{"id":5389,"date":"2026-06-01T09:00:15","date_gmt":"2026-06-01T01:00:15","guid":{"rendered":"https:\/\/1ow.top\/index.php\/2026\/06\/01\/%e5%88%ab%e5%85%89%e7%bb%99agent%e5%8a%a0tool%e4%ba%86%ef%bc%8c%e5%ae%83%e6%a0%b9%e6%9c%ac%e9%80%89%e4%b8%8d%e6%98%8e%e7%99%bd%ef%bc%81%e5%a4%8d%e6%97%a6x%e9%80%9a%e4%b9%89%e6%8f%90%e5%87%ba\/"},"modified":"2026-06-01T09:00:15","modified_gmt":"2026-06-01T01:00:15","slug":"%e5%88%ab%e5%85%89%e7%bb%99agent%e5%8a%a0tool%e4%ba%86%ef%bc%8c%e5%ae%83%e6%a0%b9%e6%9c%ac%e9%80%89%e4%b8%8d%e6%98%8e%e7%99%bd%ef%bc%81%e5%a4%8d%e6%97%a6x%e9%80%9a%e4%b9%89%e6%8f%90%e5%87%ba","status":"publish","type":"post","link":"https:\/\/1ow.top\/index.php\/2026\/06\/01\/%e5%88%ab%e5%85%89%e7%bb%99agent%e5%8a%a0tool%e4%ba%86%ef%bc%8c%e5%ae%83%e6%a0%b9%e6%9c%ac%e9%80%89%e4%b8%8d%e6%98%8e%e7%99%bd%ef%bc%81%e5%a4%8d%e6%97%a6x%e9%80%9a%e4%b9%89%e6%8f%90%e5%87%ba\/","title":{"rendered":"\u522b\u5149\u7ed9Agent\u52a0Tool\u4e86\uff0c\u5b83\u6839\u672c\u9009\u4e0d\u660e\u767d\uff01\u590d\u65e6\u00d7\u901a\u4e49\u63d0\u51fa\u5168\u65b0CUA\u8bad\u7ec3\u8303\u5f0f"},"content":{"rendered":"<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/06\/8e2f1aea4f0bc98b3bbfccee4ad63f23.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/06\/8e2f1aea4f0bc98b3bbfccee4ad63f23.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\u522b\u5149\u7ed9Agent\u52a0Tool\u4e86\uff0c\u5b83\u6839\u672c\u9009\u4e0d\u660e\u767d\uff01\u590d\u65e6\u00d7\u901a\u4e49\u63d0\u51fa\u5168\u65b0CUA\u8bad\u7ec3\u8303\u5f0f\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/06\/eb90b8fdae55c7c42f5dde09425d7b0a.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/06\/eb90b8fdae55c7c42f5dde09425d7b0a.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\u522b\u5149\u7ed9Agent\u52a0Tool\u4e86\uff0c\u5b83\u6839\u672c\u9009\u4e0d\u660e\u767d\uff01\u590d\u65e6\u00d7\u901a\u4e49\u63d0\u51fa\u5168\u65b0CUA\u8bad\u7ec3\u8303\u5f0f\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/06\/02613396e0642e5b88a928a788f48b03.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/06\/02613396e0642e5b88a928a788f48b03.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\u522b\u5149\u7ed9Agent\u52a0Tool\u4e86\uff0c\u5b83\u6839\u672c\u9009\u4e0d\u660e\u767d\uff01\u590d\u65e6\u00d7\u901a\u4e49\u63d0\u51fa\u5168\u65b0CUA\u8bad\u7ec3\u8303\u5f0f\"\/><\/div><\/figure>\n<figure class=\"wp-block-image\"><div class='fancybox-wrapper lazyload-container-unload' data-fancybox='post-images' href='https:\/\/1ow.top\/wp-content\/uploads\/2026\/06\/300207877a49c7efec1ce32d2e9a6a7b.png'><img class=\"lazyload lazyload-style-1\" src=\"data:image\/svg+xml;base64,PCEtLUFyZ29uTG9hZGluZy0tPgo8c3ZnIHdpZHRoPSIxIiBoZWlnaHQ9IjEiIHhtbG5zPSJodHRwOi8vd3d3LnczLm9yZy8yMDAwL3N2ZyIgc3Ryb2tlPSIjZmZmZmZmMDAiPjxnPjwvZz4KPC9zdmc+\"  decoding=\"async\" data-original=\"https:\/\/1ow.top\/wp-content\/uploads\/2026\/06\/300207877a49c7efec1ce32d2e9a6a7b.png\" src=\"data:image\/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAAAXNSR0IArs4c6QAAAARnQU1BAACxjwv8YQUAAAAJcEhZcwAADsQAAA7EAZUrDhsAAAANSURBVBhXYzh8+PB\/AAffA0nNPuCLAAAAAElFTkSuQmCC\" alt=\"\u522b\u5149\u7ed9Agent\u52a0Tool\u4e86\uff0c\u5b83\u6839\u672c\u9009\u4e0d\u660e\u767d\uff01\u590d\u65e6\u00d7\u901a\u4e49\u63d0\u51fa\u5168\u65b0CUA\u8bad\u7ec3\u8303\u5f0f\"\/><\/div><\/figure>\n<p>\u590d\u65e6\u00d7\u901a\u4e49\u56e2\u961f \u6295\u7a3f<\/p>\n<p>\u7ed9Agent\u540c\u65f6\u63a5\u4e0aGUI\u64cd\u4f5c\u548c\u5de5\u5177\u8c03\u7528\uff0c\u51c6\u786e\u7387\u53cd\u800c\u4e0b\u964d\u4e86\u3002<\/p>\n<p>\u6a21\u578b\u6839\u672c\u4e0d\u4f1a\u5728GUI\u548cTool\u4e4b\u95f4\u9009\u62e9\u3002\u8be5\u70b9\u6309\u94ae\u7684\u65f6\u5019\u53bb\u8c03API\uff0c\u8be5\u8c03API\u7684\u65f6\u5019\u53c8\u6b7b\u78d5\u83dc\u5355\uff0c\u4e24\u5934\u4e71\u7a9c\uff0c\u8d8a\u5e2e\u8d8a\u5fd9\u3002<\/p>\n<p>\u4e3a\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\uff0c\u590d\u65e6\u5927\u5b66\u548c\u901a\u4e49\u5b9e\u9a8c\u5ba4MobileAgent\u56e2\u961f\u8054\u5408\u63d0\u51faToolCUA\uff0c\u4e00\u4e2a\u9762\u5411GUI-Tool\u6df7\u5408\u52a8\u4f5c\u7a7a\u95f4\u7684Computer