<\/section>\n

<\/p>\n

2024\u5e7409\u670822\u65e5 00:02<\/em>\u00a0<\/span>\u9655\u897f<\/span><\/em><\/span><\/section>\n

\n
$\"\"$ <\/p>\n
\u6700\u8fd1\uff0c\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u667a\u80fd\u4f53\u5728\u5404\u4e2a\u9886\u57df\u90fd\u53d6\u5f97\u4e86\u91cd\u5927\u8fdb\u5c55\u3002\u6700\u70ed\u95e8\u7684\u7814\u7a76\u9886\u57df\u4e4b\u4e00\u662f\u5c06\u8fd9\u4e9b\u667a\u80fd\u4f53\u5e94\u7528\u4e8e\u89c6\u9891\u6e38\u620f\u4e2d\u3002\u8fd9\u4e9b\u65b9\u6cd5\u901a\u5e38\u4f9d\u8d56\u4e8e\u6e38\u620f API \u6765\u8bbf\u95ee\u6e38\u620f\u4e2d\u7684\u73af\u5883\u548c\u52a8\u4f5c\u6570\u636e\u3002\u7136\u800c\uff0c\u8fd9\u79cd\u65b9\u6cd5\u53d7\u9650\u4e8e API \u7684\u53ef\u7528\u6027\uff0c\u65e0\u6cd5\u53cd\u6620\u4eba\u7c7b\u73a9\u6e38\u620f\u7684\u65b9\u5f0f\u3002<\/span><\/section>\n
\u968f\u7740\u89c6\u89c9\u8bed\u8a00\u6a21\u578b\uff08VLM\uff09\u7684\u51fa\u73b0\uff0c\u667a\u80fd\u4f53\u73b0\u5728\u62e5\u6709\u4e86\u66f4\u5f3a\u7684\u89c6\u89c9\u7406\u89e3\u80fd\u529b\uff0c\u4f7f\u5176\u80fd\u591f\u4ec5\u4f7f\u7528\u89c6\u89c9\u8f93\u5165\u4e0e\u6e38\u620f\u8fdb\u884c\u4ea4\u4e92\u3002\u5c3d\u7ba1\u53d6\u5f97\u4e86\u8fd9\u4e9b\u8fdb\u6b65\uff0c\u4f46\u76ee\u524d\u7684\u65b9\u6cd5\u5728\u9762\u5411\u52a8\u4f5c\u7684\u4efb\u52a1\u4e2d\u4ecd\u9762\u4e34\u6311\u6218\uff0c\u7279\u522b\u662f\u5728\u52a8\u4f5c\u89d2\u8272\u626e\u6f14\u6e38\u620f\uff08ARPG\uff09\u4e2d\uff0c\u5f3a\u5316\u5b66\u4e60\u65b9\u6cd5\u975e\u5e38\u666e\u904d\uff0c\u4f46\u901a\u7528\u6027\u5dee\uff0c\u9700\u8981\u5927\u91cf\u8bad\u7ec3\u3002<\/span><\/section>\n
\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e9b\u95ee\u9898\uff0c\u963f\u91cc\u56e2\u961f\u9009\u62e9\u4ee5\u300a\u9ed1\u795e\u8bdd\uff1a\u609f\u7a7a\u300b\u4f5c\u4e3a\u7814\u7a76\u5e73\u53f0\uff0c\u63a2\u7d22\u73b0\u6709 VLM \u5728\u9700\u8981\u7eaf\u89c6\u89c9\u8f93\u5165\u548c\u590d\u6742\u52a8\u4f5c\u8f93\u51fa\u7684\u573a\u666f\u4e2d\u7684\u80fd\u529b\u8fb9\u754c\u3002\u4ed6\u4eec\u5728\u6e38\u620f\u4e2d\u5b9a\u4e49\u4e86 12 \u9879\u4efb\u52a1\uff0c\u5176\u4e2d 75% \u4ee5\u6218\u6597\u4e3a\u4e3b\uff0c\u5e76\u5c06\u51e0\u79cd SOTA \u89c6\u89c9\u8bed\u8a00\u6a21\u578b\u7eb3\u5165\u8fd9\u4e00\u57fa\u51c6\u3002\u6b64\u5916\uff0c\u4ed6\u4eec\u8fd8\u5c06\u53d1\u5e03\u4e00\u4e2a\u4eba\u5de5\u64cd\u4f5c\u6570\u636e\u96c6\uff0c\u5176\u4e2d\u5305\u542b\u5f55\u5236\u7684\u6e38\u620f\u89c6\u9891\u548c\u64cd\u4f5c\u65e5\u5fd7\uff0c\u5305\u62ec\u9f20\u6807\u548c\u952e\u76d8\u64cd\u4f5c\u3002\u4ed6\u4eec\u8fd8\u63d0\u51fa\u4e86\u4e00\u4e2a VARP\uff08\u89c6\u89c9\u52a8\u4f5c\u89d2\u8272\u626e\u6f14\uff09\u667a\u80fd\u4f53\u6846\u67b6\uff0c\u7531\u52a8\u4f5c\u89c4\u5212\u7cfb\u7edf\u548c\u89c6\u89c9\u8f68\u8ff9\u7cfb\u7edf\u7ec4\u6210\u3002\u8fd9\u4e00\u6846\u67b6\u5c55\u793a\u4e86\u6267\u884c\u57fa\u672c\u4efb\u52a1\u7684\u80fd\u529b\uff0c\u5e76\u5728 90% \u7684\u7b80\u5355\u548c\u4e2d\u7b49\u6c34\u5e73\u7684\u6218\u6597\u573a\u666f\u4e2d\u53d6\u5f97\u4e86\u6210\u529f\u3002<\/span><\/section>\n
\u8bba\u6587\u94fe\u63a5\uff1a<\/span><\/section>\n
https:\/\/arxiv.org\/abs\/2409.12889<\/span><\/section>\n
GitHub \u5730\u5740\uff1a<\/span><\/section>\n
https:\/\/varp-agent.github.io<\/span><\/section>\n
$\"\"$ <\/p>\n
\u5c0f\u7ea2\u4e66\u63a8\u51fa StoryMaker\uff1a\u5b9e\u73b0\u201c\u6587\u751f\u56fe\u201d\u7684\u7279\u5f81\u6574\u4f53\u4e00\u81f4<\/span><\/section>\n
\u65e0\u9700\u989d\u5916\u5fae\u8c03\uff08Tuning-free\uff09\u7684\u4e2a\u6027\u5316\u56fe\u50cf\u751f\u6210\u65b9\u6cd5\u5728\u4fdd\u6301\u9762\u90e8\u4e00\u81f4\u6027\u65b9\u9762\u53d6\u5f97\u4e86\u5de8\u5927\u6210\u529f\u3002\u7136\u800c\uff0c\u5728\u6709\u591a\u4e2a\u89d2\u8272\u7684\u573a\u666f\u4e2d\uff0c\u7f3a\u4e4f\u6574\u4f53\u4e00\u81f4\u6027\u963b\u788d\u4e86\u8fd9\u4e9b\u65b9\u6cd5\u521b\u9020\u8fde\u8d2f\u53d9\u4e8b\u7684\u80fd\u529b\u3002<\/span><\/section>\n
\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u5c0f\u7ea2\u4e66\u56e2\u961f\u63a8\u51fa\u4e86\u4e00\u79cd\u4e2a\u6027\u5316\u89e3\u51b3\u65b9\u6848\u2014\u2014StoryMaker\uff0c\u5b83\u4e0d\u4ec5\u80fd\u4fdd\u6301\u9762\u90e8\u7684\u4e00\u81f4\u6027\uff0c\u8fd8\u80fd\u4fdd\u6301\u670d\u88c5\u3001\u53d1\u578b\u548c\u8eab\u4f53\u7684\u4e00\u81f4\u6027\uff0c\u4ece\u800c\u901a\u8fc7\u4e00\u7cfb\u5217\u56fe\u50cf\u4fc3\u8fdb\u6545\u4e8b\u7684\u521b\u4f5c\u3002StoryMaker \u878d\u5408\u4e86\u57fa\u4e8e\u9762\u90e8\u8eab\u4efd\u7684\u6761\u4ef6\u548c\u88c1\u526a\u540e\u7684\u4eba\u7269\u56fe\u50cf\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u4ed6\u4eec\u4f7f\u7528\u4f4d\u7f6e\u611f\u77e5\u611f\u77e5\u5668\u91cd\u91c7\u6837\u5668\uff08<\/span>PPR<\/span>\uff09\u5c06\u9762\u90e8\u8eab\u4efd\u4fe1\u606f\u4e0e\u88c1\u526a\u540e\u7684\u4eba\u7269\u56fe\u50cf\u6574\u5408\u5728\u4e00\u8d77\uff0c\u4ece\u800c\u83b7\u5f97\u9c9c\u660e\u7684\u4eba\u7269\u7279\u5f81\u3002\u4e3a\u4e86\u9632\u6b62\u591a\u4e2a\u4eba\u7269\u548c\u80cc\u666f\u6df7\u6742\u5728\u4e00\u8d77\uff0c\u4ed6\u4eec\u4f7f\u7528\u5e26\u6709\u5206\u5272\u63a9\u7801\u7684 MSE \u635f\u5931\u5206\u522b\u9650\u5236\u4e0d\u540c\u4eba\u7269\u548c\u80cc\u666f\u7684\u4ea4\u53c9\u6ce8\u610f\u529b\u5f71\u54cd\u533a\u57df\u3002<\/span><\/span><\/section>\n
\u6b64\u5916\uff0c\u4ed6\u4eec\u4ee5\u59ff\u52bf\u4e3a\u6761\u4ef6\u8bad\u7ec3\u751f\u6210\u7f51\u7edc\uff0c\u4ece\u800c\u4fc3\u8fdb\u4e0e\u59ff\u52bf\u7684\u89e3\u8026\u3002\u4ed6\u4eec\u8fd8\u91c7\u7528\u4e86\u00a0<\/span>LoRA<\/span>\u00a0\u6765\u63d0\u9ad8\u4fdd\u771f\u5ea6\u548c\u8d28\u91cf\u3002<\/span><\/span><\/section>\n
\u8bba\u6587\u94fe\u63a5\uff1a<\/span><\/section>\n
https:\/\/arxiv.org\/abs\/2409.12576<\/span><\/section>\n
GitHub \u5730\u5740\uff1a<\/span><\/section>\n
https:\/\/github.com\/RedAIGC\/StoryMaker<\/span><\/section>\n
$\"\"$ <\/p>\n
\u5b57\u8282\u3001\u4e2d\u79d1\u9662\u56e2\u961f\u63a8\u51fa\u591a\u6a21\u6001\u6570\u5b66\u9884\u8bad\u7ec3\u6570\u636e\u96c6 InfiMM-WebMath-40B<\/span><\/section>\n
\u5728\u5927\u89c4\u6a21\u3001\u9ad8\u8d28\u91cf\u7684\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\u5bf9\u4e8e\u63d0\u9ad8\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u63a8\u7406\u80fd\u529b\u81f3\u5173\u91cd\u8981\uff0c\u5c24\u5176\u662f\u5728\u6570\u5b66\u7b49\u4e13\u4e1a\u9886\u57df\u3002\u5c3d\u7ba1\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\uff08MLLMs\uff09\u7684\u91cd\u8981\u6027\u5df2\u5f97\u5230\u516c\u8ba4\uff0c\u4f46\u8be5\u9886\u57df\u76ee\u524d\u4ecd\u7f3a\u4e4f\u4e13\u95e8\u9488\u5bf9\u6570\u5b66\u63a8\u7406\u7684\u5168\u9762\u5f00\u6e90\u9884\u8bad\u7ec3\u6570\u636e\u96c6\u3002<\/span><\/section>\n
