{"id":38019,"date":"2025-02-08T12:01:11","date_gmt":"2025-02-08T04:01:11","guid":{"rendered":"https:\/\/17aitech.com\/?p=38019"},"modified":"2025-02-08T12:01:11","modified_gmt":"2025-02-08T04:01:11","slug":"iclr-2025%ef%bd%9c%e5%a4%a7%e6%a8%a1%e5%9e%8b%e4%b9%9f%e9%9c%80%e8%a6%81%e5%a5%bd%e5%a5%87%e5%bf%83%ef%bc%8cteleai-%e6%8f%90%e5%87%ba%e6%8e%a2%e7%b4%a2%e9%a9%b1%e5%8a%a8%e7%9a%84%e5%af%b9%e9%bd%90","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=38019","title":{"rendered":"ICLR 2025\uff5c\u5927\u6a21\u578b\u4e5f\u9700\u8981\u597d\u5947\u5fc3\uff0cTeleAI \u63d0\u51fa\u63a2\u7d22\u9a71\u52a8\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c8B\u8d8a\u7ea7\u80dc70B"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-01-30-3\" target=\"_blank\">ICLR 2025\uff5c\u5927\u6a21\u578b\u4e5f\u9700\u8981\u597d\u5947\u5fc3\uff0cTeleAI \u63d0\u51fa\u63a2\u7d22\u9a71\u52a8\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c8B\u8d8a\u7ea7\u80dc70B<\/a><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-0ed59df52df4e822f2f40bcc90d9d0e5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-0ed59df52df4e822f2f40bcc90d9d0e5.png\"><\/a><\/p>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"166\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<p>AIxiv\u4e13\u680f\u662f\u673a\u5668\u4e4b\u5fc3\u53d1\u5e03\u5b66\u672f\u3001\u6280\u672f\u5185\u5bb9\u7684\u680f\u76ee\u3002\u8fc7\u53bb\u6570\u5e74\uff0c\u673a\u5668\u4e4b\u5fc3AIxiv\u4e13\u680f\u63a5\u6536\u62a5\u9053\u4e862000\u591a\u7bc7\u5185\u5bb9\uff0c\u8986\u76d6\u5168\u7403\u5404\u5927\u9ad8\u6821\u4e0e\u4f01\u4e1a\u7684\u9876\u7ea7\u5b9e\u9a8c\u5ba4\uff0c\u6709\u6548\u4fc3\u8fdb\u4e86\u5b66\u672f\u4ea4\u6d41\u4e0e\u4f20\u64ad\u3002\u5982\u679c\u60a8\u6709\u4f18\u79c0\u7684\u5de5\u4f5c\u60f3\u8981\u5206\u4eab\uff0c\u6b22\u8fce\u6295\u7a3f\u6216\u8005\u8054\u7cfb\u62a5\u9053\u3002\u6295\u7a3f\u90ae\u7bb1\uff1aliyazhou@jiqizhixin.com\uff1bzhaoyunfeng@jiqizhixin.com<\/p>\n<\/section>\n<\/blockquote>\n<section><\/section>\n<section>\u5728\u81ea\u7136\u754c\u4e2d\uff0c\u597d\u5947\u5fc3\u9a71\u4f7f\u7740\u751f\u7269\u63a2\u7d22\u672a\u77e5\uff0c\u662f\u751f\u5b58\u548c\u8fdb\u5316\u7684\u5173\u952e\u3002\u4eba\u7c7b\uff0c\u4f5c\u4e3a\u5730\u7403\u4e0a\u6700\u5177\u667a\u80fd\u7684\u7269\u79cd\uff0c\u5176\u63a2\u7d22\u7cbe\u795e\u5f15\u9886\u4e86\u79d1\u6280\u3001\u6587\u5316\u548c\u793e\u4f1a\u7684\u8fdb\u6b65\u30021492 \u5e74\uff0c\u54e5\u4f26\u5e03\u6000\u63e3\u63a2\u7d22\u672a\u77e5\u7684\u7406\u60f3\uff0c\u52c7\u6562\u5730\u5411\u897f\u822a\u884c\uff0c\u6700\u7ec8\u53d1\u73b0\u4e86\u65b0\u5927\u9646\u3002<\/section>\n<section><\/section>\n<section>\u6b63\u5982<strong>\u4eba\u7c7b\u5728\u9762\u5bf9\u672a\u77e5\u65f6\u5c55\u73b0\u51fa\u7684\u63a2\u7d22\u884c\u4e3a<\/strong>\uff0c\u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\uff0c\u5c24\u5176\u5728\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7406\u89e3\u8bed\u8a00\u548c\u77e5\u8bc6\u4e2d\uff0c\u7814\u7a76\u4eba\u5458\u6b63\u5c1d\u8bd5<strong>\u8d4b\u4e88 LLM \u7c7b\u4f3c\u7684\u63a2\u7d22\u80fd\u529b\uff0c\u4ece\u800c\u7a81\u7834\u5176\u5728\u7ed9\u5b9a\u6570\u636e\u96c6\u4e2d\u5b66\u4e60\u7684\u80fd\u529b\u8fb9\u754c<\/strong>\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347\u6027\u80fd\u548c\u5b89\u5168\u6027\u3002<\/section>\n<section><strong><br \/><\/strong><\/section>\n<section>\u8fd1\u671f\uff0c<strong>\u4e2d\u56fd\u7535\u4fe1\u96c6\u56e2 CTO\u3001\u9996\u5e2d\u79d1\u5b66\u5bb6\u3001\u4e2d\u56fd\u7535\u4fe1\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u9662\uff08TeleAI\uff09\u9662\u957f\u674e\u5b66\u9f99\u6559\u6388<\/strong>\u5e26\u9886\u56e2\u961f\u5728\u5168\u6a21\u6001\u661f\u8fb0\u5927\u6a21\u578b\u4f53\u7cfb\u6df1\u8015\u7684\u57fa\u7840\u4e4b\u4e0a\uff0c\u8054\u5408\u6e05\u534e\u5927\u5b66\u3001\u9999\u6e2f\u57ce\u5e02\u5927\u5b66\u3001\u4e0a\u6d77\u4eba\u5de5\u667a\u80fd\u5b9e\u9a8c\u5ba4\u7b49\u5355\u4f4d\u63d0\u51fa\u4e86\u4e00\u79cd<strong>\u65b0\u7684\u63a2\u7d22\u9a71\u52a8\u7684\u5927\u6a21\u578b\u5bf9\u9f50\u65b9\u6cd5 Count-based Online Preference