{"id":1144532,"date":"2025-07-10T19:29:00","date_gmt":"2025-07-11T02:29:00","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&p=1144532"},"modified":"2025-07-10T19:31:13","modified_gmt":"2025-07-11T02:31:13","slug":"new-arrival-in-research-28","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/new-arrival-in-research-28\/","title":{"rendered":"ICML\u4e0a\u65b0 | \u63a2\u7d22\u667a\u80fd\u751f\u6210\u4e0e\u51b3\u7b56\u4f18\u5316\u65b0\u8def\u5f84"},"content":{"rendered":"\n

\u7f16\u8005\u6309\uff1a\u6b22\u8fce\u9605\u8bfb\u201c\u79d1\u7814\u4e0a\u65b0\u201d\u680f\u76ee\uff01\u201c\u79d1\u7814\u4e0a\u65b0\u201d\u6c47\u805a\u4e86\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6700\u65b0\u7684\u521b\u65b0\u6210\u679c\u4e0e\u79d1\u7814\u52a8\u6001\u3002\u5728\u8fd9\u91cc\uff0c\u4f60\u53ef\u4ee5\u5feb\u901f\u6d4f\u89c8\u7814\u7a76\u9662\u7684\u4eae\u70b9\u8d44\u8baf\uff0c\u4fdd\u6301\u5bf9\u524d\u6cbf\u9886\u57df\u7684\u654f\u9510\u55c5\u89c9\u3002<\/p>\n\n\n\n

7\u670813\u65e5\u81f37\u670819\u65e5\uff0c\u4eba\u5de5\u667a\u80fd\u3001\u673a\u5668\u5b66\u4e60\u9886\u57df\u5168\u7403\u9876\u7ea7\u7684\u5b66\u672f\u76db\u4f1a\u4e4b\u4e00 ICML \u5c06\u5728\u6e29\u54e5\u534e\u4e3e\u529e\u3002\u6765\u81ea\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u591a\u7bc7\u8bba\u6587\u5165\u9009\u3002\u5728\u63a5\u4e0b\u6765\u7684\u51e0\u5929\u91cc\uff0c\u6211\u4eec\u5c06\u901a\u8fc7\u4e24\u671f\u201c\u79d1\u7814\u4e0a\u65b0\u201d\u4e3a\u5927\u5bb6\u5e26\u6765\u591a\u7bc7\u7814\u7a76\u9662\u5165\u9009 ICML 2025 \u7684\u7cbe\u9009\u8bba\u6587\u89e3\u8bfb\u3002\u7b2c\u4e00\u671f\u7684\u7814\u7a76\u5de5\u4f5c\u4e3b\u8981\u5173\u6ce8\u51b3\u7b56\u6a21\u578b\uff0c\u6db5\u76d6\u5f3a\u5316\u5b66\u4e60\u3001RLHF\u3001\u6269\u6563\u6a21\u578b\u5efa\u6a21\u7b49\u65b9\u5411\u3002<\/p>\n\n\n\n

\n\n\n\n

\u672c\u671f\u5185\u5bb9\u901f\u89c8<\/strong><\/p>\n\n\n\n

01. BRIDGE\uff1a\u901a\u8fc7\u591a\u667a\u80fd\u4f53\u8fed\u4ee3\u4f18\u5316\u4e0e\u6269\u6563\u5efa\u6a21\uff0c\u5b9e\u73b0\u6587\u672c\u5f15\u5bfc\u7684\u65f6\u95f4\u5e8f\u5217\u751f\u6210<\/p>\n\n\n\n

02. DPO\u4e0ePPO\u878d\u5408\uff1a\u7528\u4e8eRLHF\u7684\u5f3a\u5316\u8bcd\u5143\u4f18\u5316<\/p>\n\n\n\n

03. \u5c06\u6269\u6563\u89c4\u5212\u4e60\u60ef\u5316\uff0c\u5b9e\u73b0\u9ad8\u6548\u4e14\u6709\u6548\u7684\u51b3\u7b56\u5236\u5b9a<\/p>\n\n\n\n

04. \u7ec4\u5408\u591a\u81c2\u8001\u864e\u673a\u7684\u79bb\u7ebf\u5b66\u4e60\u65b9\u6cd5<\/p>\n\n\n\n

05. \u7b56\u7565\u8fc7\u6ee4\u673a\u5236\u63d0\u5347RLHF\u5956\u52b1\u4fe1\u53f7\u8d28\u91cf<\/p>\n\n\n\n

06. \u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u4e2d\u9a6c\u5c14\u53ef\u592b\u6ce8\u610f\u529b\u673a\u5236\u5728\u79bb\u7ebf\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u4f5c\u7528\u63a2\u6790<\/p>\n\n\n\n

01. BRIDGE\uff1a\u901a\u8fc7\u591a\u667a\u80fd\u4f53\u8fed\u4ee3\u4f18\u5316\u4e0e\u6269\u6563\u5efa\u6a21\uff0c\u5b9e\u73b0\u6587\u672c\u5f15\u5bfc\u7684\u65f6\u95f4\u5e8f\u5217\u751f\u6210<\/h3>\n\n\n\n
\"paper\"<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2503.02445 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u9ad8\u8d28\u91cf\u7684\u65f6\u95f4\u5e8f\u5217\u751f\u6210\uff08Time-Series Generation, TSG\uff09\u5728\u91d1\u878d\u3001\u533b\u7597\u3001\u6c14\u8c61\u7b49\u8bf8\u591a\u9886\u57df\u90fd\u6709\u7740\u6781\u4e3a\u5e7f\u6cdb\u7684\u5e94\u7528\u3002\u7136\u800c\uff0c\u5f53\u524d\u7684\u8bb8\u591a\u65b9\u6cd5\u4e3b\u8981\u96c6\u4e2d\u5728\u65e0\u6761\u4ef6\u7684\u5355\u4e00\u9886\u57df\u751f\u6210\u4e0a\uff0c\u96be\u4ee5\u6ee1\u8db3\u73b0\u5b9e\u573a\u666f\u4e2d\u5bf9\u4e8e\u8de8\u9886\u57df\u3001\u53ef\u63a7\u6027\u5f3a\u7684\u65f6\u95f4\u5e8f\u5217\u751f\u6210\u9700\u6c42\u3002\u7279\u522b\u662f\u5728\u90a3\u4e9b\u9700\u8981\u7ed3\u5408\u9886\u57df\u77e5\u8bc6\u3001\u5b9e\u4f8b\u7279\u5f81\u6216\u64cd\u4f5c\u7ea6\u675f\u7684\u4efb\u52a1\u4e2d\uff0c\u4f20\u7edf\u65b9\u6cd5\u5f80\u5f80\u5728\u7075\u6d3b\u6027\u4e0e\u7cbe\u5ea6\u65b9\u9762\u5b58\u5728\u660e\u663e\u4e0d\u8db3\u3002<\/p>\n\n\n\n

