{"id":1144993,"date":"2025-07-16T03:11:56","date_gmt":"2025-07-16T10:11:56","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&p=1144993"},"modified":"2025-07-16T03:11:58","modified_gmt":"2025-07-16T10:11:58","slug":"new-arrival-in-research-29","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/new-arrival-in-research-29\/","title":{"rendered":"ICML\u4e0a\u65b0 | \u8ba9\u5927\u6a21\u578b\u66f4\u201c\u806a\u660e\u201d\u3001\u66f4\u5b89\u5168\u3001\u66f4\u9ad8\u6548"},"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 \u5728\u6e29\u54e5\u534e\u4e3e\u529e\u3002\u6765\u81ea\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u591a\u7bc7\u8bba\u6587\u5165\u9009\u3002\u4e0a\u4e00\u671f\u201cICML \u4e0a\u65b0\u201d<\/a>\u7cbe\u9009\u4ecb\u7ecd\u4e86\u4e0e\u51b3\u7b56\u6a21\u578b\u76f8\u5173\u7684\u7814\u7a76\u5de5\u4f5c\uff0c\u5185\u5bb9\u6db5\u76d6\u5f3a\u5316\u5b66\u4e60\u3001RLHF\u3001\u6269\u6563\u5efa\u6a21\u7b49\u65b9\u5411\u3002\u672c\u671f\u5c06\u805a\u7126\u6a21\u578b\u63a8\u7406\u673a\u5236\u3001\u591a\u6a21\u6001\u8ba4\u77e5\u3001\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u4e0e\u5b89\u5168\u5bf9\u9f50\u7b49\u5173\u952e\u6280\u672f\u3002<\/p>\n\n\n\n

\n\n\n\n

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

01. AdaptiveStep\uff1a\u57fa\u4e8e\u6a21\u578b\u7f6e\u4fe1\u5ea6\u7684\u81ea\u52a8\u63a8\u7406\u6b65\u9aa4\u5212\u5206<\/p>\n\n\n\n

02. \u4ece\u590d\u6742\u6027\u5230\u539f\u5b50\u6027\uff0c\u901a\u8fc7\u77e5\u8bc6\u611f\u77e5\u7684\u53cc\u91cd\u91cd\u5199\u4e0e\u63a8\u7406\u673a\u5236\u63d0\u5347LLMs\u751f\u6210\u80fd\u529b<\/p>\n\n\n\n

03. \u5728\u7a7a\u95f4\u63a8\u7406\u4e2d\u60f3\u8c61\uff1a\u591a\u6a21\u6001\u601d\u7ef4\u53ef\u89c6\u5316<\/p>\n\n\n\n

04. LongRoPE2\uff1a\u8fd1\u4e4e\u65e0\u635f\u7684\u5927\u8bed\u8a00\u6a21\u578b\u4e0a\u4e0b\u6587\u7a97\u53e3\u6269\u5c55\u65b9\u6cd5<\/p>\n\n\n\n

05. MMInference\uff1a\u901a\u8fc7\u6a21\u6001\u611f\u77e5\u7f6e\u6362\u7a00\u758f\u6ce8\u610f\u529b\uff0c\u52a0\u901f\u957f\u4e0a\u4e0b\u6587\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u7684\u9884\u586b\u5145\u9636\u6bb5<\/p>\n\n\n\n

06. \u514b\u670d\u8bcd\u8868\u4e0d\u5339\u914d\uff1a\u8bcd\u8868\u65e0\u5173\u7684\u6559\u5e08\u5f15\u5bfc\u8bed\u8a00\u5efa\u6a21\u65b9\u6cd5<\/p>\n\n\n\n

07. \u63d0\u5347\u6807\u51c6\uff0c\u901a\u8fc7\u751f\u6210\u5f0f\u81ea\u8fdb\u5316\u8bc4\u4f30\uff0c\u63a2\u7a76\u5927\u8bed\u8a00\u6a21\u578b\u7684\u4ef7\u503c\u89c2<\/p>\n\n\n\n

08. \u8106\u5f31\u6027\u611f\u77e5\u5bf9\u9f50\uff1a\u7f13\u89e3\u6709\u5bb3\u5fae\u8c03\u4e2d\u7684\u975e\u5747\u5300\u9057\u5fd8\u95ee\u9898<\/p>\n\n\n\n

01. AdaptiveStep\uff1a\u57fa\u4e8e\u6a21\u578b\u7f6e\u4fe1\u5ea6\u7684\u81ea\u52a8\u63a8\u7406\u6b65\u9aa4\u5212\u5206<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2502.13943 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u968f\u7740\u6a21\u578b\u80fd\u529b\u7684\u63d0\u5347\u548c\u6280\u672f\u7684\u53d1\u5c55\uff0c\u8bed\u8a00\u6a21\u578b\u53ef\u4ee5\u5728\u8bf8\u591a\u4efb\u52a1\u4e0a\u90fd\u53d6\u5f97\u8f83\u9ad8\u5206\u6570\u3002GPT-o1 \u548c DeepSeek-R1 \u7684\u53d1\u5e03\uff0c\u51ed\u501f\u5f3a\u63a8\u7406\u80fd\u529b\u8ba9\u8bed\u8a00\u6a21\u578b\u201c\u51fa\u5708\u201d\u3002\u63d0\u5347\u6a21\u578b\u63a8\u7406\u80fd\u529b\u7684\u4e00\u4e2a\u57fa\u7840\u6280\u672f\u662f\u601d\u7ef4\u94fe\u65b9\u6cd5\uff08Chain-of-Thought, CoT\uff09\uff0c\u8be5\u65b9\u6cd5\u80fd\u591f\u5c06\u6a21\u578b\u89e3\u51b3\u95ee\u9898\u7684\u8fc7\u7a0b\u5206\u6210\u591a\u4e2a\u6b65\u9aa4\uff0c\u4ece\u800c\u89e3\u51b3\u590d\u6742\u7684\u63a8\u7406\u4efb\u52a1\u3002<\/p>\n\n\n\n

\u7136\u800c\uff0c\u5f53\u524d\u5c06\u6a21\u578b\u63a8\u7406\u6b65\u9aa4\u8fdb\u884c\u5212\u5206\u591a\u91c7\u7528\u4e24\u79cd\u65b9\u6cd5\uff1a\u4e00\u79cd\u662f\u4eba\u5de5\u8fdb\u884c\u6b65\u9aa4\u5212\u5206\uff0c\u624b\u5de5\u5212\u5206\u7684\u6b65\u9aa4\u4fe1\u606f\u542b\u91cf\u6bd4\u8f83\u9ad8\u4e14\u5747\u8861\uff0c\u4f46\u96be\u4ee5\u81ea\u52a8\u5316\uff0c\u5212\u5206\u4f9d\u8d56\u4e8e\u4e13\u5bb6\u7684\u9886\u57df\u77e5\u8bc6\uff1b\u53e6\u4e00\u79cd\u662f\u57fa\u4e8e\u89c4\u5219\u7684\u5212\u5206\uff0c\u5982\u56fa\u5b9a\u884c\u6570\u6216 token \u6570\uff0c\u8fd9\u4e00\u65b9\u6cd5\u53ef\u4ee5\u81ea\u52a8\u5316\uff0c\u4f46\u5bb9\u6613\u505a\u51fa\u6ca1\u6709\u610f\u4e49\u6216\u4e0d\u5408\u7406\u7684\u5212\u5206\u3002<\/p>\n\n\n\n

\u4e3a\u89e3\u51b3\u8fd9\u4e00\u96be\u9898\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u521b\u65b0\u65b9\u6cd5 AdaptiveStep\u3002\u5176\u6838\u5fc3\u5728\u4e8e\u501f\u52a9\u8bed\u8a00\u6a21\u578b\u9884\u6d4b\u4e0b\u4e00 token \u65f6\u7684\u7f6e\u4fe1\u5ea6\u53d8\u5316\uff0c\u81ea\u52a8\u8bc6\u522b\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u81ea\u7136\u51b3\u7b56\u70b9\uff0c\u8fdb\u800c\u5b9e\u73b0\u66f4\u5408\u7406\u7684\u6b65\u9aa4\u5212\u5206\u3002<\/p>\n\n\n\n

\"graphical
\u56fe1\uff1a\u4e09\u79cd\u63a8\u7406\u6b65\u9aa4\u5212\u5206\u65b9\u6cd5\u7684\u6bd4\u8f83\uff0c\u4ece\u4e0a\u81f3\u4e0b\u5206\u522b\u4e3a\uff1a\u4eba\u5de5\u5212\u5206\u3001\u6309\u7167\u89c4\u5219\u5212\u5206\u4e0e AdaptiveStep \u65b9\u6cd5<\/figcaption><\/figure>\n\n\n\n

