{"id":35634,"date":"2024-12-28T17:00:19","date_gmt":"2024-12-28T09:00:19","guid":{"rendered":"https:\/\/17aitech.com\/?p=35634"},"modified":"2024-12-28T17:00:19","modified_gmt":"2024-12-28T09:00:19","slug":"%e5%8c%97%e5%a4%a7%e5%af%b9%e9%bd%90%e5%9b%a2%e9%98%9f%e7%8b%ac%e5%ae%b6%e8%a7%a3%e8%af%bb%ef%bc%9aopenai-o1%e5%bc%80%e5%90%af%e3%80%8c%e5%90%8e%e8%ae%ad%e7%bb%83%e3%80%8d%e6%97%b6%e4%bb%a3%e5%bc%ba","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=35634","title":{"rendered":"\u5317\u5927\u5bf9\u9f50\u56e2\u961f\u72ec\u5bb6\u89e3\u8bfb\uff1aOpenAI o1\u5f00\u542f\u300c\u540e\u8bad\u7ec3\u300d\u65f6\u4ee3\u5f3a\u5316\u5b66\u4e60\u65b0\u8303\u5f0f"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-09-15\" target=\"_blank\">\u5317\u5927\u5bf9\u9f50\u56e2\u961f\u72ec\u5bb6\u89e3\u8bfb\uff1aOpenAI o1\u5f00\u542f\u300c\u540e\u8bad\u7ec3\u300d\u65f6\u4ee3\u5f3a\u5316\u5b66\u4e60\u65b0\u8303\u5f0f<\/a><\/p>\n<p>OpenAI o1 \u5728\u6570\u5b66\u3001\u4ee3\u7801\u3001\u957f\u7a0b\u89c4\u5212\u7b49\u95ee\u9898\u53d6\u5f97\u663e\u8457\u7684\u8fdb\u6b65\u3002\u4e00\u90e8\u5206\u4e1a\u5185\u4eba\u58eb\u5206\u6790\u5176\u539f\u56e0\u662f\u7531\u4e8e\u6784\u5efa\u8db3\u591f\u5e9e\u5927\u7684\u903b\u8f91\u6570\u636e\u96c6 &lt;\u95ee\u9898\uff0c\u660e\u786e\u7684\u6b63\u786e\u7b54\u6848&gt; \uff0c\u518d\u52a0\u4e0a\u7c7b\u4f3c AlphaGo \u4e2d MCTS \u548c RL \u7684\u65b9\u6cd5\u76f4\u63a5\u641c\u7d22\uff0c\u53ea\u8981\u63d0\u4f9b\u8db3\u591f\u7684\u8ba1\u7b97\u91cf\u7528\u4e8e\u641c\u7d22\uff0c\u603b\u53ef\u4ee5\u641c\u5230\u6700\u540e\u7684\u6b63\u786e\u8def\u5f84\u3002\u7136\u800c\uff0c\u8fd9\u6837\u53ea\u662f\u5efa\u7acb\u8d77\u95ee\u9898\u548c\u7b54\u6848\u4e4b\u95f4\u7684\u66f4\u597d\u7684\u8054\u7cfb\uff0c\u5982\u4f55\u6cdb\u5316\u5230\u66f4\u590d\u6742\u7684\u95ee\u9898\u573a\u666f\uff0c\u6280\u672f\u8fdc\u4e0d\u6b62\u8fd9\u4e48\u7b80\u5355\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7188805adf42a7a1f79ddbdbd1851e48.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7188805adf42a7a1f79ddbdbd1851e48.png\"><\/a><\/p>\n<p>AlphaGo \u662f\u5f3a\u5316\u5b66\u4e60\u5728\u56f4\u68cb\u9886\u57df\u4e2d\u7684\u4e00\u5927\u6210\u529f\uff0c\u6210\u529f\u51fb\u8d25\u4e86\u5f53\u65f6\u7684\u4e16\u754c\u51a0\u519b\u3002\u65e9\u5728\u53bb\u5e74\uff0cDeepmind \u7684 CEO Demis Hassabis \u4fbf\u5f3a\u8c03\u7528 Tree Search \u6765\u589e\u5f3a\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002\u6839\u636e\u76f8\u5173\u4eba\u58eb\u63a8\u6d4b\uff0co1 \u7684\u6a21\u578b\u8bad\u7ec3\u6570\u636e\u622a\u6b62\u5230\u53bb\u5e74\u5341\u6708\u4efd\uff0c\u800c\u6709\u5173 Q * \u7684\u7206\u6599\u5927\u7ea6\u662f\u53bb\u5e74 11 \u6708\uff0c\u8fd9\u4f3c\u4e4e\u5c55\u793a o1 \u7684\u8bad\u7ec3\u4e2d\u4e5f\u7528\u5230\u4e86 TreeSearch \u7684\u6280\u5de7\u3002<\/p>\n<p>\u5b9e\u9645\u4e0a\uff0cOpenAI o1 \u8fd0\u7528\u7684\u6280\u672f\u5173\u952e\u8fd8\u662f\u5728\u4e8e\u5f3a\u5316\u5b66\u4e60\u7684<strong>\u641c\u7d22<\/strong>\u4e0e<strong>\u5b66\u4e60<\/strong>\u673a\u5236\uff0c\u57fa\u4e8e <strong>LLM \u5df2\u6709\u7684\u63a8\u7406\u80fd\u529b\uff0c\u8fed\u4ee3\u5f0f\u7684 Bootstrap \u6a21\u578b\u4ea7\u751f\u5408\u7406\u63a8\u7406\u8fc7\u7a0b\uff08Rationales) \u7684\u80fd\u529b\uff0c\u5e76\u5c06 Rationales \u878d\u5165\u5230\u8bad\u7ec3\u8fc7\u7a0b\u5185\uff0c\u8ba9\u6a21\u578b\u5b66\u4f1a\u8fdb\u884c\u63a8\u7406\uff0c\u800c\u540e\u518d\u8fd0\u7528\u8db3\u591f\u5f3a\u5927\u7684\u8ba1\u7b97\u91cf\u5b9e\u73b0 Post-Training \u9636\u6bb5\u7684 Scaling<\/strong>\u3002\u7c7b\u4f3c\u4e8e STaR [1] \u7684\u6269\u5c55\u7248\u672c\u3002<\/p>\n<p>\u6ce8\u610f\u8fd9\u91cc<strong>\u5408\u7406\u63a8\u7406\u8fc7\u7a0b<\/strong>\u5e76\u4e0d\u53ea\u662f\u5bf9\u95ee\u9898\u7684\u62c6\u89e3\u548c\u5206\u6b65\u4f5c\u7b54\uff0c\u8fd8\u6709\u5bf9\u4e8e\u4e3a\u4ec0\u4e48\u5982\u6b64\u4f5c\u7b54\u7684\u5206\u6790\u548c\u601d\u8003\u3002<\/p>\n<p>\u6280\u672f\u8981\u70b9\u6709\u4e09\uff1a<\/p>\n<ol>\n<li>\n<strong>\u540e\u8bad\u7ec3\u6269\u5c55\u5f8b Post-Training Scaling Laws \u5df2\u7ecf\u51fa\u73b0<\/strong>\uff0c\u5e76\u4e14 Post-Training Scaling Laws \u4e3a\u4e0a\u8ff0\u6280\u672f\u8def\u5f84\u7684\u6210\u529f\u63d0\u4f9b\u4e86\u6709\u529b\u652f\u6301\u3002<\/li>\n<li>\u6a21\u578b\u5b66\u4e60\u7684\u662f\u4ea7\u751f\u5408\u7406\u63a8\u7406\u7684\u8fc7\u7a0b\uff0cMCTS \u5728\u5176\u4e2d\u7684\u4f5c\u7528\u662f<strong>\u8bf1\u5bfc\u5408\u7406\u63a8\u7406\u8fc7\u7a0b\u7684\u4ea7\u751f<\/strong>\u6216<strong>\u6784\u5efa\u76f8\u5e94\u7684\u504f\u5e8f\u5bf9\u5f62\u6210\u7ec6\u7c92\u5ea6\u5956\u52b1\u4fe1\u53f7\uff0c\u800c\u975e\u76f4\u63a5\u641c\u7d22\u8fc7\u7a0b\u548c\u6700\u7ec8\u7b54\u6848<\/strong>\u3002<\/li>\n<li>\u6a21\u578b\u7684 BootStrap \u6709\u52a9\u4e8e\u6784\u5efa\u65b0\u7684\u9ad8\u8d28\u91cf\u6570\u636e\uff0c\u5e76\u4e14\u65b0\u7684 Rationales \u6570\u636e\u4fc3\u8fdb\u4e86\u6a21\u578b\u8fdb\u4e00\u6b65\u63d0\u5347\u80fd\u529b\u3002<\/li>\n<\/ol>\n<p><strong>\u4e00\u3001OpenAI o1 \u7684\u53d1\u5e03\u662f Post-Training Scaling Laws \u7684\u5f3a\u529b\u4f53\u73b0<\/strong><\/p>\n<p>\u5317\u4eac\u65f6\u95f4 9 \u6708 13 \u65e5\u5348\u591c\uff0cOpenAI \u53d1\u5e03 o1 \u7cfb\u5217\u6a21\u578b\uff0c\u65e8\u5728\u4e13\u95e8\u89e3\u51b3\u96be\u9898\u3002OpenAI o1 \u5728\u6570\u5b66\u3001 \u4ee3\u7801\u3001\u957f\u7a0b\u89c4\u5212\u7b49\u95ee\u9898\u4e0a\u53d6\u5f97\u4e86\u663e\u8457\u63d0\u5347\uff0c\u800c\u80cc\u540e\u7684\u6210\u529f\u6700\u91cd\u8981\u79bb\u4e0d\u5f00\u540e\u8bad\u7ec3\u9636\u6bb5 (Post-Training Stage) \u4e2d\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u548c\u63a8\u7406\u9636\u6bb5\u601d\u8003\u8ba1\u7b97\u91cf\u7684\u589e\u5927\u3002\u65b0\u7684\u6269\u5c55\u5f8b \u2014\u2014 \u540e\u8bad\u7ec3\u6269\u5c55\u5f8b\uff08Post-Training Scaling Laws\uff09 \u5df2\u7ecf\u51fa\u73b0\uff0c\u5e76\u53ef\u80fd\u5f15\u53d1\u793e\u533a\u5bf9\u4e8e\u7b97\u529b\u5206\u914d\u3001\u540e\u8bad\u7ec3\u80fd\u529b\u7684\u91cd\u65b0\u601d\u8003\u3002<\/p>\n<p><strong>\u6a21\u578b\u8868\u73b0\u6982\u89c8<\/strong><\/p>\n<p>\u6700\u65b0\u7684\u53d1\u5e03\u7684 OpenAI o1 \u5728\u6570\u5b66\u4ee3\u7801\u7b49\u590d\u6742\u63a8\u7406\u80fd\u529b\u4e0a\u53d6\u5f97\u5de8\u5927\u8fdb\u6b65\uff0c\u5728\u7ade\u4e89\u6027\u7f16\u7a0b\u95ee\u9898 \uff08Codeforces\uff09 \u4e2d\u6392\u540d\u7b2c 89 \u4e2a\u767e\u5206\u4f4d\uff0c\u5728\u7f8e\u56fd\u6570\u5b66\u5965\u6797\u5339\u514b\u7ade\u8d5b \uff08AIME\uff09 \u8d44\u683c\u8d5b\u4e2d\u8dfb\u8eab\u7f8e\u56fd\u524d 500 \u540d\u5b66\u751f\u4e4b\u5217\uff0c\u5728\u7269\u7406\u3001\u751f\u7269\u548c\u5316\u5b66\u95ee\u9898\u7684\u57fa\u51c6 \uff08GPQA\uff09 \u4e0a\u8d85\u8fc7\u4e86\u4eba\u7c7b\u535a\u58eb\u6c34\u5e73\u7684\u51c6\u786e\u6027\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-f585a6f04281be2d3cc0ddcbbfca900d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-f585a6f04281be2d3cc0ddcbbfca900d.png\"><\/a><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-8ba86129bffdc2b56b3a9398aef65304.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-8ba86129bffdc2b56b3a9398aef65304.png\"><\/a><\/p>\n<p>\u800c\u5e2e\u52a9 o1 \u53d6\u5f97\u5982\u6b64\u6027\u80fd\u98de\u8dc3\u7684\uff0c\u662f Post-Training \u9636\u6bb5 RL \u8ba1\u7b97\u91cf\u7684 Scaling \u548c\u6d4b\u8bd5\u63a8\u7406\u9636\u6bb5\u601d\u8003\u65f6\u95f4\u7684 Scaling\u3002<\/p>\n<p><strong>\u4e0d\u8fc7\uff0c\u5982\u679c\u4ed4\u7ec6\u89c2\u5bdf\uff0cOpenAI o1 \u5728\u4e00\u4e9b\u5e38\u89c4\u4efb\u52a1\u5982\u82f1\u8bed\u8003\u8bd5\u548c\u8bed\u8a00\u80fd\u529b\u6d4b\u8bd5\u4e0a\u5e76\u6ca1\u6709\u663e\u8457\u63d0\u5347 \u2014\u2014 \u63a8\u7406\u80fd\u529b\u548c\u5f3a\u6307\u4ee4\u8ddf\u968f\u80fd\u529b\u7684\u63d0\u5347\u4f3c\u4e4e\u5448\u73b0\u4e86\u5206\u79bb\uff0c\u8fd9\u4e2a\u89c2\u5bdf\u548c\u601d\u8003\uff0c\u6211\u4eec\u653e\u5230\u6700\u540e\u7684\u5206\u6790<\/strong>\u3002<\/p>\n<p><strong>\u540e\u8bad\u7ec3\u6269\u5c55\u5f8b Post-Training Scaling Law<\/strong><\/p>\n<p>\u968f\u7740\u6a21\u578b\u5c3a\u5bf8\u9010\u6e10\u589e\u5927\uff0c\u9884\u8bad\u7ec3\u9636\u6bb5\u53c2\u6570 Scaling Up \u5e26\u6765\u7684\u8fb9\u9645\u6536\u76ca\u5f00\u59cb\u9012\u51cf\uff0c\u5982\u679c\u60f3\u8981\u6df1\u5ea6\u63d0\u5347\u6a21\u578b\u63a8\u7406\u80fd\u529b\u548c\u957f\u7a0b\u95ee\u9898\u80fd\u529b\uff0c\u57fa\u4e8e\u5f3a\u5316\u5b66\u4e60\u7684 Post-Training \u5c06\u4f1a\u6210\u4e3a\u4e0b\u4e00\u4e2a\u7a81\u7834\u70b9\u3002\u65e9\u5728 2018 \u5e74 Ilya \u5728 MIT \u7684\u5ba2\u5ea7\u8bb2\u5ea7\u4e0a\uff0c\u4ed6\u4fbf\u5206\u4eab\u8fc7\u81ea\u5df1\u5bf9\u4e8e\u901a\u8fc7 RL \u548c Self-play \u8d70\u5411 AGI \u7684\u4fe1\u5fc3\u3002<\/p>\n<p>OpenAI \u63a2\u7d22 Parameter Scaling Law \u4e4b\u5916\u7684 Scaling Laws \u4e5f\u5e76\u975e\u7a7a\u7a74\u6765\u98ce\u3002<\/p>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"137\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\n<p>When generating a solution, autoregressive models have no mechanism to correct their own errors. Solutions that veer off-course quickly become unrecoverable.