\u9274\u4e8e\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u521b\u65b0\u601d\u8def\u2014\u2014\u5229\u7528\u6587\u672c\u4f5c\u4e3a\u63a7\u5236\u4fe1\u53f7\u6765\u5f15\u5bfc\u65f6\u95f4\u5e8f\u5217\u751f\u6210\uff0c\u4ece\u800c\u63d0\u9ad8\u751f\u6210\u7ed3\u679c\u7684\u8bed\u4e49\u4e00\u81f4\u6027\u4e0e\u5b9a\u5236\u5316\u80fd\u529b\u3002\u6570\u636e\u7a00\u7f3a\u662f\u76ee\u524d\u6587\u672c\u5f15\u5bfc\u751f\u6210\u8fc7\u7a0b\u4e2d\u6240\u9762\u4e34\u7684\u5173\u952e\u6311\u6218\uff0c\u5bf9\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u591a\u667a\u80fd\u4f53\u6846\u67b6\uff0c\u4e13\u95e8\u7528\u4e8e\u81ea\u52a8\u5408\u6210\u4e30\u5bcc\u591a\u6837\u4e14\u771f\u5b9e\u53ef\u9760\u7684\u6587\u672c\u5230\u65f6\u95f4\u5e8f\u5217\uff08Text-to-TS\uff09\u6570\u636e\u96c6\u3002\u901a\u8fc7\u591a\u4e2a\u667a\u80fd\u4f53\u4e4b\u95f4\u7684\u534f\u540c\u4f18\u5316\uff0c\u8be5\u6846\u67b6\u53ef\u4ee5\u6a21\u62df\u4eba\u7c7b\u4e13\u5bb6\u7684\u6807\u6ce8\u8fc7\u7a0b\uff0c\u663e\u8457\u63d0\u5347\u6570\u636e\u7684\u8d28\u91cf\u4e0e\u591a\u6837\u6027\u3002<\/p>\n\n\n\n

\u5728\u6b64\u57fa\u7840\u4e4b\u4e0a\uff0c\u7814\u7a76\u5458\u4eec\u53c8\u8fdb\u4e00\u6b65\u63d0\u51fa\u4e86 BRIDGE \u6846\u67b6\uff0c\u8fd9\u662f\u4e00\u79cd\u6df7\u5408\u5f0f\u7684\u6587\u672c\u63a7\u5236\u65f6\u95f4\u5e8f\u5217\u751f\u6210\u65b9\u6cd5\u3002BRIDGE \u6846\u67b6\u901a\u8fc7\u878d\u5408\u8bed\u4e49\u539f\u578b\u4e0e\u6587\u672c\u63cf\u8ff0\uff0c\u4e3a\u751f\u6210\u8fc7\u7a0b\u63d0\u4f9b\u4e86\u9886\u57df\u7ea7\u522b\u7684\u7cbe\u51c6\u5f15\u5bfc\u4fe1\u53f7\uff0c\u5e76\u4e14\u5de7\u5999\u5730\u7ed3\u5408\u4e86\u6269\u6563\u6a21\u578b\u4e0e\u4f18\u5316\u673a\u5236\uff0c\u5b9e\u73b0\u4e86\u9ad8\u4fdd\u771f\u5ea6\u4e0e\u9ad8\u53ef\u63a7\u6027\u7684\u65f6\u95f4\u5e8f\u5217\u751f\u6210\u3002<\/p>\n\n\n\n

\"diagram,
\u56fe1\uff1aBRIDGE \u6846\u67b6\u56fe<\/figcaption><\/figure>\n\n\n\n

\u7814\u7a76\u5458\u4eec\u572812\u4e2a\u4e0d\u540c\u7684\u6570\u636e\u96c6\u4e0a\u5bf9 BRIDGE \u6846\u67b6\u8fdb\u884c\u4e86\u5168\u9762\u8bc4\u4f30\u3002\u7ed3\u679c\u663e\u793a\uff0cBRIDGE \u5728\u5176\u4e2d\u768411\u4e2a\u6570\u636e\u96c6\u4e0a\u5747\u53d6\u5f97\u4e86\u6700\u5148\u8fdb\u7684\u751f\u6210\u8d28\u91cf\u8868\u73b0\u3002\u5728\u63a7\u5236\u80fd\u529b\u65b9\u9762\uff0c\u76f8\u8f83\u4e8e\u4f20\u7edf\u65e0\u6587\u672c\u8f93\u5165\u7684\u751f\u6210\u65b9\u5f0f\uff0cBRIDGE \u5728\u5747\u65b9\u8bef\u5dee\uff08MSE\uff09\u4e0a\u5b9e\u73b0\u4e8612.52%\u7684\u663e\u8457\u63d0\u5347\uff0c\u5728\u5e73\u5747\u7edd\u5bf9\u8bef\u5dee\uff08MAE\uff09\u4e0a\u4e5f\u8fbe\u5230\u4e866.34%\u7684\u63d0\u5347\u3002<\/p>\n\n\n\n

\u8fd9\u4e9b\u7ed3\u679c\u5145\u5206\u8868\u660e\uff0c\u6587\u672c\u4e0d\u4ec5\u80fd\u591f\u63d0\u4f9b\u4e30\u5bcc\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u8fd8\u80fd\u6709\u6548\u5730\u5d4c\u5165\u5b9e\u4f8b\u7ea7\u7684\u65f6\u95f4\u6a21\u5f0f\uff0c\u8fdb\u800c\u6781\u5927\u5730\u589e\u5f3a\u751f\u6210\u6a21\u578b\u7684\u8868\u73b0\u529b\u4e0e\u9002\u5e94\u6027\u3002<\/p>\n\n\n\n

02. DPO\u4e0ePPO\u878d\u5408\uff1a\u7528\u4e8eRLHF\u7684\u5f3a\u5316\u8bcd\u5143\u4f18\u5316<\/h3>\n\n\n\n
\"paper\"<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2404.18922 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u5927\u8bed\u8a00\u6a21\u578b\u901a\u8fc7\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning from Human Feedback, RLHF\uff09\u5b9e\u73b0\u4e86\u548c\u4eba\u7c7b\u7684\u4ef7\u503c\u5bf9\u9f50\u3002\u5728\u8fd9\u4e00\u8fc7\u7a0b\u4e2d\uff0c\u4f7f\u7528\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff08Proximal Policy Optimization, PPO\uff09\u53ef\u4ee5\u6700\u5927\u5316\u7a00\u758f\u7684\u53e5\u5b50\u7ea7\u5956\u52b1\uff0c\u4f46 PPO \u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u5e38\u5e38\u9762\u4e34\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u3001\u6837\u672c\u6548\u7387\u4f4e\u4e0b\u7b49\u95ee\u9898\u3002<\/p>\n\n\n\n

\u4e3a\u89e3\u51b3\u8fd9\u4e00\u6311\u6218\uff0c\u7814\u7a76\u5458\u4eec\u5c1d\u8bd5\u5c06 RLHF \u95ee\u9898\u5efa\u6a21\u4e3a token \u7ea7\u7684\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b\uff08Markov Decision Process, MDP\uff09\uff0c\u5e76\u4f7f\u7528\u7ec6\u7c92\u5ea6\u7684 token \u7ea7\u5956\u52b1\u4fe1\u606f\u6765\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u3002\u57fa\u4e8e\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u201c\u5f3a\u5316\u8bcd\u5143\u4f18\u5316\u201d\uff08Reinforced Token Optimization, RTO\uff09\u7b97\u6cd5\u3002\u8be5\u7b97\u6cd5\u901a\u8fc7\u4eba\u7c7b\u504f\u597d\u6570\u636e\u5b66\u4e60 token \u7ea7\u5956\u52b1\u51fd\u6570\uff0c\u7136\u540e\u6839\u636e\u5956\u52b1\u4fe1\u53f7\u8fdb\u884c\u7b56\u7565\u4f18\u5316\u3002RTO \u8bc1\u660e\u4e86\u53e6\u4e00\u79cd RLHF \u7b97\u6cd5\u76f4\u63a5\u504f\u597d\u4f18\u5316\uff08Direct Preference Optimization, DPO\uff09\u8bad\u7ec3\u5f97\u5230\u7684\u6a21\u578b\uff0c\u53ef\u4ee5\u9690\u5f0f\u5730\u63d0\u4f9b token \u7ea7\u5956\u52b1\u4fe1\u606f\u3002\u56e0\u6b64\uff0cRTO \u5c06 DPO \u548c PPO \u878d\u5408\uff0c\u80fd\u591f\u4f7f\u7528 PPO \u4f18\u5316 DPO \u63d0\u4f9b\u7684 token \u7ea7\u5956\u52b1\u3002<\/p>\n\n\n\n