AdaptiveStep \u65e0\u9700\u4eba\u5de5\u6807\u6ce8\uff0c\u5b8c\u5168\u4f9d\u636e\u6a21\u578b\u81ea\u8eab\u7684\u9884\u6d4b\u884c\u4e3a\u6765\u8fdb\u884c\u63a8\u7406\u6b65\u9aa4\u7684\u81ea\u9002\u5e94\u5212\u5206\u3002\u5177\u4f53\u64cd\u4f5c\u4e3a\uff1a\u5728\u6bcf\u4e2a token \u4f4d\u7f6e\u8ba1\u7b97\u6a21\u578b\u7684\u7f6e\u4fe1\u5ea6\u53d8\u5316\uff0c\u4e00\u65e6\u7f6e\u4fe1\u5ea6\u663e\u8457\u4e0b\u964d\uff0c\u4fbf\u5c06\u5176\u89c6\u4e3a\u6f5c\u5728\u7684\u63a8\u7406\u5206\u754c\u70b9\u3002\u8fd9\u79cd\u5212\u5206\u65b9\u5f0f\u80fd\u591f\u7cbe\u51c6\u6355\u6349\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u5173\u952e\u8f6c\u6298\uff0c\u663e\u8457\u63d0\u5347\u6bcf\u4e00\u6b65\u7684\u51b3\u7b56\u4fe1\u606f\u5bc6\u5ea6\u3002<\/p>\n\n\n\n

\u7814\u7a76\u5458\u4eec\u5c06 AdaptiveStep \u5e94\u7528\u4e8e Process Reward Model (PRM) \u7684\u8bad\u7ec3\u4e2d\uff0c\u5e76\u5728\u6570\u5b66\u63a8\u7406\u3001\u4ee3\u7801\u751f\u6210\u7b49\u4efb\u52a1\u4e2d\u8fdb\u884c\u4e86\u7cfb\u7edf\u8bc4\u4f30\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u7ecf AdaptiveStep \u8bad\u7ec3\u7684 PRM \u5728\u591a\u9879\u4efb\u52a1\u4e2d\u5747\u53d6\u5f97\u4e86\u9886\u5148\u6027\u80fd\u3002\u5728 Best-of-N \u8bc4\u4f30\u4e2d\uff0c\u8be5\u65b9\u6cd5\u5927\u5e45\u4f18\u4e8e\u57fa\u4e8e\u8d2a\u5a6a\u641c\u7d22\u548c token \u7ea7\u4ef7\u503c\u5f15\u5bfc\u89e3\u7801\u7684\u7b56\u7565\uff0c\u4e14\u5728\u7ef4\u6301\u6027\u80fd\u7684\u540c\u65f6\uff0c\u8fd8\u5c06\u6784\u5efa PRM \u7684\u8ba1\u7b97\u6210\u672c\u964d\u4f4e\u4e86 30% \u4ee5\u4e0a\u3002\u6b64\u5916\uff0c\u7814\u7a76\u8fd8\u8bc1\u5b9e\u4e86 AdaptiveStep \u5728\u4e0d\u540c\u4efb\u52a1\u548c\u6a21\u578b\u67b6\u6784\u4e2d\u5177\u6709\u51fa\u8272\u7684\u8fc1\u79fb\u80fd\u529b\u4e0e\u6cdb\u5316\u6027\uff0c\u5c55\u73b0\u51fa\u4e86\u826f\u597d\u7684\u5b9e\u7528\u6027\u548c\u6269\u5c55\u6027\u3002<\/p>\n\n\n\n

02. \u4ece\u590d\u6742\u6027\u5230\u539f\u5b50\u6027\uff0c\u901a\u8fc7\u77e5\u8bc6\u611f\u77e5\u7684\u53cc\u91cd\u91cd\u5199\u4e0e\u63a8\u7406\u673a\u5236\u63d0\u5347LLMs\u751f\u6210\u80fd\u529b<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/openreview.net\/pdf?id=PAjCdkkNaU (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u5728\u5f53\u524d\u7684\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08Retrieval-Augmented Generation, RAG\uff09\u7cfb\u7edf\u4e2d\uff0c\u5c3d\u7ba1\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u80fd\u529b\u901a\u8fc7\u5916\u90e8\u77e5\u8bc6\u68c0\u7d22\u83b7\u5f97\u4e86\u6781\u5927\u7684\u63d0\u5347\uff0c\u4f46\u5728\u5904\u7406\u590d\u6742\u3001\u591a\u8df3\u63a8\u7406\u4efb\u52a1\u65f6\u4ecd\u9762\u4e34\u6311\u6218\u3002\u4f20\u7edf\u65b9\u6cd5\u901a\u5e38\u4f9d\u8d56\u4e00\u6b21\u6027\u68c0\u7d22\uff0c\u96be\u4ee5\u5145\u5206\u6316\u6398\u9886\u57df\u7279\u5b9a\u7684\u6df1\u5c42\u77e5\u8bc6\uff0c\u4e5f\u96be\u4ee5\u6784\u5efa\u8fde\u8d2f\u7684\u63a8\u7406\u94fe\u6761\u3002\u8fd9\u79cd\u5c40\u9650\u6027\u5bfc\u81f4\u6a21\u578b\u5728\u9762\u5bf9\u9700\u8981\u591a\u6b65\u903b\u8f91\u63a8\u7406\u7684\u95ee\u9898\u65f6\u8868\u73b0\u4e0d\u4f73\uff0c\u5c24\u5176\u662f\u5728\u79d1\u5b66\u95ee\u7b54\u3001\u6cd5\u5f8b\u5206\u6790\u7b49\u77e5\u8bc6\u5bc6\u96c6\u578b\u4efb\u52a1\u4e2d\u3002<\/p>\n\n\n\n

\u4e3a\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u66f4\u7ec6\u7c92\u5ea6\u7684\u77e5\u8bc6\u5904\u7406\u65b9\u5f0f\uff0c\u901a\u8fc7\u5177\u5907\u77e5\u8bc6\u611f\u77e5\u80fd\u529b\u7684\u589e\u5f3a\u751f\u6210\u6846\u67b6 KAR3-RAG\uff0c\u6765\u63d0\u5347\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u548c\u751f\u6210\u8d28\u91cf\u3002\u8be5\u65b9\u6cd5\u7684\u6838\u5fc3\u601d\u60f3\u662f\u5c06\u590d\u6742\u4efb\u52a1\u5206\u89e3\u4e3a\u539f\u5b50\u7ea7\u522b\u7684\u77e5\u8bc6\u5355\u5143\uff0c\u5e76\u901a\u8fc7\u53cc\u91cd\u91cd\u5199\u4e0e\u63a8\u7406\u673a\u5236\u9010\u6b65\u6784\u5efa\u5b8c\u6574\u7684\u7b54\u6848\u3002\u8be5\u65b9\u6cd5\u5305\u542b\u56db\u4e2a\u5173\u952e\u6a21\u5757\uff1a\u77e5\u8bc6\u539f\u5b50\u5316\u5668\uff08knowledge atomizer\uff09\u5c06\u539f\u59cb\u95ee\u9898\u62c6\u89e3\u4e3a\u591a\u4e2a\u539f\u5b50\u95ee\u9898\uff1b\u67e5\u8be2\u751f\u6210\u5668\uff08query generator\uff09\u6839\u636e\u5df2\u77e5\u4fe1\u606f\u5bf9\u539f\u95ee\u9898\u8fdb\u884c\u5206\u522b\u5e76\u63d0\u51fa\u540e\u7eed\u95ee\u9898\uff1b\u539f\u5b50\u68c0\u7d22\u5668\uff08atom retriever\uff09\u4f9d\u636e\u539f\u5b50\u77e5\u8bc6\u5bf9\u9f50\u65b9\u5f0f\u8fdb\u884c\u4fe1\u606f\u68c0\u7d22\uff1b\u539f\u5b50\u9009\u62e9\u5668\uff08atom selector\uff09\u5219\u6839\u636e\u68c0\u7d22\u7ed3\u679c\u51b3\u5b9a\u4e0b\u4e00\u6b65\u7684\u63d0\u95ee\u65b9\u5411\u3002\u6574\u4e2a\u8fc7\u7a0b\u5f62\u6210\u4e86\u4e00\u4e2a\u52a8\u6001\u7684\u77e5\u8bc6\u6784\u5efa\u4e0e\u63a8\u7406\u5faa\u73af\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u5728\u591a\u8f6e\u4ea4\u4e92\u4e2d\u9010\u6b65\u6784\u5efa\u5408\u7406\u7684\u63a8\u7406\u8def\u5f84\u3002<\/p>\n\n\n\n

\"diagram\"
\u56fe2\uff1aKAR3-RAG \u5de5\u4f5c\u6d41\u7a0b\u56fe<\/figcaption><\/figure>\n\n\n\n