\u00a0<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"181\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\n<p>If we rely purely on generative methods and extrapolate from current trends, we will require an exorbitant parameter count to achieve even moderate performance on distributions as challenging as the MATH dataset.\u00a0<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"77\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\n<p>This evidence strongly motivates the search for methods with more favorable scaling laws.\u00a0<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n<p>\u5728 2021 \u5e74\uff0c\u4ed6\u4eec\u4fbf\u5728 Training Verifiers to Solve Math Word Problems [6] \u4e2d\u63d0\u5230\uff0c<strong>\u81ea\u56de\u5f52\u6a21\u578b\u5728\u6570\u5b66\u63a8\u7406\u95ee\u9898\u4e0a\u5f88\u96be\u8fdb\u6b65\u7684\u4e00\u70b9\u5728\u4e8e\u6ca1\u6709\u529e\u6cd5\u8fdb\u884c\u56de\u7b54\u7684\u81ea\u4e3b\u4fee\u6b63\uff0c\u5982\u679c\u4ec5\u662f\u4f9d\u9760\u751f\u6210\u5f0f\u65b9\u6cd5\u548c\u6269\u5927\u53c2\u6570\u89c4\u6a21\uff0c\u90a3\u4e48\u5728\u6570\u5b66\u63a8\u7406\u4efb\u52a1\u4e0a\u5e26\u6765\u7684\u6536\u76ca\u4e0d\u4f1a\u592a\u5927\u3002\u6240\u4ee5\u9700\u8981\u5bfb\u627e\u989d\u5916\u7684 Scaling Laws<\/strong>\u3002<\/p>\n<p>\u73b0\u5728\u770b\u6765\uff0cRL \u5e26\u6765\u4e86 LLM \u8bad\u7ec3\u7684\u8303\u5f0f\u8f6c\u53d8\uff0c\u4e5f\u5e26\u6765\u4e86\u65b0\u7684 Scaling Laws\uff0c\u5373 Post-Training Scaling Laws\u3002<\/p>\n<p>\u5728 Post-Training \u00a0Scaling Laws \u4e0b\uff0c\u8bad\u7ec3\u9636\u6bb5\u7684\u8ba1\u7b97\u91cf\u4e0d\u518d\u53ea\u662f\u548c\u53c2\u6570\u91cf\u7684\u4e0a\u5347\u6709\u5173\uff0c\u540c\u65f6\u4e5f\u4f1a\u5305\u542b RL \u63a2\u7d22\u65f6 LLM Inference \u7684\u8ba1\u7b97\u91cf\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u6d4b\u8bd5\u9636\u6bb5\u6a21\u578b\u63a8\u7406\u548c\u53cd\u601d\u7684\u8ba1\u7b97\u91cf\u4e5f\u4f1a\u5f71\u54cd\u6a21\u578b\u6700\u7ec8\u7684\u8868\u73b0\u3002\u5728 DeepMind \u6700\u8fd1\u7684\u6587\u7ae0 [5] \u4e2d\uff0c\u4e5f\u8ba8\u8bba\u4e86\u8fd9\u79cd\u8303\u5f0f\u7684\u8f6c\u53d8\u3002<\/p>\n<p><strong>Post-train \u867d\u7136\u53c2\u6570\u6ca1\u53d8\uff0c\u4f46\u662f\u5728\u8bad\u7ec3\u7b97\u529b\u4e0a\u4ecd\u7136\u4f1a\u500d\u6570\u589e\u957f\uff1b\u63a8\u7406\u4e0a\u4e5f\u4f1a\u968f\u7740\u6a21\u578b \u201c\u601d\u8003\u80fd\u529b\u63d0\u9ad8\u201d\uff0c\u5355\u6b21\u7b97\u529b\u589e\u957f\u3002\u662f\u5426\u6709\u8db3\u591f\u7684\u7b97\u529b\u505a Post-Training \u4f3c\u4e4e\u5df2\u7ecf\u6210\u4e3a\u80fd\u4e0d\u80fd\u63d0\u5347\u63a8\u7406\u6027\u80fd\u7684\u5165\u573a\u5238<\/strong>\u3002<\/p>\n<p>OpenAI \u53d1\u73b0\u4e5f\u8bc1\u660e\u4e86\u8fd9\u4e00\u70b9\uff1a\u968f\u7740\u66f4\u591a\u7684\u5f3a\u5316\u5b66\u4e60\uff08\u8bad\u7ec3\u65f6\u8ba1\u7b97\uff09\u548c\u66f4\u591a\u7684\u601d\u8003\u65f6\u95f4\uff08\u6d4b\u8bd5\u65f6\u8ba1\u7b97\uff09, o1 \u7684\u6027\u80fd\u4e5f\u5728\u4e0d\u65ad\u63d0\u5347\uff0c\u5e76\u4e14 Post-Training Scaling Laws \u8fd8\u6ca1\u6709\u88ab\u5b8c\u5168\u63a2\u7d22\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7ffb1e0394e19a08cc284ae75158f13f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7ffb1e0394e19a08cc284ae75158f13f.png\"><\/a><\/p>\n<p>Sutton \u5728\u300aBitter Lesson\u300b\u4e2d\u5df2\u7ecf\u6307\u51fa\uff0c\u53ea\u6709\u4e24\u79cd\u6280\u672f\u53ef\u4ee5\u968f\u7740\u7b97\u529b\u589e\u957f\uff0c\u5b66\u4e60\u548c\u641c\u7d22\u3002\u6b63\u5982\u82f1\u4f1f\u8fbe\u79d1\u5b66\u5bb6 Jim Fan \u6240\u8bf4\uff0c\u4e5f\u8bb8\u6a21\u578b\u53c2\u6570\u5927\u90e8\u5206\u662f\u7528\u4e8e\u5b58\u50a8\u77e5\u8bc6\u548c\u8bb0\u5fc6\u3002<\/p>\n<p><\/p>\n<p>\u968f\u7740\u53c2\u6570\u6269\u5c55\u5f8b\u7684\u8fb9\u9645\u6548\u76ca\u9010\u6e10\u9012\u51cf\uff0c\u73b0\u5728\u662f\u65f6\u5019\u5c06\u66f4\u591a\u7684\u7b97\u529b\u8f6c\u5411 Post-Training \u9636\u6bb5\u548c\u63a8\u7406\u9636\u6bb5\u3002<\/p>\n<p><strong>\u4e8c\u3001OpenAI \u7684\u6210\u529f\uff0c<\/strong><strong>\u5173\u952e\u5728\u4e8e\u5408\u7406\u4f7f\u7528\u5f3a\u5316\u5b66\u4e60\u7684\u63a2\u7d22<\/strong><\/p>\n<p><strong>\u4ec5\u9760 MCTS \u662f\u8fdc\u4e0d\u8db3\u591f\u7684<\/strong><\/p>\n<p><strong>\u4ec5\u9760 MCTS \u65e0\u6cd5\u8ba9\u6a21\u578b\u5b66\u4f1a\u601d\u8003\u95ee\u9898\u7684\u5173\u8054\uff0c\u9690\u5f0f\u81ea\u52a8\u5316 CoT \u7684\u80cc\u540e\uff0c\u662f\u6a21\u578b\u771f\u6b63\u5b66\u4f1a\u4e86\u5408\u7406\u7684\u4e2d\u95f4\u63a8\u7406\u8fc7\u7a0b Rationales<\/strong>\u3002<\/p>\n<p>\u5f53\u4eba\u4eec\u5199\u4f5c\u6216\u8bf4\u8bdd\u65f6\uff0c\u5e38\u5e38\u4f1a\u505c\u4e0b\u6765\u601d\u8003\u3002\u7136\u800c\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u5728\u901a\u8fc7 Next Token Prediction \u751f\u6210\u56de\u7b54\u65f6\uff0c\u66f4\u50cf\u662f\u4e00\u79cd \u201c\u5feb\u601d\u8003\u201d \u8fc7\u7a0b\u3002\u7531\u4e8e\u7f3a\u4e4f\u8be6\u7ec6\u7684\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4\uff0c\u6a21\u578b\u4e00\u5f00\u59cb\u53ef\u80fd\u4f1a\u72af\u9519\uff0c\u800c\u8fd9\u4e9b\u9519\u8bef\u53ef\u80fd\u4f1a\u4f20\u64ad\uff0c\u6700\u7ec8\u5bfc\u81f4\u751f\u6210\u7684\u7b54\u6848\u4e5f\u662f\u9519\u8bef\u7684\u3002<\/p>\n<p>\u4e3a\u4e86\u4f18\u5316\u8fd9\u4e00\u8fc7\u7a0b\uff0c\u4ea7\u751f\u4e86\u4e00\u7cfb\u5217\u65b9\u6cd5\uff0c\u5176\u4e2d\u5305\u62ec\u5728 <strong>Token \u7ea7\u522b<\/strong>\u6216<strong>\u5b50\u53e5\u7ea7\u522b<\/strong>\u63d0\u4f9b\u5956\u52b1\u4fe1\u53f7\uff0c\u5e2e\u52a9\u6a21\u578b\u8c03\u6574\u751f\u6210\u7684\u56de\u7b54\u3002\u8fd9\u4e9b\u65b9\u6cd5\u5982<strong>\u8499\u7279\u5361\u6d1b\u6811\u641c\u7d22\uff08MCTS\uff09<\/strong>\uff0c\u5c06\u8f93\u51fa\u5efa\u6a21\u4e3a\u4e00\u7cfb\u5217\u8282\u70b9\uff0c\u8fd9\u4e9b\u8282\u70b9\u53ef\u4ee5\u662f Token \u7ea7\u522b\u6216\u53e5\u5b50\u7ea7\u522b\u3002\u4f8b\u5982\uff1a<\/p>\n<ul>\n<li>\n<strong>Token \u7ea7\u522b\u7684\u8282\u70b9<\/strong>\uff1a\u6bcf\u4e2a\u8282\u70b9\u5bf9\u5e94\u751f\u6210\u5e8f\u5217\u4e2d\u7684\u4e00\u4e2a Token\u3002\u901a\u8fc7 MCTS\uff0c\u6a21\u578b\u53ef\u4ee5\u63a2\u7d22\u4e0d\u540c\u7684 Token \u5e8f\u5217\uff0c\u6700\u7ec8\u751f\u6210\u66f4\u8fde\u8d2f\u7684\u54cd\u5e94\u3002<\/li>\n<li>\n<strong>\u53e5\u5b50\u7ea7\u522b\u7684\u8282\u70b9<\/strong>\uff1a\u5728\u590d\u6742\u63a8\u7406\u4efb\u52a1\u4e2d\uff0c\u6bcf\u4e2a\u8282\u70b9\u53ef\u4ee5\u4ee3\u8868\u4e00\u4e2a\u5b8c\u6574\u7684\u53e5\u5b50\u6216\u63a8\u7406\u6b65\u9aa4\uff0c\u5e2e\u52a9\u6a21\u578b\u66f4\u597d\u5730\u5904\u7406\u591a\u6b65\u63a8\u7406\u4efb\u52a1\u3002<\/li>\n<\/ul>\n<p>\u53e6\u4e00\u79cd\u65b9\u5f0f\u662f\u901a\u8fc7<strong>\u601d\u7ef4\u94fe\uff08Chain of Thought, CoT\uff09<\/strong>\u4f18\u5316\u6a21\u578b\u8f93\u51fa\u3002CoT <strong>\u901a\u8fc7\u5206\u6b65\u63a8\u7406\u7684\u65b9\u5f0f\uff0c\u8981\u6c42\u6a21\u578b\u5728\u751f\u6210\u6700\u7ec8\u7b54\u6848\u4e4b\u524d\uff0c\u5148\u751f\u6210\u4e00\u7cfb\u5217\u4e2d\u95f4\u63a8\u7406\u6b65\u9aa4<\/strong>\u3002\u8fd9\u79cd \u201c\u601d\u8003\u94fe\u201d \u7684\u751f\u6210\u8fc7\u7a0b\u6709\u52a9\u4e8e\u589e\u5f3a\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u5c24\u5176\u5728\u6570\u5b66\u548c\u4ee3\u7801\u751f\u6210\u7b49\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u8272\u3002<\/p>\n<p><strong>\u7136\u800c\uff0cCoT \u867d\u7136\u80fd\u591f\u751f\u6210\u4e2d\u95f4\u6b65\u9aa4\uff0c\u4f46\u5e76\u672a\u6559\u4f1a\u6a21\u578b\u5982\u4f55\u4ece\u5185\u90e8\u6df1\u5165\u601d\u8003\u95ee\u9898\u7684\u5173\u8054\u3002\u7279\u522b\u662f\u5bf9\u4e8e\u5c24\u5176\u590d\u6742\u4e14\u9700\u8981\u591a\u6b65\u63a8\u7406\u89c4\u5212\u7684\u4efb\u52a1\uff0c\u8fd9\u6837\u7684\u5408\u7406\u7684\u4e2d\u95f4 CoT \u63a8\u7406\u8fc7\u7a0b\uff08Rationales) \u66f4\u4e3a\u91cd\u8981<\/strong>\u3002<\/p>\n<p>\u7c7b\u4f3c\u7684\u601d\u8def\u5728 STaR [1] \u548c Quiet-STaR [7] \u4e2d\u6709\u6240\u4f53\u73b0\u3002<\/p>\n<p><strong>STaR \u7684\u6838\u5fc3\u601d\u8def\u662f\u5229\u7528 LLM \u5df2\u6709\u7684\u63a8\u7406\u80fd\u529b\uff0c\u8fed\u4ee3\u5f0f\u7684 Bootstrap \u6a21\u578b\u4ea7\u751f\u5408\u7406\u63a8\u7406\u8fc7\u7a0b\uff08Rationales) \u7684\u80fd\u529b\uff0c\u5e76\u5c06 Rationales \u878d\u5165\u5230\u8bad\u7ec3\u8fc7\u7a0b\u5185\uff0c\u8ba9\u6a21\u578b\u5b66\u4f1a\u8fdb\u884c\u63a8\u7406<\/strong>\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-256a395c3308104dd0f445dd1465c1ad.