\u4e3a\u4e86\u586b\u8865\u8fd9\u4e00\u7a7a\u767d\uff0c\u6765\u81ea\u5b57\u8282\u8df3\u52a8\u548c\u4e2d\u79d1\u9662\u7684\u7814\u7a76\u56e2\u961f\u63a8\u51fa\u4e86\u4e00\u4e2a\u9ad8\u8d28\u91cf\u7684\u4ea4\u9519\u56fe\u50cf-\u6587\u672c\u6587\u6863\u6570\u636e\u96c6\u2014\u2014InfiMM-WebMath-40B\u3002\u8be5\u6570\u636e\u96c6\u7531 2400 \u4e07\u4e2a\u7f51\u9875\u30018500 \u4e07\u4e2a\u76f8\u5173\u56fe\u7247 URL \u548c 400 \u4ebf\u4e2a\u6587\u672c token \u7ec4\u6210\u3002\u4e3a\u4e86\u8bc1\u660e InfiMM-WebMath-40B \u7684\u9c81\u68d2\u6027\uff0c\u4ed6\u4eec\u5728\u7eaf\u6587\u672c\u548c\u591a\u6a21\u6001\u73af\u5883\u4e2d\u8fdb\u884c\u4e86\u8bc4\u4f30\u3002\u5728\u7eaf\u6587\u672c\u57fa\u51c6\u4e0a\u7684\u8bc4\u4f30\u7ed3\u679c\u8868\u660e\uff0c\u5c3d\u7ba1\u53ea\u4f7f\u7528\u4e86 400 \u4ebf\u4e2a token\uff0c\u4f46\u6570\u636e\u96c6\u663e\u8457\u63d0\u9ad8\u4e86 1.3B \u53c2\u6570\u6a21\u578b\u7684\u6027\u80fd\uff0c\u5176\u7ed3\u679c\u53ef\u4e0e DeepSeekMath-1.3B \u76f8\u5ab2\u7f8e\uff0c\u540e\u8005\u5728\u76f8\u540c\u7684\u6a21\u578b\u89c4\u6a21\u4e0b\u4f7f\u7528\u4e86 1200 \u4ebf\u4e2a token\u3002<\/span><\/section>\n
\u5c3d\u7ba1\u5982\u6b64\uff0c\u968f\u7740\u591a\u6a21\u6001\u6570\u5b66\u9884\u8bad\u7ec3\u6570\u636e\u96c6\u7684\u5f15\u5165\uff0c\u4ed6\u4eec\u7684\u6a21\u578b\u5728\u591a\u6a21\u6001\u6570\u5b66\u57fa\u51c6\uff08\u5982 MathVerse \u548c We-Math\uff09\u4e0a\u521b\u9020\u4e86\u8fbe\u5230\u4e86\u5f00\u6e90\u6a21\u578b SOTA\u3002<\/span><\/section>\n
\u8bba\u6587\u94fe\u63a5\uff1a<\/span><\/section>\n
https:\/\/arxiv.org\/abs\/2409.12568<\/span><\/section>\n
$\"\"$ <\/p>\n
\u7efc\u8ff0\uff1a\u8bed\u8a00\u3001\u8bed\u97f3\u548c\u89c6\u89c9\u4efb\u52a1\u4e2d\u7684\u504f\u597d\u5fae\u8c03<\/span><\/section>\n
\u504f\u597d\u5fae\u8c03\u662f\u4f7f\u6df1\u5ea6\u751f\u6210\u6a21\u578b\u4e0e\u4eba\u7c7b\u504f\u597d\u76f8\u4e00\u81f4\u7684\u5173\u952e\u8fc7\u7a0b\u3002\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u6765\u81ea\u00a0<\/span>Capital One<\/span>\u00a0\u548c\u54e5\u4f26\u6bd4\u4e9a\u5927\u5b66\u7684\u7814\u7a76\u56e2\u961f\u5168\u9762\u6982\u8ff0\u4e86\u504f\u597d\u5fae\u8c03\u548c\u4eba\u7c7b\u53cd\u9988\u6574\u5408\u65b9\u9762\u7684\u6700\u65b0\u8fdb\u5c55\u3002\u5206\u4e3a\u4e09\u4e2a\u4e3b\u8981\u90e8\u5206\uff1a1\uff09\u5f15\u8a00\u548c\u524d\u8a00\uff1a\u4ecb\u7ecd\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\u3001\u504f\u597d\u5fae\u8c03\u4efb\u52a1\u3001\u6a21\u578b\u548c\u5404\u79cd\u6a21\u5f0f\u7684\u6570\u636