{"id":7738,"date":"2024-10-27T20:06:05","date_gmt":"2024-10-27T12:06:05","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=7738"},"modified":"2024-10-27T20:14:29","modified_gmt":"2024-10-27T12:14:29","slug":"toolgen","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/ja\/toolgen\/","title":{"rendered":"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528"},"content":{"rendered":"<blockquote><p>ToolGen \u662f\u4e00\u4e2a\u5c06\u5de5\u5177\u77e5\u8bc6\u76f4\u63a5\u96c6\u6210\u5230\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u4e2d\u7684\u6846\u67b6\uff0c\u901a\u8fc7\u5c06\u6bcf\u4e2a\u5de5\u5177\u8868\u793a\u4e3a\u72ec\u7279\u7684\u6807\u8bb0\uff0c\u5b9e\u73b0\u65e0\u7f1d\u7684\u5de5\u5177\u8c03\u7528\u548c\u8bed\u8a00\u751f\u6210\u3002\u8be5\u9879\u76ee\u7531 Renxi Wang \u7b49\u4eba\u5f00\u53d1\uff0c\u65e8\u5728\u63d0\u5347\u5de5\u5177\u68c0\u7d22\u548c\u4efb\u52a1\u5b8c\u6210\u7684\u6027\u80fd\u3002<\/p>\n<ul>\n<li><span style=\"color: #333333;\">\u5de5\u5177\u6807\u8bb0\u5316\uff1a\u5c06\u5de5\u5177\u8f6c\u6362\u4e3a\u72ec\u7279\u7684\u6807\u8bb0\uff0c\u4fbf\u4e8e\u6a21\u578b\u8c03\u7528\u3002<\/span><\/li>\n<li><span style=\"color: #333333;\">\u5de5\u5177\u8c03\u7528\u751f\u6210\uff1a\u6a21\u578b\u80fd\u591f\u751f\u6210\u5de5\u5177\u8c03\u7528\u548c\u53c2\u6570\u3002<\/span><\/li>\n<li><span style=\"color: #333333;\">\u4efb\u52a1\u5b8c\u6210\uff1a\u901a\u8fc7\u5de5\u5177\u8c03\u7528\u5b9e\u73b0\u590d\u6742\u4efb\u52a1\u7684\u81ea\u52a8\u5316\u3002<\/span><\/li>\n<li><span style=\"color: #333333;\">\u6570\u636e\u96c6\u652f\u6301\uff1a\u63d0\u4f9b\u4e30\u5bcc\u7684\u6570\u636e\u96c6\u4ee5\u652f\u6301\u6a21\u578b\u8bad\u7ec3\u548c\u8bc4\u4f30\u3002<\/span><\/li>\n<\/ul>\n<\/blockquote>\n<p>&nbsp;<\/p>\n<h6>\u6458\u8981<\/h6>\n<p>\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u53d1\u5c55\uff0c\u5176\u65e0\u6cd5\u901a\u8fc7\u76f4\u63a5\u4e0e\u5916\u90e8\u5de5\u5177\u4ea4\u4e92\u6765\u81ea\u4e3b\u6267\u884c\u4efb\u52a1\u7684\u5c40\u9650\u6027\u53d8\u5f97\u5c24\u4e3a\u660e\u663e\u3002\u4f20\u7edf\u65b9\u6cd5\u4f9d\u8d56\u5c06\u5de5\u5177\u63cf\u8ff0\u4f5c\u4e3a\u4e0a\u4e0b\u6587\u8f93\u5165\uff0c\u8fd9\u53d7\u9650\u4e8e\u4e0a\u4e0b\u6587\u957f\u5ea6\uff0c\u4e14\u9700\u8981\u5355\u72ec\u7684\u68c0\u7d22\u673a\u5236\uff0c\u5f80\u5f80\u6548\u7387\u4f4e\u4e0b\u3002\u6211\u4eec\u63d0\u51fa\u4e86ToolGen\uff0c\u4e00\u79cd\u901a\u8fc7\u5c06\u6bcf\u4e2a\u5de5\u5177\u8868\u793a\u4e3a\u552f\u4e00\u7684 <a href=\"https:\/\/www.kdjingpai.com\/en\/tokenization\/\">Token<\/a> \uff0c\u76f4\u63a5\u5c06\u5de5\u5177\u77e5\u8bc6\u96c6\u6210\u5230LLM\u53c2\u6570\u4e2d\u7684\u8303\u5f0f\u3002\u8fd9\u4f7f\u5f97LLM\u80fd\u591f\u5c06\u5de5\u5177\u8c03\u7528\u548c\u53c2\u6570\u4f5c\u4e3a\u5176\u4e0b\u4e00\u4e2a Token \u7684\u9884\u6d4b\u80fd\u529b\u7684\u4e00\u90e8\u5206\uff0c\u4ece\u800c\u5c06\u5de5\u5177\u8c03\u7528\u4e0e\u8bed\u8a00\u751f\u6210\u65e0\u7f1d\u7ed3\u5408\u3002\u6211\u4eec\u7684\u6846\u67b6\u5141\u8bb8LLM\u5728\u65e0\u9700\u989d\u5916\u68c0\u7d22\u6b65\u9aa4\u7684\u60c5\u51b5\u4e0b\u8bbf\u95ee\u548c\u4f7f\u7528\u5927\u91cf\u5de5\u5177\uff0c\u663e\u8457\u63d0\u5347\u4e86\u6027\u80fd\u548c\u53ef\u6269\u5c55\u6027\u3002\u57fa\u4e8e\u8d85\u8fc747,000\u4e2a\u5de5\u5177\u7684\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cToolGen\u4e0d\u4ec5\u5728\u5de5\u5177\u68c0\u7d22\u548c\u81ea\u4e3b\u4efb\u52a1\u5b8c\u6210\u65b9\u9762\u53d6\u5f97\u4e86\u5353\u8d8a\u7684\u6548\u679c\uff0c\u8fd8\u4e3a\u65b0\u4e00\u4ee3\u80fd\u591f\u9002\u5e94\u5404\u79cd\u9886\u57df\u5de5\u5177\u7684AI\u4ee3\u7406\u5960\u5b9a\u4e86\u57fa\u7840\u3002\u901a\u8fc7\u4ece\u6839\u672c\u4e0a\u5c06\u5de5\u5177\u68c0\u7d22\u8f6c\u53d8\u4e3a\u751f\u6210\u5f0f\u8fc7\u7a0b\uff0cToolGen\u4e3a\u66f4\u7075\u6d3b\u3001\u9ad8\u6548\u548c\u81ea\u4e3b\u7684AI\u7cfb\u7edf\u94fa\u5e73\u4e86\u9053\u8def\u3002ToolGen\u652f\u6301\u7aef\u5230\u7aef\u7684\u5de5\u5177\u5b66\u4e60\uff0c\u5e76\u4e3a\u4e0e\u94fe\u5f0f\u601d\u7ef4\u548c\u5f3a\u5316\u5b66\u4e60\u7b49\u5176\u4ed6\u5148\u8fdb\u6280\u672f\u7684\u96c6\u6210\u63d0\u4f9b\u4e86\u673a\u4f1a\uff0c\u4ece\u800c\u6269\u5c55\u4e86LLM\u7684\u5b9e\u7528\u529f\u80fd\u3002<\/p>\n<h2>1 \u5f15\u8a00<\/h2>\n<p>\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5728\u5904\u7406\u5916\u90e8\u8f93\u5165\u3001\u6267\u884c\u64cd\u4f5c\u548c\u81ea\u4e3b\u5b8c\u6210\u4efb\u52a1\u65b9\u9762\u5c55\u793a\u4e86\u4ee4\u4eba\u5370\u8c61\u6df1\u523b\u7684\u80fd\u529b\uff08Gravitas\uff0c\u00a02023; Qin\u7b49\uff0c\u00a02023; Yao\u7b49\uff0c\u00a02023; Shinn\u7b49\uff0c\u00a02023; Wu\u7b49\uff0c\u00a02024a; Liu\u7b49\uff0c\u00a02024\uff09\u3002\u5728\u5404\u79cd\u4f7fLLMs\u4e0e\u5916\u90e8\u4e16\u754c\u4ea4\u4e92\u7684\u65b9\u6cd5\u4e2d\uff0c\u901a\u8fc7API\u8fdb\u884c\u5de5\u5177\u8c03\u7528\u5df2\u6210\u4e3a\u6700\u5e38\u7528\u548c\u6709\u6548\u7684\u65b9\u6cd5\u4e4b\u4e00\u3002\u7136\u800c\uff0c\u968f\u7740\u5de5\u5177\u6570\u91cf\u589e\u52a0\u5230\u6570\u4e07\u7ea7\uff0c\u73b0\u6709\u7684\u5de5\u5177\u68c0\u7d22\u548c\u6267\u884c\u65b9\u6cd5\u96be\u4ee5\u9ad8\u6548\u6269\u5c55\u3002<\/p>\n<p>\u5728\u5b9e\u9645\u573a\u666f\u4e2d\uff0c\u4e00\u79cd\u5e38\u89c1\u7684\u65b9\u6cd5\u662f\u5c06\u5de5\u5177\u68c0\u7d22\u4e0e\u5de5\u5177\u6267\u884c\u7ed3\u5408\u8d77\u6765\uff0c\u5373\u68c0\u7d22\u6a21\u578b\u9996\u5148\u7b5b\u9009\u51fa\u76f8\u5173\u5de5\u5177\uff0c\u518d\u4ea4\u7ed9LLM\u8fdb\u884c\u6700\u7ec8\u9009\u62e9\u548c\u6267\u884c\uff08Qin\u7b49\uff0c\u00a02023; Patil\u7b49\uff0c\u00a02023\uff09\u3002\u867d\u7136\u8fd9\u79cd\u7ec4\u5408\u65b9\u6cd5\u5728\u5904\u7406\u5927\u91cf\u5de5\u5177\u65f6\u6709\u6240\u5e2e\u52a9\uff0c\u4f46\u4e5f\u5b58\u5728\u660e\u663e\u7684\u5c40\u9650\uff1a\u68c0\u7d22\u6a21\u578b\u901a\u5e38\u4f9d\u8d56\u5c0f\u578b\u7f16\u7801\u5668\uff0c\u96be\u4ee5\u5168\u9762\u6355\u6349\u590d\u6742\u5de5\u5177\u548c\u67e5\u8be2\u7684\u8bed\u4e49\uff0c\u800