0f\u7684\u6a21\u578b\u4e2d\u3002\u4e3a\u4e86\u7814\u7a76\u8fd9\u4e00\u70b9\uff0c\u4ed6\u4eec\u5c06 DeepSeek-R1-Zero \u4e2d\u76f8\u540c\u7684\u7eaf RL \u65b9\u6cd5\u76f4\u63a5\u5e94\u7528\u4e8e Qwen-32B\u3002<\/section>\n<section><\/section>\n<section>\u4e0b\u8868\u5c55\u793a\u4e86\u5b9e\u9a8c\u7684\u7ed3\u679c\uff0c\u5176\u4e2d QwQ-32B-Preview \u662f\u5343\u95ee\u56e2\u961f\u57fa\u4e8e Qwen 2.5 32B \u5f00\u53d1\u7684\u53c2\u8003\u63a8\u7406\u6a21\u578b\uff08\u6211\u8ba4\u4e3a\u8bad\u7ec3\u7ec6\u8282\u4ece\u672a\u62ab\u9732\u8fc7\uff09\u3002\u6b64\u6bd4\u8f83\u63d0\u4f9b\u4e86\u4e00\u4e9b\u989d\u5916\u7684\u6d1e\u5bdf\uff0c\u5373\u7eaf RL \u662f\u5426\u53ef\u4ee5\u5728\u6bd4 DeepSeek-R1-Zero \u5c0f\u5f97\u591a\u7684\u6a21\u578b\u4e2d\u5f15\u5bfc\u63a8\u7406\u80fd\u529b\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-4be7b9f5211cb958337a35d5e68a1193.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-4be7b9f5211cb958337a35d5e68a1193.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u5728\u8f83\u5c0f\u7684 32B \u6a21\u578b\u4e0a\u5bf9\u84b8\u998f\u548c RL \u8fdb\u884c\u57fa\u51c6\u6bd4\u8f83\u3002\u6ce8\u91ca\u56fe\u6765\u81ea DeepSeek-R1 \u6280\u672f\u62a5\u544a\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u6709\u8da3\u7684\u662f\uff0c\u7ed3\u679c\u8868\u660e\uff0c\u5bf9\u4e8e\u8f83\u5c0f\u7684\u6a21\u578b\uff0c\u84b8\u998f\u6bd4\u7eaf RL \u66f4\u6709\u6548\u3002\u8fd9\u4e0e\u4ee5\u4e0b\u89c2\u70b9\u4e00\u81f4\uff1a\u5355\u9760 RL \u53ef\u80fd\u4e0d\u8db3\u4ee5\u5728\u8fd9\u79cd\u89c4\u6a21\u7684\u6a21\u578b\u4e2d\u4ea7\u751f\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\uff0c\u800c\u4f7f\u7528\u9ad8\u8d28\u91cf\u63a8\u7406\u6570\u636e\u8fdb\u884c SFT \u5728\u4f7f\u7528\u5c0f\u6a21\u578b\u65f6\u53ef\u80fd\u662f\u4e00\u79cd\u66f4\u6709\u6548\u7684\u7b56\u7565\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u5b8c\u6574\u6027\uff0c\u67e5\u770b\u8868\u683c\u4e2d\u7684\u5176\u4ed6\u6bd4\u8f83\u5c06\u4f1a\u5f88\u6709\u7528\uff1a<\/section>\n<section><\/section>\n<ol>\n<li>\n<section>Qwen-32B \u4f7f\u7528 SFT + RL \u8fdb\u884c\u8bad\u7ec3\uff0c\u7c7b\u4f3c\u4e8e DeepSeek-R1 \u7684\u5f00\u53d1\u65b9\u5f0f\u3002\u8fd9\u5c06\u6709\u52a9\u4e8e\u786e\u5b9a\u5f53 RL \u4e0e SFT \u7ed3\u5408\u65f6\uff0c\u4e0e\u7eaf RL \u548c\u7eaf SFT \u76f8\u6bd4\u53ef\u4ee5\u53d6\u5f97\u591a\u5927\u7684\u6539\u8fdb\u3002<\/section>\n<\/li>\n<li>\n<section>DeepSeek-V3 \u4f7f\u7528\u7eaf SFT \u8fdb\u884c\u8bad\u7ec3\uff0c\u4e0e\u521b\u5efa\u84b8\u998f\u6a21\u578b\u7684\u65b9\u5f0f\u7c7b\u4f3c\u3002\u8fd9\u6837\u53ef\u4ee5\u76f4\u63a5\u6bd4\u8f83\uff0c\u770b\u770b RL + SFT \u76f8\u5bf9\u4e8e\u7eaf SFT \u7684\u6548\u679c\u5982\u4f55\u3002<\/section>\n<\/li>\n<\/ol>\n<section><\/section>\n<section><strong>\u603b\u7ed3<\/strong><\/section>\n<section><\/section>\n<section>\u672c\u8282\u63a2\u8ba8\u4e86\u6784\u5efa\u548c\u6539\u8fdb\u63a8\u7406\u6a21\u578b\u7684\u56db\u79cd\u4e0d\u540c\u7b56\u7565\uff1a<\/section>\n<section><\/section>\n<ol>\n<li>\n<section>\u63a8\u7406\u65f6\u95f4\u6269\u5c55\u4e0d\u9700\u8981\u989d\u5916\u7684\u8bad\u7ec3\uff0c\u4f46\u4f1a\u589e\u52a0\u63a8\u7406\u6210\u672c\uff0c\u968f\u7740\u7528\u6237\u6570\u91cf\u6216\u67e5\u8be2\u91cf\u7684\u589e\u52a0\uff0c\u5927\u89c4\u6a21\u90e8\u7f72\u7684\u6210\u672c\u4f1a\u66f4\u9ad8\u3002\u4e0d\u8fc7\uff0c\u5bf9\u4e8e\u5df2\u7ecf\u5f88\u5f3a\u5927\u7684\u6a21\u578b\u6765\u8bf4\uff0c\u63d0\u9ad8\u6027\u80fd\u4ecd\u7136\u662f\u660e\u667a\u4e4b\u4e3e\u3002\u6211\u5f3a\u70c8\u6000\u7591 o1 \u5229\u7528\u4e86\u63a8\u7406\u65f6\u95f4\u6269\u5c55\uff0c\u8fd9\u6709\u52a9\u4e8e\u89e3\u91ca\u4e3a\u4ec0\u4e48\u4e0e DeepSeek-R1 \u76f8\u6bd4\uff0c\u5b83\u5728\u6bcf token \u57fa\u7840\u4e0a\u7684\u6210\u672c\u66f4\u9ad8\u3002<\/section>\n<\/li>\n<li>\n<section>\u7eaf RL \u5bf9\u4e8e\u7814\u7a76\u76ee\u6807\u6765\u8bf4\u5f88\u6709\u8da3\uff0c\u56e0\u4e3a\u5b83\u53ef\u4ee5\u63d0\u4f9b\u63a8\u7406\u4f5c\u4e3a\u4e00\u79cd\u6d8c\u73b0\u884c\u4e3a\u7684\u6d1e\u5bdf\u3002\u7136\u800c\uff0c\u5728\u5b9e\u9645\u7684\u6a21\u578b\u5f00\u53d1\u4e2d\uff0cRL + SFT \u662f\u9996\u9009\u65b9\u6cd5\uff0c\u56e0\u4e3a\u5b83\u53ef\u4ee5\u4ea7\u751f\u66f4\u5f3a\u5927\u7684\u63a8\u7406\u6a21\u578b\u3002\u6211\u5f3a\u70c8\u6000\u7591 o1 \u4e5f\u662f\u4f7f\u7528 RL + SFT \u8fdb\u884c\u8bad\u7ec3\u7684\u3002\u66f4\u51c6\u786e\u5730\u8bf4\uff0c\u6211\u76f8\u4fe1 o1 \u4ece\u6bd4 DeepSeek-R1 \u66f4\u5f31\u3001\u66f4\u5c0f\u7684\u57fa\u7840\u6a21\u578b\u5f00\u59cb\uff0c\u4f46\u901a\u8fc7 RL + SFT \u548c\u63a8\u7406\u65f6\u95f4\u6269\u5c55\u8fdb\u884c\u4e86\u8865\u507f\u3002<\/section>\n<\/li>\n<li>\n<section>\u5982\u4e0a\u6240\u8ff0\uff0cRL + SFT \u662f\u6784\u5efa\u9ad8\u6027\u80fd\u63a8\u7406\u6a21\u578b\u7684\u5173\u952e\u65b9\u6cd5\u3002DeepSeek-R1 \u662f\u4e00\u4e2a\u5f88\u597d\u7684\u84dd\u56fe\uff0c\u5c55\u793a\u4e86\u5982\u4f55\u505a\u5230\u8fd9\u4e00\u70b9\u3002<\/section>\n<\/li>\n<li>\n<section>\u84b8\u998f\u662f\u4e00\u79cd\u6709\u5438\u5f15\u529b\u7684\u65b9\u6cd5\uff0c\u5c24\u5176\u662f\u7528\u4e8e\u521b\u5efa\u66f4\u5c0f\u3001\u66f4\u9ad8\u6548\u7684\u6a21\u578b\u3002\u7136\u800c\uff0c\u84b8\u998f\u7684\u5c40\u9650\u6027\u5728\u4e8e\u5b83\u4e0d\u4f1a\u63a8\u52a8\u521b\u65b0\u6216\u4ea7\u751f\u4e0b\u4e00\u4ee3\u63a8\u7406\u6a21\u578b\u3002\u4f8b\u5982\uff0c\u84b8\u998f\u603b\u662f\u4f9d\u8d56\u4e8e\u73b0\u6709\u7684\u3001\u66f4\u5f3a\u5927\u7684\u6a21\u578b\u6765\u751f\u6210\u76d1\u7763\u5fae\u8c03 (SFT) \u6570\u636e\u3002<\/section>\n<\/li>\n<\/ol>\n<section><\/section>\n<section>\u6211\u9884\u8ba1\u63a5\u4e0b\u6765\u4f1a\u770b\u5230\u7684\u4e00\u4e2a\u6709\u8da3\u7684\u65b9\u9762\u662f\u5c06 RL + SFT\uff08\u65b9\u6cd5 3\uff09\u4e0e\u63a8\u7406\u65f6\u95f4\u6269\u5c55\uff08\u65b9\u6cd5 