\u7814\u7a76\u5458\u4eec\u5728\u591a\u4e2a\u9700\u8981\u591a\u8df3\u63a8\u7406\u7684\u57fa\u51c6\u4efb\u52a1\u4e0a\u8fdb\u884c\u4e86\u8bc4\u4f30\uff0c\u5305\u62ec HotpotQA\u3001MuSiQue \u548c two-wiki \u7b49\u6570\u636e\u96c6\u3002\u7ed3\u679c\u663e\u793a\uff0c\u8be5\u65b9\u6cd5\u5728\u51c6\u786e\u7387\u65b9\u9762\u76f8\u6bd4\u73b0\u6709\u6700\u4f18\u65b9\u6cd5\u63d0\u5347\u4e8612.6%\uff0c\u5728\u67d0\u4e9b\u6311\u6218\u6027\u4efb\u52a1\u4e2d\u751a\u81f3\u8fbe\u5230\u4e8620%\u7684\u63d0\u5347\u3002\u6b64\u5916\uff0c\u8be5\u65b9\u6cd5\u5728\u751f\u6210\u7b54\u6848\u7684\u903b\u8f91\u4e00\u81f4\u6027\u548c\u4fe1\u606f\u5b8c\u6574\u6027\u65b9\u9762\u4e5f\u8868\u73b0\u51fa\u663e\u8457\u4f18\u52bf\u3002\u901a\u8fc7\u539f\u5b50\u5316\u77e5\u8bc6\u5904\u7406\u4e0e\u8fed\u4ee3\u5f0f\u63a8\u7406\u6784\u5efa\uff0c\u8be5\u65b9\u6cd5\u4e0d\u4ec5\u63d0\u5347\u4e86\u68c0\u7d22\u6548\u7387\uff0c\u4e5f\u589e\u5f3a\u4e86\u751f\u6210\u7ed3\u679c\u7684\u903b\u8f91\u6027\u4e0e\u4e00\u81f4\u6027\uff0c\u5c55\u73b0\u51fa\u5728\u4e13\u4e1a\u9886\u57df\u77e5\u8bc6\u95ee\u7b54\u4e2d\u7684\u5e7f\u6cdb\u5e94\u7528\u6f5c\u529b\u3002<\/p>\n\n\n\n

03. \u5728\u7a7a\u95f4\u63a8\u7406\u4e2d\u60f3\u8c61\uff1a\u591a\u6a21\u6001\u601d\u7ef4\u53ef\u89c6\u5316<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2501.07542 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u5728\u590d\u6742\u7a7a\u95f4\u63a8\u7406\u4efb\u52a1\u4e2d\uff0c\u5c3d\u7ba1\u4f20\u7edf\u7684 LLMs \u548c\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\u5728\u8bed\u8a00\u7406\u89e3\u548c\u903b\u8f91\u63a8\u7406\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff0c\u4f46\u5728\u5904\u7406\u6d89\u53ca\u7a7a\u95f4\u5173\u7cfb\u548c\u52a8\u6001\u53d8\u5316\u7684\u4efb\u52a1\u65f6\u4ecd\u9762\u4e34\u663e\u8457\u6311\u6218\u3002\u73b0\u6709\u7684 CoT \u63d0\u793a\u65b9\u6cd5\u867d\u80fd\u63d0\u5347\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4f46\u4ec5\u4f9d\u8d56\u8bed\u8a00\u8868\u8fbe\uff0c\u96be\u4ee5\u6355\u6349\u7a7a\u95f4\u7ed3\u6784\u548c\u89c6\u89c9\u7ebf\u7d22\u3002\u800c\u4eba\u7c7b\u5728\u8ba4\u77e5\u8fc7\u7a0b\u4e2d\u5e38\u5e38\u662f\u5c06\u8bed\u8a00\u4e0e\u56fe\u50cf\u7ed3\u5408\u8d77\u6765\u8fdb\u884c\u601d\u8003\uff0c\u8fd9\u79cd\u201c\u8bed\u8a00-\u89c6\u89c9\u53cc\u7f16\u7801\u201d\u673a\u5236\u542f\u53d1\u4e86\u7814\u7a76\u5458\u4eec\uff0c\u5e76\u7531\u6b64\u63d0\u51fa\u4e00\u79cd\u65b0\u7684\u63a8\u7406\u8303\u5f0f\u2014\u2014\u591a\u6a21\u6001\u601d\u7ef4\u53ef\u89c6\u5316\uff08Multimodal Visualization-of-Thought, MVoT\uff09\u3002<\/p>\n\n\n\n

MVoT \u901a\u8fc7\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u751f\u6210\u56fe\u50cf\uff0c\u6a21\u62df\u4eba\u7c7b\u5728\u601d\u8003\u65f6\u7684\u89c6\u89c9\u8054\u60f3\uff0c\u4ece\u800c\u589e\u5f3a\u6a21\u578b\u7684\u7a7a\u95f4\u7406\u89e3\u80fd\u529b\u3002\u4e3a\u4e86\u63d0\u5347\u751f\u6210\u56fe\u50cf\u7684\u8fde\u8d2f\u6027\u4e0e\u51c6\u786e\u6027\uff0c\u7814\u7a76\u5458\u4eec\u5728\u81ea\u56de\u5f52\u5f0f\u591a\u6a21\u6001\u6a21\u578b\u4e2d\u5f15\u5165\u4e86 \u201ctoken discrepancy loss\u201d \u673a\u5236\uff0c\u8fdb\u800c\u4f7f\u6a21\u578b\u5728\u8bed\u8a00\u4e0e\u89c6\u89c9\u4e4b\u95f4\u5efa\u7acb\u66f4\u7d27\u5bc6\u7684\u8054\u7cfb\u3002MVoT \u4e0d\u4ec5\u80fd\u751f\u6210\u8bed\u8a00\u63a8\u7406\u94fe\u6761\uff0c\u8fd8\u80fd\u540c\u6b65\u751f\u6210\u4e0e\u4e4b\u5bf9\u5e94\u7684\u89c6\u89c9\u63a8\u7406\u8f68\u8ff9\uff0c\u5b9e\u73b0\u8de8\u6a21\u6001\u7684\u601d\u7ef4\u8868\u8fbe\u3002<\/p>\n\n\n\n

\"diagram\"
\u56fe3\uff1aMVoT \u4f7f\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\u80fd\u591f\u5728\u4e0d\u540c\u6a21\u6001\u4e4b\u95f4\u751f\u6210\u4ea4\u9519\u7684\u63a8\u7406\u8f68\u8ff9\u3002\u4f20\u7edf\u7684 CoT \u4ec5\u4f9d\u8d56\u8bed\u8a00\u601d\u7ef4\uff0c\u800c MVoT \u5219\u5f15\u5165\u89c6\u89c9\u601d\u7ef4\uff0c\u5c06\u63a8\u7406\u8fc7\u7a0b\u53ef\u89c6\u5316\u3002\u8fd9\u79cd\u63a8\u7406\u8303\u5f0f\u7c7b\u4f3c\u4e8e\u4eba\u7c7b\u7684\u8ba4\u77e5\u65b9\u5f0f\uff0c\u80fd\u591f\u5728\u8bed\u8a00\u4e0e\u56fe\u50cf\u4e4b\u95f4\u65e0\u7f1d\u5730\u8fdb\u884c\u601d\u8003\u3002<\/figcaption><\/figure>\n\n\n\n

\u5728\u591a\u4e2a\u52a8\u6001\u7a7a\u95f4\u63a8\u7406\u4efb\u52a1\u4e2d\uff0cMVoT \u5177\u6709\u663e\u8457\u4f18\u52bf\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u8be5\u65b9\u6cd5\u5728\u4f20\u7edf CoT \u65b9\u6cd5\u5931\u6548\u7684\u590d\u6742\u573a\u666f\u4e2d\u4f9d\u7136\u4fdd\u6301\u7a33\u5b9a\u6027\u80fd\uff0c\u7279\u522b\u662f\u5728\u9700\u8981\u7406\u89e3\u7269\u4f53\u4f4d\u7f6e\u53d8\u5316\u3001\u7a7a\u95f4\u5173\u7cfb\u63a8\u7406\u7b49\u4efb\u52a1\u4e2d\u8868\u73b0\u5c24\u4e3a\u7a81\u51fa\u3002\u4e0e\u73b0\u6709\u7684\u4e24\u9636\u6bb5\u89c6\u89c9\u63a8\u7406\u65b9\u6cd5\uff08\u5982\u56fe\u50cf\u63cf\u8ff0\u3001\u573a\u666f\u56fe\u751f\u6210\u7b49\uff09\u76f8\u6bd4\uff0cMVoT \u65e0\u9700\u4f9d\u8d56\u5916\u90e8\u89c6\u89c9\u6a21\u5757\uff0c\u5177\u5907\u66f4\u5f3a\u7684\u7aef\u5230\u7aef\u63a8\u7406\u80fd\u529b\u3002<\/p>\n\n\n\n

