
{"id":5515,"date":"2025-04-29T02:32:26","date_gmt":"2025-04-28T18:32:26","guid":{"rendered":"https:\/\/infernews.com\/?page_id=5515"},"modified":"2025-04-29T02:34:28","modified_gmt":"2025-04-28T18:34:28","slug":"deepseek-r1-zero-%e7%8d%8e%e5%8b%b5%e5%87%bd%e6%95%b8","status":"publish","type":"page","link":"https:\/\/infernews.com\/blog\/deepseek-r1-zero-%e7%8d%8e%e5%8b%b5%e5%87%bd%e6%95%b8\/","title":{"rendered":"DeepSeek-R1-Zero \u734e\u52f5\u51fd\u6578"},"content":{"rendered":"\n<p><a href=\"https:\/\/huggingface.co\/deepseek-ai\/DeepSeek-R1-Zero\">DeepSeek-R1-Zero<\/a> \u6700\u986f\u8457\u7684\u7279\u9ede\u662f\u5b83\u5b8c\u5168\u50c5\u4f7f\u7528\u5f37\u5316\u5b78\u7fd2 (RL) \u9032\u884c\u8a13\u7df4\uff0c\u800c\u6c92\u6709\u7d93\u904e\u76e3\u7763\u5f0f\u5fae\u8abf (SFT) \u7684\u521d\u6b65\u968e\u6bb5\u3002\u5b83\u901a\u904e\u8a66\u932f\u4f86\u5b78\u7fd2\uff0c\u4e26\u6839\u64da\u5176\u751f\u6210\u7684\u89e3\u6c7a\u65b9\u6848\u6240\u7372\u5f97\u7684\u56de\u994b\u4f86\u8abf\u6574\u81ea\u8eab\u7684\u884c\u70ba\u548c\u53c3\u6578\u3002<\/p>\n\n\n\n<p>\u5728 DeepSeek-R1-Zero \u7528\u65bc\u63a8\u7406\u4efb\u52d9 (\u4f8b\u5982\u6578\u5b78\u554f\u984c\u3001\u7a0b\u5f0f\u7de8\u5beb\u7b49) \u7684\u5f37\u5316\u5b78\u7fd2\u8a13\u7df4\u4e2d\uff0c\u5176 Reward (\u734e\u52f5) \u6a5f\u5236\u4e3b\u8981\u662f\u57fa\u65bc\u5c0d\u6a21\u578b\u8f38\u51fa\u7684<strong>\u81ea\u52d5\u8a55\u4f30<\/strong>\uff0c\u9019\u7a2e\u8a55\u4f30\u901a\u5e38\u662f<strong>\u898f\u5247\u5c0e\u5411 (rule-based)<\/strong> \u7684\u3002<\/p>\n\n\n\n<p>\u5177\u9ad4\u4f86\u8aaa\uff0c\u734e\u52f5\u7684\u8a2d\u8a08\u6703\u76f4\u63a5\u53cd\u6620\u6a21\u578b\u751f\u6210\u7684\u89e3\u6c7a\u65b9\u6848\u7684<strong>\u6b63\u78ba\u6027\u6216\u54c1\u8cea<\/strong>\u3002\u8003\u616e\u5230 DeepSeek-R1-Zero \u5c08\u6ce8\u65bc\u63a8\u7406\u4efb\u52d9\uff0c\u734e\u52f5\u53ef\u80fd\u4f86\u81ea\u65bc\u4ee5\u4e0b\u5e7e\u500b\u65b9\u9762\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u7b54\u6848\u7684\u6b63\u78ba\u6027\uff1a<\/strong> \u5c0d\u65bc\u6709\u660e\u78ba\u6b63\u78ba\u7b54\u6848\u7684\u554f\u984c (\u4f8b\u5982\u6578\u5b78\u984c)\uff0c\u5982\u679c\u6a21\u578b\u8f38\u51fa\u7684\u7b54\u6848\u6b63\u78ba\uff0c\u5b83\u5c07\u7372\u5f97\u6b63\u5411\u734e\u52f5\uff1b\u5982\u679c\u932f\u8aa4\uff0c\u5247\u53ef\u80fd\u7372\u5f97\u8ca0\u5411\u734e\u52f5\u6216\u96f6\u734e\u52f5\u3002<\/li>\n\n\n\n<li><strong>\u89e3\u984c\u6b65\u9a5f\u7684\u6709\u6548\u6027\/\u908f\u8f2f\u6027\uff1a<\/strong> \u5c0d\u65bc\u9700\u8981\u591a\u6b65\u9a5f\u63a8\u7406\u7684\u554f\u984c\uff0c\u734e\u52f5\u53ef\u80