Use Agent\u3002<\/p>\n<p>\u6838\u5fc3\u76ee\u6807\u5c31\u4e00\u4e2a\uff1a\u8ba9\u6a21\u578b\u5b66\u4f1a\u4ec0\u4e48\u65f6\u5019\u8d70GUI\uff0c\u4ec0\u4e48\u65f6\u5019\u5207Tool\uff0c\u4ec0\u4e48\u65f6\u5019\u4e0d\u8be5\u8c03\u5de5\u5177\u3002<\/p>\n<p>\u7ed3\u679c\u76f8\u5f53\u80fd\u6253\u3002<\/p>\n<p>ToolCUA-8B\u5728OSWorld-MCP\u4e0a\u62ff\u523046.85%\u51c6\u786e\u7387\uff0c\u8d85\u8fc7Claude-4-Sonnet\uff0c\u903c\u8fd1Claude-4.5-Sonnet\u3002<\/p>\n<p>\u4ee3\u7801\u3001\u6a21\u578b\u6743\u91cd\u5df2\u5168\u9762\u5f00\u6e90\u3002<\/p>\n<p>\u4f20\u7edf\u7684CUA\u4e3b\u8981\u4f9d\u8d56\u539f\u5b50\u5316GUI\u64cd\u4f5c\uff0c\u4f8b\u5982\u70b9\u51fb\u3001\u8f93\u5165\u3001\u62d6\u62fd\u3001\u6eda\u52a8\u3002\u8fd9\u7c7b\u64cd\u4f5c\u6cdb\u5316\u6027\u5f3a\uff0c\u53ea\u8981\u754c\u9762\u4e0a\u80fd\u770b\u5230\u6309\u94ae\uff0c\u7406\u8bba\u4e0a\u6a21\u578b\u5c31\u80fd\u70b9\uff1b\u4f46\u5b83\u4e5f\u6709\u660e\u663e\u77ed\u677f\uff1a\u6b65\u9aa4\u957f\u3001\u8bef\u5dee\u5bb9\u6613\u7d2f\u79ef\uff0c\u5728\u590d\u6742\u4efb\u52a1\u4e2d\u5f88\u5bb9\u6613\u51fa\u73b0cascading errors\u3002<\/p>\n<p>\u76f8\u53cd\uff0ctool calls\u6216API-based operations\u5f80\u5f80\u66f4\u9ad8\u6548\u3001\u66f4\u7cbe\u786e\u3002\u4f8b\u5982\u5728LibreOffice\u91cc\u6279\u91cf\u5904\u7406\u8868\u683c\uff0cGUI-only\u65b9\u6848\u53ef\u80fd\u9700\u8981\u4e00\u4e32\u5197\u957f\u7684\u83dc\u5355\u70b9\u51fb\u548c\u53c2\u6570\u914d\u7f6e\uff0c\u800c\u5de5\u5177\u8c03\u7528\u53ef\u80fd\u4e00\u4e2aAPI\u5c31\u80fd\u5b8c\u6210\u3002<\/p>\n<p>\u770b\u8d77\u6765\u6700\u81ea\u7136\u7684\u65b9\u6848\uff0c\u662f\u8ba9Agent\u540c\u65f6\u62e5\u6709GUI\u548cTool\u3002\u4f46\u5b9e\u9a8c\u53d1\u73b0\u4e00\u4e2a\u975e\u5e38\u53cd\u76f4\u89c9\u7684\u4e8b\u5b9e\uff1a\u76f4\u63a5\u628atools\u63a5\u5230\u5f3a\u6a21\u578b\u4e0a\uff0c\u5e76\u4e0d\u4f1a\u81ea\u52a8\u63d0\u5347\u6027\u80fd\u3002<\/p>\n<p>\u5728hybrid GUI-Tool action space\u4e2d\uff0cAgent\u6bcf\u4e00\u6b65\u90fd\u7ad9\u5728\u4e00\u4e2a\u5c94\u8def\u53e3\uff1a\u5de6\u8fb9\u662fGUI\uff0c\u53f3\u8fb9\u662fTool\u3002GUI\u6cdb\u5316\u5f3a\u4f46\u6162\uff0cTool\u5feb\u4f46\u4f9d\u8d56\u8986\u76d6\u4e0e\u4e0a\u4e0b\u6587\u6761\u4ef6\u3002\u5982\u679c\u6a21\u578b\u7f3a\u5c11\u8def\u5f84\u9009\u62e9\u80fd\u529b\uff0c\u5c31\u4f1a\u51fa\u73b0\u4e24\u7c7b\u5178\u578b\u5931\u8d25:<\/p>\n<p>Tool underuse\uff1a\u660e\u660e\u6709\u66f4\u9ad8\u6548\u7684\u5de5\u5177\uff0c\u6a21\u578b\u4ecd\u7136\u51e0\u4e4e\u53ea\u8d70GUI\u8def\u7ebf\u3002<\/p>\n<p>Tool overuse\uff1a\u6a21\u578b\u9891\u7e41\u8c03\u7528\u5de5\u5177\uff0c\u4f46\u8c03\u7528\u65f6\u673a\u4e0d\u5bf9\u3001\u8c03\u7528\u7c92\u5ea6\u4e0d\u5bf9\uff0c\u53cd\u800c\u964d\u4f4e\u4efb\u52a1\u6210\u529f\u7387\u3002<\/p>\n<p>\u8bba\u6587\u5c06\u8fd9\u4e2a\u95ee\u9898\u5b9a\u4e49\u4e3aoptimal GUI-Tool path selection\uff1a\u5728\u957f\u7a0b\u4efb\u52a1\u4e2d\u52a8\u6001\u51b3\u5b9a\u4f55\u65f6\u4f7f\u7528GUI actions\u3001\u4f55\u65f6\u8c03\u7528tools\uff0c\u4ece\u800c\u5f62\u6210\u66f4\u9ad8\u6548\u3001\u66f4\u53ef\u9760\u7684\u6267\u884c\u8def\u5f84\u3002<\/p>\n<p>\u4e0a\u56fe\u5de6\u4fa7\u7684\u8868\u683c\u76f4\u63a5\u7ed9\u51fa\u4e86\u8fd9\u4e2a\u53cd\u76f4\u89c9\u73b0\u8c61\uff1a\u4e00\u65e6\u628atools\u63a5\u5230\u5f3a\u6a21\u578b\u4e0a\uff0c\u7ed3\u679c\u5e76\u4e0d\u603b\u662f\u66f4\u597d\u3002<\/p>\n<p>Qwen3VL-8B\u51e0\u4e4e\u4e0d\u4f7f\u7528\u5de5\u5177\uff0c\u5e73\u5747tool calls\u53ea\u67090.003\uff0c\u51c6\u786e\u7387\u4ece29.0%\u964d\u523028.2%\uff1bQwen3VL-235B\u5219\u660e\u663e\u66f4\u503e\u5411\u4e8e\u8c03\u7528\u5de5\u5177\uff0c\u5e73\u5747tool