Optimization\uff08COPO\uff09\uff0c\u5c06\u4eba\u7c7b\u63a2\u7d22\u7684\u672c\u80fd\u878d\u5165\u5230\u5927\u8bed\u8a00\u6a21\u578b\u7684\u540e\u8bad\u7ec3\uff08Post-Training\uff09\u4e2d<\/strong>\uff0c\u5f15\u5bfc\u6a21\u578b\u5728<strong>\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09<\/strong>\u6846\u67b6\u4e0b\u4e3b\u52a8\u63a2\u7d22\u5c1a\u672a\u5145\u5206\u7406\u89e3\u7684\u77e5\u8bc6\uff0c\u89e3\u51b3\u4e86\u73b0\u6709\u5bf9\u9f50\u6846\u67b6\u53d7\u9650\u4e8e\u504f\u597d\u6570\u636e\u96c6\u8986\u76d6\u8303\u56f4\u7684\u95ee\u9898\u3002<\/section>\n<section><\/section>\n<section><strong>\u8fd9\u4e00\u521b\u65b0\u6210\u679c\u4e3a\u667a\u4f20\u7f51\uff08AI Flow\uff09\u4e2d \u201c\u57fa\u4e8e\u8fde\u63a5\u4e0e\u4ea4\u4e92\u7684\u667a\u80fd\u6d8c\u73b0\u201d \u63d0\u4f9b\u4e86\u91cd\u8981\u6280\u672f\u652f\u6491\uff0c\u4f7f\u5f97\u6a21\u578b\u5728\u52a8\u6001\u4ea4\u4e92\u4e2d\u4e0d\u65ad\u5b66\u4e60\u548c\u8fdb\u6b65\uff0c\u5728\u63a2\u7d22\u7684\u8fc7\u7a0b\u4e2d\u5b9e\u73b0\u667a\u80fd\u7684\u6301\u7eed\u6d8c\u73b0<\/strong>\u3002\u8bba\u6587\u88ab\u56fd\u9645\u8868\u5f81\u5b66\u4e60\u5927\u4f1a ICLR 2025 \u5f55\u7528\uff0c\u5b9e\u73b0\u4e86\u5927\u6a21\u578b\u591a\u8f6e\u4ea4\u4e92\u63a2\u7d22\u4e2d\u7684\u80fd\u529b\u6301\u7eed\u63d0\u5347\u3002TeleAI \u7814\u7a76\u79d1\u5b66\u5bb6\u767d\u8fb0\u7532\u4e3a\u8bba\u6587\u7684\u7b2c\u4e00\u4f5c\u8005\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-7cb6176fc2fc6e159945b5af5b1d4d39.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-7cb6176fc2fc6e159945b5af5b1d4d39.png\"><\/a><\/p>\n<ul>\n<li>\n<section>\u8bba\u6587\u6807\u9898\uff1aOnline Preference Alignment for Language Models via Count-based Exploration<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/abs\/2501.12735<\/section>\n<\/li>\n<li>\n<section>\u5f00\u6e90\u4ee3\u7801\uff1ahttps:\/\/github.com\/Baichenjia\/COPO<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u7814\u7a76\u52a8\u673a<\/strong><\/section>\n<section><\/section>\n<section>\u867d\u7136\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5728\u8fdb\u884c\u591a\u79cd\u8bed\u8a00\u4efb\u52a1\u4e2d\u5df2\u7ecf\u6709\u51fa\u8272\u7684\u8868\u73b0\uff0c\u4f46\u5b83\u4eec\u5728\u4e0e\u4eba\u7c7b\u4ef7\u503c\u89c2\u548c\u610f\u56fe\u5bf9\u9f50\u65b9\u9762\u4ecd\u9762\u4e34\u7740\u5f88\u591a\u6311\u6218\u3002<strong>\u73b0\u6709\u7684\u5927\u6a21\u578b RLHF \u6846\u67b6\u4e3b\u8981\u4f9d\u8d56\u4e8e\u9884\u5148\u6536\u96c6\u7684\u504f\u597d\u6570\u636e\u96c6\u8fdb\u884c\u5bf9\u9f50\uff0c\u5176\u6027\u80fd\u53d7\u9650\u4e8e\u79bb\u7ebf\u504f\u597d\u6570\u636e\u96c6\u5bf9\u63d0\u793a &#8211; \u56de\u590d\uff08Prompt-Response\uff09\u7684\u8986\u76d6\u8303\u56f4<\/strong>\uff0c\u5bf9\u6570\u636e\u96c6\u8986\u76d6\u4e4b\u5916\u7684\u8bed\u8a00\u96be\u4ee5\u8fdb\u884c\u6709\u6548\u6cdb\u5316\u3002\u7136\u800c\uff0c\u4eba\u7c7b\u504f\u597d\u6570\u636e\u96c6\u7684\u6536\u96c6\u662f\u8f83\u4e3a\u6602\u8d35\u7684\uff0c\u4e14\u73b0\u6709\u7684\u504f\u597d\u6570\u636e\u96be\u4ee5\u8986\u76d6\u6240\u6709\u53ef\u80fd\u7684\u63d0\u793a\u548c\u56de\u590d\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u5c31\u5f15\u51fa\u4e86\u4e00\u4e2a\u5173\u952e\u95ee\u9898\uff1a<strong>\u662f\u5426\u53ef\u4ee5\u4f7f LLM \u5728\u5bf9\u9f50\u8fc7\u7a0b\u4e2d\u5bf9\u8bed\u8a00\u7a7a\u95f4\u8fdb\u884c\u81ea\u4e3b\u63a2\u7d22\uff0c\u4ece\u800c\u7a81\u7834\u79bb\u7ebf\u6570\u636e\u96c6\u7684\u7ea6\u675f\uff0c\u4e0d\u65ad\u63d0\u5347\u6cdb\u5316\u80fd\u529b<\/strong>\uff1f<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u8fd1\u671f\u7684\u5927\u6a21\u578b\u76f8\u5173\u7814\u7a76\u5f00\u59cb\u7531<strong>\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\u9a71\u52a8\u7684\u79bb\u7ebf\u5bf9\u9f50\uff08Offline RLHF\uff09\u8f6c\u5411\u5728\u7ebf\u5bf9\u9f50\uff08Online RLHF\uff09<\/strong>\uff0c\u901a\u8fc7\u8fed\u4ee3\u5f0f\u5730\u6536\u96c6\u63d0\u793a\u548c\u56de\u590d\uff0c\u5141\u8bb8\u5927\u6a21\u578b\u5728\u4e0e\u8bed\u8a00\u73af\u5883\u7684\u4e92\u52a8\u4e2d\u4e0d\u65ad\u5b66\u4e60\u548c\u8fdb\u6b65\uff0c\u4ece\u800c\u5728\u504f\u597d\u6570\u636e\u96c6\u7684\u8986\u76d6\u4e4b\u5916\u8fdb\u884c\u63a2\u7d22\u3002<\/section>\n<section><\/section>\n<section>\u672c\u7814\u7a76\u65e8\u5728<strong>\u89e3\u51b3\u5728\u7ebf