{"id":1184665,"date":"2026-07-09T18:38:00","date_gmt":"2026-07-10T01:38:00","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-blog-post&#038;p=1184665"},"modified":"2026-09-01T02:42:33","modified_gmt":"2026-09-01T09:42:33","slug":"acl-icml-%e4%b8%8a%e6%96%b0-%e4%bb%8e%e4%b8%96%e7%95%8c%e5%bb%ba%e6%a8%a1%e5%88%b0%e9%ab%98%e6%95%88%e8%ae%ad%e7%bb%83%ef%bc%9a%e8%ae%a9%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%88%90%e4%b8%ba%e6%9b%b4","status":"publish","type":"msr-blog-post","link":"https:\/\/www.microsoft.com\/en-us\/research\/articles\/acl-icml-%e4%b8%8a%e6%96%b0-%e4%bb%8e%e4%b8%96%e7%95%8c%e5%bb%ba%e6%a8%a1%e5%88%b0%e9%ab%98%e6%95%88%e8%ae%ad%e7%bb%83%ef%bc%9a%e8%ae%a9%e5%a4%a7%e6%a8%a1%e5%9e%8b%e6%88%90%e4%b8%ba%e6%9b%b4\/","title":{"rendered":"ACL + ICML \u4e0a\u65b0 | \u4ece\u4e16\u754c\u5efa\u6a21\u5230\u9ad8\u6548\u8bad\u7ec3\uff1a\u8ba9\u5927\u6a21\u578b\u6210\u4e3a\u66f4\u597d\u7684\u667a\u80fd\u4f53"},"content":{"rendered":"\n<p class=\"wp-block-paragraph\"><strong>\u672c\u671f\u5185\u5bb9\u901f\u89c8\uff1a<\/strong><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">1.\u9762\u5411\u89c6\u89c9-\u8bed\u8a00\u63a8\u7406\u7684\u5206\u89e3\u5f0f\u5728\u7ebf\u84b8\u998f\uff1a\u901a\u8fc7\u68af\u5ea6\u5f15\u5bfc\u5b9e\u73b0\u89c6\u89c9\u5b9a\u4f4d\uff08ICML 2026 Spotlight\uff09<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">2.\u4ece\u6587\u5b57\u5230\u4e16\u754c\uff1a\u5927\u8bed\u8a00\u6a21\u578b\u80fd\u5426\u6210\u4e3a\u9690\u5f0f\u7684\u6587\u672c\u4e16\u754c\u6a21\u578b\uff1f\uff08ACL 2026 Oral\uff09<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">3.InfiniteWeb\uff1a\u53ef\u6269\u5c55\u7684\u7f51\u9875\u8bad\u7ec3\u73af\u5883\u5408\u6210\u00a0<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">4.\u8fc8\u5411\u9ad8\u6548\u7684\u5927\u8bed\u8a00\u6a21\u578b\u9000\u706b\u8bad\u7ec3\uff1a\u57fa\u4e8e\u539f\u5219\u6027\u6837\u672c\u9009\u62e9\u7684\u65b9\u6cd5\uff08ICML\u00a02026\u00a0Spotlight\uff09<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">5.\u901a\u5f80\u5728\u7ebf\u76d1\u7763\u5fae\u8c03\uff08On-Policy SFT\uff09\uff1a\u5206\u5e03\u5224\u522b\u7406\u8bba\u53ca\u5176\u5728\u5927\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u4e2d\u7684\u5e94\u7528<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"1-\u9762\u5411\u89c6\u89c9-\u8bed\u8a00\u63a8\u7406\u7684\u5206\u89e3\u5f0f\u5728\u7ebf\u84b8\u998f-\u901a\u8fc7\u68af\u5ea6\u5f15\u5bfc\u5b9e\u73b0\u89c6\u89c9\u5b9a\u4f4d-icml-2026-spotlight\" class=\"wp-block-heading\">1.\u9762\u5411\u89c6\u89c9-\u8bed\u8a00\u63a8\u7406\u7684\u5206\u89e3\u5f0f\u5728\u7ebf\u84b8\u998f\uff1a\u901a\u8fc7\u68af\u5ea6\u5f15\u5bfc\u5b9e\u73b0\u89c6\u89c9\u5b9a\u4f4d\uff08ICML 2026 Spotlight\uff09<\/h2>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"790\" height=\"233\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-1.png\" alt=\"title,image\" class=\"wp-image-1184764\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-1.png 790w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-1-300x88.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-1-768x227.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-1-240x71.png 240w\" sizes=\"auto, (max-width: 790px) 100vw, 790px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2606.00564\">https:\/\/arxiv.org\/abs\/2606.00564<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\u5728\u5229\u7528\u7b56\u7565\u84b8\u998f\u8fdb\u884c\u63a8\u7406\u8bad\u7ec3\u65f6\uff0c\u5f80\u5f80\u53d7\u9650\u4e8e\u5355\u4e00\u7684\u5355\u7247\u5f0f\u76ee\u6807\u51fd\u6570\u3002\u8fd9\u79cd\u65b9\u6cd5\u672a\u80fd\u5145\u5206\u8003\u8651\u591a\u6a21\u6001\u4efb\u52a1\u4e2d\u8bed\u8a00\u5efa\u6a21\u4e0e\u89c6\u89c9\u611f\u77e5\u4e4b\u95f4\u7684\u5185\u5728\u5f20\u529b\uff0c\u5bfc\u81f4\u6a21\u578b\u5728\u5904\u7406\u9700\u6df1\u5165\u7406\u89e3\u56fe\u50cf\u7684\u590d\u6742\u63a8\u7406\u4efb\u52a1\u65f6\uff0c\u5176\u89c6\u89c9\u63a5\u5730\u80fd\u529b\uff08visual grounding\uff09\u5e38\u6210\u4e3a\u6027\u80fd\u63d0\u5347\u7684\u4e3b\u8981\u74f6\u9888\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u89e3\u51b3\u8fd9\u4e00\u96be\u9898\uff0c\u7814\u7a76\u5458\u4eec\u901a\u8fc7\u6570\u5b66\u5206\u89e3\u635f\u5931\u51fd\u6570\uff0c\u63ed\u793a\u4e86\u8bed\u8a00\u5148\u9a8c\u4e0e\u89c6\u89c9\u63a5\u5730\u76ee\u6807\u5728\u51e0\u4f55\u7a7a\u95f4\u4e0a\u8fd1\u4e4e\u6b63\u4ea4\u7684\u7279\u6027\u3002\u5e38\u89c4\u7684\u5355\u7247\u5f0f\u84b8\u998f\u88ab\u52a8\u5e73\u8861\u4e24\u8005\uff0c\u672a\u80fd\u6709\u6548\u5f3a\u5236\u6a21\u578b\u589e\u5f3a\u611f\u77e5\u4fdd\u771f\u5ea6\u3002\u5bf9\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e00\u79cd\u89c6\u89c9\u68af\u5ea6\u5f15\u5bfc\u65b9\u6cd5VGS\uff08Visual Gradient Steering\uff09\u3002\u8be5\u65b9\u6cd5\u5f15\u5165\u8f85\u52a9\u89c6\u89c9\u76ee\u6807\uff0c\u901a\u8fc7\u52a8\u6001\u91cd\u5b9a\u5411\u4f18\u5316\u8f68\u8ff9\uff0c\u663e\u5f0f\u4f18\u5148\u5904\u7406\u89c6\u89c9\u5b50\u7a7a\u95f4\uff0c\u5e76\u7ed3\u5408\u8bed\u8a00\u4fdd\u6301\u6b63\u5219\u5316\u5668\uff0c\u5728\u5f3a\u5316\u89c6\u89c9\u63a5\u5730\u7684\u540c\u65f6\uff0c\u9632\u6b62\u4e86\u5bf9\u539f\u6709\u8bed\u8a00\u63a8\u7406\u80fd\u529b\u7684\u8d1f\u9762\u5f71\u54cd\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"217\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-2.png\" alt=\"diagram\" class=\"wp-image-1184765\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-2.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-2-300x75.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-2-768x193.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-2-240x60.