{"id":39475,"date":"2025-03-18T01:02:29","date_gmt":"2025-03-17T17:02:29","guid":{"rendered":"https:\/\/17aitech.com\/?p=39475"},"modified":"2025-03-18T01:02:29","modified_gmt":"2025-03-17T17:02:29","slug":"%e5%8c%97%e5%a4%a7%e5%9b%a2%e9%98%9f%e6%8f%90%e5%87%balift%ef%bc%9a%e5%b0%86%e9%95%bf%e4%b8%8a%e4%b8%8b%e6%96%87%e7%9f%a5%e8%af%86%e6%b3%a8%e5%85%a5%e6%a8%a1%e5%9e%8b%e5%8f%82%e6%95%b0%ef%bc%8c","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=39475","title":{"rendered":"\u5317\u5927\u56e2\u961f\u63d0\u51faLIFT\uff1a\u5c06\u957f\u4e0a\u4e0b\u6587\u77e5\u8bc6\u6ce8\u5165\u6a21\u578b\u53c2\u6570\uff0c\u63d0\u5347\u5927\u6a21\u578b\u957f\u6587\u672c\u80fd\u529b"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-17-4\" target=\"_blank\">\u5317\u5927\u56e2\u961f\u63d0\u51faLIFT\uff1a\u5c06\u957f\u4e0a\u4e0b\u6587\u77e5\u8bc6\u6ce8\u5165\u6a21\u578b\u53c2\u6570\uff0c\u63d0\u5347\u5927\u6a21\u578b\u957f\u6587\u672c\u80fd\u529b<\/a><\/p>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-94db76ba847b5872b0faf357d0f2ef16.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-94db76ba847b5872b0faf357d0f2ef16.png\"><\/a><\/section>\n<section><strong>\u673a\u6784: \u5317\u4eac\u5927\u5b66\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u9662 \u5317\u4eac\u901a\u7528\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u9662<\/strong><\/section>\n<section><strong>\u4f5c\u8005: \u6bdb\u5f66\u5347 \u5f90\u5b87\u98de \u674e\u4f73\u742a \u5b5f\u7e41\u7eed \u6768\u660a\u6850 \u90d1\u5b50\u9686 \u738b\u5e0c\u5143 \u5f20\u7267\u6db5<\/strong><\/section>\n<section><\/section>\n<section>\u957f\u6587\u672c\u4efb\u52a1\u662f\u5f53\u4e0b\u5927\u6a21\u578b\u7814\u7a76\u7684\u91cd\u70b9\u4e4b\u4e00\u3002\u5728\u5b9e\u9645\u573a\u666f\u548c\u5e94\u7528\u4e2d\uff0c\u666e\u904d\u5b58\u5728\u5927\u91cf\u957f\u5e8f\u5217\uff08\u6587\u672c\u3001\u8bed\u97f3\u3001\u89c6\u9891\u7b49\uff09\uff0c\u6709\u4e9b\u751a\u81f3\u957f\u8fbe\u767e\u4e07\u7ea7 tokens\u3002\u6269\u5145\u6a21\u578b\u7684\u957f\u6587\u672c\u80fd\u529b\u4e0d\u4ec5\u610f\u5473\u7740\u53ef\u4ee5\u5728\u4e0a\u4e0b\u6587\u7a97\u53e3\u4e2d\u88c5\u5165\u66f4\u957f\u7684\u6587\u672c\uff0c\u66f4\u662f\u80fd\u591f\u66f4\u597d\u5730\u5efa\u6a21\u6587\u672c\u6bb5\u843d\u95f4\u4fe1\u606f\u7684<strong>\u957f\u7a0b\u4f9d\u8d56\u5173\u7cfb<\/strong>\uff0c\u589e\u5f3a\u5bf9\u957f\u6587\u7684\u9605\u8bfb\u7406\u89e3\u548c\u63a8\u7406\u3002<\/section>\n<section><\/section>\n<section>\u73b0\u6709\u5927\u6a21\u578b\u89e3\u51b3\u957f\u6587\u672c\u4efb\u52a1\u7684\u96be\u70b9\u4e4b\u4e00\u662f\u4f20\u7edf\u7684 dot-product attention \u5bf9\u8f93\u5165\u957f\u5ea6\u5448\u5e73\u65b9\u590d\u6742\u5ea6\uff0c\u4e14\u5b58\u50a8 KV cache \u7684\u5f00\u9500\u968f\u8f93\u5165\u957f\u5ea6\u589e\u52a0\uff0c<strong>\u65f6\u95f4\u548c\u7a7a\u95f4\u5f00\u9500\u90fd\u8f83\u9ad8<\/strong><strong>\u3002<\/strong><\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c\u6a21\u578b\u96be\u4ee5\u771f\u6b63\u7406\u89e3\u6563\u843d\u5728\u957f\u6587\u672c\u5404\u5904\u4fe1\u606f\u95f4\u7684<strong>\u957f\u7a0b\u4f9d\u8d56<\/strong>\u3002\u4e3b\u6d41\u7684\u957f\u6587\u672c\u89e3\u51b3\u65b9\u6cd5\u5305\u62ec Retrieval-Augmented Generation\uff08RAG\uff09[1]\u3001long-context adaption \u7b49\u3002<\/section>\n<section><\/section>\n<section>RAG \u4ece\u957f\u6587\u672c\u4e2d\u62bd\u53d6\u4e0e\u95ee\u9898\u76f8\u5173\u7684\u4fe1\u606f\u653e\u5165 context window \u8fdb\u884c\u63a8\u7406\uff0c\u4f46\u5b83\u4f9d\u8d56\u51c6\u786e\u7684\u68c0\u7d22\u65b9\u6cd5\uff0c\u5927\u91cf\u7684\u566a\u58f0\u548c\u65e0\u5173\u4fe1\u606f\u4f1a\u8fdb\u4e00\u6b65\u5f15\u8d77\u6a21\u578b\u5e7b\u89c9\u3002<\/section>\n<section><\/section>\n<section>long-context adaption \u901a\u8fc7\u5728\u5927\u91cf\u957f\u6587\u672c\u7684\u6570\u636e\u96c6\u4e0a\u540e\u8bad\u7ec3[2]\u6269\u5c55\u6a21\u578b\u7684 context window\uff0c\u4f46\u5176\u63a8\u7406\u590d\u6742\u5ea6\u968f\u6587\u672c\u957f\u5ea6\u5e73\u65b9\u589e\u957f\u3001\u663e\u5b58\u5360\u7528\u9ad8\uff0c\u4e14 context window \u4ecd\u7136\u6709\u9650\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u5e94\u5bf9\u957f\u6587\u672c\u5f00\u9500\u5927\u3001\u96be\u4ee5\u5efa\u7acb\u957f\u7a0b\u4f9d\u8d56\u7684\u6311\u6218\uff0c\u5317\u4eac\u5927\u5b66\u5f20\u7267\u6db5\u56e2\u961f\u63d0\u51fa\u5168\u65b0\u7684\u6846\u67b6 <strong>L<\/strong>ong <strong>I<\/strong>nput <strong>F<\/strong>ine-<strong>T<\/strong>uning\uff08<strong>LIFT<\/strong>\uff09\u3002\u901a\u8fc7\u5c06\u957f\u8f93\u5165\u6587\u672c\u8bad\u7ec3\u8fdb\u6a21\u578b\u53c2\u6570\u4e2d\uff0cLIFT \u53ef\u4ee5\u4f7f\u4efb\u610f\u77ed\u4e0a\u4e0b\u6587\u7a97\u53e3\u6a21\u578b\u83b7\u5f97\u957f\u6587\u672c\u80fd\u529b\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-9e01817cd8c32a5f4894e111915e8950.