{"id":668,"date":"2025-11-28T02:00:15","date_gmt":"2025-11-27T18:00:15","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=668"},"modified":"2025-11-28T09:32:57","modified_gmt":"2025-11-28T01:32:57","slug":"a-survey-of-generative-categories-and-techniques-in-multimodal-generative-models","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=668","title":{"rendered":"A Survey of Generative Categories and Techniques in Multimodal Generative Models"},"content":{"rendered":"\n

\u8bba\u6587\u76ee\u7684<\/h2>\n\n\n\n

\u7cfb\u7edf\u56de\u987e \u201c\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b (Multimodal Large Language Models, MLLMs)\u201d \u7684\u7814\u7a76\u8fdb\u5c55 \u2014 \u5373\u90a3\u4e9b\u4e0d\u4ec5\u5904\u7406\u6587\u672c (text)\uff0c\u8fd8\u80fd\u5904\u7406 \/ \u751f\u6210\u56fe\u50cf (image)\u3001\u97f3\u4e50 (music)\u3001\u89c6\u9891 (video)\u3001\u4eba\u4f53\u52a8\u4f5c (human motion)\u30013D \u5bf9\u8c61 (3D) \u7b49\u591a\u79cd\u6a21\u6001 (modalities) \u7684\u6a21\u578b\u3002<\/p>\n\n\n\n

\u8bba\u6587\u901a\u8fc7\u5206\u7c7b\u6a21\u6001 (modalities)\u3001\u603b\u7ed3\u57fa\u7840\u6280\u672f (foundation techniques)\u3001\u5206\u6790\u5178\u578b\u6a21\u578b \/ \u67b6\u6784 \/\u8bad\u7ec3 \/\u878d\u5408\u65b9\u5f0f\uff0c\u4ee5\u53ca\u8ba8\u8bba\u6311\u6218\u4e0e\u53d1\u5c55\u8d8b\u52bf\uff0c\u4e3a\u7814\u7a76\u8005\u63d0\u4f9b\u201c\u4ece\u6587\u672c LLM \u2192 \u901a\u7528\u591a\u6a21\u6001\u6a21\u578b\u201d\u7684\u603b\u4f53\u8def\u7ebf\u56fe (map) \u3002<\/p>\n\n\n\n

\"A<\/figure>\n\n\n\n

\u4e3a\u4ec0\u4e48\u5173\u6ce8 MLLM \/ \u8de8\u6a21\u6001\u63a8\u7406<\/p>\n\n\n\n

\u4f20\u7edf\u7684\u201c\u5927\u8bed\u8a00\u6a21\u578b (LLM)\u201d\u53ea\u5904\u7406\u6587\u672c\uff0c\u7f3a\u4e4f\u5bf9\u56fe\u50cf\u3001\u89c6\u9891\u3001\u8bed\u97f3\u30013D \u7b49\u5176\u4ed6\u6a21\u6001\u7684\u611f\u77e5\u548c\u63a8\u7406\u80fd\u529b\u3002<\/p>\n\n\n\n

\u8fd9\u516d\u7c7b\uff08\u89c1\u56fe 1\uff09\u4ee3\u8868\u4e86\u5f53\u524d\u751f\u6210\u6a21\u578b\u64cd\u4f5c\u7684\u4e3b\u8981\u6a21\u5f0f\uff0c\u6bcf\u4e00\u7c7b\u90fd\u6db5\u76d6\u4e86\u4e0d\u540c\u5f62\u5f0f\u7684\u6570\u636e\u8f93\u51fa\u548c\u72ec\u7279\u7684\u5e94\u7528\u573a\u666f\u3002<\/p>\n\n\n\n

\u591a\u6a21\u6001\u751f\u6210\u6a21\u578b\u7684\u516d\u5927\u7c7b\u522b<\/strong><\/strong><\/h3>\n\n\n\n

\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08MLLMs\uff09\u7684\u751f\u6210\u80fd\u529b\u5df2\u6269\u5c55\u5230\u6587\u672c\u3001\u56fe\u50cf\u3001\u97f3\u9891\u3001\u89c6\u9891\u7b49\u591a\u79cd\u6a21\u6001\u3002\u6839\u636e\u8f93\u5165\u548c\u8f93\u51fa\u6a21\u6001\u7684\u4e0d\u540c\uff0c\u53ef\u4ee5\u5c06\u5176\u4e3b\u8981\u751f\u6210\u7c7b\u522b\u5f52\u7eb3\u4e3a\u4ee5\u4e0b\u516d\u79cd<\/a>\uff1a<\/p>\n\n\n\n

