{"id":34677,"date":"2024-12-10T23:57:07","date_gmt":"2024-12-10T15:57:07","guid":{"rendered":"https:\/\/17aitech.com\/?p=34677"},"modified":"2024-12-10T23:57:07","modified_gmt":"2024-12-10T15:57:07","slug":"icml-2024-oral-dpo%e6%98%af%e5%90%a6%e6%af%94ppo%e6%9b%b4%e9%80%82%e5%90%88llm%ef%bc%8c%e6%b8%85%e5%8d%8e%e5%90%b4%e7%bf%bc%e5%9b%a2%e9%98%9f%e6%9c%80%e6%96%b0%e6%8f%ad%e7%a7%98-2","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=34677","title":{"rendered":"ICML 2024 Oral | DPO\u662f\u5426\u6bd4PPO\u66f4\u9002\u5408LLM\uff0c\u6e05\u534e\u5434\u7ffc\u56e2\u961f\u6700\u65b0\u63ed\u79d8"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-22-7\" target=\"_blank\">ICML 2024 Oral | DPO\u662f\u5426\u6bd4PPO\u66f4\u9002\u5408LLM\uff0c\u6e05\u534e\u5434\u7ffc\u56e2\u961f\u6700\u65b0\u63ed\u79d8<\/a><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7f7e81ff0ab737b24ac6e6afc20a56f2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7f7e81ff0ab737b24ac6e6afc20a56f2.png\"><\/a><\/section>\n<blockquote data-author-name=\"\" data-content-utf8-length=\"166\" data-source-title=\"\" data-type=\"2\" data-url=\"\">\n<section>\n<section>AIxiv\u4e13\u680f\u662f<mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>\u53d1\u5e03\u5b66\u672f\u3001\u6280\u672f\u5185\u5bb9\u7684\u680f\u76ee\u3002\u8fc7\u53bb\u6570\u5e74\uff0c<mark data-type=\"institutions\" data-id=\"9e96434f-5827-41e3-9b3c-36d39bc0d446\">\u673a\u5668\u4e4b\u5fc3<\/mark>AIxiv\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<\/section>\n<\/section>\n<\/blockquote>\n<section><\/section>\n<section><strong><mark data-type=\"experts\" data-id=\"25b47a57-a97f-41b8-9115-c3043ad9c925\">\u5434\u7ffc<\/mark>\uff0c<mark data-type=\"institutions\" data-id=\"4349d6b9-40ff-4e23-a237-f9ace91c6c81\">\u6e05\u534e\u5927\u5b66<\/mark>\u4ea4\u53c9\u4fe1\u606f\u9662\u52a9\u7406\u6559\u6388\uff0c\u66fe\u4efb OpenAI \u5168\u804c\u7814\u7a76\u5458\uff0c\u7814\u7a76\u9886\u57df\u4e3a<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\uff0c\u5927\u6a21\u578b\u5bf9\u9f50\uff0c<mark data-type=\"tech_tasks\" data-id=\"f6e8725e-6ffd-494f-91d1-ba8214e76434\">\u4eba\u673a\u4ea4\u4e92<\/mark>\uff0c\u673a\u5668\u4eba\u5b66\u4e60\u7b49\u30022019 \u5e74\u5728\u7f8e\u56fd\u52a0\u5dde\u5927\u5b66\u4f2f\u514b\u5229\u5206\u6821\u83b7\u5f97\u535a\u58eb\u5b66\u4f4d\uff0c\u5e08\u4ece\u00a0Stuart Russell \u6559\u6388\uff1b2014 \u5e74\u672c\u79d1\u6bd5\u4e1a\u4e8e<mark data-type=\"institutions\" data-id=\"4349d6b9-40ff-4e23-a237-f9ace91c6c81\">\u6e05\u534e\u5927\u5b66<\/mark>\u4ea4\u53c9\u4fe1\u606f\u9662\uff08\u59da\u73ed\uff09\u3002\u5176\u4ee3\u8868\u4f5c\u5305\u62ec\uff1aNIPS2016 \u6700\u4f73\u8bba\u6587\uff0cValue Iteration Network\uff1b\u591a\u667a\u80fd\u4f53\u6df1\u5ea6<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u9886\u57df\u6700\u9ad8\u5f15\u7528\u8bba\u6587\uff0cMADDPG \u7b97\u6cd5\uff1bOpenAI hide-and-seek \u9879\u76ee\u7b49\u3002<\/strong><\/section>\n<section><\/section>\n<section>\u5982\u4f55\u8ba9\u5927\u6a21\u578b\u66f4\u597d\u7684\u9075\u4ece\u4eba\u7c7b\u6307\u4ee4\u548c\u610f\u56fe\uff1f\u5982\u4f55\u8ba9\u5927\u6a21\u578b\u6709\u66f4\u597d\u7684\u63a8\u7406\u80fd\u529b\uff1f\u5982\u4f55\u8ba9\u5927\u6a21\u578b\u907f\u514d\u5e7b\u89c9\uff1f\u80fd\u5426\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u662f\u8ba9\u5927\u6a21\u578b\u771f\u6b63\u5e7f\u6cdb\u53ef\u7528\uff0c\u751a\u81f3\u5b9e\u73b0\u8d85\u7ea7\u667a\u80fd\uff08Super Intelligence\uff09\u6700\u4e3a\u5173\u952e\u7684\u6280\u672f\u6311\u6218\u3002\u8fd9\u4e9b\u6700\u56f0\u96be\u7684\u6311\u6218\u4e5f\u662f<mark data-type=\"experts\" data-id=\"25b47a57-a97f-41b8-9115-c3043ad9c925\">\u5434\u7ffc<\/mark>\u56e2\u961f\u957f\u671f\u4ee5\u6765\u7684\u7814\u7a76\u91cd\u70b9\uff0c\u5927\u6a21\u578b\u5bf9\u9f50\u6280\u672f\uff08Alignment\uff09\u6240\u8981\u653b\u514b\u7684\u96be\u9898\u3002<\/section>\n<section><\/section>\n<section>\u5bf9\u9f50\u6280\u672f\u4e2d\uff0c\u6700\u91cd\u8981\u7684\u7b97\u6cd5\u6846\u67b6\u5c31\u662f\u6839\u636e\u4eba\u7c7b\u53cd\u9988\u7684<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\uff08RLHF, Reinforcement Learning from Human Feedback\uff09\u3002RLHF \u6839\u636e\u4eba\u7c7b\u5bf9\u5927\u6a21\u578b\u8f93\u51fa\u7684\u504f\u597d\u53cd\u9988\uff0c\u6765\u5b66\u4e60\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5956\u52b1\u51fd\u6570\uff08Reward Model\uff09\uff0c\u5e76\u8fdb\u4e00\u6b65\u5bf9\u5927\u6a21\u578b\u8fdb\u884c<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u8bad\u7ec3\uff0c\u8ba9\u5927\u6a21\u578b\u5728\u53cd\u590d\u8fed\u4ee3\u4e2d\u5b66\u4f1a\u8fa8\u522b\u56de\u590d\u7684\u597d\u574f\uff0c\u5e76\u5b9e\u73b0\u6a21\u578b\u80fd\u529b\u63d0\u5347\u3002\u76ee\u524d\u4e16\u754c\u4e0a\u6700\u5f3a\u7684<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\uff0c\u6bd4\u5982 OpenAI \u7684 GPT \u6a21\u578b\u548c Anthropic \u7684 Claude \u6a21\u578b\uff0c\u90fd\u6781\u5176\u5f3a\u8c03 RLHF \u8bad\u7ec3\u7684\u91cd\u8981\u6027\u3002OpenAI \u548c Anthropic \u5185\u90e8\u4e5f\u90fd\u5f00\u53d1\u4e86\u57fa\u4e8e\u5927\u89c4\u6a21 PPO \u7b97\u6cd5\u7684 RLHF \u8bad\u7ec3\u7cfb\u7edf\u8fdb\u884c\u5927\u6a21\u578b\u5bf9\u9f50\u3002<\/section>\n<section><\/section>\n<section>\u7136\u800c\uff0c\u7531\u4e8e PPO \u7b97\u6cd5\u6d41\u7a0b\u590d\u6742\uff0c\u7b97\u529b\u6d88\u8017\u5927\uff0c\u7f8e\u56fd AI \u516c\u53f8\u7684\u5927\u89c4\u6a21 RLHF \u8bad\u7ec3\u7cfb\u7edf\u4e5f\u4ece\u4e0d\u5f00\u6e90\uff0c\u6240\u4ee5\u5c3d\u7ba1 PPO \u7b97\u6cd5\u975e\u5e38\u5f3a\u5927\uff0c\u5b66\u672f\u754c\u7684\u5bf9\u9f50\u5de5\u4f5c\u5374\u4e00\u76f4\u5f88\u5c11\u91c7\u7528\u590d\u6742\u7684 PPO \u7b97\u6cd5\u8fdb\u884c RLHF \u7814\u7a76\uff0c\u8f6c\u800c\u666e\u904d\u4f7f\u7528 SFT\uff08\u76d1\u7763\u5fae\u8c03\uff09\u6216\u8005 DPO\uff08Direct Policy Optimization\uff09\u7b49\u66f4\u7b80\u5316\u3001\u66f4\u76f4\u63a5\u3001\u5bf9\u8bad\u7ec3\u7cfb\u7edf\u8981\u6c42\u66f4\u4f4e\u7684\u5bf9\u9f50\u7b97\u6cd5\u3002<\/section>\n<section><\/section>\n<section>\u90a3\u4e48\uff0c\u7b80\u5355\u7684\u5bf9\u9f50\u7b97\u6cd5\u4e00\u5b9a\u6548\u679c\u66f4\u597d\u5417\uff1f<mark data-type=\"experts\" data-id=\"25b47a57-a97f-41b8-9115-c3043ad9c925\">\u5434\u7ffc<\/mark>\u56e2\u961f\u53d1\u8868\u5728 ICML 2024 \u7684\u5de5\u4f5c \u201cIs DPO Superior to PPO for LLM Alignment? A