c\u5c06\u68c0\u7d22\u4e0e\u6267\u884c\u5206\u79bb\u53ef\u80fd\u4f1a\u5bfc\u81f4\u4efb\u52a1\u5b8c\u6210\u8fc7\u7a0b\u4e2d\u7684\u6548\u7387\u4f4e\u4e0b\u548c\u9636\u6bb5\u6027\u504f\u5dee\u3002<\/p>\n<p>\u6b64\u5916\uff0cLLMs\u53ca\u5176\u5206\u8bcd\u5668\u4e3b\u8981\u5728\u81ea\u7136\u8bed\u8a00\u6570\u636e\u4e0a\u8fdb\u884c\u9884\u8bad\u7ec3\uff08Brown\u7b49\uff0c\u00a02020; Touvron\u7b49\uff0c\u00a02023\uff09\uff0c\u5176\u81ea\u8eab\u5bf9\u5de5\u5177\u76f8\u5173\u529f\u80fd\u7684\u5185\u5728\u77e5\u8bc6\u6709\u9650\u3002\u8fd9\u79cd\u77e5\u8bc6\u5dee\u8ddd\u5bfc\u81f4\u4e86\u6027\u80fd\u4e0d\u4f73\uff0c\u5c24\u5176\u662f\u5728LLM\u5fc5\u987b\u4f9d\u8d56\u68c0\u7d22\u7684\u5de5\u5177\u63cf\u8ff0\u8fdb\u884c\u51b3\u7b56\u65f6\u3002<\/p>\n<p>\u5728\u672c\u7814\u7a76\u4e2d\uff0c\u6211\u4eec\u5f15\u5165ToolGen\uff0c\u4e00\u79cd\u5c06\u73b0\u5b9e\u5de5\u5177\u77e5\u8bc6\u76f4\u63a5\u96c6\u6210\u5230LLM\u53c2\u6570\u4e2d\u5e76\u5c06\u5de5\u5177\u68c0\u7d22\u4e0e\u6267\u884c\u8f6c\u53d8\u4e3a\u7edf\u4e00\u751f\u6210\u4efb\u52a1\u7684\u65b0\u6846\u67b6\u3002\u5177\u4f53\u800c\u8a00\uff0cToolGen\u901a\u8fc7\u6269\u5c55LLM\u8bcd\u6c47\uff0c\u5c06\u5de5\u5177\u8868\u793a\u4e3a\u7279\u5b9a\u7684\u865a\u62dfToken\uff0c\u5e76\u8bad\u7ec3\u6a21\u578b\u5728\u5bf9\u8bdd\u4e0a\u4e0b\u6587\u4e2d\u751f\u6210\u8fd9\u4e9bToken\uff0c\u4ece\u800c\u66f4\u6709\u6548\u5730\u5229\u7528LLM\u7684\u9884\u5b58\u77e5\u8bc6\u6765\u5b9e\u73b0\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528\u3002<\/p>\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u6bcf\u4e2a\u5de5\u5177\u5728LLM\u8bcd\u6c47\u4e2d\u8868\u793a\u4e3a\u4e00\u4e2a\u72ec\u7279\u7684\u865a\u62dfToken\u3002\u5728\u9884\u8bad\u7ec3LLM\u7684\u57fa\u7840\u4e0a\uff0cToolGen\u7684\u8bad\u7ec3\u8fc7\u7a0b\u5305\u62ec\u4e09\u4e2a\u9636\u6bb5\uff1a\u5de5\u5177\u8bb0\u5fc6\u3001\u68c0\u7d22\u8bad\u7ec3\u548c\u4ee3\u7406\u8bad\u7ec3\u3002\u5728\u5de5\u5177\u8bb0\u5fc6\u9636\u6bb5\uff0c\u6a21\u578b\u5c06\u6bcf\u4e2a\u865a\u62df\u5de5\u5177Token\u4e0e\u5176\u6587\u6863\u5173\u8054\u3002\u5728\u68c0\u7d22\u8bad\u7ec3\u671f\u95f4\uff0c\u6a21\u578b\u5b66\u4e60\u57fa\u4e8e\u7528\u6237\u67e5\u8be2\u751f\u6210\u76f8\u5173\u7684\u5de5\u5177Token\u3002\u6700\u540e\uff0c\u5728\u7aef\u5230\u7aef\u7684\u4ee3\u7406\u8c03\u4f18\u4e2d\uff0c\u6a21\u578b\u88ab\u8bad\u7ec3\u4e3a\u81ea\u4e3b\u4ee3\u7406\uff0c\u751f\u6210\u8ba1\u5212\u548c\u5de5\u5177\uff0c\u5e76\u786e\u5b9a\u5b8c\u6210\u4efb\u52a1\u7684\u9002\u5f53\u53c2\u6570\u3002\u901a\u8fc7\u8c03\u7528\u5de5\u5177\u5e76\u4ece\u5916\u90e8\u73af\u5883\u83b7\u53d6\u53cd\u9988\uff0c\u6a21\u578b\u53ef\u4ee5\u9ad8\u6548\u800c\u4e00\u4f53\u5316\u5730\u5904\u7406\u7528\u6237\u67e5\u8be2\u3002\u56fe\u00a01\u00a0\u663e\u793a\u4e86ToolGen\u4e0e\u4f20\u7edf\u8303\u5f0f\u7684\u5bf9\u6bd4\u3002<\/p>\n<p>\u6211\u4eec\u5728\u4e24\u4e2a\u573a\u666f\u4e2d\u9a8c\u8bc1\u4e86ToolGen\u7684\u4f18\u8d8a\u6027\uff1a\u4e00\u4e2a\u5de5\u5177\u68c0\u7d22\u4efb\u52a1\u4e2d\uff0c\u6a21\u578b\u53ef\u4ee5\u4e3a\u7ed9\u5b9a\u67e5\u8be2\u68c0\u7d22\u5230\u6b63\u786e\u7684\u5de5\u5177\uff1b\u53e6\u4e00\u4e2a\u57fa\u4e8eLLM\u7684\u4ee3\u7406\u4efb\u52a1\u4e2d\uff0c\u6a21\u578b\u5b8c\u6210\u4e86\u6d89\u53ca\u5b9e\u9645API\u8c03\u7528\u7684\u590d\u6742\u4efb\u52a1\u3002\u5229\u7528\u5305\u542b47,000\u4e2a\u73b0\u5b9e\u5de5\u5177\u7684\u6570\u636e\u96c6\uff0cToolGen\u7684\u8868\u73b0\u4e0e\u9886\u5148\u7684\u5de5\u5177\u68c0\u7d22\u65b9\u6cd5\u76f8\u5f53\uff0c\u4f46\u6210\u672c\u663e\u8457\u964d\u4f4e\u4e14\u6548\u7387\u66f4\u9ad8\u3002\u6b64\u5916\uff0c\u5b83\u8d85\u8d8a\u4e86\u4f20\u7edf\u7684\u5de5\u5177\u5b66\u4e60\u8303\u5f0f\uff0c\u7a81\u663e\u4e86\u5176\u5728\u63a8\u8fdb\u66f4\u6709\u6548\u7684\u5de5\u5177\u4f7f\u7528\u7cfb\u7edf\u4e2d\u7684\u6f5c\u529b\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-7744\" title=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/35c7fe9dbae6458.jpg\" alt=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-1\" width=\"1495\" height=\"829\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/35c7fe9dbae6458.jpg 1495w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/35c7fe9dbae6458-300x166.jpg 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/35c7fe9dbae6458-1024x568.jpg 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/35c7fe9dbae6458-768x426.jpg 768w\" sizes=\"auto, (max-width: 1495px) 100vw, 1495px\" \/>\u56fe 1\uff1aToolGen\u4e0e\u4ee5\u5f80\u57fa\u4e8e\u68c0\u7d22\u7684\u65b9\u6cd5\u7684\u5bf9\u6bd4\u3002\u4ee5\u5f80\u65b9\u6cd5\u901a\u8fc7\u76f8\u4f3c\u6027\u5339\u914d\u7684\u65b9\u5f0f\u4f7f\u7528\u68c0\u7d22\u5668\u6765\u68c0\u7d22\u76f8\u5173\u5de5\u5177\uff0c\u7136\u540e\u5c06\u8fd9\u4e9b\u5de5\u5177\u653e\u5165LLM\u7684\u63d0\u793a\u4e2d\u8fdb\u884c\u9009\u62e9\u3002ToolGen\u53ef\u4ee5\u901a\u8fc7\u76f4\u63a5\u751f\u6210\u5de5\u5177Token\u6765\u5b9e\u73b0\u5de5\u5177\u68c0\u7d22\uff0c\u5e76\u4e14\u53ef\u4ee5\u5728\u4e0d\u4f9d\u8d56\u4efb\u4f55\u5916\u90e8\u68c0\u7d22\u5668\u7684\u60c5\u51b5\u4e0b\u5b8c\u6210\u4efb\u52a1\u3002<\/p>\n<p>ToolGen \u4ee3\u8868\u4e86\u4e00\u79cd\u5de5\u5177\u4ea4\u4e92\u7684\u65b0\u8303\u5f0f\uff0c\u5c06\u68c0\u7d22\u548c\u751f\u6210\u878d\u5408\u4e3a\u4e00\u4e2a\u7edf\u4e00\u7684\u6a21\u578b\u3002\u8fd9\u4e00\u521b\u65b0\u4e3a\u65b0\u4e00\u4ee3\u80fd\u591f\u9002\u5e94\u5404\u79cd\u9886\u57df\u5de5\u5177\u7684 AI \u4ee3\u7406\u5960\u5b9a\u4e86\u57fa\u7840\u3002\u6b64\u5916\uff0cToolGen \u4e3a\u5c06\u94fe\u5f0f\u601d\u7ef4\u63a8\u7406\u548c\u5f3a\u5316\u5b66\u4e60\u7b49\u5148\u8fdb\u6280\u672f\u4e0e\u5de5\u5177\u4f7f\u7528\u7684\u7edf\u4e00\u751f\u6210\u65b9\u5f0f\u76f8\u7ed3\u5408\u521b\u9020\u4e86\u65b0\u673a\u9047\uff0c\u6269\u5c55\u4e86\u5927\u8bed\u8a00\u6a21\u578b\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u7684\u80fd\u529b\u3002<\/p>\n<p>\u603b\u7ed3\u6765\u8bf4\uff0c\u6211\u4eec\u7684\u8d21\u732e\u5305\u62ec\uff1a<\/p>\n<ul>\n<li>\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u6846\u67b6 