\u751f\u6210\u7c7b\u522b<\/td>\u6838\u5fc3\u529f\u80fd\u4e0e\u7279\u70b9<\/td>\u5178\u578b\u5e94\u7528\u573a\u666f<\/td><\/tr>
\u6587\u672c\u751f\u6210\u6587\u672c (T2T)<\/strong><\/a><\/td>\u6240\u6709\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u7684\u57fa\u7840\uff0c\u652f\u6301\u4fe1\u606f\u68c0\u7d22\u3001\u6458\u8981\u3001\u7ffb\u8bd1\u548c\u5bf9\u8bdd\u7cfb\u7edf\u3002<\/td>\u667a\u80fd\u5ba2\u670d\u3001\u5185\u5bb9\u521b\u4f5c\u3001\u673a\u5668\u7ffb\u8bd1<\/td><\/tr>
\u6587\u672c\u751f\u6210\u56fe\u50cf (T2I)<\/strong><\/a><\/td>\u6839\u636e\u6587\u672c\u63cf\u8ff0\u751f\u6210\u89c6\u89c9\u5185\u5bb9\uff0c\u662f\u89c6\u89c9\u751f\u6210\u4efb\u52a1\u7684\u6838\u5fc3\u3002\u8fd1\u5e74\u6765\uff0c\u6269\u6563\u6a21\u578b (Diffusion Model)<\/strong> \u5df2\u6210\u4e3a\u8be5\u9886\u57df\u7684\u4e3b\u6d41\u6280\u672f<\/a>\u3002<\/td>\u827a\u672f\u521b\u4f5c\u3001\u8bbe\u8ba1\u8f85\u52a9\u3001\u6e38\u620f\u5f00\u53d1<\/td><\/tr>
\u6587\u672c\u751f\u6210\u89c6\u9891 (T2V)<\/strong><\/a><\/td>\u7ed3\u5408\u65f6\u95f4\u4e0e\u89c6\u89c9\u4fe1\u606f\u751f\u6210\u52a8\u6001\u573a\u666f\uff0c\u5176\u6a21\u578b\u9700\u8981\u5bf9\u73b0\u5b9e\u7269\u7406\u89c4\u5f8b\u6709\u4e00\u5b9a\u7406\u89e3\uff0c\u7c7b\u4f3c\u4e8e\u4e00\u4e2a”\u4e16\u754c\u6a21\u578b”\u3002<\/td>\u77ed\u89c6\u9891\u5236\u4f5c\u3001\u7535\u5f71\u9884\u6f14\u3001\u5e7f\u544a\u521b\u610f<\/td><\/tr>
\u6587\u672c\u751f\u6210\u4eba\u7c7b\u52a8\u4f5c (T2HM)<\/strong><\/a><\/td>\u6839\u636e\u6587\u672c\u6307\u4ee4\u751f\u6210\u4eba\u4f53\u52a8\u4f5c\u5e8f\u5217\uff0c\u662f\u5b9e\u73b0\u5177\u8eab\u667a\u80fd<\/strong>\uff08\u5982\u673a\u5668\u4eba\u63a7\u5236\uff09\u548c\u865a\u62df\u4eba\u4ea4\u4e92\u7684\u5173\u952e<\/a>\u3002<\/td>\u52a8\u753b\u5236\u4f5c\u3001\u673a\u5668\u4eba\u6307\u4ee4\u3001\u865a\u62df\u5076\u50cf<\/td><\/tr>
\u6587\u672c\u751f\u62103D\u7269\u4f53 (T2-3D)<\/strong><\/a><\/td>\u6839\u636e\u6587\u672c\u751f\u6210\u4e09\u7ef4\u6a21\u578b\uff0c\u5bf9\u865a\u62df\u73b0\u5b9e\u3001\u6e38\u620f\u548c\u5de5\u4e1a\u8bbe\u8ba1\u7b49\u6c89\u6d78\u5f0f\u5e94\u7528\u81f3\u5173\u91cd\u8981\u3002<\/td>\u865a\u62df\u73b0\u5b9e\u573a\u666f\u6784\u5efa\u3001\u4ea7\u54c1\u539f\u578b\u8bbe\u8ba1<\/td><\/tr>
\u6587\u672c\u751f\u6210\u97f3\u4e50 (T2M)<\/strong><\/a><\/td>\u97f3\u4e50\u5305\u542b\u590d\u6742\u7684\u4e50\u5668\u3001\u8282\u594f\u4e0e\u60c5\u611f\uff0c\u5efa\u6a21\u96be\u5ea6\u9ad8\u4e8e\u4e00\u822c\u8bed\u97f3\u3002\u6b64\u7c7b\u522b\u4e13\u6ce8\u4e8e\u751f\u6210\u5177\u6709\u97f3\u4e50\u6027\u7684\u97f3\u9891\u5185\u5bb9\u3002<\/td>\u914d\u4e50\u751f\u6210\u3001\u97f3\u4e50\u521b\u4f5c\u8f85\u52a9<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

MLLM\u7684\u6838\u5fc3\u7ec4\u6210\u4e0e\u6280\u672f\u673a\u5236<\/h2>\n\n\n\n

MLLM \u901a\u8fc7\u628a\u8bed\u8a00\u6a21\u578b\u4e0e\u89c6\u89c9 \/ \u5176\u4ed6\u6a21\u6001\u8f93\u5165\u7ed3\u5408\u8d77\u6765\uff0c\u80fd\u591f\u6267\u884c\u8bf8\u5982 \u201c\u6839\u636e\u56fe\u7247\u5199\u6545\u4e8b\u201d\u3001 \u201c\u4ece\u56fe\u50cf\u76f4\u63a5\u7b97\u6570 (OCR free math reasoning)\u201d\u3001 \u201c\u89c6\u89c9\u95ee\u7b54 (VQA)\u201d\u3001 \u201c\u591a\u6a21\u6001\u6307\u4ee4\u6267\u884c (vision + text \u2192 action)\u201d \u7b49\u4efb\u52a1\uff0c\u8fd9\u4e9b\u80fd\u529b\u662f\u4f20\u7edf unimodal \u6a21\u578b\u96be\u4ee5\u5b9e\u73b0\u7684\u3002<\/p>\n\n\n\n