Comprehensive Study\u201d \u4ed4\u7ec6\u63a2\u8ba8\u4e86 DPO \u4e0e PPO \u7b97\u6cd5\u7684\u7279\u70b9\uff0c\u5e76\u6307\u51fa\u63d0\u5347 RLHF \u7b97\u6cd5\u6548\u679c\u7684\u5173\u952e\u70b9\u3002\u5728\u8be5\u5de5\u4f5c\u4e2d\uff0c<mark data-type=\"experts\" data-id=\"25b47a57-a97f-41b8-9115-c3043ad9c925\">\u5434\u7ffc<\/mark>\u56e2\u961f\u57fa\u4e8e\u81ea\u7814\u7684\u5927\u89c4\u6a21 RLHF \u8bad\u7ec3\u7cfb\u7edf\uff0c\u9996\u6b21\u91c7\u7528 PPO \u7b97\u6cd5\u53ca<mark data-type=\"concepts\" data-id=\"2e982b73-88e2-41e8-a430-f7ae5a9af4bf\">\u53c2\u6570<\/mark>\u91cf\u66f4\u5c11\u7684\u5f00\u6e90\u6a21\u578b\uff0c\u5728\u516c\u8ba4\u6700\u56f0\u96be\u7684\u6311\u6218 \u2014\u2014 \u4ee3\u7801\u751f\u6210\u4efb\u52a1 CodeContest\u2014\u2014 \u4e0a\u8d85\u8fc7\u95ed\u6e90\u5927\u6a21\u578b AlphaCode 41B\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-e35119b0b354deca7f404bc1d9f31839.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-e35119b0b354deca7f404bc1d9f31839.png\"><\/a><\/p>\n<section>\u76f8\u5173\u6210\u679c\u88ab ICML 2024 \u5f55\u7528\u4e3a Oral Presentation\uff0c\u5e76\u5c06\u5728 7 \u6708 23 \u65e5\u4e8e ICML 2024 \u7b2c\u4e00\u4e2a Oral session Alignment-1 \u4e0a\u548c OpenAI\u3001Anthropic \u7b49\u77e5\u540d\u673a\u6784\u7684\u5de5\u4f5c\u4e00\u8d77\u8fdb\u884c\u516c\u5f00\u6c47\u62a5\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8d4d466ba6ae7bd6f36c762a3dae3a0d.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-8d4d466ba6ae7bd6f36c762a3dae3a0d.png\"><\/a><\/p>\n<section>\u63a5\u4e0b\u6765\u8ba9\u6211\u4eec\u5bf9\u6bd4\u4e00\u4e0b GPT-4 \u548c\u7ecf\u8fc7 PPO \u7b97\u6cd5\u8bad\u7ec3\u7684 CodeLlama 34B \u6a21\u578b\u5728\u4ee3\u7801\u751f\u6210\u4e0a\u7684\u6548\u679c\uff0c\u5728\u4f8b\u5b50 1 \u4e2d\uff0c\u7ecf\u8fc7 PPO \u7b97\u6cd5\u8bad\u7ec3\u7684 CodeLlama 34B \u6a21\u578b\u4e0e GPT-4 \u6a21\u578b\u751f\u6210\u4e86\u8d28\u91cf\u76f8\u5f53\u7684\u4ee3\u7801\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ceabb560c1f90aa1f87c8b24e35f1609.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ceabb560c1f90aa1f87c8b24e35f1609.png\"><\/a><\/p>\n<section>\u5728\u793a\u4f8b 2 \u4e2d\uff0c\u53ef\u4ee5\u770b\u5230\u7ecf\u8fc7 PPO \u7b97\u6cd5\u8bad\u7ec3\u7684 CodeLlama 34B \u6a21\u578b\u4e0e GPT-4 \u6a21\u578b\u90fd\u80fd\u751f\u6210\u5b8c\u6574\u5e76\u4e14\u53ef\u8fd0\u884c\u7684 python \u4ee3\u7801\u3002\u7136\u800c\uff0c\u5728\u8fd9\u4e2a\u4f8b\u5b50\u4e0b\uff0cGPT-4 \u751f\u6210\u4e86\u9519\u8bef\u7684\u4ee3\u7801\uff0c\u5728\u6d4b\u8bd5\u6570\u636e\u4e0a\u65e0\u6cd5\u6b63\u786e\u8f93\u51fa\u3002\u800c\u7ecf\u8fc7 PPO \u7b97\u6cd5\u8bad\u7ec3\u7684 CodeLlama 34B \u6a21\u578b\u751f\u6210\u7684\u4ee3\u7801\u53ef\u4ee5\u901a\u8fc7\u6d4b\u8bd5\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-13670e20459248c2f04605a592defc92.