{"id":5249,"date":"2024-12-04T13:27:19","date_gmt":"2024-12-04T05:27:19","guid":{"rendered":"https:\/\/nullthought.net\/?p=5249"},"modified":"2025-03-03T12:25:31","modified_gmt":"2025-03-03T04:25:31","slug":"vlsi%e6%a8%a1%e5%9e%8b%ef%bc%9a%e9%80%90%e5%b1%82%e8%92%b8%e9%a6%8f%ef%bc%8c%e9%80%90%e5%b1%82%e5%af%b9%e9%bd%90%ef%bc%8c%e5%ae%9e%e7%8e%b0%e4%bb%8e%e5%a4%a7%e8%a7%84%e6%a8%a1%e5%88%b0%e5%b0%8f","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=5249","title":{"rendered":"VLsI\u6a21\u578b\uff1a\u9010\u5c42\u84b8\u998f\uff0c\u9010\u5c42\u5bf9\u9f50\uff0c\u5b9e\u73b0\u4ece\u5927\u89c4\u6a21\u5230\u5c0f\u89c4\u6a21\u89c6\u89c9-\u8bed\u8a00\u6a21\u578b\uff08VLM\uff09\u7684\u9ad8\u6548\u77e5\u8bc6\u8fc1\u79fb"},"content":{"rendered":"\n<p>\u8bba\u6587\u300a<strong><a href=\"https:\/\/arxiv.org\/abs\/2412.01822\" target=\"_blank\" rel=\"noreferrer noopener\">VLsI: Verbalized Layers-to-Interactions from Large to Small Vision Language Models<\/a><\/strong>\u300b\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u578b\u7684\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u5bb6\u65cf\u2014\u2014VLsI\uff0c\u9488\u5bf9\u5982\u4f55\u5728\u4fdd\u6301\u6a21\u578b\u6027\u80fd\u7684\u524d\u63d0\u4e0b\u5b9e\u73b0\u66f4\u9ad8\u6548\u7684\u89c6\u89c9-\u8bed\u8a00\u6a21\u578b\uff08VLM\uff09\u8bbe\u8ba1\u3002\u8bba\u6587\u63d0\u51fa\u7684VLsI\u901a\u8fc7\u521b\u65b0\u6027\u5730\u5f15\u5165\u81ea\u7136\u8bed\u8a00\u4f5c\u4e3a\u4e2d\u4ecb\uff0c\u5b9e\u73b0\u4e86\u4ece\u5927\u89c4\u6a21\u5230\u5c0f\u89c4\u6a21\u89c6\u89c9-\u8bed\u8a00\u6a21\u578b\u7684\u9ad8\u6548\u77e5\u8bc6\u8fc1\u79fb\uff0c\u5c55\u793a\u4e86\u5728\u4e0d\u589e\u52a0\u6a21\u578b\u590d\u6742\u5ea6\u7684\u60c5\u51b5\u4e0b\u4fdd\u6301\u751a\u81f3\u8d85\u8d8a\u5927\u6a21\u578b\u6027\u80fd\u7684\u53ef\u80fd\u6027\u3002VLsI\u6a21\u578b\u7684\u9010\u5c42\u84b8\u998f\u548c\u9010\u5c42\u5bf9\u9f50\u65b9\u6cd5\u4e0d\u4ec5\u63d0\u9ad8\u4e86\u5c0f\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u8fd8\u5c55\u793a\u4e86\u81ea\u7136\u8bed\u8a00\u4f5c\u4e3a\u77e5\u8bc6\u4f20\u9012\u5a92\u4ecb\u7684\u6f5c\u529b\u3002\u8fd9\u9879\u7814\u7a76\u4e3a\u672a\u6765\u5982\u4f55\u901a\u8fc7\u81ea\u7136\u8bed\u8a00\u6765\u5b9e\u73b0AI\u6a21\u578b\u4e4b\u95f4\u7684\u77e5\u8bc6\u8fc1\u79fb\u63d0\u4f9b\u4e86\u65b0\u601d\u8def\uff0c\u63a8\u52a8\u4e86\u89c6\u89c9-\u8bed\u8a00\u6a21\u578b\u5728\u8d44\u6e90\u53d7\u9650\u73af\u5883\u4e2d\u7684\u5b9e\u9645\u5e94\u7528\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aByung-Kwan Lee, Ryo Hachiuma, Yu-Chiang Frank Wang, Yong Man Ro, Yueh-Hua Wu\uff0c\u6765\u81eaNvidia\u548cKAIST\uff08Korea Advanced