\u56e0\u4e3a\u7ee7\u627f\u4e86 LLM \u7684\u63a8\u7406\u548c\u751f\u6210\u80fd\u529b + \u65b0\u589e\u7684\u8de8\u6a21\u6001\u7406\u89e3\u80fd\u529b\uff0cMLLM \u88ab\u89c6\u4e3a\u901a\u5411\u66f4\u901a\u7528\u3001\u66f4\u7075\u6d3b AI \u7cfb\u7edf\u7684\u91cd\u8981\u8def\u5f84\u4e4b\u4e00\u3002<\/p>\n\n\n\n

\"A<\/figure>\n\n\n\n

\u67b6\u6784\u7ec4\u6210<\/h2>\n\n\n\n

\u6839\u636e survey\uff0c\u603b\u4f53\u6765\u8bf4\uff0c\u4e00\u4e2a\u5178\u578b\u7684 MLLM \u5305\u542b\u4ee5\u4e0b\u4e3b\u8981\u7ec4\u6210\u90e8\u5206 (\u7ec4\u4ef6) \uff1a<\/p>\n\n\n\n

\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) backbone \u2014 \u7528\u4f5c\u201c\u8111 (brain)\u201d\u6216\u201c\u901a\u7528\u7406\u89e3 \/\u751f\u6210 \/\u63a8\u7406\u6a21\u5757 (core)\u201d<\/p>\n\n\n\n

\u6a21\u6001\u7f16\u7801\u5668 (Modality encoder) \u2014 \u5c06\u975e\u8bed\u8a00\u6a21\u6001 (\u5982\u56fe\u50cf\u3001\u97f3\u9891\u3001\u89c6\u9891\u30013D\u3001\u52a8\u4f5c\u7b49) \u8f6c\u6362\u4e3a feature embeddings\uff0c\u4f7f\u5176\u80fd\u591f\u88ab LLM \u201c\u7406\u89e3 \/\u5904\u7406 \/\u878d\u5408\u201d\u3002<\/p>\n\n\n\n

\u8fde\u63a5 \/ \u9002\u914d\u6a21\u5757 (connector \/ adapter \/ vision-to-language adapter \/ modality adapter) \u2014 \u7528\u4e8e\u5c06 Encoder \u8f93\u51fa (\u975e\u8bed\u8a00\u7279\u5f81) \u8f6c\u6362 \/\u5bf9\u9f50 (alignment) \u5230 LLM \u7684 embedding \/ token \u7a7a\u95f4\uff0c\u4f7f LLM \u80fd\u591f\u8de8\u6a21\u6001\u63a8\u7406 \/\u751f\u6210 \/\u878d\u5408\u3002\u901a\u5e38\u6709\u6295\u5f71 (projection-based)\u3001\u67e5\u8be2 (query-based)\u3001\u878d\u5408 (fusion-based) \u7b49\u4e0d\u540c\u8fde\u63a5\u65b9\u5f0f\u3002<\/p>\n\n\n\n

(\u53ef\u9009) \u751f\u6210\u6a21\u5757 \/ decoding module \u2014 \u5f53\u9700\u8981\u751f\u6210\u975e\u6587\u672c\u6a21\u6001 (\u4f8b\u5982\u56fe\u50cf\u3001\u89c6\u9891\u30013D\u3001\u52a8\u4f5c\u7b49) \u8f93\u51fa\u65f6\uff0c\u53ef\u80fd\u989d\u5916\u9644\u5e26\u4e00\u4e2a generator \/ decoder\uff0c\u7528\u4e8e\u5c06 LLM + multimodal embedding \u8f6c\u6362\u4e3a\u76ee\u6807\u6a21\u6001\u8f93\u51fa\u3002<\/p>\n\n\n\n

\u8fd9\u4e2a\u67b6\u6784\u8bbe\u8ba1\u4f7f\u5f97 MLLM \u5728\u591a\u6a21\u6001\u4efb\u52a1\u4e0a\u80fd\u591f\u7edf\u4e00\u4f7f\u7528\u4e00\u4e2a\u201c\u6838\u5fc3 + \u611f\u77e5 + \u9002\u914d + \u751f\u6210\u201d\u7684 pipeline\u3002<\/p>\n\n\n\n

\u901a\u7528\u6280\u672f (Foundation Techniques)<\/h2>\n\n\n\n

\u4e3a\u4e86\u5b9e\u73b0\u8de8\u6a21\u6001\u80fd\u529b\u4e0e\u901a\u7528\u6027\uff0c\u8bba\u6587\u603b\u7ed3\u4e86\u5f53\u524d MLLM \u4e3b\u8981\u4f9d\u8d56\u7684\u51e0\u7c7b\u57fa\u7840\u65b9\u6cd5 \/\u6280\u672f\uff1a<\/p>\n\n\n\n

\u81ea\u76d1\u7763\u5b66\u4e60 (Self-Supervised Learning, SSL) \u2014 \u7528\u4e8e\u9884\u8bad\u7ec3\u6a21\u6001\u7f16\u7801\u5668 \/ \u7279\u5f81\u63d0\u53d6\u5668\uff0c\u4f7f\u5176\u80fd\u591f\u4ece\u5927\u91cf\u672a\u6807\u6ce8 \/\u5f31\u6807\u6ce8\u7684\u6570\u636e\u4e2d\u5b66\u4e60\u901a\u7528\u8868\u793a (representation)\u3002<\/p>\n\n\n\n

