{"id":262,"date":"2025-09-04T12:00:11","date_gmt":"2025-09-04T04:00:11","guid":{"rendered":"http:\/\/159.223.199.105\/?p=262"},"modified":"2025-09-15T12:02:34","modified_gmt":"2025-09-15T04:02:34","slug":"%e5%a4%9a%e6%a8%a1%e6%80%81%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%8c%81%e7%bb%ad%e5%ad%a6%e4%b9%a0%e7%b3%bb%e5%88%97%e7%a0%94%e7%a9%b6%ef%bc%8c%e7%bb%bc%e8%bf%b0benchmark%e6%96%b9%e6%b3%95codebase%e4%b8%80","status":"publish","type":"post","link":"http:\/\/www.applicationsofllm.com\/index.php\/2025\/09\/04\/%e5%a4%9a%e6%a8%a1%e6%80%81%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%8c%81%e7%bb%ad%e5%ad%a6%e4%b9%a0%e7%b3%bb%e5%88%97%e7%a0%94%e7%a9%b6%ef%bc%8c%e7%bb%bc%e8%bf%b0benchmark%e6%96%b9%e6%b3%95codebase%e4%b8%80\/","title":{"rendered":"\u591a\u6a21\u6001\u5927\u6a21\u578b\u6301\u7eed\u5b66\u4e60\u7cfb\u5217\u7814\u7a76\uff0c\u7efc\u8ff0+Benchmark+\u65b9\u6cd5+Codebase\u4e00\u7f51\u6253\u5c3d"},"content":{"rendered":"\n

\u8fd1\u5e74\u6765\uff0c\u751f\u6210\u5f0f AI \u548c\u591a\u6a21\u6001\u5927\u6a21\u578b\u5728\u5404\u9886\u57df\u53d6\u5f97\u4e86\u4ee4\u4eba\u77a9\u76ee\u7684\u8fdb\u5c55\u3002\u7136\u800c\uff0c\u5728\u73b0\u5b9e\u4e16\u754c\u5e94\u7528\u4e2d\uff0c\u52a8\u6001\u73af\u5883\u4e0b\u7684\u6570\u636e\u5206\u5e03\u548c\u4efb\u52a1\u9700\u6c42\u4e0d\u65ad\u53d8\u5316\uff0c\u5927\u6a21\u578b\u5982\u4f55\u5728\u6b64\u80cc\u666f\u4e0b\u5b9e\u73b0\u6301\u7eed\u5b66\u4e60\u6210\u4e3a\u4e86\u91cd\u8981\u6311\u6218\u3002\u4e3a\u4e86\u5e94\u5bf9\u8fd9\u4e00\u95ee\u9898\uff0c\u4e2d\u56fd\u79d1\u5b66\u9662\u81ea\u52a8\u5316\u7814\u7a76\u6240\u8054\u5408\u4e2d\u56fd\u79d1\u5b66\u9662\u9999\u6e2f\u9662 AI \u4e2d\u5fc3<\/strong>\u7cfb\u7edf\u6027\u5730\u7814\u7a76\u4e86\u751f\u6210\u5f0f AI \u548c\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u6301\u7eed\u5b66\u4e60\uff0c\u63d0\u51fa\u4e86\u4e00\u7cfb\u5217\u7efc\u8ff0\u3001\u65b9\u6cd5\u3001Benchmark \u548c Codebase<\/strong>\uff0c\u4e3a\u76f8\u5173\u9886\u57df\u7684\u7814\u7a76\u8005\u548c\u5b9e\u8df5\u8005\u63d0\u4f9b\u4e86\u5168\u9762\u652f\u6301\u3002<\/p>\n\n\n\n

\u751f\u6210\u5f0f AI \u7684\u6301\u7eed\u5b66\u4e60\u7efc\u8ff0<\/strong><\/p>\n\n\n\n

\u8bba\u6587\uff1aContinual Learning for Generative AI: From LLMs to MLLMs and Beyond<\/strong><\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n
    \n
  • \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2506.13045<\/li>\n\n\n\n
  • \u9879\u76ee\u4e3b\u9875\uff1ahttps:\/\/github.com\/Ghy0501\/Awesome-Continual-Learning-in-Generative-Models<\/li>\n<\/ul>\n\n\n\n

    \u7814\u7a76\u52a8\u673a<\/strong>\uff1a\u4ee5\u5927\u6a21\u578b\u4e3a\u4ee3\u8868\u7684\u751f\u6210\u5f0f AI \u6a21\u578b\u7684\u5feb\u901f\u53d1\u5c55\u8ba9\u73b0\u4ee3\u667a\u80fd\u7cfb\u7edf\u5177\u5907\u4e86\u7406\u89e3\u548c\u751f\u6210\u590d\u6742\u5185\u5bb9\u7684\u80fd\u529b\uff0c\u751a\u81f3\u5728\u90e8\u5206\u9886\u57df\u8fbe\u5230\u4e86\u63a5\u8fd1\u4eba\u7c7b\u7684\u8868\u73b0\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u6a21\u578b\u4f9d\u65e7\u9762\u4e34\u7740\u201c\u707e\u96be\u6027\u9057\u5fd8\u201d\u95ee\u9898\uff0c\u5373\u5728\u5b66\u4e60\u65b0\u4efb\u52a1\u65f6\uff0c\u5f80\u5f80\u4f1a\u5bfc\u81f4\u5df2\u5b66\u4efb\u52a1\u6027\u80fd\u7684\u663e\u8457\u4e0b\u964d\u3002\u4e3a\u89e3\u51b3\u8fd9\u4e00\u6311\u6218\uff0c\u5927\u91cf\u7684\u7814\u7a76\u63d0\u51fa\u4e86\u591a\u79cd\u65b9\u6cd5\u4ee5\u589e\u5f3a\u751f\u6210\u5f0f AI \u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u7684\u9002\u5e94\u6027\u548c\u6269\u5c55\u6027\u3002\u672c\u6587\u7cfb\u7edf\u6027\u5730\u7efc\u8ff0\u4e86\u751f\u6210\u5f0f AI \u7684\u6301\u7eed\u5b66\u4e60\u65b9\u6cd5\uff0c\u6db5\u76d6\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u3001\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\uff08MLLMs\uff09\u3001\u89c6\u89c9\u8bed\u8a00\u52a8\u4f5c\u6a21\u578b\uff08VLA\uff09\u548c\u6269\u6563\u6a21\u578b\uff08Diffusion Models\uff09\u3002<\/p>\n\n\n\n

