{"id":1163619,"date":"2026-03-16T01:26:36","date_gmt":"2026-03-16T08:26:36","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&p=1163619"},"modified":"2026-03-18T02:32:59","modified_gmt":"2026-03-18T09:32:59","slug":"llm2clip%ef%bc%9a%e4%bb%a5%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e9%87%8d%e5%a1%91%e8%b7%a8%e6%a8%a1%e6%80%81%e8%a1%a8%e5%be%81%e5%ad%a6%e4%b9%a0%e7%9a%84%e6%96%87%e6%9c%ac%e5%9f%ba%e7%9f%b3","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/llm2clip%ef%bc%9a%e4%bb%a5%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e9%87%8d%e5%a1%91%e8%b7%a8%e6%a8%a1%e6%80%81%e8%a1%a8%e5%be%81%e5%ad%a6%e4%b9%a0%e7%9a%84%e6%96%87%e6%9c%ac%e5%9f%ba%e7%9f%b3\/","title":{"rendered":"LLM2CLIP\uff1a\u4ee5\u5927\u8bed\u8a00\u6a21\u578b\u91cd\u5851\u8de8\u6a21\u6001\u8868\u5f81\u5b66\u4e60\u7684\u6587\u672c\u57fa\u77f3"},"content":{"rendered":"\n

\u5728\u8de8\u6a21\u6001\u8868\u5f81\u5b66\u4e60\u4e2d\uff0c\u5c06\u4e0d\u540c\u6a21\u6001\u7684\u4fe1\u53f7\u6620\u5c04\u81f3\u7edf\u4e00\u7684\u5171\u4eab\u8868\u793a\u7a7a\u95f4\uff0c\u8fdb\u800c\u9a71\u52a8\u68c0\u7d22\u3001\u7406\u89e3\u4e0e\u751f\u6210\u7b49\u4e00\u7cfb\u5217\u4e0b\u6e38\u4efb\u52a1\uff0c\u662f\u5176\u6838\u5fc3\u76ee\u6807\u3002<\/p>\n\n\n\n

\u800c\u6587\u672c\u5728\u6240\u6709\u6a21\u6001\u4e2d\u5177\u6709\u5929\u7136\u7684\u72ec\u7279\u6027\u3002\u5b83\u4e0d\u4ec5\u662f\u4e00\u79cd\u8f93\u5165\u4fe1\u53f7\uff0c\u66f4\u662f\u4eba\u7c7b\u5bf9\u4e16\u754c\u8fdb\u884c\u7ed3\u6784\u5316\u8ba4\u77e5\u3001\u68b3\u7406\u6982\u5ff5\u4f53\u7cfb\u4ee5\u53ca\u5f00\u653e\u4e16\u754c\u77e5\u8bc6\u7684\u8f7d\u4f53\u3002\u6b63\u56e0\u4e3a\u6b64\uff0c\u6587\u672c\u76d1\u7763\u7684\u8d28\u91cf\uff0c\u5f80\u5f80\u51b3\u5b9a\u4e86\u8de8\u6a21\u6001\u8868\u5f81\u7a7a\u95f4\u7684\u4e0a\u9650\u3002\u66f4\u5f3a\u5927\u7684\u6587\u672c\u76d1\u7763\u4fe1\u53f7\u80fd\u591f\u5f15\u5bfc\u6a21\u578b\u5728\u5bf9\u9f50\u8fc7\u7a0b\u4e2d\uff0c\u5b66\u4e60\u5230\u66f4\u590d\u6742\u3001\u66f4\u7ec6\u7c92\u5ea6\uff0c\u4e5f\u66f4\u5177\u8bed\u4e49\u7ed3\u6784\u7684\u8868\u5f81\u7a7a\u95f4\u3002<\/p>\n\n\n\n

\u5982\u4f55\u7cfb\u7edf\u6027\u5730\u91ca\u653e\u8bed\u8a00\u6240\u8574\u542b\u7684\u77e5\u8bc6\u6f5c\u529b\uff0c\u5e76\u5c06\u5176\u6709\u6548\u8f6c\u5316\u4e3a\u8de8\u6a21\u6001\u8868\u5f81\u80fd\u529b\uff0c\u662f\u8fd1\u5e74\u6765\u591a\u6a21\u6001\u7814\u7a76\u4e2d\u7684\u5173\u952e\u95ee\u9898\u4e4b\u4e00\u3002\u56f4\u7ed5\u8fd9\u4e00\u6311\u6218\uff0c\u5b66\u672f\u754c\u5f00\u59cb\u91cd\u65b0\u5ba1\u89c6\u8bed\u8a00\u6a21\u578b\u5728\u8de8\u6a21\u6001\u5b66\u4e60\u4f53\u7cfb\u4e2d\u7684\u5b9a\u4f4d\u4e0e\u4f5c\u7528\u3002<\/p>\n\n\n\n

\u5728 AAAI 2026 \u5927\u4f1a\u4e0a\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa\u7684 LLM2CLIP \u8363\u83b7\u4e86 Outstanding Paper Award\uff08\u6770\u51fa\u8bba\u6587\u5956\uff09<\/strong>\u3002\u8fd9\u4e00\u6210\u679c\u4e0d\u4ec5\u5728\u5de5\u7a0b\u5b9e\u8df5\u4e0a\u663e\u8457\u7a81\u7834\u4e86 CLIP (Contrastive Language\u2013Image Pretraining\uff0c\u8bed\u8a00-\u56fe\u50cf\u5bf9\u6bd4\u9884\u8bad\u7ec3) \u7684\u957f\u671f\u74f6\u9888\uff0c\u4e5f\u5728\u7406\u8bba\u5c42\u9762\u91cd\u65b0\u5b9a\u4e49\u4e86\u8bed\u8a00\u5728\u8de8\u6a21\u6001\u5b66\u4e60\u4e2d\u7684\u89d2\u8272\u5b9a\u4f4d\u3002<\/p>\n\n\n\n