\u4e13\u5bb6\u6df7\u5408 (Mixture of Experts, MoE) \u2014 \u5728\u591a\u6a21\u6001 + \u591a\u4efb\u52a1 + \u5927\u89c4\u6a21\u6a21\u578b\u573a\u666f\u4e0b\uff0c\u901a\u8fc7\u4e13\u5bb6 (experts) \u5206\u652f \/\u6a21\u5757\u5316\u8bbe\u8ba1\uff0c\u4f7f\u6a21\u578b\u5177\u6709\u826f\u597d\u6269\u5c55\u6027\u4e0e\u4e13\u95e8\u5316\u80fd\u529b (\u67d0\u4e9b\u4e13\u5bb6\u8d1f\u8d23\u67d0\u4e9b\u6a21\u6001 \/\u4efb\u52a1)\uff0c\u540c\u65f6\u63a7\u5236\u8ba1\u7b97 \/\u53c2\u6570\u590d\u6742\u5ea6\u3002<\/p>\n\n\n\n

\u4ece\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60 (Reinforcement Learning from Human Feedback, RLHF) \u2014 \u7528\u4e8e\u5fae\u8c03 (fine-tune) \u6a21\u578b\uff0c\u4f7f\u5176\u8f93\u51fa\u7b26\u5408\u4eba\u7c7b\u504f\u597d \/\u5b89\u5168 \/\u8d28\u91cf \/\u4e00\u81f4\u6027 \/\u53ef\u63a7\u6027 (\u5c24\u5176\u591a\u6a21\u6001\u751f\u6210 \/\u63a8\u7406\u8f93\u51fa\u7684\u5bf9\u9f50\u4e0e\u5b89\u5168\u6027)\u3002<\/p>\n\n\n\n

\u63d0\u793a + \u63a8\u7406\u94fe (Prompting & Chain-of-Thought, CoT) \u2014 \u6269\u5c55\u5230\u591a\u6a21\u6001\u573a\u666f (\u5982\u89c6\u89c9 + \u6587\u672c + \u63a8\u7406)\uff0c\u8ba9\u6a21\u578b\u901a\u8fc7\u201c\u591a\u6a21\u6001 CoT \/ multimodal CoT \/ M-CoT\u201d\u3001\u201c\u591a\u6a21\u6001 in-context learning (M-ICL)\u201d\u7b49\u65b9\u5f0f\uff0c\u589e\u5f3a\u8de8\u6a21\u6001\u63a8\u7406\u80fd\u529b\u4e0e\u901a\u7528\u6027 (few-shot \/ zero-shot \/ instruction-following) \u3002<\/p>\n\n\n\n

\u8fd9\u4e9b\u57fa\u7840\u6280\u672f \/\u65b9\u6cd5\u6784\u6210\u4e86 MLLM \u4ece\u5355\u6a21\u6001 \u2192 \u591a\u6a21\u6001\u3001\u4ece\u8bc6\u522b \/\u5206\u7c7b \u2192 \u751f\u6210 \/\u63a8\u7406 \/\u901a\u7528\u80fd\u529b (generalist) \u7684\u57fa\u77f3\u3002<\/p>\n\n\n\n

\u6280\u672f\u7684\u534f\u540c\u7ec4\u5408\uff1a\u5c3d\u7ba1\u6bcf\u79cd\u6280\u672f\u5404\u81ea\u5177\u6709\u72ec\u7279\u7684\u4f18\u52bf\uff0c\u5982\u6570\u636e\u6548\u7387\u3001\u6a21\u5757\u5316\u53ef\u6269\u5c55\u6027\u3001\u4e00\u81f4\u6027\u6216\u7ed3\u6784\u5316\u63a8\u7406\uff0c\u4f46\u5b83\u4eec\u6700\u5927\u7684\u6f5c\u529b\u5728\u4e8e\u7ec4\u5408\u4f7f\u7528\u3002\u6574\u5408\u81ea\u76d1\u7763\u5b66\u4e60\uff08SSL\uff09\u3001\u4e13\u5bb6\u6a21\u578b\uff08MoE\uff09\u3001\u5f3a\u5316\u5b66\u4e60\u4e0e\u4eba\u7c7b\u53cd\u9988\uff08RLHF\uff09\u4ee5\u53ca\u94fe\u5f0f\u63a8\u7406\uff08CoT\uff09\u53ef\u4ee5\u57f9\u517b\u51fa\u4e0d\u4ec5\u66f4\u8fde\u8d2f\u3001\u66f4\u53ef\u63a7\uff0c\u800c\u4e14\u5728\u8de8\u6a21\u6001\u548c\u8de8\u4efb\u52a1\u7684\u6cdb\u5316\u80fd\u529b\u4e0a\u66f4\u5f3a\u7684\u6a21\u578b\u3002\u8fd9\u79cd\u6c47\u805a\u4f7f\u591a\u6a21\u6001\u751f\u6210\u6a21\u578b\uff08MGM\uff09\u80fd\u591f\u8fdb\u884c\u66f4\u6df1\u5165\u7684\u63a8\u7406\u3001\u4ee5\u66f4\u7ed3\u6784\u5316\u7684\u65b9\u5f0f\u8fdb\u884c\u89c4\u5212\uff0c\u5e76\u4ee5\u66f4\u7cbe\u7ec6\u7684\u7c92\u5ea6\u54cd\u5e94\uff0c\u4ece\u800c\u4e3a\u5b9e\u73b0\u66f4\u53ef\u89e3\u91ca\u548c\u81ea\u9002\u5e94\u7684\u591a\u6a21\u6001\u667a\u80fd\u94fa\u5e73\u9053\u8def\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6bcf\u79cd\u6280\u672f\u5728\u4e0d\u540c\u6a21\u6001\u4e0a\u7684\u53d1\u5c55\u548c\u91c7\u7528\u7a0b\u5ea6\u5dee\u5f02\u8f83\u5927\uff0c\u8fd9\u53cd\u6620\u4e86\u6570\u636e\u53ef\u7528\u6027\u3001\u67b6\u6784\u9650\u5236\u548c\u4efb\u52a1\u7279\u5b9a\u6311\u6218\u7684\u4e0d\u540c\uff0c\u5982\u8868 2 \u6240\u603b\u7ed3\u3002<\/p>\n\n\n\n