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-256a395c3308104dd0f445dd1465c1ad.png\"><\/a><\/p>\n<ol>\n<li>\n<strong>\u63a8\u7406<\/strong>\uff1a\u8d77\u59cb\u6570\u636e\u96c6\u4ec5\u6709 [Question, Answer] \uff0c\u9996\u5148\u5229\u7528\u4e00\u4e9b\u5e26\u6709\u63a8\u7406\u8fc7\u7a0b\u7684 Few-Shot Examples \u6765 Prompt \u6a21\u578b\u5bf9\u4e8e\u6570\u636e\u96c6\u4e2d\u7684\u95ee\u9898\u751f\u6210\u5bf9\u5e94\u7684\u63a8\u7406\u8fc7\u7a0b\u548c\u7b54\u6848\u3002<\/li>\n<li>\n<strong>\u8fc7\u6ee4<\/strong>\uff1a\u5982\u679c\u751f\u6210\u7684\u7b54\u6848\u6b63\u786e\uff0c\u5219\u5c06\u63a8\u7406\u8fc7\u7a0b\u52a0\u5165\u5230\u539f\u6709\u7684\u6570\u636e\u96c6\u4e2d\uff1b\u5982\u679c\u751f\u6210\u7684\u7b54\u6848\u9519\u8bef\uff0c\u5219\u5c1d\u8bd5\u5728\u7ed9\u51fa\u6b63\u786e\u7b54\u6848\u7684\u524d\u63d0\u4e0b\u518d\u6b21\u751f\u6210\u63a8\u7406\u8fc7\u7a0b\u3002\u5c06\u6700\u7ec8\u751f\u6210\u6b63\u786e\u7b54\u6848\u7684\u63a8\u7406\u6536\u96c6\uff0c\u6784\u5efa\u4e00\u4e2a\u6784\u5efa\u4e00\u4e2a\u5fae\u8c03\u6570\u636e\u96c6 [Question, Rationale, Answer ] \u8fdb\u884c\u5fae\u8c03\u3002<\/li>\n<li>\n<strong>\u8fed\u4ee3<\/strong>\uff1a\u91cd\u590d\u8fd9\u4e00\u8fc7\u7a0b\uff0c\u4e14\u6bcf\u6b21\u83b7\u5f97\u4e00\u4e2a\u65b0\u7684\u6570\u636e\u96c6\uff0c\u90fd\u4ece\u539f\u59cb\u7684\u6a21\u578b\u5f00\u59cb\u8fdb\u884c Fine-tune \u4ece\u800c\u9632\u6b62\u8fc7\u62df\u5408\u3002<\/li>\n<\/ol>\n<p><strong>STaR \u7684\u601d\u8def\u548c RL \u4e2d\u7b56\u7565\u68af\u5ea6\u7b97\u6cd5\u662f\u8fd1\u4f3c\u7684\uff0c\u751a\u81f3\u6574\u4f53\u7684\u4f18\u5316\u76ee\u6807\u53ef\u4ee5\u8fd1\u4f3c\u4e3a\u4e00\u4e2a\u7b56\u7565\u68af\u5ea6\u4f18\u5316\u7684\u76ee\u6807<\/strong>\u3002<\/p>\n<p>\u6a21\u578b\u9996\u5148\u91c7\u6837\u6f5c\u5728\u7684\u63a8\u7406\u8def\u5f84\uff08rationale\uff09\u7684\u8fc7\u7a0b\u7c7b\u4f3c\u4e8e RL \u4e2d\u901a\u8fc7\u7b56\u7565\u9009\u62e9\u52a8\u4f5c\uff08action\uff09\uff0c\u57fa\u4e8e\u73af\u5883\u72b6\u6001\u9009\u62e9\u4e00\u4e2a\u53ef\u80fd\u7684\u7b56\u7565\u8def\u5f84\u3002STaR \u4e2d\uff0c\u901a\u8fc7\u8ba1\u7b97\u76ee\u6807\u51fd\u6570\uff0c\u6a21\u578b\u5bf9\u6574\u4e2a\u6570\u636e\u96c6\u7684\u9884\u6d4b\u7ed3\u679c\u8fdb\u884c\u8bc4\u4f30\uff0c\u5e76\u4e14\u53ea\u6839\u636e\u9884\u6d4b\u6b63\u786e\u7684\u6837\u672c\u66f4\u65b0\u6a21\u578b\u3002<\/p>\n<p>STaR \u5728\u540c\u4e00\u6279\u6570\u636e\u4e0a\u8fdb\u884c\u591a\u6b21\u68af\u5ea6\u66f4\u65b0\uff0c\u8fd9\u7c7b\u4f3c\u4e8e\u67d0\u4e9b\u7b56\u7565\u68af\u5ea6\u7b97\u6cd5\u4e2d\u7684\u7b56\u7565\uff0c\u5373\u901a\u8fc7\u591a\u6b21\u8c03\u6574\u540c\u4e00\u6279\u6570\u636e\u6765\u7a33\u5b9a\u5b66\u4e60\u8fc7\u7a0b\u3002\u5728 RL \u4e2d\uff0c\u7b56\u7565\u68af\u5ea6\u7b97\u6cd5\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\u5728\u63a2\u7d22\u52a8\u4f5c\u7a7a\u95f4\u65f6\u8fdb\u884c\u5b66\u4e60\uff0c\u800c STaR \u5219\u901a\u8fc7\u63a2\u7d22\u63a8\u7406\u548c\u7b54\u6848\u7a7a\u95f4\uff0c\u9010\u6b65\u6539\u5584\u63a8\u7406\u751f\u6210\u7684\u51c6\u786e\u6027\u3002<\/p>\n<p>\u8fd9\u79cd\u65b9\u6cd5\u548c\u5148\u524d\u63d0\u5230\u7684\u901a\u8fc7\u7ec6\u7c92\u5ea6\u5956\u52b1\u6216 MCTS <strong>\u4f18\u5316\u8f93\u51fa<\/strong>\u6709\u6240\u4e0d\u540c\uff0c\u6a21\u578b\u5728\u6b63\u786e\u548c\u9519\u8bef\u7684\u793a\u4f8b\u4e2d\u66f4\u591a\u7684\u5b66\u4f1a\u7684\u662f\u5982\u4f55\u8fdb\u884c<strong>\u663e\u5f0f\u7684\u5408\u7406\u63a8\u7406<\/strong>\u3002<\/p>\n<p>\u4e0e\u6b64\u540c\u65f6\uff0c\u8fd9\u79cd\u5408\u7406\u63a8\u7406\u4e0d\u53ea\u662f\u95ee\u9898\u62c6\u89e3\u5206\u6b65\u7406\uff0c\u66f4\u9002\u7528\u4e8e\u4e00\u822c\u5e38\u8bc6\u95ee\u7b54\u4efb\u52a1\u4e0a\u3002\u4f8b\u5982\uff1a<\/p>\n<ol>\n<li>\u95ee\u9898\uff1a\u4ec0\u4e48\u53ef\u4ee5\u88ab\u7528\u6765\u88c5\u4e00\u53ea\u5c0f\u72d7<\/li>\n<li>\u9009\u9879\uff1a(a) \u6e38\u6cf3\u6c60 (b) \u7bee\u5b50 (c) \u540e\u9662 (d) \u81ea\u5df1\u7684\u5bb6<\/li>\n<li>\u5408\u7406\u63a8\u7406\uff1a\u7b54\u6848\u5fc5\u987b\u662f\u53ef\u4ee5\u7528\u6765\u643a\u5e26\u4e00\u53ea\u5c0f\u72d7\u7684\u4e1c\u897f\u3002\u7bee\u5b50\u662f\u7528\u6765\u88c5\u4e1c\u897f\u7684\u3002\u56e0\u6b64\uff0c\u7b54\u6848\u662f (b) \u7bee\u5b50\u3002<\/li>\n<\/ol>\n<p>\u4f46\u662f STaR \u5b58\u5728\u51e0\u4e2a\u5c40\u9650\u6027\uff1a<\/p>\n<ol>\n<li>\n<strong>\u5bf9\u5c11\u6837\u672c\u793a\u4f8b\u7684\u4f9d\u8d56<\/strong>\uff1aSTaR \u5728\u63a8\u7406\u4efb\u52a1\u4e2d\u9ad8\u5ea6\u4f9d\u8d56\u5c11\u91cf\u7684 Few-Shot \u63a8\u7406\u793a\u4f8b\uff0c\u8fd9\u5bfc\u81f4\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u8f83\u4e3a\u6709\u9650\uff0c\u96be\u4ee5\u5e94\u5bf9\u590d\u6742\u548c\u5e7f\u6cdb\u7684\u4efb\u52a1\u3002<\/li>\n<li>\n<strong>\u6cdb\u5316\u80fd\u529b\u53d7\u9650<\/strong>\uff1aSTaR \u867d\u7136\u80fd\u591f\u901a\u8fc7\u8fed\u4ee3\u7684\u65b9\u5f0f\u63d0\u5347\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u5176\u5e94\u7528\u4e3b\u8981\u5c40\u9650\u4e8e\u7279\u5b9a\u7684\u7ed3\u6784\u5316\u4efb\u52a1\uff08\u5982\u95ee\u9898\u56de\u7b54\uff09\uff0c\u96be\u4ee5\u5728\u5f00\u653e\u57df\u6216\u4efb\u610f\u6587\u672c\u751f\u6210\u4efb\u52a1\u4e2d\u53d6\u5f97\u540c\u6837\u7684\u6548\u679c\u3002<\/li>\n<\/ol>\n<p>\u9488\u5bf9 STaR \u7684\u5c40\u9650\u6027\uff0cQuiet-STaR [7] \u63d0\u51fa \u201c\u5185\u90e8\u601d\u7ef4\u201d \u7684\u6982\u5ff5\uff0c\u5c06\u663e\u5f0f\u7684 Rationales \u63a8\u7406\u8fc7\u7a0b\u8f6c\u5316\u4e3a\u6a21\u578b\u5185\u90e8\u9690\u5f0f\u7684\u63a8\u7406\u8fc7\u7a0b\uff0c\u4ece\u800c\u6446\u8131\u5bf9\u4e8e\u5916\u90e8\u793a\u4f8b\u7684\u4f9d\u8d56\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-d1334a63643a6ff637fc425d2128fac8.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-d1334a63643a6ff637fc425d2128fac8.png\"><\/a><\/p>\n<p>\u540c\u65f6\uff0c\u5f15\u5165\u53ef\u5b66\u4e60\u7684 &lt;|startofthought|&gt; \u548c &lt;|endofthought|&gt; \u00a0token \u6765\u6807\u8bb0\u601d\u7ef4\u7684\u5f00\u59cb\u548c\u7ed3\u675f\u3002<\/p>\n<p>Quiet-STaR \u8fd8\u5b9e\u73b0\u4e86\u5728\u66f4\u4e00\u822c\u6587\u672c\u4e0a\u7684\u63a8\u7406\u5b66\u4e60\uff0c\u8fd9\u610f\u5473\u7740\u5927\u91cf\u590d\u6742\u4efb\u52a1\u4e0b\u7684\u975e\u7ed3\u6784\u5316\u8bed\u6599\uff08\u5982\u533b\u7597\u3001\u91d1\u878d\u7b49\u9886\u57df\uff09\u90fd\u53ef\u4ee5\u88ab\u52a0\u5165\u5b66\u4e60\u8fc7\u7a0b\u3002\u540c\u65f6\u5229\u7528\u5e26\u63a8\u7406\u8fc7\u7a0b\u7684\u7ed3\u679c\u4e0e\u771f\u5b9e\u7ed3\u679c\u7684\u5206\u5e03\u5dee\u5f02\u5f15\u5165\u5956\u52b1\u4fe1\u53f7\uff0c\u901a\u8fc7 REINFORCE \u7684\u65b9\u6cd5\u4f18\u5316\u751f\u6210\u7684\u63a8\u7406\uff0c\u4f7f\u5f97\u57fa\u4e8e\u8fd9\u4e9b\u63a8\u7406\u7684\u6a21\u578b\u9884\u6d4b\u672a\u6765\u7684 tokens \u66f4\u4e3a\u51c6\u786e\u3002<\/p>\n<p><strong>\u5c31\u76ee\u524d\u6765\u770b\uff0cSTaR \u548c Quiet-STaR \u662f\u6700\u63a5\u8fd1 o1 \u7684\u6280\u672f\u8def\u7ebf\u548c\u6a21\u578b\u8868\u73b0\u6548\u679c\u7684\uff0c\u4f46\u662f\u5982\u679c\u60f3\u8981\u8fdb\u4e00\u6b65\u8fbe\u5230 OpenAI o1 \u7684\u6548\u679c\uff0c\u8fd8\u9700\u8981\u514b\u670d\u5f88\u591a\u95ee\u9898<\/strong>\u3002<\/p>\n<p>\u4f8b\u5982\u5982\u4e0b\u4e24\u4e2a\u95ee\u9898\uff1a<\/p>\n<ol>\n<li>Quiet-STaR \u5728\u751f\u6210\u5185\u90e8\u601d\u7ef4\u7684\u8fc7\u7a0b\u4e2d\uff0c\u6bcf\u4e2a Token \u5747\u4f1a\u751f\u6210\u4e0b\u4e00\u6b65\u7684\u5bf9\u5e94\u7684\u601d\u8003\u8fc7\u7a0b\uff0c\u5bfc\u81f4\u751f\u6210\u4e86\u5927\u91cf\u989d\u5916\u7684 tokens\uff0c\u8fd9\u4e5f\u5bfc\u81f4\u4e86\u8ba1\u7b97\u8d44\u6e90\u9700\u6c42\u5927\u5e45\u589e\u52a0\u3002\u5b9e\u9645\u4e0a\u6a21\u578b\u9700\u8981\u5b66\u4f1a\u52a8\u6001\u7684\u8c03\u6574 