1\uff09\u76f8\u7ed3\u5408\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u5f88\u53ef\u80fd\u662f OpenAI o1 \u6b63\u5728\u505a\u7684\u4e8b\u60c5\uff0c\u53ea\u4e0d\u8fc7\u5b83\u53ef\u80fd\u57fa\u4e8e\u6bd4 DeepSeek-R1 \u66f4\u5f31\u7684\u57fa\u7840\u6a21\u578b\uff0c\u8fd9\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48 DeepSeek-R1 \u8868\u73b0\u5982\u6b64\u51fa\u8272\uff0c\u540c\u65f6\u5728\u63a8\u7406\u65f6\u95f4\u4e0a\u4fdd\u6301\u76f8\u5bf9\u4f4e\u7684\u6210\u672c\u3002<\/section>\n<section><\/section>\n<section><strong>\u5173\u4e8e DeepSeek R1 \u7684\u601d\u8003<\/strong><\/section>\n<section><\/section>\n<section>\u6700\u8fd1\u51e0\u5468\uff0c\u5f88\u591a\u4eba\u90fd\u95ee\u6211\u5bf9 DeepSeek-R1 \u6a21\u578b\u7684\u770b\u6cd5\u3002<\/section>\n<section><\/section>\n<section>\u7b80\u800c\u8a00\u4e4b\uff0c\u6211\u8ba4\u4e3a\u5b83\u4eec\u662f\u4e00\u9879\u4e86\u4e0d\u8d77\u7684\u6210\u5c31\u3002\u4f5c\u4e3a\u4e00\u540d\u7814\u7a76\u5de5\u7a0b\u5e08\uff0c\u6211\u7279\u522b\u6b23\u8d4f\u8fd9\u4efd\u8be6\u7ec6\u7684\u6280\u672f\u62a5\u544a\uff0c\u5b83\u63d0\u4f9b\u4e86\u6211\u53ef\u4ee5\u4ece\u4e2d\u5b66\u4e60\u7684\u65b9\u6cd5\u8bba\u89c1\u89e3\u3002<\/section>\n<section><\/section>\n<section>\u6700\u4ee4\u4eba\u7740\u8ff7\u7684\u6536\u83b7\u4e4b\u4e00\u662f\uff1a\u63a8\u7406\u53ef\u4ee5\u57fa\u4e8e\u7eaf RL \u6d8c\u73b0\u51fa\u6765\u3002\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u662f\uff0cDeepSeek \u5df2\u6839\u636e\u5bbd\u677e\u7684 MIT \u5f00\u6e90\u8bb8\u53ef\u8bc1\u5bf9\u5176\u6a21\u578b\u8fdb\u884c\u4e86\u5f00\u6e90\uff0c\u8be5\u8bb8\u53ef\u8bc1\u7684\u9650\u5236\u751a\u81f3\u6bd4 Meta \u7684 Llama \u6a21\u578b\u8fd8\u8981\u5c11\u3002<\/section>\n<section><\/section>\n<section><strong>\u4e0e o1 \u76f8\u6bd4\u5982\u4f55\uff1f<\/strong><\/section>\n<section><\/section>\n<section>DeepSeek-R1 \u6bd4 o1 \u597d\u5417\uff1f\u6211\u8ba4\u4e3a\u4e24\u8005\u5927\u81f4\u76f8\u540c\u3002\u7136\u800c\uff0c\u6700\u7a81\u51fa\u7684\u662f DeepSeek-R1 \u5728\u63a8\u7406\u65f6\u95f4\u4e0a\u66f4\u9ad8\u6548\u3002\u8fd9\u8868\u660e DeepSeek \u53ef\u80fd\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6295\u5165\u4e86\u66f4\u591a\uff0c\u800c OpenAI \u53ef\u80fd\u66f4\u591a\u5730\u4f9d\u8d56\u4e8e o1 \u7684\u63a8\u7406\u65f6\u95f4\u6269\u5c55\u3002<\/section>\n<section><\/section>\n<section>\u5c3d\u7ba1\u5982\u6b64\uff0c\u5f88\u96be\u76f4\u63a5\u6bd4\u8f83 o1 \u548c DeepSeek-R1\uff0c\u56e0\u4e3a OpenAI \u5c1a\u672a\u62ab\u9732\u6709\u5173 o1 \u7684\u592a\u591a\u4fe1\u606f\u3002<\/section>\n<section><\/section>\n<section>\u4f8b\u5982\uff0c\u6211\u4eec\u4e0d\u77e5\u9053\u4e00\u4e9b\u4fe1\u606f\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>o1 \u4e5f\u662f MoE \u5417\uff1f<\/section>\n<\/li>\n<li>\n<section>o1 \u6709\u591a\u5927\uff1f<\/section>\n<\/li>\n<li>\n<section>o1 \u53ef\u80fd\u53ea\u662f GPT-4o \u7684\u7a0d\u5fae\u6539\u8fdb\u7248\u672c\uff0c\u5177\u6709\u6700\u5c11\u7684 