\"A<\/figure>\n\n\n\n

\u6a21\u578b\u80fd\u529b & \u5e94\u7528 \/ Emergent \u529f\u80fd<\/h2>\n\n\n\n

\u5f97\u76ca\u4e8e\u4e0a\u8ff0\u67b6\u6784\u8bbe\u8ba1 + \u6280\u672f \/\u8bad\u7ec3\u65b9\u6cd5\uff0cMLLM \u5c55\u73b0\u51fa\u4f20\u7edf \u201c\u6587\u672c\u6a21\u578b + \u5355\u6a21\u6001 (\u56fe\u50cf\u5206\u7c7b \/ caption \/ \u68c0\u7d22)\u201d \u65e0\u6cd5\u8f7b\u6613\u5b9e\u73b0\u7684\u529f\u80fd \/ emergent \u80fd\u529b\uff0c\u5305\u62ec\uff1a<\/p>\n\n\n\n

\u56fe\u7247 \u2192 \u6545\u4e8b \/ \u63cf\u8ff0 \/\u81ea\u7136\u8bed\u8a00\u751f\u6210\uff1a\u4f8b\u5982\u7ed9\u5b9a\u56fe\u50cf (\u6216\u591a\u56fe\u50cf)\uff0c\u751f\u6210\u8fde\u8d2f\u3001\u5bcc\u8bed\u4e49\u3001\u5177\u4e0a\u4e0b\u6587 \/\u60c5\u8282 \/\u903b\u8f91\u7684\u6587\u672c (\u6545\u4e8b\u3001\u63cf\u8ff0\u3001\u89e3\u91ca\u7b49)\u3002\u8fd9\u6bd4\u4f20\u7edf captioning \u66f4\u7075\u6d3b \/\u4e30\u5bcc\u3002<\/p>\n\n\n\n

\u56fe\u50cf + \u6587\u672c \u2192 \u8de8\u6a21\u6001\u63a8\u7406 \/\u56de\u7b54 (Visual Question Answering, VQA)\uff1a\u7ed3\u5408\u89c6\u89c9 + \u8bed\u8a00\u7406\u89e3 + \u63a8\u7406 (\u903b\u8f91\u3001\u5e38\u8bc6\u3001\u4e0a\u4e0b\u6587) \u6765\u56de\u7b54\u4e0e\u56fe\u50cf\u5185\u5bb9\u76f8\u5173\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n

\u65e0\u9700 OCR \u7684\u56fe\u50cf\u6570\u5b66 \/\u63a8\u7406\u80fd\u529b (OCR-free math reasoning)\uff1a\u8fd9\u610f\u5473\u7740\u6a21\u578b\u53ef\u4ee5\u76f4\u63a5\u201c\u770b\u5230\u201d\u56fe\u50cf (\u6bd4\u5982\u56fe\u8868 \/ \u6570\u5b66\u9898 \/\u624b\u5199\u5b57 \/\u516c\u5f0f)\uff0c\u7406\u89e3\u5185\u5bb9\u5e76\u8fdb\u884c\u63a8\u7406 \/\u8ba1\u7b97\uff0c\u800c\u4e0d\u5fc5\u987b\u5148\u901a\u8fc7\u6587\u5b57\u63d0\u53d6 \/ OCR\u3002<\/p>\n\n\n\n

\u591a\u6a21\u6001\u751f\u6210 (Cross-modal generation)\uff1a\u4e0d\u4ec5\u4ec5\u662f\u56fe\u50cf \u2192 \u6587\u672c\uff0c\u8fd8\u53ef\u80fd\u662f\u6587\u672c \u2192 \u56fe\u50cf \/\u89c6\u9891 \/3D \/\u52a8\u4f5c \/\u97f3\u9891\u7b49 (\u5f53\u6a21\u578b\u96c6\u6210 generator \/ decoder \u65f6)\uff0c\u8ba9\u6a21\u578b\u5177\u5907\u591a\u6837\u7684\u611f\u77e5 + \u751f\u6210\u80fd\u529b\uff0c\u5b9e\u73b0\u771f\u6b63\u7684\u201c\u901a\u7528\u591a\u6a21\u6001\u5927\u6a21\u578b (generalist multimodal model)\u201d\u3002 \u8fd9\u662f\u8be5 survey \u6240\u5f3a\u8c03\u7684 MLLM \u53d1\u5c55\u65b9\u5411\u4e4b\u4e00\u3002<\/p>\n\n\n\n