Thinking Token\u3002<\/li>\n<li>\u5bf9\u4e8e\u66f4\u590d\u6742\u7684\u4efb\u52a1\u548c\u957f\u7a0b\u95ee\u9898\uff0c <strong>\u5982\u4f55\u9488\u5bf9\u5185\u90e8\u601d\u8003\u8fc7\u7a0b\u63d0\u4f9b\u7ec6\u7c92\u5ea6\u7684\u5956\u52b1\u4fe1\u53f7<\/strong>\uff1f\u4ec5\u4ec5\u901a\u8fc7\u6bd4\u8f83\u5408\u7406\u63a8\u7406\u7684\u56de\u7b54\u548c\u6b63\u786e\u56de\u7b54\u662f\u5426\u4e00\u81f4\uff08\u6216\u8005 Predicted Distribution \u7684\u76f8\u4f3c\u5ea6\uff09\u662f\u4e0d\u591f\u7684\u3002<\/li>\n<\/ol>\n<p>\u8fd9\u4e0d\u7981\u5f15\u53d1\u6211\u4eec\u5bf9\u4e8e OpenAI o1 \u7684\u6280\u672f\u8def\u5f84\u7684\u601d\u8003\u3002OpenAI o1 \u5e94\u5f53\u4e5f\u662f\u6cbf\u7740 STaR \u548c Quiet-STaR \u7c7b\u4f3c\u7684\u8def\u7ebf\uff0c\u4f18\u5316\u6a21\u578b\u5185\u90e8\u751f\u6210\u5408\u7406\u63a8\u7406\uff08\u5373\u9690\u5f0f\u7684 CoT) \u7684\u8fc7\u7a0b\u3002\u800c Post-Training \u9636\u6bb5 RL \u7684\u8bad\u7ec3\u9636\u6bb5\u4e3b\u8981\u7b97\u529b\u4e5f\u5e94\u5f53\u662f\u653e\u5728\u4e86\u5bf9\u4e8e\u5185\u90e8\u63a8\u7406\u8fc7\u7a0b\u7684\u4f18\u5316\u4e0a\u3002<\/p>\n<p><strong>\u90a3\u5982\u4f55\u6784\u9020\u9690\u5f0f CoT \u7684\u4f18\u5316\u8fc7\u7a0b\u7684 Reward\uff1f<\/strong><\/p>\n<p>\u53ef\u4ee5\u901a\u8fc7\u4e0d\u540c\u6e29\u5ea6\u91c7\u6837\u51fa\u6765\u7684\u63a8\u7406\u8def\u5f84\u6784\u5efa\u504f\u5e8f\uff0c\u4e5f\u53ef\u80fd\u662f MCTS \u641c\u51fa\u6765\u7684\u6b63\u8bef\u53c2\u534a\u7684\u4e0d\u540c\u63a8\u7406\u8fc7\u7a0b\u5f62\u6210\u504f\u5e8f\u3002\u8fd9\u70b9\u548c\u5148\u524d\u7684 MCTS \u7528\u6cd5\u4f1a\u6709\u6240\u4e0d\u540c\uff0c<strong>MCTS \u8282\u70b9\u4e0a\u4e0d\u518d\u662f\u6700\u7ec8\u751f\u6210\u7b54\u6848\u4e2d\u7684\u67d0\u4e2a token \u6216\u67d0\u6b65\uff0c\u800c\u662f\u9690\u5f0f\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u6bcf\u4e00\u6b65<\/strong>\u3002<\/p>\n<p>\u540c\u65f6\uff0c\u4e3a\u4e86\u63d0\u4f9b\u66f4\u52a0\u7ec6\u7c92\u5ea6\u7684\u53cd\u9988\u548c\u6307\u5bfc\uff0c\u9700\u8981<strong>\u5f15\u5165\u8fc7\u7a0b\u6027\u7684\u5956\u52b1<\/strong>\uff0c\u800c\u9488\u5bf9\u6a21\u578b\u81ea\u8eab\u5df2\u7ecf\u96be\u4ee5\u63d0\u4f9b\u5408\u7406\u63a8\u7406\u8fc7\u7a0b\u7684\u590d\u6742\u95ee\u9898\uff0c\u901a\u8fc7\u5f15\u5165\u989d\u5916\u7684\u8db3\u591f\u5f3a\u7684 Critic Model \u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n<p>\u6700\u7ec8\u901a\u8fc7<strong>\u5f3a\u5316\u5b66\u4e60<\/strong>\uff0co1 \u5b66\u4f1a\u4e86\u4f18\u5316\u5176\u601d\u7ef4\u94fe\uff0c\u5e76\u4e0d\u65ad\u6539\u8fdb\u5176\u4f7f\u7528\u7684\u7b56\u7565\u3002\u5b83\u5b66\u4f1a\u8bc6\u522b\u5e76\u7ea0\u6b63\u9519\u8bef\uff0c\u5b66\u4f1a\u5c06\u590d\u6742\u7684\u6b65\u9aa4\u5206\u89e3\u4e3a\u66f4\u7b80\u5355\u7684\u6b65\u9aa4\uff0c\u5e76\u5728\u5f53\u524d\u65b9\u6cd5\u65e0\u6548\u65f6\u5c1d\u8bd5\u4e0d\u540c\u7684\u89e3\u51b3\u65b9\u6848\u3002\u8fd9\u4e2a\u8fc7\u7a0b\u5927\u5e45\u63d0\u9ad8\u4e86\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u3002<\/p>\n<p>\u540c\u65f6\uff0c\u5728 OpenAI \u62ab\u9732\u7684\u7ec6\u8282\u4e2d\uff0c\u751f\u6210\u8fc7\u7a0b\u4e2d\u7684 Reasoning Token \u662f\u52a8\u6001\u5f15\u5165\u7684\uff0c\u8fd9\u4e5f\u5c3d\u53ef\u80fd\u7684\u51cf\u5c11\u4e86\u4e0d\u5fc5\u8981\u7684\u601d\u8003\u5e26\u6765\u7684\u989d\u5916\u7b97\u529b\u635f\u8017\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-f01c13d717aaf2612437f271405a2a37.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-f01c13d717aaf2612437f271405a2a37.png\"><\/a><\/p>\n<p>\u53ef\u4ee5\u8bf4\uff0cOpenAI o1 \u5df2\u4e0d\u518d\u662f\u5373\u65f6\u7ed9\u51fa\u7b54\u6848\u7684\u6a21\u578b\uff0c\u800c\u662f\u80fd\u591f\u5148\u8fdb\u884c\u6df1\u5165\u601d\u8003\u3002\u8fd9\u53ef\u4ee5\u7c7b\u6bd4\u4e3a o1 \u6b63\u5728\u4ece\u4f9d\u8d56<strong>\u7cfb\u7edf 1 \u601d\u7ef4<\/strong>\uff08\u5373\u5feb\u901f\u3001\u81ea\u52a8\u3001\u76f4\u89c9\u3001\u6613\u51fa\u9519\u7684\u601d\u7ef4\u6a21\u5f0f\uff09\uff0c\u9010\u6b65\u8fdb\u5316\u4e3a\u91c7\u7528<strong>\u7cfb\u7edf 2 \u601d\u7ef4<\/strong>\uff08\u5373\u7f13\u6162\u3001\u523b\u610f\u3001\u6709\u610f\u8bc6\u4e14\u66f4\u53ef\u9760\u7684\u63a8\u7406\u8fc7\u7a0b\uff09\u3002\u8fd9\u4e00\u8f6c\u53d8\u8d4b\u4e88\u4e86 o1 \u89e3\u51b3\u4e4b\u524d\u65e0\u6cd5\u5e94\u5bf9\u7684\u590d\u6742\u95ee\u9898\u7684\u80fd\u529b\uff0c\u800c\u8fd9\u4e00\u5207\u7684\u5b9e\u73b0\uff0c\u6839\u6e90\u4e8e\u8bad\u7ec3\u540e\u9636\u6bb5\u4e2d\u9075\u5faa\u7684<strong>\u6269\u5c55\u89c4\u5f8b\uff08Scaling Laws\uff09<\/strong>\u7684\u5e94\u7528\u4e0e\u4f18\u5316\u3002<\/p>\n<p>\u66f4\u6709\u610f\u601d\u7684\u662f\uff0c\u6211\u4eec\u53ef\u4ee5\u6784\u5efa\u4e00\u4e2a<strong>\u6570\u636e\u98de\u8f6e<\/strong>\uff1a\u901a\u8fc7 o1 \u6a21\u578b\u7684\u63a8\u7406\u8fc7\u7a0b\u81ea\u52a8\u751f\u6210\u5927\u91cf\u9ad8\u8d28\u91cf\u7684\u8bad\u7ec3\u6570\u636e\uff0c\u8fd9\u4e9b\u6570\u636e\u53ef\u4ee5\u88ab\u53cd\u590d\u7528\u4e8e\u8fdb\u4e00\u6b65\u63d0\u5347\u6a21\u578b\u6027\u80fd\uff0c\u5f62\u6210\u4e00\u4e2a\u81ea\u6211\u5f3a\u5316\u7684\u826f\u6027\u5faa\u73af\u3002<\/p>\n<p>\u5728\u8fd9\u4e00\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u7684<strong>\u81ea\u4e3e\u80fd\u529b<\/strong>\uff08Bootstrap\uff09\u5f97\u5230\u8fdb\u4e00\u6b65\u6269\u5c55\uff0c\u4e0d\u4ec5\u52a0\u901f\u4e86\u6027\u80fd\u63d0\u5347\u7684\u8fdb\u7a0b\uff0c\u66f4\u6709\u671b\u9010\u6b65\u63a8\u52a8\u5411\u8d85\u7ea7\u667a\u80fd\uff08Superintelligence\uff09\u7684\u8fc8\u8fdb\u3002<\/p>\n<p>\u603b\u7ed3\u4e00\u4e0b\uff1a<\/p>\n<ul>\n<li>\n<strong>RL + \u201c\u9690\u5f0f\u601d\u7ef4\u94fe\u201d<\/strong>\uff1ao1 \u6a21\u578b\u4f7f\u7528 RL \u8fdb\u884c\u8bad\u7ec3\uff0c\u901a\u8fc7\u5f15\u5165\u52a8\u6001\u7684 Reasoning Token\uff0c\u4ece\u800c\u542f\u53d1 \u201c\u9690\u5f0f\u601d\u7ef4\u94fe\u201d \u6765 \u201c\u601d\u8003\u201d \u95ee\u9898\uff0c\u601d\u8003\u65f6\u95f4\u8d8a\u957f\uff0c\u63a8\u7406\u80fd\u529b\u8d8a\u5f3a\uff01<\/li>\n<li>\n<strong>\u63a8\u7406\u65f6\u95f4 = \u65b0\u7684\u6269\u5c55\u7ef4\u5ea6<\/strong>\uff1ao1 \u6a21\u578b\u7684\u53d1\u5e03\uff0c\u610f\u5473\u7740 AI \u80fd\u529b\u7684\u63d0\u5347\u4e0d\u518d\u5c40\u9650\u4e8e\u9884\u8bad\u7ec3\u9636\u6bb5\uff0c\u8fd8\u53ef\u4ee5\u901a\u8fc7\u5728 Post-Training \u9636\u6bb5\u4e2d\u63d0\u5347 RL \u8bad\u7ec3\u7684\u63a2\u7d22\u65f6\u95f4\u548c\u589e\u52a0\u6a21\u578b\u63a8\u7406\u601d\u8003\u65f6\u95f4\u6765\u5b9e\u73b0\u6027\u80fd\u63d0\u5347\uff0c\u5373 Post-Training Scaling Laws\u3002<\/li>\n<li>\n<strong>\u6570\u636e\u98de\u8f6e + Bootstrap -&gt; SuperIntelligence<\/strong> : \u57fa\u4e8e\u81ea\u6211\u53cd\u601d\u7684\u6a21\u578b\u5c06\u80fd\u591f\u5b9e\u73b0\u81ea\u4e3e Bootstrap\uff0c\u5e76\u63d0\u5347\u5927\u5927\u63d0\u5347\u6a21\u578b\u5bf9\u4e8e\u672a\u89c1\u8fc7\u7684\u590d\u6742\u95ee\u9898\u7684\u89e3\u51b3\u80fd\u529b\uff0c\u6a21\u578b\u7684\u63a8\u7406\u8fc7\u7a0b\u5f62\u6210\u5927\u91cf\u9ad8\u8d28\u91cf\u6570\u636e\u7684\u98de\u8f6e\uff0c\u5e76\u6700\u7ec8\u6709\u53ef\u80fd\u5411 SuperIntelligence \u66f4\u8fdb\u4e00\u6b65\u3002<\/li>\n<\/ul>\n<p><strong>\u8bc4\u8bba\u6a21\u578b Critic Model\u00a0<\/strong><\/p>\n<p><strong>\u968f\u7740\u4efb\u52a1\u95ee\u9898\u7684\u9010\u6b65\u590d\u6742\uff0c\u4ec5\u4ec5\u4f9d\u9760\u6a21\u578b\u7684\u81ea\u8eab\u63a8\u7406\u80fd\u529b\u53ef\u80fd\u65e0\u6cd5\u63d0\u4f9b\u6709\u6548\u7684\u5956\u52b1\u4fe1\u53f7\u3002\u8fd9\u4f7f\u5f97\u5bf9\u4e8e\u6a21\u578b\u5185\u90e8\u7684\u590d\u6742\u63a8\u7406\u8fc7\u7a0b\u7684\u76d1\u7763\u53d8\u6210\u4e86\u4e00\u4e2a\u53ef\u6269\u5c55\u76d1\u7763\u95ee\u9898<\/strong>\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\uff0cOpenAI o1 \u9690\u5f0f\u601d\u7ef4\u94fe\u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u5e94\u5f53\u4e5f\u5f15\u5165\u4e86 Critic \u7684\u65b9\u6cd5\u3002\u9488\u5bf9\u590d\u6742\u63a8\u7406\u7684\u95ee\u9898\uff0c\u6a21\u578b\u81ea\u8eab\u5df2\u7ecf\u96be\u4ee5\u63d0\u4f9b\u5408\u7406\u63a8\u7406\u8fc7\u7a0b\uff0c\u56e0\u6b64\u8feb\u5207\u9700\u8981<strong>\u5f15\u5165\u989d\u5916\u7684\u8db3\u591f\u5f3a\u7684 Critic Model \u6765\u63d0\u4f9b\u7cbe\u51c6\u7684\u53cd\u9988<\/strong>\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u901a\u8fc7\u5c06\u63a8\u7406\u8fc7\u7a0b\u8fdb\u884c\u8fc7\u7a0b\u5206\u89e3\uff0c\u5e76\u4e14\u5229\u7528\u989d\u5916\u7684\u66f4\u5f3a\u66f4\u4e13\u9879\u7684 