RL + SFT \u548c\u4ec5\u5e7f\u6cdb\u7684\u63a8\u7406\u65f6\u95f4\u6269\u5c55\u5417\uff1f<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u5982\u679c\u4e0d\u77e5\u9053\u8fd9\u4e9b\u7ec6\u8282\uff0c\u76f4\u63a5\u6bd4\u8f83\u5c31\u662f\u9a74\u5507\u4e0d\u5bf9\u9a6c\u5634\u4e86\u3002<\/section>\n<section><\/section>\n<section><strong>\u8bad\u7ec3 DeepSeek-R1 \u7684\u6210\u672c<\/strong><\/section>\n<section><\/section>\n<section>\u53e6\u4e00\u4e2a\u8ba8\u8bba\u70b9\u662f\u5f00\u53d1 DeepSeek-R1 \u7684\u6210\u672c\u3002\u6709\u4eba\u63d0\u5230\u8bad\u7ec3\u6210\u672c\u7ea6\u4e3a 600 \u4e07\u7f8e\u5143\uff0c\u4f46\u4ed6\u4eec\u53ef\u80fd\u5c06 DeepSeek-V3\uff08\u53bb\u5e74 12 \u6708\u53d1\u5e03\u7684\u57fa\u7840\u6a21\u578b\uff09\u548c DeepSeek-R1 \u6df7\u4e3a\u4e00\u8c08\u3002<\/section>\n<section><\/section>\n<section>600 \u4e07\u7f8e\u5143\u7684\u4f30\u7b97\u662f\u57fa\u4e8e\u6bcf GPU \u5c0f\u65f6 2 \u7f8e\u5143\u7684\u5047\u8bbe\u4ee5\u53ca DeepSeek-V3 \u6700\u7ec8\u8bad\u7ec3\u8fd0\u884c\u6240\u9700\u7684 GPU \u5c0f\u65f6\u6570\uff0c\u8be5\u4f30\u7b97\u6700\u521d\u4e8e 2024 \u5e74 12 \u6708\u8fdb\u884c\u8ba8\u8bba\u3002<\/section>\n<section><\/section>\n<section>\u7136\u800c\uff0cDeepSeek \u56e2\u961f\u4ece\u672a\u900f\u9732 R1 \u7684\u5177\u4f53 GPU \u5c0f\u65f6\u6570\u6216\u5f00\u53d1\u6210\u672c\uff0c\u56e0\u6b64\u4efb\u4f55\u6210\u672c\u4f30\u7b97\u90fd\u8fd8\u53ea\u662f\u7eaf\u7cb9\u7684\u731c\u6d4b\u3002<\/section>\n<section><\/section>\n<section>\u65e0\u8bba\u5982\u4f55\uff0c\u6700\u7ec8 DeepSeek-R1 \u6210\u4e3a\u4e86\u5f00\u653e\u6743\u91cd\u63a8\u7406\u6a21\u578b\u7684\u4e00\u4e2a\u91cd\u8981\u91cc\u7a0b\u7891\uff0c\u5e76\u4e14\u5176\u63a8\u7406\u65f6\u7684\u6548\u7387\u4f7f\u5176\u6210\u4e3a OpenAI o1 \u7684\u4e00\u4e2a\u6709\u8da3\u66ff\u4ee3\u54c1\u3002<\/section>\n<section><\/section>\n<section><strong>\u5728\u6709\u9650\u7684\u9884\u7b97\u4e0b\u5f00\u53d1\u63a8\u7406\u6a21\u578b<\/strong><\/section>\n<section><\/section>\n<section>\u5f00\u53d1 DeepSeek-R1 \u7ea7\u63a8\u7406\u6a21\u578b\u53ef\u80fd\u9700\u8981\u6570\u5341\u4e07\u5230\u6570\u767e\u4e07\u7f8e\u5143\uff0c\u5373\u4f7f\u4ece\u50cf DeepSeek-V3 \u8fd9\u6837\u7684\u5f00\u653e\u6743\u91cd\u57fa\u7840\u6a21\u578b\u5f00\u59cb\u4e5f\u662f\u5982\u6b64\u3002\u5bf9\u4e8e\u9884\u7b97\u6709\u9650\u7684\u7814\u7a76\u4eba\u5458\u6216\u5de5\u7a0b\u5e08\u6765\u8bf4\uff0c\u8fd9\u53ef\u80fd\u4f1a\u4ee4\u4eba\u6cae\u4e27\u3002<\/section>\n<section><\/section>\n<section><strong>\u597d\u6d88\u606f\uff1a\u84b8\u998f\u53ef\u4ee5\u53d1\u6325\u5f88\u5927\u7684\u4f5c\u7528<\/strong><\/section>\n<section><\/section>\n<section>\u5e78\u8fd0\u7684\u662f\uff0c\u6a21\u578b\u84b8\u998f\u63d0\u4f9b\u4e86\u4e00\u79cd\u66f4\u5177\u6210\u672c\u6548\u76ca\u7684\u66ff\u4ee3\u65b9\u6848\u3002DeepSeek \u56e2\u961f\u901a\u8fc7 R1 \u84b8\u998f\u6a21\u578b\u8bc1\u660e\u4e86\u8fd9\u4e00\u70b9\uff0c\u5c3d\u7ba1\u5176\u6bd4 DeepSeek-R1 \u5c0f\u5f97\u591a\uff0c\u4f46\u63a8\u7406\u6027\u80fd\u5374\u51fa\u5947\u5730\u5f3a\u5927\u3002\u7136\u800c\uff0c\u5373\u4f7f\u662f\u8fd9\u79cd\u65b9\u6cd5\u4e5f\u5e76\u4e0d\u5b8c\u5168\u4fbf\u5b9c\u3002\u4ed6\u4eec\u7684\u84b8\u998f\u8fc7\u7a0b\u4f7f\u7528\u4e86 800K SFT \u6837\u672c\uff0c\u8fd9\u9700\u8981\u5927\u91cf\u8ba1\u7b97\u3002<\/section>\n<section><\/section>\n<section>\u6709\u8da3\u7684\u662f\uff0c\u5c31\u5728 DeepSeek-R1 \u53d1\u5e03\u524d\u51e0\u5929\uff0c\u6211\u5076\u7136\u770b\u5230\u4e86\u4e00\u7bc7\u5173\u4e8e Sky-T1 \u7684\u6587\u7ae0\uff0c\u8fd9\u662f\u4e00\u4e2a\u4ee4\u4eba\u7740\u8ff7\u7684\u9879\u76ee\uff0c\u4e00\u4e2a\u5c0f\u56e2\u961f\u4ec5\u4f7f\u7528 17K \u7684 SFT \u6837\u672c\u8bad\u7ec3\u4e86\u4e00\u4e2a\u5f00\u653e\u6743\u91cd 32B \u6a21\u578b\u3002<\/section>\n<section><\/section>\n<section>\u603b\u6210\u672c\u662f\u591a\u5c11\uff1f\u4ec5\u9700 450 \u7f8e\u5143\uff0c\u8fd9\u6bd4\u5927\u591a\u6570 AI \u4f1a\u8bae\u7684\u6ce8\u518c\u8d39\u8fd8\u4f4e\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u4e2a\u4f8b\u5b50\u8868\u660e\uff0c\u5c3d\u7ba1\u5927\u89c4\u6a21\u8bad\u7ec3\u4ecd\u7136\u6602\u8d35\uff0c\u4f46\u8f83\u5c0f\u89c4\u6a21\u3001\u6709\u9488\u5bf9\u6027\u7684\u5fae\u8c03\u5de5\u4f5c\u4ecd\u7136\u53ef\u4ee5\u4ee5\u6781\u4f4e\u7684\u6210\u672c\u4ea7\u751f\u51fa\u8272\u7684\u7ed3\u679c\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-630b661845e89d991e82d72870bc91a1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-630b661845e89d991e82d72870bc91a1.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe\u6e90\uff1a\u300aSky-T1\uff1a\u5728 450 \u7f8e\u5143\u4ee5\u5185\u8bad\u7ec3\u4f60\u81ea\u5df1\u7684 O1 \u9884\u89c8\u6a21\u578b\u300b\u4e00\u6587\u3002https:\/\/novasky-ai.github.io\/posts\/sky-t1\/<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u6839\u636e\u4ed6\u4eec\u7684\u57fa\u51c6\u6d4b\u8bd5\uff0cSky-T1 \u7684\u8868\u73b0\u4e0e OpenAI o1 \u5927\u81f4\u76f8\u5f53\uff0c\u8003\u8651\u5230\u5176\u4f4e\u5ec9\u7684\u8bad\u7ec3\u6210\u672c\uff0c\u8fd9\u4e00\u8868\u73b0\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u3002<\/section>\n<section><\/section>\n<section><strong>\u9884\u7b97\u5185\u7684\u7eaf RL\uff1aTinyZero<\/strong><\/section>\n<section><\/section>\n<section>\u867d\u7136 Sky-T1 \u4e13\u6ce8\u4e8e\u6a21\u578b\u84b8\u998f\uff0c\u4f46\u6211\u4e5f\u5728\u300c\u7eaf RL\u300d\u9886\u57df\u53d1\u73b0\u4e86\u4e00\u4e9b\u6709\u8da3\u7684\u5de5\u4f5c\u3002\u4e00\u4e2a\u503c\u5f97\u6ce8\u610f\u7684\u4f8b\u5b50\u662f TinyZero\uff0c\u8fd9\u662f\u4e00\u4e2a 3B \u53c2\u6570\u6a21\u578b\uff0c\u5b83\u590d\u5236\u4e86 DeepSeek-R1-Zero \u65b9\u6cd5\uff08\u9644\u6ce8\uff1a\u8bad\u7ec3\u6210\u672c\u4e0d\u5230 30 \u7f8e\u5143\uff09\u3002<\/section>\n<section><\/section>\n<section>\u4ee4\u4eba\u60ca\u8bb6\u7684\u662f\uff0c\u5373\u4f7f\u53ea\u6709 