fd\u4e0d\u50c5\u50c5\u53d6\u6c7a\u65bc\u6700\u7d42\u7b54\u6848\uff0c\u9084\u6703\u8a55\u4f30\u4e2d\u9593\u63a8\u7406\u6b65\u9a5f\u7684\u908f\u8f2f\u6027\u548c\u6709\u6548\u6027\u3002\u67d0\u4e9b\u81ea\u52d5\u8a55\u4f30\u7cfb\u7d71\u53ef\u4ee5\u6aa2\u67e5\u63a8\u7406\u904e\u7a0b\u662f\u5426\u9075\u5faa\u4e86\u5408\u7406\u7684\u898f\u5247\u6216\u908f\u8f2f\u3002<\/li>\n\n\n\n<li><strong>\u7a0b\u5f0f\u78bc\u7684\u53ef\u57f7\u884c\u6027\u548c\u6b63\u78ba\u6027\uff1a<\/strong> \u5c0d\u65bc\u7a0b\u5f0f\u7de8\u5beb\u4efb\u52d9\uff0c\u6a21\u578b\u751f\u6210\u7684\u7a0b\u5f0f\u78bc\u662f\u5426\u80fd\u5920\u6210\u529f\u7de8\u8b6f\u548c\u904b\u884c\uff0c\u4e26\u4e14\u5728\u7d66\u5b9a\u7684\u6e2c\u8a66\u6848\u4f8b\u4e0a\u7522\u751f\u6b63\u78ba\u7684\u8f38\u51fa\uff0c\u5c07\u76f4\u63a5\u5f71\u97ff\u6240\u7372\u5f97\u7684\u734e\u52f5\u3002<\/li>\n\n\n\n<li><strong>\u8207\u9810\u671f\u683c\u5f0f\u6216\u7d04\u675f\u7684\u7b26\u5408\u7a0b\u5ea6\uff1a<\/strong> \u5728\u67d0\u4e9b\u63a8\u7406\u4efb\u52d9\u4e2d\uff0c\u8f38\u51fa\u9700\u8981\u7b26\u5408\u7279\u5b9a\u7684\u683c\u5f0f\u6216\u6eff\u8db3\u67d0\u4e9b\u7d04\u675f\u689d\u4ef6\u3002\u6a21\u578b\u8f38\u51fa\u7b26\u5408\u9019\u4e9b\u8981\u6c42\u7684\u7a0b\u5ea6\u4e5f\u53ef\u80fd\u88ab\u7d0d\u5165\u734e\u52f5\u8a08\u7b97\u3002<\/li>\n<\/ol>\n\n\n\n<p>DeepSeek-R1-Zero \u4f7f\u7528\u4e86 <strong>Group Relative Policy Optimization (GRPO)<\/strong> \u9019\u500b RL \u6846\u67b6\u3002\u5728\u9019\u500b\u6846\u67b6\u4e0b\uff0c\u6a21\u578b\u6703\u751f\u6210\u4e00\u7d44\u53ef\u80fd\u7684\u8f38\u51fa (solutions)\uff0c\u7136\u5f8c\u4f7f\u7528\u4e00\u500b\u734e\u52f5\u51fd\u6578\u4f86\u5c0d\u9019\u4e9b\u8f38\u51fa\u9032\u884c\u8a55\u5206\u3002\u9019\u500b\u8a55\u5206\u5c31\u662f\u7528\u4f86\u8a08\u7b97\u734e\u52f5\u7684\u57fa\u790e\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0cDeepSeek-R1-Zero \u5728\u5f37\u5316\u5b78\u7fd2\u4e2d\u7684 Reward \u8a2d\u8a08\u662f\u5176\u80fd\u5920\u5728\u6c92\u6709\u76e3\u7763\u6578\u64da\u7684\u60c5\u6cc1\u4e0b\u5b78\u7fd2\u8907\u96dc\u63a8\u7406\u80fd\u529b\u7684\u95dc\u9375\u3002\u901a\u904e\u7cbe\u5fc3\u8a2d\u8a08\u7684\u898f\u5247\u5c0e\u5411\u734e\u52f5\u51fd\u6578\uff0c\u6a21\u578b\u80fd\u5920\u8fa8\u5225\u51fa\u54ea\u4e9b\u751f\u6210\u7684\u89e3\u6c7a\u65b9\u6848\u662f\u300c\u597d\u300d\u7684\uff0c\u54ea\u4e9b\u662f\u300c\u4e0d\u597d\u300d\u7684\uff0c\u4e26\u4ee5\u6b64\u70ba\u4f9d\u64da\u4e0d\u65b7\u6539\u9032\u5176\u751f\u6210\u63a8\u7406\u904e\u7a0b\u7684\u7b56\u7565\u3002\u9019\u7a2e\u7d14\u7cb9\u4f9d\u8cf4 RL \u548c\u81ea\u52d5\u5316\u734e\u52f5\u7684\u8a13\u7df4\u65b9\u5f0f\uff0c\u4f7f\u5f97 