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><figcaption class=\"wp-element-caption\">\u56fe1\uff1a\u89c6\u89c9\u68af\u5ea6\u5f15\u5bfcVGS\u7684\u6982\u8ff0\u3002\uff08\u5de6\uff09\u9488\u5bf9\u591a\u6a21\u6001\u67e5\u8be2\uff08\u56fe\u50cfI\u548c\u63d0\u793ax\uff09\uff0c\u4ece\u5b66\u751f\u6a21\u578b\u7b56\u7565\u4e2d\u91c7\u6837\u4e00\u4e2arollout \u03c4\uff0c\u5e76\u5206\u522b\u5728\u591a\u6a21\u6001\uff08I\uff0cx\uff09\u548c\u5355\u6a21\u6001\uff08\u4ec5\u6587\u672c\uff09\u573a\u666f\u4e0b\u8ba1\u7b97\u5b66\u751f\u6a21\u578b\u4e0e\u6559\u5e08\u6a21\u578b\u7684\u5bf9\u6570\u6982\u7387\u3002\uff08\u4e2d\uff09\u5c06\u6807\u51c6\u7684\u6574\u4f53\u76ee\u6807\u5206\u89e3\u4e3a\u4e24\u4e2a\u72ec\u7acb\u7ec4\u4ef6\uff1a\u8bed\u8a00\u5148\u9a8c\uff08LLang\uff09\u7528\u4e8e\u5339\u914d\u6559\u5e08\u6a21\u578b\u7684\u7eaf\u6587\u672c\u5206\u5e03\uff1b\u89c6\u89c9\u951a\u5b9a\u76ee\u6807\uff08LVis\uff09\u5219\u65e8\u5728\u6784\u5efaq\u2217T\u5206\u5e03\u4ee5\u63d0\u53d6\u6559\u5e08\u6a21\u578b\u7684\u89c6\u89c9\u4fe1\u606f\u589e\u76ca\u3002\uff08\u53f3\uff09\u51e0\u4f55\u5206\u6790\u8868\u660e\uff0c\u6807\u51c6\u84b8\u998f\u65b9\u6cd5\uff08LStandard\uff09\u5728\u8fd9\u4e24\u4e2a\u901a\u5e38\u76f8\u4e92\u72ec\u7acb\u7684\u76ee\u6807\u4e4b\u95f4\u8d77\u5230\u4e86\u88ab\u52a8\u6298\u8877\u7684\u4f5c\u7528\u3002\u800cVGS\u5219\u660e\u786e\u5c06\u68af\u5ea6\u66f4\u65b0\u5bfc\u5411\u89c6\u89c9\u5b50\u7a7a\u95f4\uff08LVis\uff09\uff0c\u4f18\u5148\u8003\u8651\u611f\u77e5\u4fdd\u771f\u5ea6\u800c\u975e\u901a\u7528\u8bed\u8a00\u5efa\u6a21\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9a8c\u8868\u660e\uff0cVGS\u663e\u8457\u63d0\u5347\u4e86\u5b66\u751f\u6a21\u578b\u5728\u591a\u9879\u590d\u6742\u591a\u6a21\u6001\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u8868\u73b0\uff0c\u4e0d\u4ec5\u5728\u7eaf\u84b8\u998f\u8bbe\u7f6e\u4e0b\u6548\u679c\u4f18\u8d8a\uff0c\u8fd8\u53ef\u4f5c\u4e3a\u4e00\u79cd\u6709\u6548\u7684\u6b63\u5219\u5316\u624b\u6bb5\u878d\u5165\u5f3a\u5316\u5b66\u4e60\u5fae\u8c03\u6d41\u7a0b\uff0c\u5728\u4fdd\u6301\u6a21\u578b\u63a8\u7406\u94fe\u8fde\u8d2f\u6027\u7684\u57fa\u7840\u4e0a\uff0c\u663e\u8457\u589e\u5f3a\u4e86\u611f\u77e5\u51c6\u786e\u5ea6\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"2-\u4ece\u6587\u5b57\u5230\u4e16\u754c-\u5927\u8bed\u8a00\u6a21\u578b\u80fd\u5426\u6210\u4e3a\u9690\u5f0f\u7684\u6587\u672c\u4e16\u754c\u6a21\u578b-acl-2026-oral\" class=\"wp-block-heading\">2.\u4ece\u6587\u5b57\u5230\u4e16\u754c\uff1a\u5927\u8bed\u8a00\u6a21\u578b\u80fd\u5426\u6210\u4e3a\u9690\u5f0f\u7684\u6587\u672c\u4e16\u754c\u6a21\u578b\uff1f\uff08ACL 2026 Oral\uff09<\/h2>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"264\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-3.png\" alt=\"title,image\" class=\"wp-image-1184766\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-3.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-3-300x92.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-3-768x234.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-3-240x73.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2512.18832\">https:\/\/arxiv.org\/abs\/2512.18832<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u667a\u80fd\u4f53\u5b66\u4e60\u9ad8\u5ea6\u4f9d\u8d56\u4ea4\u4e92\u7ecf\u9a8c\uff0c\u4f46\u771f\u5b9e\u73af\u5883\u4ea4\u4e92\u6210\u672c\u9ad8\u6602\u3001\u8986\u76d6\u6709\u9650\u4e14\u63a8\u7406\u9636\u6bb5\u52a8\u4f5c\u4e0d\u53ef\u9006\uff0c\u4e25\u91cd\u5236\u7ea6\u4e86\u667a\u80fd\u4f53\u7684\u89c4\u6a21\u5316\u8bad\u7ec3\u4e0e\u53ef\u9760\u51b3\u7b56\u3002\u4e16\u754c\u6a21\u578b\u6709\u671b\u7f13\u89e3\u8fd9\u4e9b\u74f6\u9888\uff0c\u4f46\u5927\u8bed\u8a00\u6a21\u578b\u80fd\u5426\u4f5c\u4e3a\u53ef\u9760\u7684\u4e16\u754c\u6a21\u578b\u5e76\u4e3a\u4e0b\u6e38\u667a\u80fd\u4f53\u5e26\u6765\u5b9e\u8d28\u6536\u76ca\uff0c\u5c1a\u7f3a\u4e4f\u7cfb\u7edf\u6027\u8bc4\u4f30\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5bf9\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a\u7531\u4fdd\u771f\u5ea6\u4e0e\u4e00\u81f4\u6027\u3001\u53ef\u6269\u5c55\u6027\u4e0e\u9c81\u68d2\u6027\u4ee5\u53ca\u667a\u80fd\u4f53\u6548\u80fd\u7ec4\u6210\u7684\u4e09\u5c42\u8bc4\u4f30\u6846\u67b6\u3002\u8be5\u6846\u67b6\u805a\u7126\u6587\u672c\u73af\u5883\uff0c\u5c06\u5176\u4f5c\u4e3a\u8fde\u63a5\u8bed\u8a00\u5efa\u6a21\u4e0e\u4e16\u754c\u5efa\u6a21\u7684\u53d7\u63a7\u7edf\u4e00\u63a5\u53e3\uff0c\u5e76\u5c06\u76ee\u6807\u4ece\u5355\u7eaf\u7684\u4e0b\u4e00\u8bcd\u9884\u6d4b\uff0c\u91cd\u6784\u4e3a\u4ea4\u4e92\u534f\u8bae\u4e0b\u7684\u4e0b\u4e00\u72b6\u6001\u9884\u6d4b\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"568\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-4.png\" alt=\"diagram\" class=\"wp-image-1184767\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-4.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-4-300x197.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-4-768x504.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-4-240x158.