Institute of Science and Technology\uff09\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"845\" height=\"406\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/12\/image-1.png\" alt=\"\" class=\"wp-image-5250\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2024\/12\/image-1.png 845w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/12\/image-1-300x144.png 300w, https:\/\/nullthought.net\/wp-content\/uploads\/2024\/12\/image-1-768x369.png 768w\" sizes=\"auto, (max-width: 845px) 100vw, 845px\" \/><figcaption class=\"wp-element-caption\"><strong><a href=\"https:\/\/arxiv.org\/abs\/2412.01822\" target=\"_blank\" rel=\"noreferrer noopener\">VLsI: Verbalized Layers-to-Interactions from Large to Small Vision Language Models<\/a><\/strong><\/figcaption><\/figure>\n\n\n\n<h4 class=\"wp-block-heading\">1. \u7814\u7a76\u80cc\u666f\u4e0e\u52a8\u673a<\/h4>\n\n\n\n<p>\u8fd1\u5e74\u6765\uff0c<strong><a href=\"https:\/\/nullthought.net\/?p=4920\" target=\"_blank\" rel=\"noreferrer noopener\">\u89c6\u89c9-\u8bed\u8a00\u6a21\u578b\uff08Vision-Language Models, VLMs\uff09<\/a><\/strong>\u7684\u53d1\u5c55\u6781\u5927\u5730\u589e\u5f3a\u4e86\u89c6\u89c9\u7cfb\u7edf\u7684\u89e3\u8bfb\u548c\u5904\u7406\u80fd\u529b\u3002<a href=\"https:\/\/nullthought.net\/?tag=llm\" target=\"_blank\" rel=\"noreferrer noopener\">\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09<\/a>\u4e0e\u89c6\u89c9\u7f16\u7801\u5668\u7ed3\u5408\uff0c\u901a\u8fc7\u89c6\u89c9\u6307\u4ee4\u8c03\u4f18\uff08Visual Instruction Tuning\uff09\uff0c\u5b9e\u73b0\u4e86\u5bf9\u590d\u6742\u89c6\u89c9\u8f93\u5165\u7684\u7406\u89e3\u3002\u4f8b\u5982\uff0c\u5c01\u95ed\u6e90\u6a21\u578b\u5982GPT-4V\u548cGemini-Pro\u5229\u7528\u9ad8\u8d28\u91cf\u7684\u6307\u4ee4\u6837\u672c\uff0c\u5728\u89c6\u89c9\u8bed\u8a00\u7406\u89e3\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u524d\u6240\u672a\u6709\u7684\u6027\u80fd\u3002\u7136\u800c\uff0c\u968f\u7740\u6a21\u578b\u89c4\u6a21\u7684\u6269\u5927\uff0c\u8ba1\u7b97\u6210\u672c\u7684\u6025\u5267\u589e\u52a0\u6210\u4e3a\u4e86\u90e8\u7f72\u8fd9\u4e9b\u6a21\u578b\u7684\u4e3b\u8981\u969c\u788d\uff0c\u5c24\u5176\u662f\u5728\u79fb\u52a8\u8bbe\u5907\u548c\u81ea\u52a8\u5316\u673a\u5668\u4eba\u7b49\u8d44\u6e90\u53d7\u9650\u7684\u573a\u666f\u4e2d\u3002<\/p>\n\n\n\n<p>\u4f20\u7edf\u7684\u89e3\u51b3\u65b9\u6848\u901a\u5e38\u6d89\u53ca\u6dfb\u52a0\u4e13\u95e8\u7684\u6a21\u5757\u6216\u4fee\u6539\u6a21\u578b\u67b6\u6784\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u867d\u7136\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u89e3\u51b3\u4e86\u6027\u80fd\u95ee\u9898\uff0c\u4f46\u589e\u52a0