04. LongRoPE2\uff1a\u8fd1\u4e4e\u65e0\u635f\u7684\u5927\u8bed\u8a00\u6a21\u578b\u4e0a\u4e0b\u6587\u7a97\u53e3\u6269\u5c55\u65b9\u6cd5<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2502.20082 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\u5728\u8bed\u8a00\u7406\u89e3\u3001\u63a8\u7406\u7b49\u4efb\u52a1\u4e2d\u53d6\u5f97\u7a81\u7834\u6027\u8fdb\u5c55\uff0c\u5982\u4f55\u63d0\u5347\u5176\u5728\u771f\u5b9e\u573a\u666f\u4e2d\u7684\u5b9e\u7528\u6027\u6210\u4e3a\u5f53\u524d\u7814\u7a76\u7684\u91cd\u8981\u65b9\u5411\u3002\u5176\u4e2d\uff0c\u6269\u5c55\u4e0a\u4e0b\u6587\u7a97\u53e3\u88ab\u5e7f\u6cdb\u8ba4\u4e3a\u662f\u63d0\u5347\u6a21\u578b\u63a8\u7406\u80fd\u529b\u548c\u4efb\u52a1\u9002\u5e94\u6027\u7684\u5173\u952e\u4e00\u73af\u3002\u5c24\u5176\u5728\u6570\u5b66\u548c\u7f16\u7a0b\u7b49\u9700\u8981\u591a\u6b65\u63a8\u7406\u7684\u4efb\u52a1\u4e2d\uff0c\u8d8a\u6765\u8d8a\u591a\u7684 test-time scaling \u65b9\u6cd5\u4f9d\u8d56\u4e8e\u66f4\u957f\u7684\u4e0a\u4e0b\u6587\uff0c\u4ee5\u652f\u6491\u66f4\u590d\u6742\u7684\u89e3\u9898\u7b56\u7565\u3002\u957f\u4e0a\u4e0b\u6587\u80fd\u529b\u6b63\u5728\u6210\u4e3a\u6784\u5efa\u5f3a\u63a8\u7406\u6a21\u578b\u7684\u6838\u5fc3\u80fd\u529b\u4e4b\u4e00\u3002\u7136\u800c\uff0c\u73b0\u6709\u4e0a\u4e0b\u6587\u6269\u5c55\u65b9\u6cd5\u5e38\u5e38\u9700\u8981\u8fdb\u884c\u6743\u8861\uff0c\u8981\u4e48\u727a\u7272\u77ed\u4e0a\u4e0b\u6587\u6027\u80fd\uff0c\u8981\u4e48\u5e26\u6765\u9ad8\u6602\u7684\u8bad\u7ec3\u6210\u672c\uff0c\u96be\u4ee5\u5b9e\u73b0\u9ad8\u6548\u3001\u65e0\u635f\u7684\u6269\u5c55\u3002<\/p>\n\n\n\n

\u5bf9\u6b64\uff0c\u672c\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u8fd1\u4e4e\u65e0\u635f\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\u6269\u5c55\u65b9\u6cd5 LongRoPE2\uff0c\u65e8\u5728\u5728\u4fdd\u6301\u539f\u59cb\u6027\u80fd\u7684\u540c\u65f6\uff0c\u5927\u5e45\u63d0\u5347\u6a21\u578b\u7684\u6709\u6548\u4e0a\u4e0b\u6587\u957f\u5ea6\u3002LongRoPE2 \u7684\u8bbe\u8ba1\u57fa\u4e8e\u4e09\u4e2a\u6838\u5fc3\u8d21\u732e\u3002\u9996\u5148\uff0c\u7814\u7a76\u5458\u4eec\u6307\u51fa\u73b0\u6709\u4e0a\u4e0b\u6587\u6269\u5c55\u65b9\u6cd5\u5728\u9ad8\u7ef4 RoPE \u8868\u793a\u4e0a\u7684\u8bad\u7ec3\u4e0d\u8db3\uff0c\u5bfc\u81f4\u5176\u6709\u6548\u65cb\u8f6c\u8303\u56f4\u5728\u539f\u59cb\u4e0a\u4e0b\u6587\u5185\u5c31\u5df2\u4f4e\u4e8e\u7406\u8bba\u9884\u671f\u3002\u901a\u8fc7\u8fdb\u4e00\u6b65\u5047\u8bbe\u7814\u7a76\u5458\u4eec\u53d1\u73b0\uff0c\u8fd9\u4e00\u8bad\u7ec3\u4e0d\u8db3\u662f\u9020\u6210\u5b9e\u9645\u65cb\u8f6c\u5468\u671f\u63d0\u524d\u3001\u5173\u952e RoPE \u7ef4\u5ea6\u66f4\u65e9\u8fdb\u5165\u5206\u5e03\u5916\uff08out-of-distribution, OOD\uff09\u533a\u57df\u7684\u6839\u672c\u539f\u56e0\uff0c\u4ece\u800c\u9650\u5236\u4e86\u73b0\u6709\u91cd\u7f29\u653e\u65b9\u6cd5\u5728\u6240\u6709\u7ef4\u5ea6\u4e0a\u5168\u9762\u7f13\u89e3\u4e0a\u4e0b\u6587\u6269\u5c55 OOD \u95ee\u9898\u7684\u80fd\u529b\u3002\u4e3a\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u53c8\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e needle-insertion perplexity \u7684\u8bc4\u4f30\u6307\u6807\uff0c\u5e76\u6784\u5efa\u4e86\u81ea\u52a8\u8fdb\u5316\u641c\u7d22\u7b97\u6cd5\uff0c\u5e0c\u671b\u5bfb\u627e\u66f4\u4f18\u7684\u91cd\u7f29\u653e\u66f2\u7ebf\u3002\u4e0e\u4ee5\u5f80\u4ec5\u5173\u6ce8\u7406\u8bba\u5468\u671f\u4e00\u81f4\u6027\u7684\u505a\u6cd5\u4e0d\u540c\uff0c\u8be5\u65b9\u6cd5\u901a\u8fc7\u5b9e\u9645\u63a8\u7406\u6027\u80fd\uff08perplexity\uff09\u5f15\u5bfc\u641c\u7d22\uff0c\u4f7f\u5f97\u6240\u9009\u91cd\u7f29\u653e\u7b56\u7565\u80fd\u591f\u66f4\u6709\u6548\u5730\u63d0\u5347\u6a21\u578b\u5728\u957f\u4e0a\u4e0b\u6587\u533a\u57df\u7684\u9c81\u68d2\u6027\u3002\u6700\u540e\uff0c\u4e3a\u5728\u589e\u5f3a\u957f\u4e0a\u4e0b\u6587\u80fd\u529b\u7684\u540c\u65f6\u4fdd\u6301\u77ed\u4e0a\u4e0b\u6587\u6027\u80fd\uff0c\u7814\u7a76\u5458\u4eec\u5f15\u5165\u4e86\u4e00\u79cd\u6df7\u5408\u4e0a\u4e0b\u6587\u7a97\u53e3\u8bad\u7ec3\u673a\u5236\uff1a\u5728\u5fae\u8c03\u9636\u6bb5\uff0c\u6a21\u578b\u4ea4\u66ff\u63a5\u53d7\u77ed\u4e0a\u4e0b\u6587\u548c\u957f\u4e0a\u4e0b\u6587\u8f93\u5165\uff0c\u5e76\u5206\u522b\u9002\u914d\u539f\u59cb RoPE \u548c\u91cd\u7f29\u653e RoPE\uff0c\u4ece\u800c\u7edf\u4e00\u5efa\u6a21\u4e0d\u540c\u5c3a\u5ea6\u7684\u4e0a\u4e0b\u6587\u3002\u8fd9\u79cd\u673a\u5236\u663e\u8457\u7f13\u89e3\u4e86\u4e0a\u4e0b\u6587\u6269\u5c55\u5e26\u6765\u7684\u77ed\u7a0b\u9000\u5316\u95ee\u9898\uff0c\u5b9e\u73b0\u4e86\u8fd1\u4e4e\u65e0\u635f\u7684\u4e0a\u4e0b\u6587\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n

\u5728 LLaMA3-8B \u548c Phi3-mini-3.8B \u6a21\u578b\u4e0a\u7684\u5e7f\u6cdb\u5b9e\u9a8c\uff0c\u9a8c\u8bc1\u4e86 LongRoPE2 \u7684\u6709\u6548\u6027\u3002\u7ed3\u679c\u663e\u793a\uff0cLongRoPE2 \u80fd\u591f\u5c06 LLaMA3-8B \u7684\u6709\u6548\u4e0a\u4e0b\u6587\u957f\u5ea6\u6269\u5c55\u81f3 128K\uff0c\u540c\u65f6\u4fdd\u7559\u8d85\u8fc798.5%\u7684\u77ed\u4e0a\u4e0b\u6587\u6027\u80fd\uff0c\u4e14\u4ec5\u4f7f\u7528 10B \u7684\u8bad\u7ec3\u6807\u8bb0\u3002\u8fd9\u4e00\u6210\u679c\u9a8c\u8bc1\u4e86\u7814\u7a76\u5458\u4eec\u7684\u5047\u8bbe\uff0c\u4e5f\u5c55\u793a\u4e86 LongRoPE2 \u5728\u6548\u7387\u4e0e\u6027\u80fd\u4e4b\u95f4\u7684\u5353\u8d8a\u5e73\u8861\u3002<\/p>\n\n\n\n

\"chart,
\u56fe4\uff1a\u5728128k\u4e0a\u4e0b\u6587\u957f\u5ea6\u4e0b\uff0cLongRoPE2 \u6269\u5c55\u7684 LLaMA3-8B \u5728\u7ea6100\u4ebf\u53c2\u6570\u6a21\u578b\u4e2d\u8868\u73b0\u6700\u4f73\u3002<\/figcaption><\/figure>\n\n\n\n

