﻿{"id":9221,"date":"2025-03-02T07:18:00","date_gmt":"2025-03-01T23:18:00","guid":{"rendered":"http:\/\/www.nlpir.org\/wordpress\/?p=9221"},"modified":"2025-03-11T04:20:20","modified_gmt":"2025-03-10T20:20:20","slug":"%e3%80%90%e8%bd%ac%e8%bd%bd%e3%80%91deepseek%e5%90%af%e7%a4%ba%e5%8f%af%e4%bf%a1%e5%8f%af%e6%8e%a7%e5%8f%af%e7%94%a8%e7%9a%84%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%9c%aa%e6%9d%a5%e4%b9%8b%e8%b7%af","status":"publish","type":"post","link":"http:\/\/www.nlpir.org\/wordpress\/2025\/03\/02\/%e3%80%90%e8%bd%ac%e8%bd%bd%e3%80%91deepseek%e5%90%af%e7%a4%ba%e5%8f%af%e4%bf%a1%e5%8f%af%e6%8e%a7%e5%8f%af%e7%94%a8%e7%9a%84%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%9c%aa%e6%9d%a5%e4%b9%8b%e8%b7%af\/","title":{"rendered":"\u3010\u8f6c\u8f7d\u3011DeepSeek\u542f\u793a:\u53ef\u4fe1\u53ef\u63a7\u53ef\u7528\u7684\u5927\u6a21\u578b\u672a\u6765\u4e4b\u8def"},"content":{"rendered":"\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"has-text-align-right\">\u2014\u2014<em>\u539f\u6587\u4e8e2025\u5e743\u67081\u65e5\u53d1\u5e03\u5728\u300aIMLIP\u591a\u8bed\u79cd\u667a\u80fd\u4fe1\u606f\u5904\u7406\u300b\u5fae\u4fe1\u516c\u4f17\u53f7<\/em><\/p>\n<\/blockquote>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p class=\"has-text-align-right\">\u2014\u2014<em>\u8bba\u6587\u4e8e2025\u5e742\u670828\u65e5\u520a\u53d1\u5728\u300a\u4e2d\u56fd\u4eba\u5de5\u667a\u80fd\u5b66\u53f7\u901a\u8baf\u300b\u7b2c15\u53772025\u5e74\u7b2c2\u671f<\/em><\/p>\n<\/blockquote>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading has-text-align-center\" style=\"font-size:26px\"><strong>DeepSeek\u542f\u793a:\u53ef\u4fe1\u53ef\u63a7\u53ef\u7528\u7684\u5927\u6a21\u578b\u672a\u6765\u4e4b\u8def<\/strong><\/h1>\n\n\n\n<p class=\"has-text-align-center\" style=\"font-size:18px\"><strong>\u6587\/\u5d14\u6587\u8000\uff0c\u674e\u6797\u7ff0\uff0c\u6797\u5b50\u8d85\uff0c\u738b\u4e00\u8bfa\uff0c\u66f9\u54f2\u4fca\uff0c\u5f20\u534e\u5e73<\/strong><\/p>\n\n\n\n<div class=\"wp-block-group\"><div class=\"wp-block-group__inner-container is-layout-constrained wp-block-group-is-layout-constrained\">\n<div class=\"wp-block-group\"><div class=\"wp-block-group__inner-container is-layout-constrained wp-block-group-is-layout-constrained\">\n<h2 class=\"wp-block-heading has-text-align-center\" style=\"font-size:26px\">\u6458\u8981<\/h2>\n\n\n\n<p>\u53d7DeepSeek\u7cfb\u5217\u6a21\u578b\u7684\u6210\u529f\u5b9e\u8df5\u542f\u53d1\uff0c\u672c\u6587\u7cfb\u7edf\u63a2\u8ba8\u4e86\u53ef\u4fe1\u3001\u53ef\u63a7\u3001\u53ef\u7528\u5927\u6a21\u578b\u7684\u6838\u5fc3\u6280\u672f\u8def\u5f84\u4e0e\u672a\u6765\u53d1\u5c55\u65b9\u5411\u3002\u901a\u8fc7\u89e3\u6790DeepSeek\u5728\u6a21\u578b\u67b6\u6784\u4e0e\u8bad\u7ec3\u8303\u5f0f\u4e0a\u7684\u521b\u65b0\u7a81\u7834,\u5305\u62ecMLA\u673a\u5236\u3001GRPO\u7b97\u6cd5\u7b49\uff0c\u63ed\u793a\u5176\u5728\u63d0\u5347\u63a8\u7406\u6548\u7387\u65b9\u9762\u7684\u663e\u8457\u4f18\u52bf\u3002\u5b9e\u9a8c\u8868\u660e\uff0cDeepSeek-R1\u901a\u8fc7\u7eaf\u5f3a\u5316\u5b66\u4e60\u8303\u5f0f\u4e0e\u89c4\u5219\u5956\u52b1\u673a\u5236\uff0c\u5728\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u4e2d\u5b9e\u73b071%\u7684pass@1\u51c6\u786e\u7387\uff0c\u5e76\u5728LMSYS\u7ade\u6280\u573a\u5f00\u6e90\u6a21\u578b\u4e2d\u4f4d\u5217\u524d\u4e94\uff1b\u5176\u84b8\u998f\u6a21\u578b\u4ec5\u4ee514B\u53c2\u6570\u5373\u8d85\u8d8a\u90e8\u520670B\u7ea7\u6a21\u578b\u6027\u80fd\uff0c\u9a8c\u8bc1\u4e86\u8f7b\u91cf\u5316\u90e8\u7f72\u7684\u53ef\u884c\u6027\u3002\u6b64\u5916\u8fdb\u4e00\u6b65\u63d0\u51fa\u57fa\u4e8e\u81ea\u6211\u4e00\u81f4\u6027\u8bc4\u5224\u7684\u5956\u52b1\u673a\u5236\uff0c\u901a\u8fc7\u5185\u5316\u601d\u7ef4\u65b9\u6cd5\u4e0e\u7ec4\u5185\u7ea6\u675f\u4f18\u5316\uff0c\u5728ChatBIT\u8f7b\u91cf\u6a21\u578b\u4e2d\u5b9e\u73b0\u4e86\u6a21\u578b\u5956\u52b1\u63a8\u7406\u8fc7\u7a0b\u900f\u660e\u5316\u4e0e\u81ea\u6211\u4e00\u81f4\u6027\uff0c\u5b9e\u9a8c\u8868\u660e1.5B\u7684ChatBIT\u6a21\u578b\u4ec5\u9700\u8bad\u7ec31500\u6b65\u5373\u53ef\u5f88\u597d\u5730\u5b8c\u6210\u81ea\u6211\u4e00\u81f4\u6027\u7684\u8bc4\u5224\u4efb\u52a1\uff0c\u8fd9\u4e5f\u4e3a\u6784\u5efa\u53ef\u4fe1\u53ef\u63a7\u53ef\u7528\u7684\u5927\u6a21\u578b\u70b9\u4eae\u4e86\u4e00\u6761\u672a\u6765\u4e4b\u8def\u3002<\/p>\n<\/div><\/div>\n\n\n\n<div class=\"wp-block-group\"><div class=\"wp-block-group__inner-container is-layout-constrained wp-block-group-is-layout-constrained\">\n<h2 class=\"wp-block-heading has-text-align-center\" style=\"font-size:26px\">\u5173\u952e\u8bcd<\/h2>\n\n\n\n<p class=\"has-text-align-center\">DeepSeek\u3001ChatBIT\u3001\u5927\u6a21\u578b\u3001\u63a8\u7406\u3001GRPO\u3001\u5f3a\u5316\u5b66\u4e60<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\" style=\"font-size:26px\">0 \u5f15\u8a00<\/h2>\n\n\n\n<p>\u8fd1\u5e74\u6765\uff0c\u968f\u7740\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u6a21\u578b\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u548c\u4eba\u5de5\u667a\u80fd\u9886\u57df\u7684\u4e0d\u65ad\u7a81\u7834\uff0c\u5982\u4f55\u5728\u786e\u4fdd\u6a21\u578b\u6027\u80fd\u4e0d\u65ad\u63d0\u5347\u7684\u540c\u65f6\uff0c\u5b9e\u73b0\u53ef\u4fe1\u3001\u53ef\u63a7\u3001\u53ef\u7528\u5df2\u6210\u4e3a\u4e9f\u5f85\u89e3\u51b3\u7684\u91cd\u8981\u8bfe\u9898\u3002DeepSeek\u7cfb\u5217\u6a21\u578b\u4f5c\u4e3a\u4e2d\u56fd\u4eba\u5de5\u667a\u80fd\u521b\u65b0\u53d1\u5c55\u7684\u91cd\u8981\u4ee3\u8868\uff08\u89c1\u88681\uff09\uff0c\u4ee5\u5176\u7ecf\u6d4e\u9ad8\u6548\u7684\u8bbe\u8ba1\u7406\u5ff5\u548c\u5353\u8d8a\u7684\u6280\u672f\u8868\u73b0\uff0c\u6807\u5fd7\u7740\u5927\u6a21\u578b\u53d1\u5c55\u8fdb\u5165\u4e86\u4e00\u4e2a\u5168\u65b0\u7684\u9636\u6bb5\u3002<\/p>\n\n\n\n<p>2025\u5e74\uff0cDeepSeek-R1\u6a21\u578b\u5728DeepSeek-V3\u7684\u575a\u5b9e\u57fa\u5ea7\u4e0a\uff0c\u901a\u8fc7\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\u5bf9\u957f\u601d\u7ef4\u94fe\u63a8\u7406\u4efb\u52a1\u8fdb\u884c\u9488\u5bf9\u6027\u4f18\u5316\uff0c\u4f7f\u5f97\u5728\u63a8\u7406\u9886\u57df\u7684\u51c6\u786e\u6027\u5b9e\u73b0\u4e86\u663e\u8457\u98de\u8dc3\uff0c\u5e76\u5728LMSYS\u5927\u6a21\u578b\u7ade\u6280\u573a\u4e2d\u540d\u5217\u524d\u8305\uff08\u89c1\u88682\uff09\u3002DeepSeek\u7cfb\u5217\u6a21\u578b\u7684\u6210\u529f\uff0c\u4e0d\u4ec5\u5c55\u793a\u4e86\u5728\u6a21\u578b\u67b6\u6784\u3001\u8bad\u7ec3\u6d41\u7a0b\u548c\u63a8\u7406\u7b56\u7565\u4e0a\u7684\u591a\u7ef4\u521b\u65b0\uff0c\u66f4\u4e3a\u5927\u6a21\u578b