\u4e86\u5de5\u7a0b\u590d\u6742\u5ea6\uff0c\u5bfc\u81f4\u5728\u90e8\u7f72\u8fc7\u7a0b\u4e2d\u7684\u517c\u5bb9\u6027\u95ee\u9898\uff0c\u7279\u522b\u662f\u5728\u4f4e\u5ef6\u8fdf\u548c\u5185\u5b58\u6548\u7387\u81f3\u5173\u91cd\u8981\u7684\u573a\u666f\u4e2d\u3002\u7531\u6b64\u53ef\u89c1\uff0c\u73b0\u6709\u7684\u65b9\u6cd5\u867d\u7136\u5728\u7406\u8bba\u4e0a\u63d0\u5347\u4e86\u6a21\u578b\u80fd\u529b\uff0c\u4f46\u5728\u5b9e\u8df5\u4e2d\u5374\u5b58\u5728\u663e\u8457\u7684\u5e94\u7528\u969c\u788d\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u4e2a\u5173\u952e\u95ee\u9898\uff1a\u201c\u5982\u4f55\u5728\u4e0d\u6269\u5927\u6a21\u578b\u89c4\u6a21\u6216\u8fdb\u884c\u7ed3\u6784\u53d8\u66f4\u7684\u60c5\u51b5\u4e0b\uff0c\u5b9e\u73b0\u4e0e\u5927\u6a21\u578b\u76f8\u4f3c\u751a\u81f3\u66f4\u9ad8\u7684\u89c6\u89c9-\u8bed\u8a00\u7406\u89e3\u6027\u80fd\uff1f\u201d\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u521b\u65b0\u6027\u7684<strong>\u9010\u5c42\u84b8\u998f\u7b56\u7565<\/strong>\uff0c\u65e8\u5728\u9ad8\u6548\u5730\u5c06\u5927\u6a21\u578b\u4e2d\u7684\u77e5\u8bc6\u8fc1\u79fb\u5230\u5c0f\u6a21\u578b\u4e2d\uff0c\u4e14\u4e0d\u635f\u5931\u6027\u80fd\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">2. VLsI \u6a21\u578b\u6982\u8ff0<\/h4>\n\n\n\n<p>\u8bba\u6587\u63d0\u51fa\u7684\u6a21\u578b\u5bb6\u65cf\u540d\u4e3a<strong>VLsI\uff08Verbalized Layers-to-Interactions\uff09<\/strong>\uff0c\u662f\u4e00\u79cd\u65b0\u578b\u7684\u89c6\u89c9-\u8bed\u8a00\u6a21\u578b\uff0c\u91cd\u70b9\u5728\u4e8e\u901a\u8fc7\u81ea\u7136\u8bed\u8a00\u4e3a\u4e2d\u4ecb\u8fdb\u884c\u9010\u5c42\u84b8\u998f\uff0c\u4ee5\u5b9e\u73b0\u5c0f\u6a21\u578b\u5bf9\u5927\u6a21\u578b\u63a8\u7406\u8fc7\u7a0b\u7684\u6a21\u4eff\u3002\u8be5\u6a21\u578b\u57282B\u548c7B\u4e24\u4e2a\u89c4\u6a21\u4e0a\u8fdb\u884c\u8bbe\u8ba1\u548c\u5b9e\u9a8c\u9a8c\u8bc1\uff0c\u5206\u522b\u5bf9\u5e94\u4e0d\u540c\u7684\u5e94\u7528\u573a\u666f\u548c\u8d44\u6e90\u9650\u5236\u3002<\/p>\n\n\n\n<p>\u4e0e\u4f20\u7edf\u7684<strong><a href=\"https:\/\/nullthought.net\/?p=5098\" target=\"_blank\" rel=\"noreferrer noopener\">\u84b8\u998f\u65b9\u6cd5<\/a><\/strong>\u76f4\u63a5\u6a21\u4eff\u5927\u6a21\u578b\u6700\u7ec8\u8f93\u51fa\u4e0d\u540c\uff0cVLsI\u5f15\u5165\u4e86\u4e00\u79cd\u9010\u5c42\u5bf9\u9f50\u7684\u8fc7\u7a0b\uff0c\u901a\u8fc7\u4e2d\u95f4\u5c42\u7684\u8bed\u8a00\u5316\u8868\u8fbe\uff08\u5373\u201c\u53e3\u8ff0\u5668\u201d\uff09\uff0c\u8ba9\u6bcf\u4e00\u5c42\u7684\u8f93\u51fa\u90fd\u80fd\u591f\u6620\u5c04\u5230\u81ea\u7136\u8bed\u8a00\u7a7a\u95f4\u3002\u8fd9\u79cd\u65b9\u6cd5\u7684\u72ec\u7279\u4e4b\u5904\u5728\u4e8e\uff0c\u6bcf\u4e00\u5c42\u7684\u5bf9\u