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

\u56fe1\uff1a\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa\u7684 LLM2CLIP \u8363\u83b7 AAAI 2026 \u6770\u51fa\u8bba\u6587\u5956<\/p>\n\n\n\n

LLM2CLIP: Powerful Language Model Unlocks Richer Cross-Modality Representation<\/strong><\/p>\n\n\n\n

\u8bba\u6587\u94fe\u63a5\uff1a<\/strong>https:\/\/arxiv.org\/pdf\/2411.04997 (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

\u4ee3\u7801&\u6a21\u578b\u94fe\u63a5\uff1a<\/strong>https:\/\/microsoft.github.io\/LLM2CLIP\/ (opens in new tab)<\/span><\/a><\/p>\n\n\n\n

CLIP \u7684\u6210\u529f\u4e0e\u9690\u5fe7\uff1a\u6587\u672c\u80fd\u529b\u6b63\u5728\u6210\u4e3a\u8de8\u6a21\u6001\u7684\u201c\u5929\u82b1\u677f\u201d<\/h2>\n\n\n\n

CLIP \u662f\u53d1\u8868\u4e8e2021 \u5e74\u7684\u4e00\u9879\u91cc\u7a0b\u7891\u5f0f\u7684\u5de5\u4f5c\u3002\u901a\u8fc7\u5728\u6d77\u91cf\u56fe\u6587\u5bf9\u4e0a\u8fdb\u884c\u5bf9\u6bd4\u5b66\u4e60\uff0cCLIP\u9996\u6b21\u7cfb\u7edf\u6027\u5730\u6784\u5efa\u4e86\u7edf\u4e00\u7684\u6587\u672c\u2013\u89c6\u89c9\u8de8\u6a21\u6001\u8bed\u4e49\u7a7a\u95f4\uff0c\u5960\u5b9a\u4e86\u5176\u4f5c\u4e3a\u591a\u6a21\u6001\u57fa\u7840\u6a21\u578b\u7684\u5730\u4f4d\uff0c\u5e76\u88ab\u5e7f\u6cdb\u7528\u4e8e\u56fe\u50cf\u7406\u89e3\u3001\u8de8\u6a21\u6001\u68c0\u7d22\u3001\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08VLLM\uff09\u4ee5\u53ca\u6587\u751f\u56fe\u6269\u6563\u6a21\u578b\u7b49\u524d\u6cbf\u7cfb\u7edf\u4e2d\u3002<\/p>\n\n\n\n

\u7136\u800c\uff0c\u968f\u7740\u5e94\u7528\u573a\u666f\u4e0d\u65ad\u62d3\u5c55\uff0cCLIP \u8303\u5f0f\u4e2d\u4e00\u4e2a\u6df1\u5c42\u9650\u5236\u9010\u6e10\u663e\u73b0\u4e86\u51fa\u6765\u2014\u2014\u6587\u672c\u80fd\u529b\u7684\u6ede\u540e\uff0c\u6210\u4e3a\u8de8\u6a21\u6001\u7cfb\u7edf\u6027\u80fd\u8fdb\u4e00\u6b65\u63d0\u5347\u7684\u201c\u5929\u82b1\u677f\u201d\u3002<\/p>\n\n\n\n

\u8fd9\u4e00\u74f6\u9888\u4e3b\u8981\u4f53\u73b0\u5728\u4ee5\u4e0b\u4e09\u4e2a\u65b9\u9762\uff1a<\/p>\n\n\n\n

    \n
  • \u6587\u672c\u4e0a\u4e0b\u6587\u7a97\u53e3\u6781\u7a84\uff1a<\/strong><\/li>\n<\/ul>\n\n\n\n

    CLIP \u7684\u6587\u672c\u7f16\u7801\u5668\u901a\u5e38\u4ec5\u652f\u6301 77 \u4e2a token\uff0c\u8d85\u51fa\u7684\u90e8\u5206\u4f1a\u88ab\u622a\u65ad\uff0c\u8fd9\u5bfc\u81f4\u5b83\u5728\u5904\u7406\u771f\u5b9e\u4e16\u754c\u4e2d\u5e38\u89c1\u7684\u957f\u6587\u672c\u3001\u590d\u6742\u63cf\u8ff0\u548c\u591a\u5b9e\u4f53\u5173\u7cfb\u65f6\uff0c\u5f80\u5f80\u663e\u5f97\u529b\u4e0d\u4ece\u5fc3\u3002<\/p>\n\n\n\n

      \n
    • \u6587\u672c\u7f16\u7801\u5668\u5bb9\u91cf\u6709\u9650<\/strong><\/li>\n<\/ul>\n\n\n\n

      CLIP \u7684 \u6587\u672c\u7f16\u7801\u5668\uff08text encoder\uff09 \u53c2\u6570\u91cf\u901a\u5e38\u53ea\u6709\u89c6\u89c9\u7f16\u7801\u5668\u7684\u5341\u5206\u4e4b\u4e00\u5de6\u53f3\u3002\u7531\u4e8e\u6587\u672c\u7f16\u7801\u5668\u662f\u4e0e\u89c6\u89c9\u7f16\u7801\u5668\u540c\u6b65\u4ece\u96f6\u5f00\u59cb\u8bad\u7ec3\u7684\uff0c\u6240\u4ee5\u8fd9\u79cd\u6709\u9650\u7684\u5bb9\u91cf\u4f7f\u5176\u96be\u4ee5\u5145\u5206\u5438\u6536\u4e30\u5bcc\u7684\u8bed\u8a00\u77e5\u8bc6\u3002<\/p>\n\n\n\n