RLHF \u8fc7\u7a0b\u4e2d\u7684\u6838\u5fc3\u95ee\u9898\uff1a\u5982\u4f55\u4f7f LLM \u9ad8\u6548\u5728\u8bed\u8a00\u7a7a\u95f4\uff08\u7c7b\u6bd4\u4e8e\u5f3a\u5316\u5b66\u4e60\u52a8\u4f5c\u7a7a\u95f4\uff09\u4e2d\u8fdb\u884c\u63a2\u7d22<\/strong>\u3002\u5177\u4f53\u5730\uff0c\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u5728\u8fdb\u884c\u5927\u89c4\u6a21\u7684\u72b6\u6001\u52a8\u4f5c\u7a7a\u95f4\uff08\u7c7b\u6bd4\u4e8e LLM \u4e2d\u7684\u8bed\u8a00\u751f\u6210\u7a7a\u95f4\uff09\u4e2d\u7684\u6700\u4f18\u7b56\u7565\u6c42\u89e3\u65f6\uff0c<strong>\u7cfb\u7edf\u6027\u63a2\u7d22\uff08Systematic Exploration\uff09\u5bf9\u4e8e\u6536\u96c6\u6709\u76ca\u7684\u7ecf\u9a8c\u81f3\u5173\u91cd\u8981\uff0c\u4f1a\u76f4\u63a5\u5173\u7cfb\u5230\u7b56\u7565\u5b66\u4e60\u7684\u6548\u679c<\/strong>\u3002\u5728 LLM \u5bf9\u9f50\u4e2d\uff0c\u5982\u679c\u7f3a\u4e4f\u6709\u6548\u7684\u63a2\u7d22\u673a\u5236\uff0c\u53ef\u80fd\u4f1a\u5bfc\u81f4\u6a21\u578b\u5bf9\u9f50\u9677\u5165\u5c40\u90e8\u6700\u4f18\u7b56\u7565\uff1b\u540c\u65f6\uff0c\u6709\u6548\u7684\u63a2\u7d22\u53ef\u4ee5\u5e2e\u52a9\u5927\u6a21\u578b\u66f4\u597d\u5730\u7406\u89e3\u8bed\u8a00\u73af\u5883\u7684\u77e5\u8bc6\uff0c\u4ece\u800c\u5728\u5e7f\u9614\u7684\u8bed\u8a00\u7a7a\u95f4\u4e2d\u627e\u5230\u6700\u4f18\u56de\u590d\u7b56\u7565\u3002<\/section>\n<section><\/section>\n<section>\u672c\u7814\u7a76\u7684\u76ee\u6807\u5728\u4e8e\u89e3\u51b3\u5728\u7ebf RLHF \u4e2d\u7684\u63a2\u7d22\u95ee\u9898\uff0c\u5373\u5982\u4f55\u5728\u6bcf\u6b21\u8fed\u4ee3\u4e2d\u6709\u6548\u5730\u63a2\u7d22\u63d0\u793a &#8211; \u56de\u590d\u7a7a\u95f4\uff0c\u4ee5\u6269\u5927\u504f\u597d\u6570\u636e\u8986\u76d6\u8303\u56f4\uff0c\u63d0\u9ad8\u6a21\u578b\u5bf9\u4eba\u7c7b\u504f\u597d\u7684\u5b66\u4e60\u548c\u9002\u5e94\u80fd\u529b\u3002\u5177\u4f53\u5730\uff0c<strong>COPO \u7b97\u6cd5\u901a\u8fc7\u7ed3\u5408\u57fa\u4e8e\u8ba1\u6570\u7684\u63a2\u7d22\uff08Count-based Exploration\uff09\u548c\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff09\u6846\u67b6\uff0c\u5229\u7528\u4e00\u4e2a\u8f7b\u91cf\u7ea7\u7684\u4f2a\u8ba1\u6570\u6a21\u5757\u6765\u5e73\u8861\u63a2\u7d22\u548c\u504f\u597d\u4f18\u5316\uff0c\u5e76\u5728\u7ebf\u6027\u5956\u52b1\u51fd\u6570\u8fd1\u4f3c\u548c\u79bb\u6563\u72b6\u6001\u7a7a\u95f4\u4e2d\u63d0\u4f9b\u4e86\u7406\u8bba\u6846\u67b6<\/strong>\u3002<\/section>\n<section><\/section>\n<section>\u5b9e\u9a8c\u4e2d\uff0c\u5728 Zephyr \u548c Llama-3 \u6a21\u578b\u4e0a\u8fdb\u884c\u7684 RLHF \u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cCOPO \u5728\u6307\u4ee4\u9075\u5faa\u548c\u5b66\u672f\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u6027\u80fd\u4f18\u4e8e\u5176\u4ed6 RLHF \u57fa\u7ebf\u3002<\/section>\n<section><\/section>\n<section><strong>\u7406\u8bba\u6846\u67b6<\/strong><\/section>\n<section><\/section>\n<section>\u7814\u7a76\u7684\u7406\u8bba\u6846\u67b6\u57fa\u4e8e\u5927\u6a21\u578b\u5956\u52b1\u7684\u7ebf\u6027\u5047\u8bbe\uff0c<strong>\u5c06\u5956\u52b1\u51fd\u6570\u7b80\u5316\u4e3a\u53c2\u6570\u5411\u91cf\u548c\u7279\u5f81\u5411\u91cf\u7684\u5185\u79ef\u5f62\u5f0f<\/strong>\u3002\u5728\u6b64\u5047\u8bbe\u4e0b\uff0c\u53ef\u4ee5\u5c06\u590d\u6742\u5927\u6a21\u578b\u5bf9\u8bed\u8a00\u63d0\u53d6\u7684\u7279\u5f81\u4f5c\u4e3a\u4e00\u4e2a\u4f4e\u7ef4\u7684\u5411\u91cf\uff0c\u5c06 RLHF \u8fc7\u7a0b\u4e2d\u6784\u5efa\u7684\u663e\u5f0f\u6216\u9690\u5f0f\u7684\u5927\u6a21\u578b\u5956\u52b1\u89c6\u4e3a\u5411\u91cf\u7684\u7ebf\u6027\u51fd\u6570\uff0c\u5177\u4f53\u5730\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-3fd954a0deb4d434694c028ba31ef222.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-3fd954a0deb4d434694c028ba31ef222.png\"><\/a><\/section>\n<section>\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u7ed9\u5b9a\u5927\u6a21\u578b\u504f\u597d\u6570\u636e\u96c6<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-1bfea19b118b2afc4e9f9070fd0ac9c8.