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-13670e20459248c2f04605a592defc92.png\"><\/a><\/p>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7c08f1d60d5c7788dbb7f337d90c500e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7c08f1d60d5c7788dbb7f337d90c500e.png\"><\/a><\/p>\n<section>\u5728 ICML 2024 \u7684\u8fd9\u7bc7\u8bba\u6587\u4e2d\uff0c\u7814\u7a76\u56e2\u961f\u8be6\u7ec6\u63a2\u8ba8\u4e86 DPO \u4e0e PPO \u7b97\u6cd5\u7684\u7279\u70b9\uff0c\u5e76\u6307\u51fa\u63d0\u5347 DPO \u548c PPO \u80fd\u529b\u7684\u5173\u952e\u70b9\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-76d3c7d4f9c6108f4140ed1131789ee5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-76d3c7d4f9c6108f4140ed1131789ee5.png\"><\/a><\/p>\n<ul>\n<li>\n<section>\u8bba\u6587\u6807\u9898\uff1aIs DPO Superior to PPO for LLM Alignment? A Comprehensive Study<\/section>\n<\/li>\n<li>\n<section>\u8bba\u6587\u5730\u5740\uff1ahttps:\/\/arxiv.org\/pdf\/2404.10719<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>DPO \u7b97\u6cd5\u7684\u5c40\u9650\u6027<\/strong><\/section>\n<section><\/section>\n<section>\u76f8\u6bd4\u4e8e PPO\uff0cDPO \u4f7f\u7528\u79bb\u7ebf\u6570\u636e\u800c\u975e\u5728\u7ebf\u91c7\u6837\u6570\u636e\u8bad\u7ec3\u3002\u7ecf\u5206\u6790\uff0cDPO \u7b97\u6cd5\u4f1a\u5bfc\u81f4\u8bad\u7ec3\u51fa\u7684\u6a21\u578b\u5bf9\u8bad\u7ec3\u6570\u636e\u5206\u5e03\u4e4b\u5916\u7684\u8f93\u51fa\u4ea7\u751f\u504f\u597d\uff0c\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\u4ea7\u751f\u4e0d\u53ef\u9884\u6599\u7684\u56de\u590d\u3002<\/section>\n<section><\/section>\n<section>\u4e8e\u662f\uff0c\u4e3a\u4e86\u63d0\u5347 DPO \u7b97\u6cd5\u7684\u80fd\u529b\uff0c\u7814\u7a76\u56e2\u961f\u603b\u7ed3\u4e86\u4e24\u4e2a\u5173\u952e\u6280\u672f\uff1a<strong>\u5728 RLHF \u8bad\u7ec3\u524d\u8fdb\u884c\u989d\u5916\u7684 SFT \u8bad\u7ec3\uff0c\u4ee5\u53ca\u4f7f\u7528\u5728\u7ebf\u91c7\u6837\u6570\u636e\u800c\u975e\u79bb\u7ebf\u6570\u636e<\/strong>\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ff17b4ca0cb5520e691d6628459ee564.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-ff17b4ca0cb5520e691d6628459ee564.png\"><\/a><\/p>\n<section>\u5b9e\u9a8c\u8868\u660e\uff0c\u4f7f\u7528\u989d\u5916\u7684 SFT \u8bad\u7ec3\u53ef\u4ee5\u4f7f base \u6a21\u578b\u4ee5\u53ca reference \u6a21\u578b\u66f4\u504f\u5411\u4e8e\u6570\u636e\u96c6\u5185\u7684\u5206\u5e03\uff0c\u5927\u5927\u63d0\u5347 DPO \u7b97\u6cd5\u6548\u679c\uff1b\u53e6\u4e00\u65b9\u9762\uff0c\u4f7f\u7528\u5728\u7ebf\u91c7\u6837\u6570\u636e\u8fdb\u884c\u8fed\u4ee3\u8bad\u7ec3\u7684 DPO \u7b97\u6cd5\u53ef\u4ee5\u5f97\u5230\u7a33\u6b65\u63d0\u5347\uff0c\u8868\u73b0\u8fdc\u8fdc\u4f18\u4e8e\u57fa\u7840\u7684 DPO \u7b97\u6cd5\u3002<\/section>\n<section><\/section>\n<section><strong>PPO \u7b97\u6cd5\u7684\u5173\u952e\u70b9<\/strong><\/section>\n<section><\/section>\n<section>\u9664\u53bb DPO\uff0c\u8bba\u6587\u4e2d\u4e5f\u603b\u7ed3\u4e86\u53d1\u6325 PPO \u6700\u5927\u80fd\u529b\u7684\u4e09\u4e2a\u5173\u952e\u70b9\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u4f7f\u7528\u5927\u7684\u6279\u5927\u5c0f\uff08large batch size\uff09<\/section>\n<\/li>\n<li>\n<section>\u4f18\u52bf\u5f52\u4e00\u5316\uff08advantage normalization\uff09<\/section>\n<\/li>\n<li>\n<section>\u4ee5\u53ca\u5bf9 reference model \u4f7f\u7528\u6307\u6570\u79fb\u52a8\u5e73\u5747\u8fdb\u884c\u66f4\u65b0\uff08exponential moving average for the reference model\uff09\u3002<\/section>\n<\/li>\n<\/ul>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-37afcd3f946f9247167e446ed6e86f44.