        \n
      • \u6587\u672c\u5148\u9a8c\u5229\u7528\u6548\u7387\u4f4e\u4e0b<\/strong><\/li>\n<\/ul>\n\n\n\n

        \u4e0e\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u76f8\u6bd4\uff0cCLIP \u7684\u6587\u672c\u5efa\u6a21\u80fd\u529b\u5728\u8bed\u4e49\u7406\u89e3\u3001\u77e5\u8bc6\u8986\u76d6\u548c\u63a8\u7406\u80fd\u529b\u4e0a\u5747\u5b58\u5728\u663e\u8457\u5dee\u8ddd\u3002\u7531\u4e8e\u7f3a\u4e4f\u6df1\u539a\u7684\u8bed\u8a00\u5148\u9a8c\uff0c\u6a21\u578b\u5728\u7406\u89e3\u590d\u6742\u6307\u4ee4\u548c\u62bd\u8c61\u6982\u5ff5\u65f6\u8868\u73b0\u4e0d\u4f73\u3002<\/p>\n\n\n\n

        \u76f4\u89c2\u6765\u770b\uff0c\u66f4\u5f3a\u7684\u6587\u672c\u7406\u89e3\u80fd\u529b\u7406\u5e94\u5e26\u6765\u66f4\u901a\u7528\u7684\u8de8\u6a21\u6001\u8868\u5f81\u3002\u4f46\u8fd9\u4e2d\u95f4\u7684\u771f\u6b63\u6311\u6218\u5728\u4e8e\uff1a\u5982\u4f55\u5728\u4e0d\u91cd\u65b0\u8fdb\u884c\u6602\u8d35\u7684\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u7684\u524d\u63d0\u4e0b\uff0c\u628a\u5927\u8bed\u8a00\u6a21\u578b\u7684\u80fd\u529b\u9ad8\u6548\u5730\u5f15\u5165\u5230 CLIP\u4e4b\u4e2d\uff1f<\/p>\n\n\n\n

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

        \u56fe2\uff1aLLM2CLIP \u65b9\u6cd5\u6846\u67b6\u793a\u610f\u56fe\uff0c\u901a\u8fc7\u5f15\u5165\u7ecf Caption-Contrastive \u5fae\u8c03\u7684\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u4e3a CLIP \u6784\u5efa\u66f4\u5177\u5224\u522b\u6027\u7684\u6587\u672c\u8868\u5f81\u7a7a\u95f4<\/p>\n\n\n\n

        \u4e00\u4e2a\u770b\u4f3c\u76f4\u63a5\u7684\u601d\u8def\u662f\uff0c\u7528\u5927\u8bed\u8a00\u6a21\u578b\u76f4\u63a5\u66ff\u6362 CLIP \u7684\u6587\u672c\u7f16\u7801\u5668\u3002\u7136\u800c\uff0c\u5b9e\u9a8c\u7ed3\u679c\u5374\u51e0\u4e4e\u4e00\u81f4\u5730\u6307\u5411\u5931\u8d25\u3002\u6839\u672c\u539f\u56e0\u5728\u4e8e\uff0cLLMs \u672c\u8d28\u4e0a\u662f Decoder-only \u7684\u751f\u6210\u6a21\u578b\uff0c\u5176\u8f93\u51fa\u7279\u5f81\u5e76\u975e\u4e3a\u5ea6\u91cf\u5b66\u4e60\uff08Metric Learning\uff09\u800c\u8bbe\u8ba1\u7684\u3002<\/p>\n\n\n\n

        \u5728\u8de8\u6a21\u6001\u5bf9\u9f50\u4efb\u52a1\u4e2d\uff0c\u539f\u59cb LLMs \u751f\u6210\u7684\u6587\u672c\u7279\u5f81\u96be\u4ee5\u6e05\u6670\u5730\u533a\u5206\u8bed\u4e49\u76f8\u8fd1\u4f46\u6307\u5411\u4e0d\u540c\u56fe\u50cf\u7684\u63cf\u8ff0\uff0c\u8fd9\u79cd\u201c\u6a21\u7cca\u8868\u5f81\u201d\u4f1a\u76f4\u63a5\u7834\u574f\u56fe\u6587\u5bf9\u9f50\u6240\u9700\u7684\u5224\u522b\u6027\u7ed3\u6784\uff0c\u5bfc\u81f4\u8bad\u7ec3\u6548\u679c\u65ad\u5d16\u5f0f\u4e0b\u8dcc\u3002\u8fd9\u63ed\u793a\u4e86\u4e00\u4e2a\u5173\u952e\u4e8b\u5b9e\uff0cLLMs \u5e76\u975e\u5929\u7136\u9002\u914d\u8de8\u6a21\u6001\u5bf9\u9f50\u7684\u5d4c\u5165\uff08embedding\uff09 \u6a21\u578b\u3002<\/p>\n\n\n\n

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

        \u56fe3\uff1aCaption-Contrastive\u5fae\u8c03\u663e\u8457\u63d0\u5347\u4e86 LLMs \u5bf9\u76f8\u4f3c\u56fe\u50cf\u63cf\u8ff0\u7684\u533a\u5206\u80fd\u529b\uff0c\u4e3a\u9ad8\u8d28\u91cf\u56fe\u6587\u5bf9\u9f50\u5960\u5b9a\u4e86\u57fa\u7840<\/p>\n\n\n\n

        \u5148\u8d4b\u4e88 LLMs \u201c\u5224\u522b\u6027\u201d\uff0c\u518d\u8fdb\u884c\u8de8\u6a21\u6001\u5bf9\u9f50<\/h2>\n\n\n\n