calls\u8fbe\u52306.10\uff0c\u6b65\u9aa4\u6570\u4ece25.9\u964d\u523017.4\uff0c\u4f46\u51c6\u786e\u7387\u53cd\u800c\u4ece41.1%\u964d\u523038.1%\u3002<\/p>\n<p>Claude\u7cfb\u5217\u540c\u6837\u8bf4\u660e\u4e86\u8fd9\u4e00\u70b9\u3002<\/p>\n<p>Claude-4-sonnet\u5728\u52a0\u5165\u5de5\u5177\u540e\u6b65\u9aa4\u6570\u4ece23.6\u964d\u523019.2\uff0c\u4f46\u51c6\u786e\u7387\u4ece47.7%\u964d\u523043.5%\uff1bClaude-4.5-sonnet\u7684\u6b65\u9aa4\u6570\u4ece23.3\u964d\u523019.1\uff0c\u4f46\u51c6\u786e\u7387\u4ece61.9%\u964d\u523048.4%\u3002<\/p>\n<p>\u8fd9\u8bf4\u660e\uff0c\u6df7\u5408\u52a8\u4f5c\u7a7a\u95f4\u771f\u6b63\u96be\u7684\u4e0d\u662f\u6709\u6ca1\u6709\u5de5\u5177\uff0c\u800c\u662f\u6a21\u578b\u5728GUI\u548cTool\u4e4b\u95f4\u4f1a\u4e0d\u4f1a\u9009\u8def\u3002<\/p>\n<p>\u8981\u8ba9\u6a21\u578b\u5b66\u4f1aGUI-Tool path selection\uff0c\u9996\u5148\u9700\u8981\u9ad8\u8d28\u91cf\u7684interleaved GUI-Tool trajectories\u3002\u4f46\u73b0\u5b9e\u4e2d\uff0c\u8fd9\u7c7b\u6570\u636e\u975e\u5e38\u7a00\u7f3a\u3002<\/p>\n<p>\u771f\u5b9e\u5de5\u5177\u63a5\u53e3\u5f80\u5f80\u5e94\u7528\u76f8\u5173\u3001\u8986\u76d6\u4e0d\u5b8c\u6574\uff0c\u800c\u4e14\u7ef4\u62a4\u6210\u672c\u9ad8\uff1b\u800c\u6536\u96c6\u771f\u5b9eGUI-Tool\u6df7\u5408\u8f68\u8ff9\u53c8\u9700\u8981\u590d\u6742\u7684\u73af\u5883\u63a5\u5165\u548c\u4eba\u5de5\u6807\u6ce8\u3002<\/p>\n<p>\u5df2\u6709GUI\u6570\u636e\u867d\u7136\u89c4\u6a21\u5f88\u5927\uff0c\u4f46\u5927\u591a\u662fGUI-only trajectories\uff0c\u53ea\u6559\u6a21\u578b\u5982\u4f55\u70b9\u51fb\u548c\u8f93\u5165\uff0c\u5e76\u6ca1\u6709\u544a\u8bc9\u6a21\u578b\u4f55\u65f6\u5e94\u8be5\u7528\u5de5\u5177\u66ff\u4ee3\u5197\u957fGUI\u64cd\u4f5c\u3002<\/p>\n<p>ToolCUA\u7684\u7b2c\u4e00\u6b65\uff0c\u5c31\u662f\u628a\u8fd9\u4e9bGUI-only\u6570\u636e\u76d8\u6d3b\uff0c\u5e76\u987a\u52bf\u5b8c\u6210\u7b2c\u4e00\u9636\u6bb5\u7684hybrid bootstrapping\u3002<\/p>\n<p>\u8bba\u6587\u63d0\u51faInterleaved GUI-Tool Trajectory Scaling Pipeline\uff1a\u4ece\u5df2\u6709GUI\u8f68\u8ff9\u51fa\u53d1\uff0c\u5229\u7528MLLM\u5408\u6210grounded tool library\uff0c\u518d\u5c06GUI-only trajectories\u8f6c\u6362\u6210interleaved GUI-Tool trajectories\u3002<\/p>\n<p>\u6574\u4e2apipeline\u53ef\u4ee5\u6982\u62ec\u4e3a\u4e09\u4e2a\u6b65\u9aa4:<\/p>\n<p>1\u3001Trajectory-aware synthetic tool library construction\u3002<\/p>\n<p>\u5bf9\u6bcf\u6761GUI\u8f68\u8ff9\uff0c\u6a21\u578b\u4f1a\u5206\u6790\u4efb\u52a1\u76ee\u6807\u3001\u52a8\u4f5c\u5e8f\u5217\u548c\u622a\u56fe\u63cf\u8ff0\uff0c\u4ece\u771f\u5b9e\u64cd\u4f5c\u6d41\u7a0b\u4e2d\u62bd\u8c61\u51fa\u53ef\u8c03\u7528\u7684\u5de5\u5177\u3002<\/p>\n<p>\u4f8b\u5982\u4eceChrome\u8bbe\u7f6e\u6d41\u7a0b\u4e2d\u62bd\u8c61\u51fachrome_open_language_settings\uff0c\u4eceLibreOffice\u8868\u683c\u64cd\u4f5c\u4e2d\u62bd\u8c61\u51fa\u8bfb\u53d6\u5de5\u4f5c\u7c3f\u4fe1\u606f\u3001\u521b\u5efa\u900f\u89c6\u8868\u7b49\u5de5\u5177\u3002<\/p>\n<p>\u8fd9\u4e9b\u5de5\u5177\u4e0d\u662f\u51ed\u7a7a\u751f\u6210\u7684API\u6a21\u677f\uff0c\u800c\u662fgrounded in concrete trajectory behavior\uff0c\u4e5f\u5c31\u662f\u4ece\u771f\u5b9eGUI\u884c\u4e3a\u4e2d\u62bd\u8c61\u51fa\u6765\u7684\u5de5\u5177\u80fd\u529b\u3002<\/p>\n<p>2\u3001Tool trajectory generation with next-state grounding\u3002<\/p>\n<p>\u7ed9\u5b9a\u5408\u6210\u5de5\u5177\u5e93\u548c\u539f\u59cbGUI\u8f68\u8ff9\uff0cMLLM\u751f\u6210\u4e00\u4e2a\u529f\u80fd\u7b49\u4ef7\u7684tool-only trajectory\uff0c\u5e76\u4e3a\u6bcf\u4e00\u6b65\u9884\u6d4btool response\u3002<\/p>\n<p>\u968f\u540e\u901a\u8fc7next-state grounding\uff0c\u5c06\u5de5\u5177\u6267\u884c\u6548\u679c\u951a\u5b9a\u5230\u539f\u59cbGUI\u8f68\u8ff9\u4e2d\u7684\u4e0b\u4e00\u5e27\u622a\u56fe\uff0c\u9a8c\u8bc1\u5de5\u5177\u6b65\u9aa4\u548c\u53ef\u89c1\u72b6\u6001\u53d8\u5316\u662f\u5426\u4e00\u81f4\u3002<\/p>\n<p>3\u3001Interleaved GUI-Tool trajectory