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-9e01817cd8c32a5f4894e111915e8950.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u9898\u76ee: LIFT: Improving Long Context Understanding of Large Language Models through Long Input Fine-Tuning<\/section>\n<\/li>\n<li>\n<section>\u6587\u7ae0\u94fe\u63a5: https:\/\/arxiv.org\/abs\/2502.14644<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u8868 1 \u662f LIFT \u548c\u73b0\u6709\u5e38\u89c1\u65b9\u6cd5\u7684\u5bf9\u6bd4\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e992307527faeb4b40340a735223fb04.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-e992307527faeb4b40340a735223fb04.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u8868 1. LIFT \u4e0e\u4f20\u7edf\u957f\u6587\u672c\u5904\u7406\u65b9\u6cd5\u7684\u5bf9\u6bd4<\/sup><\/em><\/section>\n<section><em><br \/><\/em><\/section>\n<section>LIFT \u9996\u6b21\u63d0\u51fa\u5c06\u957f\u6587\u672c\u77e5\u8bc6\u5b58\u50a8\u5728<strong>\u6a21\u578b\u53c2\u6570<\/strong>\u4e2d\uff0c\u800c\u4e0d\u662f<strong>\u5916\u90e8\u6570\u636e\u5e93<\/strong>\u6216<strong>\u4e0a\u4e0b\u6587\u7a97\u53e3<\/strong>\u4e2d\uff0c\u7c7b\u6bd4\u4eba\u7c7b\u5c06 working memory \u8f6c\u6210 long-term memory\uff0c\u5b9e\u73b0\u77e5\u8bc6\u7684\u5185\u5316\u3002<\/section>\n<section><\/section>\n<section>\u4e0e\u6b64\u76f8\u6bd4\uff0c\u6211\u4eec\u8ba4\u4e3a\u65e0\u9650\u5730\u6269\u5145 context window \u65e0\u6cd5\u771f\u6b63\u89e3\u51b3\u957f\u6587\u672c\u3001\u957f\u5386\u53f2\u7684\u6311\u6218\uff0c\u56e0\u4e3a\u65e0\u8bba\u518d\u957f\u7684 context window \u4ecd\u7136\u6709\u8017\u5c3d\u7684\u4e00\u5929\uff0c\u800c\u53ea\u6709\u5c06\u4e0a\u4e0b\u6587\u6301\u7eed\u5730\u8f6c\u53d8\u6210 parametric knowledge\uff0c\u624d\u80fd\u5b9e\u73b0\u65e0\u9650\u5730\u5b66\u4e60\u3002<\/section>\n<section><\/section>\n<section><strong>\u7814\u7a76\u521b\u65b0<\/strong><\/section>\n<section><\/section>\n<section>\u6211\u4eec\u7684\u65b9\u6848\u5177\u6709\u4ee5\u4e0b\u4f18\u52bf\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section><strong>\u52a8\u6001\u9ad8\u6548\u7684\u957f\u8f93\u5165\u8bad\u7ec3<\/strong>\u3002LIFT \u80fd\u591f\u901a\u8fc7\u8c03\u6574\u6a21\u578b\u53c2\u6570\uff0c\u52a8\u6001\u9002\u5e94\u65b0\u7684\u957f\u8f93\u5165\u6587\u672c\uff0c\u5c06\u5176\u4f5c\u4e3a\u65b0\u7684\u77e5\u8bc6\u6e90\uff0c\u65e0\u9700\u8fdb\u884c\u8d44\u6e90\u5bc6\u96c6\u578b\u7684 long-context adaptation\u3002\u9488\u5bf9\u6bcf\u4e00\u7bc7\u9700\u8981\u5904\u7406\u7684\u957f\u6587\u672c\uff0cLIFT \u901a\u8fc7\u5206\u6bb5\u7684 language modeling \u4ee5\u53ca\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u8f85\u52a9\u4efb\u52a1\u6765\u5fae\u8c03\u6a21\u578b\uff0c\u5b9e\u73b0\u7528\u6a21\u578b\u53c2\u6570\u6765\u8bb0\u5fc6\u548c\u7406\u89e3\u957f\u6587\u672c\uff0c\u4ece\u800c\u907f\u514d\u8fc7\u957f\u7684 context \u9020\u6210\u7684\u63a8\u7406\u590d\u6742\u5ea6\u63d0\u5347\u548c\u957f\u7a0b\u4f9d\u8d56\u4e22\u5931\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<ul>\n<li>\n<section><strong>\u5e73\u8861\u6a21\u578b\u53c2\u6570\u77e5\u8bc6\u548c\u539f\u6709\u80fd\u529b<\/strong>\u3002\u7531\u4e8e\u6a21\u578b\u539f\u6709\u53c2\u6570\uff08\u6bd4\u5982 Llama 3 8B\uff09\u901a\u5e38\u663e\u8457\u5927\u4e8e\u8bb0\u5fc6\u957f\u6587\u672c\u6240\u9700\u7684\u53c2\u6570\u91cf\uff0c\u5168\u53c2\u6570\u5fae\u8c03\u9762\u4e34\u8fc7\u62df\u5408\u957f\u6587\u672c\u800c\u635f\u5931\u6a21\u578b\u57fa\u7840\u80fd\u529b\u7684\u98ce\u9669\u3002\u4e3a\u4e86\u5728\u6a21\u578b\u539f\u6709\u80fd\u529b\u548c\u5fae\u8c03\u540e\u65b0\u7684\u53c2\u6570\u5185\u77e5\u8bc6\u4e4b\u95f4\u627e\u5230\u5e73\u8861\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u4e13\u95e8\u7684\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u6a21\u5757\u2014\u2014\u95e8\u63a7\u8bb0\u5fc6\u9002\u914d\u5668\uff08Gated Memory Adapter\uff09\uff0c\u5b83\u80fd\u5e73\u8861\u539f\u59cb\u6a21\u578b\u7684 