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

          \u56fe 1\uff1a\u751f\u6210\u5f0f AI \u4e2d\u7684\u6301\u7eed\u5b66\u4e60\u793a\u610f\u56fe<\/sup><\/p>\n\n\n\n

    \u7814\u7a76\u5185\u5bb9<\/strong>\uff1a\u672c\u6587\u56f4\u7ed5\u751f\u6210\u5f0f AI \u7684\u6301\u7eed\u5b66\u4e60\u95ee\u9898\uff0c\u7cfb\u7edf\u6027\u5730\u7efc\u8ff0\u4e86\u4e0d\u540c\u6a21\u578b\u7684\u8bad\u7ec3\u76ee\u6807\u3001\u5e94\u7528\u573a\u666f\u53ca\u6280\u672f\u65b9\u6cd5\u3002\u7814\u7a76\u6db5\u76d6\u5927\u8bed\u8a00\u6a21\u578b\u5728\u7406\u89e3\u4e0e\u751f\u6210\u4e2d\u7684\u77e5\u8bc6\u4fdd\u7559\u4e0e\u4efb\u52a1\u9002\u5e94\u3001\u591a\u6a21\u6001\u5927\u6a21\u578b\u5904\u7406\u8de8\u6a21\u6001\u6570\u636e\u65f6\u7684\u6297\u9057\u5fd8\u80fd\u529b\u3001\u89c6\u89c9\u8bed\u8a00\u52a8\u4f5c\u6a21\u578b\u5728\u673a\u5668\u4eba\u52a8\u6001\u73af\u5883\u4e2d\u7684\u884c\u4e3a\u8fc1\u79fb\u4e0e\u9002\u5e94\uff0c\u4ee5\u53ca\u6269\u6563\u6a21\u578b\u9488\u5bf9\u4e2a\u6027\u5316\u751f\u6210\u9700\u6c42\u7684\u589e\u91cf\u5b66\u4e60\u3002\u8fd9\u4e9b\u6a21\u578b\u7684\u6301\u7eed\u5b66\u4e60\u65b9\u6cd5\u4e3b\u8981\u5305\u62ec\u67b6\u6784\u6269\u5c55\u3001\u6b63\u5219\u5316\u548c\u56de\u653e\u7b56\u7565\uff0c\u65e8\u5728\u5e73\u8861\u65b0\u4efb\u52a1\u5b66\u4e60\u4e0e\u65e7\u4efb\u52a1\u6027\u80fd\u7684\u4fdd\u6301\u3002\u6b64\u5916\uff0c\u7814\u7a76\u8fd8\u63a2\u8ba8\u4e86\u8bc4\u4f30\u6307\u6807\uff08\u6574\u4f53\u6027\u80fd\u3001\u9057\u5fd8\u7a0b\u5ea6\u3001\u6cdb\u5316\u80fd\u529b\uff09\u548c\u672a\u6765\u65b9\u5411\uff08\u9ad8\u6548\u673a\u5236\u3001\u5f3a\u5316\u5b66\u4e60\u8303\u5f0f\u3001\u591a\u6a21\u6001\u6269\u5c55\u7b49\uff09\uff0c\u4e3a\u751f\u6210\u5f0f AI \u7684\u6301\u7eed\u5b66\u4e60\u63d0\u4f9b\u4e86\u5168\u9762\u53c2\u8003\u3002<\/p>\n\n\n\n

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

          \u56fe 2\uff1a\u6301\u7eed\u5b66\u4e60\u65b9\u6cd5\u5206\u7c7b\u56fe<\/sup><\/p>\n\n\n\n

    \u591a\u6a21\u6001\u5927\u6a21\u578b\u6301\u7eed\u5b66\u4e60\uff1aBenchmark \u4e0e\u65b9\u6cd5<\/strong><\/p>\n\n\n\n

    \u4f20\u7edf\u7684\u6301\u7eed\u5b66\u4e60\u4efb\u52a1\u591a\u805a\u7126\u4e8e\u5355\u6a21\u6001\u573a\u666f\uff0c\u5982\u56fe\u50cf\u6216\u6587\u672c\u5206\u7c7b\uff0c\u4f46\u968f\u7740\u5e94\u7528\u9700\u6c42\u7684\u590d\u6742\u5316\uff0c\u591a\u6a21\u6001\u4efb\u52a1\u9010\u6e10\u6210\u4e3a\u6838\u5fc3\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u7cfb\u5217\u65b0\u7684 Benchmark \u548c\u65b9\u6cd5\uff0c\u65e8\u5728\u63a8\u52a8\u591a\u6a21\u6001\u5927\u6a21\u578b\u6301\u7eed\u5b66\u4e60\u7684\u53d1\u5c55\u3002<\/p>\n\n\n\n