DeepSeek-R1-Zero \u80fd\u5920\u5c55\u73fe\u51fa\u81ea\u884c\u6f14\u5316\u548c\u767c\u73fe\u65b0\u63a8\u7406\u7b56\u7565\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n<p>DeepSeek-R1-Zero \u96d6\u7136\u5728\u8a13\u7df4\u6642<strong>\u6c92\u6709\u4f7f\u7528\u5927\u91cf\u7684\u4eba\u5de5\u6a19\u8a3b\u7684\u76e3\u7763\u5f0f\u6578\u64da (supervised data)<\/strong> \u4f86\u6307\u5c0e\u6bcf\u4e00\u6b65\u7684\u8f38\u51fa\u6216\u63d0\u4f9b\u6a19\u6e96\u7b54\u6848\uff0c\u4f46\u300c\u7cbe\u5fc3\u8a2d\u8a08\u7684\u898f\u5247\u5c0e\u5411\u734e\u52f5\u51fd\u6578\u300d\u662f\u901a\u904e\u4e00\u5957<strong>\u9810\u5148\u5b9a\u7fa9\u7684\u3001\u81ea\u52d5\u5316\u7684\u898f\u5247\u6216\u6f14\u7b97\u6cd5<\/strong>\u4f86\u8a55\u4f30\u6a21\u578b\u751f\u6210\u7684\u300c\u89e3\u6c7a\u65b9\u6848\u300d\u7684\u54c1\u8cea\u6216\u6b63\u78ba\u6027\u3002\u9019\u4e9b\u898f\u5247\u672c\u8eab\u860a\u542b\u4e86\u7279\u5b9a\u9818\u57df\uff08\u5982\u6578\u5b78\u3001\u7a0b\u5f0f\u3001\u908f\u8f2f\u63a8\u7406\uff09\u7684\u77e5\u8b58\uff0c\u4e26\u4f5c\u70ba\u4e00\u7a2e<strong>\u81ea\u52d5\u5316\u7684\u8a55\u5224\u6a19\u6e96<\/strong>\u4f86\u63d0\u4f9b\u56de\u994b\u4fe1\u865f\uff0c\u800c\u975e\u4eba\u5de5\u6a19\u8a3b\u3002<\/p>\n\n\n\n<p>\u4ee5\u4e0b\u662f\u9019\u7a2e\u898f\u5247\u5c0e\u5411\u734e\u52f5\u51fd\u6578\u5982\u4f55\u904b\u4f5c\u7684\u5177\u9ad4\u65b9\u5f0f\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u5b9a\u7fa9\u8a55\u4f30\u898f\u5247 (Defining Evaluation Rules):<\/strong> \u5c0d\u65bc\u4e0d\u540c\u7684\u63a8\u7406\u4efb\u52d9\uff0c\u8a2d\u8a08\u76f8\u61c9\u7684\u81ea\u52d5\u5316\u8a55\u4f30\u898f\u5247\u3002\u9019\u4e9b\u898f\u5247\u57fa\u65bc\u8a72\u9818\u57df\u7684\u5ba2\u89c0\u6a19\u6e96\u3002\n<ul class=\"wp-block-list\">\n<li><strong>\u6578\u5b78\uff1a<\/strong> \u898f\u5247\u53ef\u4ee5\u6aa2\u67e5\u6700\u7d42\u8a08\u7b97\u7d50\u679c\u662f\u5426\u8207\u901a\u904e\u7368\u7acb\u8a08\u7b97\u7a0b\u5e8f\u5f97\u51fa\u7684\u7d50\u679c\u4e00\u81f4\u3002\u5c0d\u65bc\u66f4\u8907\u96dc\u7684\u63a8\u7406\u6b65\u9a5f\uff0c\u53ef\u4ee5\u6709\u898f\u5247\u6aa2\u67e5\u4ee3\u6578\u8f49\u63db\u662f\u5426\u6709\u6548\uff0c\u6216\u662f\u5426\u9075\u5faa\u4e86\u5df2\u77e5\u7684\u6578\u5b78\u5b9a\u7406\u6216\u904b\u7b97\u9806\u5e8f\u3002<\/li>\n\n\n\n<li><strong>\u7a0b\u5f0f\u7de8\u5beb\uff1a<\/strong> \u898f\u5247\u53ef\u4ee5\u5305\u62ec\u5617\u8a66\u7de8\u8b6f\u6a21\u578b\u751f\u6210\u7684\u7a0b\u5f0f\u78bc\u3002\u5982\u679c\u7de8\u8b6f\u6210\u529f\uff0c\u5247\u9032\u4e00\u6b65\u904b\u884c\u7a0b\u5f0f\u78bc\u4e26\