generation\u3002<\/p>\n<p>\u6700\u540e\uff0c\u7cfb\u7edf\u4e0d\u4f1a\u7b80\u5355\u5730\u628a\u6240\u6709GUI\u64cd\u4f5c\u90fd\u66ff\u6362\u6210\u5de5\u5177\uff0c\u800c\u662f\u968f\u673a\u91c7\u6837\u90e8\u5206\u5de5\u5177\u8c03\u7528\uff0c\u518d\u66ff\u6362\u56de\u5bf9\u5e94GUI\u5b50\u5e8f\u5217\uff0c\u5f62\u6210\u591a\u79cdGUI\u4e0eTool\u4ea4\u9519\u7684\u8f68\u8ff9\u3002<\/p>\n<p>\u8fd9\u4e2a\u8bbe\u8ba1\u975e\u5e38\u5173\u952e\uff1a\u5b83\u8ba9\u6a21\u578b\u770b\u5230\u4e0d\u540ctool availability\u4e0b\u7684\u51b3\u7b56\u8fb9\u754c\uff0c\u4e5f\u81ea\u7136\u4ea7\u751fGUI -> Tool\u548cTool -> GUI\u7684critical switching steps\u3002<\/p>\n<p>\u6700\u7ec8\uff0cToolCUA\u7684\u6570\u636e\u4e2d\u5927\u7ea6\u5305\u62ec\u4e864k\u4e2aunique tools\uff0c\u8986\u76d6fine-grained\u3001mid-grained\u3001coarse-grained\u591a\u7ea7\u7c92\u5ea6\uff0c\u5927\u7ea6\u6709180k steps\u6570\u636e\u7528\u4e8ewarmup SFT\uff0c\u8fd8\u4ececritical steps\u4e2dsample\u51fa5k\u6761\u7528\u4e8esingle-turn RL\u3002<\/p>\n<p>\u57fa\u4e8e\u8fd9\u4e9b\u6570\u636e\uff0cToolCUA\u8fdb\u4e00\u6b65\u6267\u884cTool-Bootstrapped GUI RFT\u3002\u8fd9\u4e00\u9636\u6bb5\u7684\u76ee\u6807\uff0c\u4e0d\u662f\u76f4\u63a5\u5b66\u5b8c\u6574\u957f\u7a0b\u7b56\u7565\uff0c\u800c\u662f\u5148\u7ed9\u6a21\u578b\u6253\u4e0b\u4e00\u4e2a\u53ef\u7528\u7684hybrid foundation\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\uff0cToolCUA\u5148\u5728D_all\u4e0a\u8fdb\u884cwarmup SFT\uff0c\u5b66\u4e60\u591a\u6a21\u6001\u5de5\u5177\u8c03\u7528\u77e5\u8bc6\uff0c\u5305\u62ec\u5de5\u5177\u7528\u9014\u3001\u53c2\u6570\u3001\u8fd4\u56de\u7ed3\u679c\uff0c\u4ee5\u53ca\u5de5\u5177\u6267\u884c\u540e\u7684\u72b6\u6001\u53d8\u5316\u3002<\/p>\n<p>\u968f\u540e\uff0c\u6a21\u578b\u5728D_critical\u4e0a\u8fdb\u884csingle-turn RL\uff0c\u5728\u660e\u786e\u7684GUI-Tool switching steps\u4e0a\u91c7\u6837\u591a\u4e2acompletion\uff0c\u5e76\u901a\u8fc7\u53cd\u9988\u6821\u51c6\u6a21\u578b\u5728\u5c40\u90e8\u8fb9\u754c\u4e0a\u7684\u9009\u62e9\u3002<\/p>\n<p>\u8fd9\u4e00\u9636\u6bb5\u505a\u7684\u4e8b\u60c5\u662f\uff1a\u5148\u628ainterleaved GUI-Tool\u6570\u636e\u5408\u6210\u51fa\u6765\uff0c\u518d\u8ba9\u6a21\u578b\u5148\u5b66\u4f1a\u4f1a\u7528\u5de5\u5177\u548c\u5728\u5c40\u90e8\u5207\u6362\u70b9\u4e0a\u522b\u9009\u9519\u3002<\/p>\n<p>\u5982\u679c\u8bf4\u7b2c\u4e00\u9636\u6bb5\u89e3\u51b3\u7684\u662f\u6a21\u578b\u5148\u8981\u8fdb\u5165hybrid action space\uff0c\u90a3\u4e48\u7b2c\u4e8c\u9636\u6bb5\u89e3\u51b3\u7684\u5c31\u662f\uff1a\u6a21\u578b\u5982\u4f55\u5728\u771f\u5b9e\u73af\u5883\u91cc\u5b66\u4f1atrajectory-level\u7684\u8def\u5f84\u9009\u62e9\u3002<\/p>\n<p>ToolCUA\u7684\u7b2c\u4e8c\u9636\u6bb5\u662fOnline Agentic RL\u3002\u8fd9\u4e00\u6b65\u4e0d\u518d\u53ea\u4f18\u5316\u5355\u6b65\u52a8\u4f5c\uff0c\u800c\u662f\u5728\u771f\u5b9eGUI-Tool environment\u4e2d\u8fdb\u884clong-horizon rollout\uff0c\u8ba9\u6a21\u578b\u5b66\u4e60\u5b8c\u6574\u4efb\u52a1\u8f68\u8ff9\u4e0a\u7684\u8def\u5f84\u9009\u62e9\u3002<\/p>\n<p>\u56e2\u961f\u9996\u5148\u6784\u5efa\u4e86\u540c\u65f6\u5177\u5907GUI actions\u548cTool calls\u7684\u9ad8\u53ef\u7528Sandbox\u7528\u4e8eagentic RL\uff0c\u5e76\u4e14\u4e3a\u5de5\u5177\u8fd4\u56de\u7ed3\u679c\u8bbe\u8ba1\u4e86\u66f4\u52a0\u7ed3\u6784\u5316\u7684\u683c\u5f0f\u4fbf\u4e8e\u6a21\u578b\u7406\u89e3\u3002<\/p>\n<p>Agentic RL\u4f18\u5316\u7684\u6838\u5fc3\u662fTool-Efficient Path Reward:<\/p>\n<p>\u5176\u4e2d\uff0cR_fmt\u548cR_acc\u5206\u522b\u662f\u6807\u51c6\u683c\u5f0f\u5956\u52b1\u4e0e\u4efb\u52a1\u6210\u529f\u5956\u52b1\uff1bR_tool\u548cR_length\u5219\u662fToolCUA\u4e13\u95e8\u8bbe\u8ba1\u7684\u4e24\u9879\u8f68\u8ff9\u7ea7\u5956\u52b1\uff0c\u5e76\u4e14\u5b83\u4eec\u53ea\u5728\u6210\u529f\u8f68\u8ff9\u4e0a\u6fc0\u6d3b\uff0c\u907f\u514d\u6a21\u578b\u4ece\u5931\u8d25\u6267\u884c\u91cc\u5b66\u5230\u9519\u8bef\u504f\u597d\u3002<\/p>\n<p>\u7b2c\u4e00\u9879\u662fTool Appropriateness Reward (R_tool)\u3002<\/p>\n<p>\u5728\u6570\u636e\u6784\u5efa\u65f6\uff0c\u6bcf\u4e2a\u4efb\u52a1\u4f1a\u5e26\u4e00\u4e2atask-level\u7684tool-beneficial\u6807\u8bb0\uff1at_b = 1\u8868\u793a\u8fd9\u4e2a\u4efb\u52a1\u9002\u5408\u7528\u5de5\u5177\uff0ct_b = -1\u8868\u793a\u8fd9\u4e2a\u4efb\u52a1\u4e0d\u9002\u5408\u7528\u5de5\u5177\u3002\u4e0e\u6b64\u540c\u65f6\uff0cc\u8868\u793a\u6574\u6761\u8f68\u8ff9\u91cc\u7684tool calls\u6570\u3002<\/p>\n<p>\u4e8e\u662f\uff0cR_tool\u5956\u52b1\u7684\u4e0d\u662f\u5de5\u5177\u8c03\u7528\u66f4\u591a\uff0c\u800c\u662f\u66f4\u7cbe\u786e\u7684\u4e24\u79cd\u884c\u4e3a:<\/p>\n<p>\u5bf9\u4e8e\u9002\u5408\u5de5\u5177\u7684\u4efb\u52a1\uff0c\u6210\u529f\u8f68\u8ff9\u91cc\u786e\u5b9e\u8c03\u7528\u4e86\u5de5\u5177\u3002<\/p>\n<p>\u5bf9\u4e8e\u4e0d\u9002\u5408\u5de5\u5177\u7684\u4efb\u52a1\uff0c\u6210\u529f\u8f68\u8ff9\u91cc\u53cd\u800c\u6ca1\u6709\u4e71\u7528\u5de5\u5177\u3002<\/p>\n<p>\u5b83\u8981\u89e3\u51b3\u7684\u6b63\u662f\u524d\u9762\u63d0\u5230\u7684hybrid confusion\uff1a\u6709\u4e9b\u6a21\u578b\u660e\u660e\u8be5\u7528\u5de5\u5177\u5374\u4e0d\u7528\uff0c\u6709\u4e9b\u6a21\u578b\u5219\u5728\u4e0d\u8be5\u7528\u7684\u65f6\u5019\u4e71\u7528\u3002R_tool\u7684\u4f5c\u7528\uff0c\u5c31\u662f\u628a\u5de5\u5177\u662f\u5426\u5408\u9002\u8fd9\u4ef6\u4e8b\u4ece\u4efb\u52a1\u6210\u529f\u91cc\u5355\u72ec\u62ce\u51fa\u6765\u8bad\u7ec3\u3002<\/p>\n<p>\u7b2c\u4e8c\u9879\u662fPath Efficiency Reward (R_length)\u3002<\/p>\n<p>\u8fd9\u91cc\uff0cs\u662f\u5f53\u524d\u8f68\u8ff9\u7684\u6b65\u6570\uff0c\\bar{s}\u662f\u540c\u7ec4rollout\u7684\u5e73\u5747\u6b65\u957f\uff0cS_max\u662f\u6700\u5927\u6267\u884c\u6b65\u6570\u3002ToolCUA\u4e0d\u62ff\u4e00\u4e2a\u56fa\u5b9a\u9608\u503c\u6765\u5224\u5b9a\u957f\u8fd8\u662f\u77ed\uff0c\u800c\u662f\u505agroup-relative comparison:<\/p>\n<p>\u5982\u679c\u67d0\u6761\u6210\u529f\u8f68\u8ff9\u6bd4\u7ec4\u5185\u5e73\u5747\u66f4\u77ed\uff0c\u5c31\u7ed9\u7ebf\u6027bonus\u3002<\/p>\n<p>\u5982\u679c\u66f4\u957f\uff0c\u5c31\u505a\u8870\u51cf\u3002<\/p>\n<p>\u8fd9\u6837\u8bbe\u8ba1\u7684\u597d\u5904\u662f\uff0c\u6a21\u578b\u4f1a\u81ea\u7136\u503e\u5411\u4e8e\u63a2\u7d22\u66f4\u77ed\u7684\u6210\u529f\u8def\u5f84\u3002\u800c\u5728\u5f88\u591a\u573a\u666f\u91cc\uff0c\u66f4\u77ed\u7684\u8def\u5f84\u6070\u6070\u610f\u5473\u7740\uff1a\u7528\u4e00\u4e2a\u9ad8\u5c42\u5de5\u5177\u66ff\u4ee3\u4e00\u957f\u4e32\u5197\u4f59GUI\u64cd\u4f5c\u3002\u56e0\u6b64\uff0cR_length\u672c\u8d28\u4e0a\u662f\u5728\u9f13\u52b1\u6a21\u578b\u53d1\u73b0\u66f4\u9ad8\u6548\u7684GUI-Tool execution path\u3002<\/p>\n<p>\u6240\u4ee5\uff0c\u8fd9\u4e00\u9636\u6bb5\u7684\u6838\u5fc3\u5e76\u4e0d\u662f\u8ba9\u6a21\u578b\u8c03\u7528\u66f4\u591a\u5de5\u5177\uff0c\u800c\u662f\u8ba9\u5b83\u5b66\u4f1a\u4e24\u4ef6\u4e8b\uff1a\u4ec0\u4e48\u65f6\u5019\u5de5\u5177\u771f\u7684\u5408\u9002\uff0c\u4ec0\u4e48\u65f6\u5019\u8fd9\u6761\u6267\u884c\u8def\u5f84\u771f\u7684\u66f4\u77ed\u3002<\/p>\n<p>ToolCUA\u4e3b\u8981\u5728OSWorld-MCP\u4e0a\u8bc4\u6d4b\u3002\u8fd9\u4e2abenchmark\u5728\u4f20\u7edfOSWorld\u7684\u57fa\u7840\u4e0a\u5f15\u5165\u4e86hybrid GUI-Tool action space\uff0c\u8986\u76d6\u5178\u578bGUI actions\u3001150+ tools\u548c\u4e3b\u6d41\u684c\u9762\u5e94\u7528\uff0c\u9002\u5408\u8861\u91cf\u6a21\u578b\u5728\u771f\u5b9e\u6df7\u5408\u52a8\u4f5c\u7a7a\u95f4\u4e2d\u7684\u6267\u884c\u80fd\u529b\u3002<\/p>\n<p>\u8bc4\u6d4b\u6307\u6807\u5305\u62ec:<\/p>\n<p>ToolCUA-8B\u5728OSWorld-MCP\u4e0a\u53d6\u5f9746.85%accuracy\uff0c\u76f8\u6bd4Qwen3-VL-8B-Instruct baseline\u768428.23%\uff0c\u76f8\u5bf9\u63d0\u5347\u7ea666%\u3002<\/p>\n<p>\u540c\u65f6\uff0cToolCUA\u8d85\u8fc7\u4e86GUI-Owl-1.5-8B(43.84%)\u3001Gemini-3.1-Pro(41.14%)\u548cClaude-4-Sonnet(43.54%)\uff0c\u5e76\u63a5\u8fd1Claude-4.5-Sonnet(48.35%)\u4e0eGUI-Owl-1.5-32B(48.05%)\u3002<\/p>\n<p>\u66f4\u91cd\u8981\u7684\u662f\u6548\u7387\u6307\u6807\u3002ToolCUA\u7684ACS\u4ec5\u4e3a14.93 steps\uff0c\u662f\u8868\u4e2d\u6240\u6709\u6a21\u578b\u91cc\u6700\u4f4e\u7684\u3002\u8fd9\u8bf4\u660eToolCUA\u4e0d\u53ea\u662f\u5b8c\u6210\u4e86\u66f4\u591a\u4efb\u52a1\uff0c\u4e5f\u5b66\u4f1a\u4e86\u7528\u66f4\u77ed\u8def\u5f84\u5b8c\u6210\u4efb\u52a1\u3002<\/p>\n<p>\u4e0eQwen3-VL-8B-Instruct\u76f8\u6bd4\uff0cToolCUA\u7684overall