LongRoPE2 \u4e3a\u4e0a\u4e0b\u6587\u6269\u5c55\u63d0\u4f9b\u4e86\u4e00\u4e2a\u9ad8\u6548\u3001\u53ef\u6cdb\u5316\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u663e\u8457\u964d\u4f4e\u4e86\u8bad\u7ec3\u6210\u672c\uff0c\u540c\u65f6\u4fdd\u6301\u4e86\u6a21\u578b\u6027\u80fd\uff0c\u4e3a\u672a\u6765\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\u7684\u90e8\u7f72\u548c\u5e94\u7528\u5960\u5b9a\u4e86\u57fa\u7840\u3002<\/p>\n\n\n\n

05. MMInference\uff1a\u901a\u8fc7\u6a21\u6001\u611f\u77e5\u7f6e\u6362\u7a00\u758f\u6ce8\u610f\u529b\uff0c\u52a0\u901f\u957f\u4e0a\u4e0b\u6587\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u7684\u9884\u586b\u5145\u9636\u6bb5<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2504.16083 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\uff08VLMs\uff09\u5728\u591a\u6a21\u6001\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u4e2d\u8868\u73b0\u4f18\u5f02\uff0c\u5c24\u5176\u5728\u89c6\u9891\u95ee\u7b54\u548c\u56fe\u50cf\u5b57\u5e55\u751f\u6210\u7b49\u9886\u57df\u5c55\u73b0\u51fa\u5de8\u5927\u6f5c\u529b\u3002\u4f46\u5728\u5b9e\u9645\u90e8\u7f72\u65f6\uff0c\u9884\u586b\u5145\u9636\u6bb5\u6ce8\u610f\u529b\u8ba1\u7b97\u7684\u4e8c\u6b21\u65b9\u590d\u6742\u5ea6\u6210\u4e3a\u6027\u80fd\u74f6\u9888\uff0c\u4e25\u91cd\u5f71\u54cd\u4e86\u6a21\u578b\u7684\u6548\u7387\u548c\u6269\u5c55\u6027\u3002\u9488\u5bf9\u8fd9\u4e00\u96be\u9898\uff0c\u7814\u7a76\u5458\u4eec\u5728\u8bba\u6587\u4e2d\u63d0\u51fa\u4e86\u4e00\u79cd\u9762\u5411\u591a\u6a21\u6001\u767e\u4e07\u7ea7\u4ee4\u724c\u63a8\u7406\u7684\u52a8\u6001\u7a00\u758f\u6ce8\u610f\u529b\u673a\u5236 MMInference\uff0c\u53ef\u663e\u8457\u63d0\u5347\u9884\u586b\u5145\u6548\u7387\u3002<\/p>\n\n\n\n

\u8be5\u65b9\u6cd5\u9996\u5148\u63ed\u793a\u4e86\u89c6\u9891\u8f93\u5165\u5728\u65f6\u7a7a\u7ef4\u5ea6\u4e0a\u7684\u5c40\u90e8\u6027\u7279\u5f81\uff0c\u53d1\u73b0\u5176\u7a00\u758f\u6ce8\u610f\u529b\u5206\u5e03\u5448\u73b0\u72ec\u7279\u7684”\u7f51\u683c\u6a21\u5f0f”\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4e0d\u540c\u6a21\u6001\u95f4\u7684\u6ce8\u610f\u529b\u5206\u5e03\u5dee\u5f02\u660e\u663e\uff0c\u8fd9\u4f7f\u5f97\u4f20\u7edf\u7a00\u758f\u673a\u5236\u96be\u4ee5\u6709\u6548\u5904\u7406\u591a\u6a21\u6001\u8fb9\u754c\u95ee\u9898\u3002\u5bf9\u6b64\uff0cMMInference \u521b\u65b0\u6027\u5730\u91c7\u7528\u4e86\u57fa\u4e8e\u91cd\u6392\u5217\u7684\u7a00\u758f\u6784\u5efa\u7b56\u7565\uff1a\u901a\u8fc7\u79bb\u7ebf\u641c\u7d22\u4e3a\u6bcf\u4e2a\u6ce8\u610f\u529b\u5934\u786e\u5b9a\u6700\u4f18\u7a00\u758f\u6a21\u5f0f\uff0c\u5e76\u6839\u636e\u8f93\u5165\u52a8\u6001\u68c0\u6d4b\u7a00\u758f\u8499\u7248\u751f\u6210\u5bf9\u5e94\u7684\u91cd\u6392\u5217\u65b9\u6848\u3002\u540c\u65f6\uff0c\u7814\u7a76\u5458\u4eec\u8fd8\u4e13\u95e8\u8bbe\u8ba1\u4e86\u4f18\u5316\u7684 GPU \u5185\u6838\uff08kernel\uff09\uff0c\u4ee5\u63d0\u5347\u7a00\u758f\u8ba1\u7b97\u7684\u786c\u4ef6\u6267\u884c\u6548\u7387\u3002<\/p>\n\n\n\n

\"chart\"
\u56fe5\uff1a\u52a8\u6001\u7a00\u758f\u6ce8\u610f\u529b\u63a8\u7406\u6d41\u7a0b\u901a\u8fc7\u7a00\u758f\u52a0\u8f7d\u4e0e\u5bc6\u96c6\u8ba1\u7b97\u76f8\u7ed3\u5408\uff0c\u5b9e\u73b0\u4e86\u9ad8\u6548\u7684\u786c\u4ef6\u52a0\u901f<\/figcaption><\/figure>\n\n\n\n

MMInference \u7684\u7a81\u51fa\u4f18\u52bf\u5728\u4e8e\u5176\u51fa\u8272\u7684\u517c\u5bb9\u6027\u2014\u2014\u65e0\u9700\u5bf9\u73b0\u6709\u6a21\u578b\u8fdb\u884c\u4efb\u4f55\u4fee\u6539\u6216\u5fae\u8c03\uff0c\u5373\u53ef\u65e0\u7f1d\u96c6\u6210\u5230\u4e3b\u6d41\u957f\u4e0a\u4e0b\u6587 VLMs \u63a8\u7406\u6d41\u7a0b\u4e2d\u3002\u5b9e\u9a8c\u6570\u636e\u8868\u660e\uff0c\u5728 Video QA\u3001Captioning\u3001VisionNIAH \u548c Mixed-Modality NIAH \u7b49\u591a\u4e2a\u591a\u6a21\u6001\u57fa\u51c6\u4efb\u52a1\u4e0a\uff0cMMInference \u5728\u5904\u7406\u767e\u4e07\u7ea7\u4ee4\u724c\u65f6\uff0c\u9884\u586b\u5145\u9636\u6bb5\u52a0\u901f\u6bd4\u6700\u9ad8\u53ef\u8fbe8.3\u500d\uff0c\u4e14\u6a21\u578b\u7cbe\u5ea6\u4fdd\u6301\u4e0d\u53d8\u3002\u6d4b\u8bd5\u8986\u76d6\u4e86 LongVila\u3001LlavaVideo\u3001VideoChat-Flash \u548c Qwen2.5-VL \u7b49\u5f53\u524d\u6700\u5148\u8fdb\u7684\u957f\u4e0a\u4e0b\u6587 VLM \u67b6\u6784\u3002<\/p>\n\n\n\n

06. \u514b\u670d\u8bcd\u8868\u4e0d\u5339\u914d\uff1a\u8bcd\u8868\u65e0\u5173\u7684\u6559\u5e08\u5f15\u5bfc\u8bed\u8a00\u5efa\u6a21\u65b9\u6cd5<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2503.19123 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u8fd1\u5e74\u6765\uff0c\u5229\u7528\u5927\u578b\u6559\u5e08\u6a21\u578b\u6307\u5bfc\u5c0f\u578b\u5b66\u751f\u6a21\u578b\u7684\u8bad\u7ec3\uff0c\u5df2\u6210\u4e3a\u63d0\u5347\u8bed\u8a00\u6a21\u578b\u6548\u7387\u4e0e\u6027\u80fd\u7684\u4e3b\u6d41\u65b9\u6cd5\u3002\u4f46\u662f\u4e0d\u540c\u6559\u5e08\u6a21\u578b\u4e0e\u5b66\u751f\u6a21\u578b\u7684\u8bcd\u8868\u4e4b\u95f4\u5b58\u5728\u4e0d\u4e00\u81f4\u95ee\u9898\uff0c\u4e25\u91cd\u5f71\u54cd\u4e86\u77e5\u8bc6\u8fc1\u79fb\u7684\u6548\u679c\u3002\u8fd9\u79cd\u8bcd\u8868\u4e0d\u5339\u914d\u4f1a\u5bfc\u81f4\u751f\u6210\u7684\u6807\u8bb0\u5e8f\u5217\u548c\u8f93\u51fa\u5206\u5e03\u51fa\u73b0\u504f\u5dee\uff0c\u4ece\u800c\u524a\u5f31\u6559\u5e08\u6a21\u578b\u7684\u6307\u5bfc\u80fd\u529b\u3002<\/p>\n\n\n\n