    \u8bba\u6587 1\uff1a[ACL 2025] HiDe-LLaVA: Hierarchical Decoupling for Continual Instruction Tuning of Multimodal Large Language Model<\/strong><\/p>\n\n\n\n

    \"\"<\/figure>\n\n\n\n
      \n
    • \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2503.12941<\/li>\n\n\n\n
    • \u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/Ghy0501\/HiDe-LLaVA<\/li>\n\n\n\n
    • \u6570\u636e\u94fe\u63a5\uff1ahttps:\/\/huggingface.co\/datasets\/HaiyangGuo\/UCIT<\/li>\n<\/ul>\n\n\n\n

      \u7814\u7a76\u52a8\u673a<\/strong>\uff1a\u672c\u6587\u8ba4\u4e3a\u5f53\u524d\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u6301\u7eed\u5b66\u4e60\u9762\u4e34\u4e24\u5927\u5173\u952e\u6311\u6218\uff1a\u4e00\u662f\u73b0\u6709\u8bc4\u6d4b\u57fa\u51c6\u666e\u904d\u5b58\u5728\u4e0e\u9884\u8bad\u7ec3\u6570\u636e\u91cd\u53e0\u7684\u95ee\u9898\uff0c\u5bfc\u81f4\u8bc4\u4f30\u7ed3\u679c\u5931\u771f\uff1b\u4e8c\u662f\u4f20\u7edf\u65b9\u6cd5\u96be\u4ee5\u5e73\u8861\u65b0\u4efb\u52a1\u5b66\u4e60\u4e0e\u65e7\u4efb\u52a1\u9057\u5fd8\u4e4b\u95f4\u7684\u77db\u76fe\u3002\u4e3a\u6b64\uff0c\u672c\u7814\u7a76\u63d0\u51fa\u6784\u5efa\u5168\u65b0\u7684 UCIT <\/strong>\u8bc4\u6d4b\u57fa\u51c6\uff0c\u901a\u8fc7\u4e25\u683c\u7684 zero-shot \u7b5b\u9009\u673a\u5236\uff0c\u6784\u5efa\u4e86\u5168\u65b0\u7684\u65e0\u4fe1\u606f\u6cc4\u9732\u98ce\u9669\u6570\u636e\u96c6\u3002\u540c\u65f6\u521b\u65b0\u6027\u5730\u63d0\u51fa\u5c42\u6b21\u5316\u89e3\u8026\u5b66\u4e60\u7b56\u7565\uff0c\u65e8\u5728\u89e3\u51b3\u6301\u7eed\u6307\u4ee4\u5fae\u8c03\u4e2d\u7684\u707e\u96be\u6027\u9057\u5fd8\u95ee\u9898\uff0c\u4e3a\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u6301\u7eed\u5b66\u4e60\u63d0\u4f9b\u53ef\u9760\u7684\u8bc4\u4f30\u6846\u67b6\u548c\u89e3\u51b3\u65b9\u6848\u3002<\/p>\n\n\n\n

      \u7814\u7a76\u65b9\u6cd5<\/strong>\uff1a\u672c\u6587\u901a\u8fc7 CKA \u76f8\u4f3c\u6027\u5206\u6790\u63ed\u793a\u4e86\u6a21\u578b\u4e0d\u540c\u5c42\u7ea7\u7684\u4efb\u52a1\u7279\u6027\u5dee\u5f02\uff1a\u9876\u5c42\u5177\u6709\u4efb\u52a1\u7279\u5f02\u6027\uff0c\u5176\u4f59\u5c42\u5219\u4fdd\u6301\u4efb\u52a1\u901a\u7528\u6027\u3002\u57fa\u4e8e\u6b64\uff0c\u672c\u6587\u63d0\u51fa\u7684 HiDe-LLaVA <\/strong>\u91c7\u7528\u5206\u5c42\u5904\u7406\u673a\u5236\uff1a\u5728\u9876\u5c42\u5f15\u5165\u591a\u6a21\u6001\u951a\u70b9\u9a71\u52a8\u7684\u52a8\u6001\u4e13\u5bb6\u9009\u62e9\u6a21\u5757\uff0c\u5b9e\u73b0\u4efb\u52a1\u81ea\u9002\u5e94\uff1b\u5728\u5176\u4f59\u5c42\u91c7\u7528\u53c2\u6570\u878d\u5408\u7b56\u7565\u4fdd\u7559\u8de8\u4efb\u52a1\u5171\u4eab\u77e5\u8bc6\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u8be5\u65b9\u6cd5\u53ef\u4ee5\u6709\u6548\u7f13\u89e3\u6a21\u578b\u7684\u707e\u96be\u6027\u9057\u5fd8\u73b0\u8c61\uff0c\u5e76\u4e14\u6709\u6548\u5e73\u8861\u4e86\u6a21\u578b\u6027\u80fd\u4e0e\u8ba1\u7b97\u8d44\u6e90\u6548\u7387\u3002<\/p>\n\n\n\n

      \u76ee\u524d\u8be5\u7814\u7a76\u5df2\u88ab ACL 2025 <\/strong>\u4e3b\u4f1a\u63a5\u6536\uff0c\u76f8\u5173\u4ee3\u7801\u53ca\u6570\u636e\u5df2\u5168\u90e8\u5f00\u6e90\u3002<\/p>\n\n\n\n