3B \u53c2\u6570\uff0cTinyZero \u4e5f\u8868\u73b0\u51fa\u4e00\u4e9b\u6d8c\u73b0\u7684\u81ea\u6211\u9a8c\u8bc1\u80fd\u529b\uff0c\u8fd9\u652f\u6301\u4e86\u63a8\u7406\u53ef\u4ee5\u901a\u8fc7\u7eaf RL \u6d8c\u73b0\u7684\u60f3\u6cd5\uff0c\u5373\u4f7f\u5728\u5c0f\u6a21\u578b\u4e2d\u4e5f\u662f\u5982\u6b64\u3002<\/section>\n<section><\/section>\n<section>TinyZero \u5e93\u63d0\u5230\u7814\u7a76\u62a5\u544a\u4ecd\u5728\u8fdb\u884c\u4e2d\uff0c\u6211\u4e00\u5b9a\u4f1a\u5bc6\u5207\u5173\u6ce8\u66f4\u591a\u7ec6\u8282\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-d89093b7a10eea0ec2debeb56f1daabc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-d89093b7a10eea0ec2debeb56f1daabc.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0TinyZero \u5e93 (https:\/\/github.com\/Jiayi-Pan\/TinyZero) \u4e2d\u7684\u4e00\u5f20\u56fe\u7247\u8868\u660e\u8be5\u6a21\u578b\u80fd\u591f\u8fdb\u884c\u81ea\u6211\u9a8c\u8bc1\u3002\uff08\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u770b\u770b\u57fa\u7840\u6a21\u578b\u7684\u56de\u590d\u4f1a\u5f88\u6709\u8da3\u3002\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u4e0a\u8ff0\u4e24\u4e2a\u9879\u76ee\u8868\u660e\uff0c\u5373\u4f7f\u9884\u7b97\u6709\u9650\uff0c\u4e5f\u53ef\u4ee5\u5728\u63a8\u7406\u6a21\u578b\u4e0a\u5f00\u5c55\u6709\u8da3\u7684\u5de5\u4f5c\u3002\u867d\u7136\u8fd9\u4e24\u79cd\u65b9\u6cd5\u90fd\u590d\u73b0\u4e86 DeepSeek-R1 \u7684\u65b9\u6cd5\uff0c\u4e00\u79cd\u4e13\u6ce8\u4e8e\u7eaf RL\uff08TinyZero\uff09\uff0c\u53e6\u4e00\u79cd\u4e13\u6ce8\u4e8e\u7eaf SFT\uff08Sky-T1\uff09\uff0c\u4f46\u63a2\u7d22\u5982\u4f55\u8fdb\u4e00\u6b65\u6269\u5c55\u8fd9\u4e9b\u60f3\u6cd5\u5c06\u975e\u5e38\u6709\u8da3\u3002<\/section>\n<section><\/section>\n<section><strong>\u8d85\u8d8a\u4f20\u7edf SFT\uff1a\u65c5\u7a0b\u5b66\u4e60<\/strong><\/section>\n<section><\/section>\n<section>\u53bb\u5e74\u6211\u5076\u7136\u53d1\u73b0\u4e00\u79cd\u7279\u522b\u6709\u8da3\u7684\u65b9\u6cd5\uff0c\u8bba\u6587\u300aO1 Replication Journey: A Strategic Progress Report \u2013 Part 1\u300b\u4e2d\u5bf9\u6b64\u8fdb\u884c\u4e86\u63cf\u8ff0\u3002\u5c3d\u7ba1\u6807\u9898\u5982\u6b64\uff0c\u4f46\u8be5\u8bba\u6587\u5b9e\u9645\u4e0a\u5e76\u6ca1\u6709\u590d\u5236 o1\u3002\u76f8\u53cd\uff0c\u5b83\u4ecb\u7ecd\u4e86\u4e00\u79cd\u6539\u8fdb\u84b8\u998f\uff08\u7eaf SFT\uff09\u8fc7\u7a0b\u7684\u53e6\u4e00\u79cd\u65b9\u6cd5\u3002<\/section>\n<section><\/section>\n<section>\u8bba\u6587\u7684\u6838\u5fc3\u601d\u60f3\u662f\u7528\u300c\u65c5\u7a0b\u5b66\u4e60\u300d\u66ff\u4ee3\u300c\u6377\u5f84\u5b66\u4e60\u300d\u3002<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u6377\u5f84\u5b66\u4e60\u662f\u6307\u6307\u4ee4\u5fae\u8c03\u7684\u4f20\u7edf\u65b9\u6cd5\uff0c\u5176\u4e2d\u4ec5\u4f7f\u7528\u6b63\u786e\u7684\u89e3\u51b3\u65b9\u6848\u8def\u5f84\u6765\u8bad\u7ec3\u6a21\u578b\u3002<\/section>\n<\/li>\n<li>\n<section>\u53e6\u4e00\u65b9\u9762\uff0c\u65c5\u7a0b\u5b66\u4e60\u4e5f\u5305\u62ec\u9519\u8bef\u7684\u89e3\u51b3\u8def\u5f84\uff0c\u8ba9\u6a21\u578b\u4ece\u9519\u8bef\u4e2d\u5b66\u4e60\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u8fd9\u79cd\u65b9\u6cd5\u4e0e