\u5728\u53ef\u4fe1\u3001\u53ef\u63a7\u3001\u53ef\u7528\u4e09\u4e2a\u7ef4\u5ea6\u7684\u53d1\u5c55\u63d0\u4f9b\u4e86\u5b9d\u8d35\u542f\u793a\u3002<\/p>\n\n\n\n<p>\u672c\u6587\u5c06\u56f4\u7ed5DeepSeek\u7cfb\u5217\u6a21\u578b\u7684\u6280\u672f\u57fa\u7840\uff0c\u7cfb\u7edf\u5730\u89e3\u6790\u5176\u5728\u5b9e\u73b0\u9ad8\u6027\u80fd\u7684\u540c\u65f6\u517c\u987e\u7ecf\u6d4e\u6027\u4e0e\u7a33\u5b9a\u6027\u65b9\u9762\u7684\u5173\u952e\u6280\u672f\u7a81\u7834\uff0c\u5e76\u4ece\u53ef\u4fe1\u6027\u3001\u53ef\u63a7\u6027\u548c\u53ef\u7528\u6027\u4e09\u4e2a\u7ef4\u5ea6\u63a2\u8ba8\u5927\u6a21\u578b\u672a\u6765\u7684\u53d1\u5c55\u65b9\u5411\u3002\u540c\u65f6\uff0c\u7ed3\u5408\u5317\u4eac\u7406\u5de5\u5927\u5b66\u7684ChatBIT\u5728\u5e94\u5bf9\u590d\u6742\u63a8\u7406\u4efb\u52a1\u4e0e\u5956\u52b1\u673a\u5236\u4f18\u5316\u65b9\u9762\u7684\u5b9e\u8df5\u7ecf\u9a8c\uff0c\u8fdb\u4e00\u6b65\u5256\u6790\u5982\u4f55\u901a\u8fc7\u5185\u5316\u601d\u7ef4\u65b9\u6cd5\u548c\u81ea\u6211\u4e00\u81f4\u6027\u8bc4\u5224\u6765\u63d0\u5347\u6a21\u578b\u63a8\u7406\u8d28\u91cf\uff0c\u4e3a\u6784\u5efa\u66f4\u8f7b\u91cf\u5316\u3001\u53ef\u63a7\u7684\u5927\u6a21\u578b\u63d0\u4f9b\u5b9e\u8df5\u6837\u672c\u548c\u7406\u8bba\u652f\u6301\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><tbody><tr><td class=\"has-text-align-left\" data-align=\"left\"><strong>\u9636\u6bb5<\/strong><\/td><td><strong>\u9884\u8bad\u7ec3<\/strong><\/td><td><strong>\u6269\u5c55\u4e0a\u4e0b\u6587<\/strong><\/td><td><strong>\u5fae\u8c03<\/strong><\/td><td><strong>\u603b\u8ba1<\/strong><\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\">GPU\/\u5c0f\u65f6<\/td><td>2664<\/td><td>119<\/td><td>5<\/td><td>2788<\/td><\/tr><tr><td class=\"has-text-align-left\" data-align=\"left\">\u6210\u672c\/\u767e\u4e07\u7f8e\u5143<\/td><td>5.328<\/td><td>0.24<\/td><td>0.01<\/td><td>5.576<\/td><\/tr><\/tbody><\/table><figcaption class=\"wp-element-caption\">\u88681 DeepSeek-R1\u8bad\u7ec3\u6210\u672c<\/figcaption><\/figure>\n<\/div><\/div>\n<\/div><\/div>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td>\u5f00\u6e90\u6392\u540d<\/td><td>1<\/td><td>2<\/td><td>3<\/td><td>4<\/td><td>5<\/td><\/tr><tr><td>\u6574\u4f53\u6392\u540d<\/td><td>5<\/td><td>10<\/td><td>18<\/td><td>22<\/td><td>23<\/td><\/tr><tr><td>\u6a21\u578b<\/td><td>DeepSeek-R1<\/td><td>DeepSeek-V3<\/td><td>Deepseek-v2.5-1210<\/td><td>Athene-v2-Chat-72B<\/td><td>Llama-3.1-Nemotron-70B-Instruct<\/td><\/tr><tr><td>\u5f97\u5206<\/td><td>1361<\/td><td>1317<\/td><td>1279<\/td><td>1275<\/td><td>1268<\/td><\/tr><\/tbody><\/table><figcaption class=\"wp-element-caption\">\u88682 LMSYS\u5927\u6a21\u578b\u7ade\u6280\u573a\u5f00\u6e90\u6a21\u578b\u6392\u540d<\/figcaption><\/figure>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\" style=\"font-size:26px\">1 <strong>DeepSeek\u6a21\u578b\u521b\u65b0\u7b80\u4ecb<\/strong><\/h2>\n\n\n\n<p>1.1 \u6a21\u578b\u67b6\u6784\u521b\u65b0<\/p>\n\n\n\n<p>2024 \u5e74 5 \u6708\u53d1\u5e03\u7684 DeepSeek-V2 \u9996\u6b21\u63d0\u51fa\u4e86MLA\u673a\u5236\u3002\u5728\u4f20\u7edf\u591a\u5934\u81ea\u6ce8\u610f\u529b\uff08MHA\uff09\u7684\u57fa\u7840\u4e0a\uff0cMLA \u5f15\u5165\u4e86\u663e\u5f0f\u7684\u201c\u6f5c\u5728\u5411\u91cf\u201d\uff0c\u5728\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u5185\u5206\u522b\u5b8c\u6210\u5b50\u7a7a\u95f4\u7684\u6295\u5f71\u548c\u805a\u5408\u3002\u6b64\u4e3e\u4e0d\u4ec5\u4fdd\u7559\u4e86\u591a\u5934\u673a\u5236\u7684\u5e76\u884c\u8ba1\u7b97\u4f18\u52bf\uff0c\u8fd8\u80fd\u6781\u5927\u5730\u538b\u7f29 KVCache \u5bf9\u663e\u5b58\u548c\u5176\u4ed6\u8ba1\u7b97\u8d44\u6e90\u7684\u9700\u6c42\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u6a21\u578b\u5bf9\u8f93\u5165\u6570\u636e\u7684\u590d\u6742\u5173\u7cfb\u7684\u6355\u6349\u80fd\u529b\u3002\u4e0e\u4f20\u7edf\u7684 MHA\u3001MQA\uff08multi query attention\uff09\u548c GQA\uff08group query attention\uff09\u76f8\u6bd4\uff0cMLA \u5728\u57fa\u672c\u4e0d\u635f\u5931\u6027\u80fd\u7684\u524d\u63d0\u4e0b\uff0c\u5b9e\u73b0\u4e86\u5bf9\u5927\u6a21\u578b\u957f\u6587\u672c\u63a8\u7406\u8d44\u6e90\u6d88\u8017\u7684\u6709\u6548\u964d\u4f4e\uff0c\u5982\u56fe1\u6240\u793a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"315\" src=\"http:\/\/www.nlpir.org\/wordpress\/wp-content\/uploads\/2025\/03\/1.webp\" alt=\"\" class=\"wp-image-9222\"\/><figcaption class=\"wp-element-caption\">\u56fe 1 MHA\u3001GQA\u3001MQA\u4e0eMLA\u7684\u6bd4\u8f83\uff08\u6e90\u81eaDeepSeek v2\uff09<\/figcaption><\/figure>\n\n\n\n<p>\u5728 DeepSeek-V2 \u4e2d\uff0c\u7814\u7a76\u8005\u8fdb\u4e00\u6b65\u6539\u8fdb\u4e86MoE\uff0c\u901a\u8fc7\u5f15\u5165\u4e00\u79cd\u8f85\u52a9\u635f\u5931\uff0c\u80fd\u591f\u5728\u8bad\u7ec3\u9636\u6bb5\u5b9e\u73b0\u591a\u4e2a\u201c\u4e13\u5bb6\u201d\u7684\u52a8\u6001\u8def\u7531\u5206\u914d\uff0c\u4ee5\u786e\u4fdd\u5404\u4e2a\u4e13\u5bb6\u5728\u6279\u5904\u7406\u4e2d\u7684\u8d1f\u8f7d\u5e73\u8861\uff0c\u907f\u514d\u4e2a\u522b\u4e13\u5bb6\u6210\u4e3a\u201c\u74f6\u9888\u201d\u6216\u201c\u95f2\u7f6e\u201d\uff0c\u5728\u4fdd\u8bc1\u5404\u4e13\u5bb6\u53c2\u6570\u4e0d\u968f\u610f\u81a8\u80c0\u7684\u524d\u63d0\u4e0b\uff0c\u5145\u5206\u5229\u7528\u5e76\u884c\u8ba1\u7b97\u4e0e\u7a00\u758f\u6fc0\u6d3b\uff0c\u8282\u7701\u8ba1\u7b97\u5f00\u9500\u5e76\u63d0\u9ad8\u6a21\u578b\u7684\u63a8\u7406\u901f\u5ea6\u30022024 \u5e74 12 \u6708\u53d1\u5e03\u7684 DeepSeek-V3 \u5728\u7ee7\u627f DeepSeek-V2 \u7684 MoE \u6846\u67b6\u4e4b\u4e0a\uff0c\u8fdb\u4e00\u6b65\u63d0\u51fa\u4e86<strong>\u65e0\u8f85\u52a9\u635f\u5931\u7684\u4e13\u5bb6\u8d1f\u8f7d\u5747\u8861\u7b56\u7565<\/strong>\u3002\u5728 DeepSeek-V2 \u4e2d\uff0c\u867d\u7136\u8f85\u52a9\u635f\u5931\u80fd\u591f\u6709\u6548\u6307\u5bfc\u4e13\u5bb6\u8def\u7531\u7684\u5e73\u8861\u5206\u914d\uff0c\u4f46\u5728\u67d0\u4e9b\u573a\u666f\u4e0b\u53ef\u80fd\u5f15\u5165\u989d\u5916\u4f18\u5316\u76ee\u6807\u7684\u51b2\u7a81\u3002DeepSeek-V3\u901a\u8fc7\u5f15\u5165\u52a8\u6001\u8def\u7531\u4e0e\u65e0\u635f\u8017\u8d1f\u8f7d\u5747\u8861\u7b56\u7565\uff0c\u5c06\u4e13\u5bb6\u8def\u7531\u672c\u8eab\u5d4c\u5165\u5230\u4e3b\u635f\u5931\u7684\u68af\u5ea6\u66f4\u65b0\u4e4b\u4e2d\uff0c\u907f\u514d\u989d\u5916\u8f85\u52a9\u635f\u5931\u7684\u5f15\u5165\uff0c\u540c\u65f6\u4fdd\u6301\u591a\u4e13\u5bb6\u5728\u5927\u6279\u91cf\u8f93\u5165\u65f6\u7684\u9ad8\u6548\u5206\u914d\u3002\u6bcf\u4e2a Token \u6fc0\u6d3b 673 B \u53c2\u6570\uff0c\u603b\u8ba1 37 B \u53c2\u6570\uff08\u7ea6\u5360\u6574\u4f53\u53c2\u6570\u7684 