\"graphical
\u56fe2\uff1aRTO \u65b9\u6cd5\u7684\u6574\u4f53\u67b6\u6784\u56fe<\/figcaption><\/figure>\n\n\n\n

\u7406\u8bba\u5206\u6790\u8868\u660e\uff0cRTO \u53ef\u4ee5\u5bfb\u627e\u8fd1\u4f3c\u6700\u4f18\u7b56\u7565\u7684\u80fd\u529b\uff0c\u4f7f\u5176\u5728\u6837\u672c\u6548\u7387\u65b9\u9762\u5177\u6709\u4f18\u52bf\u3002\u5b9e\u8df5\u4e2d\uff0cRTO \u4f7f\u75281\/8\u7684\u6837\u672c\u6570\u91cf\u5373\u53ef\u8d85\u8d8a PPO\u3002\u7814\u7a76\u5458\u4eec\u5728\u591a\u4e2a\u57fa\u51c6\u4efb\u52a1\u4e0a\u5bf9 RTO \u8fdb\u884c\u4e86\u6d4b\u8bd5\uff0c\u5176\u8868\u73b0\u4f18\u5f02\uff0c\u5927\u5e45\u8d85\u8fc7\u4f20\u7edf PPO \u548c\u5176\u4ed6\u504f\u597d\u5b66\u4e60\u7b97\u6cd5\u3002\u5728 AlpacaEval 2 \u57fa\u51c6\u4e0a\uff0cRTO \u76f8\u8f83\u4e8e PPO \u63d0\u5347\u4e867.5\u4e2a\u767e\u5206\u70b9\uff1b\u5728 Arena-Hard \u4efb\u52a1\u4e0a\u63d0\u5347\u4e864.1\u4e2a\u767e\u5206\u70b9\uff0c\u5145\u5206\u5c55\u793a\u4e86\u5176\u5728\u771f\u5b9e\u5bf9\u9f50\u4efb\u52a1\u4e2d\u7684\u5f3a\u5927\u6027\u80fd\u3002<\/p>\n\n\n\n

03. \u5c06\u6269\u6563\u89c4\u5212\u4e60\u60ef\u5316\uff0c\u5b9e\u73b0\u9ad8\u6548\u4e14\u6709\u6548\u7684\u51b3\u7b56\u5236\u5b9a<\/h3>\n\n\n\n
\"paper\"<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2502.06401 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u8fd1\u5e74\u6765\uff0c\u6269\u6563\u6a21\u578b\u5728\u51b3\u7b56\u9886\u57df\u5c55\u73b0\u51fa\u6781\u5927\u7684\u6f5c\u529b\uff0c\u7279\u522b\u662f\u5728\u5f3a\u5316\u5b66\u4e60\u4efb\u52a1\u4e2d\u8868\u73b0\u7a81\u51fa\u3002\u7136\u800c\uff0c\u8fd9\u7c7b\u6a21\u578b\u5728\u63a8\u7406\u9636\u6bb5\u7684\u9ad8\u8ba1\u7b97\u6210\u672c\u4e25\u91cd\u5236\u7ea6\u4e86\u5176\u5b9e\u9645\u5e94\u7528\u3002\u5bf9\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86 Habi \u901a\u7528\u6846\u67b6\uff0c\u901a\u8fc7\u5c06\u8ba1\u7b97\u5bc6\u96c6\u7684\u6269\u6563\u89c4\u5212\u6a21\u578b\u8f6c\u5316\u4e3a\u9ad8\u6548\u7684\u51b3\u7b56\u6a21\u578b\u6765\u7a81\u7834\u8fd9\u4e00\u9650\u5236\u3002<\/p>\n\n\n\n

\"diagram\"
\u56fe3\uff1aHabi \u6846\u67b6\u793a\u610f\u56fe\u3002(a) \u5728\u4e60\u60ef\u5316\uff08\u8bad\u7ec3\uff09\u9636\u6bb5\uff0cHabi \u901a\u8fc7\u91cd\u5efa\u7531\u6269\u6563\u89c4\u5212\u5668\u751f\u6210\u7684\u8ba1\u5212\u4e2d\u7684\u52a8\u4f5c\u8fdb\u884c\u5b66\u4e60\uff0c\u4e60\u60ef\uff08\u5148\u9a8c\uff09\u4e0e\u89c4\u5212\uff08\u540e\u9a8c\uff09\u7684\u51b3\u7b56\u7a7a\u95f4\u901a\u8fc7\u6f5c\u5728\u7a7a\u95f4\u4e2d\u7684 KL \u6563\u5ea6\u8fdb\u884c\u5bf9\u9f50\u3002\u53ef\u8bad\u7ec3\u7684\u6a21\u5757\u5305\u62ec\u5148\u9a8c\u7f16\u7801\u5668\u3001\u540e\u9a8c\u7f16\u7801\u5668\u3001\u89e3\u7801\u5668\u548c\u8bc4\u4ef7\u5668\uff1b(b) \u5728\u4e60\u60ef\u6027\u63a8\u7406\uff08HI\uff09\u9636\u6bb5\uff0c\u4ec5\u9700\u4f7f\u7528\u8f7b\u91cf\u7ea7\u7684\u5148\u9a8c\u7f16\u7801\u5668\u548c\u6f5c\u5728\u89e3\u7801\u5668\uff0c\u5373\u53ef\u5b9e\u73b0\u5feb\u901f\u4e14\u9ad8\u8d28\u91cf\u7684\u4e60\u60ef\u6027\u884c\u4e3a\u4ee5\u652f\u6301\u51b3\u7b56\u5236\u5b9a\u3002<\/figcaption><\/figure>\n\n\n\n

Habi \u6846\u67b6\u7684\u521b\u65b0\u7075\u611f\u6e90\u81ea\u4eba\u7c7b\u8ba4\u77e5\u7684\u4e60\u60ef\u5316\u673a\u5236\u2014\u2014\u901a\u8fc7\u53cd\u590d\u8bad\u7ec3\uff0c\u5c06\u9ad8\u6210\u672c\u7684\u76ee\u6807\u5bfc\u5411\u884c\u4e3a\u8f6c\u5316\u4e3a\u9ad8\u6548\u7684\u81ea\u52a8\u5316\u54cd\u5e94\u3002\u5176\u6838\u5fc3\u5728\u4e8e\u5229\u7528\u6269\u6563\u6a21\u578b\u5f3a\u5927\u7684\u8bad\u7ec3\u9636\u6bb5\u8868\u5f81\u80fd\u529b\uff0c\u7ed3\u5408\u7b56\u7565\u84b8\u998f\u7b49\u6280\u672f\uff0c\u5c06\u5176\u51b3\u7b56\u80fd\u529b\u8fc1\u79fb\u81f3\u8f7b\u91cf\u7ea7\u7b56\u7565\u7f51\u7edc\uff0c\u4ece\u800c\u5b9e\u73b0\u63a8\u7406\u9636\u6bb5\u7684\u5feb\u901f\u54cd\u5e94\u3002<\/p>\n\n\n\n