TinyZero \u7684\u7eaf RL \u8bad\u7ec3\u4e2d\u89c2\u5bdf\u5230\u7684\u81ea\u6211\u9a8c\u8bc1\u80fd\u529b\u6709\u70b9\u76f8\u4f3c\uff0c\u4f46\u5b83\u4e13\u6ce8\u4e8e\u5b8c\u5168\u901a\u8fc7 SFT \u6539\u8fdb\u6a21\u578b\u3002\u901a\u8fc7\u8ba9\u6a21\u578b\u63a5\u89e6\u9519\u8bef\u7684\u63a8\u7406\u8def\u5f84\u53ca\u5176\u4fee\u6b63\uff0c\u65c5\u7a0b\u5b66\u4e60\u8fd8\u53ef\u4ee5\u589e\u5f3a\u81ea\u6211\u4fee\u6b63\u80fd\u529b\uff0c\u4ece\u800c\u53ef\u80fd\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\u4f7f\u63a8\u7406\u6a21\u578b\u66f4\u52a0\u53ef\u9760\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-8e3da4aa5963d317aa564a8a2698958d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-8e3da4aa5963d317aa564a8a2698958d.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u4e0e\u4f20\u7edf\u7684\u6377\u5f84\u5b66\u4e60\u4e0d\u540c\uff0c\u65c5\u7a0b\u5b66\u4e60\u5728 SFT \u6570\u636e\u4e2d\u5305\u542b\u4e86\u9519\u8bef\u7684\u89e3\u51b3\u65b9\u6848\u8def\u5f84\u3002\u6ce8\u91ca\u56fe\u6765\u81ea https:\/\/arxiv.org\/abs\/2410.18982<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u8fd9\u53ef\u80fd\u662f\u672a\u6765\u5de5\u4f5c\u7684\u4e00\u4e2a\u4ee4\u4eba\u5174\u594b\u7684\u65b9\u5411\uff0c\u7279\u522b\u662f\u5bf9\u4e8e\u4f4e\u9884\u7b97\u63a8\u7406\u6a21\u578b\u5f00\u53d1\uff0c\u56e0\u4e3a\u57fa\u4e8e RL \u7684\u65b9\u6cd5\u5728\u8ba1\u7b97\u4e0a\u53ef\u80fd\u4e0d\u5207\u5b9e\u9645\u3002<\/section>\n<section><\/section>\n<section>\u65e0\u8bba\u5982\u4f55\uff0c\u63a8\u7406\u6a21\u578b\u65b9\u9762\u76ee\u524d\u6b63\u5728\u53d1\u751f\u5f88\u591a\u6709\u8da3\u7684\u5de5\u4f5c\uff0c\u6211\u76f8\u4fe1\u6211\u4eec\u5c06\u5728\u63a5\u4e0b\u6765\u7684\u51e0\u4e2a\u6708\u91cc\u770b\u5230\u66f4\u591a\u4ee4\u4eba\u5174\u594b\u7684\u6210\u679c\uff01<\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-02-09-4\" target=\"_blank\">Sebastian Raschka\uff1a\u5173\u4e8eDeepSeek R1\u548c\u63a8\u7406\u6a21\u578b\uff0c\u6211\u6709\u51e0\u70b9\u770b\u6cd5<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:Sebastian R [&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":[71,55,65],"class_list":["post-38180","post","type-post","status-publish","format-standard","hentry","category-news","tag-rag","tag-55","tag-65"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>Sebastian Raschka\uff1a\u5173\u4e8eDeepSeek R1\u548c\u63a8\u7406\u6a21\u578b\uff0c\u6211\u6709\u51e0\u70b9\u770b\u6cd5 - \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=38180\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=38180\",\"url\":\"https:\/\/17aitech.com\/?p=38180\",\"name\":\"Sebastian Raschka\uff1a\u5173\u4e8eDeepSeek