5.5%\uff09\uff0c\u6781\u5927\u5730\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u8ba1\u7b97\u6548\u7387\u3002\u9664\u6b64\u4e4b\u5916\u591aDeepSeek\u8fd8\u5f15\u5165\u4e86Token\u9884\u6d4b\u673a\u5236\u53caFP8 \u6df7\u5408\u7cbe\u5ea6\u8bad\u7ec3\u4e0eDualPipe\u8fdb\u4e00\u6b65\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u8ba1\u7b97\u6548\u7387\u3002<\/p>\n\n\n\n<p>1.2 \u63a8\u7406\u4f18\u5316\u521b\u65b0<\/p>\n\n\n\n<p>1.2.1 GRPO \u7b97\u6cd5<\/p>\n\n\n\n<p>\u5728\u5f3a\u5316\u5b66\u4e60\u5728\u5927\u8bed\u8a00\u6a21\u578b\u540e\u8bad\u7ec3\u9636\u6bb5\u7684\u5e94\u7528\u4e2d\uff0c\u4e3b\u6d41\u65b9\u6cd5\uff08\u5982 PPO\uff09\u867d\u7136\u5df2\u7ecf\u53d6\u5f97\u4e86\u826f\u597d\u7684\u5bf9\u9f50\u4e0e\u6027\u80fd\u589e\u76ca\uff0c\u4f46\u5f80\u5f80\u9700\u8981\u989d\u5916\u7684\u4ef7\u503c\u7f51\u7edc\uff08value network\uff09\u6765\u8fdb\u884c\u7b56\u7565\u8bc4\u4f30\uff0c\u4ece\u800c\u589e\u52a0\u4e86\u6a21\u578b\u89c4\u6a21\u4e0e\u8bad\u7ec3\u590d\u6742\u5ea6\u3002DeepSeek\u4e8e2024 \u5e74 2 \u6708\u5728DeepSeek Math\u4e2d\u9996\u6b21\u63d0\u51fa\u7ec4\u5185\u76f8\u5bf9\u7b56\u7565\u4f18\u5316GRPO\u7b97\u6cd5\uff0c\u65e8\u5728\u51cf\u5c11\u5bf9\u4e8e\u5e9e\u5927 Critic \u6a21\u578b\u7684\u4f9d\u8d56\u5e76\u63d0\u5347\u8bad\u7ec3\u6548\u7387\u3002<\/p>\n\n\n\n<p>GRPO \u7684\u6838\u5fc3\u601d\u8def\u5982\u4e0b\u6240\u793a\u3002<\/p>\n\n\n\n<p>\uff081\uff09\u7fa4\u7ec4\u76f8\u5bf9\u4f18\u52bf\u4f30\u8ba1\u3002\u5728\u5bf9\u540c\u4e00\u8f93\u5165\u8fdb\u884c\u591a\u6b21\u91c7\u6837\u540e\uff0c\u5c06\u8f93\u51fa\u7ed3\u679c\u7ec4\u5408\u6210\u4e3a\u4e00\u4e2a\u5c0f\u7684\u7fa4\u7ec4\uff0c\u901a\u8fc7\u6bd4\u8f83\u7fa4\u7ec4\u5185\u90e8\u5404\u4e2a\u8f93\u51fa\u7684\u4f18\u52a3\u6765\u4f30\u7b97\u76f8\u5bf9\u4f18\u52bf\u503c\uff08advantage\uff09\u3002<\/p>\n\n\n\n<p>\uff082\uff09\u7b56\u7565\u76ee\u6807\u51fd\u6570\u91cd\u6784\u3002\u5c06\u7fa4\u7ec4\u5185\u90e8\u7684\u76f8\u5bf9\u4f18\u52bf\u800c\u975e\u5355\u4e2a\u6837\u672c\u7684\u7edd\u5bf9\u4ef7\u503c\u7eb3\u5165\u5f3a\u5316\u5b66\u4e60\u76ee\u6807\uff0c\u5728\u6bcf\u4e2a\u8bad\u7ec3\u8fed\u4ee3\u4e2d\u76f4\u63a5\u5bf9\u7b56\u7565\u7f51\u7edc\u8fdb\u884c\u66f4\u65b0\uff0c\u907f\u514d\u5728\u5927\u6a21\u578b\u89c4\u6a21\u4e0b\u91cd\u590d\u8bad\u7ec3\u4e00\u4e2a\u7b49\u89c4\u6a21\u7684 Critic \u7f51\u7edc\u3002<\/p>\n\n\n\n<p>\uff083\uff09\u5956\u52b1\u4fe1\u53f7\u5206\u53d1\u3002\u5728\u6570\u5b66\u63a8\u7406\u7b49\u53ef\u9a8c\u8bc1\u573a\u666f\u4e0b\uff0c\u901a\u8fc7\u9898\u76ee\u7b54\u6848\u7684\u6b63\u8bef\u6216\u89e3\u9898\u6b65\u9aa4\u7684\u6b63\u786e\u4e0e\u5426\u6765\u7ed9\u51fa\u7a00\u758f\u4f46\u51c6\u786e\u7684\u5956\u52b1\u4fe1\u53f7\uff1b\u7fa4\u7ec4\u5185\u90e8\u7684\u6392\u540d\u65e2\u63d0\u4f9b\u4e86\u76f8\u5bf9\u5956\u52b1\uff0c\u4e5f\u907f\u514d\u4e86\u5956\u52b1\u6a21\u578b\u53ef\u80fd\u51fa\u73b0\u7684\u5956\u52b1\u9a87\u5165\uff08reward hacking\uff09\u95ee\u9898\u3002<\/p>\n\n\n\n<p>\u5f15\u5165 GRPO \u540e\uff0cDeepSeek Math \u5728\u6570\u5b66\u63a8\u7406\u9886\u57df\u7684\u8868\u73b0\u663e\u8457\u63d0\u5347\uff0c\u90e8\u5206\u7ed3\u679c\u63a5\u8fd1 GPT-4 \u5728\u540c\u7c7b\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\uff0c\u5e76\u540c\u65f6\u8d85\u8d8a\u4e86\u4f17\u591a\u5f00\u6e90\u6a21\u578b\u3002\u8fd9\u4e00\u6210\u7ee9\u4e0d\u4ec5\u4f50\u8bc1\u4e86\u5f3a\u5316\u5b66\u4e60\u540e\u8bad\u7ec3\u5728\u5927\u6a21\u578b\u63a8\u7406\u80fd\u529b\u63d0\u5347\u4e2d\u7684\u6709\u6548\u6027\uff0c\u4e5f\u4e3a\u540e\u7eed DeepSeek \u7cfb\u5217\uff08\u5982 DeepSeek-R1\uff09\u8fdb\u4e00\u6b65\u6269\u5c55\u5230\u66f4\u590d\u6742\u63a8\u7406\u4efb\u52a1\u5960\u5b9a\u4e86\u7b97\u6cd5\u57fa\u7840\u3002<\/p>\n\n\n\n<p>1.2.2 DeepSeek-R1-Zero\u7684\u6fc0\u8fdb\u5b9e\u9a8c\u53caR1\u7684\u5de5\u7a0b\u5316\u6539\u8fdb<\/p>\n\n\n\n<p>2025 \u5e74\u521d\u53d1\u5e03\u7684 DeepSeek-R1 \u6a21\u578b\u5728 DeepSeek-V3 \u7684\u57fa\u5ea7\u4e0a\uff0c\u901a\u8fc7\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\u4e13\u95e8\u9488\u5bf9\u957f\u601d\u7ef4\u94fe\u7684\u63a8\u7406\u4efb\u52a1\u8fdb\u884c\u4e86\u5f3a\u5316\uff0c\u4ece\u800c\u663e\u8457\u63d0\u5347\u4e86\u5728 STEM \uff08\u79d1\u5b66\u3001\u6280\u672f\u3001\u5de5\u7a0b\u3001\u6570\u5b66\uff09\u9886\u57df\u7684\u63a8\u7406\u51c6\u786e\u6027\u3002\u5728\u6df1\u5ea6\u77e5\u8bc6\u95ee\u7b54\u6216\u590d\u6742\u6570\u5b66\u63a8\u7406\u573a\u666f\u4e0b\uff0c\u901a\u5e38\u9700\u8981\u591a\u6b65\u7684\u903b\u8f91\u63a8\u65ad\u548c\u53cd\u601d\uff0c\u4f20\u7edf\u81ea\u56de\u5f52\u6a21\u578b\u6613\u51fa\u73b0\u4e2d\u9014\u9519\u8bef\u96be\u4ee5\u4fee\u6b63\u7684\u95ee\u9898\u3002DeepSeek-R1-Zero\u53caDeepSeek-R1\u7ee7\u627f\u4e86 DeepSeek Math \u4e2d\u63d0\u51fa\u7684 GRPO \u601d\u8def\uff0c\u7ed3\u5408\u9ad8\u8d28\u91cf\u7684\u89c4\u5219\u5316\u5956\u52b1\uff08\u5982\u51c6\u786e\u7387\u4e0e\u683c\u5f0f\u5956\u52b1\uff09\uff0c\u9f13\u52b1\u6a21\u578b\u5728\u8bd5\u9519\u4e0e\u63a2\u7d22\u4e2d\u81ea\u4e3b\u62c9\u957f\u201c\u601d\u8003\u94fe\u201d\uff0c\u5e76\u5177\u5907\u81ea\u6211\u68c0\u67e5\u4e0e\u4fee\u6b63\u80fd\u529b\u3002\u88683\u793a\u51fa\u8fd9\u4e24\u4e2a\u6a21\u578b\u7684\u5bf9\u6bd4\u7ed3\u679c\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><tbody><tr><td>\u7ef4\u5ea6<\/td><td>\u8bad\u7ec3\u8d77\u70b9<\/td><td>\u6570\u636e\u4f9d\u8d56<\/td><td>RL\u7b97\u6cd5<\/td><td>\u6838\u5fc3\u4f18\u52bf<\/td><\/tr><tr><td>DeepSeek-R1-Zero<\/td><td>\u76f4\u63a5\u57fa\u4e8eV3\u57fa\u7840\u6a21\u578b\u542f\u52a8RL<\/td><td>\u96f6SFT\u6570\u636e\uff0c\u4ec5\u4f9d\u8d56\u89c4\u5219\u5956\u52b1<\/td><td>GRPO\uff0c\u65e0\u9700Critic\u7f51\u7edc<\/td><td>\u81ea\u6211\u8fdb\u5316\u6f5c\u529b\u5f3a\uff0c\u63a8\u7406\u80fd\u529b\u63a5\u8fd1\u9876\u5c16\u6a21\u578b<\/td><\/tr><tr><td>DeepSeek-R1<\/td><td>\u51b7\u542f\u52a8\u6570\u636e\u5fae\u8c03V3\u540e\u542f\u52a8RL<\/td><td>\u5f15\u5165\u6570\u5343\u6761\u9ad8\u8d28\u91cfCoT\u6570\u636e\u4f18\u5316\u53ef\u8bfb\u6027<\/td><td>\u7ee7\u627fGRPO\uff0c\u589e\u52a0\u591a\u9636\u6bb5\u5956\u52b1\u51fd\u6570<\/td><td>\u53ef\u8bfb\u6027\u3001\u901a\u7528\u6027\u3001\u7a33\u5b9a\u6027\u663e\u8457\u63d0\u5347<\/td><\/tr><\/tbody><\/table><figcaption class=\"wp-element-caption\">\u8868 3 DeepSeek-R1-Zero \u4e0e