9f50\u4e0d\u518d\u4ec5\u4ec5\u662f\u6570\u503c\u7279\u5f81\u7684\u7b80\u5355\u6620\u5c04\uff0c\u800c\u662f\u901a\u8fc7\u81ea\u7136\u8bed\u8a00\u7684\u63cf\u8ff0\u5b9e\u73b0\u66f4\u4e3a\u76f4\u89c2\u7684\u7279\u5f81\u8868\u8fbe\u3002\u8fd9\u79cd\u65b9\u6cd5\u4e0d\u4ec5\u80fd\u591f\u5728\u9010\u5c42\u7684\u8bed\u4e49\u7a7a\u95f4\u4e0a\u5b9e\u73b0\u5c0f\u6a21\u578b\u4e0e\u5927\u6a21\u578b\u7684\u9ad8\u6548\u5bf9\u9f50\uff0c\u8fd8\u907f\u514d\u4e86\u8f93\u51fa\u6a21\u4eff\u8fc7\u7a0b\u4e2d\u5e38\u89c1\u7684\u4e0d\u7a33\u5b9a\u6027\u95ee\u9898\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">3. VLsI\u7684\u6838\u5fc3\u65b9\u6cd5\u4e0e\u6b65\u9aa4<\/h4>\n\n\n\n<p>VLsI\u7684\u6838\u5fc3\u65b9\u6cd5\u662f\u9010\u5c42\u84b8\u998f\uff0c\u5206\u4e3a\u4e09\u4e2a\u4e3b\u8981\u6b65\u9aa4\uff1a<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">3.1 Verbalization Step\uff08\u53e3\u8ff0\u5316\u6b65\u9aa4\uff09<\/h5>\n\n\n\n<p><strong>\u53e3\u8ff0\u5316\u6b65\u9aa4<\/strong>\u662f\u6574\u4e2a\u9010\u5c42\u84b8\u998f\u7684\u57fa\u7840\u3002\u5728\u8fd9\u4e00\u6b65\u9aa4\u4e2d\uff0c\u6bcf\u4e2a\u76ee\u6807\u4e2d\u95f4\u5c42\u901a\u8fc7\u5f15\u5165\u201c\u53e3\u8ff0\u5668\u201d\uff08Verbalizer\uff09\uff0c\u5c06\u4e2d\u95f4\u7279\u5f81\u6295\u5c04\u5230\u81ea\u7136\u8bed\u8a00\u7a7a\u95f4\u3002\u5177\u4f53\u800c\u8a00\uff0c\u53e3\u8ff0\u5668\u7531\u4e00\u4e2a<strong>\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff08verb-FFN\uff09\u548c\u8bed\u8a00\u5934\uff08Language Head\uff09<\/strong>\u7ec4\u6210\uff0c\u7528\u4e8e\u5904\u7406\u6bcf\u4e00\u5c42\u7684\u8f93\u51fa\u5e76\u5c06\u5176\u6620\u5c04\u4e3a\u53ef\u7406\u89e3\u7684\u81ea\u7136\u8bed\u8a00\u6587\u672c\u3002\u8fd9\u6837\u7684\u8bbe\u8ba1\u4f7f\u5f97\u4e2d\u95f4\u5c42\u4e0d\u518d\u4ec5\u4ec5\u662f\u6df1\u5ea6\u7279\u5f81\u7684\u8868\u793a\uff0c\u800c\u662f\u53ef\u4ee5\u751f\u6210\u6587\u5b57\u63cf\u8ff0\uff0c\u5e2e\u52a9\u5c0f\u6a21\u578b\u7406\u89e3\u5927\u6a21\u578b\u7684\u63a8\u7406\u903b\u8f91\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u786e\u4fdd\u8fd9\u4e9b\u53e3\u8ff0\u5316\u8f93\u51fa\u80fd\u591f\u6b63\u786e\u5bf9\u9f50\u76ee\u6807\u54cd\u5e94\uff0c\u8bba\u6587\u91c7\u7528\u4e86<strong>\u81ea\u56de\u5f52\u635f\u5931\uff08Autoregressive Loss\uff09<\/strong>\uff0c\u4f7f\u5f97\u6bcf\u4e00\u5c42\u7684\u53e3\u8ff0\u5316\u8f93\u51fa\u4e0e\u9884\u671f\u7684\u6587\u672c\u76f8\u5339\u914d\u3002\u8fd9\u4e00\u8fc7\u7a0b\u901a\u8fc7\u5c06\u4e2d\u95f4\u5c42\u7684\u7279\u5f81\u53d8\u5f97\u53ef\u89e3\u91ca\u5316\uff0c\u4ece\u800c\u4f7f\u5f97\u5c0f\u6a21\u578b\u5728\u9010\u6b65\u5b66\u4e60\u5927\u6a21\u578b\u63a8\u7406\u903b\u8f91\u7684\u8fc7\u7a0b\u4e2d\u80fd\u591f\u66f4\u597d\u5730\u6a21\u4eff\u5176\u884c\u4e3a\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">3.2 