In-Context Learning\uff08ICL\uff09\u80fd\u529b\u548c LIFT \u8bad\u7ec3\u540e\u5bf9\u957f\u8f93\u5165\u7684\u8bb0\u5fc6\u7406\u89e3\u80fd\u529b\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<ul>\n<li>\n<section><strong>\u5728\u6d41\u884c\u7684\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u4e0a\u53d6\u5f97\u4e86\u5de8\u5927\u63d0\u5347<\/strong>\u3002\u5728\u51e0\u4e2a\u5e7f\u6cdb\u8ba4\u53ef\u7684\u957f\u4e0a\u4e0b\u6587\u57fa\u51c6\u96c6\uff08\u4f8b\u5982 LooGLE [3]\u3001Longbench [4]\uff09\u4e0a\u7684\u8bc4\u4f30\u8868\u660e\uff0c\u4e0d\u540c LLM \u59cb\u7ec8\u80fd\u901a\u8fc7 LIFT \u5728\u5e38\u89c1\u7684\u957f\/\u77ed\u4f9d\u8d56\u95ee\u7b54\u548c\u6458\u8981\u7b49\u901a\u7528\u4efb\u52a1\u4e0a\u53d7\u76ca\u3002\u4f8b\u5982\uff0c\u5728\u975e\u5e38\u5177\u6709\u6311\u6218\u6027\u7684 LooGLE \u957f\u4f9d\u8d56\u95ee\u7b54\u4e0a\uff0c\u76f8\u8f83\u4ec5\u901a\u8fc7 ICL\uff0cLIFT \u8fc7\u540e\u7684 Llama 3 8B \u7684\u6b63\u786e\u7387\u4ece 15.44% \u63d0\u5347\u81f3 29.97%\u3002\u5728 LooGLE \u77ed\u4f9d\u8d56\u95ee\u7b54\u4e0a\uff0cLIFT \u5c06 Gemma 2 9B \u7684\u6b63\u786e\u7387\u4ece 37.37% \u63d0\u5347\u81f3 50.33%\u3002<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-539ce226363ed820985706ff09b85866.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-539ce226363ed820985706ff09b85866.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 1.LIFT \u6846\u67b6\u4ee5\u53ca\u548c\u73b0\u6709\u65b9\u6cd5\u5bf9\u6bd4<\/sup><\/em><\/section>\n<section><em><br \/><\/em><\/section>\n<section><strong>LIFT \u65b9\u6cd5<\/strong><\/section>\n<section><\/section>\n<section><strong>\u957f\u6587\u672c\u5207\u6bb5\u8bad\u7ec3<\/strong><\/section>\n<section><\/section>\n<section>\u53d7 LLM \u9884\u8bad\u7ec3\u7684\u542f\u53d1\uff0cLIFT \u5c06\u300c\u8bb0\u5fc6\u957f\u6587\u672c\u300d\u7684\u4efb\u52a1\u5efa\u6a21\u4e3a\u8bed\u8a00\u5efa\u6a21\uff08Language Modeling\uff09\u4efb\u52a1\u3002\u4f46\u5728\u6574\u7bc7\u957f\u6587\u672c\u4e0a\u8fdb\u884c\u8bed\u8a00\u5efa\u6a21\u8bad\u7ec3\u5f00\u9500\u8fc7\u5927\uff0c\u4e14\u77ed\u4e0a\u4e0b\u6587\u6a21\u578b\u4e0d\u5177\u5907\u76f4\u63a5\u5728\u957f\u6587\u672c\u4e0a\u8bad\u7ec3\u7684\u80fd\u529b\u3002\u4e3a\u6b64\uff0c<strong>LIFT \u5c06\u957f\u6587\u672c\u5207\u5206\u4e3a\u56fa\u5b9a\u957f\u5ea6\u7684\u7247\u6bb5\uff0c\u5bf9\u6240\u6709\u7247\u6bb5\u5e76\u884c\u8fdb\u884c\u8bed\u8a00\u5efa\u6a21\u8bad\u7ec3<\/strong>\u3002<\/section>\n<section><\/section>\n<section>\u5982\u679c\u5c06\u957f\u6587\u672c\u5207\u5206\u4e3a\u4e92\u4e0d\u76f8\u4ea4\u7684\u7247\u6bb5\uff08\u5982\u56fe 2 \u4e2d Trivial segmentation \u6240\u793a\uff09\uff0c\u6a21\u578b\u5c06\u4e22\u5931\u7247\u6bb5\u95f4\u7684\u6b63\u786e\u987a\u5e8f\uff0c\u800c\u987a\u5e8f\u5bf9\u4e8e\u957f\u6587\u672c\u4e2d\u7684\u957f\u7a0b\u4f9d\u8d56\u548c\u603b\u4f53\u7406\u89e3\u975e\u5e38\u91cd\u8981\u3002\u56e0\u6b64\uff0cLIFT \u8981\u6c42<strong>\u76f8\u90bb\u7247\u6bb5\u6709\u4e00\u5b9a\u91cd\u53e0<\/strong>\uff08\u5982\u56fe 2 \u4e2d\u7684 Our segmentation \u6240\u793a\uff09\u2014\u2014\u6bcf\u4e2a\u7247\u6bb5\u7684\u672b\u5c3e\u5c31\u662f\u4e0b\u4e00\u4e2a\u7247\u6bb5\u7684\u5f00\u5934\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u6837\uff0c\u5982\u679c\u6a21\u578b\u80fd\u591f\u8bb0\u5fc6\u67d0\u4e2a\u7247\u6bb5\uff0c\u90a3\u4e48\u5b83\u5c31\u80fd\u591f\u7eed\u5199\u51fa\u4e0b\u4e00\u4e2a\u7247\u6bb5\uff0c\u76f4\u5230\u6309\u987a\u5e8f\u7eed\u5199\u51fa\u5168\u6587\u3002\u5728\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u53d6\u91cd\u53e0\u7684\u957f\u5ea6\u4e3a\u7247\u6bb5\u957f\u5ea6\u7684 5\/8\uff0c\u56e0\u6b64\u8bad\u7ec3\u7684\u590d\u6742\u5ea6\u5bf9\u957f\u6587\u672c\u7684\u957f\u5ea6\u5448\u7ebf\u6027\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-41309a7141bd94c41a4f5c487b6d9e6a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-41309a7141bd94c41a4f5c487b6d9e6a.