DeepSeek-R1\u7684\u5bf9\u6bd4<\/figcaption><\/figure>\n\n\n\n<p><strong>\u7eafRL\u8bad\u7ec3\u8303\u5f0f<\/strong>&nbsp;DeepSeek-R1-Zero\u5b8c\u5168\u8df3\u8fc7\u4f20\u7edf\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\uff0c\u76f4\u63a5\u57fa\u4e8eDeepSeek-V3\u57fa\u7840\u6a21\u578b\u8fdb\u884c\u5927\u89c4\u6a21\u5f3a\u5316\u5b66\u4e60\uff0c\u4f9d\u8d56\u89c4\u5219\u5956\u52b1\u7cfb\u7edf\uff08\u5982\u51c6\u786e\u6027\u5956\u52b1\u548c\u683c\u5f0f\u5956\u52b1\uff09\u9a71\u52a8\u6a21\u578b\u8fdb\u5316\u3002\u8fd9\u79cd\u8bbe\u8ba1\u9a8c\u8bc1\u4e86\u201cRL is all you need\u201d\u7684\u53ef\u80fd\u6027\uff0c\u5728\u6570\u5b66\u63a8\u7406\u4efb\u52a1\uff08\u5982AIME 2024\uff09\u4e2d\uff0c\u5176pass@1\u5206\u6570\u4ece15.6%\u8dc3\u5347\u81f371.0%\uff0c\u63a5\u8fd1OpenAI-o1-0912\u6c34\u5e73\u3002<\/p>\n\n\n\n<p><strong>\u6d8c\u73b0\u7684\u81ea\u6211\u53cd\u601d\u80fd\u529b<\/strong>&nbsp;\u6a21\u578b\u5728\u8bad\u7ec3\u4e2d\u81ea\u53d1\u8868\u73b0\u51fa\u5ef6\u957f\u601d\u7ef4\u94fe\u3001\u91cd\u8bc4\u89e3\u9898\u601d\u8def\u3001\u4e3b\u52a8\u4fee\u6b63\u9519\u8bef\u7b49\u884c\u4e3a\uff0c\u88ab\u79f0\u4e3a\u201caha\u987f\u609f\u65f6\u523b\u201d\u3002\u4f8b\u5982\uff0c\u5728ARC-AGI-1\u6d4b\u8bd5\u4e2d\uff0cDeepSeek-R1-Zero\u5c55\u73b0\u51fa\u9ad8\u5ea6\u8fde\u8d2f\u7684\u53cd\u601d\u903b\u8f91\uff0c\u90e8\u5206\u4efb\u52a1\u5f97\u5206\u4e0eR1\u4ec5\u76f8\u5dee1.8%\u3002<\/p>\n\n\n\n<p>\u4e3a\u89e3\u51b3DeepSeek-R1-Zero\u7684\u53ef\u8bfb\u6027\u5dee\u548c\u8bed\u8a00\u6df7\u6742\u95ee\u9898\uff0cR1\u5f15\u5165\u51b7\u542f\u52a8\u9636\u6bb5\u2014\u2014\u901a\u8fc7\u5c11\u91cf\u4eba\u5de5\u4f18\u9009\u7684\u957f\u601d\u7ef4\u94fe\uff08CoT\uff09\u6570\u636e\u5fae\u8c03\u57fa\u7840\u6a21\u578b\uff0c\u4f5c\u4e3aRL\u8bad\u7ec3\u7684\u7a33\u5b9a\u8d77\u70b9\u3002\u540e\u7eed\u5206\u56db\u9636\u6bb5\u4f18\u5316\uff1a\u51b7\u542f\u52a8\u2192\u63a8\u7406\u5bfc\u5411RL\u2192\u901a\u7528\u80fd\u529b\u5f3a\u5316\u2192\u62d2\u7edd\u91c7\u6837\u5fae\u8c03\uff0c\u9010\u6b65\u63d0\u5347\u6a21\u578b\u7684\u591a\u4efb\u52a1\u9002\u5e94\u6027\u3002\u65b0\u589e\u5956\u52b1\u51fd\u6570\u60e9\u7f5a\u8bed\u8a00\u6df7\u5408\uff08\u5982\u4e2d\u82f1\u6587\u6df7\u6742\u8f93\u51fa\uff09\u7684\u8bed\u8a00\u4e00\u81f4\u6027\u673a\u5236\uff0c\u5e76\u901a\u8fc7\u683c\u5f0f\u89c4\u8303\u63d0\u5347\u7ed3\u679c\u53ef\u8bfb\u6027\u3002<\/p>\n\n\n\n<p>1.2.3 \u9ad8\u6548\u7684R1\u84b8\u998f\u5c0f\u6a21\u578b<\/p>\n\n\n\n<p>\u5c06\u6a21\u578b\u84b8\u998f\u4e3a\u66f4\u5c0f\u3001\u66f4\u9ad8\u6548\u7684\u7248\u672c\uff0c\u53ef\u4ee5\u66f4\u52a0\u4fbf\u6377\u5730\u5728PC\u548c\u624b\u673a\u7b49\u7ec8\u7aef\u8bbe\u5907\u4e0a\u90e8\u7f72\u3002\u8fd9\u4e9b\u84b8\u998f\u6a21\u578b\u9700\u8981\u8f83\u5c11\u7684\u8ba1\u7b97\u80fd\u529b\u548c\u5185\u5b58\uff0c\u9002\u5408\u5728\u8d44\u6e90\u53d7\u9650\u7684\u8bbe\u5907\u4e0a\u5b9e\u65f6\u5e94\u7528\u3002\u4ec5\u901a\u8fc7\u5bf9DeepSeek-R1\u7684\u8f93\u51fa\u8fdb\u884c\u84b8\u998f\uff0c\u5c31\u80fd\u8ba9DeepSeek-R1-Distill-Qwen-7B\u5728\u5404\u65b9\u9762\u7684\u8868\u73b0\u5168\u9762\u8d85\u8d8a\u8bf8\u5982GPT-4o-0513\u7b49\u975e\u63a8\u7406\u6a21\u578b\u3002DeepSeek-R1-14B\u5728\u6240\u6709\u8bc4\u4f30\u6307\u6807\u4e0a\u5747\u8d85\u8fc7\u4e86QwQ-32B-Preview\uff0c\u800cDeepSeek-R1-Distill-Qwen-32B\u548cDeepSeek-R1-Distill-70B\u5728\u5927\u591a\u6570\u57fa\u51c6\u4e0a\u90fd\u663e\u8457\u4f18\u4e8eo1-mini\u3002\u4e0a\u8ff0\u7ed3\u679c\u5145\u5206\u5c55\u793a\u4e86\u84b8\u998f\u65b9\u6cd5\u6240\u5177\u6709\u7684\u5f3a\u5927\u6f5c\u529b\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\" style=\"font-size:26px\">2 <strong>DeepSeek\u5bf9\u5927\u6a21\u578b\u53ef\u4fe1\u53ef\u63a7\u53ef\u7528\u7684\u542f\u793a<\/strong><\/h2>\n\n\n\n<p>DeepSeek\u7cfb\u5217\u6a21\u578b\u5728\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u9886\u57df\u7684\u4e0d\u65ad\u8fed\u4ee3\uff0c\u4e0d\u4ec5\u5728\u6027\u80fd\u548c\u6548\u7387\u4e0a\u53d6\u5f97\u4e86\u663e\u8457\u7a81\u7834\uff0c\u66f4\u4e3a\u5927\u6a21\u578b\u7cfb\u7edf\u5728\u201c\u53ef\u4fe1\u201d\u201c\u53ef\u63a7\u201d\u4e0e\u201c\u53ef\u7528\u201d\u8fd9\u4e09\u4e2a\u5173\u952e\u7ef4\u5ea6\u7684\u53d1\u5c55\u6307\u660e\u4e86\u524d\u884c\u65b9\u5411\u3002\u4e0b\u9762\u4ece\u8fd9\u4e09\u4e2a\u89d2\u5ea6\u5c55\u5f00\u6df1\u5165\u63a2\u8ba8\u3002<\/p>\n\n\n\n<p>2.1 \u53ef\u4fe1\u6027\uff1a\u900f\u660e\u63a8\u7406\u4e0e\u9ad8\u6548\u9a8c\u8bc1<\/p>\n\n\n\n<p>\u5728\u5927\u6a21\u578b\u7684\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u53ef\u4fe1\u6027\u59cb\u7ec8\u662f\u7814\u7a76\u8005\u548c\u5e94\u7528\u8005\u6700\u4e3a\u5173\u6ce8\u7684\u6838\u5fc3\u95ee\u9898\u3002DeepSeek\u7cfb\u5217\u6a21\u578b\u901a\u8fc7\u5f15\u5165\u900f\u660e\u7684\u601d\u7ef4\u94fe\u63a8\u7406\u673a\u5236\uff0c\u6210\u529f\u5b9e\u73b0\u4e86\u6a21\u578b\u51b3\u7b56\u8fc7\u7a0b\u7684\u53ef\u89e3\u91ca\u6027\u548c\u81ea\u6211\u53cd\u601d\u80fd\u529b\u3002\u5177\u4f53\u800c\u8a00\uff0c\u8be5\u7cfb\u5217\u6a21\u578b\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u4e0d\u4ec5\u5c55\u793a\u4e86\u903b\u8f91\u63a8\u65ad\u7684\u663e\u5f0f\u53ef\u9605\u8bfb\u7684\u5185\u90e8\u601d\u7ef4\u94fe\uff0c\u800c\u4e14\u7ecf\u7531GRPO\u8bad\u7ec3\u51fa\u6765\u7684\u6a21\u578b\u63a8\u7406\u80fd\u529b\u6781\u5f3a\uff0c\u6bcf\u4e00\u6b65\u63a8\u7406\u5177\u6709\u5f88\u9ad8\u7684\u5408\u7406\u6027\u548c\u51c6\u786e\u6027\u3002\u663e\u5f0f\u601d\u7ef4\u94fe\u4e0d\u4ec5\u6709\u52a9\u4e8e\u63d0\u5347\u6a21\u578b\u5728\u590d\u6742\u4efb\u52a1\uff08\u5982\u6570\u5b66\u63a8\u7406\u3001\u4ee3\u7801\u751f\u6210\u548c\u6df1\u5ea6\u77e5\u8bc6\u95ee\u7b54\uff09\u4e2d\u7684\u8868\u73b0\uff0c\u800c\u4e14\u4e3a\u6a21\u578b\u8f93\u51fa\u63d0\u4f9b\u4e86\u660e\u786e\u7684\u89e3\u91ca\u8def\u5f84\uff0c\u4fbf\u4e8e\u4e13\u5bb6\u5bf9\u6a21\u578b\u51b3\u7b56\u8fdb\u884c\u8ffd\u8e2a\u548c\u9a8c\u8bc1\u3002\u672a\u6765\uff0c\u5927\u6a21\u578b\u5728\u4fdd\u6301\u9ad8\u63a8\u7406\u80fd\u529b\u7684\u540c\u65f6\uff0c\u5e94\u7ee7\u7eed\u671d\u5411\u6807\u51c6\u5316\u7684\u900f\u660e\u63a8\u7406\u94fe\u6784\u5efa\u65b9\u5411\u53d1\u5c55\uff0c\u901a\u8fc7\u5efa\u7acb\u4e00\u6574\u5957\u53ef\u9a8c\u8bc1\u7684\u63a8\u7406\u6807\u51c6\u548c\u8bc4\u4ef7\u4f53\u7cfb\uff0c\u4f8b\u5982\u7c7b\u4f3clean4\u7684\u5f62\u5f0f\u5316\u63a8\u7406\u8bc1\u660e\u5e93\uff0c\u8fdb\u4e00\u6b65\u589e\u5f3a\u7cfb\u7edf\u7684\u53ef\u4fe1\u8f93\u51fa\uff0c\u6ee1\u8db3\u5b9e\u9645\u5e94\u7528\u4e2d\u5bf9\u5b89\u5168\u3001\u53ef\u63a7\u3001\u53ef\u8ffd\u6eaf\u51b3\u7b56\u8fc7\u7a0b\u7684\u9700\u6c42\u3002<\/p>\n\n\n\n<p>2.2 \u53ef\u63a7\u6027\uff1a\u5168\u6d41\u7a0b\u5f00\u6e90\u4e0e\u5b89\u5168\u90e8\u7f72<\/p>\n\n\n\n<p>\u53ef\u63a7\u6027\u662f\u786e\u4fdd\u5927\u6a21\u578b\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u4e0d\u4f1a\u56e0\u9ed1\u7bb1\u6548\u5e94\u800c\u5f15\u53d1\u4e0d\u53ef\u9884\u671f\u98ce\u9669\u7684\u5173\u952e\u3002DeepSeek\u7cfb\u5217\u6a21\u578b\u5168\u7cfb\u5217\u5f00\u6e90\u7b56\u7565\u4e3a\u7814\u7a76\u8005\u548c\u5f00\u53d1\u8005\u63d0\u4f9b\u4e86\u6df1\u5165\u4e86\u89e3\u6a21\u578b\u5185\u90e8\u67b6\u6784\u3001\u8bad\u7ec3\u65b9\u6cd5\u53ca\u63a8\u7406\u7b56\u7565\u7684\u673a\u4f1a\uff0c\u6253\u7834\u4e86\u95ed\u6e90\u5927\u6a21\u578b\u4f8b\u5982ChatGPT\u7b49\u7684\u4e0d\u900f\u660e\u5c40\u9762\u3002\u6a21\u578b\u53c2\u6570\u7684\u516c\u5f00\u4f7f\u5f97\u7814\u7a76\u8005\u53ef\u6839\u636e\u81ea\u8eab\u9700\u6c42\u81ea\u884c\u4f18\u5316\u6a21\u578b\u9886\u57df\u53ef\u63a7\u6027\uff0c\u5185\u7f51\u90e8\u7f72\u7684\u5e94\u7528\u573a\u666f\u5219\u8fdb\u4e00\u6b65\u5f3a\u5316\u4e86\u6a21\u578b\u7684\u53ef\u63a7\u6027\u3002\u5728\u5c01\u95ed\u5b89\u5168\u7684\u73af\u5883\u4e2d\u8fd0\u884c\u5927\u6a21\u578b\uff0c\u53ef\u4ee5\u6709\u6548\u9632\u6b62\u5916\u90e8\u6076\u610f\u5229\u7528\uff0c\u540c\u65f6\u5728\u5185\u90e8\u8bbe\u5b9a\u7ec6\u81f4\u7684\u6743\u9650\u7ba1\u7406\u548c\u6570\u636e\u9694\u79bb\u673a\u5236\uff0c\u4e3a\u6a21\u578b\u7684\u5546\u4e1a\u5316\u548c\u5927\u89c4\u6a21\u90e8\u7f72\u63d0\u4f9b\u4e86\u4fdd\u969c\u3002\u672a\u6765\u7684\u7814\u7a76\u5e94\u5728\u7ee7\u7eed\u63a8\u52a8\u6a21\u578b\u5f00\u6e90\u900f\u660e\u7684\u540c\u65f6\uff0c\u6ce8\u91cd\u5efa\u7acb\u4e00\u5957\u5b8c\u5907\u7684\u5185\u90e8\u76d1\u7ba1\u548c\u5916\u90e8\u5ba1\u8ba1\u673a\u5236\uff0c\u786e\u4fdd\u5927\u6a21\u578b\u5728\u5404\u79cd\u5e94\u7528\u73af\u5883\u4e0b\u7684\u53ef\u63a7\u3001\u5b89\u5168\u4e0e\u7a33\u5b9a\u3002<\/p>\n\n\n\n<p>2.3 \u53ef\u7528\u6027\uff1a\u9ad8\u6027\u4ef7\u6bd4\u4e0e\u8f7b\u91cf\u5316\u90e8\u7f72<\/p>\n\n\n\n<p>\u5927\u6a21\u578b\u7684\u53ef\u7528\u6027\u4e0d\u4ec5\u4f53\u73b0\u5728\u5176\u5f3a\u5927\u7684\u63a8\u7406\u80fd\u529b\u4e0a\uff0c\u66f4\u5728\u4e8e\u5982\u4f55\u4ee5\u66f4\u4f4e\u7684\u8d44\u6e90\u6d88\u8017\u548c\u66f4\u9ad8\u7684\u6027\u4ef7\u6bd4\uff0c\u5b9e\u73b0\u5e7f\u6cdb\u90e8\u7f72\u548c\u666e\u53ca\u5e94\u7528\u3002DeepSeek\u7cfb\u5217\u6a21\u578b\u901a\u8fc7\u5f15\u5165\u6a21\u578b\u84b8\u998f\u6280\u672f\uff0c\u5c06\u539f\u672c\u53c2\u6570\u5e9e\u5927\u7684\u6a21\u578b\u77e5\u8bc6\u6d53\u7f29\u4e3a\u8f7b\u91cf\u5316\u7248\u672c\uff0c\u5982DeepSeek-R1\u7684\u84b8\u998f\u5c0f\u6a21\u578b\uff08\u4ece1.5B\u523070B\uff09\uff0c\u5373\u4fbf\u5728\u53c2\u6570\u89c4\u6a21\u5927\u5e45\u7f29\u51cf\u7684\u60c5\u51b5\u4e0b\uff0c\u4f9d\u7136\u4fdd\u6301\u4e86\u63a5\u8fd1\u751a\u81f3\u8d85\u8d8a\u5927\u578b\u6a21\u578b\uff08\u4f8b\u5982ChatGPT o1 mini\uff09\u7684\u63a8\u7406\u6027\u80fd\u3002\u8fd9\u79cd\u6280\u672f\u7a81\u7834\u5927\u5e45\u964d\u4f4e\u4e86\u786c\u4ef6\u95e8\u69db\uff0c\u4f7f\u5f97\u9ad8\u6027\u80fd\u5927\u6a21\u578b\u80fd\u591f\u5728PC\u3001\u624b\u673a\u7b49\u8d44\u6e90\u53d7\u9650\u7684\u7ec8\u7aef\u8bbe\u5907\u4e0a\u5b9e\u65f6\u8fd0\u884c\uff0c\u4e3a\u5404\u7c7b\u5e94\u7528\u573a\u666f\uff08\u4f8b\u5982\u5b9e\u65f6\u7ffb\u8bd1\u3001\u667a\u80fd\u52a9\u624b\u3001\u8fb9\u7f18\u8ba1\u7b97\u7b49\uff09\u63d0\u4f9b\u4e86\u5f3a\u5927\u7684\u6280\u672f\u652f\u6301\u3002\u540c\u65f6\uff0c\u84b8\u998f\u6280\u672f\u4e0d\u4ec5\u63d0\u5347\u4e86\u6a21\u578b\u7684\u90e8\u7f72\u7075\u6d3b\u6027\uff0c\u4e5f\u4e3a\u4f01\u4e1a\u964d\u4f4e\u4e86\u7814\u53d1\u4e0e\u7ef4\u62a4\u6210\u672c\uff0c\u5b9e\u73b0\u4e86\u7ecf\u6d4e\u6027\u4e0e\u5b9e\u7528\u6027\u4e4b\u95f4\u7684\u6700\u4f73\u5e73\u8861\u3002\u5c55\u671b\u672a\u6765\uff0c\u5927\u6a21\u578b\u7684\u53d1\u5c55\u9700\u8981\u66f4\u591a\u805a\u7126\u4e8e\u9ad8\u6027\u4ef7\u6bd4\u7684\u8f7b\u91cf\u5316\u8bbe\u8ba1\uff0c\u4f7f\u5f97\u6280\u672f\u521b\u65b0\u80fd\u591f\u8fc5\u901f\u8f6c\u5316\u4e3a\u5b9e\u9645\u5e94\u7528\uff0c\u63a8\u52a8\u4eba\u5de5\u667a\u80fd\u666e\u60e0\u5316\u7684\u8fdb\u7a0b\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\" style=\"font-size:26px\"><strong>3 ChatBIT\u5e94\u5bf9\u5b9e\u8df5<\/strong><\/h2>\n\n\n\n<p>ChatBIT\u660e\u7406\u5bf9\u8bdd\u5927\u6a21\u578b\u662f\u4e00\u6b3e\u9762\u5411\u519b\u4e8b\u60c5\u62a5\u9886\u57df\u7684\u591a\u8bed\u79cd\u53ef\u63a7\u8f7b\u91cf\u7ea7\u5927\u6a21\u578b\u7cfb\u7edf\uff0c\u7531\u5317\u4eac\u7406\u5de5\u5927\u5b66NLPIR\u5b9e\u9a8c\u5ba4\u81ea\u4e3b\u7814\u53d1\uff0c\u652f\u6301\u5305\u62ec\u519b\u4e8b\u60c5\u62a5\u95ee\u7b54\u3001\u60c5\u62a5\u6574\u5408\u5206\u6790\u3001\u8f85\u52a9\u51b3\u7b56\u7b49\u9488\u5bf9\u519b\u4e8b\u60c5\u62a5\u7684\u529f\u80fd\uff0c\u652f\u6301\u85cf\u8bed\u3001\u8499\u8bed\u7b49\u5c0f\u8bed\u79cd\uff0c\u540c\u65f6\u9762\u5411\u5927\u4f17\u7528\u6237\u63d0\u4f9b\u5305\u62ec\u4ee3\u7801\u751f\u6210\u3001\u667a\u80fd\u6587\u6863\u5199\u4f5c\u3001\u667a\u80fd\u7ffb\u8bd1\u5728\u5185\u7684\u901a\u7528\u529f\u80fd\u3002\u672c\u6587\u5728ChatBIT\u5927\u6a21\u578b\u5e73\u53f0\u4e0a\u8fdb\u884c\u4e86\u5b9e\u8df5\u9a8c\u8bc1\u3002<\/p>\n\n\n\n<p>\u5728\u5927\u8bed\u8a00\u6a21\u578b\u7684\u63a8\u7406\u4e0e\u4f18\u5316\u8fc7\u7a0b\u4e2d\uff0c\u5956\u52b1\u5efa\u6a21\u59cb\u7ec8\u626e\u6f14\u7740\u5173\u952e\u89d2\u8272\u3002\u7136\u800c\uff0c\u4f20\u7edf\u57fa\u4e8e\u89c4\u5219\u6216\u9884\u8bbe\u5956\u52b1\u6a21\u677f\u7684\u5956\u52b1\u5efa\u6a21\u65b9\u6cd5\u5b58\u5728\u9002\u7528\u6027\u6709\u9650\u53careward hacking\u95ee\u9898\uff0c\u8fd9\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u5236\u7ea6\u4e86\u6a21\u578b\u63a8\u7406\u80fd\u529b\u7684\u8fdb\u4e00\u6b65\u63d0\u5347\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u56f0\u5883\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u81ea\u6211\u4e00\u81f4\u6027\u8bc4\u5224\u7684\u5927\u8bed\u8a00\u6a21\u578b\u5956\u52b1\u673a\u5236\uff1a\u5185\u5316\u601d\u7ef4\u65b9\u6cd5\u4e0e\u9c81\u68d2\u63a8\u7406\u7684\u81ea\u6211\u4f18\u5316\uff0c\u901a\u8fc7\u5185\u5316\u5178\u578b\u601d\u7ef4\u65b9\u6cd5\u5e76\u5b9e\u73b0\u9c81\u68d2\u63a8\u7406\u7684\u81ea\u6211\u4f18\u5316\uff0c\u4e3a\u5927\u89c4\u6a21\u63a8\u7406\u6a21\u578b\u63d0\u4f9b\u4e86\u4e00\u6761\u5168\u65b0\u7684\u4f18\u5316\u8def\u5f84\u3002<\/p>\n\n\n\n<p>\u5728\u5f53\u524d\u7684LRM\uff08large reasoning models\uff09\u5e94\u7528\u4e2d\uff0c\u8bf8\u5982DeepSeek DeepSeek-R1-Zero\u6240\u4f7f\u7528\u7684\u89c4\u5219\u5956\u52b1\u673a\u5236\u5f80\u5f80\u663e\u5f97\u8fc7\u4e8e\u6b7b\u677f\uff0c\u4e3b\u8981\u5b58\u5728\u4e09\u4e2a\u5c40\u9650\u6027\u3002<strong>\u2460 \u683c\u5f0f\u4e0e\u6b63\u786e\u6027\u7684\u8131\u8282\u3002<\/strong>\u4ec5\u4ec5\u7b26\u5408\u9884\u5b9a\u683c\u5f0f\u7684\u56de\u7b54\u5e76\u4e0d\u5fc5\u7136\u4ee3\u8868\u63a8\u7406\u8fc7\u7a0b\u7684\u6b63\u786e\u6027\uff0c\u8fc7\u5206\u4f9d\u8d56\u683c\u5f0f\u5316\u6807\u51c6\u53ef\u80fd\u5bfc\u81f4\u6a21\u578b\u5ffd\u7565\u6838\u5fc3\u903b\u8f91\u3002<strong>\u2461 \u5956\u52b1\u6a21\u677f\u9002\u5e94\u6027\u4e0d\u8db3\u3002<\/strong>\u7531\u4eba\u5de5\u8bbe\u5b9a\u7684\u5956\u52b1\u6a21\u677f\u5f80\u5f80\u65e0\u6cd5\u5168\u9762\u6db5\u76d6\u6a21\u578b\u5728\u591a\u6837\u5316\u63a8\u7406\u4efb\u52a1\u4e2d\u6240\u9700\u8981\u7684\u7075\u6d3b\u601d\u8003\u4e0e\u591a\u5143\u5316\u8f93\u51fa\u3002<strong>\u2462 \u683c\u5f0f\u8986\u76d6\u4e0d\u5168\u9762\u3002<\/strong>\u9884\u8bbe\u6a21\u677f\u7684\u521a\u6027\u8bbe\u8ba1\u96be\u4ee5\u5e94\u5bf9\u590d\u6742\u95ee\u9898\u60c5\u5883\u4e2d\u591a\u53d8\u7684\u63a8\u7406\u9700\u6c42\uff0c\u5bb9\u6613\u51fa\u73b0reward hacking\u73b0\u8c61\uff0c\u5373\u6a21\u578b\u4e3a\u4e86\u8fce\u5408\u5956\u52b1\u6807\u51c6\u800c\u504f\u79bb\u771f\u6b63\u5408\u7406\u7684\u63a8\u7406\u8fc7\u7a0b\u3002\u56e0\u6b64\u6709\u5fc5\u8981\u6784\u5efa\u4e00\u79cd\u65b0\u7684\u8bc4\u5224\u6807\u51c6\uff0c\u8be5\u6807\u51c6\u65e2\u80fd\u5168\u9762\u8986\u76d6\u6a21\u578b\u7684\u601d\u7ef4\u7a7a\u95f4\uff0c\u53c8\u80fd\u51c6\u786e\u8868\u5f81\u6b63\u786e\u7684\u601d\u7ef4\u65b9\u6cd5\uff0c\u540c\u65f6\u907f\u514d\u5bf9\u6a21\u578b\u63a8\u7406\u8fc7\u7a0b\u8fdb\u884c\u786c\u6027\u7ea6\u675f\uff0c\u4ece\u800c\u5b9e\u73b0\u66f4\u4e3a\u7a33\u5065\u548c\u7075\u6d3b\u7684\u5956\u52b1\u53cd\u9988\u3002<\/p>\n\n\n\n<p>\u4e3a\u5e94\u5bf9\u4e0a\u8ff0\u6311\u6218\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e\u81ea\u6211\u4e00\u81f4\u6027\u8bc4\u5224\u7684\u5927\u8bed\u8a00\u6a21\u578b\u5956\u52b1\u673a\u5236\uff0c\u5176\u6838\u5fc3\u601d\u60f3\u5728\u4e8e\u5185\u5316\u5178\u578b\u601d\u7ef4\u65b9\u6cd5\uff0c\u5e76\u5229\u7528\u5927\u6a21\u578b\u672c\u8eab\u7684\u8bc4\u5224\u80fd\u529b\u5bf9\u751f\u6210\u8f93\u51fa\u8fdb\u884c\u81ea\u6211\u68c0\u6d4b\u548c\u4f18\u5316\u3002\u5177\u4f53\u800c\u8a00\u6211\u4eec\u7684\u65b9\u6cd5\u5305\u542b\uff1a<strong>\u2460 \u57fa\u7840\u601d\u7ef4\u65b9\u6cd5\u4f5c\u4e3a\u8bc4\u5224\u6807\u51c6\u3002<\/strong>\u6211\u4eec\u9009\u53d6\u4e86\u8bf8\u5982backtracking\uff08\u56de\u6eaf\uff09\u3001abandoning a reasoning chain\uff08\u653e\u5f03\u65e0\u6548\u63a8\u7406\u94fe\uff09\u4ee5\u53caself-verification\uff08\u81ea\u6211\u9a8c\u8bc1\uff09\u7b49\u5178\u578b\u601d\u7ef4\u65b9\u6cd5\uff0c\u4f5c\u4e3a\u5956\u52b1\u6a21\u578b\u7684\u6838\u5fc3\u8bc4\u5224\u6807\u51c6\u3002\u901a\u8fc7\u5f15\u5165\u8fd9\u4e9b\u4eba\u7c7b\u5e38\u7528\u7684\u8ba4\u77e5\u7b56\u7565\uff0c\u6a21\u578b\u80fd\u591f\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u5b66\u4e60\u5e76\u5185\u5316\u5408\u7406\u7684\u601d\u8003\u8def\u5f84\uff0c\u800c\u4e0d\u518d\u4f9d\u8d56\u4e8e\u4eba\u4e3a\u8bbe\u8ba1\u7684\u50f5\u786c\u5956\u52b1\u89c4\u5219\u3002<strong>\u2461 \u7ec4\u5185\u4e00\u81f4\u6027\u7ea6\u675f\u3002<\/strong>\u4e3a\u786e\u4fdd\u6a21\u578b\u5728\u8f93\u51fa\u8fc7\u7a0b\u4e2d\u4fdd\u6301\u81ea\u6211\u4e00\u81f4\u6027\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u7ec4\u5185\u4e00\u81f4\u6027\u7ea6\u675f\u673a\u5236\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5bf9\u4e8e\u540c\u4e00\u95ee\u9898\uff0c\u6a21\u578b\u5728\u591a\u6b21\u751f\u6210\u8f93\u51fa\u65f6\uff0c\u5176\u8bc4\u4ef7\u5f97\u5206\u5e94\u5c3d\u53ef\u80fd\u4fdd\u6301\u4e00\u81f4\u3002\u5229\u7528\u5927\u6a21\u578b\u4f5c\u4e3a\u8bc4\u5224\u8005\uff0c\u5bf9\u5404\u6b21\u8f93\u51fa\u8fdb\u884c\u6253\u5206\u548c\u6bd4\u5bf9\uff0c\u4ece\u800c\u786e\u4fdd\u6a21\u578b\u5728\u81ea\u6211\u8bc4\u5224\u8fc7\u7a0b\u4e2d\u5f62\u6210\u7a33\u5b9a\u800c\u9c81\u68d2\u7684\u63a8\u7406\u94fe\u3002\u8fd9\u79cd\u673a\u5236\u4e0d\u4ec5\u6709\u52a9\u4e8e\u8fc7\u6ee4\u51fa\u771f\u6b63\u5177\u5907\u9ad8\u8d28\u91cf\u63a8\u7406\u903b\u8f91\u7684\u7b54\u6848\uff0c\u4e5f\u4e3a\u6a21\u578b\u5728\u81ea\u6211\u4f18\u5316\u4e2d\u63d0\u4f9b\u4e86\u53ef\u9760\u7684\u53cd\u9988\u4fe1\u53f7\u3002<strong>\u2462 \u5f15\u5165\u683c\u5f0f\u5316\u7ea6\u675f\u4e0e\u5956\u52b1\u6743\u91cd\u8bbe\u8ba1\u3002<\/strong>\u7ed3\u5408DeepSeek-R1\u4e2dGRPO\u63d0\u51fa\u7684\u683c\u5f0f\u5316\u7ea6\u675f\u624b\u6bb5\uff0c\u6211\u4eec\u5728\u5956\u52b1\u673a\u5236\u4e2d\u52a0\u5165\u4e86\u989d\u5916\u7684\u683c\u5f0f\u8981\u6c42\u3002\u901a\u8fc7\u5bf9\u8f93\u51fa\u8fdb\u884c\u683c\u5f0f\u5316\u68c0\u6d4b\uff0c\u786e\u4fdd\u751f\u6210\u5185\u5bb9\u5728\u7b26\u5408\u903b\u8f91\u7684\u540c\u65f6\uff0c\u5177\u5907\u89c4\u8303\u7684\u8868\u8fbe\u7ed3\u6784\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u6839\u636e\u95ee\u9898\u7684\u96be\u6613\u7a0b\u5ea6\u548c\u4efb\u52a1\u590d\u6742\u6027\uff0c\u4e3a\u5404\u9879\u5956\u52b1\u6307\u6807\u8bbe\u5b9a\u4e0d\u540c\u6743\u91cd\uff0c\u7cbe\u7ec6\u5316\u8c03\u63a7\u5956\u52b1\u5206\u914d\uff0c\u4f7f\u5f97\u6a21\u578b\u5728\u9762\u5bf9\u4e0d\u540c\u96be\u5ea6\u95ee\u9898\u65f6\u5747\u80fd\u83b7\u5f97\u5408\u7406\u800c\u6709\u6548\u7684\u4f18\u5316\u4fe1\u53f7\u3002<strong>\u2463 \u81ea\u6211\u4e00\u81f4\u6027\u8bc4\u5224\u6a21\u578b\u7684\u6784\u5efa\u3002<\/strong>\u6211\u4eec\u91c7\u7528\u5f53\u524d\u80fd\u529b\u8f83\u5f3a\u7684\u5927\u578b\u8bed\u8a00\u6a21\u578b\u6765\u5145\u5f53\u81ea\u6211\u4e00\u81f4\u6027\u8bc4\u5224\u8005\uff0c\u4f7f\u5f97\u6a21\u578b\u80fd\u591f\u81ea\u4e3b\u5bfb\u627e\u548c\u6821\u6b63\u8f93\u51fa\u4e2d\u7684\u95ee\u9898\u3002\u5177\u4f53\u505a\u6cd5\u662f\uff0c\u5728\u6bcf\u6b21\u751f\u6210\u56de\u7b54\u65f6\uff0c\u7531\u5927\u6a21\u578b\u5bf9\u8f93\u51fa\u5185\u5bb9\u8fdb\u884c\u81ea\u6211\u8bc4\u5224\uff0c\u5e76\u4f9d\u636e\u5176\u5185\u5728\u7684\u4e00\u81f4\u6027\u548c\u903b\u8f91\u6027\u7ed9\u4e88\u6253\u5206\u3002\u901a\u8fc7\u8fd9\u4e00\u8fc7\u7a0b\uff0c\u6a21\u578b\u4e0d\u4ec5\u83b7\u5f97\u4e86\u5373\u65f6\u7684\u53cd\u9988\u4fe1\u606f\uff0c\u8fd8\u5728\u53cd\u590d\u5f3a\u5316\u5b66\u4e60\u4e2d\u9010\u6b65\u4f18\u5316\u81ea\u8eab\u7684\u63a8\u7406\u8def\u5f84\uff0c\u63d0\u5347\u6574\u4f53\u63a8\u7406\u8d28\u91cf\u3002<\/p>\n\n\n\n<p>\u5728\u5b9e\u9645\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u91c7\u7528\u4e861.5B\u53c2\u6570\u7684ChatBIT\u6a21\u578b\uff0c\u5e76\u641c\u96c6ChatBIT\u3001Qwen\u3001Llama\u7b49\u6a21\u578b\u7684\u8f93\u51fa\u5171\u8ba1600K\u6761\u6570\u636e\u3002\u8be5\u6570\u636e\u96c6\u6db5\u76d6\u4e86\u4e30\u5bcc\u7684\u63a8\u7406\u4efb\u52a1\u573a\u666f\uff0c\u4f8b\u5982\u6570\u5b66\u7b49\uff0c\u901a\u8fc7\u5bf9\u6570\u636e\u4e2d\u7684\u591a\u6837\u5316\u95ee\u9898\u8fdb\u884c\u8bad\u7ec3\uff0c\u4f7f\u5f97\u6a21\u578b\u5728\u7ec4\u5185\u4e00\u81f4\u6027\u68c0\u6d4b\u548c\u5956\u52b1\u53cd\u9988\u4e2d\u9010\u6b65\u5b8c\u5584\u81ea\u8eab\u7684\u63a8\u7406\u4f53\u7cfb\uff0c\u6700\u7ec8\u5b9e\u73b0\u81ea\u6211\u4e00\u81f4\u6027\u8bc4\u5224\u4e0e\u5956\u52b1\u5efa\u6a21\u7684\u95ed\u73af\u4f18\u5316\uff08\u89c1\u56fe2\uff09\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"608\" src=\"http:\/\/www.nlpir.org\/wordpress\/wp-content\/uploads\/2025\/03\/2.webp\" alt=\"\" class=\"wp-image-9224\"\/><figcaption