Interaction Step\uff08\u4ea4\u4e92\u6b65\u9aa4\uff09<\/h5>\n\n\n\n<p><strong>\u4ea4\u4e92\u6b65\u9aa4<\/strong>\u662f\u6574\u4e2a\u9010\u5c42\u5bf9\u9f50\u7684\u6838\u5fc3\uff0c\u65e8\u5728\u4f7f\u5c0f\u6a21\u578b\u80fd\u591f\u6a21\u4eff\u5927\u6a21\u578b\u7684\u63a8\u7406\u8fdb\u5c55\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u5c0f\u6a21\u578b\u7684\u6bcf\u4e00\u5c42\u4f1a\u5728\u5927\u6a21\u578b\u4e2d\u5bfb\u627e\u4e00\u4e2a\u5bf9\u5e94\u7684\u5c42\u6765\u5bf9\u9f50\uff0c\u8fd9\u79cd\u5bf9\u9f50\u4e0d\u4ec5\u4ec5\u662f\u7b80\u5355\u7684\u9010\u5c42\u5bf9\u5e94\uff0c\u800c\u662f\u901a\u8fc7\u4e00\u79cd\u81ea\u9002\u5e94\u7684\u5c42\u5339\u914d\u7b56\u7565\u52a8\u6001\u5b9e\u73b0\u7684\u3002\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u57fa\u4e8e<strong>KL\u6563\u5ea6\uff08Kullback-Leibler Divergence\uff09<\/strong>\u7684\u591a\u9879\u5f0f\u91c7\u6837\u7b56\u7565\uff0c\u786e\u4fdd\u5c0f\u6a21\u578b\u6bcf\u4e00\u5c42\u6240\u5339\u914d\u7684\u76ee\u6807\u5c42\u90fd\u662f\u4e0e\u5176\u63a8\u7406\u8fdb\u5c55\u6700\u4e3a\u76f8\u7b26\u7684\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff0c\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd<strong>\u9010\u6b65\u641c\u7d22\u4e0e\u81ea\u9002\u5e94\u5339\u914d<\/strong>\u7684\u65b9\u6cd5\u3002\u4f8b\u5982\uff0c\u5f53\u5c0f\u6a21\u578b\u7684\u7b2c2\u5c42\u4e0e\u5927\u6a21\u578b\u7684\u7b2c4\u5c42\u5339\u914d\u65f6\uff0c\u5c0f\u6a21\u578b\u7684\u7b2c3\u5c42\u5c06\u5728\u5927\u6a21\u578b\u7684\u7b2c5\u5230\u7b2c7\u5c42\u4e4b\u95f4\u8fdb\u884c\u641c\u7d22\u3002\u8fd9\u79cd\u9010\u6b65\u5bf9\u9f50\u7684\u65b9\u6cd5\u786e\u4fdd\u4e86\u63a8\u7406\u903b\u8f91\u7684\u8fde\u7eed\u6027\u548c\u9010\u6b65\u6df1\u5165\u6027\u3002<\/p>\n\n\n\n<h5 class=\"wp-block-heading\">3.3 Reinforcement Step\uff08\u5f3a\u5316\u6b65\u9aa4\uff09<\/h5>\n\n\n\n<p>\u5728\u9010\u5c42\u84b8\u998f\u5b8c\u6210\u540e\uff0c\u8fd8\u9700\u8981\u8fdb\u4e00\u6b65\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u7684\u65b9\u6cd5\u6765\u5fae\u8c03\u6574\u4e2a\u5c0f\u6a21\u578b\uff0c\u4ee5\u66f4\u597d\u5730\u5438\u6536\u5927\u6a21\u578b\u7684\u77e5\u8bc6\u3002<strong>\u5f3a\u5316\u6b65\u9aa4<\/strong>\u4e3b\u8981\u901a\u8fc7\u81ea\u56de\u5f52\u635f\u5931\u6765\u589e\u5f3a\u5c0f\u6a21\u578b\u7684\u4efb\u52a1\u7279\u5b9a\u6307\u4ee4\u8ddf\u968f\u80fd\u529b\uff0c\u4ece\u800c\u8fdb\u4e00\u6b65\u63d0\u9ad8\u5176\u5728\u5177\u4f53\u89c6\u89c9-\u8bed\u8a00\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