\u7efc\u4e0a\uff0cMLLM \u7684\u80fd\u529b\u5df2\u8d85\u51fa\u4e86\u4f20\u7edf vision-language \/ multimodal classification \/ retrieval \/ captioning \u7684\u8303\u7574\uff0c\u66f4\u63a5\u8fd1\u4e8e\u4e00\u79cd\u201c\u901a\u7528\u667a\u80fd (generalist intelligence)\u201d\u7cfb\u7edf\u3002<\/p>\n\n\n\n

 \u6311\u6218\u3001\u95ee\u9898\u4e0e\u5f00\u653e\u65b9\u5411<\/h2>\n\n\n\n

\u8bba\u6587\u4e5f\u5bf9 MLLM \u5f53\u524d \/\u672a\u6765\u53d1\u5c55\u6240\u9762\u4e34\u7684\u95ee\u9898\u8fdb\u884c\u4e86\u5206\u6790\u4e0e\u603b\u7ed3\u3002\u4e3b\u8981\u5305\u62ec\uff1a<\/p>\n\n\n\n

\u591a\u6a21\u6001 \u201c\u5e7b\u89c9 (hallucination)\u201d \u2014 \u5f53\u6a21\u578b\u5728\u89c6\u89c9 + \u8bed\u8a00\u878d\u5408 \/\u751f\u6210 \/\u63a8\u7406\u65f6\uff0c\u6709\u53ef\u80fd \u201c\u60f3\u8c61 \/\u7f16\u9020 (hallucinate)\u201d \u51fa\u4e0e\u8f93\u5165\u6a21\u6001 (\u4f8b\u5982\u56fe\u7247) \u4e0d\u7b26\u7684\u5185\u5bb9 (objects \/\u7ec6\u8282 \/\u903b\u8f91)\uff0c\u5bfc\u81f4\u8f93\u51fa\u4e0d\u53ef\u9760 \/\u4e0d\u771f\u5b9e\u3002<\/p>\n\n\n\n

\u8bc4\u4f30 (Evaluation) \u7684\u56f0\u96be \u2014 \u4f20\u7edf\u5bf9\u89c6\u89c9\u8bed\u8a00\u6a21\u578b (vision-language models) \u7684\u8bc4\u4f30 (\u5982 captioning\u3001VQA) \u4efb\u52a1 \/ benchmark \u4e0d\u8db3\u4ee5\u8986\u76d6 MLLM \u7684\u201c\u8de8\u6a21\u6001 + \u63a8\u7406 + \u751f\u6210 + \u590d\u6742\u8f93\u51fa + \u591a\u4efb\u52a1 \/\u591a\u573a\u666f\u201d\u80fd\u529b\u3002\u9700\u8981\u65b0\u7684\u3001\u66f4\u901a\u7528 \/ \u7efc\u5408 \/\u7cfb\u7edf\u5316\u7684 benchmark \u548c\u8bc4\u6d4b\u6307\u6807\u3002<\/p>\n\n\n\n

\u6a21\u6001 \/\u4efb\u52a1 \/\u751f\u6210\u7c7b\u578b\u7684\u6269\u5c55 & \u901a\u7528\u6027 (Modality \/ Task \/ Output Generalization) \u96be\u9898 \u2014 \u867d\u7136 MLLM \u652f\u6301\u591a\u6a21\u6001\uff0c\u4f46\u4ece\u201c\u89c6\u89c9 + \u6587\u672c\u201d\u6269\u5c55\u5230\u201c\u89c6\u9891 \/ \u52a8\u4f5c \/3D \/\u97f3\u9891 \/\u97f3\u4e50 \/\u590d\u6742\u751f\u6210 \/\u8de8\u4efb\u52a1\u201d\u7b49\u65f6\uff0c\u8bbe\u8ba1\u3001\u8bad\u7ec3\u3001\u5bf9\u9f50 (alignment)\u3001\u6570\u636e\u652f\u6301\u3001\u751f\u6210\u5668 \/ decoder \u7b49\u90fd\u5f88\u590d\u6742\uff0c\u76ee\u524d\u8fd8\u7f3a\u4e4f\u6210\u719f\u7edf\u4e00\u6846\u67b6 \/\u7cfb\u7edf\u3002<\/p>\n\n\n\n

\u6548\u7387 \/\u8d44\u6e90 \/\u90e8\u7f72\u95ee\u9898 (Cost & Deployment) \u2014 \u591a\u6a21\u6001 + \u5927\u6a21\u578b + \u591a\u4efb\u52a1 \/ \u591a\u8f93\u51fa = \u9ad8\u7b97\u529b\u3001\u6d77\u91cf\u6570\u636e\u9700\u6c42\u3002\u8bad\u7ec3\u6210\u672c\u9ad8\uff0c\u63a8\u7406 \/\u90e8\u7f72 (\u5c24\u5176\u5728\u8d44\u6e90\u53d7\u9650\u73af\u5883 \/ edge \/ mobile) \u96be\u5ea6\u5927\u3002\u8bba\u6587\u8ba4\u4e3a\u7814\u7a76 \u201c\u9ad8\u6548 \/\u8f7b\u91cf \/\u6a21\u5757\u5316 \/\u53ef\u6269\u5c55 \/\u8d44\u6e90\u53cb\u597d\u201d \u7684 MLLM \u662f\u4e00\u4e2a\u91cd\u8981\u65b9\u5411\u3002<\/p>\n\n\n\n