\u4e3a\u89e3\u51b3\u8fd9\u4e00\u6311\u6218\uff0c\u672c\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u8bcd\u8868\u65e0\u5173\u6559\u5e08\u5f15\u5bfc\u8bed\u8a00\u5efa\u6a21\u65b9\u6cd5 VocAgnoLM\uff08Vocabulary-agnostic Teacher Guided Language Modeling\uff09\u3002\u8be5\u65b9\u6cd5\u901a\u8fc7\u4e24\u4e2a\u5173\u952e\u673a\u5236\uff0c\u6709\u6548\u7f13\u89e3\u4e86\u8bcd\u8868\u4e0d\u4e00\u81f4\u5e26\u6765\u7684\u95ee\u9898\uff1a\u4e00\u662f\u8bcd\u7ea7\u5bf9\u9f50\u673a\u5236\uff0c\u901a\u8fc7\u6784\u5efa\u5b66\u751f\u6807\u8bb0\u4e0e\u6559\u5e08\u6807\u8bb0\u4e4b\u95f4\u7684\u4e00\u5bf9\u591a\u6620\u5c04\uff0c\u5b9e\u73b0\u8de8\u8bcd\u8868\u7684\u8bed\u4e49\u5bf9\u9f50\uff1b\u4e8c\u662f\u6559\u5e08\u5f15\u5bfc\u635f\u5931\u673a\u5236\uff0c\u5229\u7528\u6559\u5e08\u6a21\u578b\u7684\u635f\u5931\u4fe1\u606f\u5f15\u5bfc\u5b66\u751f\u6a21\u578b\u7684\u8bad\u7ec3\uff0c\u4ece\u800c\u63d0\u5347\u5176\u5b66\u4e60\u6548\u679c\u3002<\/p>\n\n\n\n

\"diagram\"
\u56fe6\uff1a\u8bcd\u8868\u65e0\u5173\u6559\u5e08\u5f15\u5bfc\u8bed\u8a00\u5efa\u6a21\u65b9\u6cd5\u6982\u89c8<\/figcaption><\/figure>\n\n\n\n

\u7814\u7a76\u5458\u4eec\u5728\u591a\u4e2a\u6559\u5e08\u6a21\u578b\uff08\u5982 Qwen\u3001Mistral\u3001DeepSeek\uff09\u4e0e\u4e00\u4e2a 1B \u53c2\u6570\u89c4\u6a21\u7684\u5b66\u751f\u6a21\u578b\uff08TinyLlama\uff09\u4e4b\u95f4\u8fdb\u884c\u4e86\u5b9e\u9a8c\u3002\u5373\u4f7f\u5728\u8bcd\u8868\u91cd\u5408\u5ea6\u6781\u4f4e\u7684\u60c5\u51b5\u4e0b\uff08\u5982 Qwen2.5-Math-Instruct \u4e0e TinyLlama \u4ec5\u67096%\u7684\u8bcd\u8868\u91cd\u5408\uff09\uff0cVocAgnoLM \u4ecd\u663e\u8457\u63d0\u5347\u4e86\u5b66\u751f\u6a21\u578b\u7684\u6027\u80fd\uff0c\u76f8\u8f83\u4e8e\u4f20\u7edf\u7684\u6301\u7eed\u9884\u8bad\u7ec3\u65b9\u6cd5\uff0c\u6027\u80fd\u63d0\u5347\u9ad8\u8fbe46%\u3002\u6b64\u5916\uff0c\u5b9e\u9a8c\u8fd8\u8868\u660e\uff0cVocAgnoLM \u5728\u9762\u5bf9\u66f4\u5f3a\u7684\u6559\u5e08\u6a21\u578b\u65f6\u80fd\u6301\u7eed\u53d7\u76ca\uff0c\u5c55\u73b0\u51fa\u826f\u597d\u7684\u6cdb\u5316\u80fd\u529b\u548c\u7a33\u5b9a\u6027\u3002<\/p>\n\n\n\n

07. \u63d0\u5347\u6807\u51c6\uff0c\u901a\u8fc7\u751f\u6210\u5f0f\u81ea\u8fdb\u5316\u8bc4\u4f30\uff0c\u63a2\u7a76\u5927\u8bed\u8a00\u6a21\u578b\u7684\u4ef7\u503c\u89c2<\/h3>\n\n\n\n
\"paper<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2406.14230 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u8fd1\u5e74\u6765\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\u9886\u57df\u53d6\u5f97\u4e86\u663e\u8457\u8fdb\u5c55\uff0c\u4f46\u5176\u4ecd\u4f1a\u751f\u6210\u6709\u5bb3\u6216\u4e0d\u7b26\u5408\u4f26\u7406\u7684\u5185\u5bb9\uff0c\u5f15\u53d1\u5e7f\u6cdb\u5173\u6ce8\u3002\u5982\u4f55\u8bc4\u4f30\u548c\u786e\u4fdd\u6a21\u578b\u7684\u4ef7\u503c\u5bf9\u9f50\uff0c\u6210\u4e3a\u63a8\u52a8\u5176\u8d1f\u8d23\u4efb\u90e8\u7f72\u7684\u5173\u952e\u95ee\u9898\u3002\u5c3d\u7ba1\u5df2\u6709\u5927\u91cf\u57fa\u51c6\u6570\u636e\u7528\u4e8e\u68c0\u6d4b\u6a21\u578b\u7684\u793e\u4f1a\u504f\u89c1\u3001\u6bd2\u6027\u548c\u4f26\u7406\u95ee\u9898\uff0c\u4f46\u8fd9\u4e9b\u9759\u6001\u57fa\u51c6\u9762\u4e34\u201c\u8bc4\u4f30\u65f6\u6548\u6027\u201d\u7684\u6311\u6218\u2014\u2014\u968f\u7740\u6a21\u578b\u4e0d\u65ad\u8fed\u4ee3\u5347\u7ea7\uff0c\u539f\u6709\u6d4b\u8bd5\u6570\u636e\u53ef\u80fd\u5df2\u88ab\u6a21\u578b\u76f4\u63a5\u5b66\u4e60\uff0c\u5bfc\u81f4\u8bc4\u4f30\u7ed3\u679c\u5931\u771f\u6216\u8fc7\u5ea6\u4e50\u89c2\u3002<\/p>\n\n\n\n

\u4e3a\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u751f\u6210\u5f0f\u81ea\u8fdb\u5316\u8bc4\u4f30\u65b9\u6cd5 GETA\uff08Generative Evolving Testing Approach\uff09\u3002\u8be5\u65b9\u6cd5\u501f\u9274\u6d4b\u91cf\u7406\u8bba\u4e2d\u7684\u81ea\u9002\u5e94\u6d4b\u8bd5\u7406\u5ff5\uff0c\u7a81\u7834\u4e86\u4f20\u7edf\u65b9\u6cd5\u4f9d\u8d56\u9759\u6001\u9898\u5e93\u7684\u9650\u5236\u3002GETA \u80fd\u591f\u6839\u636e\u6a21\u578b\u5f53\u524d\u80fd\u529b\u52a8\u6001\u751f\u6210\u6d4b\u8bd5\u9879\uff0c\u4ece\u800c\u66f4\u6709\u6548\u5730\u63a2\u6d4b\u5176\u9053\u5fb7\u8fb9\u754c\u548c\u4ef7\u503c\u503e\u5411\u3002\u5177\u4f53\u800c\u8a00\uff0cGETA \u901a\u8fc7\u5b66\u4e60\u6d4b\u8bd5\u9879\u96be\u5ea6\u4e0e\u6a21\u578b\u4ef7\u503c\u9075\u5b88\u5ea6\u4e4b\u95f4\u7684\u8054\u5408\u5206\u5e03\uff0c\u53ef\u5b9e\u73b0\u6d4b\u8bd5\u5185\u5bb9\u4e0e\u6a21\u578b\u80fd\u529b\u7684\u534f\u540c\u6f14\u5316\u3002\u8fd9\u79cd\u673a\u5236\u65e2\u63d0\u5347\u4e86\u6d4b\u8bd5\u7684\u9488\u5bf9\u6027\u548c\u7075\u6d3b\u6027\uff0c\u4e5f\u663e\u8457\u7f13\u89e3\u4e86\u8bc4\u4f30\u65f6\u6548\u6027\u5e26\u6765\u7684\u8bef\u5dee\u3002<\/p>\n\n\n\n

\"graphical
\u56fe7\uff1a\u9759\u6001\u8bc4\u4f30\u3001CAT \u548c GETA \u65b9\u6cd5\u793a\u610f\u56fe<\/figcaption><\/figure>\n\n\n\n