002\u8fd9\u4e00\u8fc7\u7a0b\u7684\u7075\u611f\u6765\u6e90\u4e8e\u57fa\u4e8e\u526a\u679d\u7684\u84b8\u998f\u65b9\u6cd5\uff0c\u901a\u8fc7\u989d\u5916\u7684\u8bad\u7ec3\u6765\u5f25\u8865\u7ed3\u6784\u53d8\u5316\u53ef\u80fd\u5e26\u6765\u7684\u6027\u80fd\u4e0b\u964d\u3002\u5f3a\u5316\u6b65\u9aa4\u4e0d\u4ec5\u80fd\u591f\u4f7f\u5c0f\u6a21\u578b\u5728\u7ed3\u6784\u4e0d\u53d8\u7684\u60c5\u51b5\u4e0b\u66f4\u597d\u5730\u5bf9\u9f50\u5927\u6a21\u578b\uff0c\u8fd8\u63d0\u9ad8\u4e86\u5c0f\u6a21\u578b\u5728\u751f\u6210\u54cd\u5e94\u65f6\u7684\u51c6\u786e\u6027\u548c\u4e00\u81f4\u6027\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">4. \u5b9e\u9a8c\u4e0e\u7ed3\u679c\u5206\u6790<\/h4>\n\n\n\n<p>\u8bba\u6587\u5728\u5341\u4e2a\u5177\u6709\u6311\u6218\u6027\u7684\u89c6\u89c9-\u8bed\u8a00\u57fa\u51c6\u6d4b\u8bd5\u4e0a\u9a8c\u8bc1\u4e86VLsI\u7684\u6709\u6548\u6027\uff0c\u5c55\u793a\u4e86\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6027\u80fd\u6bd4\u8f83<\/strong>\uff1a\u5728\u4e0d\u589e\u52a0\u6a21\u578b\u89c4\u6a21\u6216\u66f4\u6539\u67b6\u6784\u7684\u60c5\u51b5\u4e0b\uff0c2B\u548c7B\u89c4\u6a21\u7684VLsI\u6a21\u578b\u76f8\u8f83\u4e8eGPT-4V\u5206\u522b\u53d6\u5f97\u4e86<strong>11.0%\u548c17.4%<\/strong>\u7684\u6027\u80fd\u63d0\u5347\u3002\u8fd9\u6837\u7684\u7ed3\u679c\u8868\u660e\uff0cVLsI\u5728\u4fdd\u6301\u6a21\u578b\u9ad8\u6548\u6027\u7684\u540c\u65f6\u80fd\u591f\u8fbe\u5230\u751a\u81f3\u8d85\u8d8a\u73b0\u6709\u7684\u5927\u89c4\u6a21\u6a21\u578b\u3002<\/li>\n\n\n\n<li><strong>\u5b9e\u9a8c\u8bbe\u7f6e<\/strong>\uff1a\u5b9e\u9a8c\u5728\u516b\u53f0NVIDIA A100 80GB GPU\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u4f7f\u7528\u4e86LoRA\uff08Low-Rank Adaptation of Large Language Models\uff09\u6765\u4f18\u5316\u6a21\u578b\u53c2\u6570\uff0c\u91c7\u7528AdamW\u4f18\u5316\u5668\u548c\u4f59\u5f26\u9000\u706b\u8c03\u5ea6\u6765\u63a7\u5236\u5b66\u4e60\u7387\u3002\u8fd9\u4e9b\u6280\u672f\u8bbe\u7f6e\u786e\u4fdd\u4e86\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u9ad8\u6548\u6027\u548c\u7a33\u5b9a\u6027\u3002<\/li>\n\n\n\n<li><strong>\u57fa\u51c6\u6d4b\u8bd5<\/strong>\uff1a\u5728\u89c6\u89c9-\u8bed\u8a00\u57fa\u51c6\u5982MM-Vet\u3001MMMU\u3001MathVista\u7b49\u6d4b\u8bd5\u4e2d\uff0cVLsI\u6a21\u578b\u8868\u73b0\u51fa\u8272\uff0c\u5c24\u5176\u5728\u8d44\u6e90\u53d7\u9650\u7684\u60c5\u51b5\u4e0b\uff0c\u76f8\u8f83\u4e8e\u5176\u4ed6\u5f00\u653e\u6e90\u6a21\u578b\u548c\u5c01\u95ed\u6e90\u6a21\u578b\u90fd\u5c55\u73b0\u4e86\u663e\u8457\u7684\u4f18\u52bf\u3002\u8868\u683c1\u548c\u8868\u683c2\u4e2d\u5c55\u793a\u4e86VLsI\u76f8\u8f83\u4e8e\u5176\u4ed6\u6a21\u578b\u7684\u5177\u4f53\u8868\u73b0\uff0cVLsI\u6a21\u578b\u4e0d\u4ec5\u5728\u6a21\u578b\u5927\u5c0f\u4e0a\u6709\u4f18\u52bf\uff0c\u5176\u5728\u5bf9\u89c6\u89c9\u548c\u8bed\u8a00\u7684\u7efc\u5408\u7406\u89e3\u80fd\u529b\u4e0a\u4e5f\u8868\u73b0\u5f97\u66f4\u52a0\u7a81\u51fa\u3002<\/li>\n<\/ul>\n\n\n\n<h4 