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-37afcd3f946f9247167e446ed6e86f44.png\"><\/a><\/p>\n<section>\u7814\u7a76\u56e2\u961f\u6210\u529f\u4f7f\u7528 PPO \u7b97\u6cd5\u5728\u5bf9\u8bdd\u4efb\u52a1 Safe-RLHF\/HH-RLHF \u4ee5\u53ca\u4ee3\u7801\u751f\u6210\u4efb\u52a1 APPS\/CodeContest \u4e0a\u8fbe\u5230\u4e86 SOTA \u7684\u6548\u679c\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-319a54e03410bf9725d22f286d959d78.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-319a54e03410bf9725d22f286d959d78.png\"><\/a><\/p>\n<section>\u5728\u5bf9\u8bdd\u4efb\u52a1\u4e0a\uff0c\u7814\u7a76\u56e2\u961f\u53d1\u73b0\u7efc\u5408\u4e86\u4e09\u4e2a\u5173\u952e\u70b9\u7684 PPO \u7b97\u6cd5\u663e\u8457\u4f18\u4e8e DPO \u7b97\u6cd5\u4ee5\u53ca\u5728\u7ebf\u91c7\u6837\u7684 DPO \u7b97\u6cd5 DPO-Iter\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-e0dca65bb98254b9eba4b27b1bcde95e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-e0dca65bb98254b9eba4b27b1bcde95e.png\"><\/a><\/p>\n<section>\u5728\u4ee3\u7801\u751f\u6210\u4efb\u52a1 APPS \u548c CodeContest \u4e0a\uff0c\u57fa\u4e8e\u5f00\u6e90\u6a21\u578b Code Llama 34B\uff0cPPO \u7b97\u6cd5\u4e5f\u8fbe\u5230\u4e86\u6700\u5f3a\u7684\u6c34\u5e73\uff0c\u5728 CodeContest \u4e0a\u8d85\u8d8a\u4e86\u4e4b\u524d\u7684 SOTA\uff0cAlphaCode 41B\u3002<\/section>\n<section><\/section>\n<section>\u60f3\u8981\u5b9e\u73b0\u6548\u679c\u8f83\u597d\u7684\u5927\u6a21\u578b\u5bf9\u9f50\uff0c\u9ad8\u6548\u7387\u7684\u8bad\u7ec3\u7cfb\u7edf\u662f\u4e0d\u53ef\u7f3a\u5c11\u7684\uff0c\u5728\u5b9e\u73b0\u5927\u89c4\u6a21<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u8bad\u7ec3\u4e0a\uff0c<mark data-type=\"experts\" data-id=\"25b47a57-a97f-41b8-9115-c3043ad9c925\">\u5434\u7ffc<\/mark>\u56e2\u961f\u6709\u957f\u671f\u7684\u79ef\u7d2f\uff0c\u4ece 2021 \u5e74\u5f00\u59cb\u5c31\u642d\u5efa\u4e86\u4e13\u5c5e\u7684\u5206\u5e03\u5f0f<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u6846\u67b6\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-c99403655b76a1cf874bc9d5f1517217.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-c99403655b76a1cf874bc9d5f1517217.png\"><\/a><\/p>\n<ul>\n<li>\n<section>NeurIPS 2022 The Surprising Effectiveness of PPO in Cooperative\u00a0Multi-Agent Games [1]\uff1a\u63d0\u51fa\u5e76\u5f00\u6e90\u4e86\u7528\u4e8e\u591a\u667a\u80fd\u4f53\u7684<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u5e76\u884c\u8bad\u7ec3\u6846\u67b6 MAPPO\uff0c\u652f\u6301\u5408\u4f5c\u573a\u666f\u4e0b\u7684\u591a\u667a\u80fd\u4f53\u8bad\u7ec3\uff0c\u8be5\u5de5\u4f5c\u88ab\u5927\u91cf\u591a\u667a\u80fd\u4f53\u9886\u57df\u5de5\u4f5c\u91c7\u7528\uff0c\u76ee\u524d\u8bba\u6587\u5f15\u7528\u91cf\u5df2\u8d85\u8fc7 1k\u3002<\/section>\n<\/li>\n<li>\n<section>ICLR 2024 Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores [2]: \u63d0\u51fa\u4e86\u7528\u4e8e<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u7684\u5206\u5e03\u5f0f\u8bad\u7ec3\u6846\u67b6\uff0c\u53ef\u8f7b\u677e\u6269\u5c55\u81f3\u4e0a\u4e07\u4e2a\u6838\u5fc3\uff0c\u52a0\u901f\u6bd4\u8d85\u8d8a OpenAI \u7684\u5927\u89c4\u6a21<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u7cfb\u7edf