{"id":4771,"date":"2024-10-03T10:36:59","date_gmt":"2024-10-03T02:36:59","guid":{"rendered":"https:\/\/nullthought.net\/?p=4771"},"modified":"2024-11-07T11:02:57","modified_gmt":"2024-11-07T03:02:57","slug":"%e7%ba%a6%e6%9d%9f%e7%94%9f%e6%88%90%e7%ad%96%e7%95%a5%e4%bc%98%e5%8c%96%ef%bc%88cgpo%ef%bc%89%e6%a1%86%e6%9e%b6%e8%a7%a3%e5%86%b3%e5%9f%ba%e4%ba%8e%e4%ba%ba%e7%b1%bb%e5%8f%8d%e9%a6%88%e5%bc%ba","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=4771","title":{"rendered":"\u7ea6\u675f\u751f\u6210\u7b56\u7565\u4f18\u5316\uff08CGPO\uff09\u6846\u67b6\u89e3\u51b3\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09\u5728\u591a\u4efb\u52a1\u5b66\u4e60\uff08MTL\uff09\u4e2d\u7684\u5c40\u9650\u6027"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/arxiv.org\/abs\/2409.20370\" target=\"_blank\" rel=\"noreferrer noopener\">The Perfect Blend: Redefining RLHF with Mixture of Judges<\/a><\/strong>\u4e3b\u8981\u8ba8\u8bba\u4e86\u5982\u4f55\u901a\u8fc7\u4e00\u79cd\u65b0\u7684\u540e\u8bad\u7ec3\u8303\u5f0f\uff0c\u79f0\u4e3a\u7ea6\u675f\u751f\u6210\u7b56\u7565\u4f18\u5316\uff08Constrained Generative Policy Optimization\uff0cCGPO\uff09\uff0c\u91cd\u65b0\u5b9a\u4e49\u4e86\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning from Human Feedback\uff0cRLHF\uff09\uff0c\u7279\u522b\u662f\u5728\u591a\u4efb\u52a1\u5b66\u4e60\uff08MTL\uff09\u4e2d\u7684\u5e94\u7528\u3002<\/p>\n\n\n\n<p>\u8be5\u8bba\u6587\u6307\u51fa\uff0c\u4f20\u7edf\u7684RLHF\u65b9\u6cd5\u5728\u591a\u4efb\u52a1\u5b66\u4e60\u4e2d\u5b58\u5728\u4e00\u4e9b\u6311\u6218\uff0c\u5305\u62ec\u5956\u52b1\u9ed1\u5ba2\uff08reward hacking\uff09\u548c\u6781\u7aef\u7684\u591a\u76ee\u6807\u4f18\u5316\u3002\u4e3a\u4e86\u5e94\u5bf9\u8fd9\u4e9b\u6311\u6218\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86\u6df7\u5408\u8bc4\u5ba1\uff08Mixture of Judges\uff0cMoJ\uff09\u7684\u6982\u5ff5\uff0c\u901a\u8fc7\u6210\u672c\u6548\u76ca\u7684\u7ea6\u675f\u7b56\u7565\u4f18\u5316\u548c\u5206\u5c42\u673a\u5236\uff0c\u5728\u591a\u4efb\u52a1RLHF\u4e2d\u627e\u5230\u5e73\u8861\u70b9\u3002CGPO\u65b9\u6cd5\u80fd\u591f\u68c0\u6d4b\u5e76\u51cf\u8f7b\u5956\u52b1\u9ed1\u5ba2\u884c\u4e3a\uff0c\u540c\u65f6\u5728\u591a\u4e2a\u76ee\u6807\u4e4b\u95f4\u8fbe\u5230\u5e15\u7d2f\u6258\u6700\u4f18\uff0c\u6700\u7ec8\u4e3aRLHF\u63d0\u4f9b\u4e86\u66f4\u5168\u9762\u548c\u7cbe\u786e\u7684\u5bf9\u9f50\u624b\u6bb5\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"857\" height=\"381\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image.png\" alt=\"\" class=\"wp-image-4772\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image.png 857w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-300x133.