class=\"wp-element-caption\">\u56fe 2 \u81ea\u6211\u5956\u52b1\u4e00\u81f4\u6027\u5728ChatBIT1.5B\u7684\u5b9e\u9645\u8bad\u7ec3\u6548\u679c<\/figcaption><\/figure>\n\n\n\n<p>\u8be5\u5956\u52b1\u673a\u5236\u7684\u521b\u65b0\u4e3b\u8981\u4f53\u73b0\u5728\u56db\u4e2a\u65b9\u9762\u3002<strong>\u2460 \u81ea\u6211\u5bfb\u627e\u8bc4\u5224\u6807\u51c6\u3002<\/strong>\u4e0d\u540c\u4e8e\u4f20\u7edf\u4f9d\u8d56\u4eba\u4e3a\u8bbe\u5b9a\u7684\u5956\u52b1\u6a21\u677f\uff0c\u6211\u4eec\u8ba9\u6a21\u578b\u81ea\u6211\u53d1\u73b0\u548c\u5185\u5316\u5408\u7406\u7684\u8bc4\u5224\u6807\u51c6\uff0c\u4ece\u800c\u907f\u514d\u4e86\u5956\u52b1\u6a21\u677f\u50f5\u5316\u7684\u95ee\u9898\u3002<strong>\u2461 \u81ea\u6211\u4e00\u81f4\u6027\u5f3a\u5316\u5b66\u4e60\u3002<\/strong>\u901a\u8fc7\u5f15\u5165\u7ec4\u5185\u4e00\u81f4\u6027\u7ea6\u675f\uff0c\u5229\u7528\u5927\u6a21\u578b\u8fdb\u884c\u591a\u6b21\u81ea\u6211\u8bc4\u5224\uff0c\u6709\u6548\u7f13\u89e3\u4e86reward hacking\u73b0\u8c61\uff0c\u4f7f\u6a21\u578b\u5728\u8ffd\u6c42\u5956\u52b1\u6700\u4f18\u5316\u65f6\u59cb\u7ec8\u4fdd\u6301\u5408\u7406\u7684\u63a8\u7406\u94fe\u8054\u3002<strong>\u2462 \u5408\u683c\u5f0f\u5316\u4e0e\u96be\u5ea6\u6743\u91cd\u8bbe\u8ba1\u3002<\/strong>\u5728\u5956\u52b1\u673a\u5236\u4e2d\u7efc\u5408\u8003\u8651\u8f93\u51fa\u683c\u5f0f\u4e0e\u95ee\u9898\u96be\u5ea6\uff0c\u901a\u8fc7\u52a8\u6001\u8c03\u6574\u5956\u52b1\u6743\u91cd\uff0c\u5b9e\u73b0\u4e86\u591a\u7ef4\u5ea6\u7684\u7cbe\u7ec6\u5316\u63a7\u5236\uff0c\u4e3a\u590d\u6742\u63a8\u7406\u4efb\u52a1\u63d0\u4f9b\u4e86\u66f4\u4e3a\u7cbe\u51c6\u7684\u53cd\u9988\u4fe1\u53f7\u3002<strong>\u2463 \u81ea\u6211\u8fdb\u5316\u7684\u5956\u52b1\u6a21\u578b\u6784\u5efa\u3002<\/strong>\u5728\u81ea\u6211\u4e00\u81f4\u6027\u8bc4\u5224\u6a21\u578b\u57fa\u7840\u4e0a\uff0c\u8fdb\u4e00\u6b65\u8bad\u7ec3\u51fa\u66f4\u4e3a\u5f3a\u5927\u7684\u5956\u52b1\u6a21\u578b\uff0c\u4f7f\u5f97\u63a8\u7406\u6a21\u578b\u4e0e\u8bc4\u5224\u6a21\u578b\u76f8\u4e92\u4fc3\u8fdb\uff0c\u5f62\u6210\u534f\u540c\u8fdb\u5316\u7684\u95ed\u73af\u4f53\u7cfb\u3002<\/p>\n\n\n\n<p>\u672a\u6765\uff0c\u6211\u4eec\u8ba1\u5212\u8fdb\u4e00\u6b65\u878d\u5408\u4e24\u9636\u6bb5\u5b66\u4e60\u7b56\u7565\uff0c\u540c\u65f6\u8bad\u7ec3\u8bc4\u5224\u6a21\u578b\u4e0e\u63a8\u7406\u6a21\u578b\uff0c\u63a8\u52a8\u4e8c\u8005\u7684\u534f\u540c\u8fdb\u5316\u3002\u5177\u4f53\u800c\u8a00\uff0c\u53ef\u4ee5\u63a2\u7d22\u5728\u8054\u5408\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5982\u4f55\u66f4\u597d\u5730\u5e73\u8861\u81ea\u6211\u4e00\u81f4\u6027\u4e0e\u5916\u90e8\u683c\u5f0f\u7ea6\u675f\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u6a21\u578b\u5728\u591a\u4efb\u52a1\u3001\u591a\u573a\u666f\u4e0b\u7684\u81ea\u9002\u5e94\u80fd\u529b\u4e0e\u9c81\u68d2\u6027\u3002\u6b64\u5916\uff0c\u8fd8\u53ef\u8003\u8651\u5f15\u5165\u66f4\u591a\u5143\u5316\u7684\u8ba4\u77e5\u65b9\u6cd5\u4f5c\u4e3a\u8bc4\u5224\u4f9d\u636e\uff0c\u62d3\u5c55\u5956\u52b1\u673a\u5236\u7684\u8986\u76d6\u8303\u56f4\uff0c\u4f7f\u5176\u5728\u66f4\u52a0\u590d\u6742\u548c\u591a\u53d8\u7684\u63a8\u7406\u4efb\u52a1\u4e2d\u4fdd\u6301\u7a33\u5b9a\u9ad8\u6548\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading has-text-align-center\" style=\"font-size:26px\"><strong><strong>4 \u7ed3\u8bba\u4e0e\u672a\u6765\u5c55\u671b<\/strong><\/strong><\/h2>\n\n\n\n<p>\u672c\u6587\u6df1\u5165\u63a2\u8ba8\u4e86DeepSeek\u7cfb\u5217\u6a21\u578b\u4e0eChatBIT\u5956\u52b1\u673a\u5236\u5728\u63d0\u5347\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u7684\u53ef\u4fe1\u6027\u3001\u53ef\u63a7\u6027\u548c\u53ef\u7528\u6027\u65b9\u9762\u7684\u521b\u65b0\uff0c\u5e76\u901a\u8fc7\u5b9e\u9645\u6848\u4f8b\u5c55\u793a\u4e86\u63a8\u7406\u94fe\u900f\u660e\u5316\u3001\u7ec4\u5185\u4e00\u81f4\u6027\u7ea6\u675f\u53ca\u5956\u52b1\u81ea\u9002\u5e94\u7b49\u65b9\u9762\u7684\u8fdb\u5c55\u3002\u7136\u800c\uff0c\u5f53\u524d\u7814\u7a76\u4ecd\u9762\u4e34\u4e00\u4e9b\u6311\u6218\u3002\u9996\u5148\uff0c\u6a21\u578b\u5728\u751f\u6210\u63a8\u7406\u94fe\u65f6\u53ef\u80fd\u5bfc\u81f4\u8fc7\u5ea6\u601d\u8003\uff0c\u589e\u52a0\u5197\u4f59\u4fe1\u606f\u548c\u9519\u8bef\u7d2f\u79ef\uff0c\u672a\u6765\u9700\u63a2\u7d22\u7cbe\u70bc\u601d\u7ef4\u94fe\u7684\u673a\u5236\uff0c\u4ee5\u5728\u4fdd\u6301\u5fc5\u8981\u63a8\u7406\u900f\u660e\u6027\u7684\u540c\u65f6\u63d0\u9ad8\u6548\u7387\u3002\u5176\u6b21\uff0c\u57fa\u4e8e\u6570\u503c\u7684\u5956\u52b1\u673a\u5236\u5b58\u5728\u5c40\u9650\uff0c\u672a\u6765\u53ef\u8003\u8651\u57fa\u4e8e\u81ea\u7136\u8bed\u8a00\u7684\u5956\u52b1\u4f53\u7cfb\uff0c\u4ee5\u66f4\u597d\u5730\u6355\u6349\u63a8\u7406\u4e2d\u7684\u8bed\u4e49\u4e0e\u903b\u8f91\uff0c\u907f\u514dreward hacking\u73b0\u8c61\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u5927\u6a21\u578b\u5728\u591a\u6a21\u6001\u63a8\u7406\u65b9\u9762\u7684\u80fd\u529b\u4ecd\u9700\u52a0\u5f3a\uff0c\u672a\u6765\u7684\u7814\u7a76\u5e94\u81f4\u529b\u4e8e\u5b9e\u73b0\u8de8\u6a21\u6001\u4fe1\u606f\u878d\u5408\uff0c\u6269\u5c55\u6a21\u578b\u7684\u5e94\u7528\u8fb9\u754c\uff0c\u7279\u522b\u662f\u5728\u89c6\u89c9\u3001\u97f3\u9891\u7b49\u4fe1\u606f\u4ea4\u53c9\u7684\u590d\u6742\u573a\u666f\u4e2d\u3002\u6700\u540e\uff0c\u5f3a\u5316\u5b66\u4e60\u7684\u6548\u7387\u4e9f\u5f85\u63d0\u9ad8\uff0c\u5f00\u53d1\u66f4\u9ad8\u6548\u7684\u7b97\u6cd5\u548c\u53cd\u9988\u673a\u5236\uff0c\u5c06\u6709\u52a9\u4e8e\u63a8\u52a8\u5927\u6a21\u578b\u5728\u590d\u6742\u4efb\u52a1\u4e2d\u7684\u81ea\u6211\u4f18\u5316\u548c\u9ad8\u6548\u5e94\u7528\uff0c\u4e3a\u5927\u89c4\u6a21\u63a8\u7406\u4efb\u52a1\u7684\u843d\u5730\u5e94\u7528\u63d0\u4f9b\u66f4\u575a\u5b9e\u7684\u6280\u672f\u4fdd\u969c\u3002<\/p>\n\n\n\n<p>\uff08\u53c2\u8003\u6587\u732e\u7565\uff09<\/p>\n\n\n\n<div