Rapid\u3002<\/section>\n<\/li>\n<li>\n<section>ReaLHF: Optimized RLHF Training for Large Language Models through Parameter Reallocation [3]: \u6700\u8fd1\uff0c<mark data-type=\"experts\" data-id=\"25b47a57-a97f-41b8-9115-c3043ad9c925\">\u5434\u7ffc<\/mark>\u56e2\u961f\u8fdb\u4e00\u6b65\u5b9e\u73b0\u4e86\u5206\u5e03\u5f0f RLHF \u8bad\u7ec3\u6846\u67b6 ReaLHF\u3002<mark data-type=\"experts\" data-id=\"25b47a57-a97f-41b8-9115-c3043ad9c925\">\u5434\u7ffc<\/mark>\u56e2\u961f\u7684 ICML Oral \u8bba\u6587\u6b63\u662f\u57fa\u4e8e ReaLHF \u7cfb\u7edf\u4ea7\u51fa\u7684\u3002ReaLHF \u7cfb\u7edf\u7ecf\u8fc7\u957f\u65f6\u95f4\u7684\u5f00\u53d1\uff0c\u7ecf\u5386\u5927\u91cf\u7684\u7ec6\u8282\u6253\u78e8\uff0c\u8fbe\u5230\u6700\u4f18\u6027\u80fd\u3002\u76f8\u6bd4\u4e8e\u4e4b\u524d\u7684\u5f00\u6e90\u5de5\u4f5c\uff0cReaLHF \u53ef\u4ee5\u5728 RLHF \u8fd9\u4e2a\u6bd4\u9884\u8bad\u7ec3\u66f4\u590d\u6742\u7684\u573a\u666f\u4e0b\u8fbe\u5230\u8fd1\u4e4e\u7ebf\u6027\u7684\u62d3\u5c55\u6027\uff0c\u540c\u65f6\u5177\u6709\u66f4\u9ad8\u7684\u8d44\u6e90\u5229\u7528\u7387\uff0c\u5728 128 \u5757 A100 GPU \u4e0a\u4e5f\u80fd\u7a33\u5b9a\u5feb\u901f\u5730\u8fdb\u884c RLHF \u8bad\u7ec3\uff0c\u76f8\u5173\u5de5\u4f5c\u5df2\u5f00\u6e90\uff1ahttps:\/\/github.com\/openpsi-project\/ReaLHF<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u9664\u4e86\u63d0\u5347\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u4ee3\u7801\u80fd\u529b\u4e4b\u5916\uff0c<mark data-type=\"experts\" data-id=\"25b47a57-a97f-41b8-9115-c3043ad9c925\">\u5434\u7ffc<\/mark>\u56e2\u961f\u8fd8\u91c7\u7528\u591a\u79cd\u5c06<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u7b97\u6cd5\u548c\u5927\u6a21\u578b\u7ed3\u5408\u7684\u65b9\u5f0f\uff0c\u5b9e\u73b0\u4e86\u591a\u79cd\u590d\u6742 LLM Agent\uff0c\u5e76\u53ef\u4ee5\u548c\u4eba\u7c7b\u8fdb\u884c\u590d\u6742\u4ea4\u4e92\u3002<\/section>\n<section><\/section>\n<section>\u5728 MiniRTS \u4e2d\u4f7f\u7528<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u65e2\u80fd\u542c\u4ece\u4eba\u7c7b\u6307\u4ee4\u4e5f\u80fd\u505a\u51fa\u6700\u4f18\u51b3\u7b56\u7684\u8bed\u8a00\u667a\u80fd\u4f53 [4]\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-a84ad62d9cd235ec21a04878a696aa9c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-a84ad62d9cd235ec21a04878a696aa9c.png\"><\/a><\/p>\n<section>\u5728\u72fc\u4eba\u6740\u4e2d\u8bad\u7ec3\u7b56\u7565\u591a\u6837\u5316\u7684<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u7b56\u7565\u4ee5\u63d0\u5347\u5927\u6a21\u578b\u7684\u51b3\u7b56\u80fd\u529b [5]\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-c5f369d3da0923b6b0bf8bfaaf173be5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-c5f369d3da0923b6b0bf8bfaaf173be5.png\"><\/a><\/p>\n<section>\u5728 Overcooked \u6e38\u620f\u4e2d\u7ed3\u5408\u5c0f\u6a21\u578b\u4e0e\u5927\u6a21\u578b\u5b9e\u73b0\u80fd\u8fdb\u884c\u5b9e\u65f6\u53cd\u9988\u7684\u5408\u4f5c Language Agent [6]\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d1e62177ff5bb8158967ed822ddea0c7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-d1e62177ff5bb8158967ed822ddea0c7.png\"><\/a><\/p>\n<section>\u7ed3\u5408<mark