\u53ef\u63a7\u6027 \/\u5bf9\u9f50 \/\u5b89\u5168 \/\u9c81\u68d2\u6027 \/\u53ef\u89e3\u91ca\u6027 (Alignment, Safety, Interpretability, Robustness) \u2014 \u968f\u7740\u6a21\u578b\u80fd\u529b\u8d8a\u6765\u8d8a\u6cdb\u3001\u8f93\u51fa\u8d8a\u6765\u8d8a\u591a\u6837 (\u6587\u672c \/\u56fe\u50cf \/3D \/\u52a8\u4f5c\u7b49)\uff0c\u5982\u4f55\u4fdd\u8bc1\u6a21\u578b\u8f93\u51fa\u5b89\u5168 \/\u53ef\u9760 \/\u7b26\u5408\u4f26\u7406 \/\u6ca1\u6709\u504f\u5dee \/\u5bf9\u6297 \/\u9690\u79c1 \/\u8bef\u7528\u98ce\u9669\uff0c\u662f\u91cd\u5927\u6311\u6218\u3002\u5c24\u5176\u5f53\u8f93\u51fa\u5305\u62ec\u56fe\u50cf \/\u89c6\u9891 \/3D \/\u52a8\u4f5c \/\u5176\u4ed6\u6a21\u6001\u65f6\u3002 \u8bba\u6587\u5efa\u8bae\u672a\u6765\u7814\u7a76\u5bf9\u6b64\u7ed9\u4e88\u66f4\u591a\u5173\u6ce8\u3002<\/p>\n\n\n\n

\u672a\u6765\u65b9\u5411 & \u4f5c\u8005\u5efa\u8bae<\/h2>\n\n\n\n

\u57fa\u4e8e\u5bf9\u73b0\u72b6\u4e0e\u6311\u6218\u7684\u5206\u6790\uff0c\u8bba\u6587\u63d0\u51fa\u4e86\u82e5\u5e72\u672a\u6765\u7814\u7a76 \/\u53d1\u5c55\u5efa\u8bae \/\u53ef\u80fd\u65b9\u5411 (critical paths toward general-purpose multimodal systems) \uff1a<\/p>\n\n\n\n

\u6269\u5c55\u65b0\u7684\u6a21\u6001 & \u8f93\u51fa\u7c7b\u578b \u2014 \u9664\u4e86\u56fe\u50cf + \u6587\u672c + \u89c6\u9891 + 3D + \u52a8\u4f5c\uff0c\u8fd8\u53ef\u80fd\u52a0\u5165 \u97f3\u9891 \/\u97f3\u4e50 \/\u8bed\u97f3 \/\u591a\u8bed\u79cd \/\u8de8\u6587\u5316 \/\u8de8\u9886\u57df \u6a21\u6001\u3002\u6784\u5efa\u66f4\u901a\u7528\u3001\u66f4\u7075\u6d3b\u3001\u66f4\u5f3a\u5927\u7684\u591a\u6a21\u6001\u7cfb\u7edf\u3002<\/p>\n\n\n\n

\u6784\u5efa\u66f4\u7cfb\u7edf \/\u7efc\u5408 \/\u6807\u51c6\u5316\u7684\u8bc4\u4f30 benchmark \u2014 \u8986\u76d6\u591a\u6a21\u6001\u7406\u89e3 (perception)\u3001\u63a8\u7406 (reasoning)\u3001\u751f\u6210 (generation)\u3001\u591a\u4efb\u52a1 (\u591a\u6a21\u6001 +\u591a\u4efb\u52a1)\u3001\u901a\u7528\u6027 (generalization)\u3001\u5b89\u5168 \/\u5065\u58ee\u6027 \/\u5bf9\u9f50 \/\u53ef\u63a7\u6027 \u7b49\u7ef4\u5ea6\u3002\u89e3\u51b3\u5f53\u524d benchmark \/\u8bc4\u4f30\u4f53\u7cfb\u788e\u7247\u5316 \/\u4e0d\u5168\u9762\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n

\u63d0\u5347\u6a21\u578b\u6548\u7387 \/\u8f7b\u91cf\u5316 \/\u8d44\u6e90\u53cb\u597d\u578b\u8bbe\u8ba1 \u2014 \u4f8b\u5982\u6a21\u5757\u5316\u3001\u4e13\u5bb6\u6df7\u5408 (MoE)\u3001\u77e5\u8bc6\u84b8\u998f (distillation)\u3001\u91cf\u5316 (quantization)\u3001\u526a\u679d (pruning)\u3001efficient architecture \/ adapter \/ connector \u8bbe\u8ba1\u7b49\uff0c\u4ee5\u4fbf\u66f4\u597d\u5730\u90e8\u7f72\u5230\u8fb9\u7f18 \/\u79fb\u52a8\u8bbe\u5907 \/\u8d44\u6e90\u53d7\u9650\u73af\u5883\u3002<\/p>\n\n\n\n