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><figcaption class=\"wp-element-caption\">\u56fe2\uff1a\u5927\u8bed\u8a00\u6a21\u578b\u4f5c\u4e3a\u57fa\u4e8e\u6587\u672c\u7684\u667a\u80fd\u4f53\u5b66\u4e60\u4e16\u754c\u6a21\u578b\u3002\u73b0\u5b9e\u4e16\u754c\u7684\u4ea4\u4e92\u8fc7\u7a0b\u6210\u672c\u9ad8\u6602\u3001\u901f\u5ea6\u7f13\u6162\u4e14\u5f80\u5f80\u4e0d\u53ef\u9006\u3002\u901a\u8fc7\u6839\u636e\u7528\u6237\u64cd\u4f5c\u9884\u6d4b\u73af\u5883\u52a8\u6001\uff0c\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u7684\u4e16\u754c\u6a21\u578b\u80fd\u591f\u5b9e\u73b0\u4f4e\u6210\u672c\u7684\u6570\u636e\u5408\u6210\u3001\u53ef\u6269\u5c55\u7684\u5b66\u4e60\u4ee5\u53ca\u53ef\u56de\u6eaf\u7684\u51b3\u7b56\u8fc7\u7a0b\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u901a\u8fc7\u5728\u4e94\u79cd\u4ee3\u8868\u6027\u6587\u672c\u73af\u5883\u4e2d\u5bf9\u8be5\u6846\u67b6\u7684\u8bc4\u4f30\uff0c\u7814\u7a76\u5458\u4eec\u53d1\u73b0\uff1a\u9996\u5148\uff0c\u7ecf\u8fc7\u5145\u5206\u8bad\u7ec3\u7684\u6a21\u578b\u80fd\u591f\u6355\u6349\u73af\u5883\u7684\u5185\u9690\u52a8\u6001\uff0c\u5e76\u5728\u7ed3\u6784\u5316\u8bbe\u7f6e\u4e2d\u8868\u73b0\u51fa\u9ad8\u5ea6\u7684\u4fdd\u771f\u5ea6\uff0c\u800c\u76d1\u7763\u5fae\u8c03\u5bf9\u5b9e\u73b0\u9ad8\u4fdd\u771f\u5efa\u6a21\u81f3\u5173\u91cd\u8981\u3002\u5176\u6b21\uff0c\u4e16\u754c\u6a21\u578b\u6027\u80fd\u968f\u6570\u636e\u91cf\u3001\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u53ca\u73af\u5883\u590d\u6742\u5ea6\u7684\u589e\u52a0\u7cfb\u7edf\u6027\u5730\u6269\u5c55\uff0c\u5c55\u73b0\u51fa\u826f\u597d\u7684\u53ef\u6269\u5c55\u6027\u3002\u6700\u540e\uff0c\u9ad8\u8d28\u91cf\u7684\u4e16\u754c\u6a21\u578b\u80fd\u4e3a\u4e0b\u6e38\u667a\u80fd\u4f53\u5e26\u6765\u660e\u786e\u63d0\u5347\uff0c\u4f8b\u5982\uff0c\u5728WebShop\u73af\u5883\u4e2d\uff0c\u884c\u52a8\u9a8c\u8bc1\u4f7fGPT-4o\u7684\u6027\u80fd\u63d0\u9ad8\u4e865.5%\uff0c\u800c\u5728SciWorld\u73af\u5883\u4e2d\uff0c\u91c7\u7528\u70ed\u542f\u52a8\u5f3a\u5316\u5b66\u4e60\u5b9e\u73b0\u4e86\u7ea615%\u7684\u589e\u76ca\u3002\u8fd9\u8bc1\u5b9e\u4e86\u5c06\u8bed\u8a00\u6a21\u578b\u4f5c\u4e3a\u6587\u672c\u73af\u5883\u4e16\u754c\u6a21\u578b\u7684\u6f5c\u529b\uff0c\u540c\u65f6\u660e\u786e\u4e86\u5176\u5728\u6570\u636e\u5408\u6210\u3001\u6a21\u578b\u5185\u63a8\u7406\u53ca\u53ef\u9760\u51b3\u7b56\u5236\u5b9a\u65b9\u9762\u7684\u5e94\u7528\u4ef7\u503c\uff0c\u4e3a\u63a8\u52a8\u667a\u80fd\u4f53\u5b66\u4e60\u4ece\u6587\u672c\u9886\u57df\u5411\u66f4\u5e7f\u6cdb\u7684\u573a\u666f\u6f14\u8fdb\u5960\u5b9a\u4e86\u5b9e\u8bc1\u57fa\u7840\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"3-infiniteweb-\u53ef\u6269\u5c55\u7684\u7f51\u9875\u8bad\u7ec3\u73af\u5883\u5408\u6210\" class=\"wp-block-heading\">3.InfiniteWeb\uff1a\u53ef\u6269\u5c55\u7684\u7f51\u9875\u8bad\u7ec3\u73af\u5883\u5408\u6210\u00a0<\/h2>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"249\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-5.png\" alt=\"title,image\" class=\"wp-image-1184768\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-5.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-5-300x86.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-5-768x221.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-5-240x69.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2601.04126\">https:\/\/arxiv.org\/abs\/2601.04126<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">GUI\u667a\u80fd\u4f53\u5728\u4ee3\u8868\u7528\u6237\u4e0e\u56fe\u5f62\u754c\u9762\u4ea4\u4e92\u65b9\u9762\u5c55\u73b0\u51fa\u4e86\u5de8\u5927\u7684\u6f5c\u529b\uff0c\u4e5f\u662f\u6784\u5efa\u5b9e\u7528AI\u52a9\u624b\u7684\u91cd\u8981\u65b9\u5411\u3002\u7136\u800c\uff0c\u8bad\u7ec3\u8fd9\u4e9b\u667a\u80fd\u4f53\u7f3a\u4e4f\u80fd\u591f\u652f\u6301\u5927\u89c4\u6a21\u8bad\u7ec3\u7684\u5408\u9002\u73af\u5883\u3002\u5c3d\u7ba1\u5df2\u6709\u4e00\u4e9b\u57fa\u51c6\u6d4b\u8bd5\u63d0\u4f9b\u4e86\u5b9d\u8d35\u7684\u6d4b\u8bd5\u5e73\u53f0\uff0c\u4f46\u5b83\u4eec\u4f7f\u7528\u7684\u771f\u5b9e\u7f51\u7ad9\u7531\u4e8e\u591a\u6837\u6027\u5355\u4e00\u3001\u7f51\u7ad9\u9650\u6d41\u3001\u4efb\u52a1\u4e0d\u53ef\u64a4\u9500\u7b49\u7b49\u95ee\u9898\uff0c\u96be\u4ee5\u4f5c\u4e3a\u8bad\u7ec3\u73af\u5883\u8fdb\u884c\u7a33\u5b9a\u7684\u5f3a\u5316\u5b66\u4e60\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63a8\u51faInfiniteWeb\u7cfb\u7edf\uff0c\u80fd\u591f\u5927\u89c4\u6a21\u81ea\u52a8\u751f\u6210\u9002\u7528\u4e8eGUI\u667a\u80fd\u4f53\u8bad\u7ec3\u7684\u529f\u80fd\u6027\u7f51\u9875\u865a\u62df\u73af\u5883\u3002\u8be5\u7cfb\u7edf\u4ee5\u7f51\u7ad9\u79cd\u5b50\u63cf\u8ff0\u548c\u53c2\u8003\u8bbe\u8ba1\u56fe\u4e3a\u8f93\u5165\uff0c\u901a\u8fc7\u7edf\u4e00\u63a5\u53e3\u89c4\u8303\uff0c\u7531\u4efb\u52a1\u63a8\u5bfc\u5171\u4eab\u6570\u636e\u6a21\u578b\u4e0eAPI\uff0c\u786e\u4fdd\u8de8\u9875\u6570\u636e\u7684\u4e00\u81f4\u6027\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u529f\u80fd\u5c42\u9762\uff0c\u7cfb\u7edf\u91c7\u7528\u4efb\u52a1\u9a71\u52a8\u6d4b\u8bd5\u5f00\u53d1\u6a21\u5f0f\uff0c\u8fed\u4ee3\u4fee\u6b63\u4efb\u52a1\u76f8\u5173\u7684\u4e1a\u52a1\u903b\u8f91\u4ee5\u4fdd\u8bc1\u529f\u80fd\u6b63\u786e\u6027\u3002\u89c6\u89c9\u4e0a\uff0c\u7cfb\u7edf\u7ed3\u5408\u4ece\u771f\u5b9e\u7f51\u9875\u622a\u56fe\u4e2d\u63d0\u53d6\u7684\u89c6\u89c9\u7279\u5f81\u5f15\u5bfc\u524d\u7aef\u751f\u6210\uff0c\u7ef4\u6301\u9875\u9762\u98ce\u683c\u7684\u591a\u6837\u6027\u3002\u6b64\u5916\uff0c\u7cfb\u7edf\u8fd8\u80fd\u81ea\u52a8\u4ea7\u751f\u5e26\u6709\u4ee3\u7801\u63d2\u6869\u7684\u53ef\u9a8c\u8bc1\u4efb\u52a1\u8bc4\u6d4b\u5668\uff0c\u8f93\u51fa\u7a20\u5bc6\u5956\u52b1\u4fe1\u53f7\u4f9b\u667a\u80fd\u4f53\u8bad\u7ec3\u4f7f\u7528\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"354\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-6.png\" alt=\"graphical user interface, website\" class=\"wp-image-1184769\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-6.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-6-300x123.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-6-768x314.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-6-240x98.