e\u96c6\uff1a\u8bed\u8a00\u3001\u8bed\u97f3\u548c\u89c6\u89c9\uff0c\u4ee5\u53ca\u4e0d\u540c\u7684\u7b56\u7565\u65b9\u6cd5\uff1b2\uff09\u6df1\u5165\u7814\u7a76\u6bcf\u79cd\u504f\u597d\u5fae\u8c03\u65b9\u6cd5\uff1a\u8be6\u7ec6\u5206\u6790\u504f\u597d\u5fae\u8c03\u4e2d\u4f7f\u7528\u7684\u65b9\u6cd5\uff1b3\uff09\u5e94\u7528\u3001\u8ba8\u8bba\u548c\u672a\u6765\u65b9\u5411\uff1a\u63a2\u8ba8\u504f\u597d\u5fae\u8c03\u5728\u4e0b\u6e38\u4efb\u52a1\u4e2d\u7684\u5e94\u7528\uff0c\u5305\u62ec\u4e0d\u540c\u6a21\u6001\u7684\u8bc4\u4f30\u65b9\u6cd5\uff0c\u4ee5\u53ca\u5bf9\u672a\u6765\u7814\u7a76\u65b9\u5411\u7684\u5c55\u671b\u3002<\/span><\/span><\/section>\n
\u8bba\u6587\u94fe\u63a5\uff1a<\/span><\/section>\n
https:\/\/arxiv.org\/abs\/2409.11564<\/span><\/section>\n
$\"\"$ <\/p>\n
\u5fae\u8f6f\u63d0\u51fa MoE \u8bad\u7ec3\u65b0\u65b9\u6cd5 GRIN<\/span><\/section>\n
\u6df7\u5408\u4e13\u5bb6\u6a21\u578b\uff08MoE\uff09\u901a\u8fc7\u4e13\u5bb6\u8def\u7531\u8fdb\u884c\u7a00\u758f\u8ba1\u7b97\uff0c\u53ea\u9009\u62e9\u6027\u5730\u6fc0\u6d3b\u4e00\u5c0f\u90e8\u5206\u4e13\u5bb6\u6a21\u5757\uff0c\u56e0\u6b64\u6bd4\u7a20\u5bc6\u6a21\u578b\u66f4\u80fd\u6709\u6548\u6269\u5c55\u3002\u7136\u800c\uff0c\u7a00\u758f\u8ba1\u7b97\u5bf9\u4f20\u7edf\u7684\u8bad\u7ec3\u65b9\u6cd5\u63d0\u51fa\u4e86\u6311\u6218\uff0c\u56e0\u4e3a\u79bb\u6563\u7684\u4e13\u5bb6\u8def\u7531\u4f1a\u963b\u788d\u6807\u51c6\u53cd\u5411\u4f20\u64ad\uff0c\u4ece\u800c\u963b\u788d\u57fa\u4e8e\u68af\u5ea6\u7684\u4f18\u5316\uff0c\u800c\u68af\u5ea6\u4f18\u5316\u662f\u6df1\u5ea6\u5b66\u4e60\u7684\u91cd\u8981\u4e00\u73af\u3002<\/span><\/section>\n
\u4e3a\u4e86\u66f4\u597d\u5730\u53d1\u6325 MoE \u7684\u6269\u5c55\u80fd\u529b\uff0c\u5fae\u8f6f\u56e2\u961f\u63d0\u51fa\u4e86 GRIN\uff08GRadient-INformed MoE training\uff09\uff0c\u5b83\u5c06\u7a00\u758f\u68af\u5ea6\u4f30\u8ba1\u7528\u4e8e\u4e13\u5bb6\u8def\u7531\uff0c\u5e76\u914d\u7f6e\u6a21\u578b\u5e76\u884c\u6027\u4ee5\u907f\u514d token \u4e22\u5931\u3002\u5c06 GRIN \u5e94\u7528\u4e8e\u81ea\u56de\u5f52\u8bed\u8a00\u5efa\u6a21\uff0c\u4ed6\u4eec\u5f00\u53d1\u51fa\u4e86\u4e00\u4e2a top-2 16\u00d73.8B MoE \u6a21\u578b\u3002\u8fd9\u4e00\u6a21\u578b\u4ec5\u6709 6.6B \u6fc0\u6d3b\u53c2\u6570\uff0c\u5176\u6027\u80fd\u8d85\u8fc7\u4e86 7B \u7a20\u5bc6\u6a21\u578b\uff0c\u5e76\u4e0e\u5728\u76f8\u540c\u6570\u636e\u4e0a\u8bad\u7ec3\u7684 14B \u5bc6\u96c6\u6a21\u578b\u4e0d\u76f8\u4e0a\u4e0b\u3002<\/span><\/section>\n