\u7814\u7a76\u5458\u4eec\u5728 D4RL \u7b49\u6807\u51c6\u79bb\u7ebf\u5f3a\u5316\u5b66\u4e60\u57fa\u51c6\u4e0a\u8fdb\u884c\u4e86\u7cfb\u7edf\u9a8c\u8bc1\u3002\u7ed3\u679c\u663e\u793a\uff0c\u5373\u4f7f\u5728\u666e\u901a\u7b14\u8bb0\u672c CPU \u73af\u5883\u4e0b\uff0c\u7ecf\u8fc7\u4e60\u60ef\u5316\u5904\u7406\u7684\u6a21\u578b\u4ecd\u80fd\u4fdd\u6301800Hz\u4ee5\u4e0a\u7684\u51b3\u7b56\u9891\u7387\uff0c\u663e\u8457\u8d85\u8d8a\u73b0\u6709\u6269\u6563\u89c4\u5212\u65b9\u6cd5\u7684\u63a8\u7406\u6548\u7387\u3002\u66f4\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6a21\u578b\u5728\u52a0\u901f\u63a8\u7406\u7684\u540c\u65f6\uff0c\u6027\u80fd\u4e0e\u539f\u59cb\u6269\u6563\u6a21\u578b\u57fa\u672c\u6301\u5e73\uff0c\u90e8\u5206\u4efb\u52a1\u4e2d\u751a\u81f3\u5c55\u73b0\u51fa\u66f4\u4f18\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n

\u8be5\u7814\u7a76\u8fd8\u4ece\u8de8\u5b66\u79d1\u89c6\u89d2\u8fdb\u884c\u4e86\u6df1\u5165\u63a2\u8ba8\uff1a\u5728\u751f\u7269\u5b66\u5c42\u9762\uff0c\u901a\u8fc7\u7c7b\u6bd4\u5927\u8111\u4e60\u60ef\u5316\u673a\u5236\u9610\u91ca\u4e86\u6a21\u578b\u7684\u884c\u4e3a\u8fc1\u79fb\u539f\u7406\uff1b\u5728\u5de5\u7a0b\u5c42\u9762\uff0c\u7cfb\u7edf\u8bc4\u4f30\u4e86\u65b0\u65b9\u6cd5\u5728\u4e0d\u540c\u4efb\u52a1\u548c\u67b6\u6784\u4e0b\u7684\u9c81\u68d2\u6027\u3002\u8fd9\u9879\u7814\u7a76\u4e0d\u4ec5\u4e3a\u6269\u6563\u89c4\u5212\u7684\u6548\u7387\u4f18\u5316\u63d0\u4f9b\u4e86\u521b\u65b0\u89e3\u51b3\u65b9\u6848\uff0c\u66f4\u4e3a\u590d\u6742\u51b3\u7b56\u6a21\u578b\u7684\u5b9e\u7528\u5316\u843d\u5730\u5f00\u8f9f\u4e86\u65b0\u7684\u6280\u672f\u8def\u5f84\u3002<\/p>\n\n\n\n

04. \u7ec4\u5408\u591a\u81c2\u8001\u864e\u673a\u7684\u79bb\u7ebf\u5b66\u4e60\u65b9\u6cd5<\/h3>\n\n\n\n
\"paper\"<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2501.19300 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u7ec4\u5408\u591a\u81c2\u8001\u864e\u673a\uff08Combinatorial Multi-armed Bandit, CMAB\uff09\u662f\u4e00\u4e2a\u91cd\u8981\u7684\u5e8f\u8d2f\u51b3\u7b56\u6846\u67b6\uff0c\u8fc7\u53bb\u5341\u5e74\u4e2d\u53d7\u5230\u5e7f\u6cdb\u5173\u6ce8\u3002\u4f46\u73b0\u6709\u7814\u7a76\u5927\u591a\u96c6\u4e2d\u5728\u5728\u7ebf\u5b66\u4e60\u573a\u666f\uff0c\u5f80\u5f80\u5ffd\u89c6\u4e86\u5728\u7ebf\u4ea4\u4e92\u5e26\u6765\u7684\u9ad8\u6602\u6210\u672c\u4ee5\u53ca\u73b0\u5b9e\u4e2d\u5927\u91cf\u53ef\u7528\u7684\u79bb\u7ebf\u6570\u636e\u3002\u4e3a\u6b64\uff0c\u672c\u8bba\u6587\u63d0\u51fa\u4e86\u9996\u4e2a\u4e13\u4e3a CMAB \u8bbe\u8ba1\u7684\u79bb\u7ebf\u5b66\u4e60\u6846\u67b6 Off-CMAB\uff0c\u65e8\u5728\u5145\u5206\u5229\u7528\u5df2\u6709\u6570\u636e\uff0c\u964d\u4f4e\u5b66\u4e60\u6210\u672c\uff0c\u540c\u65f6\u4fdd\u6301\u9ad8\u6548\u7684\u7b56\u7565\u4f18\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n

Off-CMAB \u7684\u6838\u5fc3\u662f\u7ec4\u5408\u4e0b\u7f6e\u4fe1\u754c\uff08CLCB\uff09\u7b97\u6cd5\u3002\u8be5\u7b97\u6cd5\u7ed3\u5408\u4e86\u60b2\u89c2\u7684\u5956\u52b1\u4f30\u8ba1\u548c\u7ec4\u5408\u4f18\u5316\u6c42\u89e3\u5668\uff0c\u80fd\u591f\u5728\u4e0d\u8fdb\u884c\u5728\u7ebf\u4ea4\u4e92\u7684\u524d\u63d0\u4e0b\uff0c\u4ece\u79bb\u7ebf\u6570\u636e\u4e2d\u5b66\u4e60\u51fa\u63a5\u8fd1\u6700\u4f18\u7684\u7b56\u7565\u3002\u4e3a\u8bc4\u4f30\u79bb\u7ebf\u6570\u636e\u7684\u8d28\u91cf\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e24\u4e2a\u65b0\u7684\u6570\u636e\u8986\u76d6\u6761\u4ef6\uff0c\u5e76\u5728\u7406\u8bba\u4e0a\u8fdb\u884c\u4e86\u8bc1\u660e\uff1a\u5728\u6ee1\u8db3\u8fd9\u4e9b\u6761\u4ef6\u7684\u60c5\u51b5\u4e0b\uff0cCLCB \u7b97\u6cd5\u53ef\u4ee5\u8fbe\u5230\u8fd1\u4e4e\u6700\u4f18\u7684\u6b21\u4f18\u6027\u95f4\u9699\uff0c\u5176\u6027\u80fd\u4e0e\u7406\u8bba\u4e0b\u754c\u76f8\u5dee\u4ec5\u4e00\u4e2a\u5bf9\u6570\u56e0\u5b50\u3002<\/p>\n\n\n\n

\u8be5\u65b9\u6cd5\u5728\u591a\u4e2a\u5b9e\u9645\u5e94\u7528\u4e2d\u5f97\u5230\u4e86\u9a8c\u8bc1\uff0c\u5305\u62ec\u6392\u5e8f\u5b66\u4e60\u3001LLMs \u7f13\u5b58\u7b56\u7565\u4f18\u5316\uff0c\u4ee5\u53ca\u793e\u4ea4\u5f71\u54cd\u6700\u5927\u5316\u7b49\u4efb\u52a1\u3002\u5b9e\u9a8c\u8868\u660e\uff0cCLCB \u4e0d\u4ec5\u80fd\u5904\u7406\u975e\u7ebf\u6027\u5956\u52b1\u51fd\u6570\u548c\u590d\u6742\u7684\u53cd\u9988\u673a\u5236\uff0c\u8fd8\u80fd\u5e94\u5bf9\u5206\u5e03\u5916\u7684\u52a8\u4f5c\u6837\u672c\uff0c\u5373\u4f7f\u8fd9\u4e9b\u6837\u672c\u4e0d\u5305\u542b\u6700\u4f18\u751a\u81f3\u53ef\u884c\u7684\u52a8\u4f5c\u3002\u5728\u591a\u4e2a\u5408\u6210\u548c\u771f\u5b9e\u6570\u636e\u96c6\u4e0a\u7684\u5e7f\u6cdb\u5b9e\u9a8c\u8fdb\u4e00\u6b65\u8bc1\u660e\u4e86 CLCB \u7684\u4f18\u8d8a\u6027\u80fd\uff0c\u5176\u5728\u51c6\u786e\u6027\u548c\u7a33\u5b9a\u6027\u65b9\u9762\u5747\u4f18\u4e8e\u73b0\u6709\u65b9\u6cd5\u3002<\/p>\n\n\n\n

