{"id":795,"date":"2025-12-27T01:47:09","date_gmt":"2025-12-26T17:47:09","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=795"},"modified":"2025-12-27T01:47:09","modified_gmt":"2025-12-26T17:47:09","slug":"mme-cot-benchmarking-chain-of-thought-in-large-multimodal-models-for-reasoning-quality-robustness-and-efficiency","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=795","title":{"rendered":"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency"},"content":{"rendered":"\n<p><strong>\u4f1a\u8bae<\/strong>\uff1aICML 2025\uff08\u7b2c42\u5c4a\u56fd\u9645\u673a\u5668\u5b66\u4e60\u5927\u4f1a\uff09<\/p>\n\n\n\n<p>Dongzhi Jiang\u2217 1 , Renrui Zhang\u2217\u2020 1 , Ziyu Guo2 , Yanwei Li\u20213 , Yu Qi\u20214 , Xinyan Chen\u20211<br>Liuhui Wang\u20215 , Jianhan Jin\u20216 , Claire Guo\u20217 , Shen Yan3 , Bo Zhang8<br>Chaoyou Fu6 , Peng Gao8 , Hongsheng Li1<br>arXiv:2502.09621v1 [cs.CV] 13 Feb 2025<br>1 CUHKMMLab 2CUHKMiuLarLab 3 ByteDance 4 NEU 5 UPenn<br>6 NJU 7CUHK(Shenzhen) 8 Shanghai AI Laboratory<br>{dzjiang,renruizhang}@link.cuhk.edu.hk<br>\u2217 Core contribution \u2020 Project lead \u2021 Equal contribution<br>Project Page: https:\/\/mmecot.github.io\/<\/p>\n\n\n\n<p>\u8fd9\u662f\u76ee\u524d\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\uff08MLLM\uff09\u9886\u57df\u9996\u4e2a\u7cfb\u7edf\u8bc4\u4f30**\u601d\u7ef4\u94fe\uff08Chain-of-Thought\uff0cCoT\uff09**\u5728\u591a\u6a21\u6001\u63a8\u7406\u4efb\u52a1\u4e0a\u7684\u8d28\u91cf\u4e0e\u5c40\u9650\u6027\u7684\u7814\u7a76\u3002\u8be5\u8bba\u6587\u63d0\u51fa\u65b0\u7684\u8bc4\u6d4b\u6846\u67b6\uff0c\u5e76\u4ece\u591a\u4e2a\u7ef4\u5ea6\u5206\u6790\u6700\u5148\u8fdb\u6a21\u578b\u7684\u8868\u73b0\uff0c\u4e3a\u540e\u7eed\u8de8\u6a21\u6001\u63a8\u7406\u7814\u7a76\u63d0\u4f9b\u4e86\u91cd\u8981\u53c2\u8003\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e00\u3001\u7814\u7a76\u52a8\u673a\u4e0e\u80cc\u666f<\/h1>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"535\" height=\"735\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-45.png\"  class=\"wp-image-796\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-45.png 535w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-45-218x300.png 218w\" sizes=\"auto, (max-width: 535px) 100vw, 535px\" title=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe\" alt=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe\" \/><\/figure>\n\n\n\n<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u4e2d**\u601d\u7ef4\u94fe\u63d0\u793a\uff08CoT prompting\uff09<strong>\u901a\u8fc7\u9010\u6b65\u751f\u6210\u63a8\u7406\u6b65\u9aa4\u663e\u8457\u589e\u5f3a\u8bed\u8a00\u63a8\u7406\u80fd\u529b\u3002\u7136\u800c\uff0c\u5728<\/strong>\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\uff08LMMs\uff09**\u4e2d\uff0c\u8fd9\u4e00\u673a\u5236\u7684\u4f5c\u7528\u5c1a\u672a\u88ab\u7cfb\u7edf\u7814\u7a76\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4f20\u7edf CoT \u5728\u6587\u672c\u63a8\u7406\u4e2d\u80fd\u63d0\u9ad8\u903b\u8f91\u6e05\u6670\u5ea6\u548c\u63a8\u7406\u80fd\u529b\uff0c<\/li>\n\n\n\n<li>\u4f46\u5bf9\u56fe\u50cf\u3001\u7a7a\u95f4\u4fe1\u606f\u6216\u89c6\u89c9\u903b\u8f91\u63a8\u7406\u662f\u5426\u540c\u6837\u6709\u6548\u5e76\u4e0d\u6e05\u695a\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u56e0\u6b64\uff0c\u8be5\u8bba\u6587\u63d0\u51fa\u4e13\u95e8\u7528\u4e8e\u591a\u6a21\u6001\u63a8\u7406\u7684\u57fa\u51c6\u4e0e\u8bc4\u4f30\u4f53\u7cfb\uff0c\u4ee5<strong>\u5b9e\u8bc1\u65b9\u5f0f\u68c0\u6d4b CoT \u5728\u89c6\u89c9-\u6587\u672c\u63a8\u7406\u4e2d\u7684\u6548\u529b\u4e0e\u5f0a\u7aef\u3002<\/strong><\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e8c\u3001\u6838\u5fc3\u8d21\u732e<\/h1>\n\n\n\n<h2 class=\"wp-block-heading\">1) <strong>MME-CoT \u57fa\u51c6\u4f53\u7cfb<\/strong><\/h2>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"1013\" height=\"954\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-46.png\"  class=\"wp-image-797\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-46.png 1013w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-46-300x283.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-46-768x723.png 768w\" sizes=\"auto, (max-width: 1013px) 100vw, 1013px\" title=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe1\" alt=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe1\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u4e2a\u4e13\u95e8\u7684 benchmark \u2014\u2014 <strong>MME-CoT<\/strong>\uff0c\u91cd\u70b9\u8bc4\u4f30 LMMs \u5728\u591a\u6a21\u6001\u63a8\u7406\u4efb\u52a1\u4e2d\u7684\u8d28\u91cf\u3001\u7a33\u5065\u6027\u4e0e\u6548\u7387\u3002\u8fd9\u4e2a\u4f53\u7cfb\u5305\u542b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u516d\u7c7b\u4efb\u52a1\u57df<\/strong>\uff1a\u5305\u62ec\u6570\u5b66\u3001\u79d1\u5b66\u3001OCR\uff08\u5149\u5b66\u5b57\u7b26\u8bc6\u522b\uff09\u3001\u903b\u8f91\u3001\u65f6\u7a7a\u5206\u6790\u4e0e\u5e38\u89c4\u573a\u666f\u7406\u89e3\uff0c\u8986\u76d6\u8bed\u8a00\u3001\u89c6\u89c9\u4e0e\u903b\u8f91\u4ea4\u4e92\u7684\u590d\u6742\u63a8\u7406\u573a\u666f\uff1b<\/li>\n\n\n\n<li><strong>\u4e09\u4e2a\u5168\u65b0\u8bc4\u4f30\u6307\u6807<\/strong>\uff1a\n<ul class=\"wp-block-list\">\n<li><strong>\u63a8\u7406\u8d28\u91cf\uff08Quality\uff09<\/strong>\uff1a\u8861\u91cf\u6a21\u578b\u8f93\u51fa\u63a8\u7406\u7684\u903b\u8f91\u6027\u4e0e\u51c6\u786e\u6027\uff1b<\/li>\n\n\n\n<li><strong>\u7a33\u5065\u6027\uff08Robustness\uff09<\/strong>\uff1a\u6d4b\u8bd5\u6a21\u578b\u5bf9\u8f93\u5165\u6270\u52a8\uff08\u5982\u566a\u97f3\u56fe\u50cf\u3001\u6587\u672c\u53d8\u5316\uff09\u7684\u654f\u611f\u5ea6\uff1b<\/li>\n\n\n\n<li><strong>\u63a8\u7406\u6548\u7387\uff08Efficiency\uff09<\/strong>\uff1a\u8bc4\u4f30\u63a8\u7406\u6b65\u9aa4\u6570\u3001\u751f\u6210\u65f6\u95f4\u4e0e\u8d44\u6e90\u6d88\u8017\u3002<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e00\u5168\u9762\u6307\u6807\u4f53\u7cfb\u53ef\u8f83\u7ec6\u7c92\u5ea6\u5730\u8861\u91cf\u591a\u6a21\u6001\u63a8\u7406\u80fd\u529b\uff0c\u800c\u4e0d\u4ec5\u4ec5\u4f9d\u8d56\u6700\u7ec8\u7b54\u6848\u51c6\u786e\u7387\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"481\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-47-1024x481.png\"  class=\"wp-image-798\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-47-1024x481.