ToolGen\uff0c\u5b83\u901a\u8fc7\u865a\u62df Token \u5c06\u5de5\u5177\u68c0\u7d22\u548c\u6267\u884c\u6574\u5408\u5230\u5927\u8bed\u8a00\u6a21\u578b\u7684\u751f\u6210\u8fc7\u7a0b\u4e2d\u3002<\/li>\n<li>\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u4e09\u9636\u6bb5\u7684\u8bad\u7ec3\u8fc7\u7a0b\uff0c\u4f7f ToolGen \u80fd\u591f\u5728\u5de5\u5177\u68c0\u7d22\u548c API \u8c03\u7528\u65b9\u9762\u5b9e\u73b0\u9ad8\u6548\u548c\u53ef\u6269\u5c55\u6027\u3002<\/li>\n<li>\u5b9e\u9a8c\u9a8c\u8bc1\u8868\u660e\uff0cToolGen \u5728\u5927\u89c4\u6a21\u5de5\u5177\u5e93\u4e2d\u6bd4\u73b0\u6709\u6700\u4f73\u5de5\u5177\u68c0\u7d22\u65b9\u6cd5\u8868\u73b0\u51fa\u76f8\u5f53\u7684\u6027\u80fd\uff0c\u4f46\u6210\u672c\u66f4\u4f4e\u3001\u6548\u7387\u66f4\u9ad8\uff0c\u5e76\u4e14\u8d85\u8d8a\u4e86\u4f20\u7edf\u7684\u5de5\u5177\u5b66\u4e60\u8303\u5f0f\u3002<\/li>\n<\/ul>\n<h2>2 \u76f8\u5173\u5de5\u4f5c<\/h2>\n<h3>2.1 \u5de5\u5177\u68c0\u7d22<\/h3>\n<p>\u5de5\u5177\u68c0\u7d22\u5728\u5927\u8bed\u8a00\u6a21\u578b\u4ee3\u7406\u7684\u5b9e\u9645\u4efb\u52a1\u6267\u884c\u4e2d\u81f3\u5173\u91cd\u8981\uff0c\u5176\u4e2d\u5de5\u5177\u901a\u5e38\u7531\u5176\u6587\u6863\u63cf\u8ff0\u3002\u4f20\u7edf\u65b9\u6cd5\u5982\u7a00\u758f\u68c0\u7d22\uff08\u4f8b\u5982 <a href=\"https:\/\/www.kdjingpai.com\/en\/bm25\/\">BM25<\/a> (Robertson et al.,\u00a02009\uff09\uff09\u548c\u5bc6\u96c6\u68c0\u7d22\uff08\u4f8b\u5982 DPR (Karpukhin et al.,\u00a02020\uff09\u3001ANCE (Xiong et al.,\u00a02021\uff09\uff09\u4f9d\u8d56\u4e8e\u5927\u578b\u6587\u6863\u7d22\u5f15\u548c\u5916\u90e8\u6a21\u5757\uff0c\u5bfc\u81f4\u6548\u7387\u4f4e\u4e0b\u4e14\u96be\u4ee5\u5728\u7aef\u5230\u7aef\u4ee3\u7406\u6846\u67b6\u4e2d\u4f18\u5316\u3002\u4e00\u4e9b\u7814\u7a76\u63a2\u8ba8\u4e86\u66ff\u4ee3\u65b9\u6cd5\u3002\u4f8b\u5982\uff0cChen \u7b49\u4eba (2024b) \u91cd\u5199\u67e5\u8be2\u5e76\u63d0\u53d6\u5176\u610f\u56fe\uff0c\u9762\u5411\u65e0\u76d1\u7763\u68c0\u7d22\u8bbe\u7f6e\uff0c\u5c3d\u7ba1\u7ed3\u679c\u4e0d\u5982\u6709\u76d1\u7763\u65b9\u6cd5\u3002Xu \u7b49\u4eba (2024) \u63d0\u51fa\u4e00\u79cd\u65b9\u6cd5\uff0c\u901a\u8fc7\u57fa\u4e8e\u5de5\u5177\u53cd\u9988\u8fed\u4ee3\u5730\u4f18\u5316\u67e5\u8be2\uff0c\u63d0\u9ad8\u4e86\u68c0\u7d22\u7cbe\u5ea6\u4f46\u589e\u52a0\u4e86\u5ef6\u8fdf\u3002<\/p>\n<p>\u8fd1\u6765\uff0c\u751f\u6210\u5f0f\u68c0\u7d22\u6210\u4e3a\u4e00\u79cd\u6709\u524d\u666f\u7684\u65b0\u8303\u5f0f\uff0c\u6a21\u578b\u76f4\u63a5\u751f\u6210\u76f8\u5173\u7684\u6587\u6863\u6807\u8bc6\u7b26\uff0c\u800c\u975e\u4f9d\u8d56\u4f20\u7edf\u68c0\u7d22\u673a\u5236 (Wang et al.,\u00a02022; Sun et al.,\u00a02023; Kishore et al.,\u00a02023; Mehta et al.,\u00a02023; Chen et al.,\u00a02023b)\u3002\u53d7\u5230\u6b64\u542f\u53d1\uff0cToolGen \u5c06\u6bcf\u4e2a\u5de5\u5177\u8868\u793a\u4e3a\u4e00\u4e2a\u552f\u4e00\u7684 Token\uff0c\u4f7f\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528\u53ef\u4ee5\u4f5c\u4e3a\u751f\u6210\u4efb\u52a1\u6765\u5904\u7406\u3002\u9664\u4e86\u7b80\u5316\u68c0\u7d22\u5916\uff0c\u8fd9\u4e00\u8bbe\u8ba1\u8fd8\u80fd\u4e0e\u5176\u4ed6\u5927\u8bed\u8a00\u6a21\u578b\u548c\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u7684\u4ee3\u7406\u529f\u80fd\uff08\u5982\u94fe\u5f0f\u601d\u7ef4\u63a8\u7406 (Wei et al.,\u00a02023\uff09\u548c <a href=\"https:\/\/www.kdjingpai.com\/en\/react\/\">ReAct<\/a> (Yao et al.,\u00a02023\uff09\uff09\u5e73\u6ed1\u96c6\u6210\u3002\u901a\u8fc7\u5c06\u68c0\u7d22\u548c\u4efb\u52a1\u6267\u884c\u6574\u5408\u5230\u5355\u4e00\u7684\u5927\u8bed\u8a00\u6a21\u578b\u4ee3\u7406\u4e2d\uff0c\u5b83\u964d\u4f4e\u4e86\u5ef6\u8fdf\u548c\u8ba1\u7b97\u5f00\u9500\uff0c\u63d0\u9ad8\u4e86\u4efb\u52a1\u5b8c\u6210\u7684\u6548\u7387\u548c\u6548\u679c\u3002<\/p>\n<h3>2.2 \u5e26\u6709\u5de5\u5177\u8c03\u7528\u7684\u5927\u8bed\u8a00\u6a21\u578b\u4ee3\u7406<\/h3>\n<p>\u5927\u8bed\u8a00\u6a21\u578b\u5728\u638c\u63e1\u5404\u79cd\u4efb\u52a1\u6240\u9700\u7684\u5de5\u5177\u65b9\u9762\u5c55\u73b0\u4e86\u5f88\u5f3a\u7684\u6f5c\u529b\u3002\u7136\u800c\uff0c\u5927\u591a\u6570\u73b0\u6709\u7814\u7a76\u96c6\u4e2d\u5728\u6709\u9650\u7684\u4e00\u7ec4\u52a8\u4f5c\u4e0a (Chen et al.,\u00a02023a; Zeng et al.,\u00a02023; Yin et al.,\u00a02024; Wang et al.,\u00a02024)\u3002\u4f8b\u5982\uff0cToolformer (Schick et al.,\u00a02023) \u5fae\u8c03\u4e86 GPT-J \u4ee5\u5904\u7406\u4ec5\u4e94\u79cd\u5de5\u5177\uff08\u5982\u8ba1\u7b97\u5668\uff09\u3002\u867d\u7136\u8fd9\u79cd\u65b9\u6cd5\u5728\u72ed\u7a84\u4efb\u52a1\u4e2d\u6548\u679c\u826f\u597d\uff0c\u4f46\u5728\u5305\u542b\u5927\u91cf\u52a8\u4f5c\u7a7a\u95f4\u7684\u73b0\u5b9e\u573a\u666f\u4e2d\u5b58\u5728\u56f0\u96be\u3002ToolBench (Qin et al.,\u00a02023) \u6269\u5c55\u4e86\u7814\u7a76\u8303\u56f4\uff0c\u5f15\u5165\u4e86\u8d85\u8fc7 16000 \u79cd\u5de5\u5177\uff0c\u5f3a\u8c03\u4e86\u5728\u590d\u6742\u73af\u5883\u4e2d\u8fdb\u884c\u5de5\u5177\u9009\u62e9\u7684\u6311\u6218\u3002<\/p>\n<p>\u4e3a\u6267\u884c\u5de5\u5177\u9009\u62e9\uff0c\u5f53\u524d\u65b9\u6cd5\u901a\u5e38\u91c7\u7528\u4e00\u4e2a\u68c0\u7d22-\u751f\u6210\u6d41\u6c34\u7ebf\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u5148\u68c0\u7d22\u76f8\u5173\u5de5\u5177\uff0c\u7136\u540e\u8fdb\u884c\u5229\u7528 (Patil et al.,\u00a02023; Qin et al.,\u00a02023)\u3002\u7136\u800c\uff0c\u6d41\u6c34\u7ebf\u65b9\u6cd5\u9762\u4e34\u4e24\u4e2a\u4e3b\u8981\u95ee\u9898\uff1a\u68c0\u7d22\u6b65\u9aa4\u7684\u9519\u8bef\u4f20\u9012\uff0c\u4ee5\u53ca\u5927\u8bed\u8a00\u6a21\u578b\u96be\u4ee5\u901a\u8fc7\u7b80\u5355\u7684\u63d0\u793a\u5b8c\u5168\u7406\u89e3\u548c\u4f7f\u7528\u5de5\u5177\u3002<\/p>\n<p>\u4e3a\u7f13\u89e3\u8fd9\u4e9b\u95ee\u9898\uff0c\u7814\u7a76\u4eba\u5458\u5c1d\u8bd5\u5c06\u52a8\u4f5c\u8868\u793a\u4e3a Token\uff0c\u5c06\u52a8\u4f5c\u9884\u6d4b\u8f6c\u5316\u4e3a\u751f\u6210\u4efb\u52a1\u3002\u4f8b\u5982\uff0cRT2 (Brohan et