\"chart\"
\u56fe4\uff1aLLMs \u7f13\u5b58\u7b56\u7565\u4f18\u5316\u95ee\u9898\u548c\u6392\u5e8f\u5b66\u4e60\u95ee\u9898\u4e2d\uff0c\u5404\u7b97\u6cd5\u7684\u6b21\u4f18\u6027\u95f4\u9699\u6bd4\u8f83<\/figcaption><\/figure>\n\n\n\n

05. \u7b56\u7565\u8fc7\u6ee4\u673a\u5236\u63d0\u5347RLHF\u5956\u52b1\u4fe1\u53f7\u8d28\u91cf<\/h3>\n\n\n\n
\"paper\"<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2409.06957 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

RLHF \u5df2\u6210\u4e3a\u5bf9\u5927\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u5bf9\u9f50\u5fae\u8c03\u7684\u6838\u5fc3\u6280\u672f\u3002\u5176\u5173\u952e\u5728\u4e8e\u901a\u8fc7\u5956\u52b1\u6a21\u578b\u6a21\u62df\u4eba\u7c7b\u504f\u597d\uff0c\u5f15\u5bfc\u6a21\u578b\u751f\u6210\u66f4\u7b26\u5408\u4eba\u7c7b\u4ef7\u503c\u89c2\u7684\u54cd\u5e94\u3002\u7136\u800c\uff0c\u5728\u4ee3\u7801\u751f\u6210\u548c\u6570\u5b66\u63a8\u7406\u7b49\u9700\u8981\u7cbe\u786e\u5224\u65ad\u7684\u4efb\u52a1\u4e2d\uff0c\u5956\u52b1\u6a21\u578b\u7ed9\u51fa\u7684\u4fe1\u53f7\u4e2d\u901a\u5e38\u5939\u6742\u4e86\u5927\u91cf\u7684\u566a\u58f0\uff0c\u8fd9\u662f\u56e0\u4e3a\u5956\u52b1\u6a21\u578b\u5bb9\u6613\u53d7\u5230\u8f93\u5165\u5206\u5e03\u3001\u6a21\u578b\u504f\u5dee\u548c\u8bad\u7ec3\u6570\u636e\u7a00\u758f\u6027\u7b49\u65b9\u9762\u7684\u5f71\u54cd\u3002\u8fd9\u79cd\u566a\u58f0\u4e0d\u4ec5\u964d\u4f4e\u4e86\u8bad\u7ec3\u6548\u7387\uff0c\u8fd8\u53ef\u80fd\u8bef\u5bfc\u7b56\u7565\u5b66\u4e60\uff0c\u6700\u7ec8\u5f71\u54cd\u6a21\u578b\u6027\u80fd\u3002<\/p>\n\n\n\n

\u7814\u7a76\u5458\u4eec\u53d1\u73b0\uff0c\u5956\u52b1\u6a21\u578b\u7ed9\u51fa\u7684\u5956\u52b1\u4fe1\u53f7\u7684\u53ef\u9760\u6027\u968f\u7740\u5956\u52b1\u533a\u95f4\u7684\u4e0d\u540c\u800c\u5177\u6709\u660e\u663e\u5dee\u5f02\u3002\u7531\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a\u201c\u7b56\u7565\u8fc7\u6ee4\u201d\uff08Policy Filtration\uff09\u7684\u65b0\u65b9\u6cd5\uff0c\u5e76\u5c06\u5176\u96c6\u6210\u5230\u4e3b\u6d41\u7684 PPO \u7b97\u6cd5\u4e2d\uff0c\u5f62\u6210 PF-PPO\u3002\u8be5\u65b9\u6cd5\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u52a8\u6001\u8bc6\u522b\u5e76\u8fc7\u6ee4\u6389\u90a3\u4e9b\u5956\u52b1\u4e0d\u53ef\u9760\u7684\u6837\u672c\uff0c\u4ec5\u4fdd\u7559\u5956\u52b1\u6a21\u578b\u5728\u5176\u201c\u4fe1\u4efb\u533a\u57df\u201d\u5185\u7684\u6837\u672c\u7528\u4e8e\u7b56\u7565\u66f4\u65b0\u3002\u5177\u4f53\u505a\u6cd5\u662f\u901a\u8fc7\u8ba1\u7b97\u5956\u52b1\u4e0e\u53c2\u8003\u5f97\u5206\u4e4b\u95f4\u7684\u51b3\u5b9a\u7cfb\u6570\uff08R\u00b2\uff09\u6765\u8bc4\u4f30\u5956\u52b1\u7684\u53ef\u4fe1\u5ea6\uff0c\u5e76\u636e\u6b64\u786e\u5b9a\u6700\u4f18\u7684\u5956\u52b1\u8fc7\u6ee4\u533a\u95f4\u3002\u7814\u7a76\u53d1\u73b0\uff0c\u5956\u52b1\u6a21\u578b\u5728\u9ad8\u5206\u6bb5\u6216\u4e2d\u95f4\u5206\u6bb5\u7684\u7a33\u5b9a\u6027\u66f4\u5f3a\uff0c\u56e0\u6b64\u7b56\u7565\u8fc7\u6ee4\u673a\u5236\u80fd\u591f\u6709\u6548\u63d0\u5347\u8bad\u7ec3\u4fe1\u53f7\u7684\u8d28\u91cf\u3002<\/p>\n\n\n\n

\"chart,
\u56fe5\uff1a\u5956\u52b1\u6a21\u578b\u7684\u53ef\u9760\u6027\u5728\u4e0d\u540c\u5956\u52b1\u533a\u95f4\u5b58\u5728\u663e\u8457\u5dee\u5f02<\/figcaption><\/figure>\n\n\n\n

\u7814\u7a76\u5458\u4eec\u5728\u591a\u4e2a\u771f\u5b9e\u4efb\u52a1\u4e0a\u7684\u5b9e\u9a8c\u9a8c\u8bc1\u4e86PF-PPO\u7684\u6709\u6548\u6027\u3002PF-PPO \u5728 HumanEval\u3001MBPP \u548c LeetCode Contest \u7b49\u4ee3\u7801\u751f\u6210\u57fa\u51c6\u6d4b\u8bd5\u4ee5\u53ca Ape210K \u548c CMATH \u7b49\u6570\u5b66\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u90fd\u53d6\u5f97\u4e86\u660e\u663e\u7684\u6027\u80fd\u63d0\u5347\u3002\u800c\u4e14\uff0c\u8be5\u7b97\u6cd5\u80fd\u591f\u9002\u7528\u4e8e\u4e0d\u540c\u7684\u6a21\u578b\u67b6\u6784\uff08\u5982 deepseek-7B\uff0cqwen1.5-7B \u548c doubao-25B\uff09\uff0c\u5e76\u5728\u4e0d\u540c\u6a21\u578b\u4e0b\u5747\u8868\u73b0\u51fa\u663e\u8457\u7684\u80fd\u529b\u63d0\u5347\u3002\u6b64\u5916\uff0cPF-PPO \u8fd8\u5c55\u73b0\u51fa\u66f4\u5f3a\u7684\u8bad\u7ec3\u7a33\u5b9a\u6027\u548c\u66f4\u5feb\u7684\u6536\u655b\u901f\u5ea6\uff0c\u8868\u660e\u5176\u5728\u5b9e\u9645\u90e8\u7f72\u4e2d\u5177\u6709\u8f83\u9ad8\u7684\u5b9e\u7528\u4ef7\u503c\u3002<\/p>\n\n\n\n