class=\"wp-block-columns is-layout-flex wp-container-core-columns-is-layout-1 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<figure class=\"wp-block-image size-full is-style-default\"><img loading=\"lazy\" decoding=\"async\" width=\"295\" height=\"413\" src=\"http:\/\/www.nlpir.org\/wordpress\/wp-content\/uploads\/2025\/03\/11-1.webp\" alt=\"\" class=\"wp-image-9226\"\/><\/figure>\n<\/div>\n\n\n\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<p class=\"has-text-align-center\"><strong>\u5d14\u6587\u8000<\/strong><\/p>\n\n\n\n<p>&nbsp;\u5317\u4eac\u7406\u5de5\u5927\u5b66\u8ba1\u7b97\u673a\u5b66\u9662\u535a\u58eb\u7814\u7a76\u751f\u3002\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u4e3aLLM\u3001AI For Science\u3002<\/p>\n<\/div>\n<\/div>\n\n\n\n<div class=\"wp-block-columns is-layout-flex wp-container-core-columns-is-layout-2 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"390\" height=\"567\" src=\"http:\/\/www.nlpir.org\/wordpress\/wp-content\/uploads\/2025\/03\/12.webp\" alt=\"\" class=\"wp-image-9227\"\/><\/figure>\n<\/div>\n\n\n\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<p class=\"has-text-align-center\"><strong>\u674e\u6797\u7ff0<\/strong><\/p>\n\n\n\n<p>\u5317\u4eac\u7406\u5de5\u5927\u5b66\u8ba1\u7b97\u673a\u5b66\u9662\u7855\u58eb\u7814\u7a76\u751f\u3002\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u4e3a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001AI For Math\u3002<\/p>\n<\/div>\n<\/div>\n\n\n\n<div class=\"wp-block-columns is-layout-flex wp-container-core-columns-is-layout-3 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"600\" height=\"800\" src=\"http:\/\/www.nlpir.org\/wordpress\/wp-content\/uploads\/2025\/03\/13.webp\" alt=\"\" class=\"wp-image-9228\"\/><\/figure>\n<\/div>\n\n\n\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<p class=\"has-text-align-center\"><strong>\u6797\u5b50\u8d85<\/strong><\/p>\n\n\n\n<p>\u5317\u4eac\u7406\u5de5\u5927\u5b66\u8ba1\u7b97\u673a\u5b66\u9662\u7855\u58eb\u7814\u7a76\u751f\u3002\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u4e3a\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u3002<\/p>\n<\/div>\n<\/div>\n\n\n\n<div class=\"wp-block-columns is-layout-flex wp-container-core-columns-is-layout-4 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"1512\" src=\"http:\/\/www.nlpir.org\/wordpress\/wp-content\/uploads\/2025\/03\/14.webp\" alt=\"\" class=\"wp-image-9229\"\/><\/figure>\n<\/div>\n\n\n\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<p class=\"has-text-align-center\"><strong>\u738b\u4e00\u8bfa<\/strong><\/p>\n\n\n\n<p>\u5317\u4eac\u7406\u5de5\u5927\u5b66\u8ba1\u7b97\u673a\u79d1\u5b66\u4e0e\u6280\u672f2023\u7ea7\u7814\u7a76\u751f\uff0c\u7814\u7a76\u65b9\u5411\u4e3a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3002<\/p>\n<\/div>\n<\/div>\n\n\n\n<div class=\"wp-block-columns is-layout-flex wp-container-core-columns-is-layout-5 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"540\" height=\"720\" src=\"http:\/\/www.nlpir.org\/wordpress\/wp-content\/uploads\/2025\/03\/15.webp\" alt=\"\" class=\"wp-image-9230\"\/><\/figure>\n<\/div>\n\n\n\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<p class=\"has-text-align-center\"><strong>\u66f9\u54f2\u4fca\uff08Zhejun Cao\uff09<\/strong><\/p>\n\n\n\n<p>\u5317\u4eac\u7406\u5de5\u5927\u5b66\u8ba1\u7b97\u673a\u6280\u672f\u4e13\u4e1aNLPIR\u5b9e\u9a8c\u5ba4\u535a\u58eb\u7814\u7a76\u751f\uff0c\u7814\u7a76\u65b9\u5411\u4e3a\u5927\u6a21\u578b\u4e0e\u4eba\u5de5\u667a\u80fd\u5b89\u5168\u3002<\/p>\n<\/div>\n<\/div>\n\n\n\n<div class=\"wp-block-columns is-layout-flex wp-container-core-columns-is-layout-6 wp-block-columns-is-layout-flex\">\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1080\" height=\"1638\" src=\"http:\/\/www.nlpir.org\/wordpress\/wp-content\/uploads\/2025\/03\/16.webp\" alt=\"\" class=\"wp-image-9231\"\/><\/figure>\n<\/div>\n\n\n\n<div class=\"wp-block-column is-layout-flow wp-block-column-is-layout-flow\">\n<p class=\"has-text-align-center\"><strong>\u5f20\u534e\u5e73<\/strong><\/p>\n\n\n\n<p>\u5317\u4eac\u7406\u5de5\u5927\u5b66\u8ba1\u7b97\u673a\u5b66\u9662\u7279\u8058\u6559\u6388\uff0c\u56fd\u5bb6\u7ea7\u9886\u519b\u4eba\u624d\u3002\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u4e3a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u3001\u5927\u6a21\u578b\u3001\u667a\u80fd\u60c5\u62a5\u5206\u6790\u4e0e\u6570\u636e\u6316\u6398\u3002<\/p>\n<\/div>\n<\/div>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"382\" src=\"http:\/\/www.nlpir.org\/wordpress\/wp-content\/uploads\/2025\/03\/3.webp\" alt=\"\" class=\"wp-image-9232\"\/><\/figure>\n","protected":false},"excerpt":{"rendered":"<p>\u2014\u2014\u539f\u6587\u4e8e2025\u5e743\u67081\u65e5\u53d1\u5e03\u5728\u300aIMLIP\u591a\u8bed\u79cd\u667a\u80fd\u4fe1\u606f\u5904\u7406\u300b\u5fae\u4fe1\u516c\u4f17\u53f7 \u2014 &hellip; <a href=\"http:\/\/www.nlpir.org\/wordpress\/2025\/03\/02\/%e3%80%90%e8%bd%ac%e8%bd%bd%e3%80%91deepseek%e5%90%af%e7%a4%ba%e5%8f%af%e4%bf%a1%e5%8f%af%e6%8e%a7%e5%8f%af%e7%94%a8%e7%9a%84%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%9c%aa%e6%9d%a5%e4%b9%8b%e8%b7%af\/\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[38],"tags":[],"_links":{"self":[{"href":"http:\/\/www.nlpir.org\/wordpress\/wp-json\/wp\/v2\/posts\/9221"}],"collection":[{"href":"http:\/\/www.nlpir.org\/wordpress\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.nlpir.org\/wordpress\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.nlpir.org\/wordpress\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.nlpir.org\/wordpress\/wp-json\/wp\/v2\/comments?post=9221"}],"version-history":[{"count":3,"href":"http:\/\/www.nlpir.org\/wordpress\/wp-json\/wp\/v2\/posts\/9221\/revisions"}],"predecessor-version":[{"id":9234,"href":"http:\/\/www.nlpir.org\/wordpress\/wp-json\/wp\/v2\/posts\/9221\/revisions\/9234"}],"wp:attachment":[{"href":"http:\/\/www.nlpir.org\/wordpress\/wp-json\/wp\/v2\/media?parent=9221"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.nlpir.org\/wordpress\/wp-json\/wp\/v2\/categories?post=9221"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.nlpir.org\/wordpress\/wp-json\/wp\/v2\/tags?post=9221"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}