\u5bf9\u4e0d\u540c\u4efb\u52a1\u7684\u5e7f\u6cdb\u8bc4\u4f30\u8868\u660e\uff0cGRIN \u6709\u6f5c\u529b\u63d0\u9ad8 MoE \u7684\u6548\u7387\uff0c\u5728 MMLU\u3001HellaSwag\u3001HumanEval \u548c MATH \u4e0a\u5206\u522b\u53d6\u5f97\u4e86 79.4\u300183.7\u300174.4 \u548c 58.9 \u7684\u5206\u6570\u3002<\/span><\/section>\n
\u8bba\u6587\u94fe\u63a5\uff1a<\/span><\/section>\n
https:\/\/arxiv.org\/abs\/2409.12136<\/span><\/section>\n
$\"\"$ <\/p>\n
JourneyBench\uff1a\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\u7684\u89c6\u89c9\u7406\u89e3\u8bc4\u4f30\u57fa\u51c6<\/span><\/section>\n
\u6700\u8fd1\u7684\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\u53ea\u80fd\u4f9d\u9760\u80cc\u666f\u8bed\u8a00\u504f\u5dee\uff0c\u5728\u6d45\u5c42\u89c6\u89c9\u7406\u89e3\u7684\u57fa\u7840\u4e0a\u5b9e\u73b0\u826f\u597d\u7684\u6027\u80fd\u3002\u56e0\u6b64\uff0c\u5728\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u8868\u73b0\u4f18\u5f02\u5e76\u4e0d\u4e00\u5b9a\u4e0e\u89c6\u89c9\u7406\u89e3\u80fd\u529b\u5f3a\u6709\u5173\u3002<\/span><\/section>\n
\u6765\u81ea\u54e5\u4f26\u6bd4\u4e9a\u5927\u5b66\u3001\u5f17\u5409\u5c3c\u4e9a\u7406\u5de5\u5927\u5b66\u548c\u52a0\u5dde\u5927\u5b66\u6d1b\u6749\u77f6\u5206\u6821\u7684\u7814\u7a76\u56e2\u961f\uff0c\u63a8\u51fa\u4e86\u4e00\u4e2a\u7531\u4eba\u7c7b\u6807\u6ce8\u7684\u751f\u6210\u56fe\u50cf\u7684\u7efc\u5408\u57fa\u51c6\u2014\u2014JourneyBench\uff0c\u65e8\u5728\u8bc4\u4f30\u6a21\u578b\u5728\u4ee5\u4e0b\u4e94\u9879\u4efb\u52a1\u4e2d\u7684\u7ec6\u7c92\u5ea6\u591a\u6a21\u6001\u63a8\u7406\u80fd\u529b\uff1a\u4e92\u8865\u591a\u6a21\u6001\u601d\u7ef4\u94fe\u3001\u591a\u56fe\u50cf VQA\u3001\u865a\u6784\u56fe\u50cf\u63cf\u8ff0\u3001\u5e26\u6709\u5e7b\u89c9\u89e6\u53d1\u5668\u7684 VQA\uff0c\u4ee5\u53ca\u5e26\u6709\u7279\u5b9a\u6837\u672c\u5e72\u6270\u9879\u7684\u7ec6\u7c92\u5ea6\u68c0\u7d22\u3002\u4e0e\u73b0\u6709\u57fa\u51c6\u4e0d\u540c\u7684\u662f\uff0cJourneyBench \u660e\u786e\u8981\u6c42\u5728\u4e0d\u5bfb\u5e38\u7684\u60f3\u8c61\u573a\u666f\u4e2d\u8fdb\u884c\u7ec6\u7c92\u5ea6\u591a\u6a21\u6001\u63a8\u7406\uff0c\u800c\u5728\u8fd9\u4e9b\u573a\u666f\u4e2d\uff0c\u8bed\u8a00\u504f\u5dee\u548c\u6574\u4f53\u56fe\u50cf\u8981\u70b9\u662f\u4e0d\u591f\u7684\u3002<\/span><\/section>\n