al.,\u00a02023) \u751f\u6210\u4ee3\u8868\u673a\u5668\u4eba\u52a8\u4f5c\u7684 Token\uff0cSelf-RAG (Asai et al.,\u00a02023) \u4f7f\u7528\u7279\u6b8a Token \u51b3\u5b9a\u4f55\u65f6\u68c0\u7d22\u6587\u6863\u3002ToolkenGPT (Hao et al.,\u00a02023) \u5f15\u5165\u4e86\u5de5\u5177\u7279\u5b9a\u7684 Token \u4ee5\u89e6\u53d1\u5de5\u5177\u4f7f\u7528\uff0c\u8fd9\u4e00\u6982\u5ff5\u6700\u63a5\u8fd1\u6211\u4eec\u7684\u65b9\u6cd5\u3002<\/p>\n<p>\u6211\u4eec\u7684\u65b9\u6cd5\u4e0e ToolkenGPT \u5b58\u5728\u663e\u8457\u4e0d\u540c\u3002\u9996\u5148\uff0c\u6211\u4eec\u4e13\u6ce8\u4e8e\u9700\u8981\u7075\u6d3b\u53c2\u6570\u7528\u4e8e\u590d\u6742\u4efb\u52a1\u7684\u771f\u5b9e\u5de5\u5177\uff08\u4f8b\u5982\uff0cYouTube \u9891\u9053\u641c\u7d22\uff09\uff0c\u800c ToolkenGPT \u9650\u4e8e\u8f83\u7b80\u5355\u3001\u8f93\u5165\u8f83\u5c11\u7684\u5de5\u5177\uff08\u4f8b\u5982\uff0c\u5305\u542b\u4e24\u4e2a\u6570\u7684\u6570\u5b66\u51fd\u6570\uff09\u3002\u6b64\u5916\uff0cToolkenGPT \u4f9d\u8d56\u4e8e\u5c11\u6837\u672c\u63d0\u793a\uff0c\u800c ToolGen \u901a\u8fc7\u5168\u53c2\u6570\u5fae\u8c03\u5c06\u5de5\u5177\u77e5\u8bc6\u76f4\u63a5\u6574\u5408\u5230\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\uff0c\u4f7f\u5176\u80fd\u591f\u81ea\u4e3b\u68c0\u7d22\u548c\u6267\u884c\u4efb\u52a1\u3002\u6700\u540e\uff0c\u6211\u4eec\u7684\u5b9e\u9a8c\u6d89\u53ca\u66f4\u5927\u89c4\u6a21\u7684\u5de5\u5177\u96c6\u2014\u201447000 \u4e2a\u5de5\u5177\uff0c\u76f8\u8f83\u4e8e ToolkenGPT \u7684 13\u2013300 \u4e2a\u3002<\/p>\n<p>\u5176\u4ed6\u7814\u7a76\u5982 ToolPlanner (Wu et al.,\u00a02024b) \u548c AutoACT (Qiao et al.,\u00a02024) \u4f7f\u7528\u4e86\u5f3a\u5316\u5b66\u4e60\u6216\u591a\u4ee3\u7406\u7cfb\u7edf\u6765\u589e\u5f3a\u5de5\u5177\u5b66\u4e60\u6216\u4efb\u52a1\u5b8c\u6210 (Qiao et al.,\u00a02024; Liu et al.,\u00a02023; Shen et al.,\u00a02024; Chen et al.,\u00a02024a)\u3002\u6211\u4eec\u4e0d\u5c06\u8fd9\u4e9b\u65b9\u6cd5\u4e0e\u6211\u4eec\u7684\u6a21\u578b\u8fdb\u884c\u5bf9\u6bd4\uff0c\u539f\u56e0\u6709\u4e8c\uff1a(1) \u8fd9\u4e9b\u5de5\u4f5c\u5927\u591a\u4f9d\u8d56\u53cd\u9988\u673a\u5236\uff0c\u4f8b\u5982 <a href=\"https:\/\/www.kdjingpai.com\/en\/reflection-2\/\">Reflection<\/a> (Shinn et al.,\u00a02023) \u6216\u5956\u52b1\u6a21\u578b\uff0c\u8fd9\u7c7b\u4f3c\u4e8e ToolBench \u7684\u8bc4\u4f30\u8bbe\u8ba1\uff0c\u5176\u4e2d\u5927\u8bed\u8a00\u6a21\u578b\u4f5c\u4e3a\u8bc4\u4f30\u5668\u4e14\u65e0\u6cd5\u8bbf\u95ee\u771f\u5b9e\u7b54\u6848\u3002\u7136\u800c\uff0c\u8fd9\u4e0d\u662f\u6211\u4eec\u7814\u7a76\u7684\u91cd\u70b9\uff0c\u6211\u4eec\u7684\u7aef\u5230\u7aef\u5b9e\u9a8c\u4e0d\u4f9d\u8d56\u4e8e\u6b64\u7c7b\u53cd\u9988\u673a\u5236\u3002 (2) \u6211\u4eec\u7684\u65b9\u6cd5\u4e0e\u8fd9\u4e9b\u65b9\u6cd5\u4e0d\u51b2\u7a81\uff0c\u53cd\u800c\u53ef\u4ee5\u96c6\u6210\u3002\u5bf9\u8fd9\u79cd\u96c6\u6210\u7684\u63a2\u7d22\u7559\u5f85\u672a\u6765\u7814\u7a76\u3002<\/p>\n<h2>3 ToolGen<\/h2>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u9996\u5148\u4ecb\u7ecd\u8bba\u6587\u4e2d\u4f7f\u7528\u7684\u7b26\u53f7\u8868\u793a\u3002\u7136\u540e\uff0c\u6211\u4eec\u8be6\u7ec6\u63cf\u8ff0\u4e86 ToolGen \u7684\u5177\u4f53\u65b9\u6cd5\uff0c\u5305\u62ec\u5de5\u5177\u865a\u62df\u5316\u3001\u5de5\u5177\u8bb0\u5fc6\u3001\u68c0\u7d22\u8bad\u7ec3\u548c\u7aef\u5230\u7aef\u4ee3\u7406\u8c03\u4f18\uff0c\u5982\u56fe\u00a02\u00a0\u6240\u793a\u3002\u6700\u540e\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86\u6211\u4eec\u7684\u63a8\u7406\u65b9\u6cd5\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-7745\" title=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/a4e7067aed5d076.jpg\" alt=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-1\" width=\"1494\" height=\"995\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/a4e7067aed5d076.jpg 1494w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/a4e7067aed5d076-300x200.jpg 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/a4e7067aed5d076-1024x682.jpg 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/a4e7067aed5d076-768x511.jpg 768w\" sizes=\"auto, (max-width: 1494px) 100vw, 1494px\" \/>\u56fe 2: ToolGen \u6846\u67b6\u793a\u610f\u56fe\u3002\u5728\u5de5\u5177\u865a\u62df\u5316\u9636\u6bb5\uff0c\u5de5\u5177\u88ab\u6620\u5c04\u5230\u865a\u62df Token\u3002\u5728\u63a5\u4e0b\u6765\u7684\u4e09\u9636\u6bb5\u8bad\u7ec3\u4e2d\uff0cToolGen \u9996\u5148\u901a\u8fc7\u5de5\u5177\u6587\u6863\u9884\u6d4b\u5de5\u5177 Token \u6765\u8bb0\u5fc6\u5de5\u5177\u3002\u63a5\u7740\u901a\u8fc7\u4ece\u67e5\u8be2\u4e2d\u9884\u6d4b\u5de5\u5177 Token \u6765\u5b66\u4e60\u68c0\u7d22\u5de5\u5177\u3002\u6700\u540e\uff0c\u7ba1\u9053\u6570\u636e (\u5373\u8f68\u8ff9) \u88ab\u7528\u4e8e\u5fae\u8c03\u6700\u540e\u4e00\u9636\u6bb5\u7684\u68c0\u7d22\u6a21\u578b\uff0c\u751f\u6210 ToolGen Agent \u6a21\u578b\u3002<\/p>\n<h3>3.1 \u9884\u5907\u77e5\u8bc6<\/h3>\n<p>\u5bf9\u4e8e\u7ed9\u5b9a\u7684\u7528\u6237\u67e5\u8be2 q\uff0c\u5de5\u5177\u5b66\u4e60\u7684\u76ee\u6807\u662f\u901a\u8fc7\u4f7f\u7528\u6765\u81ea\u5927\u578b\u5de5\u5177\u96c6 D={d1,d2,\u2026,dN} \u7684\u5de5\u5177\u6765\u89e3\u51b3 q\uff0c\u5176\u4e2d |D|=N \u662f\u4e00\u4e2a\u5927\u6570\u91cf\uff0c\u8fd9\u4f7f\u5f97\u5728\u5927\u8bed\u8a00\u6a21\u578b\u73af\u5883\u4e2d\u5305\u542b D \u4e2d\u7684\u6240\u6709\u5de5\u5177\u53d8\u5f97\u4e0d\u5207\u5b9e\u9645\u3002\u56e0\u6b64\uff0c\u5f53\u524d\u7814\u7a76\u901a\u5e38\u4f7f\u7528\u68c0\u7d22\u5668 R \u4ece D \u4e2d\u68c0\u7d22 k \u4e2a\u76f8\u5173\u5de5\u5177\uff0c\u8bb0\u4e3a Dk,R={dr1,dr2,\u2026,drk}=R\u2062(q,k,D)\uff0c\u5176\u4e2d |Dk,R|\u226aN\u3002\u6700\u7ec8\u7684\u63d0\u793a\u8bcd\u662f q \u548c Dk,R \u7684\u4e32\u8054\uff0c\u8bb0\u4e3a Prompt=[q,Dk,R]\u3002\u4e3a\u4e86\u5b8c\u6210\u4e00\u4e2a\u4efb\u52a1 (\u67e5\u8be2) \uff0c\u57fa\u4e8e\u5927\u8bed\u8a00\u6a21\u578b\u7684\u4ee3\u7406\u901a\u5e38\u91c7\u7528\u56db\u9636\u6bb5\u8fed\u4ee3 (Qu \u7b49\u4eba\uff0c\u00a02024\u00a0)\uff1a\u751f\u6210\u4e00\u4e2a\u8ba1\u5212 