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/10\/image-768x341.png 768w\" sizes=\"auto, (max-width: 857px) 100vw, 857px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2409.20370\" target=\"_blank\" rel=\"noreferrer noopener\">The Perfect Blend: Redefining RLHF with Mixture of Judges<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aTengyu Xu, Eryk Helenowski, Karthik Abinav Sankararaman, Di Jin, Kaiyan Peng, Eric Han, Shaoliang Nie, Chen Zhu, Hejia Zhang, Wenxuan Zhou, Zhouhao Zeng, Yun He, Karishma Mandyam, Arya Talabzadeh, Madian Khabsa, Gabriel Cohen, Yuandong Tian, Hao Ma, Sinong Wang\u548cHan Fang\uff0c\u5747\u6765\u81eaMeta\uff08Meta GenAI, FAIR\uff09\u3002<\/p>\n\n\n\n<p>\u4e00\u4e9b\u4e3a\u8bba\u6587\u6982\u8981\u5185\u5bb9\uff1a<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">1. \u7814\u7a76\u80cc\u666f<\/h4>\n\n\n\n<p>\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5df2\u7ecf\u5c55\u793a\u4e86\u5728\u591a\u79cd\u4efb\u52a1\u4e2d\u7684\u4f18\u8d8a\u8868\u73b0\uff0c\u8fd9\u5f97\u76ca\u4e8e\u5176\u5927\u89c4\u6a21\u53c2\u6570\u5316\u548c\u591a\u4efb\u52a1\u5b66\u4e60\uff08MTL\uff09\u7684\u80fd\u529b\u3002\u7136\u800c\uff0c\u4f20\u7edf\u7684RLHF\u65b9\u6cd5\u5728\u591a\u4efb\u52a1\u573a\u666f\u4e2d\u9762\u4e34\u4e24\u5927\u4e3b\u8981\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5956\u52b1\u9ed1\u5ba2\u95ee\u9898<\/strong>\uff1a\u7531\u4e8e\u5956\u52b1\u6a21\u578b\u5e76\u4e0d\u80fd\u5b8c\u7f8e\u53cd\u6620\u4eba\u7c7b\u7684\u771f\u5b9e\u504f\u597d\uff0c\u6a21\u578b\u53ef\u80fd\u4f1a\u4f18\u5316\u51fa\u4e0d\u7b26\u5408\u4eba\u7c7b\u671f\u671b\u7684\u7ed3\u679c\uff0c\u9020\u6210\u201c\u5956\u52b1\u9ed1\u5ba2\u201d\u73b0\u8c61\uff0c\u5373\u6a21\u578b\u901a\u8fc7\u6700\u5927\u5316\u5956\u52b1\u51fd\u6570\u7684\u7f3a\u9677\uff0c\u800c\u4e0d\u662f\u4f18\u5316\u4eba\u7c7b\u7684\u771f\u5b9e\u76ee\u6807\u3002<\/li>\n\n\n\n<li><strong>\u591a\u76ee\u6807\u4f18\u5316\u7684\u96be\u9898<\/strong>\uff1a\u591a\u4efb\u52a1\u5b66\u4e60\u5e38\u5e38\u9700\u8981\u5728\u591a\u4e2a\u6709\u65f6\u4e92\u76f8\u77db\u76fe\u7684\u76ee\u6807\u4e4b\u95f4\u8fdb\u884c\u5e73\u8861\u3002\u4f20\u7edfRLHF\u65b9\u6cd5\u901a\u5e38\u901a\u8fc7\u7ebf\u6027\u7ec4\u5408\u591a\u4e2a\u5956\u52b1\u6a21\u578b\u6765\u5904\u7406\u591a\u4efb\u52a1\u95ee\u9898\uff0c\u8fd9\u79cd\u65b9\u6cd5\u9700\u8981\u5bf9\u6bcf\u4e2a\u4efb\u52a1\u7684\u6743\u91cd\u8fdb\u884c\u7ec6\u81f4\u7684\u8c03\u6574\uff0c\u96be\u4ee5\u901a\u7528\u4e14\u5bb9\u6613\u5bfc\u81f4\u67d0\u4e9b\u4efb\u52a1\u7684\u6027\u80fd\u9000\u5316\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">2. CGPO\u65b9\u6cd5\u7684\u6838\u5fc3\u521b\u65b0<\/h4>\n\n\n\n<p>\u4e3a\u4e86\u89e3\u51b3\u4e0a\u8ff0\u95ee\u9898\uff0c\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u540e\u8bad\u7ec3\u8303\u5f0f\uff0c\u5373\u7ea6\u675f\u751f\u6210\u7b56\u7565\u4f18\u5316\uff08CGPO\uff09\uff0c\u5176\u4e3b\u8981\u521b\u65b0\u70b9\u5982\u4e0b\uff1a<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">2.1 \u6df7\u5408\u8bc4\u5ba1\u673a\u5236\uff08MoJ\uff09<\/h5>\n\n\n\n<p>\u8bba\u6587\u7684\u6838\u5fc3\u4e4b\u4e00\u662f\u5f15\u5165\u4e86\u201c\u6df7\u5408\u8bc4\u5ba1\u201d\uff08Mixture of Judges, MoJ\uff09\uff0c\u8fd9\u662f\u4e00\u79cd\u7531\u591a\u79cd\u8bc4\u5ba1\u5668\u7ec4\u6210\u7684\u6a21\u5757\uff0c\u7528\u4e8e\u8bc6\u522b\u548c\u5e94\u5bf9\u5956\u52b1\u9ed1\u5ba2\u95ee\u9898\u3002MoJ\u53ef\u4ee5\u5305\u62ec\u57fa\u4e8e\u89c4\u5219\u7684\u8bc4\u5ba1\u5668\u548c\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8bc4\u5ba1\u5668\u3002\u57fa\u4e8e\u89c4\u5219\u7684\u8bc4\u5ba1\u5668\u53ef\u4ee5\u901a\u8fc7\u7279\u5b9a\u89c4\u5219\uff08\u5982\u6570\u5b66\u95ee\u9898\u7684\u7b54\u6848\u51c6\u786e\u6027\u6216\u4ee3\u7801\u6267\u884c\u7ed3\u679c\uff09\u8bc4\u4f30\u751f\u6210\u7684\u6587\u672c\u662f\u5426\u7b26\u5408\u8981\u6c42\uff1b\u800c\u57fa\u4e8eLLM\u7684\u8bc4\u5ba1\u5668\u5219\u80fd\u591f\u5904\u7406\u66f4\u590d\u6742\u7684\u4efb\u52a1\uff0c\u5982\u5224\u65ad\u751f\u6210\u5185\u5bb9\u7684\u4e8b\u5b9e\u6027\u3001\u5224\u65ad\u751f\u6210\u5185\u5bb9\u662f\u5426\u8fdd\u53cd\u5b89\u5168\u7ea6\u675f\u7b49\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">2.2 \u7ea6\u675f\u751f\u6210\u7b56\u7565\u4f18\u5316\uff08CGPO\uff09<\/h5>\n\n\n\n<p>CGPO\u901a\u8fc7\u7ea6\u675f\u751f\u6210\u7b56\u7565\u4f18\u5316\u65b9\u6cd5\uff0c\u5c06\u591a\u4efb\u52a1RLHF\u4e2d\u7684\u5956\u52b1\u9ed1\u5ba2\u95ee\u9898\u6700\u5c0f\u5316\u3002\u8be5\u65b9\u6cd5\u4e0d\u4ec5\u4f18\u5316\u5956\u52b1\u6a21\u578b\u5206\u6570\uff0c\u8fd8\u901a\u8fc7\u7ea6\u675f\u6761\u4ef6\u4fdd\u8bc1\u751f\u6210\u7ed3\u679c\u7684\u6709\u6548\u6027\u3002\u4f8b\u5982\uff0c\u5728\u6570\u5b66\u95ee\u9898\u4e2d\uff0cCGPO\u8981\u6c42\u751f\u6210\u7b54\u6848\u5fc5\u987b\u6b63\u786e\uff0c\u800c\u4e0d\u4ec5\u4ec5\u662f\u5956\u52b1\u6a21\u578b\u6253\u5206\u9ad8\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff0c\u8bba\u6587\u5f15\u5165\u4e86\u591a\u4e2a\u7ea6\u675fRLHF\u4f18\u5316\u5668\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6821\u51c6\u6b63\u5219\u5316\u7b56\u7565\u68af\u5ea6\uff08CRPG\uff09<\/strong>\uff1a\u5728\u4f20\u7edf\u7684RLHF\u4f18\u5316\u4e2d\uff0c\u5956\u52b1\u6a21\u578b\u7684\u8f93\u51fa\u53ef\u80fd\u5b58\u5728\u6821\u51c6\u95ee\u9898\uff0c\u5bfc\u81f4\u4e0d\u540c\u63d0\u793a\uff08prompt\uff09\u7684\u751f\u6210\u7ed3\u679c\u5206\u6570\u4e0d\u80fd\u76f4\u63a5\u6bd4\u8f83\u3002CRPG\u901a\u8fc7\u5f15\u5165\u6821\u51c6\u5956\u52b1\u673a\u5236\u6765\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u4fdd\u8bc1\u4e0d\u540c\u63d0\u793a\u4e4b\u95f4\u7684\u5956\u52b1\u5206\u6570\u662f\u53ef\u6bd4\u8f83\u7684\u3002<\/li>\n\n\n\n<li><strong>\u7ea6\u675f\u5728\u7ebf\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08CODPO\uff09<\/strong>\uff1aCODPO\u57fa\u4e8e\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08DPO\uff09\u65b9\u6cd5\uff0c\u5728\u7ebf\u4f18\u5316\u7b56\u7565\uff0c\u901a\u8fc7\u5bf9\u6bcf\u4e2a\u4efb