\u52a0\u5f3a\u5bf9\u9f50 \/\u5b89\u5168 \/\u53ef\u63a7 \/\u53ef\u89e3\u91ca \/\u53ef\u4fe1 (alignment \/ robustness \/ interpretability \/ trustworthiness) \u2014 \u5305\u62ec\u51cf\u5c11 hallucination \/\u9519\u8bef \/\u504f\u5dee \/\u4e0d\u5b89\u5168\u8f93\u51fa\uff1b\u589e\u5f3a\u8f93\u51fa\u7684\u4e00\u81f4\u6027\u3001\u53ef\u63a7\u6027\u3001\u53ef\u89e3\u91ca\u6027\uff1b\u8bbe\u8ba1\u4eba\u7c7b\u53cd\u9988 \/\u89c4\u5219 \/\u9650\u5236 \/\u5bf9\u9f50\u673a\u5236\u7b49\u3002<\/p>\n\n\n\n

\u4fc3\u8fdb\u8de8\u6a21\u6001 \/\u8de8\u4efb\u52a1 \/\u8de8\u573a\u666f \/\u8de8\u9886\u57df\u901a\u7528\u6027 (Generalist, cross-domain \/ cross-modality \/ cross-task) \u2014 \u9f13\u52b1\u7814\u7a76\u4e0d\u4ec5\u4e13\u6ce8\u4e8e\u56fe\u50cf+\u6587\u672c\uff0c\u8fd8\u517c\u987e\u89c6\u9891 \/ \u52a8\u4f5c \/3D \/\u97f3\u9891 \/\u8de8\u8bed\u79cd \/\u8de8\u6587\u5316 \/\u5b9e\u9645\u5e94\u7528 (\u533b\u5b66\u3001\u79d1\u5b66\u3001\u5de5\u7a0b\u3001\u793e\u4f1a \u2026) \u7b49\u66f4\u591a\u573a\u666f\u3002<\/p>\n\n\n\n

\u80cc\u9aa8 (backbone) \u901a\u5e38\u662f\u5927\u578b\u8bed\u8a00\u6a21\u578b (LLM) \u2014\u2014 \u8d1f\u8d23\u8bed\u8a00\u7406\u89e3\u4e0e\u751f\u6210\u3002<\/p>\n\n\n\n

\u589e\u52a0\u201c\u89c6\u89c9 (\u6216\u591a\u6a21\u6001) \u7f16\u7801\u5668 (vision \/ modality encoder)\u201d \u2014 \u5c06\u56fe\u50cf (\u6216\u89c6\u9891\uff0f\u97f3\u9891\uff0f3D\uff0f\u5176\u4ed6\u6a21\u6001) \u8f6c\u6362\u6210\u4e00\u7ec4 feature embeddings\uff0c\u4ee5\u4fbf\u4e0e\u8bed\u8a00 token \u5728\u540c\u4e00\u7a7a\u95f4 \/ \u5e8f\u5217\u4e2d\u4ea4\u4e92\u3002<\/p>\n\n\n\n

\u63a5\u7740\u662f\u201c\u8de8\u6a21\u6001\u878d\u5408 (cross-modal fusion \/ alignment)\u201d\u6a21\u5757 \u2014 \u7528\u4e8e\u5c06\u4e0d\u540c\u6a21\u6001 (\u89c6\u89c9 + \u6587\u672c + \u2026) \u7684\u8868\u793a\u6574\u5408\u8d77\u6765\uff0c\u4f7f\u6a21\u578b\u80fd\u591f\u201c\u7406\u89e3\u201d\u590d\u5408\u8f93\u5165 (\u4f8b\u5982 \u201c\u56fe\u7247 + \u63d0\u95ee\u201d -> \u201c\u56de\u7b54\u201d)\u3002<\/p>\n\n\n\n

\u6709\u4e9b\u7cfb\u7edf\u91c7\u7528\u6a21\u5757\u5316 \/ \u53ef\u6269\u5c55\u8bbe\u8ba1 (modular \/ mixture-of-experts \u7ed3\u6784)\uff0c\u4ee5\u652f\u6301\u591a\u79cd\u6a21\u6001\uff0f\u591a\u4efb\u52a1\uff0f\u591a\u8bed\u8a00\u3002<\/p>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

\u8bba\u6587\u76ee\u7684 \u7cfb\u7edf\u56de\u987e \u201c\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b (Multimodal Large Language Models, MLLMs)\u201d \u7684\u7814\u7a76\u8fdb\u5c55 \u2014 \u5373\u90a3\u4e9b\u4e0d\u4ec5\u5904\u7406\u6587\u672c (text)\uff0c\u8fd8\u80fd\u5904\u7406 \/ \u751f\u6210\u56fe\u50cf (image)\u3001\u97f3\u4e50 (music)\u3001\u89c6\u9891 (video)\u3001\u4eba\u4f53\u52a8\u4f5c (human motion)\u30013D \u5bf9\u8c61 (3D) \u7b49\u591a\u79cd\u6a21\u6001 (modalities) \u7684\u6a21\u578b\u3002 \u8bba\u6587\u901a\u8fc7\u5206\u7c7b\u6a21\u6001 (modalities)\u3001\u603b\u7ed3\u57fa\u7840\u6280\u672f (foundation techniques)\u3001\u5206\u6790 … \u7ee7\u7eed\u9605\u8bfb →<\/span><\/a><\/p>\n","protected":false},"author":2,"featured_media":670,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[],"class_list":["post-668","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-rengongzhineng"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/668","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=668"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/668\/revisions"}],"predecessor-version":[{"id":672,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/668\/revisions\/672"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/670"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=668"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=668"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=668"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}