data-type=\"tech_methods\" data-id=\"ee1a8f69-3170-4ddf-b2b6-47d91c844425\">\u5f3a\u5316\u5b66\u4e60<\/mark>\u8bad\u7ec3\u7684\u673a\u5668\u4eba\u63a7\u5236\u7b56\u7565\u4e0e\u5927<mark data-type=\"tech_tasks\" data-id=\"bf35ef94-d956-4033-a533-0c0828308c36\">\u8bed\u8a00\u6a21\u578b<\/mark>\u63a8\u7406\u80fd\u529b\u8ba9\u673a\u5668\u4eba\u80fd\u591f\u6267\u884c\u4e00\u7cfb\u5217\u590d\u6742\u4efb\u52a1 [7]\u3002<\/section>\n<p><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-db315fec761eaaa140f1af26b8dab32f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-db315fec761eaaa140f1af26b8dab32f.png\"><\/a><\/p>\n<section>\u4e3a\u4e86\u4f7f\u5927\u6a21\u578b\u80fd\u771f\u6b63\u8d70\u8fdb\u5343\u5bb6\u4e07\u6237\uff0c\u5bf9\u9f50\u6280\u672f\u662f\u81f3\u5173\u91cd\u8981\u7684\uff0c\u5bf9\u4e8e\u5b66\u672f\u754c\u548c\u5927\u6a21\u578b\u4ece\u4e1a\u8005\u6765\u8bf4\uff0c\u597d\u7684\u5f00\u6e90\u5de5\u4f5c\u548c\u8bba\u6587\u65e0\u7591\u4f1a\u5927\u5927\u964d\u4f4e\u5b9e\u9a8c\u6210\u672c\u548c\u5f00\u53d1\u96be\u5ea6\uff0c\u4e5f\u671f\u5f85\u968f\u7740\u6280\u672f\u53d1\u5c55\uff0c\u4f1a\u6709\u66f4\u591a\u670d\u52a1\u4e8e\u4eba\u7c7b\u7684\u5927\u6a21\u578b\u51fa\u73b0\u3002<\/section>\n<section><\/section>\n<section><em><sup>[1] Yu, Chao, Akash Velu, Eugene Vinitsky, Jiaxuan Gao, Yu Wang, Alexandre Bayen, and Yi Wu. &#8220;The surprising effectiveness of ppo in cooperative multi-agent games.&#8221;<\/sup><\/em><\/section>\n<section><sup><em>[2] Mei, Zhiyu, Wei Fu, Guangju Wang, Huanchen Zhang, and Yi Wu. &#8220;SRL: Scaling Distributed Reinforcement Learning to Over Ten Thousand Cores.&#8221;<\/em><\/sup><\/section>\n<section><sup><em>[3] Mei, Zhiyu, Wei Fu, Kaiwei Li, Guangju Wang, Huanchen Zhang, and Yi Wu. &#8220;ReaLHF: Optimized RLHF Training for Large Language Models through Parameter Reallocation.&#8221;<\/em><\/sup><\/section>\n<section><sup><em>[4] Xu, Shusheng, Huaijie Wang, Jiaxuan Gao, Yutao Ouyang, Chao Yu, and Yi Wu. &#8220;Language-guided generation of physically realistic robot motion and control.&#8221;<\/em><\/sup><\/section>\n<section><sup><em>[5] Xu, Zelai, Chao Yu, Fei Fang, Yu Wang, and Yi Wu. &#8220;Language agents with reinforcement learning for strategic play in the werewolf game.&#8221;<\/em><\/sup><\/section>\n<section><sup><em>[6] Liu, Jijia, Chao Yu, Jiaxuan Gao, Yuqing Xie, Qingmin Liao, Yi Wu, and Yu Wang. &#8220;Llm-powered hierarchical language agent for real-time human-ai coordination.&#8221;<\/em><\/sup><\/section>\n<section><em><sup>[7] Ouyang, Yutao, Jinhan Li, Yunfei Li, Zhongyu Li, Chao Yu, Koushil Sreenath, and Yi Wu. &#8220;Long-horizon Locomotion and Manipulation on a Quadrupedal Robot with Large Language Models.