\u4ed6\u4eec\u5728 JourneyBench \u4e0a\u5bf9 SOTA \u6a21\u578b\u8fdb\u884c\u4e86\u57fa\u51c6\u6d4b\u8bd5\uff0c\u5e76\u4ece\u591a\u4e2a\u7ec6\u7c92\u5ea6\u7ef4\u5ea6\u5bf9\u6027\u80fd\u8fdb\u884c\u4e86\u5206\u6790\u3002\u6240\u6709\u4e94\u9879\u4efb\u52a1\u7684\u7ed3\u679c\u8868\u660e\uff0c\u5373\u4f7f\u5bf9 SOTA \u6765\u8bf4\uff0cJourneyBench \u4e5f\u6781\u5177\u6311\u6218\u6027\uff0c\u8fd9\u8868\u660e\u6a21\u578b\u7684\u89c6\u89c9\u63a8\u7406\u80fd\u529b\u5e76\u4e0d\u50cf\u6700\u521d\u770b\u8d77\u6765\u90a3\u4e48\u5f3a\u3002<\/span><\/section>\n
\u8bba\u6587\u94fe\u63a5\uff1a<\/span><\/section>\n
https:\/\/arxiv.org\/abs\/2409.12953<\/span><\/section>\n
$\"\"$ <\/p>\n
Promptriever\uff1a\u9996\u4e2a\u80fd\u591f\u50cf LM \u4e00\u6837\u8fdb\u884c\u63d0\u793a\u7684\u68c0\u7d22\u6a21\u578b<\/span><\/section>\n
\u7ecf\u8fc7\u6307\u4ee4\u5fae\u8c03\u7684\u8bed\u8a00\u6a21\u578b\uff08LM\uff09\u80fd\u591f\u54cd\u5e94\u6307\u4ee4\u6027\u547d\u4ee4\uff0c\u8fdb\u800c\u63d0\u4f9b\u6bd4\u57fa\u7840\u6a21\u578b\u66f4\u81ea\u7136\u7684\u7528\u6237\u754c\u9762\u3002\u5728\u8fd9\u9879\u5de5\u4f5c\u4e2d\uff0c\u6765\u81ea\u7ea6\u7ff0\u970d\u666e\u91d1\u65af\u5927\u5b66\u548c Samaya AI \u7684\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86\u9996\u4e2a\u80fd\u591f\u50cf LM \u4e00\u6837\u8fdb\u884c\u63d0\u793a\u7684\u68c0\u7d22\u6a21\u578b\u2014\u2014Promptriever\u3002\u4e3a\u4e86\u8bad\u7ec3 Promptriever\uff0c\u4ed6\u4eec\u4ece MS MARCO \u6536\u96c6\u5e76\u53d1\u5e03\u4e86\u4e00\u4e2a\u65b0\u7684\u5b9e\u4f8b\u7ea7\u6307\u4ee4\u8bad\u7ec3\u96c6\uff0c\u6db5\u76d6\u8fd1 50 \u4e07\u4e2a\u5b9e\u4f8b\u3002Promptriever \u4e0d\u4ec5\u5728\u6807\u51c6\u68c0\u7d22\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u8272\uff0c\u800c\u4e14\u8fd8\u80fd\u8ddf\u968f\u6307\u4ee4\u3002<\/span><\/section>\n