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><figcaption class=\"wp-element-caption\">\u56fe3\uff1aInfiniteWeb\u751f\u6210\u7684\u591a\u6837\u5316\u7f51\u7ad9\u793a\u4f8b\uff0c\u6db5\u76d6\u7535\u5b50\u5546\u52a1\u3001\u9884\u8ba2\u5e73\u53f0\u3001\u793e\u4ea4\u5a92\u4f53\u7b49\u591a\u4e2a\u9886\u57df\u3002\u6587\u7ae0\u8fdb\u884c\u4e86\u7aef\u5230\u7aef\u7684\u865a\u62df\u73af\u5883\u8bad\u7ec3\u548c\u771f\u5b9e\u73af\u5883\u6d4b\u8bd5\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cInfiniteWeb\u7cfb\u7edf\u5728\u6784\u5efa\u771f\u5b9e\u7f51\u9875\u73af\u5883\u65b9\u9762\u8d85\u8d8a\u4e86\u4e3b\u6d41\u7684\u5546\u7528\u4ee3\u7801\u667a\u80fd\u4f53\u3002\u4ee5\u6b64\u5408\u6210\u73af\u5883\u8bad\u7ec3\u7684GUI\u667a\u80fd\u4f53\uff0c\u5728OSWorld\u3001Online-Mind2Web\u548cMobileWorld\u4e0a\u5206\u522b\u63d0\u53476.9%\u30015.7%\u548c3.9%\u3002\u8fd9\u8bc1\u660e\u5408\u6210\u73af\u5883\u5177\u5907\u8f83\u9ad8\u771f\u5b9e\u6027\u4e0e\u8de8\u5e73\u53f0\u8fc1\u79fb\u4ef7\u503c\uff0c\u4e3a\u5927\u89c4\u6a21GUI\u667a\u80fd\u4f53\u8bad\u7ec3\u63d0\u4f9b\u4e86\u53ef\u6269\u5c55\u7684\u73af\u5883\u5408\u6210\u65b9\u6848\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"4-\u8fc8\u5411\u9ad8\u6548\u7684\u5927\u8bed\u8a00\u6a21\u578b\u9000\u706b\u8bad\u7ec3-\u57fa\u4e8e\u539f\u5219\u6027\u6837\u672c\u9009\u62e9\u7684\u65b9\u6cd5-icml-2026-spotlight\" class=\"wp-block-heading\">4.\u8fc8\u5411\u9ad8\u6548\u7684\u5927\u8bed\u8a00\u6a21\u578b\u9000\u706b\u8bad\u7ec3\uff1a\u57fa\u4e8e\u539f\u5219\u6027\u6837\u672c\u9009\u62e9\u7684\u65b9\u6cd5\uff08ICML\u00a02026\u00a0Spotlight\uff09<\/h2>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"848\" height=\"207\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-7.png\" alt=\"title,image\" class=\"wp-image-1184770\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-7.png 848w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-7-300x73.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-7-768x187.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-7-240x59.png 240w\" sizes=\"auto, (max-width: 848px) 100vw, 848px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2605.31175\">https:\/\/arxiv.org\/abs\/2605.31175<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5927\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u7684\u9000\u706b\u9636\u6bb5\u662f\u6a21\u578b\u6700\u7ec8\u6536\u655b\u5e76\u786e\u7acb\u8d28\u91cf\u7684\u5173\u952e\u65f6\u671f\uff0c\u7136\u800c\u5982\u4f55\u5728\u8be5\u9636\u6bb5\u6709\u6548\u9009\u62e9\u8bad\u7ec3\u6570\u636e\u4ecd\u662f\u4e00\u5927\u6311\u6218\u3002\u5e38\u89c4\u7b56\u7565\u901a\u5e38\u4f9d\u8d56\u4e8e\u9886\u57df\u8fc7\u6ee4\u6216\u4e0a\u4e0b\u6587\u6269\u5c55\u7b49\u7ecf\u9a8c\u542f\u53d1\u5f0f\u65b9\u6cd5\uff0c\u7f3a\u4e4f\u57fa\u4e8e\u4f18\u5316\u7406\u8bba\u7684\u539f\u5219\u6027\u4f9d\u636e\u3002\u9488\u5bf9\u8fd9\u4e00\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u901a\u8fc7\u635f\u5931\u5730\u5f62\u7684\u8c31\u51e0\u4f55\u89c6\u89d2\u6765\u523b\u753b\u9000\u706b\u9636\u6bb5\uff0c\u6307\u51fa\u4f18\u5316\u6536\u655b\u8981\u6c42\u68af\u5ea6\u66f4\u65b0\u6ee1\u8db3\u4e0d\u540c\u7279\u5f81\u65b9\u5411\u4e0a\u7684\u5f02\u6784\u7ea6\u675f\u3002&nbsp;<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86DiReCT\u6846\u67b6\uff0c\u5c06\u9000\u706b\u9636\u6bb5\u7684\u6837\u672c\u9009\u62e9\u91cd\u65b0\u8868\u8ff0\u4e3a\u4e00\u4e2a\u7ea6\u675f\u4f18\u5316\u95ee\u9898\u3002\u8be5\u6846\u67b6\u57fa\u4e8e\u9ed1\u585e\u77e9\u9635\u7684\u8c31\u7279\u6027\uff0c\u5bf9\u6bcf\u4e2a\u6837\u672c\u7684\u68af\u5ea6\u65bd\u52a0\u660e\u786e\u7684\u65b9\u5411\u7ea6\u675f\uff0c\u4ece\u800c\u53ef\u8bc6\u522b\u51fa\u7b26\u5408\u6700\u4f18\u66f2\u7387\u611f\u77e5\u4e0b\u964d\u8def\u5f84\u7684\u6837\u672c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8003\u8651\u5230\u9ad8\u6548\u5b9e\u73b0\uff0cDiReCT\u5229\u7528\u968f\u673a\u6295\u5f71\u6280\u672f\u8fd1\u4f3c\u9ad8\u7ef4\u68af\u5ea6\uff0c\u663e\u8457\u964d\u4f4e\u4e86\u8ba1\u7b97\u5f00\u9500\u3002\u6574\u5957\u6570\u636e\u8c03\u5ea6\u6d41\u7a0b\u4ec5\u9700\u5728\u8bad\u7ec3\u524d\u5bf9\u6570\u636e\u96c6\u505a\u4e00\u6b21\u63a8\u7406\uff0c\u5373\u53ef\u5b8c\u6210\u9884\u8ba1\u7b97\u3002\u6b64\u5916\uff0c\u501f\u52a9\u8fde\u7eed\u677e\u5f1b\u548c\u9010\u6b21\u51f8\u8fd1\u4f3c\uff08SCA\uff09\u6c42\u89e3\u5668\uff0c\u8be5\u65b9\u6cd5\u6709\u6548\u89e3\u51b3\u4e86\u975e\u51f8\u5bf9\u9f50\u7ea6\u675f\u4e0b\u7684\u7ec4\u5408\u4f18\u5316\u96be\u9898\u3002&nbsp;<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"288\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-8.png\" alt=\"diagram\" class=\"wp-image-1184771\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-8.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-8-300x100.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-8-768x256.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-8-240x80.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><figcaption class=\"wp-element-caption\">\u56fe4\uff1a\u4e0d\u540c\u8bad\u7ec3\u9636\u6bb5\u4f18\u5316\u52a8\u6001\u7684\u51e0\u4f55\u5bf9\u6bd4\u3002\u5728\u7a33\u5b9a\u8bad\u7ec3\u9636\u6bb5(a)\uff0c\u8f83\u9ad8\u7684\u5b66\u4e60\u7387\u548c\u672a\u7ecf\u6574\u7406\u7684\u6570\u636e\u4f1a\u5bfc\u81f4\u6a21\u578b\u5728\u635f\u5931\u51fd\u6570\u66f2\u9762\u7684\u9661\u5ced\u533a\u57df\u4ea7\u751f\u9ad8\u65b9\u5dee\u632f\u8361\uff0c\u4ece\u800c\u5bfc\u81f4\u4e0b\u964d\u4fe1\u53f7\u8f83\u5f31\uff1b\u800c\u5728\u9000\u706b\u9636\u6bb5(b)\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u7684\u673a\u5236\u80fd\u591f\u6291\u5236\u6a2a\u5411\u566a\u58f0\u5e76\u6700\u5927\u5316\u7eb5\u5411\u4fe1\u53f7\uff0c\u4f7f\u6a21\u578b\u80fd\u6cbf\u4f4e\u66f2\u7387\u5c71\u8c37\u9ad8\u6548\u5730\u5411\u6700\u4f18\u89e3\u65b9\u5411\u4e0b\u964d\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u5b9e\u9a8c\u8868\u660e\uff0cDiReCT\u5728\u5404\u79cd\u6a21\u578b\u89c4\u6a21\u4e0b\u5747\u80fd\u4e00\u81f4\u8fbe\u5230\u4e1a\u754c\u9886\u5148\u7684\u6027\u80fd\u3002\u5176\u6838\u5fc3\u4f18\u52bf\u5728\u4e8e\u80fd\u591f\u4e3b\u52a8\u9009\u62e9\u90a3\u4e9b\u5728\u5e73\u5766\u5b50\u7a7a\u95f4\u5185\u653