        \u9488\u5bf9\u4e0a\u8ff0\u74f6\u9888\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86 LLM2CLIP\uff0c\u65e8\u5728\u901a\u8fc7\u5c06\u5927\u8bed\u8a00\u6a21\u578b\u5f15\u5165 CLIP \u8bad\u7ec3\u6d41\u7a0b\u4e2d\uff0c\u4ece\u800c\u7cfb\u7edf\u6027\u63d0\u5347\u8de8\u6a21\u6001\u8868\u5f81\u80fd\u529b\u3002AAAI 2026\u8bc4\u5ba1\u5bf9LLM2CLIP\u7ed9\u51fa\u9ad8\u5ea6\u8bc4\u4ef7\u7684\u539f\u56e0\u4e4b\u4e00\u5c31\u662fLLM2CLIP\u5e76\u672a\u505c\u7559\u5728\u201c\u66f4\u5927\u7684\u6a21\u578b\u201d\u4e0a\uff0c\u800c\u662f\u7cbe\u51c6\u51fb\u4e2d\u4e86\u8de8\u6a21\u6001\u5b66\u4e60\u4e2d\u4e00\u4e2a\u957f\u671f\u88ab\u5ffd\u89c6\u7684\u7ed3\u6784\u6027\u95ee\u9898\u3002<\/p>\n\n\n\n

        \u7814\u7a76\u5458\u4eec\u4e3aLLM2CLIP \u8bbe\u8ba1\u4e86\u4e00\u5957\u7cbe\u51c6\u7684\u4e24\u9636\u6bb5\u8bad\u7ec3\u7b56\u7565\uff1a<\/p>\n\n\n\n

        \u7b2c\u4e00\u9636\u6bb5\uff1aCaption-to-Caption \u5bf9\u6bd4\u5b66\u4e60\uff0c\u91cd\u5851\u6587\u672c\u5d4c\u5165\u7a7a\u95f4<\/strong><\/p>\n\n\n\n

        \u5728\u56fe\u6587\u5bf9\u9f50\u4e4b\u524d\uff0c\u7814\u7a76\u5458\u4eec\u9996\u5148\u5bf9 LLMs \u8fdb\u884c\u4e86 caption-to-caption \u7684\u5bf9\u6bd4\u5b66\u4e60\u5fae\u8c03\u3002\u901a\u8fc7\u5f15\u5165\u9ad8\u8d28\u91cf\u7684\u56fe\u50cf\u63cf\u8ff0 \uff08caption\uff09 \u6570\u636e\uff0c\u5f15\u5bfc LLMs \u5728\u5d4c\u5165\u7a7a\u95f4\u4e2d\u7cbe\u51c6\u6355\u6349\u8bed\u4e49\u95f4\u7684\u7ec6\u5fae\u5dee\u522b\uff0c\u4ece\u800c\u6709\u6548\u533a\u5206\u90a3\u4e9b\u8bed\u4e49\u76f8\u8fd1\u4f46\u6307\u5411\u4e0d\u540c\u56fe\u50cf\u7684\u6587\u672c\u63cf\u8ff0\u3002<\/p>\n\n\n\n

        \u8fd9\u4e00\u9636\u6bb5\u7684\u8bad\u7ec3\u81f3\u5173\u91cd\u8981\u3002\u5b83\u663e\u8457\u589e\u5f3a\u4e86 LLMs \u5728 caption \u68c0\u7d22\u4efb\u52a1\u4e2d\u7684\u5224\u522b\u80fd\u529b\u3002\u5b9e\u9a8c\u8bc1\u5b9e\uff0c\u82e5\u8df3\u8fc7\u6b64\u6b65\u9aa4\u76f4\u63a5\u4f7f\u7528 LLMs \u8fdb\u884c CLIP \u5fae\u8c03\uff0c\u6a21\u578b\u5c06\u56e0\u65e0\u6cd5\u6536\u655b\u5230\u6709\u6548\u7684\u5224\u522b\u7a7a\u95f4\u800c\u5bfc\u81f4\u6027\u80fd\u4e25\u91cd\u9000\u5316\u3002<\/p>\n\n\n\n

        \u8fd9\u4e00\u6b65\u6210\u529f\u9a71\u52a8\u4e86 LLM s\u4ece\u201c\u751f\u6210\u6a21\u578b\u201d\u8f6c\u53d8\u4e3a\u4e00\u4e2a\u9002\u5408\u8de8\u6a21\u6001\u5b66\u4e60\u7684\u6587\u672c\u5d4c\u5165\u6a21\u578b\uff0c\u4e3a\u540e\u7eed\u7684\u56fe\u6587\u5bf9\u9f50\u592f\u5b9e\u4e86\u6587\u672c\u5e95\u5ea7\u3002<\/p>\n\n\n\n

        \u7b2c\u4e8c\u9636\u6bb5\uff1a\u9ad8\u6548\u7684 LLM\u2013\u89c6\u89c9\u7f16\u7801\u5668\u5bf9\u9f50<\/strong><\/p>\n\n\n\n

        \u83b7\u5f97\u7ecf\u8fc7 caption \u5bf9\u6bd4\u5b66\u4e60\u7684 LLMs \u540e\uff0c\u7814\u7a76\u5458\u4eec\u5c06\u5176\u4e0e CLIP \u7684\u89c6\u89c9\u7f16\u7801\u5668\u91cd\u65b0\u8fdb\u884c\u8de8\u6a21\u6001\u5bf9\u9f50\uff1a<\/p>\n\n\n\n