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-47-300x141.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-47-768x361.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-47.png 1030w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe2\" alt=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe2\" \/><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h2 class=\"wp-block-heading\">2) <strong>\u5bf9\u5148\u8fdb LMMs \u7684\u7cfb\u7edf\u8bc4\u4f30\u4e0e\u6d1e\u5bdf<\/strong><\/h2>\n\n\n\n<p>\u8bba\u6587\u901a\u8fc7 MME-CoT \u5bf9\u5f53\u65f6\u6700\u5f3a\u7684\u591a\u6a21\u6001\u6a21\u578b\uff08\u5305\u62ec\u5e26\u6709 reflection \u673a\u5236\u7684\u6a21\u578b\u3001GPT-4o \u7b49\uff09\u8fdb\u884c\u4e86\u5927\u89c4\u6a21\u6d4b\u8bd5\uff0c\u5f97\u5230\u4e09\u5927\u5173\u952e\u7ed3\u8bba\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\uff081\uff09<strong>\u5e26\u53cd\u601d\uff08reflection\uff09\u673a\u5236\u7684\u6a21\u578b\u62e5\u6709\u66f4\u9ad8\u7684 CoT \u8d28\u91cf<\/strong><\/h3>\n\n\n\n<p>\u4e00\u4e9b\u6700\u65b0\u67b6\u6784\u5f15\u5165\u4e86\u53cd\u601d\u673a\u5236\uff08reflection\uff0c\u4f7f\u6a21\u578b\u5728\u751f\u6210\u7b54\u6848\u65f6\u540c\u65f6\u201c\u5ba1\u89c6\u201d\u751f\u6210\u8fc7\u7a0b\uff09\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u8fd9\u7c7b\u6a21\u578b\u5728\u63a8\u7406\u8d28\u91cf\u4e0a\u663e\u8457\u4f18\u4e8e\u5176\u4ed6\u6a21\u578b\uff0c\u5728\u67d0\u4e9b\u4efb\u52a1\u4e0a\u751a\u81f3\u8d85\u8fc7 GPT-4o\u3002\u8fd9\u663e\u793a\u589e\u5f3a\u5185\u90e8\u63a8\u7406\u8fc7\u7a0b\u6bd4\u5355\u7eaf\u589e\u52a0\u53c2\u6570\u6709\u6548\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">\uff082\uff09<strong>CoT \u63d0\u793a\u5728\u611f\u77e5\u5bc6\u96c6\u4efb\u52a1\u4e0a\u6548\u679c\u4e0d\u4f73<\/strong><\/h3>\n\n\n\n<p>\u5728\u6d89\u53ca\u89c6\u89c9\u7ec6\u8282\u5904\u7406\u7684\u63a8\u7406\u4efb\u52a1\u4e2d\uff08\u5982\u9700\u8981\u51c6\u786e\u9605\u8bfb\u56fe\u50cf\u5185\u5bb9\u6216\u7a7a\u95f4\u5173\u7cfb\u8fa8\u8ba4\uff09\uff0c\u4f7f\u7528 CoT \u63d0\u793a\u53cd\u800c<strong>\u964d\u4f4e\u4e86\u8868\u73b0<\/strong>\u3002\u8bba\u6587\u6307\u51fa\uff0c\u8fd9\u53ef\u80fd\u662f\u56e0\u4e3a\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>CoT \u6269\u5927\u4e86\u201c\u8fc7\u5ea6\u601d\u8003\uff08overthinking\uff09\u201d\u884c\u4e3a\uff0c\u5bf9\u611f\u77e5\u7ec6\u8282\u654f\u611f\u7684\u4efb\u52a1\u4e0d\u9002\u5408\u5197\u957f\u63a8\u7406\uff1b<\/li>\n\n\n\n<li>\u591a\u6a21\u6001\u8f93\u5165\u5bfc\u81f4\u7684\u566a\u97f3\u5e72\u6270 CoT \u63a8\u7406\u94fe\u7684\u7a33\u5b9a\u6027\u3002<\/li>\n<\/ul>\n\n\n\n<h3 class=\"wp-block-heading\">\uff083\uff09<strong>\u53cd\u601d\u673a\u5236\u63d0\u5347\u8d28\u91cf\u4f46\u964d\u4f4e\u6548\u7387<\/strong><\/h3>\n\n\n\n<p>\u53cd\u601d\u673a\u5236\u53ef\u63d0\u5347\u591a\u6a21\u6001 CoT \u7684\u8f93\u51fa\u8d28\u91cf\uff0c\u4f46\u5728\u63a8\u7406\u6548\u7387\u4e0a\u5b58\u5728\u660e\u663e\u6298\u8877\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u751f\u6210\u94fe\u8d8a\u957f\u3001\u6a21\u578b\u8d44\u6e90\u6d88\u8017\u8d8a\u5927\uff1b<\/li>\n\n\n\n<li>\u50cf\u53cd\u601d\u9636\u6bb5\u7684\u81ea\u6211\u68c0\u67e5\u6b65\u6570\u589e\u52a0\u5ef6\u8fdf\u4e0e\u6210\u672c\uff0c\u4f7f\u6a21\u578b\u5728\u5b9e\u9645\u90e8\u7f72\u4e2d\u6548\u7387\u8f83\u4f4e\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e9b\u53d1\u73b0\u4e3a\u5982\u4f55\u4f7f\u7528 CoT \u63d0\u793a\u548c\u67b6\u6784\u6539\u8fdb\u63d0\u4f9b\u4e86\u5b9e\u8bc1\u6027\u6307\u5bfc\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e09\u3001\u7814\u7a76\u65b9\u6cd5\u4e0e\u6280\u672f\u7ec6\u8282<\/h1>\n\n\n\n<h2 class=\"wp-block-heading\">1) <strong>MME-CoT Bench Design<\/strong><\/h2>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"661\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-48-1024x661.png\"  class=\"wp-image-799\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-48-1024x661.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-48-300x194.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-48-768x496.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-48.png 1035w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe3\" alt=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe3\" \/><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4efb\u52a1\u57df\u6db5\u76d6\u591a\u7c7b\u578b\u8de8\u6a21\u6001\u63a8\u7406\u573a\u666f\uff0c\u4ece<strong>\u6570\u5b66\u56fe\u8868\u7406\u89e3<\/strong>\u3001<strong>\u79d1\u5b66\u5e38\u8bc6\u63a8\u7406<\/strong>\u5230<strong>OCR \u56fe\u50cf\u9605\u8bfb\u518d\u63a8\u7406<\/strong>\u7b49\uff1b<\/li>\n\n\n\n<li>\u6bcf\u7c7b\u4efb\u52a1\u5305\u542b\u9ad8\u8d28\u91cf\u3001\u591a\u6837\u5316\u7684\u6837\u672c\u4ee5\u8bc4\u6d4b\u4e0d\u540c\u5c42\u9762\u7684\u80fd\u529b\uff08\u903b\u8f91\u3001\u6570\u5b66\u3001\u89c6\u89c9\u7406\u89e3\u7b49\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">2) <strong>\u4e09\u5927\u8bc4\u4f30\u6307\u6807<\/strong><\/h2>\n\n\n\n<p>\u8bba\u6587\u63d0\u51fa\u7684\u6307\u6807\u4e0d\u53ea\u662f\u770b\u6700\u7ec8\u6b63\u786e\u7387\uff0c\u8fd8\u5728\u63a8\u7406\u7ed3\u6784\u4e0a\u505a\u6df1\u5165\u5206\u6790\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"765\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-50-1024x765.png\"  class=\"wp-image-801\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-50-1024x765.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-50-300x224.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-50-768x574.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-50.png 1075w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe4\" alt=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe4\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"777\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-51-1024x777.png\"  class=\"wp-image-802\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-51-1024x777.