e\u5927\u4e0b\u964d\u4fe1\u53f7\uff0c\u540c\u65f6\u6291\u5236\u5c16\u9510\u65b9\u5411\u4e0a\u566a\u58f0\u7684\u6837\u672c\uff0c\u4ece\u800c\u5728\u5bfc\u822a\u51e0\u4f55\u74f6\u9888\u65f6\u4fdd\u6301\u7a33\u5b9a\u6536\u655b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u7406\u8bba\u5206\u6790\u8fdb\u4e00\u6b65\u901a\u8fc7PAC-Bayes\u8fb9\u754c\u8bc1\u660e\u4e86\u8fd9\u79cd\u57fa\u4e8e\u5e73\u5766\u65b9\u5411\u9009\u62e9\u7684\u66f4\u65b0\u80fd\u591f\u6536\u7f29\u8f68\u8ff9\u66f2\u7387\uff0c\u6539\u5584\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u3002\u540c\u65f6\u6570\u636e\u5206\u6790\u63ed\u793a\uff0c\u7b97\u6cd5\u503e\u5411\u4e8e\u9009\u62e9\u90a3\u4e9b\u5177\u6709\u66f4\u9ad8\u9884\u8bad\u7ec3\u635f\u5931\u4e14\u5e8f\u5217\u957f\u5ea6\u8f83\u957f\u7684\u6837\u672c\uff1a\u8fd9\u7c7b\u6837\u672c\u80fd\u63d0\u4f9b\u5145\u8db3\u68af\u5ea6\u80fd\u91cf\uff0c\u63a8\u52a8\u4f18\u5316\u6cbf\u5e73\u5766\u635f\u5931\u66f2\u9762\u6301\u7eed\u52a0\u901f\u6536\u655b\u3002\u8fd9\u4e9b\u7ed3\u679c\u8bc1\u5b9e\uff0c\u76f8\u6bd4\u4e8e\u5355\u7eaf\u8ffd\u6c42\u4fe1\u606f\u5bc6\u5ea6\u7684\u7ecf\u9a8c\u65b9\u6cd5\uff0c\u4f18\u5148\u786e\u4fdd\u66f4\u65b0\u65b9\u5411\u5bf9\u9f50\uff0c\u5bf9\u4e8e\u63d0\u5347\u5927\u6a21\u578b\u6f5c\u80fd\u66f4\u4e3a\u91cd\u8981\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n\n\n\n<h2 id=\"5-\u901a\u5f80\u5728\u7ebf\u76d1\u7763\u5fae\u8c03-\u5206\u5e03\u5224\u522b\u7406\u8bba\u53ca\u5176\u5728\u5927\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u4e2d\u7684\u5e94\u7528\" class=\"wp-block-heading\">5.\u901a\u5f80\u5728\u7ebf\u76d1\u7763\u5fae\u8c03\uff1a\u5206\u5e03\u5224\u522b\u7406\u8bba\u53ca\u5176\u5728\u5927\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u4e2d\u7684\u5e94\u7528<\/h2>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"793\" height=\"251\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-9.png\" alt=\"title,image\" class=\"wp-image-1184772\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-9.png 793w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-9-300x95.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-9-768x243.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-9-240x76.png 240w\" sizes=\"auto, (max-width: 793px) 100vw, 793px\" \/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8bba\u6587\u94fe\u63a5\uff1a<a class=\"msr-external-link glyph-append glyph-append-open-in-new-tab glyph-append-xsmall\" rel=\"noopener noreferrer\" target=\"_blank\" href=\"https:\/\/arxiv.org\/abs\/2602.12222\">https:\/\/arxiv.org\/abs\/2602.12222<span class=\"sr-only\"> (opens in new tab)<\/span><\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u5927\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u7684\u540e\u671f\u9636\u6bb5\uff0c\u76d1\u7763\u5fae\u8c03\u56e0\u5176\u8ba1\u7b97\u9ad8\u6548\u800c\u5e7f\u53d7\u9752\u7750\uff0c\u4f46\u76f8\u8f83\u4e8e\u5f3a\u5316\u5b66\u4e60\uff0c\u5176\u6cdb\u5316\u80fd\u529b\u901a\u5e38\u7a0d\u900a\u4e00\u7b79\u3002\u8fd9\u79cd\u5dee\u8ddd\u4e3b\u8981\u6e90\u4e8e\u5f3a\u5316\u5b66\u4e60\u91c7\u7528\u4e86\u7b56\u7565\u5185\u6570\u636e\uff0c\u80fd\u591f\u66f4\u597d\u5730\u4fdd\u7559\u6a21\u578b\u539f\u6709\u7684\u5206\u5e03\u7279\u5f81\u5e76\u7f13\u89e3\u707e\u96be\u6027\u9057\u5fd8\uff0c\u800c\u6807\u51c6\u76d1\u7763\u5fae\u8c03\u5219\u503e\u5411\u4e8e\u5f3a\u5236\u6a21\u578b\u5e73\u7b49\u5730\u62df\u5408\u6240\u6709\u5916\u90e8\u6570\u636e\uff0c\u5bfc\u81f4\u6a21\u578b\u5728\u5904\u7406\u5206\u5e03\u5dee\u5f02\u8f83\u5927\u7684\u6570\u636e\u65f6\uff0c\u5176\u9884\u8bad\u7ec3\u7684\u77e5\u8bc6\u7ed3\u6784\u4f1a\u53d7\u5230\u635f\u5bb3\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4e3a\u4e86\u5f25\u5408\u8fd9\u4e00\u5dee\u8ddd\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u5206\u5e03\u5224\u522b\u7406\u8bbaDDT\uff0c\u901a\u8fc7\u4fe1\u53f7\u68c0\u6d4b\u7406\u8bba\u63d0\u4f9b\u4e86\u4e00\u79cd\u7ebf\u6027\u9608\u503c\u7684\u51b3\u7b56\u89c6\u89d2\uff0c\u5e76\u786e\u5b9a\u4e86\u4e2d\u5fc3\u5316\u5bf9\u6570\u4f3c\u7136\uff08CLL\uff09\u4f5c\u4e3a\u8861\u91cf\u5206\u5e03\u5224\u522b\u6700\u4f18\u7684\u6807\u51c6\u3002\u5728\u5e8f\u5217\u751f\u6210\u573a\u666f\u4e0b\uff0c\u901a\u8fc7\u6784\u5efaCLL\u7edf\u8ba1\u91cf\u5316\u751f\u6210\u5e8f\u5217\u4e0e\u6a21\u578b\u5206\u5e03\u7684\u5bf9\u9f50\u60c5\u51b5\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"865\" height=\"290\" src=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-10.png\" alt=\"diagram\" class=\"wp-image-1184773\" style=\"width:770px\" srcset=\"https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-10.png 865w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-10-300x101.png 300w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-10-768x257.png 768w, https:\/\/www.microsoft.com\/en-us\/research\/wp-content\/uploads\/2026\/08\/6-10-240x80.png 240w\" sizes=\"auto, (max-width: 865px) 100vw, 865px\" \/><figcaption class=\"wp-element-caption\">\u56fe5\uff1a\u4f7f\u7528\u591a\u79cd\u5148\u8fdb\u5927\u8bed\u8a00\u6a21\u578b\u53ca\u4e0d\u540c\u6570\u636e\u7c7b\u578b\u5bf9\u8be5\u7406\u8bba\u8fdb\u884c\u7ecf\u9a8c\u9a8c\u8bc1\u3002<\/figcaption><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u57fa\u4e8e\u4e0a\u8ff0\u7406\u8bba\uff0c\u7814\u7a76\u5458\u4eec\u5f00\u53d1\u4e86\u4e24\u79cd\u4e92\u8865\u7684\u5e94\u7528\u6280\u672f\uff1a\u4e00\u662f\u5206\u5e03\u5185\u5fae\u8c03\uff08IDFT\uff09\uff0c\u5728\u635f\u5931\u51fd\u6570\u5c42\u9762\uff0c\u901a\u8fc7\u81ea\u9002\u5e94\u8c03\u8282\u673a\u5236\u8d4b\u4e88\u6a21\u578b\u5185\u90e8\u5206\u5e03\u4e2d\u7f6e\u4fe1\u5ea6\u66f4\u9ad8\u7684\u6837\u672c\u66f4\u9ad8\u7684\u6743\u91cd\uff0c\u4ece\u800c\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u4fdd\u62a4\u6a21\u578b\u7ed3\u6784\uff0c\u907f\u514d\u8fc7\u62df\u5408\u5206\u5e03\u4e4b\u5916\u7684\u566a\u58f0\u6216\u6781\u7aef\u6837\u672c\uff1b\u4e8c\u662f\u63d0\u793a\u89e3\u7801\uff08Hinted