u4f7f\u7528\u4e00\u7d44\u9810\u8a2d\u7684\u6e2c\u8a66\u6848\u4f8b\u4f86\u9a57\u8b49\u5176\u8f38\u51fa\u662f\u5426\u6b63\u78ba\u3002\u7a0b\u5f0f\u78bc\u662f\u5426\u7b26\u5408\u7279\u5b9a\u7684\u683c\u5f0f\u898f\u7bc4\u6216\u6548\u7387\u8981\u6c42\u4e5f\u53ef\u4ee5\u7d0d\u5165\u898f\u5247\u3002<\/li>\n\n\n\n<li><strong>\u908f\u8f2f\u63a8\u7406\uff1a<\/strong> \u898f\u5247\u53ef\u4ee5\u57fa\u65bc\u5f62\u5f0f\u908f\u8f2f\u7cfb\u7d71\u3002\u4f8b\u5982\uff0c\u4f7f\u7528\u81ea\u52d5\u5b9a\u7406\u8b49\u660e\u5668 (automated theorem prover) \u4f86\u9a57\u8b49\u6a21\u578b\u5f97\u51fa\u7684\u7d50\u8ad6\u662f\u5426\u53ef\u4ee5\u5f9e\u7d66\u5b9a\u7684\u524d\u63d0\u901a\u904e\u6709\u6548\u7684\u63a8\u7406\u6b65\u9a5f\u5c0e\u51fa\u3002\u6216\u8005\u6aa2\u67e5\u6a21\u578b\u751f\u6210\u7684\u63a8\u7406\u93c8\u662f\u5426\u9075\u5faa\u4e86\u9810\u5b9a\u7fa9\u7684\u908f\u8f2f\u63a8\u7406\u898f\u5247 (\u5982\u80af\u5b9a\u524d\u4ef6\u3001\u5426\u5b9a\u5f8c\u4ef6\u7b49)\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u57f7\u884c\u6a21\u578b\u4e26\u751f\u6210\u89e3\u6c7a\u65b9\u6848 (Executing the Model and Generating Solutions):<\/strong> \u5728 RL \u8a13\u7df4\u904e\u7a0b\u4e2d\uff0cDeepSeek-R1-Zero \u6703\u6839\u64da\u7576\u524d\u7684\u7b56\u7565 (policy) \u91dd\u5c0d\u7d66\u5b9a\u7684\u554f\u984c\u751f\u6210\u4e00\u500b\u89e3\u6c7a\u65b9\u6848\u3002\u9019\u500b\u89e3\u6c7a\u65b9\u6848\u53ef\u80fd\u662f\u4e00\u500b\u6700\u7d42\u7b54\u6848\u3001\u4e00\u500b\u63a8\u7406\u6b65\u9a5f\u5e8f\u5217\uff0c\u6216\u8005\u4e00\u6bb5\u7a0b\u5f0f\u78bc\u3002<\/li>\n\n\n\n<li><strong>\u61c9\u7528\u898f\u5247\u5c0e\u5411\u734e\u52f5\u51fd\u6578\u8a55\u4f30 (Applying the Rule-Based Reward Function for Evaluation):<\/strong> \u5c07\u6a21\u578b\u751f\u6210\u7684\u89e3\u6c7a\u65b9\u6848\u8f38\u5165\u5230\u9810\u5148\u8a2d\u8a08\u597d\u7684\u898f\u5247\u5c0e\u5411\u734e\u52f5\u51fd\u6578\u4e2d\u3002\u9019\u500b\u51fd\u6578\u6703\u6839\u64da\u4e0a\u8ff0\u5b9a\u7fa9\u7684\u898f\u5247\u5c0d\u89e3\u6c7a\u65b9\u6848\u9032\u884c\u81ea\u52d5\u8a55\u4f30\u3002\n<ul class=\"wp-block-list\">\n<li>\u5982\u679c\u89e3\u6c7a\u65b9\u6848\u7b26\u5408\u300c\u6b63\u78ba\u300d\u7684\u6a19\u6e96 (\u4f8b\u5982\u6578\u5b78\u7b54\u6848\u6b63\u78ba\u3001\u7a0b\u5f0f\u78bc\u901a\u904e\u6240\u6709\u6e2c\u8a66\u6848\u4f8b)\uff0c\u734e\u52f5\u51fd\u6578\u6703\u8fd4\u56de\u4e00\u500b\u8f03\u9ad8\u7684\u6b63\u5411\u734e\u52f5\u503c\u3002<\/li>\n\n\n\n<li>\u5982\u679c\u89e3\u6c7a\u65b9\u6848\u932f\u8aa4\u6216\u4e0d\u7b26\u5408\u6a19\u6e96 (\u4f8b\u5982\u6578\u5b78\u7b54\u6848\u932f\u8aa4\u3001\u7a0b\u5f0f\u78bc\u7121\u6cd5\u7de8\u8b6f\u6216\u6e2c\u8a66\u5931\u6557)\uff0c\u734e\u52f5\u51fd\u6578\u6703\u8fd4\u56de\u96f6\u734e\u52f5\u6216\u8ca0\u5411\u61f2\u7f70\u503c\u3002<\/li>\n\n\n\n<li>\u5c0d\u65bc\u90e8\u5206\u6b63\u78ba\u6216\u5c55\u793a\u51fa\u4e00\u5b9a\u9032\u5c55\u7684\u89e3\u6c7a\u65b9\u6848\uff0c\u53ef\u80fd\u6703\u6709\u4e00\u500b\u68af\u5ea6\u7684\u734e\u52f5\uff0c\u4f8b\u5982\u5b8c\u6210\u90e8\u5206\u63a8\u7406\u6b65\u9a5f\u6216\u901a\u904e\u4e86\u90e8\u5206\u6e2c\u8a66\u6848\u4f8b\u3002<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>\u5c07\u734e\u52f5\u4fe1\u865f\u7528\u65bc\u7b56\u7565\u66f4\u65b0 (Using the Reward Signal for Policy Update):<\/strong> \u734e\u52f5\u51fd\u6578\u8fd4\u56de\u7684\u6a19\u91cf\u734e\u52f5\u503c\u88ab\u7528\u4f5c\u5f37\u5316\u5b78\u7fd2\u6f14\u7b97\u6cd5 (\u5982 GRPO) \u7684\u56de\u994b\u4fe1\u865f\u3002RL \u6f14\u7b97\u6cd5\u6839\u64da\u9019\u500b\u734e\u52f5\u4f86\u66f4\u65b0\u6a21\u578b\u7684\u53c3\u6578 (\u5373\u8abf\u6574\u5176\u7b56\u7565)\uff0c\u4f7f\u5f97\u6a21\u578b\u5728\u672a\u4f86\u66f4\u6709\u53ef\u80fd\u751f\u6210\u80fd\u5920\u7372\u5f97\u9ad8\u734e\u52f5\u7684\u89e3\u6c7a\u65b9\u6848\u3002<\/li>\n<\/ol>\n\n\n\n<p><strong>\u95dc\u9375\u9ede\u5728\u65bc\uff1a<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u734e\u52f5\u7684\u5ba2\u89c0\u6027\uff1a<\/strong> \u898f\u5247\u5c0e\u5411\u7684\u8a55\u4f30\u662f\u5ba2\u89c0\u4e14\u53ef\u91cd\u8907\u7684\uff0c\u4e0d\u4f9d\u8cf4\u65bc\u4eba\u5de5\u7684\u4e3b\u89c0\u5224\u65b7\u3002<\/li>\n\n\n\n<li><strong>\u9818\u57df\u77e5\u8b58\u7684\u9ad4\u73fe\uff1a<\/strong> \u898f\u5247\u672c\u8eab\u662f\u5c07\u7279\u5b9a\u9818\u57df\u7684\u6b63\u78ba\u6027\u6216\u6709\u6548\u6027\u6a19\u6e96\u7de8\u78bc\u9032\u53bb\u7684\u65b9\u5f0f\u3002\u9019\u662f\u5728\u6c92\u6709\u5927\u91cf\u9010\u6b65\u9a5f\u4eba\u5de5\u6a19\u8a3b\u7684\u60c5\u6cc1\u4e0b\u5f15\u5165\u9818\u57df\u77e5\u8b58\u7684\u624b\u6bb5\u3002<\/li>\n\n\n\n<li><strong>\u7121\u9700\u6a19\u6e96\u89e3\u7684\u6bcf\u500b\u6b65\u9a5f\uff1a<\/strong> \u8207\u76e3\u7763\u5b78\u7fd2\u9700\u8981\u6bcf\u4e00\u6b65\u7684\u6b63\u78ba\u6a19\u7c64\u4e0d\u540c\uff0cRL \u5728\u9019\u88e1\u53ea\u9700\u8981\u80fd\u5920\u81ea\u52d5\u8a55\u4f30\u6700\u7d42\u6216\u4e2d\u9593\u300c\u89e3\u6c7a\u65b9\u6848\u300d\u7684\u54c1\u8cea\u5373\u53ef\u3002\u6a21\u578b\u901a\u904e\u8a66\u932f\u5b78\u7fd2\u5982\u4f55\u5f9e\u554f\u984c\u751f\u6210\u9ad8\u8a55\u50f9\u7684\u89e3\u6c7a\u65b9\u6848\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u7d50\u5408 DeepSeek-R1-Zero \u4f7f\u7528\u7684 GRPO \u7b97\u6cd5\uff0c\u5b83\u53ef\u80fd\u4e0d\u662f\u55ae\u7368\u8a55\u4f30\u4e00\u500b\u8f38\u51fa\uff0c\u800c\u662f\u751f\u6210\u4e00\u6279\u5019\u9078\u89e3\u6c7a\u65b9\u6848\uff0c\u7136\u5f8c\u6839\u64da\u898f\u5247\u5c0d\u9019\u6279\u89e3\u6c7a\u65b9\u6848\u9032\u884c\u8a55\u5206\uff0c\u4e26\u53ef\u80fd\u5229\u7528\u9019\u4e9b\u76f8\u5c0d\u5206\u6578\u4f86\u8a08\u7b97\u734e\u52f5\u6216\u66f4\u65b0\u7b56\u7565\uff0c\u9019\u6709\u52a9\u65bc\u66f4\u6709\u6548\u5730\u63a2\u7d22\u548c\u5b78\u7fd2\u3002<\/p>\n\n\n\n<p>\u7e3d\u800c\u8a00\u4e4b\uff0c\u898f\u5247\u5c0e\u5411\u734e\u52f5\u51fd\u6578\u5728 DeepSeek-R1-Zero \u4e2d\u7684\u4f5c\u7528\u662f\u63d0\u4f9b\u4e00\u500b\u81ea\u52d5\u5316\u7684\u3001\u57fa\u65bc\u9818\u57df\u5ba2\u89c0\u6a19\u6e96\u7684\u300c\u8a55\u50f9\u8005\u300d\uff0c\u8b93\u6a21\u578b\u5728\u6c92\u6709\u4eba\u5de5\u624b\u628a\u624b\u6559\u5c0e\u6bcf\u500b\u6b65\u9a5f\u7684\u60c5\u6cc1\u4e0b\uff0c\u80fd\u5920\u901a\u904e\u5617\u8a66\u751f\u6210\u89e3\u6c7a\u65b9\u6848\u4e26\u89c0\u5bdf\u5176\u300c\u5f97\u5206\u300d\u4f86\u5b78\u7fd2\u5982\u4f55\u9032\u884c\u6709\u6548\u7684\u63a8\u7406\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u53c3\u8003\uff1a<\/h2>\n\n\n<div class=\"vlp-link-container vlp-layout-spotlight-clone wp-block-visual-link-preview-link\"><a href=\"https:\/\/iaee.substack.com\/p\/deepseek-r1-intuitively-and-exhaustively\" class=\"vlp-link\" title=\"DeepSeek-R1 \u2014 Intuitively and Exhaustively Explained\" rel=\"nofollow\" target=\"_blank\"><\/a><div class=\"vlp-layout-zone-main\"><span class=\"vlp-block-0 vlp-link-title\">DeepSeek-R1 \u2014 Intuitively and Exhaustively Explained<\/span><div class=\"vlp-block-1 vlp-link-summary\">An \u201cAha\u201d moment in Artificial Intelligence<\/div><div class=\"vlp-block-2 vlp-link-image\"><img decoding=\"async\" src=\"https:\/\/substackcdn.com\/image\/fetch\/w_1200,h_600,c_fill,f_jpg,q_auto:good,fl_progressive:steep,g_auto\/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd072e245-c242-45b3-a6d1-20dca65d34ac_1050x894.png\" style=\"max-width: 1024px; max-height: 1024px\" \/><\/div><\/div><\/div>","protected":false},"excerpt":{"rendered":"<p>DeepSeek-R1-Zero \u6700\u986f\u8457\u7684\u7279\u9ede\u662f\u5b83\u5b8c\u5168\u50c5\u4f7f\u7528\u5f37\u5316\u5b78\u7fd2 (RL) \u9032\u884c\u8a13\u7df4\uff0c\u800c\u6c92\u6709\u7d93\u904e\u76e3\u7763\u5f0f\u5fae\u8abf (SFT) \u7684\u521d\u6b65\u968e\u6bb5\u3002\u5b83\u901a\u904e\u8a66\u932f\u4f86\u5b78\u7fd2\uff0c\u4e26\u6839\u64da\u5176\u751f\u6210\u7684\u89e3\u6c7a\u65b9\u6848\u6240\u7372\u5f97\u7684\u56de\u994b\u4f86\u8abf\u6574\u81ea\u8eab\u7684\u884c\u70ba\u548c\u53c3\u6578\u3002 \u5728 DeepSeek-R1-Zero \u7528\u65bc\u63a8\u7406\u4efb\u52d9 (\u4f8b\u5982\u6578\u5b78\u554f\u984c\u3001\u7a0b\u5f0f\u7de8\u5beb\u7b49) \u7684\u5f37\u5316\u5b78\u7fd2\u8a13\u7df4\u4e2d\uff0c\u5176 Reward (\u734e\u52f5) \u6a5f\u5236\u4e3b\u8981\u662f\u57fa\u65bc\u5c0d\u6a21\u578b\u8f38\u51fa\u7684\u81ea\u52d5\u8a55\u4f30\uff0c\u9019\u7a2e\u8a55\u4f30\u901a\u5e38\u662f\u898f\u5247\u5c0e\u5411 (rule-based) \u7684\u3002 \u5177\u9ad4\u4f86\u8aaa\uff0c\u734e\u52f5\u7684\u8a2d\u8a08\u6703\u76f4\u63a5\u53cd\u6620\u6a21\u578b\u751f\u6210\u7684\u89e3\u6c7a\u65b9\u6848\u7684\u6b63\u78ba\u6027\u6216\u54c1\u8cea\u3002\u8003\u616e\u5230 DeepSeek-R1-Zero \u5c08\u6ce8\u65bc\u63a8\u7406\u4efb\u52d9\uff0c\u734e\u52f5\u53ef\u80fd\u4f86\u81ea\u65bc\u4ee5\u4e0b\u5e7e\u500b\u65b9\u9762\uff1a DeepSeek-R1-Zero \u4f7f\u7528\u4e86 Group Relative Policy Optimization (GRPO) \u9019\u500b RL \u6846\u67b6\u3002\u5728\u9019\u500b\u6846\u67b6\u4e0b\uff0c\u6a21\u578b\u6703\u751f\u6210\u4e00\u7d44\u53ef\u80fd\u7684\u8f38\u51fa (solutions)\uff0c\u7136\u5f8c\u4f7f\u7528\u4e00\u500b\u734e\u52f5\u51fd\u6578\u4f86\u5c0d\u9019\u4e9b\u8f38\u51fa\u9032\u884c\u8a55\u5206\u3002\u9019\u500b\u8a55\u5206\u5c31\u662f\u7528\u4f86\u8a08\u7b97\u734e\u52f5\u7684\u57fa\u790e\u3002 \u56e0\u6b64\uff0cDeepSeek-R1-Zero \u5728\u5f37\u5316\u5b78\u7fd2\u4e2d\u7684 Reward \u8a2d\u8a08\u662f\u5176\u80fd\u5920\u5728\u6c92\u6709\u76e3\u7763\u6578\u64da\u7684\u60c5\u6cc1\u4e0b\u5b78\u7fd2\u8907\u96dc\u63a8\u7406\u80fd\u529b\u7684\u95dc\u9375\u3002\u901a\u904e\u7cbe\u5fc3\u8a2d\u8a08\u7684\u898f\u5247\u5c0e\u5411\u734e\u52f5\u51fd\u6578\uff0c\u6a21\u578b\u80fd\u5920\u8fa8\u5225\u51fa\u54ea\u4e9b\u751f\u6210\u7684\u89e3\u6c7a\u65b9\u6848\u662f\u300c\u597d\u300d\u7684\uff0c\u54ea\u4e9b\u662f\u300c\u4e0d\u597d\u300d\u7684\uff0c\u4e26\u4ee5\u6b64\u70ba\u4f9d\u64da\u4e0d\u65b7\u6539\u9032\u5176\u751f\u6210\u63a8\u7406\u904e\u7a0b\u7684\u7b56\u7565\u3002\u9019\u7a2e\u7d14\u7cb9\u4f9d\u8cf4 RL \u548c\u81ea\u52d5\u5316\u734e\u52f5\u7684\u8a13\u7df4\u65b9\u5f0f\uff0c\u4f7f\u5f97 DeepSeek-R1-Zero \u80fd\u5920\u5c55\u73fe\u51fa\u81ea\u884c\u6f14\u5316\u548c\u767c\u73fe\u65b0\u63a8\u7406\u7b56\u7565\u7684\u80fd\u529b\u3002 DeepSeek-R1-Zero \u96d6\u7136\u5728\u8a13\u7df4\u6642\u6c92\u6709\u4f7f\u7528\u5927\u91cf\u7684\u4eba\u5de5\u6a19\u8a3b\u7684\u76e3\u7763\u5f0f\u6578\u64da (supervised