R1\u548c\u63a8\u7406\u6a21\u578b\uff0c\u6211\u6709\u51e0\u70b9\u770b\u6cd5 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=38180#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=38180#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-15d4c381c2263f771daf2ecd7558e848.png\",\"datePublished\":\"2025-02-11T06:00:52+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=38180#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=38180\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=38180#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-15d4c381c2263f771daf2ecd7558e848.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-15d4c381c2263f771daf2ecd7558e848.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=38180#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Sebastian Raschka\uff1a\u5173\u4e8eDeepSeek R1\u548c\u63a8\u7406\u6a21\u578b\uff0c\u6211\u6709\u51e0\u70b9\u770b\u6cd5\"}]},{\"@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\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"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":"Sebastian Raschka\uff1a\u5173\u4e8eDeepSeek R1\u548c\u63a8\u7406\u6a21\u578b\uff0c\u6211\u6709\u51e0\u70b9\u770b\u6cd5 - \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=38180","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=38180","url":"https:\/\/17aitech.com\/?p=38180","name":"Sebastian Raschka\uff1a\u5173\u4e8eDeepSeek R1\u548c\u63a8\u7406\u6a21\u578b\uff0c\u6211\u6709\u51e0\u70b9\u770b\u6cd5 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=38180#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=38180#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-15d4c381c2263f771daf2ecd7558e848.png","datePublished":"2025-02-11T06:00:52+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=38180#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=38180"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=38180#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-15d4c381c2263f771daf2ecd7558e848.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/02\/frc-15d4c381c2263f771daf2ecd7558e848.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=38180#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"Sebastian Raschka\uff1a\u5173\u4e8eDeepSeek R1\u548c\u63a8\u7406\u6a21\u578b\uff0c\u6211\u6709\u51e0\u70b9\u770b\u6cd5"}]},{"@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":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","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\/38180","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=38180"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/38180\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=38180"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=38180"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=38180"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}