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-1bfea19b118b2afc4e9f9070fd0ac9c8.png\"><\/a>\uff0c\u5728\u73b0\u6709 Bradley-Terry (BT) \u5956\u52b1\u6a21\u578b\u7684\u57fa\u7840\u4e0a\u53ef\u4ee5\u901a\u8fc7\u6781\u5927\u4f3c\u7136\u4f30\u8ba1\u6765\u4f30\u8ba1\u5956\u52b1\u6a21\u578b\u7684\u53c2\u6570\uff0c\u5373\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-900c131a10f11ffc72a295db2aa50be9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-900c131a10f11ffc72a295db2aa50be9.png\"><\/a><\/section>\n<section>\u968f\u540e\uff0c\u6839\u636e\u7edf\u8ba1\u5b66\u4e2d\u7684\u76f8\u5173\u7406\u8bba\uff0c<strong>\u53ef\u4ee5\u5b9a\u91cf\u5730\u4e3a\u5956\u52b1\u6a21\u578b\u63d0\u4f9b\u4e86\u4e00\u4e2a\u660e\u786e\u7684\u8bef\u5dee\u754c\u9650\uff0c\u5e76\u5f97\u5230\u5173\u4e8e\u5956\u52b1\u6a21\u578b\u53c2\u6570\u7684\u7f6e\u4fe1\u96c6\u5408\uff08confidence set\uff09<\/strong>\uff0c\u4ece\u800c\u4f7f\u4f30\u8ba1\u7684\u53c2\u6570\u4ee5\u8f83\u5927\u6982\u7387\u843d\u5728\u7f6e\u4fe1\u96c6\u5408\u4e2d\u3002\u5177\u4f53\u5730\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-6c4efd26a22206f0505de6ffdcf689c1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-6c4efd26a22206f0505de6ffdcf689c1.png\"><\/a><\/section>\n<section>\u968f\u540e\uff0c\u5728\u53c2\u6570\u96c6\u5408\u4e2d\u53ef\u4ee5\u4f7f\u7528<strong>\u4e50\u89c2\u7684\u671f\u671b\u503c\u51fd\u6570\u6765\u83b7\u5f97\u503c\u51fd\u6570\u4f30\u8ba1\u7684\u7f6e\u4fe1\u4e0a\u754c\uff0c\u4ece\u800c\u5b9e\u73b0\u4e86\u5f3a\u5316\u5b66\u4e60\u63a2\u4e2d\u7684\u4e50\u89c2\u539f\u5219\uff08Optimism\uff09<\/strong>, \u4f7f\u5927\u6a21\u578b\u7b56\u7565\u5411\u4e50\u89c2\u65b9\u5411\u8fdb\u884c\u7b56\u7565\u4f18\u5316\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-11d8731b2f4910be5a9bd92982ffaaee.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-11d8731b2f4910be5a9bd92982ffaaee.png\"><\/a><\/section>\n<section>\u5728\u4e0a\u8ff0\u76ee\u6807\u4e2d\uff0c\u6700\u7ec8\u7684\u4f18\u5316\u9879\u5305\u542b\u4e24\u4e2a\u90e8\u5206\uff1a\u7b2c\u4e00\u90e8\u5206\u5bf9\u5e94\u4e8e\u7ecf\u5178\u7684\u4e24\u9636\u6bb5 RLHF \u65b9\u6cd5\uff0c\u5728 BT \u6a21\u578b\u7684\u57fa\u7840\u4e0a\u4f30\u8ba1\u5956\u52b1\uff0c\u901a\u8fc7\u6700\u5927\u5316\u5956\u52b1\u6765\u5b66\u4e60\u7b56\u7565\uff0c\u540c\u65f6\u4fdd\u6301\u548c\u539f\u59cb\u5927\u6a21\u578b\u7b56\u7565\u7684\u63a5\u8fd1\u6027\u7ea6\u675f\u3002\u7b2c\u4e8c\u90e8\u5206\u4e3a\u65b0\u5f15\u5165\u7684\u7f6e\u4fe1\u533a\u95f4\u4e0a\u754c\uff08UCB\uff09\u9879\uff0c\u7528\u4e8e\u6d4b\u91cf\u5f53\u524d\u6570\u636e\u96c6\u5bf9\u76ee\u6807\u7b56\u7565\u751f\u6210\u7684\u72b6\u6001\u5206\u5e03\u7684\u8986\u76d6\u7a0b\u5ea6\uff0c\u9f13\u52b1\u6a21\u578b\u63a2\u7d22\u90a3\u4e9b\u5c1a\u672a\u5145\u5206\u63a2\u7d22\u7684\u8bed\u8a00\u7a7a\u95f4\u3002<\/section>\n<section><\/section>\n<section>\u5177\u4f53\u6765\u8bf4\uff0c<strong>UCB \u9879\u901a\u8fc7\u589e\u52a0\u5bf9\u8f83\u5c11\u4ea7\u751f\u7684\u63d0\u95ee &#8211; \u56de\u7b54\u7684\u7ec4\u5408\u7684\u5bf9\u6570\u4f3c\u7136\uff0c\u4ece\u800c\u9f13\u52b1\u5927\u6a21\u578b\u751f\u6210\u65b0\u7684\u3001\u53ef\u80fd\u66f4\u4f18\u7684\u56de\u7b54<\/strong>\u3002\u8fd9\u5c06\u6709\u52a9\u4e8e\u5927\u6a21\u578b\u5728\u6700\u5927\u5316\u5956\u52b1\u548c\u63a2\u7d22\u65b0\u54cd\u5e94\u4e4b\u95f4\u7684\u6743\u8861\uff0c\u5373\u8457\u540d\u7684<strong>\u5f3a\u5316\u5b66\u4e60\u63a2\u7d22 &#8211; \u5229\u7528\u6743\u8861\uff08exploration-exploitation trade-off\uff09<\/strong>\u3002<\/section>\n<section><\/section>\n<section>\u6700\u7ec8\uff0c\u7814\u7a76\u8bc1\u660e\u4e86\u91c7\u7528 COPO \u7b97\u6cd5\u7684\u5728\u7ebf\u5b66\u4e60\u8303\u5f0f\u80fd\u591f\u5728 T \u6b21\u8fed\u4ee3\u540e\uff0c\u5c06<strong>\u603b\u540e\u6094\u503c\u9650\u5236\u5728 O (\u221aT) \u7684\u91cf\u7ea7\u5185<\/strong>\uff0c\u663e\u793a\u4e86\u7b97\u6cd5\u5728\u5904\u7406\u5927\u89c4\u6a21\u72b6\u6001\u7a7a\u95f4\u65f6\u7684\u6548\u7387\u548c\u7a33\u5b9a\u6027\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-953151d92ca69b32026a3a78098824a8.