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

            \u56fe 3\uff1aHiDe-LLaVA \u6a21\u578b\u6846\u67b6\u793a\u610f\u56fe\u3002<\/sup><\/p>\n\n\n\n

      \u8bba\u6587 2\uff1a[ICCV 2025] Federated Continual Instruction Tuning<\/strong><\/p>\n\n\n\n

      \"\"<\/figure>\n\n\n\n
        \n
      • \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2503.12897<\/li>\n\n\n\n
      • \u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/Ghy0501\/FCIT<\/li>\n\n\n\n
      • \u6570\u636e\u94fe\u63a5\uff1ahttps:\/\/huggingface.co\/datasets\/MLLM-CL\/FCIT<\/li>\n<\/ul>\n\n\n\n

        \u7814\u7a76\u52a8\u673a<\/strong>\uff1a\u5f53\u524d\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u6307\u4ee4\u5fae\u8c03\u9762\u4e34\u96c6\u4e2d\u8bad\u7ec3\u6210\u672c\u9ad8\u3001\u5b9e\u7528\u6027\u5dee\u7684\u95ee\u9898\uff0c\u800c\u73b0\u6709\u8054\u90a6\u5b66\u4e60\u65b9\u6848\u5728\u52a8\u6001\u73af\u5883\u4e2d\u96be\u4ee5\u5b9e\u73b0\u6301\u7eed\u5b66\u4e60\u4e0e\u77e5\u8bc6\u4fdd\u7559\u7684\u5e73\u8861\u3002\u4e3a\u6b64\uff0c\u672c\u6587\u9996\u6b21\u63d0\u51fa \u8054\u90a6\u8fde\u7eed\u6307\u4ee4\u5fae\u8c03\uff08FCIT\uff09<\/strong>\u57fa\u51c6\uff0c\u4ee5\u89e3\u51b3\u5206\u5e03\u5f0f\u73af\u5883\u4e0b\u6301\u7eed\u5b66\u4e60\u7684\u5173\u952e\u6311\u6218\u3002FCIT \u57fa\u51c6\u5305\u542b\u540c\u8d28\uff08\u5404\u5ba2\u6237\u7aef\u5b66\u4e60\u76f8\u540c\u4efb\u52a1\uff09\u548c\u5f02\u8d28\uff08\u5404\u5ba2\u6237\u7aef\u5b66\u4e60\u4e0d\u540c\u4efb\u52a1\uff09\u4e24\u79cd\u73b0\u5b9e\u573a\u666f\uff0c\u901a\u8fc7 4 \u79cd\u5b9e\u9a8c\u8bbe\u7f6e\u548c 12 \u4e2a\u7cbe\u9009\u6570\u636e\u96c6\uff0c\u5168\u9762\u8bc4\u4f30\u6a21\u578b\u5728\u975e\u72ec\u7acb\u540c\u5206\u5e03\u6570\u636e\u548c\u707e\u96be\u6027\u9057\u5fd8\u60c5\u51b5\u4e0b\u7684\u8868\u73b0\uff0c\u4e3a\u591a\u6a21\u6001\u5927\u6a21\u578b\u7684\u8054\u90a6\u6301\u7eed\u5b66\u4e60\u7814\u7a76\u63d0\u4f9b\u6807\u51c6\u5316\u8bc4\u6d4b\u6846\u67b6\u3002<\/p>\n\n\n\n

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

              \u56fe 4\uff1aFCIT \u8bbe\u5b9a\u793a\u610f\u56fe<\/sup><\/p>\n\n\n\n

        \u7814\u7a76\u65b9\u6cd5<\/strong>\uff1a\u4e3a\u5e94\u5bf9 FCIT \u4e2d\u7684\u6311\u6218\uff0c\u6211\u4eec\u63d0\u51fa\u4e86 DISCO <\/strong>\u6846\u67b6\uff0c\u7ed3\u5408\u4e86 \u52a8\u6001\u77e5\u8bc6\u68b3\u7406\uff08DKO\uff09 \u548c \u5b50\u7a7a\u95f4\u9009\u62e9\u6fc0\u6d3b\uff08SSA\uff09 \u4e24\u79cd\u7b56\u7565\u3002DKO \u5229\u7528\u5168\u5c40\u670d\u52a1\u5668\u7684\u52a8\u6001\u7f13\u5b58\uff0c\u5b58\u50a8\u5e76\u7ec4\u7ec7\u4efb\u52a1\u7279\u5b9a\u7684\u53c2\u6570\uff0c\u51cf\u5c11\u4efb\u52a1\u95f4\u548c\u9636\u6bb5\u95f4\u7684\u51b2\u7a81\uff1bSSA \u5219\u901a\u8fc7\u5339\u914d\u8f93\u5165\u7279\u5f81\u4e0e\u52a8\u6001\u7f13\u5b58\u4e2d\u7684\u4efb\u52a1\u5b50\u7a7a\u95f4\uff0c\u9009\u62e9\u6027\u6fc0\u6d3b\u76f8\u5173\u8f93\u51fa\uff0c\u540c\u65f6\u8fc7\u6ee4\u65e0\u5173\u4fe1\u606f\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cDISCO \u5728\u89e3\u51b3\u6570\u636e\u5f02\u8d28\u6027\u548c\u707e\u96be\u6027\u9057\u5fd8\u65b9\u9762\u663e\u8457\u63d0\u5347\u4e86\u6a21\u578b\u6027\u80fd\uff0c\u5168\u9762\u8d85\u8d8a\u73b0\u6709\u65b9\u6cd5\uff0c\u5e76\u5728 FCIT \u57fa\u51c6\u4e0a\u53d6\u5f97\u4e86\u6700\u597d\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n