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-51-300x228.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-51-768x582.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-51.png 1076w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe5\" alt=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe5\" \/><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Quality Metrics<\/strong>\uff1a\u7ed3\u5408\u63a8\u7406\u94fe\u6761\u4e0e\u6700\u7ec8\u7b54\u6848\u7684\u8fde\u8d2f\u6027\u3001\u6b63\u786e\u6027\u6253\u5206\uff1b<\/li>\n\n\n\n<li><strong>Robustness Metrics<\/strong>\uff1a\u901a\u8fc7\u6dfb\u52a0\u566a\u97f3\u3001\u6270\u52a8\u540e\u7684\u7a33\u5b9a\u6027\u6d4b\u8bd5\u6a21\u578b\u5bf9\u5fae\u5c0f\u53d8\u5316\u7684\u654f\u611f\u5ea6\uff1b<\/li>\n\n\n\n<li><strong>Efficiency Metrics<\/strong>\uff1a\u7edf\u8ba1\u63a8\u7406\u8fc7\u7a0b\u4e2d\u751f\u6210 token \u6570\u3001\u53cd\u601d\/\u81ea\u7ea0\u9519\u6b65\u6570\u4e0e\u8ba1\u7b97\u6d88\u8017\u3002<a href=\"https:\/\/proceedings.mlr.press\/v267\/jiang25n.html?utm_source=chatgpt.com\" target=\"_blank\" rel=\"noreferrer noopener\">P<\/a><\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u79cd\u8bc4\u4f30\u65b9\u5f0f\u5141\u8bb8\u5bf9\u6bd4\u6a21\u578b\u5728\u4e0d\u540c\u63a8\u7406\u7b56\u7565\u4e0b\uff08\u76f4\u63a5\u56de\u7b54 vs Chain-of-Thought vs Reflection\uff09\u8868\u73b0\u7684\u5dee\u5f02\uff0c\u771f\u6b63\u8861\u91cf\u63a8\u7406\u8d28\u91cf\u800c\u4e0d\u4ec5\u662f\u8f93\u51fa\u51c6\u786e\u6027\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h1 class=\"wp-block-heading\">\u56db\u3001\u5b9e\u9a8c\u4e0e\u7ed3\u679c\u603b\u7ed3<\/h1>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"444\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-52-1024x444.png\"  class=\"wp-image-803\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-52-1024x444.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-52-300x130.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-52-768x333.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-52.png 1037w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe6\" alt=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe6\" \/><\/figure>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"412\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-53-1024x412.png\"  class=\"wp-image-804\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-53-1024x412.png 1024w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-53-300x121.png 300w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-53-768x309.png 768w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-53.png 1049w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" title=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe7\" alt=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe7\" \/><\/figure>\n\n\n\n<p>\u8bba\u6587\u5728\u591a\u4e2a\u6a21\u578b\u4e0a\u505a\u4e86\u5e7f\u6cdb\u6a2a\u5411\u5bf9\u6bd4\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u53cd\u601d\u673a\u5236\u6a21\u578b<\/strong>\uff08reflection