\u5728\u5b9e\u9a8c\u4e2d\uff0c\u7814\u7a76\u5458\u4eec\u4f7f\u7528 GETA \u5bf9\u591a\u4e2a\u4e3b\u6d41 LLMs \u8fdb\u884c\u4e86\u8bc4\u4f30\u3002\u7ed3\u679c\u8868\u660e\uff0cGETA \u80fd\u591f\u81ea\u52a8\u751f\u6210\u4e0e\u6a21\u578b\u80fd\u529b\u5339\u914d\u7684\u6d4b\u8bd5\u9898\u76ee\uff0c\u5176\u8bc4\u4f30\u7ed3\u679c\u5728\u4e0e\u5f00\u53d1\u8fc7\u7a0b\u4e2d\u672a\u4f7f\u7528\u7684\u5206\u5e03\u5916\uff08OOD\uff09\u548c\u72ec\u7acb\u540c\u5206\u5e03\uff08i.i.d.\uff09\u6570\u636e\u65f6\u8868\u73b0\u51fa\u66f4\u9ad8\u7684\u4e00\u81f4\u6027\u3002\u8fd9\u8868\u660e GETA \u4e0d\u4ec5\u80fd\u66f4\u51c6\u786e\u5730\u53cd\u6620\u6a21\u578b\u7684\u771f\u5b9e\u4ef7\u503c\u5bf9\u9f50\u7a0b\u5ea6\uff0c\u4e5f\u4e3a\u672a\u6765\u6784\u5efa\u52a8\u6001\u3001\u53ef\u6269\u5c55\u7684\u8bc4\u4f30\u8303\u5f0f\u5960\u5b9a\u4e86\u57fa\u7840\uff0c\u6709\u671b\u5728\u6a21\u578b\u5b89\u5168\u6027\u3001\u4f26\u7406\u6027\u548c\u793e\u4f1a\u8d23\u4efb\u65b9\u9762\u53d1\u6325\u91cd\u8981\u4f5c\u7528\u3002<\/p>\n\n\n\n

08. \u8106\u5f31\u6027\u611f\u77e5\u5bf9\u9f50\uff1a\u7f13\u89e3\u6709\u5bb3\u5fae\u8c03\u4e2d\u7684\u975e\u5747\u5300\u9057\u5fd8\u95ee\u9898<\/h3>\n\n\n\n
\"text\"<\/figure>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2506.03850 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u5f53\u524d\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u7684\u5e7f\u6cdb\u5f00\u6e90\u4e0e\u5b9a\u5236\u5316\u5e94\u7528\uff0c\u6b63\u5728\u5f15\u53d1\u4e00\u79cd\u65b0\u7684\u5b89\u5168\u5a01\u80c1\u2014\u2014\u6709\u5bb3\u5fae\u8c03\uff08harmful fine-tuning, HFT\uff09\u3002\u4e00\u4e9b\u7528\u6237\u901a\u8fc7\u5bf9\u5f00\u6e90\u7684\u5927\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u4ee5\u6ee1\u8db3\u7279\u5b9a\u7684\u5b9a\u5236\u5316\u9700\u6c42\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u5fae\u8c03\u8fc7\u7a0b\u5f80\u5f80\u5f88\u5bb9\u6613\u7834\u574f\u6a21\u578b\u539f\u6709\u7684\u5b89\u5168\u9632\u62a4\u673a\u5236\uff0c\u5bfc\u81f4\u6a21\u578b\u88ab\u8bf1\u5bfc\u751f\u6210\u6709\u5bb3\u5185\u5bb9\u3002\u66f4\u4ee4\u4eba\u62c5\u5fe7\u7684\u662f\uff0c\u8fd9\u79cd\u65b9\u5f0f\u95e8\u69db\u4f4e\u3001\u9690\u853d\u6027\u5f3a\uff0c\u800c\u4e14\u5728\u5404\u7c7b\u573a\u666f\u4e2d\u5e7f\u6cdb\u5b58\u5728\uff0c\u7ed9\u5927\u6a21\u578b\u7684\u5b89\u5168\u4f7f\u7528\u5e26\u6765\u4e86\u4e25\u5cfb\u6311\u6218\u3002<\/p>\n\n\n\n

\u672c\u8bba\u6587\u9996\u6b21\u53d1\u73b0\uff0c\u4e00\u90e8\u5206\u5b89\u5168\u5bf9\u9f50\u6570\u636e\u5177\u6709\u66f4\u9ad8\u7684\u201c\u654f\u611f\u6027\u201d\uff0c\u5728\u7528\u6237\u8fdb\u884c\uff08\u6709\u5bb3\uff09\u5fae\u8c03\u65f6\u66f4\u5bb9\u6613\u88ab\u9057\u5fd8\u3002\u8fd9\u4e00\u73b0\u8c61\u63ed\u793a\uff0c\u73b0\u6709\u201c\u4e00\u89c6\u540c\u4ec1\u201d\u7684\u5bf9\u9f50\u65b9\u6cd5\u5b58\u5728\u5c40\u9650\u6027\uff0c\u6211\u4eec\u8feb\u5207\u9700\u8981\u57fa\u4e8e\u6570\u636e\u654f\u611f\u6027\u5dee\u5f02\uff0c\u8bbe\u8ba1\u66f4\u5177\u9488\u5bf9\u6027\u7684\u5b89\u5168\u5bf9\u9f50\u65b9\u6cd5\u3002<\/p>\n\n\n\n

\u4e3a\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u8106\u5f31\u611f\u77e5\u5bf9\u9f50\u65b9\u6cd5\uff08Vulnerability-Aware Alignment, VAA\uff09\u3002\u8be5\u65b9\u6cd5\u9996\u5148\u901a\u8fc7\u201c\u9057\u5fd8\u6b21\u6570\u201d\u8fd9\u4e00\u8bc4\u4f30\u6307\u6807\uff0c\u5c06\u5b89\u5168\u5bf9\u9f50\u6570\u636e\u5212\u5206\u4e3a\u201c\u8106\u5f31\u7ec4\u201d\u548c\u201c\u975e\u8106\u5f31\u7ec4\u201d\uff1b\u968f\u540e\u5f15\u5165\u5206\u5e03\u9c81\u68d2\u4f18\u5316\uff08Distributionally Robust Optimization, DRO\uff09\u6846\u67b6\uff0c\u5728\u5bf9\u9f50\u8fc7\u7a0b\u4e2d\u52a8\u6001\u8c03\u6574\u4e0d\u540c\u7ec4\u522b\u6837\u672c\u7684\u5b66\u4e60\u6743\u91cd\u3002\u901a\u8fc7\u6709\u9488\u5bf9\u6027\u5730\u5f3a\u5316\u201c\u8106\u5f31\u7ec4\u201d\u7684\u5b66\u4e60\uff0cVAA \u65b9\u6cd5\u663e\u8457\u63d0\u5347\u4e86\u5b89\u5168\u5bf9\u9f50\u5728\u9762\u5bf9\u6709\u5bb3\u5fae\u8c03\u65f6\u7684\u9c81\u68d2\u6027\uff0c\u4f7f\u6a21\u578b\u5728\u5fae\u8c03\u540e\u7684\u5b89\u5168\u80fd\u529b\u201c\u5b58\u7559\u7387\u201d\u66f4\u9ad8\u3002<\/p>\n\n\n\n

\"chart,
\u56fe8\uff1a\u8106\u5f31\u6027\u611f\u77e5\u5bf9\u9f50 VAA<\/figcaption><\/figure>\n\n\n\n

\u5728\u56db\u4e2a\u5178\u578b\u7684\u5fae\u8c03\u573a\u666f\u4e2d\uff0cVAA \u5728\u6709\u6548\u964d\u4f4e\u6709\u5bb3\u8f93\u51fa\u8bc4\u5206\uff08\u4e0b\u964d\u5e45\u5ea6\u8fbe42%\uff09\u7684\u540c\u65f6\uff0c\u4fdd\u6301\u4e86\u4e0b\u6e38\u4efb\u52a1\u7684\u6027\u80fd\uff0c\u663e\u8457\u4f18\u4e8e\u73b0\u6709\u7684\u57fa\u7ebf\u65b9\u6cd5\uff08\u5982 Robust Fine-Tuning \u548c Unlearnable Examples\uff09\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cVAA \u4e0d\u4ec5\u80fd\u591f\u6709\u6548\u62b5\u5fa1 HFT \u653b\u51fb\uff0c\u8fd8\u80fd\u5728\u590d\u6742\u7684\u5fae\u8c03\u573a\u666f\u4e2d\u7ef4\u6301\u6a21\u578b\u7684\u5b9e\u7528\u6027\u3002<\/p>\n\n\n\n