        \u76ee\u524d\u8be5\u7814\u7a76\u5df2\u88ab ICCV 2025<\/strong> \u63a5\u6536\uff0c\u76f8\u5173\u4ee3\u7801\u53ca\u6570\u636e\u5df2\u5168\u90e8\u5f00\u6e90\u3002<\/p>\n\n\n\n

        \u8bba\u6587 3\uff1a[EMNLP 2025] ModalPrompt: Dual-Modality Guided Prompt for Continual Learning of Large Multimodal Models<\/strong><\/p>\n\n\n\n

        \"\"<\/figure>\n\n\n\n
          \n
        • \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2410.05849<\/li>\n\n\n\n
        • \u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/AuroraZengfh\/ModalPrompt<\/li>\n<\/ul>\n\n\n\n

          \u7814\u7a76\u52a8\u673a<\/strong>\uff1a\u4e3a\u7f13\u89e3\u591a\u6a21\u6001\u5927\u6a21\u578b\u6301\u7eed\u5b66\u4e60\u4efb\u52a1\u4e2d\u7684\u707e\u96be\u6027\u9057\u5fd8\u73b0\u8c61\uff0c\u672c\u6587\u8ba4\u4e3a\u73b0\u6709\u89e3\u51b3\u65b9\u6848\u5b58\u5728\u663e\u8457\u5c40\u9650\u6027\uff1a\u57fa\u4e8e\u6570\u636e\u56de\u653e\u7684\u65b9\u6cd5\u9762\u4e34\u9690\u79c1\u6cc4\u9732\u98ce\u9669\u548c\u5b58\u50a8\u6210\u672c\u538b\u529b\uff0c\u800c\u6a21\u578b\u6269\u5c55\u7b56\u7565\u5219\u4e0d\u53ef\u907f\u514d\u5730\u5f15\u53d1\u8ba1\u7b97\u8d44\u6e90\u7684\u7ebf\u6027\u589e\u957f\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5f53\u524d\u7814\u7a76\u5c1a\u672a\u5145\u5206\u63a2\u7d22\u591a\u6a21\u6001\u6570\u636e\u5728\u6301\u7eed\u5b66\u4e60\u4e2d\u7684\u534f\u540c\u76d1\u7763\u6f5c\u529b\u3002\u800c\u5f53\u524d\u9886\u57df\u5185\u7f3a\u4e4f\u4e13\u95e8\u9488\u5bf9\u591a\u6a21\u6001\u7279\u6027\u7684\u6301\u7eed\u5b66\u4e60\u6846\u67b6\uff0c\u56e0\u6b64\u9700\u8981\u5f00\u53d1\u4e00\u79cd\u65e2\u80fd\u5229\u7528\u56fe\u50cf-\u6587\u672c\u53cc\u91cd\u76d1\u7763\u3001\u53c8\u80fd\u907f\u514d\u8ba1\u7b97\u81a8\u80c0\u7684\u65b0\u65b9\u6cd5\uff0c\u4ee5\u5b9e\u73b0\u9ad8\u6548\u4e14\u9690\u79c1\u5b89\u5168\u7684\u77e5\u8bc6\u6301\u7eed\u79ef\u7d2f\u3002<\/p>\n\n\n\n

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

                \u56fe 5\uff1aModalPrompt \u6a21\u578b\u6846\u67b6\u793a\u610f\u56fe<\/sup><\/p>\n\n\n\n

          \u7814\u7a76\u65b9\u6cd5<\/strong>\uff1a\u672c\u6587\u63d0\u51fa ModalPrompt <\/strong>\u6846\u67b6\uff0c\u5229\u7528\u591a\u6a21\u6001\u76d1\u7763\uff0c\u901a\u8fc7\u6784\u5efa\u4efb\u52a1\u7279\u5b9a\u7684\u56fe\u50cf\u6587\u672c\u539f\u578b\u63d0\u793a\uff0c\u7ed3\u5408\u53cc\u6a21\u6001\u5f15\u5bfc\u63d0\u793a\u9009\u62e9\u548c\u591a\u4efb\u52a1\u63d0\u793a\u878d\u5408\u673a\u5236\uff0c\u5b9e\u73b0\u4e86\u5728\u65e0\u56de\u653e\u6570\u636e\u7684\u60c5\u51b5\u4e0b\u6709\u6548\u4fdd\u7559\u65e7\u4efb\u52a1\u77e5\u8bc6\u5e76\u63d0\u5347\u65b0\u4efb\u52a1\u6027\u80fd\u3002\u6b64\u5916\uff0c\u8be5\u65b9\u6cd5\u901a\u8fc7\u52a8\u6001\u63d0\u793a\u9009\u62e9\u964d\u4f4e\u8ba1\u7b97\u590d\u6742\u5ea6\uff0c\u4f7f\u63a8\u7406\u901f\u5ea6\u63d0\u5347 1.42 \u500d\uff0c\u540c\u65f6\u663e\u8457\u51cf\u5c11\u5b58\u50a8\u548c\u8bad\u7ec3\u6210\u672c\u3002<\/p>\n\n\n\n

          \u76ee\u524d\u8be5\u7814\u7a76\u5df2\u88ab EMNLP 2025<\/strong> \u4e3b\u4f1a\u63a5\u6536\uff0c\u76f8\u5173\u4ee3\u7801\u5df2\u5168\u90e8\u5f00\u6e90\u3002<\/p>\n\n\n\n