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-953151d92ca69b32026a3a78098824a8.png\"><\/a><\/section>\n<section><strong>\u7b97\u6cd5\u8bbe\u8ba1<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u7406\u8bba\u6846\u67b6\u4e0b\uff0c\u5177\u4f53\u7684\u7b97\u6cd5\u8bbe\u8ba1\u4e2d\u7ed3\u5408\u4e86\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff09\u7684\u7b97\u6cd5\u6846\u67b6\u3002\u5176\u4e2d\u7b2c\u4e00\u9879\u5bf9\u5956\u52b1\u7684\u6784\u5efa\u548c\u5956\u52b1\u6700\u5927\u5316\u7684\u5b66\u4e60\u5177\u8c61\u5316\u4e3a DPO \u7684\u5b66\u4e60\u76ee\u6807\uff0c\u800c\u5c06\u4e50\u89c2\u63a2\u7d22\u7684 UCB \u9879\u8f6c\u5316\u4e3a\u66f4\u5bb9\u6613\u6c42\u89e3\u7684\u76ee\u6807\u3002\u5177\u4f53\u5730\uff0c<strong>\u5728\u6709\u9650\u72b6\u6001\u52a8\u4f5c\u7a7a\u95f4\u7684\u5047\u8bbe\u4e0b\uff0c\u4e50\u89c2\u63a2\u7d22\u9879\u53ef\u4ee5\u8868\u793a\u4e3a\u57fa\u4e8e\u72b6\u6001 &#8211; \u52a8\u4f5c\u8ba1\u6570\uff08Count\uff09\u7684\u5b66\u4e60\u76ee\u6807<\/strong>\uff0c\u5373<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-aca97c19be88666795ddff4a41bd9997.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-aca97c19be88666795ddff4a41bd9997.png\"><\/a><\/section>\n<section>\u4ece\u800c\uff0c\u6700\u7ec8\u7684\u5b66\u4e60\u76ee\u6807\u8868\u793a\u4e3a DPO \u5956\u52b1\u548c\u57fa\u4e8e\u63d0\u793a &#8211; \u56de\u7b54\u8ba1\u6570\u7684\u63a2\u7d22\u76ee\u6807\u3002\u5177\u4f53\u5730\uff0c<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-160f435ca44cd6ba9f04aaa678e20683.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-160f435ca44cd6ba9f04aaa678e20683.png\"><\/a><\/section>\n<section>\u4e0a\u5f0f\u4e2d\u7b2c\u4e8c\u9879\u901a\u8fc7\u5728\u504f\u597d\u6570\u636e\u4e2d\u5bf9\u6a21\u578b\u4ea7\u751f\u7684\u63d0\u793a &#8211; \u56de\u7b54\u8fdb\u884c\u8ba1\u6570\uff0c\u53ef\u4ee5<strong>\u9f13\u52b1\u589e\u52a0\u5bf9\u4e4b\u524d\u51fa\u73b0\u6b21\u6570\u8f83\u5c11\u7684\u63d0\u793a &#8211; \u56de\u7b54\u7684\u63a2\u7d22\u6765\u9f13\u52b1\u5927\u6a21\u578b\u7a81\u7834\u79bb\u7ebf\u6570\u636e\u96c6\u7684\u8986\u76d6\uff0c\u4f7f\u6a21\u578b\u4e3b\u52a8\u63a2\u7d22\u65b0\u7684\u3001\u53ef\u80fd\u66f4\u4f18\u7684\u56de\u590d<\/strong>\uff0c\u4ece\u800c\u5728\u8fed\u4ee3\u8fc7\u7a0b\u4e2d\u6269\u5927\u6570\u636e\u8986\u76d6\u8303\u56f4\u5e76\u63d0\u9ad8\u7b56\u7565\u7684\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section>\u8fdb\u800c\u53ef\u4ee5\u901a\u8fc7\u6c42\u89e3\u68af\u5ea6\u7684\u65b9\u5f0f\u8fdb\u4e00\u6b65\u7684\u89e3\u6790 COPO \u4f18\u5316\u76ee\u6807\u7684\u610f\u4e49\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-ce83c080ac02038a94f2bfa3d3913189.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-ce83c080ac02038a94f2bfa3d3913189.png\"><\/a><\/section>\n<section>\u7531\u4e24\u90e8\u5206\u7ec4\u6210\uff1a\u7b2c\u4e00\u90e8\u5206\u8d1f\u8d23\u4f18\u5316\u6a21\u578b\u4ee5\u6700\u5927\u5316\u504f\u597d\u6570\u636e\u4e0a\u7684\u9884\u671f\u5956\u52b1\uff1b\u7b2c\u4e8c\u90e8\u5206\u5bf9\u5e94\u4e8e\u63a2\u7d22\u9879\u7684\u68af\u5ea6\uff0c\u5b83\u6839\u636e\u63d0\u793a &#8211; \u56de\u590d\u5bf9\u7684\u5386\u53f2\u8bbf\u95ee\u6b21\u6570\u6765\u8c03\u6574\u6a21\u578b\u7684\u4f18\u5316\u65b9\u5411\u3002\u5f53\u67d0\u4e2a\u56de\u590d\u7684\u5386\u53f2\u8bbf\u95ee\u6b21\u6570\u8f83\u5c11\u65f6\uff0c\u8be5\u9879\u4f1a\u63a8\u52a8\u6a21\u578b\u589e\u52a0\u751f\u6210\u8be5\u56de\u590d\u7684\u5bf9\u6570\u4f3c\u7136\uff0c\u4ece\u800c\u9f13\u52b1\u6a21\u578b\u63a2\u7d22\u90a3\u4e9b\u8f83\u5c11\u88ab\u8bbf\u95ee\u4f46\u53ef\u80fd\u5e26\u6765\u66f4\u9ad8\u5956\u52b1\u7684\u533a\u57df\uff0c\u4f7f\u7b97\u6cd5\u80fd\u591f<strong>\u5728\u6700\u5927\u5316\u5956\u52b1\u7684\u540c\u65f6\u6709\u6548\u5730\u5e73\u8861\u63a2\u7d22\u4e0e\u5229\u7528\uff0c\u5b9e\u73b0\u66f4\u4f18\u7684\u7b56\u7565\u5b66\u4e60<\/strong>\u3002<\/section>\n<section><\/section>\n<section>\u7136\u800c\uff0c\u5728\u5bf9\u5927\u6a21\u578b\u8fdb\u884c\u4e0a\u8ff0\u76ee\u6807\u4f18\u5316\u4e2d\uff0c\u5f80\u5f80\u65e0\u6cd5\u5728\u5927\u89c4\u6a21\u8bed\u8a00\u7a7a\u95f4\u4e2d\u5b9e\u73b0\u5bf9 \u201c\u63d0\u793a &#8211; \u56de\u590d\u201d \u7684\u51c6\u786e \u201c\u8ba1\u6570\u201d\u3002\u8bed\u8a00\u7a7a\u95f4\u7684\u72b6\u6001\u901a\u5e38\u662f\u65e0\u9650\u7684\uff0c\u4e14\u5b8c\u5168\u76f8\u540c\u7684\u56de\u590d\u5f88\u5c11\u88ab\u591a\u6b21\u4ea7\u751f\uff0c\u56e0\u6b64\u9700\u8981\u4e00\u79cd\u65b9\u6cd5\u6765\u4f30\u8ba1\u6216\u6a21\u62df\u8fd9\u4e9b\u63d0\u793a &#8211; \u56de\u590d\u5bf9\u7684 \u201c\u4f2a\u8ba1\u6570\u201d\uff0c\u4ee5\u4fbf\u7b97\u6cd5\u80fd\u591f\u5728\u63a2\u7d22\u8f83\u5c11\u8bbf\u95ee\u7684\u533a\u57df\u65f6\u83b7\u5f97\u6fc0\u52b1\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0c<strong>COPO \u63d0\u51fa\u4f7f\u7528 Coin Flipping Network\uff08CFN\uff09\u6765\u9ad8\u6548\u7684\u5b9e\u73b0\u4f2a\u8ba1\u6570<\/strong>\u3002CFN \u4e0d\u4f9d\u8d56\u4e8e\u590d\u6742\u7684\u5bc6\u5ea6\u4f30\u8ba1\u6216\u5bf9\u6a21\u578b\u67b6\u6784\u548c\u8bad\u7ec3\u8fc7\u7a0b\u7684\u9650\u5236\uff0c\u800c\u662f\u901a\u8fc7\u4e00\u4e2a\u7b80\u5355\u7684\u56de\u5f52\u95ee\u9898\u6765\u9884\u6d4b\u57fa\u4e8e\u8ba1\u6570\u7684\u63a2\u7d22\u5956\u52b1\u3002<\/section>\n<section><\/section>\n<section>\u5177\u4f53\u5730\uff0c<strong>CFN \u57fa\u4e8e\u7684\u57fa\u672c\u5047\u8bbe\u662f\uff0c\u8ba1\u6570\u53ef\u4ee5\u901a\u8fc7\u4ece Rademacher \u5206\u5e03\u7684\u91c7\u6837\u6765\u4f30\u8ba1\u6765\u5f97\u5230<\/strong>\uff0c\u8003\u8651\u4ece {-1,1} \u7684\u96c6\u5408\u4e2d\u8fd1\u4f3c\u968f\u673a\u91c7\u6837\u5f97\u5230\u7684\u5206\u5e03\uff0c\u5982\u679c\u8fdb\u884c n \u6b21\u91c7\u6837\u5e76\u5bf9\u91c7\u6837\u7ed3\u679c\u53d6\u5e73\u5747\uff0c\u5219\u8be5\u53d8\u91cf\u7684\u4e8c\u9636\u77e9\u548c\u8ba1\u6570\u7684\u5012\u6570\u5448\u73b0\u51fa\u7b49\u4ef7\u7684\u5173\u7cfb\uff0c\u5373<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-1b3a21d13d5941b28dc305827033c753.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-1b3a21d13d5941b28dc305827033c753.png\"><\/a><\/section>\n<section>\u8fdb\u800c\uff0c<strong>CFN \u901a\u8fc7\u5728\u6bcf\u6b21\u9047\u5230\u72b6\u6001\u65f6\u8fdb\u884c Rademacher \u8bd5\u9a8c\uff08\u5373\u786c\u5e01\u7ffb\u8f6c\uff09\uff0c\u5e76\u5229\u7528\u8fd9\u4e9b\u8bd5\u9a8c\u7684\u5e73\u5747\u503c\u6765\u63a8\u65ad\u72b6\u6001\u7684\u8bbf\u95ee\u9891\u7387<\/strong>\u3002\u5728\u5b9e\u73b0\u4e2d\uff0cCFN \u8868\u793a\u4e3a\u4e00\u4e2a\u8f7b\u91cf\u5316\u7684\u7f51\u7edc\uff0c\u5b83\u901a\u8fc7\u6700\u5c0f\u5316\u9884\u6d4b\u503c\u548c\u5b9e\u9645 Rademacher \u6807\u7b7e\u4e4b\u95f4\u7684\u5747\u65b9\u8bef\u5dee\u6765\u8fdb\u884c\u8bad\u7ec3\u3002<\/section>\n<section><\/section>\n<section>\u5728\u5b9e\u73b0\u4e2d\uff0c<strong>CFN \u63a5\u53d7\u7531\u4e3b\u8bed\u8a00\u6a21\u578b\u63d0\u53d6\u7684\u63d0\u793a &#8211; \u56de\u590d\u5bf9\u7684\u6700\u540e\u9690\u85cf\u72b6\u6001\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u8f93\u51fa\u4e00\u4e2a\u9884\u6d4b\u503c\uff0c\u8be5\u503c\u4e0e\u72b6\u6001\u7684 \u201c\u4f2a\u8ba1\u6570\u201d \u6210\u53cd\u6bd4<\/strong>\u3002\u901a\u8fc7\u8fd9\u79cd\u65b9\u5f0f\uff0cCFN \u80fd\u591f\u4e3a\u6bcf\u4e2a\u63d0\u793a &#8211; \u54cd\u5e94\u63d0\u4f9b\u4e00\u4e2a\u63a2\u7d22\u6fc0\u52b1\uff0c\u9f13\u52b1\u6a21\u578b\u5728\u63a2\u7d22\u8fed\u4ee3\u4e2d\u6269\u5927\u6570\u636e\u8986\u76d6\u8303\u56f4\uff0c\u63d0\u9ad8\u6a21\u578b\u5bf9\u9f50\u7684\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section><strong>\u5b9e\u9a8c\u7ed3\u679c<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u5b9e\u9a8c\u4e2d\u4f7f\u7528 UltraFeedback 60K \u504f\u597d\u6570\u636e\u96c6\u6765\u5bf9 Zephyr-7B \u548c Llama3-8B \u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u6570\u636e\u96c6\u4e2d\u5305\u542b\u4e30\u5bcc\u7684\u5355\u8f6e\u5bf9\u8bdd\u504f\u597d\u5bf9\u7684\u6570\u636e\u3002\u5b9e\u9a8c\u4e2d<strong>\u4f7f\u7528\u4e86\u4e00\u4e2a\u5c0f\u578b\u7684\u5956\u52b1\u6a21\u578b PairRM 0.4B \u6765\u5bf9\u591a\u8f6e\u8fed\u4ee3\u8fc7\u7a0b\u4e2d\u6a21\u578b\u6a21\u578b\u751f\u6210\u7684\u56de\u590d\u8fdb\u884c\u504f\u597d\u6392\u5e8f<\/strong>\uff0c\u4ece\u800c\u5728\u63a2\u7d22\u4e2d\u5229\u7528\u4e0d\u65ad\u66f4\u65b0\u540e\u7684\u5927\u6a21\u578b\u6765\u4ea7\u751f\u4e0d\u65ad\u6269\u5145\u7684\u504f\u597d\u6570\u636e\uff0c\u63d0\u5347\u4e86\u6570\u636e\u96c6\u7684\u8d28\u91cf\u548c\u8986\u76d6\u7387\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c<strong>\u5b9e\u9a8c\u4e2d\u4f7f\u7528\u8f7b\u91cf\u5316\u7684 CFN \u7f51\u7edc\u5b9e\u73b0\u5bf9\u63d0\u793a &#8211; \u54cd\u5e94\u5bf9\u7684\u4f2a\u8ba1\u6570<\/strong>\uff0c\u5927\u5e45\u63d0\u5347\u4e86\u5728\u7ebf RLHF \u7b97\u6cd5\u7684\u63a2\u7d22\u80fd\u529b\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-22c83b64aceb6f31e78a2ddb593ec668.