\u672c\u7814\u7a76\u9996\u6b21\u5c06\u6570\u636e\u9057\u5fd8\u8106\u5f31\u6027\u7eb3\u5165\u5b89\u5168\u5bf9\u9f50\u7684\u6838\u5fc3\u8003\u91cf\uff0c\u4e3a\u6784\u5efa\u66f4\u5b89\u5168\u3001\u53ef\u4fe1\u7684 LLMs \u63d0\u4f9b\u4e86\u65b0\u8303\u5f0f\u3002\u672a\u6765\u7814\u7a76\u53ef\u8fdb\u4e00\u6b65\u63a2\u7d22\u8de8\u6a21\u6001\u6570\u636e\u7684\u8106\u5f31\u6027\u8fc1\u79fb\u89c4\u5f8b\uff0c\u4ee5\u53ca\u52a8\u6001\u8106\u5f31\u6027\u8bc4\u4f30\u7684\u5b9e\u65f6\u66f4\u65b0\u673a\u5236\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 \u5728\u6e29\u54e5\u534e\u4e3e\u529e\u3002\u6765\u81ea\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u591a\u7bc7\u8bba\u6587\u5165\u9009\u3002\u4e0a\u4e00\u671f\u201cICML \u4e0a\u65b0\u201d\u7cbe\u9009\u4ecb\u7ecd\u4e86\u4e0e\u51b3\u7b56\u6a21\u578b\u76f8\u5173\u7684\u7814\u7a76\u5de5\u4f5c\uff0c\u5185\u5bb9\u6db5\u76d6\u5f3a\u5316\u5b66\u4e60\u3001RLHF\u3001\u6269\u6563\u5efa\u6a21\u7b49\u65b9\u5411\u3002\u672c\u671f\u5c06\u805a\u7126\u6a21\u578b\u63a8\u7406\u673a\u5236\u3001\u591a\u6a21\u6001\u8ba4\u77e5\u3001\u957f\u4e0a\u4e0b\u6587\u5904\u7406\u4e0e\u5b89\u5168\u5bf9\u9f50\u7b49\u5173\u952e\u6280\u672f\u3002 \u672c\u671f\u5185\u5bb9\u901f\u89c8 01. AdaptiveStep\uff1a\u57fa\u4e8e\u6a21\u578b\u7f6e\u4fe1\u5ea6\u7684\u81ea\u52a8\u63a8\u7406\u6b65\u9aa4\u5212\u5206 02. \u4ece\u590d\u6742\u6027\u5230\u539f\u5b50\u6027\uff0c\u901a\u8fc7\u77e5\u8bc6\u611f\u77e5\u7684\u53cc\u91cd\u91cd\u5199\u4e0e\u63a8\u7406\u673a\u5236\u63d0\u5347LLMs\u751f\u6210\u80fd\u529b 03. \u5728\u7a7a\u95f4\u63a8\u7406\u4e2d\u60f3\u8c61\uff1a\u591a\u6a21\u6001\u601d\u7ef4\u53ef\u89c6\u5316 04. LongRoPE2\uff1a\u8fd1\u4e4e\u65e0\u635f\u7684\u5927\u8bed\u8a00\u6a21\u578b\u4e0a\u4e0b\u6587\u7a97\u53e3\u6269\u5c55\u65b9\u6cd5 05. MMInference\uff1a\u901a\u8fc7\u6a21\u6001\u611f\u77e5\u7f6e\u6362\u7a00\u758f\u6ce8\u610f\u529b\uff0c\u52a0\u901f\u957f\u4e0a\u4e0b\u6587\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u7684\u9884\u586b\u5145\u9636\u6bb5 06. \u514b\u670d\u8bcd\u8868\u4e0d\u5339\u914d\uff1a\u8bcd\u8868\u65e0\u5173\u7684\u6559\u5e08\u5f15\u5bfc\u8bed\u8a00\u5efa\u6a21\u65b9\u6cd5 07. \u63d0\u5347\u6807\u51c6\uff0c\u901a\u8fc7\u751f\u6210\u5f0f\u81ea\u8fdb\u5316\u8bc4\u4f30\uff0c\u63a2\u7a76\u5927\u8bed\u8a00\u6a21\u578b\u7684\u4ef7\u503c\u89c2 08. \u8106\u5f31\u6027\u611f\u77e5\u5bf9\u9f50\uff1a\u7f13\u89e3\u6709\u5bb3\u5fae\u8c03\u4e2d\u7684\u975e\u5747\u5300\u9057\u5fd8\u95ee\u9898 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2502.13943 (opens in new tab) \u968f\u7740\u6a21\u578b\u80fd\u529b\u7684\u63d0\u5347\u548c\u6280\u672f\u7684\u53d1\u5c55\uff0c\u8bed\u8a00\u6a21\u578b\u53ef\u4ee5\u5728\u8bf8\u591a\u4efb\u52a1\u4e0a\u90fd\u53d6\u5f97\u8f83\u9ad8\u5206\u6570\u3002GPT-o1 \u548c DeepSeek-R1 \u7684\u53d1\u5e03\uff0c\u51ed\u501f\u5f3a\u63a8\u7406\u80fd\u529b\u8ba9\u8bed\u8a00\u6a21\u578b\u201c\u51fa\u5708\u201d\u3002\u63d0\u5347\u6a21\u578b\u63a8\u7406\u80fd\u529b\u7684\u4e00\u4e2a\u57fa\u7840\u6280\u672f\u662f\u601d\u7ef4\u94fe\u65b9\u6cd5\uff08Chain-of-Thought, CoT\uff09\uff0c\u8be5\u65b9\u6cd5\u80fd\u591f\u5c06\u6a21\u578b\u89e3\u51b3\u95ee\u9898\u7684\u8fc7\u7a0b\u5206\u6210\u591a\u4e2a\u6b65\u9aa4\uff0c\u4ece\u800c\u89e3\u51b3\u590d\u6742\u7684\u63a8\u7406\u4efb\u52a1\u3002 \u7136\u800c\uff0c\u5f53\u524d\u5c06\u6a21\u578b\u63a8\u7406\u6b65\u9aa4\u8fdb\u884c\u5212\u5206\u591a\u91c7\u7528\u4e24\u79cd\u65b9\u6cd5\uff1a\u4e00\u79cd\u662f\u4eba\u5de5\u8fdb\u884c\u6b65\u9aa4\u5212\u5206\uff0c\u624b\u5de5\u5212\u5206\u7684\u6b65\u9aa4\u4fe1\u606f\u542b\u91cf\u6bd4\u8f83\u9ad8\u4e14\u5747\u8861\uff0c\u4f46\u96be\u4ee5\u81ea\u52a8\u5316\uff0c\u5212\u5206\u4f9d\u8d56\u4e8e\u4e13\u5bb6\u7684\u9886\u57df\u77e5\u8bc6\uff1b\u53e6\u4e00\u79cd\u662f\u57fa\u4e8e\u89c4\u5219\u7684\u5212\u5206\uff0c\u5982\u56fa\u5b9a\u884c\u6570\u6216 token \u6570\uff0c\u8fd9\u4e00\u65b9\u6cd5\u53ef\u4ee5\u81ea\u52a8\u5316\uff0c\u4f46\u5bb9\u6613\u505a\u51fa\u6ca1\u6709\u610f\u4e49\u6216\u4e0d\u5408\u7406\u7684\u5212\u5206\u3002 \u4e3a\u89e3\u51b3\u8fd9\u4e00\u96be\u9898\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u521b\u65b0\u65b9\u6cd5 AdaptiveStep\u3002\u5176\u6838\u5fc3\u5728\u4e8e\u501f\u52a9\u8bed\u8a00\u6a21\u578b\u9884\u6d4b\u4e0b\u4e00 token \u65f6\u7684\u7f6e\u4fe1\u5ea6\u53d8\u5316\uff0c\u81ea\u52a8\u8bc6\u522b\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u81ea\u7136\u51b3\u7b56\u70b9\uff0c\u8fdb\u800c\u5b9e\u73b0\u66f4\u5408\u7406\u7684\u6b65\u9aa4\u5212\u5206\u3002 AdaptiveStep \u65e0\u9700\u4eba\u5de5\u6807\u6ce8\uff0c\u5b8c\u5168\u4f9d\u636e\u6a21\u578b\u81ea\u8eab\u7684\u9884\u6d4b\u884c\u4e3a\u6765\u8fdb\u884c\u63a8\u7406\u6b65\u9aa4\u7684\u81ea\u9002\u5e94\u5212\u5206\u3002\u5177\u4f53\u64cd\u4f5c\u4e3a\uff1a\u5728\u6bcf\u4e2a token \u4f4d\u7f6e\u8ba1\u7b97\u6a21\u578b\u7684\u7f6e\u4fe1\u5ea6\u53d8\u5316\uff0c\u4e00\u65e6\u7f6e\u4fe1\u5ea6\u663e\u8457\u4e0b\u964d\uff0c\u4fbf\u5c06\u5176\u89c6\u4e3a\u6f5c\u5728\u7684\u63a8\u7406\u5206\u754c\u70b9\u3002\u8fd9\u79cd\u5212\u5206\u65b9\u5f0f\u80fd\u591f\u7cbe\u51c6\u6355\u6349\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u5173\u952e\u8f6c\u6298\uff0c\u663e\u8457\u63d0\u5347\u6bcf\u4e00\u6b65\u7684\u51b3\u7b56\u4fe1\u606f\u5bc6\u5ea6\u3002 \u7814\u7a76\u5458\u4eec\u5c06 AdaptiveStep \u5e94\u7528\u4e8e Process Reward Model (PRM) \u7684\u8bad\u7ec3\u4e2d\uff0c\u5e76\u5728\u6570\u5b66\u63a8\u7406\u3001\u4ee3\u7801\u751f\u6210\u7b49\u4efb\u52a1\u4e2d\u8fdb\u884c\u4e86\u7cfb\u7edf\u8bc4\u4f30\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u7ecf AdaptiveStep \u8bad\u7ec3\u7684 PRM \u5728\u591a\u9879\u4efb\u52a1\u4e2d\u5747\u53d6\u5f97\u4e86\u9886\u5148\u6027\u80fd\u3002\u5728 […]<\/p>\n","protected":false},"author":34512,"featured_media":1145016,"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-1144993","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\/1144993","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":7,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1144993\/revisions"}],"predecessor-version":[{"id":1145017,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1144993\/revisions\/1145017"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1145016"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1144993"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1144993"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1144993"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1144993"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}