          \u8bba\u6587 4. MLLM-CL: Continual Learning for Multimodal Large Language Models<\/strong><\/p>\n\n\n\n

          \"\"<\/figure>\n\n\n\n
            \n
          • \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2506.05453<\/li>\n\n\n\n
          • \u4ee3\u7801\u94fe\u63a5\uff1ahttps:\/\/github.com\/bjzhb666\/MLLM-CL<\/li>\n\n\n\n
          • \u6570\u636e\u94fe\u63a5\uff1ahttps:\/\/huggingface.co\/datasets\/Impression2805\/MLLM-CL<\/li>\n<\/ul>\n\n\n\n

            \u7814\u7a76\u52a8\u673a<\/strong>\uff1a\u672c\u6587\u8ba4\u4e3a\u73b0\u6709\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u8fde\u7eed\u6307\u4ee4\u5fae\u8c03\u8bc4\u6d4b\u57fa\u51c6\u4e3b\u8981\u5173\u6ce8\u72ec\u7acb\u540c\u5206\u5e03\uff08IID\uff09\u573a\u666f\u4e0b\u7684\u9886\u57df\u77e5\u8bc6\u8bc4\u4f30\uff0c\u7f3a\u4e4f\u5bf9\u6a21\u578b\u57fa\u7840\u80fd\u529b\uff08\u5982 OCR\u3001\u6570\u5b66\u63a8\u7406\u7b49\uff09\u5728\u975e IID \u573a\u666f\u4e0b\u7684\u7cfb\u7edf\u6027\u8bc4\u6d4b\u3002\u4e3a\u6b64\uff0c\u672c\u6587\u63d0\u51fa\u4e86\u4e00\u4e2a\u65b0\u7684\u591a\u6a21\u6001\u5927\u6a21\u578b\u6301\u7eed\u5b66\u4e60\u57fa\u51c6 MLLM-CL<\/strong>\uff0c\u6db5\u76d6\u9886\u57df\u6301\u7eed\u5b66\u4e60\uff08DCL\uff09 \u548c\u80fd\u529b\u6301\u7eed\u5b66\u4e60\uff08ACL\uff09 \u4e24\u79cd\u8bbe\u7f6e\uff0c\u5206\u522b\u9488\u5bf9\u540c\u5206\u5e03\uff08IID\uff09\u548c\u975e\u540c\u5206\u5e03\uff08non-IID\uff09\u573a\u666f\u4e0b\u7684\u9886\u57df\u77e5\u8bc6\u548c\u57fa\u7840\u80fd\u529b\u5b66\u4e60\u8fdb\u884c\u8bc4\u4f30\u3002<\/p>\n\n\n\n

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

                  \u56fe 6\uff1aMLLM-CL \u57fa\u51c6\u793a\u610f\u56fe<\/sup><\/p>\n\n\n\n

            \u7814\u7a76\u65b9\u6cd5<\/strong>\uff1a\u4e3a\u89e3\u51b3\u707e\u96be\u6027\u9057\u5fd8\u95ee\u9898\uff0c\u672c\u6587\u63d0\u51fa\u4e86 MR-LoRA<\/strong>\uff0c\u901a\u8fc7\u9886\u57df\u6216\u80fd\u529b\u7279\u5b9a\u7684 LoRA \u6a21\u5757\u5b9e\u73b0\u53c2\u6570\u9694\u79bb\uff0c\u907f\u514d\u4efb\u52a1\u95f4\u5e72\u6270\uff0c\u5e76\u8bbe\u8ba1\u4e86\u57fa\u4e8e MLLM \u81ea\u8eab\u7684\u591a\u6a21\u6001\u7406\u89e3\u80fd\u529b\u7684\u8def\u7531\u9009\u62e9\u5668\uff0c\u4ec5\u9700\u5c11\u91cf\u6837\u672c\u5fae\u8c03\u5373\u53ef\u7cbe\u51c6\u5339\u914d\u8f93\u5165\u4e0e\u6700\u4f18\u4e13\u5bb6\u6a21\u5757\u3002\u5b9e\u9a8c\u8868\u660e\uff0c\u8be5\u65b9\u6cd5\u5728\u9886\u57df\u6301\u7eed\u5b66\u4e60\uff08DCL\uff09\u548c\u80fd\u529b\u6301\u7eed\u5b66\u4e60\uff08ACL\uff09\u4efb\u52a1\u4e0a\u663e\u8457\u4f18\u4e8e\u4f20\u7edf\u56de\u653e\u6216\u6a21\u578b\u6269\u5c55\u65b9\u6cd5\u3002<\/p>\n\n\n\n

            \u8bba\u6587 5. LLaVA-c: Continual Improved Visual Instruction Tuning<\/strong><\/p>\n\n\n\n

            \"\"<\/figure>\n\n\n\n
              \n
            • \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2506.08666<\/li>\n<\/ul>\n\n\n\n