          \n
        • \u91c7\u7528\u4e0e\u539f\u59cb CLIP \u76f8\u540c\u7684\u5bf9\u6bd4\u5b66\u4e60\u76ee\u6807\uff08InfoNCE \/ Sigmoid Loss\uff09<\/li>\n\n\n\n
        • \u5f7b\u5e95\u629b\u5f03\u539f\u6709\u7684 CLIP \u6587\u672c\u7f16\u7801\u5668<\/li>\n\n\n\n
        • \u51bb\u7ed3 LLMs \u4e3b\u4f53\uff0c\u4ec5\u901a\u8fc7\u4e00\u4e2a\u8f7b\u91cf\u9002\u914d\u5668\uff08 adaptor\uff09\u6a21\u5757\u5b9e\u73b0\u6587\u672c\u7279\u5f81\u5411\u8de8\u6a21\u6001\u7a7a\u95f4\u7684\u6620\u5c04<\/li>\n\n\n\n
        • \u5bf9\u89c6\u89c9\u7f16\u7801\u5668\u8fdb\u884c\u5fae\u8c03<\/li>\n<\/ul>\n\n\n\n

          \u4e24\u9636\u6bb5\u7684\u8bbe\u8ba1\u4e3a LLM2CLIP \u5e26\u6765\u4e86\u663e\u8457\u7684\u5de5\u7a0b\u4f18\u52bf\u3002\u7531\u4e8e LLMs \u53ef\u4ee5\u5728\u8ba1\u7b97\u8fc7\u7a0b\u4e2d\u5b8c\u5168 offload\uff08\u5378\u8f7d\uff09\uff0c\u6240\u4ee5\u6574\u4f53\u8bad\u7ec3\u7684\u663e\u5b58\u548c\u8ba1\u7b97\u5f00\u9500\u88ab\u538b\u4f4e\u81f3\u4e0e\u539f\u59cb CLIP \u5fae\u8c03\u51e0\u4e4e\u76f8\u5f53\u7684\u6c34\u5e73<\/strong>\u3002<\/p>\n\n\n\n

          \u4ee5\u642d\u8f7d 8B \u53c2\u6570\u7684\u5927\u8bed\u8a00\u6a21\u578b\u7684\u5b9e\u9a8c\u914d\u7f6e\u4e3a\u4f8b\u3002LLM2CLIP\u5b9e\u73b0\u4e8613 \u500d\u7684\u8bad\u7ec3\u901f\u5ea6\u98de\u8dc3\uff0c\u663e\u5b58\u5360\u7528\u5927\u5e45\u7f29\u51cf\u4e3a\u539f\u6765\u7684 1\/20\uff0c\u800c\u4e14\u8fd8\u53ef\u4ee5\u5728\u6709\u9650\u8bbe\u5907\u4e0a\u90e8\u7f72\u66f4\u5927\u89c4\u6a21\u7684\u6279\u91cf\u5904\u7406\uff08 Batch Size\uff09\uff0c\u8fd9\u79cd\u89c4\u6a21\u6548\u5e94\u53cd\u800c\u53cd\u54fa\u4e86\u6a21\u578b\u7684\u6027\u80fd\u3002<\/p>\n\n\n\n

          \u5c11\u91cf\u6570\u636e\uff0c\u663e\u8457\u6536\u76ca\uff1a\u6253\u7834\u5bf9\u201c\u5927\u89c4\u6a21\u9884\u8bad\u7ec3\u201d\u7684\u8def\u5f84\u4f9d\u8d56<\/h2>\n\n\n\n

          \u4f20\u7edf CLIP \u8303\u5f0f\u9ad8\u5ea6\u4f9d\u8d56\u4e8e\u6d77\u91cf\u6570\u636e\uff0c\u5176\u6587\u672c\u7f16\u7801\u5668\u5f80\u5f80\u9700\u8981\u7ecf\u5386\u7ea6 40B \u7ea7\u522b\u56fe\u6587\u5bf9\u7684\u9884\u8bad\u7ec3\u3002\u7136\u800c\uff0c\u5728\u5f15\u5165 LLMs \u540e\uff0c\u7814\u7a76\u5458\u4eec\u53d1\u73b0\uff0c\u4ec5\u9700\u4f7f\u7528 3\u201315M \u7684\u56fe\u6587\u6570\u636e\u91cd\u65b0\u5bf9\u9f50\uff0c\u5373\u53ef\u5728\u6027\u80fd\u4e0a\u5168\u9762\u8d85\u8d8a\u539f\u59cb\u6a21\u578b\u3002<\/p>\n\n\n\n

          \u5728\u4e0e\u5de5\u4e1a\u7ea7 SOTA \u6a21\u578b SigLIP-2 \u7684\u5bf9\u6bd4\u5b9e\u9a8c\u4e2d\uff0cLLM2CLIP\u5c55\u73b0\u51fa\u4e86\u663e\u8457\u7684\u589e\u957f\u6f5c\u529b\uff1a<\/p>\n\n\n\n

            \n
          • \u82f1\u6587\u77ed\u6587\u672c\u68c0\u7d22\uff1a\u7a33\u6b65\u63d0\u5347 +1.0 \/ +1.9<\/li>\n\n\n\n
          • \u82f1\u6587\u957f\u6587\u672c\u68c0\u7d22\uff1a\u5b9e\u73b0 +14.8 \/ +15.8 \u7684\u7a81\u7834\u6027\u8de8\u8d8a\uff0c\u89e3\u51b3\u4e86 CLIP \u7684\u201c\u957f\u6587\u672c\u6050\u60e7\u75c7\u201d\u3002<\/li>\n\n\n\n
          • \u4e2d\u6587 COCO-CN\uff1a\u63d0\u5347 +15.5 \/ +18.5%<\/li>\n\n\n\n
          • XM3600\uff0836 \u79cd\u8bed\u8a00\uff09\uff1a\u5747\u5206\u63d0\u5347 +11.9 \/ +15.2<\/li>\n<\/ul>\n\n\n\n