models\uff09\u5728\u591a\u6570\u63a8\u7406\u4efb\u52a1\u4e0a\u8868\u73b0\u6700\u4f73\uff1b<\/li>\n\n\n\n<li><strong>\u6807\u51c6 CoT prompting<\/strong> \u4f1a\u5728\u89c6\u89c9\u611f\u77e5\u5bc6\u96c6\u4efb\u52a1\u8868\u73b0\u5f31\u4e8e\u65e0 CoT\uff1b<\/li>\n\n\n\n<li><strong>GPT-4o<\/strong> \u7b49\u5f3a\u5927\u7684\u5546\u4e1a\u6a21\u578b\u867d\u7136\u8868\u73b0\u4f18\u5f02\uff0c\u4f46\u5728\u67d0\u4e9b domain-specific \u63a8\u7406\uff08\u5982 OCR +\u903b\u8f91\uff09\u4ecd\u843d\u540e\u4e8e\u6700\u5148\u8fdb\u7814\u7a76\u6a21\u578b\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e9b\u5b9e\u9a8c\u7ed3\u679c\u63ed\u793a\u51fa\u591a\u6a21\u6001\u63a8\u7406\u4e0d\u540c\u4e8e\u7eaf\u6587\u672c\u63a8\u7406\u7684\u672c\u8d28\uff1a\u56fe\u50cf\u4e0e\u6587\u672c\u4fe1\u606f\u5bf9\u51c6\u786e\u63a8\u7406\u94fe\u6784\u5efa\u6709\u622a\u7136\u4e0d\u540c\u7684\u9700\u6c42\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"513\" height=\"390\" src=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-54.png\"  class=\"wp-image-805\" srcset=\"https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-54.png 513w, https:\/\/www.ndnlab.com\/wp-content\/uploads\/2025\/12\/image-54-300x228.png 300w\" sizes=\"auto, (max-width: 513px) 100vw, 513px\" title=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe8\" alt=\"MME-CoT: Benchmarking Chain-of-Thought in Large Multimodal Models for Reasoning Quality, Robustness, and Efficiency\u63d2\u56fe8\" \/><\/figure>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h1 class=\"wp-block-heading\">\u4e94\u3001\u521b\u65b0\u70b9\u4e0e\u610f\u4e49<\/h1>\n\n\n\n<p>\u8be5\u8bba\u6587\u7684\u4e3b\u8981\u521b\u65b0\u4e0e\u8d21\u732e\u5305\u62ec\uff1a<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>1. \u7b2c\u4e00\u4e2a\u7cfb\u7edf\u8bc4\u4f30\u591a\u6a21\u6001 LLM \u63a8\u7406\u94fe\u80fd\u529b\u7684 benchmark<\/strong><\/h3>\n\n\n\n<p>MME-CoT \u63d0\u4f9b\u4e86\u4ece\u8d28\u91cf\u3001\u7a33\u5065\u6027\u4e0e\u6548\u7387\u4e09\u4e2a\u7ef4\u5ea6\u7684\u7efc\u5408\u6d4b\u8bc4\u65b9\u6cd5\uff0c\u4e3a\u591a\u6a21\u6001\u63a8\u7406\u7814\u7a76\u63d0\u4f9b\u4e86\u6807\u51c6\u5316\u8bc4\u4ef7\u4f53\u7cfb\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>2. \u5bf9 CoT \u5728\u591a\u6a21\u6001\u4efb\u52a1\u4e2d\u7684\u9002\u7528\u6027\u63d0\u51fa\u5b9e\u8bc1\u7ed3\u8bba<\/strong><\/h3>\n\n\n\n<p>\u5b9e\u9a8c\u8bc1\u660e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u601d\u7ef4\u94fe\u5bf9\u67d0\u4e9b\u89c6\u89c9\u903b\u8f91\u4efb\u52a1\u672a\u5fc5\u6709\u76ca\uff1b<\/li>\n\n\n\n<li>\u5f15\u5165\u53cd\u601d\u673a\u5236\u53ef\u4ee5\u63d0\u5347\u903b\u8f91\u8d28\u91cf\u4f46\u727a\u7272\u6548\u7387\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u79cd\u5b9e\u8bc1\u5206\u6790\u5bf9\u4e8e\u6784\u5efa\u66f4\u64c5\u957f\u8de8\u6a21\u6001\u63a8\u7406\u7684 LMMs \u67b6\u6784\u5177\u6709\u91cd\u8981\u6307\u5bfc\u610f\u4e49\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\"><strong>3. \u88ab\u9a8c\u8bc1\u7684\u6a21\u578b\u8868\u73b0\u6d1e\u5bdf\u6307\u5bfc\u672a\u6765\u6a21\u578b\u8bbe\u8ba1<\/strong><\/h3>\n\n\n\n<p>\u8bba\u6587\u6307\u51fa\u672a\u6765\u6539\u8fdb\u65b9\u5411\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u66f4\u9002\u5408\u89c6\u89c9\u63a8\u7406\u7684 CoT \u5f62\u5f0f\uff08\u5982\u89c6\u89c9\u94fe\u63d0\u793a\uff09\uff1b<\/li>\n\n\n\n<li>\u4e0e\u53cd\u601d\u673a\u5236\u7ed3\u5408\u7684\u9ad8\u6548\u67b6\u6784\u8bbe\u8ba1\uff1b<\/li>\n\n\n\n<li>\u66f4\u7a33\u5065\u7684 multimodal grounding \u4e0e reasoning \u6a21\u5757\u3002<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\" \/>\n\n\n\n<h1 class=\"wp-block-heading\">\u516d\u3001\u7ed3\u8bba\u4e0e\u672a\u6765\u65b9\u5411<\/h1>\n\n\n\n<p><strong>\u7ed3\u8bba\uff1a<\/strong> \u8bba\u6587\u9996\u6b21\u63d0\u51fa\u4e86\u5bf9\u591a\u6a21\u6001\u5927\u6a21\u578b\u8de8\u6a21\u6001\u63a8\u7406\u4e09\u7ef4\u5ea6\u7684\u8bc4\u4ef7\u4f53\u7cfb\uff0c\u5e76\u57fa\u4e8e\u6b64\u4f53\u7cfb\u5b9e\u8bc1\u53d1\u73b0\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u73b0\u5728\u6700\u5148\u8fdb\u7684\u5927\u6a21\u578b\u5728\u601d\u7ef4\u94fe\u65b9\u9762\u5df2\u7ecf\u5177\u5907\u67d0\u4e9b\u4f18\u52bf\uff1b<\/li>\n\n\n\n<li>CoT \u5e76\u975e\u5728\u6240\u6709\u8de8\u6a21\u6001\u4efb\u52a1\u4e2d\u5747\u6709\u6548\uff1b<\/li>\n\n\n\n<li>\u53cd\u601d\u673a\u5236\u80fd\u591f\u63d0\u5347\u8d28\u91cf\uff0c\u4f46\u9700\u8981\u6743\u8861\u6548\u7387\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>\u672a\u6765\u65b9\u5411\uff1a<\/strong> \u7814\u7a76\u8005\u5e94\u91cd\u70b9\u5173\u6ce8\u5982\u4f55\u4f7f\u63a8\u7406\u94fe\u5bf9\u89c6\u89c9\u611f\u77e5\u4efb\u52a1\u66f4\u6709\u6548\u3001\u63d0\u9ad8\u8de8\u6a21\u6001\u7a33\u5b9a\u6027\u3001\u4f18\u5316\u63a8\u7406\u6548\u7387\uff0c\u4ee5\u53ca\u6784\u5efa\u66f4\u5168\u9762\u7684 reasoning \u57fa\u51c6<\/p>\n\n\n\n<p><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f1a\u8bae\uff1aICML 2025\uff08\u7b2c42\u5c4a\u56fd\u9645\u673a\u5668\u5b66\u4e60\u5927\u4f1a\uff09 Dongzhi Jiang\u2217 1 , Renrui Zhang\u2217\u2020 1 , Ziyu Guo2 , Yanwei Li\u20213 , Yu Qi\u20214 , Xinyan Chen\u20211Liuhui Wang\u20215 , Jianhan Jin\u20216 , Claire Guo\u20217 , Shen Yan3 , Bo Zhang8Chaoyou Fu6 , Peng Gao8 , Hongsheng Li1arXiv:2502.09621v1  &hellip; <a href=\"https:\/\/www.ndnlab.com\/?p=795\">\u7ee7\u7eed\u9605\u8bfb <span class=\"meta-nav\">&rarr;<\/span><\/a><\/p>\n","protected":false},"author":2,"featured_media":797,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-795","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/795","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=795"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/795\/revisions"}],"predecessor-version":[{"id":806,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/795\/revisions\/806"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/797"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=795"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=795"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=795"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}