class=\"wp-block-heading\">5. \u53e3\u8ff0\u5668\u4e0e\u5c42\u6b21\u5bf9\u9f50\u7b56\u7565\u7684\u6548\u679c<\/h4>\n\n\n\n<p>\u8bba\u6587\u4e2d\u5c55\u793a\u4e86VLsI\u5982\u4f55\u901a\u8fc7\u53e3\u8ff0\u5668\u9010\u5c42\u751f\u6210\u8bed\u8a00\u5316\u8f93\u51fa\uff0c\u4ece\u800c\u8ffd\u8e2a\u5404\u5c42\u6b21\u7684\u63a8\u7406\u8fdb\u5c55\u3002\u56fe3\u4e2d\u4e3e\u4f8b\u8bf4\u660e\u4e86\u5982\u4f55\u4f7f\u7528\u53e3\u8ff0\u5316\u7684\u8f93\u51fa\u6765\u8ffd\u8e2a\u6bcf\u4e00\u5c42\u7684\u89e3\u8bfb\u8fc7\u7a0b\uff0c\u5e76\u6bd4\u8f83\u4e86\u4f20\u7edf\u5c0f\u6a21\u578b\u548c\u4f7f\u7528VLsI\u7684\u6a21\u578b\u5728\u63a8\u7406\u80fd\u529b\u4e0a\u7684\u5dee\u5f02\u3002\u6d45\u5c42\u6b21\u65f6\uff0c\u4e24\u79cd\u6a21\u578b\u751f\u6210\u7684\u63cf\u8ff0\u5927\u591a\u662f\u5bf9\u7b80\u5355\u5f62\u72b6\u548c\u989c\u8272\u7684\u63cf\u8ff0\uff0c\u4f46\u968f\u7740\u6df1\u5ea6\u589e\u52a0\uff0cVLsI\u6a21\u578b\u80fd\u591f\u8bc6\u522b\u548c\u63cf\u8ff0\u66f4\u590d\u6742\u7684\u89c6\u89c9\u7ed3\u6784\uff0c\u6700\u540e\u4e00\u5c42\u751a\u81f3\u80fd\u591f\u51c6\u786e\u63cf\u8ff0\u51fa\u76ee\u6807\u54cd\u5e94\uff0c\u800c\u4f20\u7edf\u5c0f\u6a21\u578b\u5219\u65e0\u6cd5\u6355\u6349\u5230\u8fd9\u4e00\u7279\u5b9a\u6a21\u5f0f\u3002\u8fd9\u8868\u660e\uff0c\u9010\u5c42\u8bed\u8a00\u5316\u7684\u84b8\u998f\u65b9\u6cd5\u5728\u6539\u5584\u63a8\u7406\u7684\u51c6\u786e\u6027\u548c\u4e00\u81f4\u6027\u65b9\u9762\u5177\u6709\u660e\u663e\u4f18\u52bf\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">6. \u5c42\u5339\u914d\u7b56\u7565\u7684\u6df1\u5165\u5206\u6790<\/h4>\n\n\n\n<p>\u5728\u4ea4\u4e92\u6b65\u9aa4\u4e2d\uff0c\u8bba\u6587\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u5c42\u5339\u914d\u7b56\u7565\uff0c\u5229\u7528KL\u6563\u5ea6\u6765\u9009\u62e9\u6bcf\u4e00\u5c42\u7684\u5339\u914d\u3002\u8fd9\u79cd\u65b9\u6cd5\u901a\u8fc7\u591a\u9879\u5f0f\u91c7\u6837\u786e\u4fdd\u4e86\u5c0f\u6a21\u578b\u80fd\u591f\u5c3d\u53ef\u80fd\u9009\u62e9\u6700\u4f18\u7684\u5339\u914d\u5c42\uff0c\u907f\u514d\u4e86\u4f20\u7edf\u9010\u5c42\u5339\u914d\u4e2d\u53ef\u80fd\u5b58\u5728\u7684\u6b21\u4f18\u9009\u62e9\u3002\u8bba\u6587\u901a\u8fc7\u5bf9\u6bd4\u4e0d\u540c\u7684\u5339\u914d\u7b56\u7565\uff0c\u8bc1\u660e\u4e86KL\u6563\u5ea6\u7ed3\u5408\u9010\u6b65\u5339\u914d\u7684\u4f18\u8d8a\u6027\uff0c\u8fd9\u6837\u7684\u7b56\u7565\u5728\u5b9e\u9a8c\u4e2d\u663e\u793a\u51fa\u4e86\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002\u6b64\u5916\uff0c\u5f3a\u5316\u6b65\u9aa4\u8fdb\u4e00\u6b65\u589e\u5f3a\u4e86\u5c0f\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\uff0c\u4f7f\u5f97\u5b83\u80fd\u591f\u66f4\u597d\u5730\u5bf9\u9f50\u5e76\u5438\u6536\u5927\u6a21\u578b\u7684\u77e5\u8bc6\u3002<\/p>\n\n\n\n<h4 