            \u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cLLM2CLIP \u7684\u8de8\u6a21\u6001\u5bf9\u9f50\u8fc7\u7a0b\u5b8c\u5168\u57fa\u4e8e\u82f1\u6587\u6570\u636e\uff0c\u4f46\u5176\u5c55\u73b0\u51fa\u7684\u591a\u8bed\u8a00\u80fd\u529b\u5374\u5b9e\u73b0\u5927\u5e45\u98de\u8dc3\uff0c\u8fd9\u4f53\u73b0\u4e86 LLMs \u8bed\u8a00\u5148\u9a8c\u7684\u5f3a\u6cdb\u5316\u6027\u3002<\/p>\n\n\n\n

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

            \u56fe4\uff1aLLM2CLIP \u5728\u591a\u9879\u56fe\u6587\u68c0\u7d22\u57fa\u51c6\u4e0a\u7a33\u5b9a\u63d0\u5347\u6027\u80fd\uff0c\u5305\u62ec\u5728 SigLIP2 \u7b49\u5de5\u4e1a\u7ea7\u8de8\u6a21\u6001\u6a21\u578b\u4e0a\u7684\u8868\u73b0\u3002<\/p>\n\n\n\n

            \u4e0d\u6b62\u4e8e\u8de8\u6a21\u6001\uff1a\u89c6\u89c9\u7f16\u7801\u5668\u7684\u6df1\u5ea6\u8fdb\u5316<\/h2>\n\n\n\n

            \u5b9e\u9a8c\u6570\u636e\u8868\u660e\uff0c\u5f15\u5165 LLMs \u540e\uff0c\u6587\u672c\u76d1\u7763\u4fe1\u53f7\u53d8\u5f97\u66f4\u52a0\u7ec6\u7c92\u5ea6\u548c\u7ed3\u6784\u5316\uff0c\u53cd\u5411\u9a71\u52a8\u4e86\u89c6\u89c9\u7f16\u7801\u5668\u7684\u81ea\u6211\u8fdb\u5316\uff1a<\/p>\n\n\n\n

              \n
            • \u5728\u96f6\u6837\u672c\uff08zero-shot\uff09\u53ca\u6709\u76d1\u7763\u8bbe\u7f6e\u4e0b\u7684\u76ee\u6807\u68c0\u6d4b\u4e0e\u56fe\u50cf\u5206\u5272\u4efb\u52a1\u4e0a\uff0c\u5747\u53d6\u5f97\u4e86\u7a33\u5065\u7684\u6027\u80fd\u589e\u957f\uff1b<\/li>\n\n\n\n
            • \u5728 ImageNet linear probing \u6d4b\u8bd5\u4e2d\u8868\u73b0\u66f4\u4f18\uff1b<\/li>\n\n\n\n
            • \u5728\u5b8c\u5168\u5265\u79bb\u6587\u672c\u7f16\u7801\u5668\u7684\u7eaf\u89c6\u89c9\u573a\u666f\u4e0b\uff0c\u89c6\u89c9\u8868\u5f81\u80fd\u529b\u4e5f\u663e\u8457\u589e\u5f3a\u3002<\/li>\n<\/ul>\n\n\n\n

              \u4e3a\u4e86\u8fdb\u4e00\u6b65\u9a8c\u8bc1LLM2CLIP\u7684\u6f5c\u529b\uff0c\u7814\u7a76\u5458\u4eec\u5c06  LLaVA-1.5 \u4e2d\u7684\u539f\u59cb\u89c6\u89c9\u7f16\u7801\u5668\u66ff\u6362\u6210\u4e86 LLM2CLIP \u8bad\u7ec3\u540e\u7684 ViT-L \u7248\u672c\u3002\u5728\u8d85\u8fc7 87.5% \u7684\u591a\u6a21\u6001\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u6a21\u578b\u6027\u80fd\u5b9e\u73b0\u4e86\u5168\u9762\u8dc3\u5347\uff0c\u4ec5\u5728\u6781\u5c11\u6570\u4efb\u52a1\u4e0a\u6709\u8f7b\u5fae\u4e0b\u964d\u3002<\/strong>\u8fd9\u8868\u660e LLM2CLIP \u4e0d\u4ec5\u91cd\u5851\u4e86\u8de8\u6a21\u6001\u7a7a\u95f4\uff0c\u66f4\u4e3a\u590d\u6742\u7684\u89c6\u89c9\u7406\u89e3\u4e0e\u63a8\u7406\u4efb\u52a1\u5960\u5b9a\u4e86\u4e00\u4e2a\u66f4\u52a0\u575a\u56fa\u7684\u89c6\u89c9\u6839\u57fa\u3002<\/p>\n\n\n\n

              LLM2CLIP \u7684\u4ef7\u503c\u5e76\u4e0d\u53ea\u4f53\u73b0\u5728\u6027\u80fd\u7684\u6570\u5b57\u4e0a\uff0c\u8fd8\u6709\u5176\u65b9\u6cd5\u8bba\u7684\u610f\u4e49\uff1a<\/p>\n\n\n\n