TIR\u4ece8.41%\u63d0\u5347\u523024.32%\uff0cACS\u4ece19.34\u964d\u523014.93\u3002\u8fd9\u8bf4\u660e\u6a21\u578b\u4e0d\u4ec5\u66f4\u4f1a\u505a\u4efb\u52a1\uff0c\u4e5f\u66f4\u4f1a\u5224\u65ad\u4ec0\u4e48\u65f6\u5019\u5e94\u8be5\u8c03\u7528\u5de5\u5177\u3002<\/p>\n<p>\u5728\u8bad\u7ec3\u9636\u6bb5\uff0cOnline Agentic RL\u53ea\u4f7f\u7528\u5355\u5e94\u7528Linux\u4efb\u52a1\uff0c\u5e76\u523b\u610f\u6392\u9664\u4e86multi_apps domain\uff0c\u7528\u4e8eOOD\u9a8c\u8bc1\u3002<\/p>\n<p>\u7ed3\u679c\u663e\u793a\uff0c\u5728held-out multi_apps\u4efb\u52a1\u4e0a\uff0cToolCUA\u4ecebaseline\u76849.8%\u548cpre-online RL stage\u768418.5%\u63d0\u5347\u523023.9%\u3002<\/p>\n<p>\u5728\u5177\u4f53\u5e94\u7528\u57df\u4e0a\uff0cToolCUA\u4e5f\u6709\u660e\u663e\u63d0\u5347\u3002\u4f8b\u5982\u5728libreoffice_calculation\u4e0a\u4ece19.6%\u63d0\u5347\u523034.8%\uff0c\u5728vs_code\u4e0a\u4ece66.7%\u63d0\u5347\u523094.4%\u3002<\/p>\n<p>\u66f4\u8fdb\u4e00\u6b65\uff0cToolCUA\u8fd8\u5728WindowsAgentArena\u4e0a\u8fdb\u884c\u8bc4\u6d4b\u3002<\/p>\n<p>\u5c3d\u7ba1\u8bad\u7ec3\u6570\u636e\u548csandbox\u90fd\u6765\u81eaLinux\u684c\u9762\u73af\u5883\uff0cToolCUA\u5728unseen Windows desktop apps\u4e0a\u8fbe\u523033.8%accuracy\uff0c\u8d85\u8fc7Qwen3-VL-8B-Instruct\u768426.4%\u3001Qwen3-VL-32B-Instruct\u768430.9%\uff0c\u4e5f\u8d85\u8fc7Qwen3-VL-235B-A22B\u768432.1%\u3002<\/p>\n<p>\u8fd9\u8bf4\u660eToolCUA\u5b66\u5230\u7684\u5e76\u4e0d\u53ea\u662f\u67d0\u4e9b\u7279\u5b9a\u4efb\u52a1\u6a21\u677f\uff0c\u800c\u662f\u66f4\u63a5\u8fd1\u4e00\u79cd\u53ef\u8fc1\u79fb\u7684hybrid action orchestration\u80fd\u529b\u3002<\/p>\n<p>ToolCUA\u7684\u63d0\u5347\u5230\u5e95\u6765\u81ea\u54ea\u91cc\uff1f\u8bba\u6587\u91cc\u7684ablation\u5f88\u6e05\u695a\u5730\u7ed9\u51fa\u4e09\u6761\u7ed3\u8bba\u3002<\/p>\n<p>\u7b2c\u4e00\uff0c\u5982\u679c\u6ca1\u6709interleaved GUI-Tool trajectory data\uff0conline RL\u672c\u8eab\u5b66\u4e0d\u4f1a\u53ef\u9760\u7684tool use\u3002<\/p>\n<p>\u5f53\u53bb\u6389offline interleaved GUI-Tool bootstrapping\uff0c\u76f4\u63a5\u4eceQwen3-VL-8B-Instruct baseline\u5f00\u59cb\u505aonline agentic RL\u65f6\uff0c\u6a21\u578b\u7684overall accuracy\u867d\u7136\u4e5f\u4f1a\u7ee7\u7eed\u4e0a\u5347\uff0c\u4f46\u5b83\u5f88\u96be\u771f\u6b63\u5b66\u4f1a\u7a33\u5b9a\u7684\u5de5\u5177\u8c03\u7528\u884c\u4e3a\u3002<\/p>\n<p>\u6700\u5178\u578b\u7684\u73b0\u8c61\u662f\uff1aTIR\u957f\u671f\u504f\u4f4e\uff0c\u8bad\u7ec3\u540e\u671f\u4e5f\u53ea\u5230\u7ea615%\uff1btool calls\u5728\u5927\u90e8\u5206\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u90fd\u63a5\u8fd10\u3002<\/p>\n<p>\u8fd9\u8bf4\u660e\uff0c\u4ec5\u9760trajectory-level online reward\uff0c\u5e76\u4e0d\u8db3\u4ee5\u8ba9\u4e00\u4e2aGUI-centric base model\u81ea\u7136\u957f\u51fa\u9760\u8c31\u7684hybrid switching\u80fd\u529b\u3002\u6a21\u578b\u9700\u8981\u5148\u901a\u8fc7interleaved supervision\u83b7\u5f97\u5de5\u5177\u77e5\u8bc6\u548c\u5207\u6362\u5148\u9a8c\u3002<\/p>\n<p>\u7b2c\u4e8c\uff0c\u5982\u679c\u6ca1\u6709Tool-Efficient Path Reward\uff0c\u6a21\u578b\u5b66\u4e0d\u4f1a\u7a33\u5b9a\u4e14\u9ad8\u6548\u7684\u8def\u5f84\u3002<\/p>\n<p>\u540c\u6837\u5728rl_dynamics\u91cc\u53ef\u4ee5\u770b\u5230\uff0c\u53bb\u6389R_tool\u548cR_length\u540e\uff0c\u53ea\u4fdd\u7559\u6807\u51c6\u7684R_acc\u4e0eR_fmt\uff0caccuracy\u66f2\u7ebf\u4f1a\u660e\u663e\u66f4\u4e0d\u7a33\u5b9a\uff0c\u5728\u8bad\u7ec3step8-11\u5de6\u53f3\u51fa\u73b0\u4e0b\u964d\uff0c\u6700\u7ec8\u4e0e\u5b8c\u6574ToolCUA\u4e4b\u95f4\u6709\u5927\u7ea67\u4e2a\u70b9\u7684\u5dee\u8ddd\u3002<\/p>\n<p>\u4e0e\u6b64\u540c\u65f6\uff0cTIR\u548ctool-calls\u4e5f\u6ca1\u6709\u7a33\u5b9a\u4e0a\u5347\u8d8b\u52bf\uff0ctrajectory