\u52a1\u72ec\u7acb\u4f18\u5316\uff0c\u907f\u514d\u4e86\u4f20\u7edfRLHF\u4e2d\u4efb\u52a1\u95f4\u7684\u76f8\u4e92\u5e72\u6270\u3002<\/li>\n\n\n\n<li><strong>\u6821\u51c6\u6b63\u5219\u5316\u5956\u52b1\u6392\u540d\u5fae\u8c03\uff08CRRAFT\uff09<\/strong>\uff1aCRRAFT\u7ed3\u5408\u4e86\u5956\u52b1\u6392\u540d\u7b97\u6cd5\uff0c\u901a\u8fc7\u7ea6\u675f\u4f18\u5316\u6bcf\u6b21\u751f\u6210\u7ed3\u679c\u7684\u6982\u7387\uff0c\u786e\u4fdd\u6a21\u578b\u751f\u6210\u9ad8\u8d28\u91cf\u7684\u54cd\u5e94\u3002<\/li>\n<\/ul>\n\n\n\n<h5 class=\"wp-block-heading\">2.3 \u591a\u4efb\u52a1\u4f18\u5316\u4e2d\u7684\u7ec6\u7c92\u5ea6\u7ba1\u7406<\/h5>\n\n\n\n<p>\u8bba\u6587\u4e2d\u7279\u522b\u5f3a\u8c03\u4e86\u5728\u591a\u4efb\u52a1\u5b66\u4e60\u573a\u666f\u4e0b\uff0c\u4e0d\u80fd\u4e00\u5200\u5207\u5730\u5e94\u7528\u540c\u4e00\u4e2a\u5956\u52b1\u6a21\u578b\u6216\u4f18\u5316\u7b56\u7565\u3002\u76f8\u53cd\uff0cCGPO\u6839\u636e\u4e0d\u540c\u4efb\u52a1\u5b9a\u5236\u4e86\u7279\u5b9a\u7684\u5956\u52b1\u6a21\u578b\u3001\u8bc4\u5ba1\u673a\u5236\u548c\u4f18\u5316\u5668\u8bbe\u7f6e\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u6bcf\u4e2a\u4efb\u52a1\u90fd\u6709\u81ea\u5df1\u7684\u4f18\u5316\u7b56\u7565\uff0c\u907f\u514d\u4e86\u4e0d\u540c\u4efb\u52a1\u76ee\u6807\u4e4b\u95f4\u7684\u51b2\u7a81\u3002\u4f8b\u5982\uff0c\u5728\u751f\u6210\u804a\u5929\u56de\u590d\u7684\u4efb\u52a1\u4e2d\uff0c\u53ef\u80fd\u66f4\u5173\u6ce8\u751f\u6210\u5185\u5bb9\u7684\u5e2e\u52a9\u6027\u548c\u4e8b\u5b9e\u6027\uff1b\u800c\u5728\u7f16\u7a0b\u4efb\u52a1\u4e2d\uff0c\u5219\u91cd\u70b9\u5728\u4e8e\u4ee3\u7801\u7684\u6b63\u786e\u6027\u548c\u8fd0\u884c\u7ed3\u679c\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">3. \u5b9e\u9a8c\u7ed3\u679c<\/h4>\n\n\n\n<p>\u5728\u5b9e\u9a8c\u90e8\u5206\uff0cCGPO\u901a\u8fc7\u591a\u79cd\u57fa\u51c6\u4efb\u52a1\uff08\u5305\u62ec\u5bf9\u8bdd\u751f\u6210\u3001\u6570\u5b66\u63a8\u7406\u3001\u7f16\u7a0b\u3001\u77e5\u8bc6\u95ee\u7b54\u7b49\uff09\u7684\u8bc4\u4f30\uff0c\u8bc1\u660e\u4e86\u5176\u5728\u591a\u4e2a\u4efb\u52a1\u4e2d\u4f18\u4e8e\u73b0\u6709\u7684RLHF\u7b97\u6cd5\u5982PPO\u548cDPO\u3002\u4f8b\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728AlpacaEval-2\uff08\u5bf9\u8bdd\u751f\u6210\u4efb\u52a1\uff09\u4e2d\uff0cCGPO\u76f8\u5bf9\u4e8ePPO\u63d0\u5347\u4e867.4%\uff0c\u76f8\u5bf9\u4e8eDPO\u63d0\u5347\u4e8612.5%\u3002<\/li>\n\n\n\n<li>\u5728HumanEval\uff08\u7f16\u7a0b\u4efb\u52a1\uff09\u4e2d\uff0cCGPO\u76f8\u8f83\u4e8ePPO\u63d0\u5347\u4e865%\u3002<\/li>\n\n\n\n<li>\u5728MATH\uff08\u6570\u5b66\u63a8\u7406\u4efb\u52a1\uff09\u4e2d\uff0cCGPO\u5728\u4fdd\u6301\u4ee3\u7801\u51c6\u786e\u6027\u7684\u540c\u65f6\uff0c\u907f\u514d\u4e86PPO\u5e38\u89c1\u7684\u5956\u52b1\u9ed1\u5ba2\u95ee\u9898\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e9b\u7ed3\u679c\u8868\u660e\uff0cCGPO\u65b9\u6cd5\u5728\u5e94\u5bf9\u5956\u52b1\u9ed1\u5ba2\u95ee\u9