\u4ed6\u4eec\u89c2\u5bdf\u5230\uff1a\uff081\uff09\u5728\u8ddf\u968f\u8be6\u7ec6\u7684\u76f8\u5173\u6027\u6307\u4ee4\u65b9\u9762\u53d6\u5f97\u4e86\u5de8\u5927\u8fdb\u6b65\uff08\u8fbe\u5230\u4e86 SoTA\uff09\uff08FollowIR \u4e0a +14.3 p-MRR \/ +3.1 nDCG\uff09\uff0c\uff082\uff09\u5bf9\u67e5\u8be2\u3001\u6307\u4ee4\u4e2d\u8bcd\u6c47\u9009\u62e9\/\u63aa\u8f9e\u7684\u9c81\u68d2\u6027\u663e\u8457\u63d0\u9ad8\uff08InstructIR \u4e0a +12.9 Robustness@10\uff09\uff0c\uff083\uff09\u80fd\u591f\u901a\u8fc7\u63d0\u793a\u6267\u884c\u8d85\u53c2\u6570\u641c\u7d22\uff0c\u4ece\u800c\u53ef\u9760\u5730\u63d0\u9ad8\u68c0\u7d22\u6027\u80fd\uff08BEIR \u4e0a\u5e73\u5747\u63d0\u9ad8 +1.4\uff09\u3002Promptriever \u8bc1\u660e\u4e86\u68c0\u7d22\u6a21\u578b\u53ef\u4ee5\u5728\u6bcf\u6b21\u67e5\u8be2\u7684\u57fa\u7840\u4e0a\u901a\u8fc7\u63d0\u793a\u8fdb\u884c\u63a7\u5236\uff0c\u4e3a\u4eca\u540e\u5c06 LM \u63d0\u793a\u6280\u672f\u4e0e\u4fe1\u606f\u68c0\u7d22\u76f8\u7ed3\u5408\u7684\u5de5\u4f5c\u5960\u5b9a\u4e86\u57fa\u7840\u3002<\/span><\/section>\n
\u8bba\u6587\u94fe\u63a5\uff1a<\/span><\/section>\n
https:\/\/arxiv.org\/abs\/2409.11136<\/span><\/section>\n
$\"\"$ <\/p>\n
\n
\n
\u7279\u522b\u8bf4\u660e\uff1a\u672c\u6587\u4ec5\u7528\u4e8e\u5b66\u672f\u4ea4\u6d41\uff0c\u5982\u6709\u4fb5\u6743\u8bf7\u540e\u53f0\u8054\u7cfb\u5c0f\u7f16\u5220\u9664\u3002<\/span><\/span><\/section>\n<\/section>\n<\/blockquote>\n
<\/section>\n<\/section>\n
<\/p>\n<\/section>\n","protected":false},"excerpt":{"rendered":"
\u5927\u6a21\u578b\u73a9\u300a\u9ed1\u795e\u8bdd\uff1a\u609f\u7a7a\u300b\uff0c\u5b8c\u6210 90% \u7b80\u5355\u3001\u4e2d\u7b49\u6c34\u5e73\u6218\u6597 \u7ffb\u8bd1\u6280\u672f\u6559\u80b2\u4e0e\u7814\u7a76 2024\u5e7409\u6708 […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[391],"tags":[],"class_list":["post-34144","post","type-post","status-publish","format-standard","hentry","category-391"],"_links":{"self":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/34144","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=34144"}],"version-history":[{"count":1,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/34144\/revisions"}],"predecessor-version":[{"id":34153,"href":"https:\/\/linguaresources.com\/index.php?rest_route=\/wp\/v2\/posts\/34144\/revisions\/34153"}],"wp:attachment":[{"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=34144"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=34144"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/linguaresources.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=34144"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}