                \n
              • \u5b83\u4ece\u5e95\u5c42\u89e3\u51b3\u4e86 CLIP \u957f\u671f\u5b58\u5728\u7684\u6587\u672c\u5efa\u6a21\u77ed\u677f<\/strong>\uff0c\u5f7b\u5e95\u6253\u7834\u4e86\u4e0a\u4e0b\u6587\u957f\u5ea6\u548c\u6587\u672c\u7406\u89e3\u80fd\u529b\u7684\u74f6\u9888\uff1b<\/li>\n\n\n\n
              • \u5b83\u63d0\u51fa\u4e86\u4e00\u79cd\u9ad8\u6548\u3001\u53ef\u6269\u5c55\u4e14\u5de5\u7a0b\u53cb\u597d\u7684 LLMs \u5f15\u5165\u65b9\u5f0f<\/strong>\uff0c\u5728\u51e0\u4e4e\u4e0d\u589e\u52a0\u8bad\u7ec3\u6210\u672c\u7684\u524d\u63d0\u4e0b\uff0c\u5b9e\u73b0\u4e86\u6a21\u578b\u8de8\u6a21\u6001\u6027\u80fd\u7684\u8dc3\u5347\uff1b<\/li>\n\n\n\n
              • \u5b83\u5168\u9762\u589e\u5f3a\u4e86<\/strong>\u5305\u62ec\u5de5\u4e1a\u7ea7 SOTA\uff08\u5982EVA02\uff0c SigLIP-2\uff09\u5728\u5185\u7684 CLIP \u7cfb\u5217\u6a21\u578b<\/strong>\uff0c\u5176\u5f71\u54cd\u529b\u8986\u76d6\u4e86\u8de8\u6a21\u6001\u68c0\u7d22\u3001\u591a\u8bed\u8a00\u7406\u89e3\u548c\u89c6\u89c9\u4efb\u52a1\u7b49\u573a\u666f\u3002<\/li>\n<\/ul>\n\n\n\n

                \u7814\u7a76\u5458\u4eec\u8ba4\u4e3a\uff0c\u8bed\u8a00\u4e0d\u5e94\u4ec5\u88ab\u89c6\u4f5c\u591a\u6a21\u6001\u7cfb\u7edf\u7684\u4e00\u4e2a\u8f93\u5165\u6a21\u6001\uff0c\u800c\u5e94\u8fdb\u9636\u4e3a\u5f15\u5bfc\u8de8\u6a21\u6001\u8868\u5f81\u5b66\u4e60\u7684\u201c\u77e5\u8bc6\u4e2d\u67a2\u201d<\/strong>\u3002<\/p>\n\n\n\n

                \u901a\u8fc7\u66f4\u6df1\u5c42\u5730\u6316\u6398\u548c\u5f15\u5165\u4eba\u7c7b\u8bed\u8a00\u4e2d\u7684\u5148\u9a8c\u77e5\u8bc6\uff0c\u672a\u6765\u7684\u591a\u6a21\u6001\u6a21\u578b\u5c06\u80fd\u591f\u5efa\u7acb\u8d77\u5bf9\u771f\u5b9e\u4e16\u754c\u66f4\u6df1\u9083\u7684\u6d1e\u5bdf\uff0c\u4ece\u800c\u652f\u6491\u8d77\u66f4\u590d\u6742\u3001\u66f4\u591a\u6837\u3001\u5177\u6709\u66f4\u9ad8\u4ef7\u503c\u7684\u5e94\u7528\u573a\u666f\u3002<\/p>\n\n\n\n