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 2. LIFT \u7684\u6587\u7ae0\u5207\u6bb5\u65b9\u6cd5<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u8f85\u52a9\u4efb\u52a1\u8bad\u7ec3<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u5fae\u8c03 LLM \u901a\u5e38\u4f1a\u5bfc\u81f4\u5176\u5728\u5176\u4ed6\u4efb\u52a1\u4e0a\u80fd\u529b\u4e0b\u964d\u3002\u540c\u7406\uff0c\u957f\u6587\u672c\u5207\u6bb5\u8bad\u7ec3\u53ef\u80fd\u5bfc\u81f4 LLM \u7684 reasoning\u3001instruction-following \u7b49\u80fd\u529b\u635f\u5931\u3002<\/section>\n<section><\/section>\n<section>\u7814\u7a76\u56e2\u961f<strong>\u63d0\u51fa\u5728\u5408\u6210\u7684\u8f85\u52a9\u4efb\u52a1<\/strong>\u4e0a\u8bad\u7ec3\uff0c\u4e00\u65b9\u9762\u5f25\u8865\u6a21\u578b\u7684\u80fd\u529b\u635f\u5931\uff0c\u53e6\u4e00\u65b9\u9762\u5e2e\u52a9\u6a21\u578b<strong>\u5b66\u4f1a\u5e94\u7528\u957f\u6587\u672c\u4e2d\u7684\u4fe1\u606f\u56de\u7b54\u95ee\u9898<\/strong>\u3002\u5177\u4f53\u800c\u8a00\uff0c\u7814\u7a76\u56e2\u961f\u7528\u9884\u8bad\u7ec3\u7684 LLM \u57fa\u4e8e\u957f\u6587\u672c\u7247\u6bb5\u81ea\u52a8\u751f\u6210\u51e0\u5341\u4e2a\u95ee\u7b54\u7c7b\u578b\u7684\u8f85\u52a9\u4efb\u52a1\u3002<\/section>\n<section><\/section>\n<section>\u4e8e\u662f LIFT \u8bad\u7ec3\u5206\u4e3a\u4e24\u4e2a\u9636\u6bb5\uff0c\u7b2c\u4e00\u4e2a\u9636\u6bb5\u53ea\u5728\u957f\u6587\u672c\u5207\u6bb5\u4efb\u52a1\u4e0a\u8fdb\u884c\u8bed\u8a00\u5efa\u6a21\u8bad\u7ec3\uff0c\u7b2c\u4e8c\u4e2a\u9636\u6bb5\u5728\u8f85\u52a9\u4efb\u52a1\u4e0a\u8bad\u7ec3\u6a21\u578b\u57fa\u4e8e\u957f\u6587\u672c\u56de\u7b54\u95ee\u9898\u7684\u80fd\u529b\u3002<\/section>\n<section><\/section>\n<section><strong>Gated Memory \u67b6\u6784<\/strong><\/section>\n<section><\/section>\n<section>\u5c3d\u7ba1 LIFT \u53ef\u4ee5\u4efb\u610f\u4f7f\u7528\u5168\u53c2\u6570\u5fae\u8c03\u6216 LoRA\/PiSSA \u7b49\u53c2\u6570\u9ad8\u6548\u5fae\u8c03\u65b9\u6cd5\u6765\u8bad\u7ec3\u6a21\u578b\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a\u4e13\u7528\u7684 Gated Memory Adapter \u6765\u5e73\u8861\u957f\u6587\u672c\u8bb0\u5fc6\u548c\u80fd\u529b\u3002\u5176\u6838\u5fc3\u5728\u4e8e<strong>\u7528\u77ed\u7a97\u53e3\u6a21\u578b\u6a21\u62df\u5047\u8bbe\u957f\u6587\u672c\u5728\u4e0a\u4e0b\u6587\u7a97\u53e3\u4e2d\u65f6\u6a21\u578b\u7684\u884c\u4e3a\u548c\u5185\u90e8\u8868\u793a<\/strong>\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u6b64\u6211\u4eec\u5c06\u5047\u8bbe\u7684\u300c\u5168\u4e0a\u4e0b\u6587\u300d\u5206\u4e3a\u7a97\u53e3\u5916\uff08out-of-context\uff09\u548c\u7a97\u53e3\u4e2d\uff08in-context\uff09\u4e24\u90e8\u5206\u2014\u2014\u7a97\u53e3\u5916\u653e\u7f6e\u7684\u662f\u9884\u8ba1\u5c06\u901a\u8fc7\u5fae\u8c03\u653e\u5165\u53c2\u6570\u4e2d\u7684\u957f\u6587\u672c\uff0c\u800c\u7a97\u53e3\u4e2d\u653e\u7f6e\u7684\u662f\u5173\u4e8e\u957f\u6587\u672c\u7684\u95ee\u9898\u3002<\/section>\n<section><\/section>\n<section>\u6211\u4eec\u7684\u76ee\u7684\u662f\u8bbe\u8ba1\u4e00\u4e2a\u6a21\u578b\uff0c\u7528 LIFT \u53c2\u6570 + \u7a97\u53e3\u5916\u5185\u5bb9\uff08\u77ed\u4e0a\u4e0b\u6587\uff09\u53bb\u6a21\u62df\u5168\u4e0a\u4e0b\u6587\u7684\u884c\u4e3a\uff0c\u4ee5\u6b64\u8fbe\u6210\u53ea\u7528\u77ed\u4e0a\u4e0b\u6587\u6a21\u578b\u5b9e\u73b0\u957f\u4e0a\u4e0b\u6587\u7684\u6ce8\u610f\u529b\u673a\u5236\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u6b64\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u95e8\u63a7\u8bb0\u5fc6\u6a21\u5757\uff08Gated Memory\uff09\uff08\u89c1\u56fe 3\uff09\u3002\u8be5\u6a21\u5757\u4e3a\u6bcf\u4e2a\u6ce8\u610f\u529b\u5c42\u589e\u52a0\u4e86\u4e24\u4e2a\u7279\u6b8a\u7684 MLP\uff08\u56fe 3 \u4e2d\u7684 Memory MLP \u548c Gate MLP\uff09\uff0c\u5747\u4ee5\u6bcf\u4e2a\u4f4d\u7f6e\u7684 query vector \u4e3a\u8f93\u5165\uff0c\u5206\u522b\u7528\u4e8e\u5b66\u4e60\u300c<strong>\u7a97\u53e3\u5916\u90e8\u5206\u7684\u6743\u91cd<\/strong>\u300d\uff08gate\uff09\u548c\u300c<strong>\u7a97\u53e3\u5916\u90e8\u5206\u7684\u8bb0\u5fc6\u63d0\u53d6\u5185\u5bb9<\/strong>\u300d\uff08memory\uff09\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u6837\uff0c\u5f53\u4e00\u4e2a\u65b0\u7684 query \u8fdb\u5165\uff0c\u6a21\u578b\u53ef\u4ee5\u52a8\u6001\u5730\u8c03\u63a7\u5176\u4f7f\u7528\u591a\u5c11 LIFT \u8bb0\u5fc6\u7684\u7a97\u53e3\u5916\u5185\u5bb9\uff1a\u5f53 gate=0\uff0c\u6a21\u578b\u5c06\u6062\u590d\u4e3a\u7eaf ICL\uff0c\u4e0d\u7528\u4efb\u4f55 LIFT \u8bb0\u5fc6\u7684\u4fe1\u606f\uff1b\u5f53 gate=1\uff0c\u6a21\u578b\u5c06\u5b8c\u5168\u4f9d\u8d56 LIFT \u77e5\u8bc6\u800c\u5ffd\u7565\u5f53\u524d\u7a97\u53e3\u4e2d\u7684\u4e0a\u4e0b\u6587\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u79cd\u52a8\u6001\u5206\u914d\u673a\u5236\u53ef\u4ee5\u6709\u6548\u5730\u5e73\u8861\u5bf9\u957f\u6587\u672c\u7684\u8bb0\u5fc6\u548c\u6a21\u578b\u539f\u672c\u7684 ICL \u80fd\u529b\u3002LIFT \u8bad\u7ec3\u8fc7\u7a0b\u4e2d\uff0c\u6211\u4eec\u5c06\u53ea\u5fae\u8c03 Gated Memory \u4e2d\u7684\u53c2\u6570\uff0c\u5b9e\u73b0\u4e86\u6a21\u578b\u5728\u5fae\u8c03\u8f83\u5c0f\u53c2\u6570\u91cf\u7684\u60c5\u51b5\u4e0b\uff0c\u6709\u6548\u5730\u8bb0\u5fc6\u957f\u6587\u672c\u5185\u5bb9\u5e76\u7528\u4e8e\u4e0b\u6e38\u4efb\u52a1\u3002<\/section>\n<section><\/section>\n<section>\u5b9e\u9a8c\u8bc1\u660e\u4e86\u8fd9\u4e00\u7ed3\u6784\u7684\u6709\u6548\u6027\uff08\u89c1\u4e0b\u6587\u8868 4\uff09\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1fa9c7055f6dc9545b50a1e13cda7f38.