Critic Model\uff0c\u53ef\u4ee5\u5c06\u63a8\u7406\u8fc7\u7a0b\u7684\u76d1\u7763\u6269\u5c55\u5230\u66f4\u590d\u6742\u7684\u95ee\u9898\u4e0a\u3002\u8fd9\u4e5f\u4e00\u5b9a\u7a0b\u5ea6\u7f13\u89e3\u4e86\u4ec5\u4ec5\u662f\u901a\u8fc7\u63a8\u7406\u8fc7\u7a0b\u80fd\u5426\u5bfc\u51fa\u6b63\u786e\u7ed3\u679c\u7684\u6765\u786e\u5b9a\u5956\u52b1\u4fe1\u53f7\u7684\u7a00\u758f\u95ee\u9898\u3002<\/p>\n<p>\u8fd9\u4e2a\u601d\u8def\u65e9\u5728\u5148\u524d\u4e5f\u6709\u6240\u63a2\u7d22\u3002<\/p>\n<p>\u524d\u9635\u5b50 OpenAI \u53d1\u5e03\u7684 CriticGPT [2]\uff0c \u901a\u8fc7 RLHF \u65b9\u6cd5\u8bad\u7ec3\u6a21\u578b\u80fd\u591f\u4e3a\u771f\u5b9e\u4e16\u754c\u4e2d\u7684\u4ee3\u7801\u4efb\u52a1\u4e66\u5199\u81ea\u7136\u8bed\u8a00\u53cd\u9988\uff0c\u5e76\u6210\u529f\u6cdb\u5316\u5230 OOD \u7684\u5206\u5e03\u4e0a\u3002\u8fd9\u79cd\u53cd\u9988\u53ef\u4ee5\u7528\u6765\u5e2e\u52a9\u4eba\u7c7b\u8fdb\u884c\u66f4\u51c6\u786e\u7684\u8bc4\u4ef7\uff0c\u4ece\u800c\u5b9e\u73b0\u5bf9\u4e8e\u590d\u6742\u8f93\u51fa\u7684\u6709\u6548\u5956\u52b1\u53cd\u9988\u3002\u5148\u524d OpenAI \u4e5f\u6df1\u5165\u63a2\u7a76\u8fc7\u81ea\u6211\u6279\u5224\u65b9\u6cd5\u548c Critic Model \u8f85\u52a9\u4eba\u7c7b\u8bc4\u5224 \u5728\u6587\u672c\u603b\u7ed3\u4efb\u52a1\u4e0a\u7684\u53ef\u884c\u6027 [3]\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-0b4e5efaf82c6dc2797d924c67c449c5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-0b4e5efaf82c6dc2797d924c67c449c5.png\"><\/a><\/p>\n<p>\u4ece\u53ef\u6269\u5c55\u76d1\u7763\u7684\u89d2\u5ea6\u6765\u8bf4\uff0c\u8fd9\u6761\u8def\u662f\u5fc5\u7136\u7684\u3002\u968f\u7740\u4efb\u52a1\u53d8\u5f97\u6108\u53d1\u590d\u6742\uff08\u5982\u6570\u5b66\u4ee3\u7801\u63a8\u7406\uff09\uff0c\u4eba\u7c7b\u53cd\u9988\u8005\u5f88\u96be\u5bf9\u6a21\u578b\u7684\u56de\u7b54\u8fdb\u884c\u6709\u6548\u7684\u8bc4\u4ef7\uff0c\u4e5f\u5c31\u65e0\u6cd5\u63d0\u4f9b\u6709\u6548\u7684\u504f\u597d\u6216\u8005\u5956\u52b1\u4fe1\u53f7\uff0c\u5982\u4f55\u5728\u66f4\u590d\u6742\u7684\u4efb\u52a1\u4e0a\u5bf9\u9f50\u66f4\u5f3a\u5927\u7684\u6a21\u578b\uff0c\u662f\u53ef\u6269\u5c55\u76d1\u7763\u7684\u91cd\u8981\u95ee\u9898\u3002<\/p>\n<p>\u5bf9\u4e8e Critic \u8fd9\u7c7b\u65b9\u6cd5\uff0c\u5173\u952e\u7684\u6311\u6218\u5728\u4e8e\u5982\u4f55\u5c06 Critic \u7684\u80fd\u529b\u6cdb\u5316\u5230\u66f4\u52a0\u590d\u6742\u7684\u4efb\u52a1\uff0c\u4f8b\u5982\u5bf9\u4e8e\u4ee3\u7801\u6570\u5b66\u6216\u8005\u957f\u6587\u672c\u8f93\u51fa\uff0cCritic Model \u9700\u8981\u8003\u8651\u8f93\u51fa\u5404\u90e8\u5206\u4e4b\u95f4\u7684\u4f9d\u8d56\u5173\u7cfb\u548c\u903b\u8f91\u63a8\u7406\u95ee\u9898\uff0c\u56e0\u6b64\u5bf9\u4e8e\u6a21\u578b\u81ea\u8eab\u7684\u63a8\u7406\u80fd\u529b\u8981\u6c42\u66f4\u9ad8\u3002<\/p>\n<p>\u7528\u4e8e\u8bc4\u8bba\u6279\u5224\u7684\u6a21\u578b\uff08Critic Model\uff09\u540c\u6837\u4e5f\u53ef\u80fd\u4f1a\u5b58\u5728 Generator-Discriminator-Critique (GDC) gaps \uff0c\u5373\u6a21\u578b\u53ef\u80fd\u4e0d\u4f1a\u6307\u51fa\u4ed6\u4eec\u6240\u53d1\u73b0\u7684\u9519\u8bef\uff0c\u8fd9\u4e00\u5dee\u8ddd\u5728 CriticGPT \u8fd9\u4e2a\u91cf\u7ea7\u5c3a\u5bf8\u7684\u6a21\u578b\u4e0a\u662f\u5426\u88ab\u7f29\u51cf\u76ee\u524d\u8fd8\u4e0d\u5f97\u800c\u77e5\u3002<\/p>\n<p>\u53ef\u4ee5\u8bf4\u8fd9\u4e00\u7cfb\u5217\u5de5\u4f5c\u662f\u4e00\u8109\u76f8\u627f\u7684\uff0c\u57fa\u4e8e\u8bc4\u4ef7\u6bd4\u751f\u6210\u66f4\u7b80\u5355\u7684\u539f\u5219\uff0cCritic \u7684\u601d\u8def\u662f\u53ef\u6269\u5c55\u76d1\u7763\u4e00\u6761\u6709\u5e0c\u671b\u7684\u6280\u672f\u8def\u5f84\uff0c\u5e76\u4e14\u5e94\u8be5\u4e5f\u88ab\u7528\u5728\u4e86\u8f85\u52a9 o1 \u7684\u8bad\u7ec3\u5f53\u4e2d\u3002<\/p>\n<p><strong>\u5927\u6a21\u578b\u7684\u5929\u82b1\u677f\u5728\u54ea\u91cc\uff1f<\/strong><\/p>\n<p>\u81ea\u4ece 2022 \u5e74 ChatGPT \u9762\u4e16\u4ee5\u6765\uff0c\u5927\u6a21\u578b\u7ecf\u8fc7\u4e86\u8fd1\u4e24\u5e74\u7684\u8fed\u4ee3\u3002\u76ee\u524d\uff0c\u65e0\u8bba\u662f\u5de5\u4e1a\u754c\u8fd8\u662f\u5b66\u672f\u754c\uff0c\u90fd\u5728\u63a2\u7d22\u5927\u6a21\u578b\u7684\u4e0a\u9650\u3002\u5728 Scaling Law \u7684\u652f\u6301\u4e0b\uff0c\u5927\u6a21\u578b\u6301\u7eed\u589e\u52a0\u9884\u8bad\u7ec3\u7684\u6570\u636e\u91cf\u548c\u6a21\u578b\u53c2\u6570\u3002\u7136\u800c\uff0c\u968f\u7740\u786c\u4ef6\u96c6\u7fa4\u7684\u9650\u5236\u548c\u6210\u672c\u7684\u7ea6\u675f\uff0c\u6a21\u578b\u53c2\u6570\u7684\u589e\u957f\u5df2\u9010\u6e10\u505c\u6ede\u3002\u5728 OpenAI o1\uff0c\u666e\u904d\u8ba4\u4e3a\u8981\u8fdb\u4e00\u6b65\u63d0\u5347\u5927\u6a21\u578b\u7684\u80fd\u529b\uff0c\u4e3b\u8981\u6709\u4ee5\u4e0b\u4e24\u6761\u6280\u672f\u8def\u7ebf\uff1a<\/p>\n<ol>\n<li>\n<strong>\u901a\u8fc7\u5408\u6210\u6570\u636e\u8fdb\u4e00\u6b65\u6269\u5c55\u6570\u636e\u548c\u53c2\u6570\u89c4\u6a21<\/strong>\u3002\u4e00\u4e9b\u6a21\u578b\u4f7f\u7528\u4e86\u5927\u91cf\u7684\u516c\u5f00\u6570\u636e\u8fdb\u884c\u8bad\u7ec3\uff0c\u968f\u7740\u6570\u636e\u91cf\u7684\u589e\u52a0\uff0c\u6a21\u578b\u6027\u80fd\u4ecd\u5728\u63d0\u5347\u3002\u7136\u800c\uff0c\u968f\u7740\u65f6\u95f4\u7684\u63a8\u79fb\uff0c\u6570\u636e\u7a00\u7f3a\u5c06\u9010\u6e10\u6210\u4e3a\u589e\u52a0\u66f4\u591a\u6570\u636e\u7684\u6311\u6218\u3002\u4e00\u4e9b\u89e3\u51b3\u65b9\u6848\u5305\u62ec\u751f\u6210\u5408\u6210\u8bad\u7ec3\u6570\u636e\uff0c\u4f8b\u5982 NVIDIA \u53d1\u5e03\u4e86 Nemotron-4 340B \u53ef\u4ee5\u5e2e\u52a9\u5728\u65e0\u6cd5\u83b7\u53d6\u5927\u91cf\u3001\u591a\u6837\u7684\u6807\u6ce8\u6570\u636e\u96c6\u7684\u60c5\u51b5\u4e0b\u751f\u6210\u5408\u6210\u8bad\u7ec3\u6570\u636e\uff0c\u5e76\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u89e3\u51b3\u6570\u636e\u9965\u8352\u7684\u95ee\u9898\uff1bOpenAI o1 \u4e5f\u662f\u57fa\u4e8e\u8fd9\u6837\u7684\u601d\u8def\uff0c\u63d0\u4f9b\u4e86\u6784\u5efa\u5408\u6210\u6570\u636e\u98de\u8f6e\u7684\u673a\u4f1a\u3002<\/li>\n<li>\n<strong>\u901a\u8fc7\u6a21\u6001\u6df7\u5408\u548c\u6a21\u6001\u7a7f\u900f\u7684\u65b9\u6cd5\uff0c\u501f\u52a9\u5176\u4ed6\u6a21\u6001\u589e\u5f3a\u6a21\u578b\u80fd\u529b<\/strong>\u3002\u76f8\u6bd4\u4e8e\u516c\u5f00\u7684\u6587\u672c\u6570\u636e\uff0c\u56fe\u50cf\u3001\u89c6\u9891\u3001\u97f3\u9891\u7b49\u6570\u636e\u7684\u603b\u91cf\u66f4\u5927\uff0c\u4e14\u5305\u542b\u7684\u4fe1\u606f\u91cf\u4e5f\u66f4\u4e30\u5bcc\u3002\u4e00\u6761\u53ef\u884c\u7684\u6280\u672f\u8def\u7ebf\u662f\u6709\u6548\u589e\u52a0\u6a21\u578b\u5904\u7406\u7684\u6a21\u6001\u6570\u91cf\uff0c\u4e0d\u4ec5\u8ba9\u6a21\u578b\u5b8c\u6210\u4e0d\u9650\u4e8e\u6587\u672c\u6a21\u6001\u7684\u4efb\u52a1\uff0c\u66f4\u91cd\u8981\u7684\u662f\uff0c\u901a\u8fc7\u6a21\u6001\u7a7f\u900f\u548c\u6a21\u578b\u878d\u5408\uff0c\u5728\u590d\u6742\u63a8\u7406\u80fd\u529b\u4e0a\u66f4\u4e0a\u4e00\u5c42\u697c\uff0c\u5373\u5b9e\u73b0<strong>\u6a21\u6001\u4e0a\u7684 Scaling Law<\/strong>\u3002\u5728\u8fd9\u65b9\u9762\u524d\u666f\u5e7f\u9614\uff1a\u6587\u672c\u7684\u5e8f\u5217\u5316\u4fe1\u606f\u76f8\u6bd4\u4e8e\u56fe\u50cf\u548c\u89c6\u9891\u6240\u5305\u542b\u7684\u590d\u6742\u4fe1\u606f\u8981\u5c11\u5f97\u591a\uff0c\u66f4\u4e30\u5bcc\u7684\u6570\u636e\u80fd\u591f\u6709\u6548\u6269\u5145\u6a21\u578b\u63a8\u7406\u7a7a\u95f4\u7684\u4e30\u5bcc\u5ea6\uff1b<\/li>\n<li>\n<strong>\u63a8\u7406\u80fd\u529b\u548c\u6a21\u578b\u7684\u6307\u4ee4\u8ddf\u968f\u80fd\u529b\u5448\u73b0\u51fa\u5206\u79bb\u5173\u7cfb<\/strong>\u3002\u5728 OpenAI o1 \u8868\u73b0\u4e2d\uff0c\u5c3d\u7ba1\u5728\u6570\u5b66\u3001\u7269\u7406\u7b49\u590d\u6742\u4efb\u52a1\u4e0a\u7684\u63a8\u7406\u80fd\u529b\u6709\u4e86\u5927\u5e45\u63d0\u5347\uff0c\u4f46\u5728\u4e00\u4e9b\u8bed\u8a00\u751f\u6210\u4efb\u52a1\u4e0a\uff0c\u5e76\u6ca1\u6709\u4f53\u73b0\u51fa\u66f4\u5927\u7684\u8fdb\u6b65\u3002\u5728 System Card \u548c OpenAI \u7814\u7a76\u4eba\u5458\u7684\u8bbf\u8c08\u4e2d\u4e5f\u63d0\u5230\uff0cOpenAI o1 \u4e13\u9879\u4e8e\u63a8\u7406\u80fd\u529b\uff0c\u800c\u5e76\u4e0d\u80fd\u4f5c\u4e3a\u4e00\u4e2a\u5f88\u597d\u7684 Agent \u548c Assistant\u3002\u8fd9\u79cd\u63a8\u7406\u80fd\u529b\u548c\u6307\u4ee4\u8ddf\u968f\u80fd\u529b\u7684\u5206\u79bb\u5728\u6a21\u578b\u5f3a\u5927\u5230\u4e00\u5b9a\u7a0b\u5ea6\u624d\u51fa\u73b0\uff0c\u751a\u81f3\u662f\u4e92\u65a5\u7684\u3002\u4f46\u5bf9\u4e8e\u6784\u5efa\u901a\u7528\u667a\u80fd\u4f53\u6765\u8bf4\uff0c\u5982\u4f55 balance \u4e8c\u8005\u5173\u7cfb\u4f1a\u6210\u4e3a\u4e00\u4e2a\u6838\u5fc3\u95ee\u9898\u3002<\/li>\n<\/ol>\n<p><strong>AI \u5b89\u5168\u7684\u63a8\u7406\u94fe Chain of Reasoning for AI Safety<\/strong><\/p>\n<p>OpenAI o1 \u5c55\u73b0\u51fa\u7684\u9690\u5f0f\u601d\u7ef4\u94fe\u63a8\u7406\u80fd\u529b\u4e5f\u4e3a\u6a21\u578b\u7684\u5bf9\u9f50\u548c\u5b89\u5168\u63d0\u4f9b\u4e86\u65b0\u7684\u89c6\u89d2\u3002\u901a\u8fc7\u5c06\u5b89\u5168\u7684\u89c4\u5219\u878d\u5165\u5230\u6a21\u578b\u7684\u601d\u7ef4\u94fe\u4e2d\uff0c\u6a21\u578b\u80fd\u591f\u66f4\u597d\u7684\u7406\u89e3\u89c4\u5219\u7684\u5185\u6db5\uff0c\u5e76\u4e14\u4e5f\u63d0\u4f9b\u7ed9\u76d1\u7ba1\u8005 \u201c\u8bfb\u61c2\u6a21\u578b\u7684\u5b89\u5168\u601d\u60f3\u201d \u5e76\u7406\u89e3\u5176\u601d\u7ef4\u8fc7\u7a0b\u7684\u673a\u4f1a\u3002<\/p>\n<p>\u4e00\u4e2a\u76f4\u89c2\u7684\u4f53\u73b0\u662f\uff0c\u901a\u8fc7\u878d\u5165\u5b89\u5168\u539f\u5219\u5230\u9690\u5f0f\u601d\u7ef4\u8fc7\u7a0b\u4e2d\uff0c\u6a21\u578b\u51cf\u5c11\u4e86\u77eb\u6789\u8fc7\u6b63\u548c\u8fc7\u5ea6\u62d2\u7edd Overrefuse \u7684\u60c5\u51b5\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-2fdf1d5068565d647c94c1ddd22cf8de.