\u8be5\u7814\u7a76\u8868\u660e\uff0c\u7b56\u7565\u8fc7\u6ee4\u662f\u4e00\u79cd\u7b80\u5355\u800c\u6709\u6548\u7684\u673a\u5236\uff0c\u80fd\u591f\u663e\u8457\u7f13\u89e3\u5956\u52b1\u6a21\u578b\u566a\u58f0\u5bf9 RLHF \u8bad\u7ec3\u7684\u5e72\u6270\u3002\u8be5\u65b9\u6cd5\u4e3a\u672a\u6765\u66f4\u9ad8\u6548\u3001\u66f4\u53ef\u9760\u7684 RLHF \u8bad\u7ec3\u6846\u67b6\u63d0\u4f9b\u4e86\u91cd\u8981\u542f\u793a\u3002<\/p>\n\n\n\n

06. \u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u4e2d\u9a6c\u5c14\u53ef\u592b\u6ce8\u610f\u529b\u673a\u5236\u5728\u79bb\u7ebf\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u4f5c\u7528\u63a2\u6790<\/h3>\n\n\n\n
\"paper\"<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2409.06985 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u8fd1\u5e74\u6765\uff0c\u79bb\u7ebf\u5f3a\u5316\u5b66\u4e60\uff08Offline Reinforcement Learning, Offline RL\uff09\u56e0\u5176\u65e0\u9700\u5728\u7ebf\u4ea4\u4e92\u5373\u53ef\u5b66\u4e60\u7b56\u7565\uff0c\u53d7\u5230\u4e86\u5e7f\u6cdb\u5173\u6ce8\u3002\u51b3\u7b56 Transformer\uff08Decision Transformer, DT\uff09\u4f5c\u4e3a\u4e00\u79cd\u80fd\u591f\u5c06\u5f3a\u5316\u5b66\u4e60\u95ee\u9898\u8f6c\u5316\u4e3a\u5e8f\u5217\u5efa\u6a21\u4efb\u52a1\u7684\u65b9\u6cd5\uff0c\u5728\u591a\u4e2a\u73af\u5883\u4e2d\u53d6\u5f97\u4e86\u663e\u8457\u6210\u679c\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\uff08\u5982 GPT\uff09\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u9886\u57df\u7684\u6210\u529f\uff0c\u6fc0\u53d1\u4e86\u79d1\u7814\u4eba\u5458\u5c06\u5176\u8fc1\u79fb\u81f3\u5f3a\u5316\u5b66\u4e60\u4efb\u52a1\u7684\u5174\u8da3\uff0c\u671f\u671b\u501f\u52a9\u5176\u5f3a\u5927\u7684\u8868\u793a\u80fd\u529b\u6765\u63d0\u5347\u7b56\u7565\u5b66\u4e60\u7684\u6548\u7387\u4e0e\u6cdb\u5316\u80fd\u529b\u3002\u7136\u800c\uff0c\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\uff0c\u867d\u7136\u8fd9\u7c7b\u8de8\u9886\u57df\u9884\u8bad\u7ec3\u6a21\u578b\u5728\u77ed\u671f\u4efb\u52a1\u4e2d\u8868\u73b0\u4f18\u5f02\uff0c\u4f46\u5728\u9700\u8981\u957f\u671f\u89c4\u5212\u80fd\u529b\u7684\u4efb\u52a1\u4e2d\u5374\u5e38\u5e38\u5931\u6548\uff0c\u751a\u81f3\u4e0d\u5982\u4ece\u96f6\u5f00\u59cb\u8bad\u7ec3\u7684\u6a21\u578b\u3002<\/p>\n\n\n\n

\u4e3a\u6b64\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u5f00\u59cb\u5bf9\u5176\u5185\u90e8\u673a\u5236\u8fdb\u884c\u6df1\u5165\u63a2\u7a76\u3002\u672c\u7bc7\u8bba\u6587\u805a\u7126\u4e8e\u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u6ce8\u610f\u529b\u673a\u5236\uff0c\u9996\u6b21\u63ed\u793a\u4e86\u201c\u9a6c\u5c14\u53ef\u592b\u6ce8\u610f\u529b\u5934\u201d\uff08Markov Heads\uff09\u5728\u5f3a\u5316\u5b66\u4e60\u4efb\u52a1\u4e2d\u7684\u5173\u952e\u4f5c\u7528\u3002\u7814\u7a76\u53d1\u73b0\uff0c\u8fd9\u4e9b\u6ce8\u610f\u529b\u5934\u503e\u5411\u4e8e\u4ec5\u5173\u6ce8\u5f53\u524d\u65f6\u95f4\u6b65\u53ca\u5176\u76f8\u90bb\u72b6\u6001\uff0c\u7136\u540e\u5f62\u6210\u4e00\u79cd\u8fd1\u4f3c\u9a6c\u5c14\u53ef\u592b\u6027\u7684\u6ce8\u610f\u529b\u5206\u5e03\u3002\u8fd9\u79cd\u7ed3\u6784\u5728\u77ed\u671f\u4efb\u52a1\u4e2d\u6709\u52a9\u4e8e\u5bf9\u5c40\u90e8\u4fe1\u606f\u7684\u5229\u7528\uff0c\u4f46\u5728\u957f\u671f\u4efb\u52a1\u4e2d\u5374\u9650\u5236\u4e86\u6a21\u578b\u5bf9\u8fdc\u7a0b\u4f9d\u8d56\u7684\u5efa\u6a21\u80fd\u529b\uff0c\u4ece\u800c\u5f71\u54cd\u7b56\u7565\u7684\u6574\u4f53\u8868\u73b0\u3002<\/p>\n\n\n\n

\u4e3a\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u901a\u7528\u7684\u589e\u5f3a\u65b9\u6cd5 GPT2-DTMA\uff08General Pre-trained Transformer with Decision Transformer Mixture of Attention\uff09\u3002\u8be5\u65b9\u6cd5\u5728\u5fae\u8c03\u9636\u6bb5\u5f15\u5165\u6df7\u5408\u6ce8\u610f\u529b\u673a\u5236\uff08Mixture of Attention, MoA\uff09\uff0c\u901a\u8fc7\u7ec4\u5408\u9a6c\u5c14\u53ef\u592b\u6ce8\u610f\u529b\u4e0e\u5168\u5c40\u6ce8\u610f\u529b\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u6839\u636e\u4efb\u52a1\u9700\u6c42\u81ea\u9002\u5e94\u5730\u8c03\u6574\u6ce8\u610f\u529b\u5206\u5e03\uff0c\u4ece\u800c\u517c\u987e\u77ed\u671f\u4e0e\u957f\u671f\u4efb\u52a1\u7684\u8868\u73b0\u3002\u8be5\u673a\u5236\u65e0\u9700\u4fee\u6539\u539f\u59cb\u6a21\u578b\u7ed3\u6784\uff0c\u4e14\u5177\u6709\u826f\u597d\u7684\u901a\u7528\u6027\u4e0e\u53ef\u6269\u5c55\u6027\u3002<\/p>\n\n\n\n