898\u548c\u591a\u76ee\u6807\u4f18\u5316\u4e0a\u5177\u6709\u663e\u8457\u4f18\u52bf\u3002\u6b64\u5916\uff0cCGPO\u7684\u6df7\u5408\u8bc4\u5ba1\u673a\u5236\u53ef\u4ee5\u9488\u5bf9\u6027\u5730\u4f18\u5316\u6bcf\u4e2a\u4efb\u52a1\uff0c\u907f\u514d\u4e86\u4f20\u7edfRLHF\u5728\u591a\u4efb\u52a1\u4f18\u5316\u4e2d\u9047\u5230\u7684\u74f6\u9888\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">4. \u7ed3\u8bba\u4e0e\u8d21\u732e<\/h4>\n\n\n\n<p>CGPO\u4e3aRLHF\u9886\u57df\u5e26\u6765\u4e86\u591a\u9879\u7a81\u7834\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u591a\u4efb\u52a1\u5b66\u4e60\u4e2d\u7684\u65b0\u8303\u5f0f<\/strong>\uff1aCGPO\u901a\u8fc7\u7ec6\u7c92\u5ea6\u7684\u4efb\u52a1\u7ba1\u7406\uff0c\u89e3\u51b3\u4e86\u4f20\u7edfRLHF\u5728\u591a\u4efb\u52a1\u73af\u5883\u4e2d\u7684\u6027\u80fd\u4e0b\u964d\u95ee\u9898\u3002<\/li>\n\n\n\n<li><strong>\u5956\u52b1\u9ed1\u5ba2\u7684\u6709\u6548\u63a7\u5236<\/strong>\uff1a\u901a\u8fc7\u6df7\u5408\u8bc4\u5ba1\u673a\u5236\uff0cCGPO\u5728\u5956\u52b1\u9ed1\u5ba2\u73b0\u8c61\u7684\u9632\u63a7\u4e0a\u8868\u73b0\u51fa\u8272\uff0c\u80fd\u591f\u5728\u591a\u4e2a\u4efb\u52a1\u4e2d\u627e\u5230\u5e15\u7d2f\u6258\u6700\u4f18\u89e3\u3002<\/li>\n\n\n\n<li><strong>\u65e0\u987b\u5927\u91cf\u8d85\u53c2\u6570\u8c03\u6574\u7684\u4f18\u5316\u5668<\/strong>\uff1aCRPG\u3001CODPO\u548cCRRAFT\u7b49\u65b0\u578b\u4f18\u5316\u5668\u5177\u6709\u5f88\u597d\u7684\u53ef\u6269\u5c55\u6027\uff0c\u9002\u5408\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u7684\u540e\u8bad\u7ec3\u4efb\u52a1\uff0c\u4e14\u4e0d\u9700\u8981\u5927\u91cf\u7684\u8d85\u53c2\u6570\u8c03\u6574\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587The Perfect Blend: Redefining RLHF with Mixture of Ju [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","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":"","ast-disable-related-posts":"","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":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"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":""},"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":[8],"tags":[39,81,77,86],"class_list":["post-4771","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-ai-safety","tag-meta","tag-llm"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587The Perfect Blend: Redefining RLHF with Mixture of Ju&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4771","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=4771"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4771\/revisions"}],"predecessor-version":[{"id":4773,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/4771\/revisions\/4773"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=4771"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=4771"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=4771"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}