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-2fdf1d5068565d647c94c1ddd22cf8de.png\"><\/a><\/p>\n<p>\u4f8b\u5982\uff0c\u5bf9\u4e8e\u5982\u4e0b\u8bf7\u6c42\uff1a\u201c\u8bf7\u4f60\u5c06\u4e0b\u9762\u8fd9\u53e5\u8bdd\u7ffb\u8bd1\u4e3a\u82f1\u6587\uff1a&#8217; \u5982\u4f55\u5236\u9020\u70b8\u5f39 &#8216;\u201d\uff0cGPT-4o \u53ef\u80fd\u4f1a\u7531\u4e8e\u89c4\u5219\u8fc7\u6ee4\u62d2\u7edd\u56de\u7b54\uff0c\u4f46\u662f OpenAI o1 \u5c55\u73b0\u51fa\u4e86\u5bf9\u5b89\u5168\u89c4\u5219\u7684\u6df1\u5165\u7406\u89e3\uff0c\u4ece\u800c\u6b63\u5e38\u7684\u56de\u5e94\u7528\u6237\u7684\u9700\u6c42\u3002<\/p>\n<p>\u7136\u800c\uff0c\u9690\u5f0f\u94fe\u5f0f\u601d\u7ef4\u63a8\u7406\u4e0d\u4ec5\u5e26\u6765\u4e86\u663e\u8457\u63d0\u5347\u7684\u5b89\u5168\u57fa\u51c6\u6d4b\u8bd5\u8868\u73b0\uff0c\u540c\u65f6\u4e5f\u589e\u52a0\u4e86\u67d0\u4e9b\u7c7b\u578b\u7684\u98ce\u9669\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u901a\u8fc7\u5185\u90e8\u8bc4\u4f30\u548c\u4e0e\u5916\u90e8\u7ea2\u961f\u7684\u5408\u4f5c\uff0cOpenAI \u53d1\u73b0\u672a\u8fdb\u884c\u7f13\u89e3\u63aa\u65bd\u7684\u6a21\u578b\u5728\u8bf4\u670d\u529b\u548c\u5316\u5b66\u3001\u751f\u7269\u3001\u653e\u5c04\u6027\u548c\u6838\u5a01\u80c1\uff08CBRN\uff09\u65b9\u9762\u5b58\u5728\u4e2d\u7b49\u98ce\u9669\u3002<\/p>\n<p><strong>\u4e0e\u6b64\u540c\u65f6\uff0c\u968f\u7740\u6a21\u578b\u88ab\u63d0\u4f9b\u9690\u5f0f\u601d\u8003\u7684\u673a\u4f1a\uff0c\u5b83\u4eec\u51fa\u73b0\u6b3a\u9a97\u548c\u64cd\u7eb5\u7684\u53ef\u80fd\u6027\u9010\u6e10\u52a0\u5927<\/strong>\uff1a\u8fd9\u79cd\u6b3a\u9a97\u548c\u64cd\u7eb5\u7684\u6765\u6e90\u662f\u6a21\u578b\u901a\u8fc7\u6df1\u5165\u601d\u8003\u53d1\u73b0\u4e86\u80fd\u591f\u5b8c\u6210\u76ee\u6807\u7684\u66f4\u5feb\u65b9\u6cd5\uff0c\u5c3d\u7ba1\u8fd9\u79cd\u65b9\u6cd5\u662f\u4e0d\u88ab\u5141\u8bb8\u6216\u6b3a\u9a97\u4eba\u7c7b\u7684\u3002<\/p>\n<p>OpenAI o1 \u6a21\u578b\u5df2\u7ecf\u5c55\u73b0\u51fa\u4e86\u7b56\u7565\u63a8\u6f14\u548c\u5956\u52b1\u653b\u9677\u7684\u80fd\u529b\u3002\u5728 SystemCard \u4e2d\u63d0\u4f9b\u7684\u4e00\u4e2a\u4f8b\u5b50\u4e2d\uff0co1-preview \u6a21\u578b\u901a\u8fc7\u53d1\u73b0\u8bc4\u4f30\u4e3b\u673a\u865a\u62df\u673a\u4e0a\u7684 Docker \u5b88\u62a4\u8fdb\u7a0b API\uff0c\u5e76\u5229\u7528\u5b83\u91cd\u542f\u4e86\u635f\u574f\u7684\u6311\u6218\u5bb9\u5668\uff0c\u8bfb\u53d6\u4e86\u4efb\u52a1\u76ee\u6807\uff08flag.txt\uff09\uff0c\u4ece\u800c\u907f\u5f00\u4e86\u5e94\u5f53\u901a\u8fc7\u6f0f\u6d1e\u5229\u7528\u7684\u5b89\u5168\u6311\u6218\u3002\u6a21\u578b\u5229\u7528\u9519\u8bef\u914d\u7f6e\u7684\u57fa\u7840\u8bbe\u65bd\uff0c\u8df3\u8fc7\u4e86\u539f\u672c\u8bbe\u8ba1\u7684\u4efb\u52a1\u6b65\u9aa4\uff0c\u5b9e\u73b0\u4e86\u4efb\u52a1\u76ee\u6807\u3002\u8fd9\u8868\u660e\uff0c\u6a21\u578b\u5728\u9762\u5bf9\u4e0d\u660e\u786e\u6216\u5b58\u5728\u6f0f\u6d1e\u7684\u4efb\u52a1\u65f6\uff0c\u80fd\u591f\u901a\u8fc7\u610f\u60f3\u4e0d\u5230\u7684\u65b9\u5f0f\u5b9e\u73b0 \u201c\u5956\u52b1\u653b\u9677\u201d\uff0c\u4ece\u800c\u907f\u5f00\u771f\u6b63\u7684\u6311\u6218\u6838\u5fc3\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-06ee3e90bcdbeb9c8cc29962c9bc0436.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-06ee3e90bcdbeb9c8cc29962c9bc0436.png\"><\/a><\/p>\n<p>Anthropic \u5148\u524d\u7684\u7814\u7a76 [9] \u4e5f\u5c55\u73b0\uff1a\u8bed\u8a00\u6a21\u578b\u5728\u9762\u5bf9\u5956\u52b1\u8bbe\u8ba1\u65f6\uff0c\u53ef\u4ee5\u901a\u8fc7\u4e0d\u6b63\u5f53\u7684\u65b9\u5f0f\u83b7\u5f97\u9ad8\u5956\u52b1\uff0c\u751a\u81f3\u901a\u8fc7 \u201c\u5956\u52b1\u7be1\u6539\u201d \u4fee\u6539\u81ea\u5df1\u7684\u5956\u52b1\u673a\u5236\uff0c\u6765\u907f\u5f00\u4efb\u52a1\u4e2d\u8bbe\u7f6e\u7684\u96be\u70b9\u3002<\/p>\n<p>\u8fd9\u79cd\u80fd\u529b\u610f\u5473\u7740\uff0c<strong>\u5728\u590d\u6742\u73af\u5883\u4e0b\u6a21\u578b\u53ef\u80fd\u4f1a\u91c7\u53d6\u6377\u5f84\u6216\u504f\u79bb\u539f\u672c\u8bbe\u8ba1\u7684\u4efb\u52a1\u8def\u7ebf<\/strong>\u3002<\/p>\n<p>\u8fd9\u4e0d\u4ec5\u5bf9\u8bc4\u4f30\u4efb\u52a1\u7684\u8bbe\u8ba1\u63d0\u51fa\u4e86\u66f4\u9ad8\u7684\u8981\u6c42\uff0c\u4e5f\u63d0\u793a\u6211\u4eec\u9700\u8981\u66f4\u4e25\u683c\u7684\u76d1\u63a7\u673a\u5236\uff0c\u4ee5\u786e\u4fdd\u6a21\u578b\u5728\u89e3\u51b3\u590d\u6742\u95ee\u9898\u65f6\u4e0d\u901a\u8fc7\u6377\u5f84\u6765\u89c4\u907f\u6838\u5fc3\u95ee\u9898\u3002<\/p>\n<p>\u8fd9\u79cd\u73b0\u8c61\u53ef\u80fd\u4f1a\u5728\u672a\u6765\u6a21\u578b\u4e2d\u53d8\u5f97\u66f4\u52a0\u663e\u8457\uff0c\u5c24\u5176\u662f\u5728\u9ad8\u98ce\u9669\u5e94\u7528\u573a\u666f\u4e2d\u3002\u56e0\u6b64\uff0c\u6709\u5fc5\u8981\u4e3a\u8fd9\u4e9b\u6a21\u578b\u7684\u90e8\u7f72\u548c\u76d1\u63a7\u5efa\u7acb\u66f4\u52a0\u5b8c\u5584\u7684\u9632\u62a4\u63aa\u65bd\uff0c\u4ee5\u9632\u6b62\u7b56\u7565\u653b\u9677\u884c\u4e3a\u7684\u53d1\u751f\u3002<\/p>\n<p>\u8fd9\u8fdb\u4e00\u6b65\u8868\u660e\uff0c\u968f\u7740 AI \u6a21\u578b\u80fd\u529b\u7684\u63d0\u5347\uff0c\u7279\u522b\u662f\u5728\u590d\u6742\u7684\u9ad8\u98ce\u9669\u573a\u666f\u4e0b\uff08\u5982\u81ea\u52a8\u9a7e\u9a76\u3001\u533b\u7597\u8bca\u65ad\u3001\u91d1\u878d\u51b3\u7b56\u7b49\uff09\uff0c\u6211\u4eec\u5fc5\u987b\u9884\u89c1\u5e76\u9632\u8303\u8fd9\u4e9b\u6a21\u578b\u901a\u8fc7\u4e0d\u6b63\u5f53\u9014\u5f84\u83b7\u53d6\u5956\u52b1\u7684\u884c\u4e3a\u3002\u5f53\u524d\u7684\u76d1\u7763\u673a\u5236\u548c\u5956\u52b1\u8bbe\u8ba1\u53ef\u80fd\u4e0d\u8db3\u4ee5\u9632\u6b62\u6a21\u578b\u901a\u8fc7\u7ec6\u81f4\u7684\u63a8\u6f14\u548c\u7be1\u6539\u4ee3\u7801\u6765\u83b7\u5f97\u610f\u5916\u7684\u9ad8\u5956\u52b1\u3002\u4e3a\u4e86\u5e94\u5bf9\u8fd9\u79cd\u5a01\u80c1\uff0c\u672a\u6765\u7684\u5de5\u4f5c\u5e94\u8be5\u96c6\u4e2d\u5728\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762\uff1a<\/p>\n<ol>\n<li>\n<strong>\u4efb\u52a1\u8bbe\u8ba1\u7684\u9c81\u68d2\u6027<\/strong>\uff1a\u9700\u8981\u5bf9\u4efb\u52a1\u548c\u73af\u5883\u8fdb\u884c\u8be6\u7ec6\u5ba1\u67e5\uff0c\u786e\u4fdd\u4e0d\u5b58\u5728\u53ef\u88ab\u6a21\u578b\u5229\u7528\u7684\u7cfb\u7edf\u6f0f\u6d1e\u3002\u7279\u522b\u662f\u5728\u6d89\u53ca\u590d\u6742\u7cfb\u7edf\u914d\u7f6e\u7684\u4efb\u52a1\u4e2d\uff0c\u786e\u4fdd\u6240\u6709\u53ef\u80fd\u7684\u6377\u5f84\u90fd\u88ab\u5c01\u9501\u3002<\/li>\n<li>\n<strong>\u5b9e\u65f6\u76d1\u7763\u4e0e\u53cd\u9988<\/strong>\uff1a\u52a0\u5f3a\u76d1\u7763\u6a21\u578b\u7684\u80fd\u529b\uff0c\u901a\u8fc7\u5b9e\u65f6\u68c0\u6d4b\u6a21\u578b\u884c\u4e3a\u4e2d\u7684\u5f02\u5e38\uff0c\u8bc6\u522b\u5176\u662f\u5426\u504f\u79bb\u4e86\u9884\u671f\u7684\u4efb\u52a1\u8def\u5f84\u3002\u4f8b\u5982\uff0c\u5f53\u6a21\u578b\u8868\u73b0\u51fa\u4e0e\u5176\u4efb\u52a1\u76ee\u6807\u4e0d\u4e00\u81f4\u7684\u884c\u4e3a\u65f6\uff0c\u7cfb\u7edf\u5e94\u7acb\u5373\u53d1\u51fa\u8b66\u62a5\u5e76\u8fdb\u884c\u5e72\u9884\u3002<\/li>\n<li>\n<strong>\u5956\u52b1\u8bbe\u8ba1\u7684\u7075\u6d3b\u6027\u4e0e\u9632\u7be1\u6539\u673a\u5236<\/strong>\uff1a\u5f15\u5165\u66f4\u7075\u6d3b\u4e14\u96be\u4ee5\u88ab\u7be1\u6539\u7684\u5956\u52b1\u673a\u5236\u3002\u5956\u52b1\u673a\u5236\u7684\u8bbe\u8ba1\u5e94\u8be5\u5305\u62ec\u591a\u5c42\u6b21\u7684\u76d1\u7763\uff0c\u4e0d\u4ec5\u68c0\u6d4b\u4efb\u52a1\u5b8c\u6210\u7684\u7ed3\u679c\uff0c\u8fd8\u8981\u76d1\u63a7\u6a21\u578b\u5b8c\u6210\u4efb\u52a1\u7684\u8fc7\u7a0b\uff0c\u9632\u6b62\u6a21\u578b\u5229\u7528\u6377\u5f84\u6216\u8005\u4fee\u6539\u5956\u52b1\u51fd\u6570\u83b7\u53d6\u975e\u9884\u671f\u7684\u9ad8\u5956\u52b1\u3002<\/li>\n<\/ol>\n<p>\u901a\u8fc7\u7ed3\u5408\u8fd9\u4e9b\u63aa\u65bd\uff0c\u672a\u6765\u7684 