Decoding\uff09\uff0c\u5728\u6570\u636e\u5c42\u9762\uff0c\u901a\u8fc7\u5728\u89e3\u7801\u8fc7\u7a0b\u4e2d\u52a8\u6001\u5730\u5c06\u8bad\u7ec3\u8bed\u6599\u91cd\u65b0\u5bf9\u9f50\u81f3\u6a21\u578b\u7684\u81ea\u8eab\u5206\u5e03\uff0c\u751f\u6210\u4e0e\u6a21\u578b\u80fd\u529b\u66f4\u5339\u914d\u7684\u8bad\u7ec3\u6837\u672c\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5728\u591a\u4e2a\u524d\u6cbf\u5927\u8bed\u8a00\u6a21\u578b\u53ca\u6570\u636e\u96c6\u4e0a\u7684\u5e7f\u6cdb\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u4e0a\u8ff0\u6846\u67b6\u5728\u4fdd\u6301\u76d1\u7763\u5fae\u8c03\u9ad8\u6548\u6027\u7684\u524d\u63d0\u4e0b\uff0c\u4e0d\u4ec5\u80fd\u591f\u6709\u6548\u7f13\u89e3\u707e\u96be\u6027\u9057\u5fd8\uff0c\u8fd8\u5728\u6cdb\u5316\u6027\u80fd\u4e0a\u8d85\u8fc7\u4e86\u5e38\u7528\u7684\u79bb\u7ebf\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\uff0c\u5982DPO\u548cSimPO\u3002\u8fd9\u79cd\u65b9\u6cd5\u4e3a\u5728\u5f3a\u5316\u5b66\u4e60\u5b9e\u65bd\u96be\u5ea6\u8f83\u5927\u7684\u9886\u57df\uff0c\u63d0\u4f9b\u4e86\u4e00\u79cd\u5177\u5907\u6781\u9ad8\u5b9e\u7528\u4ef7\u503c\u548c\u6570\u636e\u6548\u7387\u7684\u66ff\u4ee3\u65b9\u6848\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u672c\u671f\u5185\u5bb9\u901f\u89c8\uff1a 1.\u9762\u5411\u89c6\u89c9-\u8bed\u8a00\u63a8\u7406\u7684\u5206\u89e3\u5f0f\u5728\u7ebf\u84b8\u998f\uff1a\u901a\u8fc7\u68af\u5ea6\u5f15\u5bfc\u5b9e\u73b0\u89c6\u89c9\u5b9a\u4f4d\uff08ICML 2026 Spotlight\uff09 2.\u4ece\u6587\u5b57\u5230\u4e16\u754c\uff1a\u5927\u8bed\u8a00\u6a21\u578b\u80fd\u5426\u6210\u4e3a\u9690\u5f0f\u7684\u6587\u672c\u4e16\u754c\u6a21\u578b\uff1f\uff08ACL 2026 Oral\uff09 3.InfiniteWeb\uff1a\u53ef\u6269\u5c55\u7684\u7f51\u9875\u8bad\u7ec3\u73af\u5883\u5408\u6210\u00a0 4.\u8fc8\u5411\u9ad8\u6548\u7684\u5927\u8bed\u8a00\u6a21\u578b\u9000\u706b\u8bad\u7ec3\uff1a\u57fa\u4e8e\u539f\u5219\u6027\u6837\u672c\u9009\u62e9\u7684\u65b9\u6cd5\uff08ICML\u00a02026\u00a0Spotlight\uff09 5.\u901a\u5f80\u5728\u7ebf\u76d1\u7763\u5fae\u8c03\uff08On-Policy SFT\uff09\uff1a\u5206\u5e03\u5224\u522b\u7406\u8bba\u53ca\u5176\u5728\u5927\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u4e2d\u7684\u5e94\u7528 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2606.00564 (opens in new tab) \u591a\u6a21\u6001\u5927\u8bed\u8a00\u6a21\u578b\u5728\u5229\u7528\u7b56\u7565\u84b8\u998f\u8fdb\u884c\u63a8\u7406\u8bad\u7ec3\u65f6\uff0c\u5f80\u5f80\u53d7\u9650\u4e8e\u5355\u4e00\u7684\u5355\u7247\u5f0f\u76ee\u6807\u51fd\u6570\u3002\u8fd9\u79cd\u65b9\u6cd5\u672a\u80fd\u5145\u5206\u8003\u8651\u591a\u6a21\u6001\u4efb\u52a1\u4e2d\u8bed\u8a00\u5efa\u6a21\u4e0e\u89c6\u89c9\u611f\u77e5\u4e4b\u95f4\u7684\u5185\u5728\u5f20\u529b\uff0c\u5bfc\u81f4\u6a21\u578b\u5728\u5904\u7406\u9700\u6df1\u5165\u7406\u89e3\u56fe\u50cf\u7684\u590d\u6742\u63a8\u7406\u4efb\u52a1\u65f6\uff0c\u5176\u89c6\u89c9\u63a5\u5730\u80fd\u529b\uff08visual grounding\uff09\u5e38\u6210\u4e3a\u6027\u80fd\u63d0\u5347\u7684\u4e3b\u8981\u74f6\u9888\u3002 \u4e3a\u89e3\u51b3\u8fd9\u4e00\u96be\u9898\uff0c\u7814\u7a76\u5458\u4eec\u901a\u8fc7\u6570\u5b66\u5206\u89e3\u635f\u5931\u51fd\u6570\uff0c\u63ed\u793a\u4e86\u8bed\u8a00\u5148\u9a8c\u4e0e\u89c6\u89c9\u63a5\u5730\u76ee\u6807\u5728\u51e0\u4f55\u7a7a\u95f4\u4e0a\u8fd1\u4e4e\u6b63\u4ea4\u7684\u7279\u6027\u3002\u5e38\u89c4\u7684\u5355\u7247\u5f0f\u84b8\u998f\u88ab\u52a8\u5e73\u8861\u4e24\u8005\uff0c\u672a\u80fd\u6709\u6548\u5f3a\u5236\u6a21\u578b\u589e\u5f3a\u611f\u77e5\u4fdd\u771f\u5ea6\u3002\u5bf9\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e00\u79cd\u89c6\u89c9\u68af\u5ea6\u5f15\u5bfc\u65b9\u6cd5VGS\uff08Visual Gradient Steering\uff09\u3002\u8be5\u65b9\u6cd5\u5f15\u5165\u8f85\u52a9\u89c6\u89c9\u76ee\u6807\uff0c\u901a\u8fc7\u52a8\u6001\u91cd\u5b9a\u5411\u4f18\u5316\u8f68\u8ff9\uff0c\u663e\u5f0f\u4f18\u5148\u5904\u7406\u89c6\u89c9\u5b50\u7a7a\u95f4\uff0c\u5e76\u7ed3\u5408\u8bed\u8a00\u4fdd\u6301\u6b63\u5219\u5316\u5668\uff0c\u5728\u5f3a\u5316\u89c6\u89c9\u63a5\u5730\u7684\u540c\u65f6\uff0c\u9632\u6b62\u4e86\u5bf9\u539f\u6709\u8bed\u8a00\u63a8\u7406\u80fd\u529b\u7684\u8d1f\u9762\u5f71\u54cd\u3002 \u5b9e\u9a8c\u8868\u660e\uff0cVGS\u663e\u8457\u63d0\u5347\u4e86\u5b66\u751f\u6a21\u578b\u5728\u591a\u9879\u590d\u6742\u591a\u6a21\u6001\u63a8\u7406\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u7684\u8868\u73b0\uff0c\u4e0d\u4ec5\u5728\u7eaf\u84b8\u998f\u8bbe\u7f6e\u4e0b\u6548\u679c\u4f18\u8d8a\uff0c\u8fd8\u53ef\u4f5c\u4e3a\u4e00\u79cd\u6709\u6548\u7684\u6b63\u5219\u5316\u624b\u6bb5\u878d\u5165\u5f3a\u5316\u5b66\u4e60\u5fae\u8c03\u6d41\u7a0b\uff0c\u5728\u4fdd\u6301\u6a21\u578b\u63a8\u7406\u94fe\u8fde\u8d2f\u6027\u7684\u57fa\u7840\u4e0a\uff0c\u663e\u8457\u589e\u5f3a\u4e86\u611f\u77e5\u51c6\u786e\u5ea6\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2512.18832 (opens in new tab) \u667a\u80fd\u4f53\u5b66\u4e60\u9ad8\u5ea6\u4f9d\u8d56\u4ea4\u4e92\u7ecf\u9a8c\uff0c\u4f46\u771f\u5b9e\u73af\u5883\u4ea4\u4e92\u6210\u672c\u9ad8\u6602\u3001\u8986\u76d6\u6709\u9650\u4e14\u63a8\u7406\u9636\u6bb5\u52a8\u4f5c\u4e0d\u53ef\u9006\uff0c\u4e25\u91cd\u5236\u7ea6\u4e86\u667a\u80fd\u4f53\u7684\u89c4\u6a21\u5316\u8bad\u7ec3\u4e0e\u53ef\u9760\u51b3\u7b56\u3002\u4e16\u754c\u6a21\u578b\u6709\u671b\u7f13\u89e3\u8fd9\u4e9b\u74f6\u9888\uff0c\u4f46\u5927\u8bed\u8a00\u6a21\u578b\u80fd\u5426\u4f5c\u4e3a\u53ef\u9760\u7684\u4e16\u754c\u6a21\u578b\u5e76\u4e3a\u4e0b\u6e38\u667a\u80fd\u4f53\u5e26\u6765\u5b9e\u8d28\u6536\u76ca\uff0c\u5c1a\u7f3a\u4e4f\u7cfb\u7edf\u6027\u8bc4\u4f30\u3002 \u5bf9\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a\u7531\u4fdd\u771f\u5ea6\u4e0e\u4e00\u81f4\u6027\u3001\u53ef\u6269\u5c55\u6027\u4e0e\u9c81\u68d2\u6027\u4ee5\u53ca\u667a\u80fd\u4f53\u6548\u80fd\u7ec4\u6210\u7684\u4e09\u5c42\u8bc4\u4f30\u6846\u67b6\u3002\u8be5\u6846\u67b6\u805a\u7126\u6587\u672c\u73af\u5883\uff0c\u5c06\u5176\u4f5c\u4e3a\u8fde\u63a5\u8bed\u8a00\u5efa\u6a21\u4e0e\u4e16\u754c\u5efa\u6a21\u7684\u53d7\u63a7\u7edf\u4e00\u63a5\u53e3\uff0c\u5e76\u5c06\u76ee\u6807\u4ece\u5355\u7eaf\u7684\u4e0b\u4e00\u8bcd\u9884\u6d4b\uff0c\u91cd\u6784\u4e3a\u4ea4\u4e92\u534f\u8bae\u4e0b\u7684\u4e0b\u4e00\u72b6\u6001\u9884\u6d4b\u3002 \u901a\u8fc7\u5728\u4e94\u79cd\u4ee3\u8868\u6027\u6587\u672c\u73af\u5883\u4e2d\u5bf9\u8be5\u6846\u67b6\u7684\u8bc4\u4f30\uff0c\u7814\u7a76\u5458\u4eec\u53d1\u73b0\uff1a\u9996\u5148\uff0c\u7ecf\u8fc7\u5145\u5206\u8bad\u7ec3\u7684\u6a21\u578b\u80fd\u591f\u6355\u6349\u73af\u5883\u7684\u5185\u9690\u52a8\u6001\uff0c\u5e76\u5728\u7ed3\u6784\u5316\u8bbe\u7f6e\u4e2d\u8868\u73b0\u51fa\u9ad8\u5ea6\u7684\u4fdd\u771f\u5ea6\uff0c\u800c\u76d1\u7763\u5fae\u8c03\u5bf9\u5b9e\u73b0\u9ad8\u4fdd\u771f\u5efa\u6a21\u81f3\u5173\u91cd\u8981\u3002\u5176\u6b21\uff0c\u4e16\u754c\u6a21\u578b\u6027\u80fd\u968f\u6570\u636e\u91cf\u3001\u6a21\u578b\u53c2\u6570\u89c4\u6a21\u53ca\u73af\u5883\u590d\u6742\u5ea6\u7684\u589e\u52a0\u7cfb\u7edf\u6027\u5730\u6269\u5c55\uff0c\u5c55\u73b0\u51fa\u826f\u597d\u7684\u53ef\u6269\u5c55\u6027\u3002\u6700\u540e\uff0c\u9ad8\u8d28\u91cf\u7684\u4e16\u754c\u6a21\u578b\u80fd\u4e3a\u4e0b\u6e38\u667a\u80fd\u4f53\u5e26\u6765\u660e\u786e\u63d0\u5347\uff0c\u4f8b\u5982\uff0c\u5728WebShop\u73af\u5883\u4e2d\uff0c\u884c\u52a8\u9a8c\u8bc1\u4f7fGPT-4o\u7684\u6027\u80fd\u63d0\u9ad8\u4e865.5%\uff0c\u800c\u5728SciWorld\u73af\u5883\u4e2d\uff0c\u91c7\u7528\u70ed\u542f\u52a8\u5f3a\u5316\u5b66\u4e60\u5b9e\u73b0\u4e86\u7ea615%\u7684\u589e\u76ca\u3002\u8fd9\u8bc1\u5b9e\u4e86\u5c06\u8bed\u8a00\u6a21\u578b\u4f5c\u4e3a\u6587\u672c\u73af\u5883\u4e16\u754c\u6a21\u578b\u7684\u6f5c\u529b\uff0c\u540c\u65f6\u660e\u786e\u4e86\u5176\u5728\u6570\u636e\u5408\u6210\u3001\u6a21\u578b\u5185\u63a8\u7406\u53ca\u53ef\u9760\u51b3\u7b56\u5236\u5b9a\u65b9\u9762\u7684\u5e94\u7528\u4ef7\u503c\uff0c\u4e3a\u63a8\u52a8\u667a\u80fd\u4f53\u5b66\u4e60\u4ece\u6587\u672c\u9886\u57df\u5411\u66f4\u5e7f\u6cdb\u7684\u573a\u666f\u6f14\u8fdb\u5960\u5b9a\u4e86\u5b9e\u8bc1\u57fa\u7840\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2601.04126 (opens in new tab) GUI\u667a\u80fd\u4f53\u5728\u4ee3\u8868\u7528\u6237\u4e0e\u56fe\u5f62\u754c\u9762\u4ea4\u4e92\u65b9\u9762\u5c55\u73b0\u51fa\u4e86\u5de8\u5927\u7684\u6f5c\u529b\uff0c\u4e5f\u662f\u6784\u5efa\u5b9e\u7528AI\u52a9\u624b\u7684\u91cd\u8981\u65b9\u5411\u3002\u7136\u800c\uff0c\u8bad\u7ec3\u8fd9\u4e9b\u667a\u80fd\u4f53\u7f3a\u4e4f\u80fd\u591f\u652f\u6301\u5927\u89c4\u6a21\u8bad\u7ec3\u7684\u5408\u9002\u73af\u5883\u3002\u5c3d\u7ba1\u5df2\u6709\u4e00\u4e9b\u57fa\u51c6\u6d4b\u8bd5\u63d0\u4f9b\u4e86\u5b9d\u8d35\u7684\u6d4b\u8bd5\u5e73\u53f0\uff0c\u4f46\u5b83\u4eec\u4f7f\u7528\u7684\u771f\u5b9e\u7f51\u7ad9\u7531\u4e8e\u591a\u6837\u6027\u5355\u4e00\u3001\u7f51\u7ad9\u9650\u6d41\u3001\u4efb\u52a1\u4e0d\u53ef\u64a4\u9500\u7b49\u7b49\u95ee\u9898\uff0c\u96be\u4ee5\u4f5c\u4e3a\u8bad\u7ec3\u73af\u5883\u8fdb\u884c\u7a33\u5b9a\u7684\u5f3a\u5316\u5b66\u4e60\u3002 \u4e3a\u6b64\uff0c\u7814\u7a76\u5458\u4eec\u63a8\u51faInfiniteWeb\u7cfb\u7edf\uff0c\u80fd\u591f\u5927\u89c4\u6a21\u81ea\u52a8\u751f\u6210\u9002\u7528\u4e8eGUI\u667a\u80fd\u4f53\u8bad\u7ec3\u7684\u529f\u80fd\u6027\u7f51\u9875\u865a\u62df\u73af\u5883\u3002\u8be5\u7cfb\u7edf\u4ee5\u7f51\u7ad9\u79cd\u5b50\u63cf\u8ff0\u548c\u53c2\u8003\u8bbe\u8ba1\u56fe\u4e3a\u8f93\u5165\uff0c\u901a\u8fc7\u7edf\u4e00\u63a5\u53e3\u89c4\u8303\uff0c\u7531\u4efb\u52a1\u63a8\u5bfc\u5171\u4eab\u6570\u636e\u6a21\u578b\u4e0eAPI\uff0c\u786e\u4fdd\u8de8\u9875\u6570\u636e\u7684\u4e00\u81f4\u6027\u3002 \u5728\u529f\u80fd\u5c42\u9762\uff0c\u7cfb\u7edf\u91c7\u7528\u4efb\u52a1\u9a71\u52a8\u6d4b\u8bd5\u5f00\u53d1\u6a21\u5f0f\uff0c\u8fed\u4ee3\u4fee\u6b63\u4efb\u52a1\u76f8\u5173\u7684\u4e1a\u52a1\u903b\u8f91\u4ee5\u4fdd\u8bc1\u529f\u80fd\u6b63\u786e\u6027\u3002\u89c6\u89c9\u4e0a\uff0c\u7cfb\u7edf\u7ed3\u5408\u4ece\u771f\u5b9e\u7f51\u9875\u622a\u56fe\u4e2d\u63d0\u53d6\u7684\u89c6\u89c9\u7279\u5f81\u5f15\u5bfc\u524d\u7aef\u751f\u6210\uff0c\u7ef4\u6301\u9875\u9762\u98ce\u683c\u7684\u591a\u6837\u6027\u3002\u6b64\u5916\uff0c\u7cfb\u7edf\u8fd8\u80fd\u81ea\u52a8\u4ea7\u751f\u5e26\u6709\u4ee3\u7801\u63d2\u6869\u7684\u53ef\u9a8c\u8bc1\u4efb\u52a1\u8bc4\u6d4b\u5668\uff0c\u8f93\u51fa\u7a20\u5bc6\u5956\u52b1\u4fe1\u53f7\u4f9b\u667a\u80fd\u4f53\u8bad\u7ec3\u4f7f\u7528\u3002 \u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cInfiniteWeb\u7cfb\u7edf\u5728\u6784\u5efa\u771f\u5b9e\u7f51\u9875\u73af\u5883\u65b9\u9762\u8d85\u8d8a\u4e86\u4e3b\u6d41\u7684\u5546\u7528\u4ee3\u7801\u667a\u80fd\u4f53\u3002\u4ee5\u6b64\u5408\u6210\u73af\u5883\u8bad\u7ec3\u7684GUI\u667a\u80fd\u4f53\uff0c\u5728OSWorld\u3001Online-Mind2Web\u548cMobileWorld\u4e0a\u5206\u522b\u63d0\u53476.9%\u30015.7%\u548c3.9%\u3002\u8fd9\u8bc1\u660e\u5408\u6210\u73af\u5883\u5177\u5907\u8f83\u9ad8\u771f\u5b9e\u6027\u4e0e\u8de8\u5e73\u53f0\u8fc1\u79fb\u4ef7\u503c\uff0c\u4e3a\u5927\u89c4\u6a21GUI\u667a\u80fd\u4f53\u8bad\u7ec3\u63d0\u4f9b\u4e86\u53ef\u6269\u5c55\u7684\u73af\u5883\u5408\u6210\u65b9\u6848\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2605.31175 (opens in new