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-1fa9c7055f6dc9545b50a1e13cda7f38.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 3.Gated Memory \u6a21\u5757<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u5b9e\u9a8c\u6d4b\u8bc4<\/strong><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u8bc4\u4f30 LIFT \u7684\u6709\u6548\u6027\uff0c\u7814\u7a76\u56e2\u961f\u5728 Llama 3 8B \u548c Gemma 2 9B \u4e24\u4e2a\u77ed\u6587\u672c\u5f00\u6e90\u6a21\u578b\uff08\u4e0a\u4e0b\u6587\u7a97\u53e3\u4e3a 8k\uff09\u4e0a\u548c GPT 3.5 \u5546\u7528\u6a21\u578b\uff08\u4e0a\u4e0b\u6587\u7a97\u53e3\u4e3a 16k\uff09\u4e0a\u6bd4\u8f83\u4e86 LIFT \u65b9\u6cd5\u548c\u4f7f\u7528\u622a\u65ad ICL \u7684 baselines\u3002<\/section>\n<section><\/section>\n<section>baselines \u4f7f\u7528\u539f\u6a21\u578b\uff0c\u5c3d\u53ef\u80fd\u5c06\u957f\u6587\u672c\u586b\u5165\u6a21\u578b\u7684\u4e0a\u4e0b\u6587\u7a97\u53e3\uff08\u4f18\u5148\u586b\u5165\u5f00\u5934\u548c\u672b\u5c3e tokens\uff0c\u5176\u4f59\u622a\u65ad\uff09\uff0c\u5e76\u4fdd\u8bc1\u95ee\u9898 prompt \u5168\u90e8\u586b\u5165\u3002LIFT \u5728\u6d4b\u8bd5\u65f6\u7684\u8f93\u5165\u4e0e baseline \u76f8\u540c\uff0c\u4f46\u4f7f\u7528\u7684\u6a21\u578b\u4e3a\u7ecf\u8fc7 LIFT \u8bad\u7ec3\u7684\u6a21\u578b\uff0c\u5e76\u9ed8\u8ba4\u4f7f\u7528 Gated Memory \u9002\u914d\u5668\u3002<\/section>\n<section><\/section>\n<section>\u5bf9\u4e8e GPT3.5\uff0c\u6211\u4eec\u76f4\u63a5\u8c03\u7528 GPT 3.5 \u7684\u8bad\u7ec3 API\u3002\u6211\u4eec\u4e3b\u8981\u5728\u4e24\u4e2a\u4ee3\u8868\u6027\u7684\u957f\u6587\u672c\u8bc4\u6d4b\u96c6 LooGLE \u548c LongBench \u4e0a\u8bc4\u6d4b\uff0c\u5176\u4e2d LooGLE \u5305\u542b\u5927\u91cf\u4eba\u5de5\u6807\u6ce8\u7684\u6781\u5177\u6311\u6218\u6027\u7684\u957f\u4f9d\u8d56\u95ee\u7b54\uff08LongQA\uff09\u548c LLM \u81ea\u52a8\u751f\u6210\u7684\u77ed\u4f9d\u8d56\u95ee\u7b54\uff08ShortQA\uff09\uff0cLongBench \u5305\u542b\u95ee\u7b54\u3001\u6458\u8981\u7b49\u591a\u79cd\u4efb\u52a1\u3002<\/section>\n<section><\/section>\n<section>\u7ed3\u679c\u5982\u8868 2\u3001\u8868 3 \u6240\u793a\uff0c\u5b9e\u9a8c\u8868\u660e\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section><strong>LIFT \u6781\u5927\u63d0\u5347\u4e86\u77ed\u6587\u672c\u6a21\u578b\u5728 LooGLE \u4e0a\u7684\u8868\u73b0<\/strong>\u3002LIFT \u7a33\u5b9a\u63d0\u5347\u4e86\u88ab\u6d4b\u6a21\u578b\u5728 ShortQA \u548c LongQA \u4e2d\u7684\u5e73\u5747\u6307\u6807\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cLlama 3 \u5728 LongQA \u4e0a\u7684\u6307\u6807\u4ece 15.44% \u63d0\u5347\u81f3 29.97%\uff0cGemma 2 \u5728 ShortQA \u4e0a\u7684\u6307\u6807\u4ece 37.37% \u63d0\u5347\u81f3 50.33%\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<ul>\n<li>\n<section><strong>LIFT \u63d0\u5347\u4e86\u77ed\u6587\u672c\u6a21\u578b\u5728 Longbench \u7684\u5927\u591a\u6570\u5b50\u4efb\u52a1\u4e0a\u7684\u8868\u73b0<\/strong>\u3002\u7814\u7a76\u56e2\u961f\u4ece LongBench \u4e2d\u9009\u53d6\u4e86 5 \u4e2a\u5177\u6709\u4ee3\u8868\u6027\u7684\u5b50\u4efb\u52a1\u8fdb\u884c\u6d4b\u8bd5\uff0c\u4efb\u52a1\u5305\u62ec\u591a\u7bc7\u6587\u7ae0\u95f4\u7684\u591a\u8df3\u63a8\u7406\u3001\u9605\u8bfb\u7406\u89e3\u548c\u6982\u62ec\u3001\u68c0\u7d22\u53ec\u56de\u7b49\uff0cLlama 3 \u901a\u8fc7 LIFT \u5728\u5176\u4e2d 4 \u4e2a\u5b50\u4efb\u52a1\u4e0a\u5747\u6709\u63d0\u5347\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<ul>\n<li>\n<section><strong>LIFT \u7684\u6548\u679c\u4e0e\u6a21\u578b\u7684\u539f\u6709\u80fd\u529b\u4ee5\u53ca\u6d4b\u8bd5\u4efb\u52a1\u6709\u5173<\/strong>\u3002LIFT \u867d\u7136\u666e\u904d\u63d0\u5347\u4e86\u6a21\u578b\u7684\u957f\u6587\u672c\u80fd\u529b\uff0c\u4f46\u5728\u90e8\u5206\u5b50\u4efb\u52a1\u4e0a\u4ecd\u6709\u6539\u8fdb\u7a7a\u95f4\u3002\u901a\u8fc7\u5206\u6790\u5404\u4e2a\u5b50\u4efb\u52a1\uff0c\u7814\u7a76\u56e2\u961f\u8ba4\u4e3a\u4e0e\u6d4b\u8bd5\u95ee\u9898\u76f8\u4f3c\u7684\u8f85\u52a9\u4efb\u52a1\u53ef\u4ee5\u4fc3\u8fdb\u6a21\u578b\u5173\u6ce8\u5bf9\u6d4b\u8bd5\u4efb\u52a1\u6709\u7528\u7684\u957f\u4e0a\u4e0b\u6587\u4fe1\u606f\uff0c\u6709\u52a9\u4e8e\u4e0b\u6e38\u4efb\u52a1\u8868\u73b0\u3002<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-b9e25eed1ab010b37e370c481bb28953.