length\u4e5f\u7f3a\u5c11\u6301\u7eed\u4e0b\u964d\u3002<\/p>\n<p>\u8fd9\u8bf4\u660e\uff0c\u4efb\u52a1\u6210\u529f\u5956\u52b1\u672c\u8eab\u4e0d\u8db3\u4ee5\u6559\u4f1a\u6a21\u578b\u4ec0\u4e48\u65f6\u5019\u5de5\u5177\u662f\u5408\u9002\u7684\u548c\u4ec0\u4e48\u8def\u5f84\u624d\u662f\u771f\u6b63\u9ad8\u6548\u7684\u3002<\/p>\n<p>\u7b2c\u4e09\uff0cHybrid GUI-Tool training\u6bd4pure GUI training\u66f4\u6709\u6548\u3002<\/p>\n<p>\u8bba\u6587\u8fdb\u4e00\u6b65\u6bd4\u8f83\u4e86pure GUI training\u548chybrid GUI-Tool training\u3002<\/p>\n<p>GUI-only pipeline\u4ecebaseline29.03%\u63d0\u5347\u5230SFT\u540e34.93%\uff0c\u518d\u5230agentic RL\u540e42.05%\uff1b\u800cGUI+Tool pipeline\u4e2d\uff0cRFT\u5df2\u7ecf\u8fbe\u523038.13%\uff0c\u5b8c\u6574ToolCUA\u8fdb\u4e00\u6b65\u8fbe\u523046.85%\u3002<\/p>\n<p>\u8fd9\u8868\u660ehybrid GUI-Tool action space\u672c\u8eab\u5c31\u662f\u4e00\u4e2a\u66f4\u9ad8\u4fdd\u771f\u7684\u8bad\u7ec3\u73af\u5883\u3002\u6a21\u578b\u4e0d\u53ea\u662f\u5b66visual grounding\uff0c\u4e5f\u5728\u8fd9\u4e2a\u8fc7\u7a0b\u4e2d\u5b66\u4f1a\u4f55\u65f6\u5e94\u8be5\u7528\u7ed3\u6784\u5316\u5de5\u5177\u66ff\u4ee3\u5197\u4f59GUI\u64cd\u4f5c\u3002<\/p>\n<p>WindowsAgentArena\u7684\u7ed3\u679c\u4e5f\u8bf4\u660e\uff0c\u8fd9\u79cd\u8bad\u7ec3\u8303\u5f0f\u5e26\u6765\u7684\u4e0d\u662f\u5355\u70b9\u6536\u76ca\uff0c\u800c\u662f\u66f4\u5f3a\u7684\u8de8\u5e73\u53f0\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n<p>\u4e3a\u4e86\u66f4\u76f4\u89c2\u5730\u7406\u89e3ToolCUA\u7684\u80fd\u529b\uff0c\u53ef\u4ee5\u770b\u4e24\u4e2a\u5b9e\u9645\u6848\u4f8b\u3002<\/p>\n<p>\u7b2c\u4e00\u4e2a\u662fLibreOffice Calc\u4efb\u52a1\uff1a\u7528\u6237\u8981\u6c42\u5728\u4e00\u4e2a\u540d\u4e3aSheet2\u7684\u65b0sheet\u4e2d\u521b\u5efa\u4e24\u4e2apivot tables\uff0c\u5206\u522b\u7edf\u8ba1product\u548csales channel\u5bf9\u5e94\u7684total revenue\u3002<\/p>\n<p>GUI-only\u65b9\u6cd5\u901a\u5e38\u9700\u8981\u9009\u62e9\u6570\u636e\u8303\u56f4\u3001\u6253\u5f00\u83dc\u5355\u3001\u914d\u7f6e\u5b57\u6bb5\u3001\u786e\u8ba4\u53c2\u6570\uff0c\u6b65\u9aa4\u5197\u957f\u4e14\u5bb9\u6613\u51fa\u9519\u3002<\/p>\n<p>ToolCUA\u5219\u5148\u8c03\u7528\u5de5\u5177\u8bfb\u53d6workbook\u4fe1\u606f\u548csheet\u5185\u5bb9\uff0c\u8bc6\u522b\u6570\u636e\u7ed3\u6784\u4e0e\u5b57\u6bb5\u4f4d\u7f6e\uff0c\u7136\u540e\u76f4\u63a5\u8c03\u7528create_pivot_table\u751f\u6210\u900f\u89c6\u8868\u3002<\/p>\n<p>\u8fd9\u4e2a\u6848\u4f8b\u5c55\u793a\u7684\u4e0d\u662f\u5de5\u5177\u6c38\u8fdc\u6bd4GUI\u597d\uff0c\u800c\u662f: \u5f53\u4efb\u52a1\u6838\u5fc3\u662f\u7ed3\u6784\u5316\u8868\u683c\u64cd\u4f5c\u65f6\uff0cTool\u53ef\u4ee5\u7ed5\u8fc7\u8106\u5f31\u7684\u9010\u6b65GUI\u5bfc\u822a\uff0c\u7528\u66f4\u786e\u5b9a\u7684\u65b9\u5f0f\u5b8c\u6210\u4efb\u52a1\u3002<\/p>\n<p>\u7b2c\u4e8c\u4e2a\u6848\u4f8b\u6765\u81eaVS Code\u3002\u4efb\u52a1\u662f\u5c06\/home\/user\/data1\u548c\/home\/user\/data2\u4e24\u4e2a\u6587\u4ef6\u5939\u52a0\u5165\u5f53\u524dworkspace\u3002<\/p>\n<p>ToolCUA\u5148\u8fde\u7eed\u8c03\u7528add_folder\u5de5\u5177\uff0c\u628a\u4e24\u4e2a\u76ee\u5f55\u52a0\u5165VS Code workspace\u3002<\/p>\n<p>\u8fd9\u4e00\u6b65\u975e\u5e38\u9002\u5408\u5de5\u5177\u8c03\u7528\uff0c\u56e0\u4e3a\u8def\u5f84\u660e\u786e\u3001\u64cd\u4f5c\u7ed3\u6784\u5316\u3001\u76ee\u6807\u53ef\u9a8c\u8bc1\u3002<\/p>\n<p>\u4f46\u5de5\u5177\u8c03\u7528\u5b8c\u6210\u540e\uff0cVS Code\u5f39\u51fa\u4e86Do you trust the authors?\u7684\u4fe1\u4efb\u786e\u8ba4\u5bf9\u8bdd\u6846\u3002<\/p>\n<p>\u8fd9\u4e2a\u72b6\u6001\u4e0d\u662f\u7b80\u5355tool call\u5c31\u80fd\u95ed\u73af\u7684\u3002<\/p>\n<p>\u6b64\u65f6ToolCUA\u5207\u6362\u56deGUI action\uff0c\u70b9\u51fbYes, I trust the authors\u3002<\/p>\n<p>\u5b8c\u6210\u754c\u9762\u4e0a\u7684\u6700\u540e\u4e00\u6b65\u3002<\/p>\n<p>\u8fd9\u6b63\u662fToolCUA\u60f3\u89e3\u51b3\u7684\u95ee\u9898\uff1a\u5b83\u4e0d\u662f\u8bd5\u56fe\u7528Tool\u66ff\u4ee3\u6240\u6709GUI\uff0c\u4e5f\u4e0d\u662f\u9000\u56de\u7eafGUI\u64cd\u4f5c\uff0c\u800c\u662f\u5728\u771f\u5b9e\u73af\u5883\u91cc\u5b66\u4e60\u4e24\u79cdaction