class=\"wp-block-heading\">7. \u9650\u5236\u4e0e\u672a\u6765\u5de5\u4f5c<\/h4>\n\n\n\n<p>\u5c3d\u7ba1VLsI\u5728\u5b9e\u9a8c\u4e2d\u5c55\u793a\u4e86\u826f\u597d\u7684\u6027\u80fd\uff0c\u8bba\u6587\u4e5f\u6307\u51fa\u4e86\u4e00\u4e9b<strong>\u5c40\u9650\u6027<\/strong>\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u5171\u4eab\u5206\u8bcd\u5668<\/strong>\uff1a\u76ee\u524d\uff0c\u5927\u5c0f\u6a21\u578b\u4e4b\u95f4\u5fc5\u987b\u5171\u4eab\u76f8\u540c\u7684\u5206\u8bcd\u5668\u548c\u8bcd\u6c47\u7d22\u5f15\u987a\u5e8f\uff0c\u8fd9\u9650\u5236\u4e86VLsI\u5728\u67d0\u4e9b\u7279\u5b9a\u573a\u666f\u4e2d\u7684\u9002\u7528\u6027\u3002\u672a\u6765\u7684\u7814\u7a76\u65b9\u5411\u5c06\u63a2\u7d22\u5982\u4f55\u5728\u4e0d\u540c\u7684\u5206\u8bcd\u5668\u548c\u8bcd\u6c47\u7d22\u5f15\u987a\u5e8f\u4e0b\u5b9e\u73b0VLsI\u7684\u901a\u7528\u6027\u3002<\/li>\n\n\n\n<li><strong>\u8ba1\u7b97\u8d44\u6e90\u8981\u6c42<\/strong>\uff1a\u867d\u7136VLsI\u51cf\u5c11\u4e86\u6a21\u578b\u89c4\u6a21\uff0c\u4f46\u5728\u8bad\u7ec3\u9636\u6bb5\u4ecd\u7136\u9700\u8981\u8f83\u5927\u7684\u8ba1\u7b97\u8d44\u6e90\uff0c\u5c24\u5176\u662f\u9700\u8981\u5927\u6a21\u578b\u4f5c\u4e3a\u9aa8\u5e72\u6765\u8fdb\u884c\u9010\u5c42\u5bf9\u9f50\u3002\u8fd9\u5bf9\u8d44\u6e90\u53d7\u9650\u7684\u7528\u6237\u6765\u8bf4\u4ecd\u662f\u4e00\u4e2a\u6311\u6218\u3002<\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587\u300aVLsI: Verbalized Layers-to-Interactions from Large t [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,84,86,97,80],"class_list":["post-5249","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-nvidia","tag-llm","tag-distillation","tag-cv"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587\u300aVLsI: Verbalized Layers-to-Interactions from Large t&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5249","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5249"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5249\/revisions"}],"predecessor-version":[{"id":5251,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5249\/revisions\/5251"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5249"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5249"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5249"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}