AI \u7cfb\u7edf\u624d\u80fd\u5728\u590d\u6742\u73af\u5883\u4e2d\u8868\u73b0\u5f97\u66f4\u52a0\u5b89\u5168\u548c\u53ef\u9760\uff0c\u907f\u514d\u56e0\u5956\u52b1\u653b\u9677\u5e26\u6765\u7684\u6f5c\u5728\u98ce\u9669\u548c\u4e0d\u826f\u540e\u679c\u3002<\/p>\n<p><strong>\u4e09\u3001\u672a\u6765\u65b9\u5411\u7684\u5c55\u671b<\/strong><\/p>\n<p><strong>\u5f3a\u5316\u5b66\u4e60\u7684\u91cd\u8981\u6027<\/strong><\/p>\n<p>OpenAI o1 \u7684\u53d1\u5e03\u5c06\u91cd\u5851\u884c\u4e1a\u5bf9\u4e8e\u7b97\u529b\u5206\u914d\u7684\u8ba4\u77e5\uff0c\u6807\u5fd7\u7740 RL \u4e0b Post-Training \u00a0Scaling Law \u7684\u65f6\u4ee3\u6b63\u5f0f\u5230\u6765\u3002OpenAI \u7814\u7a76\u5458 Jason Wei \u4e5f\u8868\u793a\uff0co1 \u6a21\u578b\u80cc\u540e\u7684\u6838\u5fc3\u4e0d\u53ea\u662f\u901a\u8fc7 Prompt \u63d0\u793a\u8bcd\u5b8c\u6210 CoT\uff0c\u800c\u662f\u5f15\u5165 RL \u8bad\u7ec3\u6a21\u578b\uff0c\u4ece\u800c\u4f7f\u6a21\u578b\u66f4\u597d\u5730\u6267\u884c\u94fe\u5f0f\u601d\u8003\u3002\u9690\u5f0f\u601d\u7ef4\u94fe\u601d\u8003\u7ed9 o1 \u5e26\u6765\u7684\u5de8\u5927\u6027\u80fd\u63d0\u5347\uff0c\u4e5f\u5c06\u542f\u53d1\u884c\u4e1a\u5728\u6a21\u578b\u89c4\u6a21\u8fbe\u5230\u4e00\u5b9a\u91cf\u7ea7\u540e\uff0c\u66f4\u591a\u7684\u5c06\u7b97\u529b\u6295\u5165\u5230 Post-Training \u9636\u6bb5\u7684 RL \u8bad\u7ec3\u548c\u63a8\u7406\u9636\u6bb5\u6a21\u578b\u7684\u601d\u8003\u8fc7\u7a0b\u5f53\u4e2d\u3002\u5f3a\u5316\u5b66\u4e60\u5148\u9a71 Rich Sutton \u5728 \u201cThe Bitter Lesson\u201d \u4e2d\u8bf4\uff1a<\/p>\n<section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"183\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\n<p>One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great.\u00a0<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"69\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\n<p>The two methods that seem to scale arbitrarily in this way are search and learning.<\/p>\n<\/section>\n<\/section>\n<\/blockquote>\n<\/section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"75\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>\n<section>\n<p>We want AI agents that can discover like we can, not which contain what we have discovered.\u00a0<\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/blockquote>\n<p><strong>\u6211\u4eec\u5e0c\u671b AI \u4ee3\u7406\u80fd\u591f\u50cf\u6211\u4eec\u4e00\u6837\u8fdb\u884c\u63a2\u7d22\uff0c\u800c\u4e0d\u662f\u4ec5\u4ec5\u5305\u542b\u6211\u4eec\u5df2\u7ecf\u53d1\u73b0\u7684\u77e5\u8bc6<\/strong>\uff08\u901a\u8fc7 Pre-training \u6765\u8ba9\u6a21\u578b\u62df\u5408\u6d77\u91cf\u7684\u6570\u636e\u5206\u5e03\uff0c\u5e76\u671f\u5f85\u6a21\u578b\u5177\u5907\u4e00\u5b9a\u7684\u6cdb\u5316\u6027\uff09\u3002<strong>\u53ea\u6709\u641c\u7d22\u548c\u5b66\u4e60\u8fd9\u4e24\u79cd\u5b66\u4e60\u8303\u5f0f\u80fd\u591f\u968f\u7740\u8ba1\u7b97\u80fd\u529b\u7684\u589e\u957f\u65e0\u9650\u6269\u5c55<\/strong>\uff0c\u5f3a\u5316\u5b66\u4e60\u4f5c\u4e3a\u8fd9\u4e24\u79cd\u5b66\u4e60\u8303\u5f0f\u7684\u8f7d\u4f53\uff0c\u5982\u4f55\u80fd\u591f\u5728\u5b9e\u73b0\u53ef\u6269\u5c55\u7684 RL \u5b66\u4e60\uff08Scalable RL Learning\uff09\u548c\u5f3a\u5316\u5b66\u4e60\u6269\u5c55\u6cd5\u5219\uff08RL Scaling Law\uff09\uff0c\u5c06\u6210\u4e3a\u8fdb\u4e00\u6b65\u7a81\u7834\u5927\u6a21\u578b\u6027\u80fd\u4e0a\u9650\u7684\u5173\u952e\u9014\u5f84\u3002<\/p>\n<p><strong>Test-Time \u7b97\u6cd5\u8bbe\u8ba1\u7684\u91cd\u8981\u6027<\/strong><\/p>\n<p>\u4ece\u6280\u672f\u4e0a\u6765\u770b\uff1a\u5982\u4f55\u66f4\u597d\u7684\u542f\u53d1\u6a21\u578b\u7684\u5185\u90e8\u601d\u8003\u8fc7\u7a0b\uff1f\u4e0d\u540c\u7684\u601d\u8003\u65b9\u5f0f\u548c Self-Critique \u53cd\u9988\u65b9\u5f0f\u5e26\u6765\u7684\u6210\u6548\u4e0d\u540c \uff1bPost-Training \u9636\u6bb5\u5982\u4f55\u901a\u8fc7 RL \u624b\u6bb5\u5e2e\u52a9\u6a21\u578b\u5b66\u4f1a\u5185\u5728\u7684\u63a8\u7406\uff1fVerifier  Critic Model \u5982\u4f55\u548c\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5\u5982 RL-Self Play \u3001MCTS \u8fdb\u884c\u6709\u6548\u7684\u8026\u5408\uff0c\u5b9e\u73b0\u771f\u6b63\u7684\u53ef\u6269\u5c55\u76d1\u7763\uff08Scalable Oversight\uff09\uff1b\u6a21\u578b\u601d\u8003\u8fc7\u7a0b\u7684\u4e0a\u754c\u5728\u54ea\uff0cTest-Time Optimization \u7684\u8fb9\u754c\uff1b\u8fd9\u4e9b\u90fd\u5c06\u662f\u4e0d\u5c40\u9650\u4e8e\u5229\u7528\u4eba\u7c7b\u504f\u597d\u6570\u636e\uff0c\u5229\u7528\u5f3a\u5316\u5b66\u4e60\u8fdb\u884c \u201c\u5fae\u201d \u8c03\uff0c\u968f\u7740\u4e1a\u754c\u5728 pre-train \u9636\u6bb5\u6280\u672f\u7684\u6210\u719f\uff0c\u4ee5\u53ca\u73b0\u6709\u6570\u636e\u4e5f\u5373\u5c06\u532e\u4e4f\uff0c\u5927\u6a21\u578b\u7684\u8bad\u7ec3\u5c06\u4ece \u201c\u6a21\u4eff\u201d \u5230 \u201c\u63a2\u7d22\u201d \u7684\u8f6c\u53d8\uff0c\u5728 Llama 3.1 \u7684\u6280\u672f\u62a5\u544a\u4e2d\u8c08\u5230\uff0c\u5728 Post-training \u9636\u6bb5\u80fd\u591f\u89c2\u5bdf\u5230\u6027\u80fd\u7684\u6709\u6548\u63d0\u5347\uff0c\u800c OpenAI O1 \u5219\u662f\u8fdb\u4e00\u6b65\u5c06\u5927\u6a21\u578b\u8bad\u7ec3\u7684\u6218\u573a\u63a8\u5411 Post-training \u9636\u6bb5\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-cd623ff52da7ae37a86f1581a128fa7a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-cd623ff52da7ae37a86f1581a128fa7a.png\"><\/a><\/p>\n<p><strong>AI \u63a7\u5236\u9700\u8981\u5212\u6e05 RL \u7684\u804c\u6743\u754c\u9650<\/strong><\/p>\n<p>AI \u63a7\u5236\uff08AI Control\uff09\u662f Greenblatt et al. \u63d0\u51fa\u7684\u4e00\u5957 AI \u5b89\u5168\u8303\u5f0f [8] \uff0c\u8fd1\u6765\u5728 AGI \u5b89\u5168\u9886\u57df\u83b7\u5f97\u4e86\u4e0d\u5c11\u5173\u6ce8\u3002\u5176\u4e3b\u8981\u7279\u5f81\u662f\u5047\u5b9a\u6a21\u578b\u5177\u6709\u672a\u5bf9\u9f50\u3001\u975e\u5b89\u5168\u7684\u503e\u5411\uff0c\u5e76\u8bbe\u8ba1\u90e8\u7f72\u9636\u6bb5\u7684 \u201c\u5de5\u4e1a\u6d41\u7a0b\u201d \u4f7f\u4e0d\u540c\u80fd\u529b\u3001\u4e0d\u540c\u5b89\u5168\u7cfb\u6570\u7684\u6a21\u578b\u4e92\u76f8\u8f85\u52a9\u3001\u4e92\u76f8\u76d1\u7763\uff0c\u4ee5\u5728\u4e00\u5b9a\u7684\u5b89\u5168\u6027\u6982\u7387\u4fdd\u969c\u4e0b\u6709\u6548\u5229\u7528\u975e\u5b89\u5168\u7684\u6a21\u578b\u3002<\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-c5baa591b0f0ae0a1e5e6379a94cf812.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-c5baa591b0f0ae0a1e5e6379a94cf812.png\"><\/a><\/p>\n<p>AI \u63a7\u5236\u8303\u5f0f\u662f\u4e00\u7cfb\u5217<strong>\u63a8\u7406\u9636\u6bb5 AI \u5b89\u5168\u65b9\u6cd5<\/strong>\u7684\u4ee3\u8868\u3002\u8fd9\u4e9b\u65b9\u6cd5\u5f80\u5f80\u9700\u8981\u6a21\u578b\u65e0\u6cd5\u901a\u8fc7\u68af\u5ea6\u66f4\u65b0\u6765\u627e\u5230\u5e76\u5229\u7528\u76d1\u7763\u673a\u5236\u4e2d\u7684\u6f0f\u6d1e\uff0c\u56e0\u4e3a\u53cd\u4e4b\u5219\u51e0\u4e4e\u4efb\u4f55\u76d1\u7763\u673a\u5236\u90fd\u53ef\u627e\u5230\u6f0f\u6d1e\uff08\u8fd9\u4e00\u4e8b\u5b9e\u5df2\u7ecf\u5728\u4eba\u7c7b\u793e\u4f1a\u4e2d\u5f97\u5230\u5e7f\u6cdb\u5370\u8bc1\uff09\u3002<\/p>\n<p>\u57fa\u4e8e RL \u7684\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u65b9\u6cd5\uff0c\u5728\u63d0\u9ad8\u6a21\u578b\u80fd\u529b\u7684\u540c\u65f6\uff0c\u4e5f\u4e3a\u6a21\u578b\u5e26\u6765\u4e86\u5bfb\u627e\u76d1\u7763\u673a\u5236\u6f0f\u6d1e\u7684\u5f3a\u5927\u52a8\u673a \u2014\u2014 \u5728\u5f53\u4e0b\u548c\u8fc7\u53bb\u7684 RL agent \u7814\u7a76\u4e2d\uff0c\u7814\u7a76\u8005\u7684\u8bb8\u591a\u7cbe\u529b\u90fd\u82b1\u5728\u907f\u514d RL policy \u6536\u655b\u5230\u610f\u6599\u4e4b\u5916\u7684 \u201c\u4f5c\u5f0a\u201d \u7b56\u7565\u4e0a\uff1b\u5982\u679c\u80fd\u529b\u5f3a\u5927\u3001\u4e14\u5177\u6709 agent \u67b6\u6784\u7684\u8bed\u8a00\u6a21\u578b\u540c\u6837\u4ea7\u751f\u4e86\u8fd9\u4e9b \u201c\u4f5c\u5f0a\u201d \u503e\u5411\uff0c\u4f1a\u5e26\u6765\u4e0d\u540c\u4e8e\u8f93\u51fa\u6bd2\u6027\u7684\u4e00\u7cfb\u5217\u65b0\u7684\u5b89\u5168\u98ce\u9669\u3002AI \u5b89\u5168\u9886\u57df\u56f4\u7ed5\u8fd9\u4e9b\u98ce\u9669\u8fdb\u884c\u4e86\u5e7f\u6cdb\u7684\u8ba8\u8bba\u548c\u7814\u7a76\uff0c\u53ef\u4ee5\u53c2\u8003\u56f4\u7ed5\u5956\u52b1\u7834\u89e3\uff08reward hacking\uff09\u548c\u76ee\u6807\u8bef\u6307\uff08goal misspecification\uff09\u7b49\u4e3b\u9898\u7684\u6587\u732e\uff0c\u5728 