&#8221;<\/sup><\/em><\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-07-22-7\" target=\"_blank\">ICML 2024 Oral | DPO\u662f\u5426\u6bd4PPO\u66f4\u9002\u5408LLM\uff0c\u6e05\u534e\u5434\u7ffc\u56e2\u961f\u6700\u65b0\u63ed\u79d8<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:ICML 2024 O [&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":[70,71,68],"class_list":["post-34677","post","type-post","status-publish","format-standard","hentry","category-news","tag-agent","tag-rag","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>ICML 2024 Oral | DPO\u662f\u5426\u6bd4PPO\u66f4\u9002\u5408LLM\uff0c\u6e05\u534e\u5434\u7ffc\u56e2\u961f\u6700\u65b0\u63ed\u79d8 - \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=34677\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=34677\",\"url\":\"https:\/\/17aitech.com\/?p=34677\",\"name\":\"ICML 2024 Oral | DPO\u662f\u5426\u6bd4PPO\u66f4\u9002\u5408LLM\uff0c\u6e05\u534e\u5434\u7ffc\u56e2\u961f\u6700\u65b0\u63ed\u79d8 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=34677#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=34677#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7f7e81ff0ab737b24ac6e6afc20a56f2.png\",\"datePublished\":\"2024-12-10T15:57:07+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=34677#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=34677\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=34677#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7f7e81ff0ab737b24ac6e6afc20a56f2.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7f7e81ff0ab737b24ac6e6afc20a56f2.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=34677#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"ICML 2024 Oral | DPO\u662f\u5426\u6bd4PPO\u66f4\u9002\u5408LLM\uff0c\u6e05\u534e\u5434\u7ffc\u56e2\u961f\u6700\u65b0\u63ed\u79d8\"}]},{\"@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":"ICML 2024 Oral | DPO\u662f\u5426\u6bd4PPO\u66f4\u9002\u5408LLM\uff0c\u6e05\u534e\u5434\u7ffc\u56e2\u961f\u6700\u65b0\u63ed\u79d8 - \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=34677","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=34677","url":"https:\/\/17aitech.com\/?p=34677","name":"ICML 2024 Oral | DPO\u662f\u5426\u6bd4PPO\u66f4\u9002\u5408LLM\uff0c\u6e05\u534e\u5434\u7ffc\u56e2\u961f\u6700\u65b0\u63ed\u79d8 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=34677#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=34677#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7f7e81ff0ab737b24ac6e6afc20a56f2.png","datePublished":"2024-12-10T15:57:07+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=34677#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=34677"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=34677#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7f7e81ff0ab737b24ac6e6afc20a56f2.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/07\/frc-7f7e81ff0ab737b24ac6e6afc20a56f2.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=34677#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"ICML 2024 Oral | DPO\u662f\u5426\u6bd4PPO\u66f4\u9002\u5408LLM\uff0c\u6e05\u534e\u5434\u7ffc\u56e2\u961f\u6700\u65b0\u63ed\u79d8"}]},{"@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\/34677","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=34677"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/34677\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34677"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34677"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34677"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}