              \u7814\u7a76\u52a8\u673a<\/strong>\uff1a\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08\u5982 LLaVA-1.5\uff09\u5728\u8fde\u7eed\u6307\u4ee4\u5fae\u8c03\u4e2d\u9762\u4e34\u7684\u4e24\u5927\u6838\u5fc3\u6311\u6218\uff1a\u9996\u5148\uff0c\u4f20\u7edf\u7684\u591a\u4efb\u52a1\u8054\u5408\u8bad\u7ec3\u5b58\u5728\u4efb\u52a1\u5e73\u8861\u56f0\u96be\uff08\u9700\u4eba\u5de5\u8c03\u6574\u6570\u636e\u6bd4\u4f8b\uff09\u548c\u6269\u5c55\u6210\u672c\u9ad8\uff08\u65b0\u589e\u4efb\u52a1\u9700\u5168\u91cf\u91cd\u8bad\u7ec3\uff09\u7684\u56fa\u6709\u7f3a\u9677\uff1b\u5176\u6b21\uff0c\u73b0\u6709\u6301\u7eed\u5b66\u4e60\u65b9\u6cd5\u867d\u80fd\u589e\u91cf\u5b66\u4e60\u65b0\u4efb\u52a1\uff0c\u4f46\u666e\u904d\u5b58\u5728 “\u57fa\u7840\u6a21\u578b\u9000\u5316” \u73b0\u8c61\u2014\u2014\u6a21\u578b\u8fc7\u5ea6\u62df\u5408\u4efb\u52a1\u7279\u5b9a\u6307\u4ee4\uff08\u5982\u5f3a\u5236\u5355\u5b57\u56de\u7b54\uff09\uff0c\u4e27\u5931\u5904\u7406\u591a\u6837\u5316\u6307\u4ee4\u7684\u901a\u7528\u80fd\u529b\u3002<\/p>\n\n\n\n

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

                    \u56fe 7\uff1aLLaVA-c \u6a21\u578b\u6846\u67b6\u793a\u610f\u56fe<\/sup><\/p>\n\n\n\n

              \u7814\u7a76\u65b9\u6cd5<\/strong>\uff1a\u672c\u6587\u63d0\u51fa\u4e86 LLaVA-c<\/strong>\uff0c\u901a\u8fc7\u4e24\u4e2a\u6838\u5fc3\u6280\u672f\u6539\u8fdb LLaVA-1.5 \u6a21\u578b\uff1a1\uff09\u8c31\u611f\u77e5\u5de9\u56fa\uff08SAC\uff09\uff0c\u57fa\u4e8e\u5947\u5f02\u503c\u5206\u89e3\u7684\u6a21\u578b\u878d\u5408\u7b56\u7565\u6709\u6548\u514b\u670d\u65b0\u65e7\u77e5\u8bc6\u51b2\u7a81\uff0c\u76f8\u6bd4\u4f20\u7edf\u7684\u6a21\u578b\u6df7\u5408\u7b56\u7565\u63d0\u5347\u4e86\u4efb\u52a1\u517c\u5bb9\u6027\uff1b2\uff09\u65e0\u76d1\u7763\u67e5\u8be2\u6b63\u5219\u5316\uff08UIR\uff09\uff0c\u901a\u8fc7\u7ea6\u675f\u672a\u6807\u6ce8\u6587\u672c\u6307\u4ee4\u7684\u7279\u5f81\u7a7a\u95f4\u504f\u79fb\uff08L2 \u8ddd\u79bb\u635f\u5931\uff09\u9632\u6b62\u57fa\u7840\u6a21\u578b\u9000\u5316\uff0c\u5728\u96f6\u989d\u5916\u6807\u6ce8\u6210\u672c\u4e0b\u4fdd\u6301\u6307\u4ee4\u8ddf\u968f\u80fd\u529b\u3002\u672c\u6587\u5728\u9884\u8bad\u7ec3\u548c\u6307\u4ee4\u5fae\u8c03\u4e24\u9636\u6bb5\u4e0a\u90fd\u9a8c\u8bc1\u4e86\u6240\u63d0\u51fa\u65b9\u6cd5\u7684\u6709\u6548\u6027\uff0c\u5728\u901a\u7528\u8bc4\u4ef7\u57fa\u51c6\u548c\u4e0b\u6e38\u4efb\u52a1\u6307\u6807\u4e0a\u5747\u53d6\u5f97\u4e86\u6700\u4f18\u7684\u6027\u80fd\uff0c\u5e76\u4e14\u9996\u6b21\u5b9e\u73b0\u6301\u7eed\u5b66\u4e60\u6548\u679c\u8d85\u8d8a\u591a\u4efb\u52a1\u8054\u5408\u8bad\u7ec3\u3002<\/p>\n\n\n\n

              \u591a\u6a21\u6001\u5927\u6a21\u578b\u6301\u7eed\u5b66\u4e60\uff1a\u4ee3\u7801\u4ed3\u5e93<\/strong><\/p>\n\n\n\n

              \u8bba\u6587\uff1aMCITlib: Multimodal Continual Instruction Tuning Library and Benchmark<\/strong><\/p>\n\n\n\n

              \"\"<\/figure>\n\n\n\n
                \n
              • \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2508.07307<\/li>\n\n\n\n
              • \u4ee3\u7801\u4ed3\u5e93\uff1ahttps:\/\/github.com\/Ghy0501\/MCITlib<\/li>\n<\/ul>\n\n\n\n