data) \u4f86\u6307\u5c0e\u6bcf\u4e00\u6b65\u7684\u8f38\u51fa\u6216\u63d0\u4f9b\u6a19\u6e96\u7b54\u6848\uff0c\u4f46\u300c\u7cbe\u5fc3\u8a2d\u8a08\u7684\u898f\u5247\u5c0e\u5411\u734e\u52f5\u51fd\u6578\u300d\u662f\u901a\u904e\u4e00\u5957\u9810\u5148\u5b9a\u7fa9\u7684\u3001\u81ea\u52d5\u5316\u7684\u898f\u5247\u6216\u6f14\u7b97\u6cd5\u4f86\u8a55\u4f30\u6a21\u578b\u751f\u6210\u7684\u300c\u89e3\u6c7a\u65b9\u6848\u300d\u7684\u54c1\u8cea\u6216\u6b63\u78ba\u6027\u3002\u9019\u4e9b\u898f\u5247\u672c\u8eab\u860a\u542b\u4e86\u7279\u5b9a\u9818\u57df\uff08\u5982\u6578\u5b78\u3001\u7a0b\u5f0f\u3001\u908f\u8f2f\u63a8\u7406\uff09\u7684\u77e5\u8b58\uff0c\u4e26\u4f5c\u70ba\u4e00\u7a2e\u81ea\u52d5\u5316\u7684\u8a55\u5224\u6a19\u6e96\u4f86\u63d0\u4f9b\u56de\u994b\u4fe1\u865f\uff0c\u800c\u975e\u4eba\u5de5\u6a19\u8a3b\u3002 \u4ee5\u4e0b\u662f\u9019\u7a2e\u898f\u5247\u5c0e\u5411\u734e\u52f5\u51fd\u6578\u5982\u4f55\u904b\u4f5c\u7684\u5177\u9ad4\u65b9\u5f0f\uff1a \u95dc\u9375\u9ede\u5728\u65bc\uff1a \u7d50\u5408 DeepSeek-R1-Zero \u4f7f\u7528\u7684 GRPO \u7b97\u6cd5\uff0c\u5b83\u53ef\u80fd\u4e0d\u662f\u55ae\u7368\u8a55\u4f30\u4e00\u500b\u8f38\u51fa\uff0c\u800c\u662f\u751f\u6210\u4e00\u6279\u5019\u9078\u89e3\u6c7a\u65b9\u6848\uff0c\u7136\u5f8c\u6839\u64da\u898f\u5247\u5c0d\u9019\u6279\u89e3\u6c7a\u65b9\u6848\u9032\u884c\u8a55\u5206\uff0c\u4e26\u53ef\u80fd\u5229\u7528\u9019\u4e9b\u76f8\u5c0d\u5206\u6578\u4f86\u8a08\u7b97\u734e\u52f5\u6216\u66f4\u65b0\u7b56\u7565\uff0c\u9019\u6709\u52a9\u65bc\u66f4\u6709\u6548\u5730\u63a2\u7d22\u548c\u5b78\u7fd2\u3002 \u7e3d\u800c\u8a00\u4e4b\uff0c\u898f\u5247\u5c0e\u5411\u734e\u52f5\u51fd\u6578\u5728 DeepSeek-R1-Zero \u4e2d\u7684\u4f5c\u7528\u662f\u63d0\u4f9b\u4e00\u500b\u81ea\u52d5\u5316\u7684\u3001\u57fa\u65bc\u9818\u57df\u5ba2\u89c0\u6a19\u6e96\u7684\u300c\u8a55\u50f9\u8005\u300d\uff0c\u8b93\u6a21\u578b\u5728\u6c92\u6709\u4eba\u5de5\u624b\u628a\u624b\u6559\u5c0e\u6bcf\u500b\u6b65\u9a5f\u7684\u60c5\u6cc1\u4e0b\uff0c\u80fd\u5920\u901a\u904e\u5617\u8a66\u751f\u6210\u89e3\u6c7a\u65b9\u6848\u4e26\u89c0\u5bdf\u5176\u300c\u5f97\u5206\u300d\u4f86\u5b78\u7fd2\u5982\u4f55\u9032\u884c\u6709\u6548\u7684\u63a8\u7406\u3002 \u53c3\u8003\uff1a<\/p>\n","protected":false},"author":1,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":{"googlesitekit_rrm_CAowvqSiDA:productID":"","footnotes":""},"class_list":["post-5515","page","type-page","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/pages\/5515","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=5515"}],"version-history":[{"count":0,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/pages\/5515\/revisions"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=5515"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}