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

                \u5728\u8de8\u6a21\u6001\u8868\u5f81\u5b66\u4e60\u4e2d\uff0c\u5c06\u4e0d\u540c\u6a21\u6001\u7684\u4fe1\u53f7\u6620\u5c04\u81f3\u7edf\u4e00\u7684\u5171\u4eab\u8868\u793a\u7a7a\u95f4\uff0c\u8fdb\u800c\u9a71\u52a8\u68c0\u7d22\u3001\u7406\u89e3\u4e0e\u751f\u6210\u7b49\u4e00\u7cfb\u5217\u4e0b\u6e38\u4efb\u52a1\uff0c\u662f\u5176\u6838\u5fc3\u76ee\u6807\u3002 \u800c\u6587\u672c\u5728\u6240\u6709\u6a21\u6001\u4e2d\u5177\u6709\u5929\u7136\u7684\u72ec\u7279\u6027\u3002\u5b83\u4e0d\u4ec5\u662f\u4e00\u79cd\u8f93\u5165\u4fe1\u53f7\uff0c\u66f4\u662f\u4eba\u7c7b\u5bf9\u4e16\u754c\u8fdb\u884c\u7ed3\u6784\u5316\u8ba4\u77e5\u3001\u68b3\u7406\u6982\u5ff5\u4f53\u7cfb\u4ee5\u53ca\u5f00\u653e\u4e16\u754c\u77e5\u8bc6\u7684\u8f7d\u4f53\u3002\u6b63\u56e0\u4e3a\u6b64\uff0c\u6587\u672c\u76d1\u7763\u7684\u8d28\u91cf\uff0c\u5f80\u5f80\u51b3\u5b9a\u4e86\u8de8\u6a21\u6001\u8868\u5f81\u7a7a\u95f4\u7684\u4e0a\u9650\u3002\u66f4\u5f3a\u5927\u7684\u6587\u672c\u76d1\u7763\u4fe1\u53f7\u80fd\u591f\u5f15\u5bfc\u6a21\u578b\u5728\u5bf9\u9f50\u8fc7\u7a0b\u4e2d\uff0c\u5b66\u4e60\u5230\u66f4\u590d\u6742\u3001\u66f4\u7ec6\u7c92\u5ea6\uff0c\u4e5f\u66f4\u5177\u8bed\u4e49\u7ed3\u6784\u7684\u8868\u5f81\u7a7a\u95f4\u3002 \u5982\u4f55\u7cfb\u7edf\u6027\u5730\u91ca\u653e\u8bed\u8a00\u6240\u8574\u542b\u7684\u77e5\u8bc6\u6f5c\u529b\uff0c\u5e76\u5c06\u5176\u6709\u6548\u8f6c\u5316\u4e3a\u8de8\u6a21\u6001\u8868\u5f81\u80fd\u529b\uff0c\u662f\u8fd1\u5e74\u6765\u591a\u6a21\u6001\u7814\u7a76\u4e2d\u7684\u5173\u952e\u95ee\u9898\u4e4b\u4e00\u3002\u56f4\u7ed5\u8fd9\u4e00\u6311\u6218\uff0c\u5b66\u672f\u754c\u5f00\u59cb\u91cd\u65b0\u5ba1\u89c6\u8bed\u8a00\u6a21\u578b\u5728\u8de8\u6a21\u6001\u5b66\u4e60\u4f53\u7cfb\u4e2d\u7684\u5b9a\u4f4d\u4e0e\u4f5c\u7528\u3002 \u5728 AAAI 2026 \u5927\u4f1a\u4e0a\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa\u7684 LLM2CLIP \u8363\u83b7\u4e86 Outstanding Paper Award\uff08\u6770\u51fa\u8bba\u6587\u5956\uff09\u3002\u8fd9\u4e00\u6210\u679c\u4e0d\u4ec5\u5728\u5de5\u7a0b\u5b9e\u8df5\u4e0a\u663e\u8457\u7a81\u7834\u4e86 CLIP (Contrastive Language\u2013Image Pretraining\uff0c\u8bed\u8a00-\u56fe\u50cf\u5bf9\u6bd4\u9884\u8bad\u7ec3) \u7684\u957f\u671f\u74f6\u9888\uff0c\u4e5f\u5728\u7406\u8bba\u5c42\u9762\u91cd\u65b0\u5b9a\u4e49\u4e86\u8bed\u8a00\u5728\u8de8\u6a21\u6001\u5b66\u4e60\u4e2d\u7684\u89d2\u8272\u5b9a\u4f4d\u3002 \u56fe1\uff1a\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u63d0\u51fa\u7684 LLM2CLIP \u8363\u83b7 AAAI 2026 \u6770\u51fa\u8bba\u6587\u5956 LLM2CLIP: Powerful Language Model Unlocks Richer Cross-Modality Representation \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2411.04997 (opens in new tab) \u4ee3\u7801&\u6a21\u578b\u94fe\u63a5\uff1ahttps:\/\/microsoft.github.io\/LLM2CLIP\/ (opens in new tab) CLIP \u662f\u53d1\u8868\u4e8e2021 \u5e74\u7684\u4e00\u9879\u91cc\u7a0b\u7891\u5f0f\u7684\u5de5\u4f5c\u3002\u901a\u8fc7\u5728\u6d77\u91cf\u56fe\u6587\u5bf9\u4e0a\u8fdb\u884c\u5bf9\u6bd4\u5b66\u4e60\uff0cCLIP\u9996\u6b21\u7cfb\u7edf\u6027\u5730\u6784\u5efa\u4e86\u7edf\u4e00\u7684\u6587\u672c\u2013\u89c6\u89c9\u8de8\u6a21\u6001\u8bed\u4e49\u7a7a\u95f4\uff0c\u5960\u5b9a\u4e86\u5176\u4f5c\u4e3a\u591a\u6a21\u6001\u57fa\u7840\u6a21\u578b\u7684\u5730\u4f4d\uff0c\u5e76\u88ab\u5e7f\u6cdb\u7528\u4e8e\u56fe\u50cf\u7406\u89e3\u3001\u8de8\u6a21\u6001\u68c0\u7d22\u3001\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08VLLM\uff09\u4ee5\u53ca\u6587\u751f\u56fe\u6269\u6563\u6a21\u578b\u7b49\u524d\u6cbf\u7cfb\u7edf\u4e2d\u3002 \u7136\u800c\uff0c\u968f\u7740\u5e94\u7528\u573a\u666f\u4e0d\u65ad\u62d3\u5c55\uff0cCLIP \u8303\u5f0f\u4e2d\u4e00\u4e2a\u6df1\u5c42\u9650\u5236\u9010\u6e10\u663e\u73b0\u4e86\u51fa\u6765\u2014\u2014\u6587\u672c\u80fd\u529b\u7684\u6ede\u540e\uff0c\u6210\u4e3a\u8de8\u6a21\u6001\u7cfb\u7edf\u6027\u80fd\u8fdb\u4e00\u6b65\u63d0\u5347\u7684\u201c\u5929\u82b1\u677f\u201d\u3002 \u8fd9\u4e00\u74f6\u9888\u4e3b\u8981\u4f53\u73b0\u5728\u4ee5\u4e0b\u4e09\u4e2a\u65b9\u9762\uff1a CLIP \u7684\u6587\u672c\u7f16\u7801\u5668\u901a\u5e38\u4ec5\u652f\u6301 77 \u4e2a token\uff0c\u8d85\u51fa\u7684\u90e8\u5206\u4f1a\u88ab\u622a\u65ad\uff0c\u8fd9\u5bfc\u81f4\u5b83\u5728\u5904\u7406\u771f\u5b9e\u4e16\u754c\u4e2d\u5e38\u89c1\u7684\u957f\u6587\u672c\u3001\u590d\u6742\u63cf\u8ff0\u548c\u591a\u5b9e\u4f53\u5173\u7cfb\u65f6\uff0c\u5f80\u5f80\u663e\u5f97\u529b\u4e0d\u4ece\u5fc3\u3002 CLIP \u7684 \u6587\u672c\u7f16\u7801\u5668\uff08text […]<\/p>\n","protected":false},"author":32169,"featured_media":1164000,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[13556],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1163619","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-research-area-artificial-intelligence","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1163619","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/users\/32169"}],"version-history":[{"count":4,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1163619\/revisions"}],"predecessor-version":[{"id":1164001,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1163619\/revisions\/1164001"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1164000"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1163619"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1163619"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1163619"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1163619"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}