                \u7814\u7a76\u52a8\u673a<\/strong>\uff1a\u968f\u7740\u591a\u6a21\u6001\u5927\u6a21\u578b\u6301\u7eed\u5b66\u4e60\u7814\u7a76\u7684\u84ec\u52c3\u53d1\u5c55\uff0c\u5404\u7c7b\u521b\u65b0\u65b9\u6cd5\u548c\u8bc4\u4f30\u57fa\u51c6\u4e0d\u65ad\u6d8c\u73b0\uff0c\u4f46\u7814\u7a76\u793e\u533a\u59cb\u7ec8\u7f3a\u4e4f\u4e00\u4e2a\u7cfb\u7edf\u5316\u3001\u6807\u51c6\u5316\u7684\u5f00\u53d1\u4e0e\u8bc4\u6d4b\u5e73\u53f0\u3002\u4e3a\u586b\u8865\u8fd9\u4e00\u5173\u952e\u7a7a\u767d\uff0c\u6211\u4eec\u63a8\u51fa\u4e86 MCITlib<\/strong>\uff0c\u4e00\u4e2a\u5f00\u6e90\u7684\u591a\u6a21\u6001\u6301\u7eed\u6307\u4ee4\u5fae\u8c03\u4ee3\u7801\u4ed3\u5e93\u3002MCITlib \u96c6\u6210\u4e86\u5f53\u524d\u9886\u57df\u5185 8 \u79cd\u4e3b\u6d41\u7b97\u6cd5\uff0c\u7cbe\u5fc3\u6311\u9009\u4e86\u4e24\u4e2a\u9ad8\u8d28\u91cf\u57fa\u51c6\uff08UCIT \u548c DCL\uff09\uff0c\u6709\u6548\u907f\u514d\u4fe1\u606f\u6cc4\u9732\uff0c\u4e3a\u7814\u7a76\u8005\u63d0\u4f9b\u4e86\u4e00\u4e2a\u7edf\u4e00\u3001\u516c\u5e73\u7684\u5b9e\u9a8c\u73af\u5883\uff0c\u4fbf\u4e8e\u5168\u9762\u8bc4\u4f30\u4e0d\u540c\u65b9\u6cd5\u7684\u4f18\u52a3\u3002<\/p>\n\n\n\n

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

                      \u56fe 8\uff1a\u5f00\u6e90\u4ee3\u7801\u4ed3\u5e93 MCITlib<\/sup><\/p>\n\n\n\n

                \u672a\u6765\uff0cMCITlib \u4e5f\u5c06\u6301\u7eed\u8fdb\u884c\u66f4\u65b0\uff0c\u6269\u5c55\u66f4\u591a\u6a21\u578b\u3001\u4efb\u52a1\u548c\u8bc4\u6d4b\u7ef4\u5ea6\uff0c\u4e3a\u591a\u6a21\u6001\u5927\u6a21\u578b\u6301\u7eed\u5b66\u4e60\u7814\u7a76\u63d0\u4f9b\u575a\u5b9e\u52a9\u529b\u3002<\/p>\n\n\n\n

                \u603b\u7ed3\u4e0e\u5c55\u671b<\/strong><\/p>\n\n\n\n

                \u8d4b\u4e88\u4ee5\u591a\u6a21\u6001\u5927\u6a21\u578b\u4e3a\u4ee3\u8868\u7684\u751f\u6210\u5f0f AI \u6301\u7eed\u5b66\u4e60\u7684\u80fd\u529b\u662f\u8fc8\u5411\u4eba\u5de5\u667a\u80fd\u901a\u7528\u5316\u7684\u91cd\u8981\u4e00\u6b65\u3002\u6211\u4eec\u5e0c\u671b\u901a\u8fc7\u7cfb\u7edf\u7684\u7efc\u8ff0\u3001\u5b8c\u5584\u7684 Benchmark\u3001\u524d\u6cbf\u7684\u65b9\u6cd5\u548c\u5f00\u6e90\u7684\u5de5\u5177\uff0c\u80fd\u591f\u4e3a\u8fd9\u4e00\u9886\u57df\u7684\u7814\u7a76\u8005\u548c\u5e94\u7528\u5f00\u53d1\u8005\u63d0\u4f9b\u66f4\u591a\u652f\u6301\u3002\u672a\u6765\uff0c\u6211\u4eec\u56e2\u961f\u5c06\u7ee7\u7eed\u6df1\u8015\u591a\u6a21\u6001\u5927\u6a21\u578b\u6301\u7eed\u5b66\u4e60\u9886\u57df\uff0c\u63a2\u7d22\u66f4\u5e7f\u6cdb\u7684\u5e94\u7528\u573a\u666f\uff0c\u6301\u7eed\u63a8\u52a8\u8be5\u9886\u57df\u6280\u672f\u7684\u53d1\u5c55\u4e0e\u521b\u65b0\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

                \u8fd1\u5e74\u6765\uff0c\u751f\u6210\u5f0f AI \u548c\u591a\u6a21\u6001\u5927\u6a21\u578b\u5728\u5404\u9886\u57df\u53d6\u5f97\u4e86\u4ee4\u4eba\u77a9\u76ee\u7684\u8fdb\u5c55\u3002\u7136\u800c\uff0c\u5728\u73b0\u5b9e\u4e16\u754c\u5e94\u7528\u4e2d\uff0c\u52a8\u6001\u73af\u5883\u4e0b\u7684\u6570\u636e\u5206\u5e03 […]<\/p>\n","protected":false},"author":1,"featured_media":277,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[],"class_list":["post-262","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-8"],"_links":{"self":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/262","targetHints":{"allow":["GET"]}}],"collection":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/comments?post=262"}],"version-history":[{"count":1,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/262\/revisions"}],"predecessor-version":[{"id":278,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/posts\/262\/revisions\/278"}],"wp:featuredmedia":[{"embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media\/277"}],"wp:attachment":[{"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/media?parent=262"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/categories?post=262"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/www.applicationsofllm.com\/index.php\/wp-json\/wp\/v2\/tags?post=262"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}