space\u7684\u534f\u540c\u4e0e\u5207\u6362\u3002<\/p>\n<p>\u5728agent\u70ed\u6f6e\u7684\u63a8\u52a8\u4e0b\uff0ccomputer use agent\u6b63\u5728\u66f4\u79ef\u6781\u5730\u63a2\u7d22\u771f\u5b9e\u4e16\u754c\u91cc\u7684\u843d\u5730\u8def\u5f84\u3002<\/p>\n<p>ToolCUA\u4e3a\u793e\u533a\u63ed\u793a\u4e86\u4e00\u4e2a\u5173\u952e\u73b0\u8c61\uff1a\u4e00\u65e6\u8fdb\u5165hybrid action space\uff0c\u73b0\u6709CUA\u548c\u90e8\u5206\u5f3a\u57fa\u5ea7\u6a21\u578b\u4f1a\u51fa\u73b0\u660e\u663e\u7684\u8def\u5f84\u56f0\u60d1\uff0c\u751a\u81f3\u5bfc\u81f4\u51c6\u786e\u7387\u4e0b\u964d\u3002<\/p>\n<p>\u56e2\u961f\u901a\u8fc7staged training paradigm\u5728hybrid action training\u4e0a\u505a\u4e86\u4e00\u6b21\u6709\u76ca\u63a2\u7d22\uff0c\u5e76\u9a8c\u8bc1\u4e86\u8fd9\u4e00\u8def\u7ebf\u7684\u6709\u6548\u6027\u3002<\/p>\n<p>\u63a5\u4e0b\u6765\uff0c\u66f4\u503c\u5f97\u7ee7\u7eed\u548c\u63a8\u8fdb\u7684\u65b9\u5411\uff0c\u662f\u6784\u5efa\u66f4\u5927\u89c4\u6a21\u7684CUA\u5de5\u5177\uff0c\u8bad\u7ec3\u66f4\u5927\u89c4\u6a21\u7684CUA\u57fa\u5ea7\u6a21\u578b\uff0c\u8ba9CUA\u539f\u751f\u5177\u6709hybrid actions\u7684\u80fd\u529b\uff0c\u66f4\u597d\u5730\u89e3\u51b3\u4eba\u7c7b\u590d\u6742\u95ee\u9898\u3002<\/p>\n<p>\u9879\u76ee\u7f51\u7ad9\uff1ahttps:\/\/x-plug.github.io\/ToolCUA\/\u4ee3\u7801\u4ed3\u5e93\uff1ahttps:\/\/github.com\/X-PLUG\/ToolCUA\u6a21\u578b\u5730\u5740\uff1ahttps:\/\/huggingface.co\/mPLUG\/ToolCUA-8BMobile-Agent\u7cfb\u5217\uff1ahttps:\/\/github.com\/X-PLUG\/MobileAgent<\/p>\n<blockquote>\n<p><strong>\u7248\u6743\u58f0\u660e\uff1a<\/strong><\/p>\n<p>\u672c\u535a\u5ba2\u90e8\u5206\u5185\u5bb9\u4e3a\u8f6c\u8f7d\u6587\u7ae0\uff0c\u65e8\u5728\u5206\u4eab\u6709\u4ef7\u503c\u7684\u4fe1\u606f\uff0c\u7248\u6743\u5f52\u539f\u4f5c\u8005\u6240\u6709\u3002<\/p>\n<p>\u8f6c\u8f7d\u4ec5\u4e3a\u4e2a\u4eba\u5b66\u4e60\u4e0e\u4ea4\u6d41\u76ee\u7684\uff0c\u4e0d\u5bf9\u6587\u7ae0\u89c2\u70b9\u8d1f\u8d23\uff0c\u4ea6\u4e0d\u7528\u4e8e\u4efb\u4f55\u5546\u4e1a\u7528\u9014\u3002<\/p>\n<p>\u5982\u6d89\u53ca\u7248\u6743\u95ee\u9898\uff0c\u8bf7\u8054\u7cfb\u672c\u4eba\u5220\u9664\u3002<\/p>\n<\/blockquote>\n","protected":false},"excerpt":{"rendered":"<p>\u590d\u65e6\u00d7\u901a\u4e49\u56e2\u961f \u6295\u7a3f<\/p>\n<p>\u7ed9Agent\u540c\u65f6\u63a5\u4e0aGUI\u64cd\u4f5c\u548c\u5de5\u5177\u8c03\u7528\uff0c\u51c6\u786e\u7387\u53cd\u800c\u4e0b\u964d\u4e86\u3002<\/p>\n<p>\u6a21\u578b\u6839\u672c\u4e0d\u4f1a\u5728GUI\u548cTool\u4e4b\u95f4\u9009\u62e9\u3002\u8be5\u70b9\u6309\u94ae\u7684\u65f6\u5019\u53bb\u8c03API\uff0c\u8be5\u8c03API\u7684\u65f6\u5019\u53c8\u6b7b\u78d5\u83dc\u5355\uff0c\u4e24\u5934\u4e71\u7a9c\uff0c\u8d8a\u5e2e\u8d8a\u5fd9\u3002<\/p>\n<p>\u4e3a\u5e94\u5bf9\u8fd9\u4e00\u6311\u6218\uff0c\u590d\u65e6\u5927\u5b66\u548c\u901a\u4e49\u5b9e\u9a8c\u5ba4MobileAgent\u56e2\u961f\u8054\u5408\u63d0\u51faToolCUA\uff0c\u4e00\u4e2a\u9762\u5411GUI-Tool\u6df7\u5408\u52a8\u4f5c\u7a7a\u95f4\u7684Computer Use Agent\u3002<\/p>\n<p>\u6838\u5fc3\u76ee\u6807\u5c31\u4e00\u4e2a\uff1a\u8ba9\u6a21\u578b\u5b66\u4f1a\u4ec0<\/p>\n","protected":false},"author":1,"featured_media":5385,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-5389","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-xwzx"],"_links":{"self":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/5389","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/comments?post=5389"}],"version-history":[{"count":0,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/posts\/5389\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media\/5385"}],"wp:attachment":[{"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/media?parent=5389"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/categories?post=5389"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/1ow.top\/index.php\/wp-json\/wp\/v2\/tags?post=5389"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}