\"chart,
\u56fe6\uff1a\uff08\u5de6\uff09DT \u5728\u8bad\u7ec3\u540e\u5bf9\u4e8e\u90bb\u8fd1\u65f6\u95f4\u6b65\u4fe1\u606f\u7684\u5173\u6ce8\u663e\u8457\u63d0\u5347\uff1b\uff08\u53f3\uff09\u9884\u8bad\u7ec3\u6a21\u578b\u5728\u5fae\u8c03\u524d\u540e\u90fd\u6709\u76f8\u4f3c\u7684\u73b0\u8c61\uff0c\u4e14\u7814\u7a76\u5458\u4eec\u4ece\u7406\u8bba\u4e0a\u8bc1\u660e\u4e86\u8fd9\u79cd\u73b0\u8c61\u662f\u5fae\u8c03\u6a21\u578b\u65e0\u6cd5\u6539\u53d8\u7684\u3002<\/figcaption><\/figure>\n\n\n\n

\u7814\u7a76\u5458\u4eec\u5728\u591a\u4e2a\u6807\u51c6\u79bb\u7ebf\u5f3a\u5316\u5b66\u4e60\u57fa\u51c6\u73af\u5883\uff08\u5982 D4RL\uff09\u4e0a\u5bf9\u8be5\u65b9\u6cd5\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u7ed3\u679c\u8868\u660e\uff0cGPT2-DTMA \u5728\u77ed\u671f\u4efb\u52a1\u4e2d\u663e\u8457\u4f18\u4e8e\u4f20\u7edf DT \u548c\u5176\u4ed6\u9884\u8bad\u7ec3\u65b9\u6cd5\uff0c\u5728\u957f\u671f\u4efb\u52a1\u4e2d\u4e5f\u6709\u6548\u7f13\u89e3\u4e86\u9a6c\u5c14\u53ef\u592b\u6ce8\u610f\u529b\u5e26\u6765\u7684\u6027\u80fd\u74f6\u9888\uff0c\u8868\u73b0\u63a5\u8fd1\u751a\u81f3\u8d85\u8fc7\u4ece\u96f6\u8bad\u7ec3\u7684\u6a21\u578b\u3002\u6b64\u5916\uff0c\u5206\u6790\u8fd8\u8868\u660e\uff0cMoA \u673a\u5236\u80fd\u591f\u52a8\u6001\u8c03\u6574\u6ce8\u610f\u529b\u5934\u7684\u6743\u91cd\u5206\u5e03\uff0c\u63d0\u5347\u4e86\u6a21\u578b\u5bf9\u4e0d\u540c\u65f6\u95f4\u5c3a\u5ea6\u4fe1\u606f\u7684\u5efa\u6a21\u80fd\u529b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u7f16\u8005\u6309\uff1a\u6b22\u8fce\u9605\u8bfb\u201c\u79d1\u7814\u4e0a\u65b0\u201d\u680f\u76ee\uff01\u201c\u79d1\u7814\u4e0a\u65b0\u201d\u6c47\u805a\u4e86\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u6700\u65b0\u7684\u521b\u65b0\u6210\u679c\u4e0e\u79d1\u7814\u52a8\u6001\u3002\u5728\u8fd9\u91cc\uff0c\u4f60\u53ef\u4ee5\u5feb\u901f\u6d4f\u89c8\u7814\u7a76\u9662\u7684\u4eae\u70b9\u8d44\u8baf\uff0c\u4fdd\u6301\u5bf9\u524d\u6cbf\u9886\u57df\u7684\u654f\u9510\u55c5\u89c9\u3002 7\u670813\u65e5\u81f37\u670819\u65e5\uff0c\u4eba\u5de5\u667a\u80fd\u3001\u673a\u5668\u5b66\u4e60\u9886\u57df\u5168\u7403\u9876\u7ea7\u7684\u5b66\u672f\u76db\u4f1a\u4e4b\u4e00 ICML \u5c06\u5728\u6e29\u54e5\u534e\u4e3e\u529e\u3002\u6765\u81ea\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u591a\u7bc7\u8bba\u6587\u5165\u9009\u3002\u5728\u63a5\u4e0b\u6765\u7684\u51e0\u5929\u91cc\uff0c\u6211\u4eec\u5c06\u901a\u8fc7\u4e24\u671f\u201c\u79d1\u7814\u4e0a\u65b0\u201d\u4e3a\u5927\u5bb6\u5e26\u6765\u591a\u7bc7\u7814\u7a76\u9662\u5165\u9009 ICML 2025 \u7684\u7cbe\u9009\u8bba\u6587\u89e3\u8bfb\u3002\u7b2c\u4e00\u671f\u7684\u7814\u7a76\u5de5\u4f5c\u4e3b\u8981\u5173\u6ce8\u51b3\u7b56\u6a21\u578b\uff0c\u6db5\u76d6\u5f3a\u5316\u5b66\u4e60\u3001RLHF\u3001\u6269\u6563\u6a21\u578b\u5efa\u6a21\u7b49\u65b9\u5411\u3002 \u672c\u671f\u5185\u5bb9\u901f\u89c8 01. BRIDGE\uff1a\u901a\u8fc7\u591a\u667a\u80fd\u4f53\u8fed\u4ee3\u4f18\u5316\u4e0e\u6269\u6563\u5efa\u6a21\uff0c\u5b9e\u73b0\u6587\u672c\u5f15\u5bfc\u7684\u65f6\u95f4\u5e8f\u5217\u751f\u6210 02. DPO\u4e0ePPO\u878d\u5408\uff1a\u7528\u4e8eRLHF\u7684\u5f3a\u5316\u8bcd\u5143\u4f18\u5316 03. \u5c06\u6269\u6563\u89c4\u5212\u4e60\u60ef\u5316\uff0c\u5b9e\u73b0\u9ad8\u6548\u4e14\u6709\u6548\u7684\u51b3\u7b56\u5236\u5b9a 04. \u7ec4\u5408\u591a\u81c2\u8001\u864e\u673a\u7684\u79bb\u7ebf\u5b66\u4e60\u65b9\u6cd5 05. \u7b56\u7565\u8fc7\u6ee4\u673a\u5236\u63d0\u5347RLHF\u5956\u52b1\u4fe1\u53f7\u8d28\u91cf 06. \u9884\u8bad\u7ec3\u8bed\u8a00\u6a21\u578b\u4e2d\u9a6c\u5c14\u53ef\u592b\u6ce8\u610f\u529b\u673a\u5236\u5728\u79bb\u7ebf\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u4f5c\u7528\u63a2\u6790 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2503.02445 (opens in new tab) \u9ad8\u8d28\u91cf\u7684\u65f6\u95f4\u5e8f\u5217\u751f\u6210\uff08Time-Series Generation, TSG\uff09\u5728\u91d1\u878d\u3001\u533b\u7597\u3001\u6c14\u8c61\u7b49\u8bf8\u591a\u9886\u57df\u90fd\u6709\u7740\u6781\u4e3a\u5e7f\u6cdb\u7684\u5e94\u7528\u3002\u7136\u800c\uff0c\u5f53\u524d\u7684\u8bb8\u591a\u65b9\u6cd5\u4e3b\u8981\u96c6\u4e2d\u5728\u65e0\u6761\u4ef6\u7684\u5355\u4e00\u9886\u57df\u751f\u6210\u4e0a\uff0c\u96be\u4ee5\u6ee1\u8db3\u73b0\u5b9e\u573a\u666f\u4e2d\u5bf9\u4e8e\u8de8\u9886\u57df\u3001\u53ef\u63a7\u6027\u5f3a\u7684\u65f6\u95f4\u5e8f\u5217\u751f\u6210\u9700\u6c42\u3002\u7279\u522b\u662f\u5728\u90a3\u4e9b\u9700\u8981\u7ed3\u5408\u9886\u57df\u77e5\u8bc6\u3001\u5b9e\u4f8b\u7279\u5f81\u6216\u64cd\u4f5c\u7ea6\u675f\u7684\u4efb\u52a1\u4e2d\uff0c\u4f20\u7edf\u65b9\u6cd5\u5f80\u5f80\u5728\u7075\u6d3b\u6027\u4e0e\u7cbe\u5ea6\u65b9\u9762\u5b58\u5728\u660e\u663e\u4e0d\u8db3\u3002 \u9274\u4e8e\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u521b\u65b0\u601d\u8def\u2014\u2014\u5229\u7528\u6587\u672c\u4f5c\u4e3a\u63a7\u5236\u4fe1\u53f7\u6765\u5f15\u5bfc\u65f6\u95f4\u5e8f\u5217\u751f\u6210\uff0c\u4ece\u800c\u63d0\u9ad8\u751f\u6210\u7ed3\u679c\u7684\u8bed\u4e49\u4e00\u81f4\u6027\u4e0e\u5b9a\u5236\u5316\u80fd\u529b\u3002\u6570\u636e\u7a00\u7f3a\u662f\u76ee\u524d\u6587\u672c\u5f15\u5bfc\u751f\u6210\u8fc7\u7a0b\u4e2d\u6240\u9762\u4e34\u7684\u5173\u952e\u6311\u6218\uff0c\u5bf9\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u591a\u667a\u80fd\u4f53\u6846\u67b6\uff0c\u4e13\u95e8\u7528\u4e8e\u81ea\u52a8\u5408\u6210\u4e30\u5bcc\u591a\u6837\u4e14\u771f\u5b9e\u53ef\u9760\u7684\u6587\u672c\u5230\u65f6\u95f4\u5e8f\u5217\uff08Text-to-TS\uff09\u6570\u636e\u96c6\u3002\u901a\u8fc7\u591a\u4e2a\u667a\u80fd\u4f53\u4e4b\u95f4\u7684\u534f\u540c\u4f18\u5316\uff0c\u8be5\u6846\u67b6\u53ef\u4ee5\u6a21\u62df\u4eba\u7c7b\u4e13\u5bb6\u7684\u6807\u6ce8\u8fc7\u7a0b\uff0c\u663e\u8457\u63d0\u5347\u6570\u636e\u7684\u8d28\u91cf\u4e0e\u591a\u6837\u6027\u3002 \u5728\u6b64\u57fa\u7840\u4e4b\u4e0a\uff0c\u7814\u7a76\u5458\u4eec\u53c8\u8fdb\u4e00\u6b65\u63d0\u51fa\u4e86 BRIDGE \u6846\u67b6\uff0c\u8fd9\u662f\u4e00\u79cd\u6df7\u5408\u5f0f\u7684\u6587\u672c\u63a7\u5236\u65f6\u95f4\u5e8f\u5217\u751f\u6210\u65b9\u6cd5\u3002BRIDGE \u6846\u67b6\u901a\u8fc7\u878d\u5408\u8bed\u4e49\u539f\u578b\u4e0e\u6587\u672c\u63cf\u8ff0\uff0c\u4e3a\u751f\u6210\u8fc7\u7a0b\u63d0\u4f9b\u4e86\u9886\u57df\u7ea7\u522b\u7684\u7cbe\u51c6\u5f15\u5bfc\u4fe1\u53f7\uff0c\u5e76\u4e14\u5de7\u5999\u5730\u7ed3\u5408\u4e86\u6269\u6563\u6a21\u578b\u4e0e\u4f18\u5316\u673a\u5236\uff0c\u5b9e\u73b0\u4e86\u9ad8\u4fdd\u771f\u5ea6\u4e0e\u9ad8\u53ef\u63a7\u6027\u7684\u65f6\u95f4\u5e8f\u5217\u751f\u6210\u3002 \u7814\u7a76\u5458\u4eec\u572812\u4e2a\u4e0d\u540c\u7684\u6570\u636e\u96c6\u4e0a\u5bf9 BRIDGE \u6846\u67b6\u8fdb\u884c\u4e86\u5168\u9762\u8bc4\u4f30\u3002\u7ed3\u679c\u663e\u793a\uff0cBRIDGE \u5728\u5176\u4e2d\u768411\u4e2a\u6570\u636e\u96c6\u4e0a\u5747\u53d6\u5f97\u4e86\u6700\u5148\u8fdb\u7684\u751f\u6210\u8d28\u91cf\u8868\u73b0\u3002\u5728\u63a7\u5236\u80fd\u529b\u65b9\u9762\uff0c\u76f8\u8f83\u4e8e\u4f20\u7edf\u65e0\u6587\u672c\u8f93\u5165\u7684\u751f\u6210\u65b9\u5f0f\uff0cBRIDGE \u5728\u5747\u65b9\u8bef\u5dee\uff08MSE\uff09\u4e0a\u5b9e\u73b0\u4e8612.52%\u7684\u663e\u8457\u63d0\u5347\uff0c\u5728\u5e73\u5747\u7edd\u5bf9\u8bef\u5dee\uff08MAE\uff09\u4e0a\u4e5f\u8fbe\u5230\u4e866.34%\u7684\u63d0\u5347\u3002 \u8fd9\u4e9b\u7ed3\u679c\u5145\u5206\u8868\u660e\uff0c\u6587\u672c\u4e0d\u4ec5\u80fd\u591f\u63d0\u4f9b\u4e30\u5bcc\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u8fd8\u80fd\u6709\u6548\u5730\u5d4c\u5165\u5b9e\u4f8b\u7ea7\u7684\u65f6\u95f4\u6a21\u5f0f\uff0c\u8fdb\u800c\u6781\u5927\u5730\u589e\u5f3a\u751f\u6210\u6a21\u578b\u7684\u8868\u73b0\u529b\u4e0e\u9002\u5e94\u6027\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2404.18922 (opens in new tab) \u5927\u8bed\u8a00\u6a21\u578b\u901a\u8fc7\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08Reinforcement Learning from Human Feedback, RLHF\uff09\u5b9e\u73b0\u4e86\u548c\u4eba\u7c7b\u7684\u4ef7\u503c\u5bf9\u9f50\u3002\u5728\u8fd9\u4e00\u8fc7\u7a0b\u4e2d\uff0c\u4f7f\u7528\u8fd1\u7aef\u7b56\u7565\u4f18\u5316\uff08Proximal Policy Optimization, PPO\uff09\u53ef\u4ee5\u6700\u5927\u5316\u7a00\u758f\u7684\u53e5\u5b50\u7ea7\u5956\u52b1\uff0c\u4f46 PPO \u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u5e38\u5e38\u9762\u4e34\u8bad\u7ec3\u4e0d\u7a33\u5b9a\u3001\u6837\u672c\u6548\u7387\u4f4e\u4e0b\u7b49\u95ee\u9898\u3002 […]<\/p>\n","protected":false},"author":34512,"featured_media":1144546,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":null,"footnotes":""},"research-area":[13556],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1144532","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-research-area-artificial-intelligence","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1144532","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":2,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1144532\/revisions"}],"predecessor-version":[{"id":1144547,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1144532\/revisions\/1144547"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1144546"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1144532"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1144532"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1144532"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1144532"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}