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-22c83b64aceb6f31e78a2ddb593ec668.png\"><\/a><\/section>\n<section>\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cCOPO \u7b97\u6cd5\u5728<strong>\u00a0AlpacaEval 2.0 \u548c MT-Bench \u57fa\u51c6\u6d4b\u8bd5<\/strong>\u53ef\u4ee5\u901a\u8fc7\u591a\u8f6e\u63a2\u7d22\u548c\u5bf9\u9f50\u6765\u4e0d\u65ad\u8fdb\u884c\u6027\u80fd\u63d0\u5347\u3002\u5177\u4f53\u5730\uff0c\u76f8\u6bd4\u4e8e\u79bb\u7ebf DPO \u7b97\u6cd5\uff0c<strong>COPO \u663e\u8457\u63d0\u5347\u4e86 Zephyr-7B \u548c Llama3-8B \u6a21\u578b\u7684 LC \u80dc\u7387<\/strong>\uff0c\u5206\u522b\u8fbe\u5230\u4e86 18.8% \u548c 7.1% \u7684\u63d0\u5347\uff0c\u9a8c\u8bc1\u4e86 LLM \u63a2\u7d22\u80fd\u529b\u63d0\u5347\u5bf9\u83b7\u53d6\u66f4\u5927\u6570\u636e\u8986\u76d6\u548c\u6700\u4f18\u7b56\u7565\u65b9\u9762\u7684\u4f18\u52bf\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0cCOPO \u8d85\u8d8a\u4e86\u5728\u7ebf DPO\u3001SELM \u7b49\u5f53\u524d\u6700\u597d\u7684\u5728\u7ebf\u5bf9\u9f50\u65b9\u6cd5\uff0c\u4ee5\u00a0<strong>8B \u7684\u6a21\u578b\u5bb9\u91cf\u8d85\u8d8a\u4e86\u8bb8\u591a\u5927\u4f53\u91cf\u6a21\u578b\uff08\u5982 Yi-34B\uff0cLlama3-70B\uff09\u7684\u6027\u80fd<\/strong>\uff0c\u63d0\u5347\u4e86\u5927\u6a21\u578b\u5728\u8bed\u8a00\u4efb\u52a1\u4e2d\u7684\u6307\u4ee4\u8ddf\u968f\u80fd\u529b\u548c\u6cdb\u5316\u80fd\u529b\u3002<\/section>\n<section><\/section>\n<section><strong>\u56e2\u961f\u8d1f\u8d23\u4eba\u4ecb\u7ecd\uff1a\u674e\u5b66\u9f99\uff0c\u4e2d\u56fd\u7535\u4fe1\u96c6\u56e2 CTO\u3001\u9996\u5e2d\u79d1\u5b66\u5bb6\u3001\u4e2d\u56fd\u7535\u4fe1\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u9662\uff08TeleAI\uff09\u9662\u957f\uff0c\u4e3b\u8981\u5173\u6ce8\u5927\u6a21\u578b\u3001\u667a\u80fd\u5149\u7535\u3001\u4e34\u5730\u5b89\u9632\u3001\u667a\u4f20\u7f51\uff08AI Flow\uff09<\/strong>\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-aa1169a4733511069eeccf6bafec806c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-aa1169a4733511069eeccf6bafec806c.png\"><\/a><\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-01-30-3\" target=\"_blank\">ICLR 2025\uff5c\u5927\u6a21\u578b\u4e5f\u9700\u8981\u597d\u5947\u5fc3\uff0cTeleAI \u63d0\u51fa\u63a2\u7d22\u9a71\u52a8\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c8B\u8d8a\u7ea7\u80dc70B<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:ICLR 2025\uff5c\u5927 [&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":[68],"class_list":["post-38019","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>ICLR 2025\uff5c\u5927\u6a21\u578b\u4e5f\u9700\u8981\u597d\u5947\u5fc3\uff0cTeleAI \u63d0\u51fa\u63a2\u7d22\u9a71\u52a8\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c8B\u8d8a\u7ea7\u80dc70B - \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=38019\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=38019#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=38019\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"ICLR 2025\uff5c\u5927\u6a21\u578b\u4e5f\u9700\u8981\u597d\u5947\u5fc3\uff0cTeleAI \u63d0\u51fa\u63a2\u7d22\u9a71\u52a8\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c8B\u8d8a\u7ea7\u80dc70B\",\"datePublished\":\"2025-02-08T04:01:11+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=38019\"},\"wordCount\":165,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=38019#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/01\\\/frc-0ed59df52df4e822f2f40bcc90d9d0e5.png\",\"keywords\":[\"\u5927\u6a21\u578b\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=38019#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=38019\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=38019\",\"name\":\"ICLR 