tab) \u5927\u8bed\u8a00\u6a21\u578b\u9884\u8bad\u7ec3\u7684\u9000\u706b\u9636\u6bb5\u662f\u6a21\u578b\u6700\u7ec8\u6536\u655b\u5e76\u786e\u7acb\u8d28\u91cf\u7684\u5173\u952e\u65f6\u671f\uff0c\u7136\u800c\u5982\u4f55\u5728\u8be5\u9636\u6bb5\u6709\u6548\u9009\u62e9\u8bad\u7ec3\u6570\u636e\u4ecd\u662f\u4e00\u5927\u6311\u6218\u3002\u5e38\u89c4\u7b56\u7565\u901a\u5e38\u4f9d\u8d56\u4e8e\u9886\u57df\u8fc7\u6ee4\u6216\u4e0a\u4e0b\u6587\u6269\u5c55\u7b49\u7ecf\u9a8c\u542f\u53d1\u5f0f\u65b9\u6cd5\uff0c\u7f3a\u4e4f\u57fa\u4e8e\u4f18\u5316\u7406\u8bba\u7684\u539f\u5219\u6027\u4f9d\u636e\u3002\u9488\u5bf9\u8fd9\u4e00\u95ee\u9898\uff0c\u7814\u7a76\u5458\u4eec\u901a\u8fc7\u635f\u5931\u5730\u5f62\u7684\u8c31\u51e0\u4f55\u89c6\u89d2\u6765\u523b\u753b\u9000\u706b\u9636\u6bb5\uff0c\u6307\u51fa\u4f18\u5316\u6536\u655b\u8981\u6c42\u68af\u5ea6\u66f4\u65b0\u6ee1\u8db3\u4e0d\u540c\u7279\u5f81\u65b9\u5411\u4e0a\u7684\u5f02\u6784\u7ea6\u675f\u3002&nbsp; \u4e3a\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u7814\u7a76\u5458\u4eec\u63d0\u51fa\u4e86DiReCT\u6846\u67b6\uff0c\u5c06\u9000\u706b\u9636\u6bb5\u7684\u6837\u672c\u9009\u62e9\u91cd\u65b0\u8868\u8ff0\u4e3a\u4e00\u4e2a\u7ea6\u675f\u4f18\u5316\u95ee\u9898\u3002\u8be5\u6846\u67b6\u57fa\u4e8e\u9ed1\u585e\u77e9\u9635\u7684\u8c31\u7279\u6027\uff0c\u5bf9\u6bcf\u4e2a\u6837\u672c\u7684\u68af\u5ea6\u65bd\u52a0\u660e\u786e\u7684\u65b9\u5411\u7ea6\u675f\uff0c\u4ece\u800c\u53ef\u8bc6\u522b\u51fa\u7b26\u5408\u6700\u4f18\u66f2\u7387\u611f\u77e5\u4e0b\u964d\u8def\u5f84\u7684\u6837\u672c\u3002 \u8003\u8651\u5230\u9ad8\u6548\u5b9e\u73b0\uff0cDiReCT\u5229\u7528\u968f\u673a\u6295\u5f71\u6280\u672f\u8fd1\u4f3c\u9ad8\u7ef4\u68af\u5ea6\uff0c\u663e\u8457\u964d\u4f4e\u4e86\u8ba1\u7b97\u5f00\u9500\u3002\u6574\u5957\u6570\u636e\u8c03\u5ea6\u6d41\u7a0b\u4ec5\u9700\u5728\u8bad\u7ec3\u524d\u5bf9\u6570\u636e\u96c6\u505a\u4e00\u6b21\u63a8\u7406\uff0c\u5373\u53ef\u5b8c\u6210\u9884\u8ba1\u7b97\u3002\u6b64\u5916\uff0c\u501f\u52a9\u8fde\u7eed\u677e\u5f1b\u548c\u9010\u6b21\u51f8\u8fd1\u4f3c\uff08SCA\uff09\u6c42\u89e3\u5668\uff0c\u8be5\u65b9\u6cd5\u6709\u6548\u89e3\u51b3\u4e86\u975e\u51f8\u5bf9\u9f50\u7ea6\u675f\u4e0b\u7684\u7ec4\u5408\u4f18\u5316\u96be\u9898\u3002&nbsp; \u5b9e\u9a8c\u8868\u660e\uff0cDiReCT\u5728\u5404\u79cd\u6a21\u578b\u89c4\u6a21\u4e0b\u5747\u80fd\u4e00\u81f4\u8fbe\u5230\u4e1a\u754c\u9886\u5148\u7684\u6027\u80fd\u3002\u5176\u6838\u5fc3\u4f18\u52bf\u5728\u4e8e\u80fd\u591f\u4e3b\u52a8\u9009\u62e9\u90a3\u4e9b\u5728\u5e73\u5766\u5b50\u7a7a\u95f4\u5185\u653e\u5927\u4e0b\u964d\u4fe1\u53f7\uff0c\u540c\u65f6\u6291\u5236\u5c16\u9510\u65b9\u5411\u4e0a\u566a\u58f0\u7684\u6837\u672c\uff0c\u4ece\u800c\u5728\u5bfc\u822a\u51e0\u4f55\u74f6\u9888\u65f6\u4fdd\u6301\u7a33\u5b9a\u6536\u655b\u3002 \u7406\u8bba\u5206\u6790\u8fdb\u4e00\u6b65\u901a\u8fc7PAC-Bayes\u8fb9\u754c\u8bc1\u660e\u4e86\u8fd9\u79cd\u57fa\u4e8e\u5e73\u5766\u65b9\u5411\u9009\u62e9\u7684\u66f4\u65b0\u80fd\u591f\u6536\u7f29\u8f68\u8ff9\u66f2\u7387\uff0c\u6539\u5584\u6a21\u578b\u7684\u6cdb\u5316\u80fd\u529b\u3002\u540c\u65f6\u6570\u636e\u5206\u6790\u63ed\u793a\uff0c\u7b97\u6cd5\u503e\u5411\u4e8e\u9009\u62e9\u90a3\u4e9b\u5177\u6709\u66f4\u9ad8\u9884\u8bad\u7ec3\u635f\u5931\u4e14\u5e8f\u5217\u957f\u5ea6\u8f83\u957f\u7684\u6837\u672c\uff1a\u8fd9\u7c7b\u6837\u672c\u80fd\u63d0\u4f9b\u5145\u8db3\u68af\u5ea6\u80fd\u91cf\uff0c\u63a8\u52a8\u4f18\u5316\u6cbf\u5e73\u5766\u635f\u5931\u66f2\u9762\u6301\u7eed\u52a0\u901f\u6536\u655b\u3002\u8fd9\u4e9b\u7ed3\u679c\u8bc1\u5b9e\uff0c\u76f8\u6bd4\u4e8e\u5355\u7eaf\u8ffd\u6c42\u4fe1\u606f\u5bc6\u5ea6\u7684\u7ecf\u9a8c\u65b9\u6cd5\uff0c\u4f18\u5148\u786e\u4fdd\u66f4\u65b0\u65b9\u5411\u5bf9\u9f50\uff0c\u5bf9\u4e8e\u63d0\u5347\u5927\u6a21\u578b\u6f5c\u80fd\u66f4\u4e3a\u91cd\u8981\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/abs\/2602.12222 (opens in new tab) \u5728\u5927\u8bed\u8a00\u6a21\u578b\u8bad\u7ec3\u7684\u540e\u671f\u9636\u6bb5\uff0c\u76d1\u7763\u5fae\u8c03\u56e0\u5176\u8ba1\u7b97\u9ad8\u6548\u800c\u5e7f\u53d7\u9752\u7750\uff0c\u4f46\u76f8\u8f83\u4e8e\u5f3a\u5316\u5b66\u4e60\uff0c\u5176\u6cdb\u5316\u80fd\u529b\u901a\u5e38\u7a0d\u900a\u4e00\u7b79\u3002\u8fd9\u79cd\u5dee\u8ddd\u4e3b\u8981\u6e90\u4e8e\u5f3a\u5316\u5b66\u4e60\u91c7\u7528\u4e86\u7b56\u7565\u5185\u6570\u636e\uff0c\u80fd\u591f\u66f4\u597d\u5730\u4fdd\u7559\u6a21\u578b\u539f\u6709\u7684\u5206\u5e03\u7279\u5f81\u5e76\u7f13\u89e3\u707e\u96be\u6027\u9057\u5fd8\uff0c\u800c\u6807\u51c6\u76d1\u7763\u5fae\u8c03\u5219\u503e\u5411\u4e8e\u5f3a\u5236\u6a21\u578b\u5e73\u7b49\u5730\u62df\u5408\u6240\u6709\u5916\u90e8\u6570\u636e\uff0c\u5bfc\u81f4\u6a21\u578b\u5728\u5904\u7406\u5206\u5e03\u5dee\u5f02\u8f83\u5927\u7684\u6570\u636e\u65f6\uff0c\u5176\u9884\u8bad\u7ec3\u7684\u77e5\u8bc6\u7ed3\u6784\u4f1a\u53d7\u5230\u635f\u5bb3\u3002 [&hellip;]<\/p>\n","protected":false},"author":44093,"featured_media":1184775,"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":[13561,13556,13545],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1184665","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-research-area-algorithms","msr-research-area-artificial-intelligence","msr-research-area-human-language-technologies","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\/1184665","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\/44093"}],"version-history":[{"count":3,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1184665\/revisions"}],"predecessor-version":[{"id":1184776,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1184665\/revisions\/1184776"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media\/1184775"}],"wp:attachment":[{"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1184665"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1184665"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1184665"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/www.microsoft.com\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1184665"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}