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-b9e25eed1ab010b37e370c481bb28953.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u8868 2. LIFT \u5728 LooGLE \u4e0a\u7684 GPT4_score \u6307\u6807<\/sup><\/em><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-fb1fd3eeb451696404553dffdbfb7da7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-fb1fd3eeb451696404553dffdbfb7da7.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u8868 3. LIFT \u5728 LongBench \u4e0a\u7684\u8868\u73b0\uff08\u8bc4\u6d4b\u6307\u6807\u548c\u539f\u6570\u636e\u96c6\u4e00\u81f4\uff09<\/sup><\/em><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-63eeb39be5ff09994f3741b46696f31a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-63eeb39be5ff09994f3741b46696f31a.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u8868 4. LIFT Gated Memory \u67b6\u6784\u7684\u6d88\u878d\u5b9e\u9a8c<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c\u6211\u4eec\u901a\u8fc7\u6d88\u878d\u5b9e\u9a8c\u9a8c\u8bc1\u4e86 Gated Memory \u9002\u914d\u5668\u7684\u4f5c\u7528\u3002\u5982\u8868 4 \u6240\u793a\uff0c\u5728 LooGLE ShortQA \u6570\u636e\u96c6\u4e0a\uff0cGated Memory \u67b6\u6784\u76f8\u6bd4\u4e8e\u4f7f\u7528 PiSSA[5]\uff08\u4e00\u79cd LoRA \u7684\u6539\u8fdb\u7248\u65b9\u6cd5\uff09\u5fae\u8c03\u7684\u539f\u6a21\u578b\uff0cGPT-4 score \u63d0\u5347\u4e86 5.48%\u3002<\/section>\n<section><\/section>\n<section><strong>\u603b\u7ed3\u3001\u5c55\u671b\u548c\u8ba8\u8bba<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u6846\u67b6\u2014\u2014LIFT\uff0c\u4ee5\u589e\u5f3a LLMs \u7684\u957f\u4e0a\u4e0b\u6587\u7406\u89e3\u80fd\u529b\u3002LIFT \u901a\u8fc7\u9ad8\u6548\u5fae\u8c03\u6a21\u578b\u53c2\u6570\uff0c\u5229\u7528<strong>\u53c2\u6570\u5185\u77e5\u8bc6<\/strong>\uff08in-parameter knowledge\uff09\u6765\u52a8\u6001\u9002\u5e94\u957f\u8f93\u5165\uff0c\u4ece\u800c\u63d0\u5347\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u7684\u80fd\u529b\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u5728 LooGLE \u548c LongBench \u7b49\u6d41\u884c\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0cLIFT \u663e\u8457\u63d0\u5347\u4e86\u77ed\u4e0a\u4e0b\u6587 LLMs \u5728\u957f\u4e0a\u4e0b\u6587\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u3002<\/section>\n<section><\/section>\n<section>\u7136\u800c\uff0cLIFT \u4ecd\u7136\u5b58\u5728\u4e00\u5b9a<strong>\u5c40\u9650\u6027<\/strong>\u3002\u9996\u5148\uff0c\u5728 context window \u4e0d\u591f\u7684\u60c5\u51b5\u4e0b\uff0c\u6211\u4eec\u7ecf\u5e38\u9700\u8981\u622a\u65ad\u4e0a\u4e0b\u6587\u6765\u505a\u957f\u6587\u672c\u63a8\u7406\uff0c\u4f46\u5bf9\u4e8e\u9700\u8981\u7cbe\u786e\u4fe1\u606f\u63d0\u53d6\u7684\u4efb\u52a1\uff0c\u5982\u300c\u5927\u6d77\u635e\u9488\u4efb\u52a1\u300d\uff08Needle in a Haystack\uff09\uff0c\u8be5\u65b9\u6cd5\u4ecd\u7136\u6027\u80fd\u6b20\u4f73\u3002<\/section>\n<section><\/section>\n<section>\u5176\u6b21\uff0cLIFT \u901a\u8fc7\u5c06\u957f\u6587\u672c\u8f93\u5165\u6ce8\u5165\u6a21\u578b\u53c2\u6570\uff0c\u589e\u5f3a\u4e86\u6a21\u578b\u5bf9\u6570\u636e\u7684\u719f\u6089\u5ea6\uff0c\u4f46\u4e0b\u6e38\u4efb\u52a1\u7684\u6548\u679c\u4ecd\u7136\u4f9d\u8d56\u4e8e\u6a21\u578b\u80fd\u5426\u81ea\u4e3b\u63d0\u53d6\u548c\u5229\u7528 LIFT \u8fc7\u7a0b\u4e2d\u83b7\u5f97\u7684\u53c2\u6570\u5316\u77e5\u8bc6\u3002\u5206\u6790\u8868\u660e\uff0c\u6a21\u578b\u5728\u300cin-context\u300d\u548c\u300cout-of-context\u300d\u95ee\u9898\u4e0a\u7684\u8868\u73b0\u5b58\u5728\u663e\u8457\u5dee\u8ddd\uff0c\u8868\u660e LIFT \u540e\u7684\u53c2\u6570\u5316\u77e5\u8bc6\u63d0\u53d6\u80fd\u529b\u4ecd\u9700\u8fdb\u4e00\u6b65\u4f18\u5316\u3002<\/section>\n<section><\/section>\n<section>\u6b64\u5916\uff0c\u6211\u4eec\u53d1\u73b0\u5728 LIFT \u8fc7\u7a0b\u4e2d\u5f15\u5165\u8f85\u52a9\u4efb\u52a1\u5e76\u4e0d\u80fd\u603b\u662f\u663e\u8457\u63d0\u9ad8\u6a21\u578b\u80fd\u529b\uff0c\u5176\u6027\u80fd\u4e25\u91cd\u4f9d\u8d56\u4e0b\u6e38\u6d4b\u8bd5\u4efb\u52a1\u548c\u8f85\u52a9\u4efb\u52a1\u7684\u76f8\u4f3c\u7a0b\u5ea6\uff0c\u751a\u81f3\u53ef\u80fd\u56e0\u8fc7\u62df\u5408\u800c\u5bfc\u81f4\u6027\u80fd\u4e0b\u964d\u3002\u56e0\u6b64\uff0c\u5982\u4f55\u8bbe\u8ba1<strong>\u66f4\u901a\u7528\u7684\u8f85\u52a9\u4efb\u52a1<\/strong>\u662f\u672a\u6765\u7684\u7814\u7a76\u91cd\u70b9\u3002<\/section>\n<section><\/section>\n<section>\u6700\u540e\uff0c\u5c3d\u7ba1 Gated Memory \u67b6\u6784\u663e\u8457\u63d0\u5347\u4e86\u957f\u6587\u672c\u8bb0\u5fc6\u548c ICL \u80fd\u529b\u7684\u5e73\u8861\uff0c\u6211\u4eec\u53d1\u73b0 LIFT \u540e\u7684\u6a21\u578b\u4ecd\u5b58\u5728\u5bf9\u539f\u6709\u80fd\u529b\u7684\u7834\u574f\uff0c\u5982\u4f55\u8bbe\u8ba1\u66f4\u597d\u7684\u9002\u914d\u5668\u6765\u5e73\u8861\u8bb0\u5fc6\u548c\u80fd\u529b\uff0c\u4e5f\u7559\u4f5c\u672a\u6765\u5de5\u4f5c\u3002<\/section>\n<section><\/section>\n<section>LIFT \u7684\u7406\u5ff5\u975e\u5e38\u6709\u8da3\uff0c\u56e0\u4e3a\u4eba\u7c7b\u7684<strong>\u77ed\u671f\u8bb0\u5fc6\u4e5f\u4f1a\u8f6c\u5316\u4e3a\u957f\u671f\u8bb0\u5fc6<\/strong>\uff0c\u8fd9\u4e00\u8fc7\u7a0b\u7c7b\u4f3c\u4e8e LIFT \u5c06<strong>\u4e0a\u4e0b\u6587\u4e2d\u7684\u77e5\u8bc6\u8f6c\u6362\u4e3a\u53c2\u6570\u5316\u77e5\u8bc6<\/strong>\u3002\u867d\u7136\u8ddd\u79bb\u5f7b\u5e95\u89e3\u51b3 LLMs \u7684\u957f\u4e0a\u4e0b\u6587\u6311\u6218\u4ecd\u7136\u4efb\u91cd\u9053\u8fdc\uff0c\u4f46\u6211\u4eec\u7684\u521d\u6b65\u7ed3\u679c\u8868\u660e\uff0cLIFT \u63d0\u4f9b\u4e86\u4e00\u4e2a\u6781\u5177\u6f5c\u529b\u548c\u524d\u666f\u7684\u7814\u7a76\u65b9\u5411\u3002<\/section>\n<section><\/section>\n<section>\u6211\u4eec\u9f13\u52b1\u793e\u533a\u4e00\u540c\u63a2\u7d22 LIFT \u5728\u66f4\u5e7f\u6cdb\u7684\u8bad\u7ec3\u6570\u636e\u3001\u66f4\u4e30\u5bcc\u7684\u6a21\u578b\u3001\u66f4\u5148\u8fdb\u7684\u8f85\u52a9\u4efb\u52a1\u8bbe\u8ba1\u4ee5\u53ca\u66f4\u5f3a\u8ba1\u7b97\u8d44\u6e90\u652f\u6301\u4e0b\u7684\u6f5c\u5728\u80fd\u529b\u3002<\/section>\n<section><\/section>\n<section><em><sup>\u53c2\u8003\u6587\u732e<\/sup><\/em><\/section>\n<section><sup><em>[1] Jiang, Ziyan, Xueguang Ma, and Wenhu Chen. &#8220;Longrag: Enhancing retrieval-augmented generation with long-context llms.&#8221; arXiv preprint arXiv:2406.15319 (2024).<\/em><\/sup><\/section>\n<section><sup><em>[2] Chen, Yukang, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. &#8220;LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models.&#8221; In The Twelfth International Conference on Learning Representations.<\/em><\/sup><\/section>\n<section><sup><em>[3] Li, Jiaqi, Mengmeng Wang, Zilong Zheng, and Muhan Zhang. &#8220;LooGLE: Can Long-Context Language Models Understand Long Contexts?.&#8221; In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 16304-16333. 2024.<\/em><\/sup><\/section>\n<section><sup><em>[4] Bai, Yushi, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du et al. &#8220;LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding.&#8221; In ACL (1). 2024.<\/em><\/sup><\/section>\n<section><sup><em>[5] Meng, Fanxu, Zhaohui Wang, and Muhan Zhang. &#8220;PiSSA: Principal singular values and singular vectors adaptation of large language models.&#8221; Advances in Neural Information Processing Systems 37 (2024): 121038-121072.<\/em><\/sup><\/section>\n<section><em><sup>[6] Hong, Junyuan, Lingjuan Lyu, Jiayu Zhou, and Michael Spranger. &#8220;Mecta: Memory-economic continual test-time model adaptation.&#8221; In 2023 International Conference on Learning Representations. 2023.<\/sup><\/em><\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2025-03-17-4\" target=\"_blank\">\u5317\u5927\u56e2\u961f\u63d0\u51faLIFT\uff1a\u5c06\u957f\u4e0a\u4e0b\u6587\u77e5\u8bc6\u6ce8\u5165\u6a21\u578b\u53c2\u6570\uff0c\u63d0\u5347\u5927\u6a21\u578b\u957f\u6587\u672c\u80fd\u529b<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u5317\u5927\u56e2\u961f\u63d0\u51faLIFT\uff1a [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","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":"","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":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","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":[27],"tags":[71,68],"class_list":["post-39475","post","type-post","status-publish","format-standard","hentry","category-news","tag-rag","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v26.4 - https:\/\/yoast.com\/wordpress\/plugins\/seo\/ -->\n<title>\u5317\u5927\u56e2\u961f\u63d0\u51faLIFT\uff1a\u5c06\u957f\u4e0a\u4e0b\u6587\u77e5\u8bc6\u6ce8\u5165\u6a21\u578b\u53c2\u6570\uff0c\u63d0\u5347\u5927\u6a21\u578b\u957f\u6587\u672c\u80fd\u529b - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=39475\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/17aitech.com\/?p=39475\",\"url\":\"https:\/\/17aitech.com\/?p=39475\",\"name\":\"\u5317\u5927\u56e2\u961f\u63d0\u51faLIFT\uff1a\u5c06\u957f\u4e0a\u4e0b\u6587\u77e5\u8bc6\u6ce8\u5165\u6a21\u578b\u53c2\u6570\uff0c\u63d0\u5347\u5927\u6a21\u578b\u957f\u6587\u672c\u80fd\u529b - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\/\/17aitech.com\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\/\/17aitech.com\/?p=39475#primaryimage\"},\"image\":{\"@id\":\"https:\/\/17aitech.com\/?p=39475#primaryimage\"},\"thumbnailUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-94db76ba847b5872b0faf357d0f2ef16.png\",\"datePublished\":\"2025-03-17T17:02:29+00:00\",\"author\":{\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\/\/17aitech.com\/?p=39475#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/17aitech.com\/?p=39475\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/?p=39475#primaryimage\",\"url\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-94db76ba847b5872b0faf357d0f2ef16.png\",\"contentUrl\":\"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-94db76ba847b5872b0faf357d0f2ef16.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/17aitech.com\/?p=39475#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/17aitech.com\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u5317\u5927\u56e2\u961f\u63d0\u51faLIFT\uff1a\u5c06\u957f\u4e0a\u4e0b\u6587\u77e5\u8bc6\u6ce8\u5165\u6a21\u578b\u53c2\u6570\uff0c\u63d0\u5347\u5927\u6a21\u578b\u957f\u6587\u672c\u80fd\u529b\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/17aitech.com\/#website\",\"url\":\"https:\/\/17aitech.com\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/17aitech.com\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/17aitech.com\/#\/schema\/person\/image\/\",\"url\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"contentUrl\":\"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\/\/17aitech.com\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u5317\u5927\u56e2\u961f\u63d0\u51faLIFT\uff1a\u5c06\u957f\u4e0a\u4e0b\u6587\u77e5\u8bc6\u6ce8\u5165\u6a21\u578b\u53c2\u6570\uff0c\u63d0\u5347\u5927\u6a21\u578b\u957f\u6587\u672c\u80fd\u529b - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=39475","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=39475","url":"https:\/\/17aitech.com\/?p=39475","name":"\u5317\u5927\u56e2\u961f\u63d0\u51faLIFT\uff1a\u5c06\u957f\u4e0a\u4e0b\u6587\u77e5\u8bc6\u6ce8\u5165\u6a21\u578b\u53c2\u6570\uff0c\u63d0\u5347\u5927\u6a21\u578b\u957f\u6587\u672c\u80fd\u529b - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=39475#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=39475#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-94db76ba847b5872b0faf357d0f2ef16.png","datePublished":"2025-03-17T17:02:29+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=39475#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=39475"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=39475#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-94db76ba847b5872b0faf357d0f2ef16.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2025\/03\/frc-94db76ba847b5872b0faf357d0f2ef16.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=39475#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u5317\u5927\u56e2\u961f\u63d0\u51faLIFT\uff1a\u5c06\u957f\u4e0a\u4e0b\u6587\u77e5\u8bc6\u6ce8\u5165\u6a21\u578b\u53c2\u6570\uff0c\u63d0\u5347\u5927\u6a21\u578b\u957f\u6587\u672c\u80fd\u529b"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/#\/schema\/person\/image\/","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39475","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=39475"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/39475\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=39475"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=39475"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=39475"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}