2025\uff5c\u5927\u6a21\u578b\u4e5f\u9700\u8981\u597d\u5947\u5fc3\uff0cTeleAI \u63d0\u51fa\u63a2\u7d22\u9a71\u52a8\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c8B\u8d8a\u7ea7\u80dc70B - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=38019#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=38019#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/01\\\/frc-0ed59df52df4e822f2f40bcc90d9d0e5.png\",\"datePublished\":\"2025-02-08T04:01:11+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=38019#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=38019\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=38019#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/01\\\/frc-0ed59df52df4e822f2f40bcc90d9d0e5.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2025\\\/01\\\/frc-0ed59df52df4e822f2f40bcc90d9d0e5.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=38019#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"ICLR 2025\uff5c\u5927\u6a21\u578b\u4e5f\u9700\u8981\u597d\u5947\u5fc3\uff0cTeleAI \u63d0\u51fa\u63a2\u7d22\u9a71\u52a8\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c8B\u8d8a\u7ea7\u80dc70B\"}]},{\"@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\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"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":"ICLR 2025\uff5c\u5927\u6a21\u578b\u4e5f\u9700\u8981\u597d\u5947\u5fc3\uff0cTeleAI \u63d0\u51fa\u63a2\u7d22\u9a71\u52a8\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c8B\u8d8a\u7ea7\u80dc70B - \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=38019","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=38019#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=38019"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"ICLR 2025\uff5c\u5927\u6a21\u578b\u4e5f\u9700\u8981\u597d\u5947\u5fc3\uff0cTeleAI \u63d0\u51fa\u63a2\u7d22\u9a71\u52a8\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c8B\u8d8a\u7ea7\u80dc70B","datePublished":"2025-02-08T04:01:11+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=38019"},"wordCount":165,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=38019#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-0ed59df52df4e822f2f40bcc90d9d0e5.png","keywords":["\u5927\u6a21\u578b"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=38019#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=38019","url":"https:\/\/17aitech.com\/?p=38019","name":"ICLR 2025\uff5c\u5927\u6a21\u578b\u4e5f\u9700\u8981\u597d\u5947\u5fc3\uff0cTeleAI \u63d0\u51fa\u63a2\u7d22\u9a71\u52a8\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c8B\u8d8a\u7ea7\u80dc70B - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=38019#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=38019#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-0ed59df52df4e822f2f40bcc90d9d0e5.png","datePublished":"2025-02-08T04:01:11+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=38019#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=38019"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=38019#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-0ed59df52df4e822f2f40bcc90d9d0e5.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/01\/frc-0ed59df52df4e822f2f40bcc90d9d0e5.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=38019#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"ICLR 2025\uff5c\u5927\u6a21\u578b\u4e5f\u9700\u8981\u597d\u5947\u5fc3\uff0cTeleAI \u63d0\u51fa\u63a2\u7d22\u9a71\u52a8\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c8B\u8d8a\u7ea7\u80dc70B"}]},{"@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":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","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\/38019","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=38019"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/38019\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=38019"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=38019"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=38019"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}