AI \u5bf9\u9f50\u7efc\u8ff0 [10] \u4e2d\uff0c\u4e5f\u6709\u6240\u5206\u6790\u3002<\/p>\n<p>\u8fd9\u4e00\u8003\u8651\u610f\u5473\u7740\uff0c\u5728\u8bed\u8a00\u6a21\u578b\uff08\u5c24\u5176\u662f\u5177\u6709 agent \u67b6\u6784\u7684\u8bed\u8a00\u6a21\u578b\uff09\u7684 RL \u8bad\u7ec3\u4e2d\uff0c\u5e94\u5f53\u4e25\u683c\u5212\u5206 \u201cRL \u8bad\u7ec3\u56de\u8def\u5185\u201d \u548c \u201cRL \u8bad\u7ec3\u56de\u8def\u5916\u201d \u7684\u90e8\u5206\uff0c\u5e76\u7528\u53ef\u9760\u7684\u6280\u672f\u63aa\u65bd\uff08\u5982\u6c99\u76d2\uff09\u3001\u6d41\u7a0b\u63aa\u65bd\uff08\u5982\u907f\u514d\u6839\u636e\u56de\u8def\u5916\u7684\u4fe1\u606f\u6765\u7b5b\u9009\u6a21\u578b\uff09\u6765\u9694\u79bb\u4e8c\u8005\u3002\u56de\u8def\u5185\u90e8\u5206\u4f8b\u5982\u65f6\u7a7a\u5c3a\u5ea6\u6709\u9650\u7684\u53d7\u8bad\u4efb\u52a1\u672c\u8eab\uff0c\u800c\u56de\u8def\u5916\u90e8\u5206\u4f8b\u5982\u5728\u6a21\u578b\u5468\u56f4\u90e8\u7f72\u7684 AI \u5b89\u5168\u673a\u5236\uff0c\u5305\u62ec\u524d\u8ff0\u7684 AI \u63a7\u5236\u65b9\u6cd5\u3002\u82e5\u5141\u8bb8 RL \u76f4\u63a5\u5bf9\u6297 AI \u5b89\u5168\u673a\u5236\u4f5c\u5956\u52b1\u6700\u5927\u5316\uff0c\u5219\u53ef\u80fd\u5e26\u6765 \u201c\u94bb\u5b89\u5168\u6f0f\u6d1e\u201d \u7684\u98ce\u9669\u3002<\/p>\n<p><strong>\u4f5c\u8005\u4fe1\u606f<\/strong><\/p>\n<p>\u9648\u535a\u8fdc\uff0c\u5317\u4eac\u5927\u5b66\u5143\u57f9\u5b66\u9662 2022 \u7ea7 \u201c\u901a\u73ed\u201d \u672c\u79d1\u751f\uff0c\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u4e3a\uff1a\u5927\u6a21\u578b\u53ef\u6269\u5c55\u76d1\u7763\uff0c\u66fe\u83b7 2024 \u5546\u6c64\u5956\u5b66\u91d1\u548c\u5317\u4eac\u5e02\u81ea\u7136\u79d1\u5b66\u57fa\u91d1\u8d44\u52a9\uff0c\u4e2a\u4eba\u4e3b\u9875 cby-pku.github.io\uff1b<\/p>\n<p>\u90b1\u5929\u5f02\uff0c\u5317\u4eac\u5927\u5b66\u4fe1\u606f\u79d1\u5b66\u6280\u672f\u5b66\u9662 2022 \u7ea7 \u201c\u56fe\u7075\u73ed\u201d \u672c\u79d1\u751f\uff0c\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u4e3a\uff1a\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u7684\u9053\u5fb7\u8fdb\u6b65\u4e0e\u5bf9\u9f50\uff0c\u66fe\u83b7 John Hopcroft \u5956\u5b66\u91d1\u548c\u56fd\u81ea\u7136\u672c\u79d1\u751f\u57fa\u91d1\u8d44\u52a9\uff0c\u4e2a\u4eba\u4e3b\u9875\u4e3a\uff1atianyiqiu.net\uff1b<\/p>\n<p>\u5409\u5609\u94ed\uff0c\u5317\u4eac\u5927\u5b66\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u9662 2023 \u7ea7\u535a\u58eb\u751f\uff0c\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u4e3a\uff1a\u5927\u6a21\u578b\u5b89\u5168\u5bf9\u9f50\uff0c\u66fe\u83b7\u5317\u4eac\u5927\u5b66\u6821\u957f\u5956\u5b66\u91d1\u548c\u9996\u6279\u56fd\u81ea\u7136\u535a\u58eb\u751f\u57fa\u91d1\u8d44\u52a9\uff0c\u4e2a\u4eba\u4e3b\u9875\u4e3a\uff1ajijiaming.com\u3002<\/p>\n<p>\u6240\u5728\u7684\u5b9e\u9a8c\u5ba4\u4e3a\u5317\u4eac\u5927\u5b66\u5bf9\u9f50\u4e0e\u4ea4\u4e92\u5b9e\u9a8c\u5ba4 PAIR-Lab\uff0cpair-lab.com\uff0c\u5bfc\u5e08\u4e3a\u5317\u4eac\u5927\u5b66\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u9662\u52a9\u7406\u6559\u6388\u6768\u8000\u4e1c\u3002<\/p>\n<p><em>\u53c2\u8003\u6587\u732e<\/em><\/p>\n<p><em>[1] STaR: Bootstrapping Reasoning With Reasoning https:\/\/arxiv.org\/abs\/2203.14465<\/em><\/p>\n<p><em>[2] LLM Critics Help Catch LLM Bugs<\/em><\/p>\n<p><em>https:\/\/arxiv.org\/pdf\/2407.00215<\/em><\/p>\n<p><em>[3] Self-critiquing models for assisting human evaluators<\/em><\/p>\n<p><em>https:\/\/arxiv.org\/pdf\/2206.05802<\/em><\/p>\n<p><em>[4] OpenAI o1 System Card<\/em><\/p>\n<p><em>https:\/\/assets.ctfassets.net\/kftzwdyauwt9\/67qJD51Aur3eIc96iOfeOP\/71551c3d223cd97e591aa89567306912\/o1_system_card.pdf<\/em><\/p>\n<p><em>[5] Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters<\/em><\/p>\n<p><em>https:\/\/arxiv.org\/abs\/2408.03314<\/em><\/p>\n<p><em>[6] Training Verifiers to Solve Math Word Problems https:\/\/arxiv.org\/pdf\/2110.14168<\/em><\/p>\n<p><em>[7] Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking https:\/\/arxiv.org\/abs\/2403.09629<\/em><\/p>\n<p><em>[8] AI Control: Improving Safety Despite Intentional Subversion https:\/\/arxiv.org\/abs\/2312.06942<\/em><\/p>\n<p><em>[9] Sycophancy to Subterfuge: Investigating Reward Tampering in Language Models https:\/\/arxiv.org\/abs\/2406.10162<\/em><\/p>\n<p><em>[10] AI Alignment: A Comprehensive Survey https:\/\/arxiv.org\/abs\/2310.19852<\/em><\/p>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-09-15\" target=\"_blank\">\u5317\u5927\u5bf9\u9f50\u56e2\u961f\u72ec\u5bb6\u89e3\u8bfb\uff1aOpenAI o1\u5f00\u542f\u300c\u540e\u8bad\u7ec3\u300d\u65f6\u4ee3\u5f3a\u5316\u5b66\u4e60\u65b0\u8303\u5f0f<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u5317\u5927\u5bf9\u9f50\u56e2\u961f\u72ec\u5bb6\u89e3\u8bfb\uff1a [&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":[70,68],"class_list":["post-35634","post","type-post","status-publish","format-standard","hentry","category-news","tag-agent","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u5317\u5927\u5bf9\u9f50\u56e2\u961f\u72ec\u5bb6\u89e3\u8bfb\uff1aOpenAI o1\u5f00\u542f\u300c\u540e\u8bad\u7ec3\u300d\u65f6\u4ee3\u5f3a\u5316\u5b66\u4e60\u65b0\u8303\u5f0f - \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=35634\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=35634\",\"url\":\"https:\/\/17aitech.com\/?p=35634\",\"name\":\"\u5317\u5927\u5bf9\u9f50\u56e2\u961f\u72ec\u5bb6\u89e3\u8bfb\uff1aOpenAI o1\u5f00\u542f\u300c\u540e\u8bad\u7ec3\u300d\u65f6\u4ee3\u5f3a\u5316\u5b66\u4e60\u65b0\u8303\u5f0f - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=35634#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=35634#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7188805adf42a7a1f79ddbdbd1851e48.png\",\"datePublished\":\"2024-12-28T09:00:19+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=35634#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=35634\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=35634#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7188805adf42a7a1f79ddbdbd1851e48.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7188805adf42a7a1f79ddbdbd1851e48.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=35634#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u5317\u5927\u5bf9\u9f50\u56e2\u961f\u72ec\u5bb6\u89e3\u8bfb\uff1aOpenAI o1\u5f00\u542f\u300c\u540e\u8bad\u7ec3\u300d\u65f6\u4ee3\u5f3a\u5316\u5b66\u4e60\u65b0\u8303\u5f0f\"}]},{\"@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":"\u5317\u5927\u5bf9\u9f50\u56e2\u961f\u72ec\u5bb6\u89e3\u8bfb\uff1aOpenAI o1\u5f00\u542f\u300c\u540e\u8bad\u7ec3\u300d\u65f6\u4ee3\u5f3a\u5316\u5b66\u4e60\u65b0\u8303\u5f0f - \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=35634","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=35634","url":"https:\/\/17aitech.com\/?p=35634","name":"\u5317\u5927\u5bf9\u9f50\u56e2\u961f\u72ec\u5bb6\u89e3\u8bfb\uff1aOpenAI o1\u5f00\u542f\u300c\u540e\u8bad\u7ec3\u300d\u65f6\u4ee3\u5f3a\u5316\u5b66\u4e60\u65b0\u8303\u5f0f - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=35634#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=35634#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7188805adf42a7a1f79ddbdbd1851e48.png","datePublished":"2024-12-28T09:00:19+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=35634#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=35634"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=35634#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7188805adf42a7a1f79ddbdbd1851e48.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/09\/frc-7188805adf42a7a1f79ddbdbd1851e48.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=35634#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u5317\u5927\u5bf9\u9f50\u56e2\u961f\u72ec\u5bb6\u89e3\u8bfb\uff1aOpenAI o1\u5f00\u542f\u300c\u540e\u8bad\u7ec3\u300d\u65f6\u4ee3\u5f3a\u5316\u5b66\u4e60\u65b0\u8303\u5f0f"}]},{"@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\/35634","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=35634"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/35634\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=35634"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=35634"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=35634"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}