pi\uff0c\u9009\u62e9\u5de5\u5177 dsi\uff0c\u786e\u5b9a\u5de5\u5177\u53c2\u6570 ci\uff0c\u5e76\u4ece\u5de5\u5177\u53cd\u9988 fi \u6536\u96c6\u4fe1\u606f\u3002\u6211\u4eec\u5c06\u8fd9\u4e9b\u6b65\u9aa4\u8bb0\u4e3a\u7b2c i \u6b21\u8fed\u4ee3\u7684 pi,dsi,ci,fi\u3002\u6a21\u578b\u5c06\u6301\u7eed\u8fed\u4ee3\u8fd9\u4e9b\u6b65\u9aa4\uff0c\u76f4\u5230\u4efb\u52a1\u5b8c\u6210\uff0c\u6b64\u65f6\u751f\u6210\u6700\u7ec8\u7b54\u6848 a\u3002\u6574\u4e2a\u8f68\u8ff9\u53ef\u4ee5\u8868\u793a\u4e3a Traj=[Prompt,(p1,ds1,c1,f1),\u2026,(pt,dst,ct,ft),a]=[q,R\u2062(q,D),(p1,ds1,c1,f1),\u2026,(pt,dst,ct,ft),a]\u3002\u8fd9\u79cd\u8fed\u4ee3\u65b9\u6cd5\u5141\u8bb8\u6a21\u578b\u5728\u6bcf\u4e00\u6b65\u6839\u636e\u6536\u5230\u7684\u53cd\u9988\u52a8\u6001\u8c03\u6574\u548c\u5b8c\u5584\u5176\u64cd\u4f5c\uff0c\u4ece\u800c\u63d0\u9ad8\u5176\u5b8c\u6210\u590d\u6742\u4efb\u52a1\u7684\u6027\u80fd\u3002<\/p>\n<h3>3.2 \u5de5\u5177\u865a\u62df\u5316<\/h3>\n<p>\u5728 ToolGen \u4e2d\uff0c\u6211\u4eec\u901a\u8fc7\u4e00\u79cd\u79f0\u4e3a\u539f\u5b50\u7d22\u5f15\u7684\u65b9\u6cd5\uff0c\u5c06\u6bcf\u4e2a\u5de5\u5177\u6620\u5c04\u5230\u4e00\u4e2a\u552f\u4e00\u7684\u65b0 Token\uff0c\u4ee5\u5b9e\u73b0\u5de5\u5177\u865a\u62df\u5316\u3002\u5728\u6b64\u65b9\u6cd5\u4e2d\uff0c\u901a\u8fc7\u6269\u5c55\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8bcd\u6c47\u8868\uff0c\u4e3a\u6bcf\u4e2a\u5de5\u5177\u5206\u914d\u4e00\u4e2a\u552f\u4e00\u7684 Token\u3002\u6bcf\u4e2a\u5de5\u5177 Token \u7684\u5d4c\u5165\u521d\u59cb\u5316\u4e3a\u5176\u5bf9\u5e94\u5de5\u5177\u540d\u79f0\u7684\u5e73\u5747\u5d4c\u5165\uff0c\u4ece\u800c\u4e3a\u6bcf\u4e2a\u5de5\u5177\u63d0\u4f9b\u4e86\u4e00\u4e2a\u8bed\u4e49\u4e0a\u6709\u610f\u4e49\u7684\u8d77\u70b9\u3002<\/p>\n<p>\u5f62\u5f0f\u4e0a\uff0cToken \u96c6\u5b9a\u4e49\u4e3a T=Index\u2062(d)|\u2200d\u2208D\uff0c\u5176\u4e2d Index \u662f\u5c06\u5de5\u5177\u6620\u5c04\u5230 Token \u7684\u51fd\u6570\u3002\u6211\u4eec\u8bc1\u660e\uff0c\u4e0e\u5176\u4ed6\u7d22\u5f15\u65b9\u6cd5 (\u4f8b\u5982\u8bed\u4e49\u548c\u6570\u5b57\u6620\u5c04\uff0c\u8ba8\u8bba\u89c1\u00a04.3\u00a0\u548c\u00a05.4\u00a0) \u76f8\u6bd4\uff0c\u539f\u5b50\u7d22\u5f15\u66f4\u4e3a\u9ad8\u6548\uff0c\u5e76\u4e14\u80fd\u591f\u51cf\u5c11\u5e7b\u89c9\u73b0\u8c61\u3002<\/p>\n<h3>3.3 \u5de5\u5177\u8bb0\u5fc6<\/h3>\n<p>\u5728\u4e3a\u5de5\u5177\u5206\u914d\u4e86 Token \u4e4b\u540e\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u4ecd\u7136\u7f3a\u4e4f\u5bf9\u5de5\u5177\u7684\u4efb\u4f55\u77e5\u8bc6\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e00\u95ee\u9898\uff0c\u6211\u4eec\u901a\u8fc7\u4f7f\u7528\u5de5\u5177\u63cf\u8ff0\u4f5c\u4e3a\u8f93\u5165\u548c\u5176\u5bf9\u5e94\u7684 Token \u4f5c\u4e3a\u8f93\u51fa\u5bf9\u5176\u8fdb\u884c\u5fae\u8c03\uff0c\u5c06\u6b64\u8fc7\u7a0b\u79f0\u4e3a\u5de5\u5177\u8bb0\u5fc6\u3002\u6211\u4eec\u4f7f\u7528\u4ee5\u4e0b\u635f\u5931\u51fd\u6570\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td>\u2112tool=\u2211d\u2208D\u2212log\u2061p\u03b8\u2062(Index\u2062(d)|ddoc)<\/td>\n<td><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5176\u4e2d \u03b8 \u8868\u793a\u5927\u8bed\u8a00\u6a21\u578b\u53c2\u6570\uff0cddoc \u8868\u793a\u5de5\u5177\u63cf\u8ff0\u3002\u6b64\u6b65\u9aa4\u4e3a\u5927\u8bed\u8a00\u6a21\u578b\u63d0\u4f9b\u4e86\u5de5\u5177\u548c\u76f8\u5173\u64cd\u4f5c\u7684\u57fa\u672c\u77e5\u8bc6\u3002<\/p>\n<h3>3.4 \u68c0\u7d22\u8bad\u7ec3<\/h3>\n<p>\u7136\u540e\u6211\u4eec\u8bad\u7ec3\u5927\u8bed\u8a00\u6a21\u578b\u5c06\u865a\u62df\u5de5\u5177 Token (\u53ca\u5176\u6587\u6863) \u7684\u9690\u85cf\u7a7a\u95f4\u4e0e\u7528\u6237\u67e5\u8be2\u7a7a\u95f4\u5173\u8054\uff0c\u4ece\u800c\u4f7f\u6a21\u578b\u80fd\u591f\u6839\u636e\u7528\u6237\u7684\u67e5\u8be2\u751f\u6210\u6b63\u786e\u7684\u5de5\u5177\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u901a\u8fc7\u7528\u6237\u67e5\u8be2\u4f5c\u4e3a\u8f93\u5165\u548c\u5bf9\u5e94\u5de5\u5177 Token \u4f5c\u4e3a\u8f93\u51fa\u6765\u5fae\u8c03\u5927\u8bed\u8a00\u6a21\u578b\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td>\u2112retrieval=\u2211q\u2208Q\u2211d\u2208Dq\u2212log\u2061p\u03b8\u2032\u2062(Index\u2062(d)|q)<\/td>\n<td><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u5176\u4e2d \u03b8\u2032 \u8868\u793a\u5de5\u5177\u8bb0\u5fc6\u540e\u7684\u5927\u8bed\u8a00\u6a21\u578b\u53c2\u6570\uff0cQ \u662f\u7528\u6237\u67e5\u8be2\u96c6\uff0cDq \u662f\u4e0e\u6bcf\u4e2a\u67e5\u8be2\u76f8\u5173\u7684\u5de5\u5177\u96c6\u3002\u8be5\u8fc7\u7a0b\u751f\u6210\u4e86 ToolGen \u68c0\u7d22\u5668\uff0c\u4f7f\u5176\u80fd\u591f\u5728\u7ed9\u5b9a\u7528\u6237\u67e5\u8be2\u65f6\u751f\u6210\u9002\u5f53\u7684\u5de5\u5177 Token\u3002<\/p>\n<h3>3.5 \u7aef\u5230\u7aef\u4ee3\u7406\u8c03\u4f18<\/h3>\n<p>\u5728\u68c0\u7d22\u8bad\u7ec3\u540e\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u80fd\u591f\u4ece\u67e5\u8be2\u4e2d\u751f\u6210\u5de5\u5177 Token\u3002\u5728\u6700\u540e\u9636\u6bb5\uff0c\u6211\u4eec\u4f7f\u7528\u4ee3\u7406\u4efb\u52a1\u5b8c\u6210\u8f68\u8ff9\u5fae\u8c03\u6a21\u578b\u3002\u6211\u4eec\u91c7\u7528\u7c7b\u4f3c Agent-Flan (Chen \u7b49\u4eba\uff0c\u00a02024c\u00a0) \u7684\u63a8\u7406\u7b56\u7565\uff0c\u6211\u4eec\u7684\u7ba1\u9053\u91c7\u7528\u8fed\u4ee3\u8fc7\u7a0b\uff0c\u6a21\u578b\u9996\u5148\u751f\u6210 Thought\uff0c\u7136\u540e\u751f\u6210\u5bf9\u5e94\u7684 Action Token\u3002\u8be5 Token \u7528\u4e8e\u83b7\u53d6\u5de5\u5177\u6587\u6863\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u5229\u7528\u8be5\u6587\u6863\u751f\u6210\u5fc5\u8981\u7684\u53c2\u6570\u3002\u8be5\u8fc7\u7a0b\u4e0d\u65ad\u8fed\u4ee3\uff0c\u76f4\u5230\u6a21\u578b\u751f\u6210\u201c\u5b8c\u6210\u201d Token \u6216\u8fbe\u5230\u6700\u5927\u56de\u5408\u6570\u3002\u751f\u6210\u7684\u8f68\u8ff9\u8868\u793a\u4e3a Traj=[q,(p1,Index\u2062(ds1),c1,f1),\u2026,(pt,Index\u2062(dst),ct,ft),a]\u3002\u5728\u6b64\u7ed3\u6784\u4e2d\uff0c\u4e0d\u518d\u9700\u8981\u5173\u8054\u5de5\u5177\u3002<\/p>\n<h3>3.6 \u63a8\u7406<\/h3>\n<p>\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u53ef\u80fd\u4f1a\u751f\u6210\u8d85\u51fa\u9884\u5b9a\u4e49\u5de5\u5177 Token \u96c6\u7684\u52a8\u4f5c Token\u3002\u4e3a\u9632\u6b62\u8fd9\u79cd\u60c5\u51b5\uff0c\u6211\u4eec\u8bbe\u8ba1\u4e86\u4e00\u79cd\u9650\u5236\u675f\u641c\u7d22\u751f\u6210\u65b9\u6cd5\uff0c\u5c06\u8f93\u51fa Token \u9650\u5236\u5728\u5de5\u5177 Token \u96c6\u4e2d\u3002\u6211\u4eec\u5c06\u6b64\u9650\u5236\u675f\u641c\u7d22\u5e94\u7528\u4e8e\u5de5\u5177\u68c0\u7d22 (\u6a21\u578b\u57fa\u4e8e\u67e5\u8be2\u9009\u62e9\u5de5\u5177) \u548c\u7aef\u5230\u7aef\u4ee3\u7406\u7cfb\u7edf\u4e2d\uff0c\u6709\u6548\u51cf\u5c11\u4e86\u52a8\u4f5c\u751f\u6210\u6b65\u9aa4\u4e2d\u7684\u5e7b\u89c9\u3002\u8be6\u7ec6\u5206\u6790\u89c1\u00a05.4\u00a0\u3002\u5b9e\u73b0\u7ec6\u8282\u89c1\u9644\u5f55\u00a0C\u00a0\u3002<\/p>\n<h2>4 \u5de5\u5177\u68c0\u7d22\u8bc4\u4f30<\/h2>\n<h3>4.1 \u5b9e\u9a8c\u8bbe\u7f6e<\/h3>\n<p>\u6211\u4eec\u4f7f\u7528\u9884\u8bad\u7ec3\u7684 Llama-3-8B \u6a21\u578b (Dubey \u7b49\u4eba\uff0c2024) \u4f5c\u4e3a\u57fa\u7840\u6a21\u578b\uff0c\u5176\u8bcd\u6c47\u91cf\u4e3a 128,256\u3002\u901a\u8fc7\u539f\u5b50\u7d22\u5f15\u65b9\u6cd5\uff0c\u5728\u5de5\u5177\u865a\u62df\u5316\u8fc7\u7a0b\u4e2d\u589e\u52a0\u4e86 46,985 \u4e2a Token\uff0c\u6700\u7ec8\u8bcd\u6c47\u91cf\u8fbe\u5230 175,241\u3002\u6211\u4eec\u4f7f\u7528 Llama-3 \u804a\u5929\u6a21\u677f\u5bf9\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u91c7\u7528\u4f59\u5f26\u5b66\u4e60\u7387\u8c03\u5ea6\u5668\u5e76\u5e94\u7528 3% \u7684\u9884\u70ed\u6b65\u957f\u3002\u6700\u5927\u5b66\u4e60\u7387\u4e3a 4\u00d710\u22125\u3002\u6240\u6709\u6a21\u578b\u90fd\u5728 4\u00d7A100 GPU \u4e0a\u901a\u8fc7 Deepspeed ZeRO 3 (Rajbhandari \u7b49\u4eba\uff0c<a href=\"https:\/\/arxiv.org\/html\/2410.03439v2#bib.bib26\">20<\/a>2<a href=\"https:\/\/arxiv.org\/html\/2410.03439v2#bib.bib26\">0<\/a>) \u8fdb\u884c\u8bad\u7ec3\u3002\u5de5\u5177\u8bb0\u5fc6\u8bad\u7ec3\u4e86 8 \u8f6e\uff0c\u68c0\u7d22\u8bad\u7ec3\u4e86 1 \u8f6e\u3002<\/p>\n<h4>\u6570\u636e\u96c6<\/h4>\n<p>\u6211\u4eec\u7684\u5b9e\u9a8c\u57fa\u4e8e ToolBench\uff0c\u8fd9\u662f\u4e00\u4e2a\u771f\u5b9e\u4e16\u754c\u7684\u5de5\u5177\u57fa\u51c6\uff0c\u5305\u542b\u8d85\u8fc7 16,000 \u4e2a\u5de5\u5177\u96c6\u5408\uff0c\u6bcf\u4e2a\u96c6\u5408\u5305\u542b\u591a\u4e2a API\uff0c\u603b\u5171\u7ea6 47,000 \u4e2a\u72ec\u7279\u7684 API\u3002\u6bcf\u4e2a API \u4f7f\u7528\u4e00\u4e2a\u5305\u542b API \u540d\u79f0\u3001\u63cf\u8ff0\u548c\u8c03\u7528\u53c2\u6570\u7684\u5b57\u5178\u8fdb\u884c\u6587\u6863\u5316\u3002\u5b9e\u9645\u793a\u4f8b\u8bf7\u89c1\u9644\u5f55\u00a0<a title=\"\u9644\u5f55 A: \u5b9e\u9645\u5de5\u5177\u793a\u4f8b \u2023 ToolGen: \u7edf\u4e00\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528\u751f\u6210\" href=\"https:\/\/arxiv.org\/html\/2410.03439v2#A1\">A<\/a>\u3002\u6211\u4eec\u5c06\u6bcf\u4e2a API \u89c6\u4e3a\u4e00\u4e2a\u64cd\u4f5c\uff0c\u5e76\u5c06\u5176\u6620\u5c04\u4e3a\u4e00\u4e2a Token\u3002\u6211\u4eec\u7684\u68c0\u7d22\u548c\u7aef\u5230\u7aef\u4ee3\u7406\u8c03\u4f18\u6570\u636e\u4ece ToolBench \u7684\u539f\u59cb\u6570\u636e\u8f6c\u6362\u800c\u6765\u3002\u8be6\u7ec6\u4fe1\u606f\u89c1\u9644\u5f55\u00a0<a title=\"\u9644\u5f55 G: ToolBench \u6570\u636e\u9002\u914d\u81f3 ToolGen \u2023 ToolGen: \u7edf\u4e00\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528\u751f\u6210\" href=\"https:\/\/arxiv.org\/html\/2410.03439v2#A7\">G<\/a>\u3002\u5c3d\u7ba1\u6bcf\u4e2a\u5de5\u5177\u53ef\u80fd\u5305\u542b\u591a\u4e2a API\uff0c\u4e3a\u7b80\u5316\u8d77\u89c1\uff0c\u672c\u6587\u5c06\u6bcf\u4e2a API \u7edf\u79f0\u4e3a\u5de5\u5177\u3002<\/p>\n<p>\u6211\u4eec\u9075\u5faa Qin \u7b49\u4eba (2023) \u7684\u6570\u636e\u5212\u5206\u65b9\u6cd5\uff0c\u5c06 200,000 \u5bf9 (\u67e5\u8be2\uff0c\u76f8\u5173 API) \u5212\u5206\u4e3a\u4e09\u7c7b\uff1aI1\uff08\u5355\u5de5\u5177\u67e5\u8be2\uff09\u3001I2\uff08\u7c7b\u5185\u591a\u5de5\u5177\u67e5\u8be2\uff09\u548c I3\uff08\u96c6\u5408\u5185\u591a\u5de5\u5177\u6307\u4ee4\uff09\uff0c\u5206\u522b\u5305\u542b 87,413\u300184,815 \u548c 25,251 \u4e2a\u5b9e\u4f8b\u3002<\/p>\n<h4>\u57fa\u51c6\u65b9\u6cd5<\/h4>\n<p>\u6211\u4eec\u5c06 ToolGen \u4e0e\u4ee5\u4e0b\u57fa\u51c6\u8fdb\u884c\u6bd4\u8f83\uff1a<\/p>\n<ul>\n<li>BM25\uff1a\u4e00\u79cd\u7ecf\u5178\u7684\u57fa\u4e8e TF-IDF \u7684\u65e0\u76d1\u7763\u68c0\u7d22\u65b9\u6cd5\uff0c\u57fa\u4e8e\u67e5\u8be2\u4e0e\u6587\u6863\u7684\u8bcd\u8bed\u76f8\u4f3c\u6027\u8fdb\u884c\u68c0\u7d22\u3002<\/li>\n<li>Embedding \u76f8\u4f3c\u5ea6 (EmbSim)\uff1a\u4f7f\u7528 OpenAI \u7684\u53e5\u5b50\u5d4c\u5165\u6a21\u578b\u751f\u6210\u7684\u53e5\u5b50\u5d4c\u5165\uff1b\u5177\u4f53\u4e3a\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\u4f7f\u7528\u7684 text-embedding-3-large\u3002<\/li>\n<li>Re-Invoke (Chen \u7b49\u4eba\uff0c2024b)\uff1a\u4e00\u79cd\u65e0\u76d1\u7763\u68c0\u7d22\u65b9\u6cd5\uff0c\u5305\u542b\u67e5\u8be2\u91cd\u5199\u548c\u6587\u6863\u6269\u5c55\u3002<\/li>\n<li>IterFeedback (Xu \u7b49\u4eba\uff0c2024)\uff1a\u57fa\u4e8e BERT \u7684\u68c0\u7d22\u5668\uff0c\u4f7f\u7528 gpt-3.5-turbo-0125 \u4f5c\u4e3a\u53cd\u9988\u6a21\u578b\uff0c\u53ef\u8fdb\u884c\u6700\u591a 10 \u8f6e\u7684\u8fed\u4ee3\u53cd\u9988\u3002<\/li>\n<li>ToolRetriever (Qin \u7b49\u4eba\uff0c2023)\uff1a\u4e00\u79cd\u901a\u8fc7\u5bf9\u6bd4\u5b66\u4e60\u8bad\u7ec3\u7684\u57fa\u4e8e BERT \u7684\u68c0\u7d22\u5668\u3002<\/li>\n<\/ul>\n<h4>\u8bbe\u7f6e<\/h4>\n<p>\u6211\u4eec\u5728\u4e24\u79cd\u8bbe\u7f6e\u4e0b\u8fdb\u884c\u5b9e\u9a8c\u3002\u5728\u7b2c\u4e00\u4e2a\u8bbe\u7f6e\u4e2d\uff0c\u57df\u5185\u68c0\u7d22\uff08In-Domain Retrieval\uff09\u9650\u5236\u68c0\u7d22\u641c\u7d22\u7a7a\u95f4\u4e3a\u540c\u4e00\u57df\u5185\u7684\u5de5\u5177\u3002\u4f8b\u5982\uff0c\u5728\u8bc4\u4f30 I1 \u57df\u7684\u67e5\u8be2\u65f6\uff0c\u4ec5\u9650\u4e8e I1 \u7684\u5de5\u5177\u3002\u6b64\u8bbe\u7f6e\u4e0e ToolBench \u7684\u8bbe\u7f6e\u4e00\u81f4\u3002\u7b2c\u4e8c\u4e2a\u8bbe\u7f6e\uff0c\u591a\u57df\u68c0\u7d22\uff08Multi-Domain Retrieval\uff09\u5219\u66f4\u4e3a\u590d\u6742\uff0c\u641c\u7d22\u7a7a\u95f4\u6269\u5c55\u5230\u6240\u6709\u4e09\u7c7b\u57df\u7684\u5de5\u5177\u3002\u5728\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u6a21\u578b\u5728\u5408\u5e76\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u589e\u52a0\u4e86\u641c\u7d22\u7a7a\u95f4\u548c\u4efb\u52a1\u590d\u6742\u6027\u3002\u4e0e ToolBench \u4e0d\u540c\uff0c\u591a\u57df\u8bbe\u7f6e\u53cd\u6620\u4e86\u68c0\u7d22\u4efb\u52a1\u53ef\u80fd\u6d89\u53ca\u91cd\u53e0\u6216\u6df7\u5408\u9886\u57df\u7684\u73b0\u5b9e\u573a\u666f\u3002\u8be5\u8bbe\u7f6e\u8bc4\u4f30\u4e86\u6a21\u578b\u5728\u8de8\u9886\u57df\u6cdb\u5316\u548c\u5904\u7406\u66f4\u590d\u6742\u591a\u6837\u7684\u68c0\u7d22\u4efb\u52a1\u7684\u80fd\u529b\u3002<\/p>\n<h4>\u6307\u6807<\/h4>\n<p>\u6211\u4eec\u4f7f\u7528\u5f52\u4e00\u5316\u6298\u6263\u7d2f\u79ef\u589e\u76ca (NDCG) (J\u00e4rvelin &amp; Kek\u00e4l\u00e4inen,\u00a02002) \u6765\u8bc4\u4f30\u68c0\u7d22\u6027\u80fd\uff0c\u8fd9\u662f\u6392\u540d\u4efb\u52a1\u4e2d\u5e7f\u6cdb\u4f7f\u7528\u7684\u6307\u6807\uff0c\u5305\u62ec\u5de5\u5177\u68c0\u7d22\u3002NDCG \u540c\u65f6\u8003\u8651\u4e86\u68c0\u7d22\u5230\u7684\u5de5\u5177\u7684\u76f8\u5173\u6027\u548c\u6392\u540d\u4f4d\u7f6e\u3002<\/p>\n<p>\u8868 1\uff1a\u5728\u4e24\u79cd\u8bbe\u7f6e\u4e0b\u7684\u5de5\u5177\u68c0\u7d22\u8bc4\u4f30\uff1a(1) \u57df\u5185\u68c0\u7d22\uff0c\u5176\u4e2d\u6a21\u578b\u5728\u76f8\u540c\u9886\u57df\u5185\u8fdb\u884c\u8bad\u7ec3\u548c\u8bc4\u4f30\uff1b(2) \u591a\u57df\u68c0\u7d22\uff0c\u6a21\u578b\u5728\u6240\u6709\u9886\u57df\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u5e76\u5728\u6240\u6709\u9886\u57df\u5de5\u5177\u7684\u5b8c\u6574\u96c6\u5408\u4e2d\u8fdb\u884c\u8bc4\u4f30\u3002BM25\u3001EmbSim \u548c Re-Invoke \u662f\u672a\u7ecf\u8bad\u7ec3\u7684\u65e0\u76d1\u7763\u57fa\u51c6\u65b9\u6cd5\u3002IterFeedback \u662f\u5e26\u6709\u591a\u6a21\u578b\u548c\u53cd\u9988\u673a\u5236\u7684\u68c0\u7d22\u7cfb\u7edf\u3002ToolRetriever \u4f7f\u7528\u5bf9\u6bd4\u5b66\u4e60\u8fdb\u884c\u8bad\u7ec3\uff0c\u800c ToolGen \u5219\u4f7f\u7528\u4e0b\u4e00\u4e2a Token \u9884\u6d4b\u8fdb\u884c\u8bad\u7ec3\u3002\u7ed3\u679c\u4e2d\u5e26 * \u7684\u9879\u76ee\u8868\u793a\u975e\u6211\u4eec\u5b9e\u73b0\u7684\u6a21\u578b\uff0c\u6570\u636e\u6765\u6e90\u4e8e\u539f\u59cb\u8bba\u6587\uff0c\u56e0\u800c\u4ec5\u5728\u57df\u5185\u8bbe\u7f6e\u4e0b\u5217\u51fa\u3002\u5bf9\u4e8e\u57df\u5185\u8bbe\u7f6e\u4e0b\u7684 ToolGen\uff0c\u6211\u4eec\u5141\u8bb8\u751f\u6210\u7a7a\u95f4\u5305\u62ec\u6240\u6709 Token\uff0c\u4e0e\u5176\u4ed6\u6a21\u578b\u76f8\u6bd4\u8fd9\u662f\u66f4\u5177\u6311\u6218\u6027\u7684\u573a\u666f\u3002\u6bcf\u4e2a\u7c7b\u522b\u4e2d\u6700\u4f18\u7684\u7ed3\u679c\u7528\u7c97\u4f53\u6807\u6ce8\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b<\/th>\n<th>I1<\/th>\n<th><\/th>\n<th><\/th>\n<th>I2<\/th>\n<th><\/th>\n<th><\/th>\n<th>I3<\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td>NDCG1<\/td>\n<td>NDCG3<\/td>\n<td>NDCG5<\/td>\n<td>NDCG1<\/td>\n<td>NDCG3<\/td>\n<td>NDCG5<\/td>\n<td>NDCG1<\/td>\n<td>NDCG3<\/td>\n<td>NDCG5<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td>\u57df\u5185<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>BM25<\/td>\n<td>29.46<\/td>\n<td>31.12<\/td>\n<td>33.27<\/td>\n<td>24.13<\/td>\n<td>25.29<\/td>\n<td>27.65<\/td>\n<td>32.00<\/td>\n<td>25.88<\/td>\n<td>29.78<\/td>\n<\/tr>\n<tr>\n<td>EmbSim<\/td>\n<td>63.67<\/td>\n<td>61.03<\/td>\n<td>65.37<\/td>\n<td>49.11<\/td>\n<td>42.27<\/td>\n<td>46.56<\/td>\n<td>53.00<\/td>\n<td>46.40<\/td>\n<td>52.73<\/td>\n<\/tr>\n<tr>\n<td>Re-Invoke*<\/td>\n<td>69.47<\/td>\n<td>\u2013<\/td>\n<td>61.10<\/td>\n<td>54.56<\/td>\n<td>\u2013<\/td>\n<td>53.79<\/td>\n<td>59.65<\/td>\n<td>\u2013<\/td>\n<td>59.55<\/td>\n<\/tr>\n<tr>\n<td>IterFeedback*<\/td>\n<td>90.70<\/td>\n<td>90.95<\/td>\n<td>92.47<\/td>\n<td>89.01<\/td>\n<td>85.46<\/td>\n<td>87.10<\/td>\n<td>91.74<\/td>\n<td>87.94<\/td>\n<td>90.20<\/td>\n<\/tr>\n<tr>\n<td>ToolRetriever<\/td>\n<td>80.50<\/td>\n<td>79.55<\/td>\n<td>84.39<\/td>\n<td>71.18<\/td>\n<td>64.81<\/td>\n<td>70.35<\/td>\n<td>70.00<\/td>\n<td>60.44<\/td>\n<td>64.70<\/td>\n<\/tr>\n<tr>\n<td>ToolGen<\/td>\n<td>89.17<\/td>\n<td>90.85<\/td>\n<td>92.67<\/td>\n<td>91.45<\/td>\n<td>88.79<\/td>\n<td>91.13<\/td>\n<td>87.00<\/td>\n<td>85.59<\/td>\n<td>90.16<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td>\u591a\u57df<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>BM25<\/td>\n<td>22.77<\/td>\n<td>22.64<\/td>\n<td>25.61<\/td>\n<td>18.29<\/td>\n<td>20.74<\/td>\n<td>22.18<\/td>\n<td>10.00<\/td>\n<td>10.08<\/td>\n<td>12.33<\/td>\n<\/tr>\n<tr>\n<td>EmbSim<\/td>\n<td>54.00<\/td>\n<td>50.82<\/td>\n<td>55.86<\/td>\n<td>40.84<\/td>\n<td>36.67<\/td>\n<td>39.55<\/td>\n<td>18.00<\/td>\n<td>17.77<\/td>\n<td>20.70<\/td>\n<\/tr>\n<tr>\n<td>ToolRetriever<\/td>\n<td>72.31<\/td>\n<td>70.30<\/td>\n<td>74.99<\/td>\n<td>64.54<\/td>\n<td>57.91<\/td>\n<td>63.61<\/td>\n<td>52.00<\/td>\n<td>39.89<\/td>\n<td>42.92<\/td>\n<\/tr>\n<tr>\n<td>ToolGen<\/td>\n<td>87.67<\/td>\n<td>88.84<\/td>\n<td>91.54<\/td>\n<td>83.46<\/td>\n<td>86.24<\/td>\n<td>88.84<\/td>\n<td>79.00<\/td>\n<td>79.80<\/td>\n<td>84.79<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>4.2 \u7ed3\u679c<\/h3>\n<p>\u8868\u683c\u00a01\u00a0\u5c55\u793a\u4e86\u5de5\u5177\u68c0\u7d22\u7684\u7ed3\u679c\u3002\u6b63\u5982\u9884\u671f\uff0c\u6240\u6709\u8bad\u7ec3\u8fc7\u7684\u6a21\u578b\u5728\u5404\u9879\u6307\u6807\u4e0a\u90fd\u660e\u663e\u4f18\u4e8e\u672a\u7ecf\u8bad\u7ec3\u7684\u57fa\u7ebf\uff08BM25\u3001EmbSim \u548c Re-Invoke\uff09\uff0c\u663e\u793a\u4e86\u5728\u5de5\u5177\u68c0\u7d22\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\u7684\u4f18\u52bf\u3002<\/p>\n<p>\u6211\u4eec\u63d0\u51fa\u7684 ToolGen \u6a21\u578b\u5728\u4e24\u79cd\u8bbe\u7f6e\u4e2d\u90fd\u6301\u7eed\u8868\u73b0\u6700\u4f73\u3002\u5728\u9886\u57df\u5185\u8bbe\u7f6e\u4e2d\uff0cToolGen \u63d0\u4f9b\u4e86\u9ad8\u5ea6\u7ade\u4e89\u529b\u7684\u7ed3\u679c\uff0c\u5176\u6027\u80fd\u4e0e\u4f7f\u7528\u591a\u4e2a\u6a21\u578b\u548c\u53cd\u9988\u673a\u5236\u7684 IterFeedback \u7cfb\u7edf\u76f8\u5f53\u3002\u4f5c\u4e3a\u5355\u4e00\u6a21\u578b\uff0cToolGen \u5728\u6240\u6709\u6307\u6807\u4e0a\u5747\u663e\u8457\u4f18\u4e8e ToolRetriever\uff0c\u751a\u81f3\u5728\u591a\u4e2a\u573a\u666f\uff08\u5982\u9886\u57df I1 \u7684 NDCG@5 \u548c I2 \u7684 NDCG@1\u3001@3\u3001@5\uff09\u4e2d\u8d85\u8fc7\u4e86 IterFeedback\u3002<\/p>\n<p>\u5728\u591a\u9886\u57df\u8bbe\u7f6e\u4e2d\uff0c\u5c3d\u7ba1\u641c\u7d22\u7a7a\u95f4\u66f4\u5927\uff0c\u6574\u4f53\u8868\u73b0\u901a\u5e38\u4e0b\u964d\uff0c\u4f46 ToolGen \u4f9d\u7136\u4fdd\u6301\u7a33\u5065\uff0c\u4f18\u4e8e ToolRetriever\uff0c\u5e76\u5728\u5404\u57fa\u7ebf\u6a21\u578b\u4e2d\u4fdd\u6301\u9886\u5148\u3002\u8fd9\u8868\u660e ToolGen \u5c3d\u7ba1\u662f\u5355\u4e00\u6a21\u578b\uff0c\u4f9d\u7136\u80fd\u4e0e\u50cf IterFeedback \u8fd9\u6837\u590d\u6742\u7684\u68c0\u7d22\u7cfb\u7edf\u7ade\u4e89\uff0c\u5c55\u73b0\u4e86\u5176\u5904\u7406\u9886\u57df\u8fb9\u754c\u4e0d\u6e05\u6670\u7684\u590d\u6742\u5b9e\u9645\u68c0\u7d22\u4efb\u52a1\u7684\u80fd\u529b\u3002<\/p>\n<h3>4.3 \u7d22\u5f15\u65b9\u6cd5\u5bf9\u6bd4<\/h3>\n<p>\u867d\u7136 ToolGen \u4f7f\u7528\u539f\u5b50\u7d22\u5f15\u8fdb\u884c\u5de5\u5177\u865a\u62df\u5316\uff0c\u6211\u4eec\u4e5f\u63a2\u8ba8\u4e86\u51e0\u79cd\u66ff\u4ee3\u7684\u751f\u6210\u5f0f\u68c0\u7d22\u65b9\u6cd5\u3002\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5c06\u5176\u4e0e\u4ee5\u4e0b\u4e09\u79cd\u65b9\u6cd5\u8fdb\u884c\u6bd4\u8f83\uff1a<\/p>\n<ul>\n<li>\u6570\u5b57\uff1a\u5c06\u6bcf\u4e2a\u5de5\u5177\u6620\u5c04\u5230\u4e00\u4e2a\u552f\u4e00\u7f16\u53f7\u3002\u751f\u6210\u7684 Token \u7eaf\u7cb9\u662f\u6570\u5b57\u5316\u7684\uff0c\u672a\u63d0\u4f9b\u5185\u5728\u7684\u8bed\u4e49\u4fe1\u606f\uff0c\u4f46\u4e3a\u6bcf\u4e2a\u5de5\u5177\u63d0\u4f9b\u4e86\u72ec\u7279\u7684\u6807\u8bc6\u7b26\u3002<\/li>\n<li>\u5206\u5c42\uff1a\u8be5\u65b9\u6cd5\u5c06\u5de5\u5177\u805a\u7c7b\u6210\u975e\u91cd\u53e0\u7ec4\uff0c\u5e76\u9012\u5f52\u5730\u5bf9\u8fd9\u4e9b\u96c6\u7fa4\u8fdb\u884c\u5212\u5206\uff0c\u5f62\u6210\u4e00\u79cd\u5c42\u7ea7\u7ed3\u6784\u3002\u8be5\u7ed3\u6784\u4e2d\u4ece\u6839\u8282\u70b9\u5230\u53f6\u8282\u70b9\u7684\u7d22\u5f15\u8868\u793a\u4e86\u6bcf\u4e2a\u5de5\u5177\uff0c\u7c7b\u4f3c\u4e8e Brown \u805a\u7c7b\u6280\u672f\u3002<\/li>\n<li>\u8bed\u4e49\uff1a\u6b64\u65b9\u6cd5\u4e2d\uff0c\u6bcf\u4e2a\u5de5\u5177\u6620\u5c04\u5230\u5176\u540d\u79f0\uff0c\u901a\u8fc7\u5de5\u5177\u540d\u79f0\u7684\u8bed\u4e49\u5185\u5bb9\u6765\u6307\u5bfc\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u3002\u5de5\u5177\u540d\u79f0\u76f4\u63a5\u63d0\u4f9b\u4e0e\u5176\u529f\u80fd\u76f8\u5173\u7684\u6709\u610f\u4e49\u7684\u8868\u793a\u3002<\/li>\n<\/ul>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-7739\" title=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-3\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/1414c9756505ddf.png\" alt=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-3\" width=\"1661\" height=\"845\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/1414c9756505ddf.png 1661w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/1414c9756505ddf-300x153.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/1414c9756505ddf-1024x521.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/1414c9756505ddf-768x391.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/1414c9756505ddf-1536x781.png 1536w\" sizes=\"auto, (max-width: 1661px) 100vw, 1661px\" \/><\/p>\n<p>\u56fe 3\uff1a\u6bcf\u79cd\u7d22\u5f15\u65b9\u6cd5\u4e0b\uff0c\u6bcf\u4e2a\u5de5\u5177\u7684\u5b50 Token \u6570\u91cf\u5206\u5e03\u4e0d\u540c\u3002\u539f\u5b50\u7d22\u5f15\u786e\u4fdd\u6bcf\u4e2a\u5de5\u5177\u662f\u4e00\u4e2a\u5355\u4e00\u7684 Token\uff0c\u800c\u6570\u5b57\u7d22\u5f15\u5c06\u5de5\u5177\u7f16\u7801\u6210 N \u4e2a Token \uff0c\u5373\u5de5\u5177\u7f16\u53f7\u4f4d\u4e8e (10N\u22121,10N] \u8303\u56f4\u5185\u3002\u76f8\u5bf9\u5730\uff0c\u8bed\u4e49\u7d22\u5f15\u548c\u5206\u5c42\u7d22\u5f15\u4ea7\u751f\u7684\u5b50 Token \u6570\u91cf\u53ef\u53d8\uff0c\u5176\u4e2d\u8bed\u4e49\u7d22\u5f15\u7684\u5b50 Token \u5e8f\u5217\u8f83\u957f\u4e14\u6709\u66f4\u591a\u5f02\u5e38\u503c\u3002<\/p>\n<p>\u5177\u4f53\u7684\u5b9e\u73b0\u7ec6\u8282\u63cf\u8ff0\u5728\u9644\u5f55\u00a0B\u00a0\u4e2d\u3002<\/p>\n<p>\u9996\u5148\uff0c\u6211\u4eec\u5206\u6790\u4e86\u6bcf\u79cd\u65b9\u6cd5\u8868\u793a\u6bcf\u4e2a\u5de5\u5177\u6240\u9700\u7684\u5b50 Token \u6570\u91cf\uff0c\u5982\u56fe\u00a03\u00a0\u6240\u793a\u3002\u56fe\u8868\u7a81\u663e\u4e86\u539f\u5b50\u7d22\u5f15\u7684\u4f18\u8d8a\u6027\uff0c\u5176\u4e2d\u6bcf\u4e2a\u5de5\u5177\u7531\u4e00\u4e2a\u5355\u4e00\u7684 Token \u8868\u793a\uff0c\u800c\u5176\u4ed6\u65b9\u6cd5\u9700\u8981\u591a\u4e2a Token\u3002\u8fd9\u79cd\u6548\u7387\u4f7f\u5f97 ToolGen \u80fd\u51cf\u5c11\u751f\u6210 Token \u7684\u6570\u91cf\uff0c\u5e76\u5728\u68c0\u7d22\u548c\u4ee3\u7406\u573a\u666f\u4e2d\u7f29\u77ed\u63a8\u7406\u65f6\u95f4\u3002<\/p>\n<p>\u63a5\u7740\uff0c\u6211\u4eec\u68c0\u9a8c\u4e86\u4e0d\u540c\u7d22\u5f15\u65b9\u6cd5\u7684\u6709\u6548\u6027\u3002\u5982\u8868\u00a02\u00a0\u6240\u793a\uff0c\u8bed\u4e49\u7d22\u5f15\u5728\u591a\u9879\u6307\u6807\u548c\u573a\u666f\u4e2d\u5c55\u73b0\u51fa\u6700\u4f73\u7684\u68c0\u7d22\u6027\u80fd\uff0c\u800c\u539f\u5b50\u7d22\u5f15\u5728\u8bb8\u591a\u60c5\u51b5\u4e0b\u7d27\u968f\u5176\u540e\u3002\u6211\u4eec\u5c06\u6b64\u5f52\u56e0\u4e8e\u8bed\u4e49\u7d22\u5f15\u66f4\u7b26\u5408\u5927\u8bed\u8a00\u6a21\u578b\u7684\u9884\u8bad\u7ec3\u6570\u636e\u3002\u7136\u800c\uff0c\u968f\u7740\u8bad\u7ec3\u6570\u636e\u548c\u7c7b\u578b\u7684\u589e\u52a0\uff0c\u8fd9\u79cd\u4f18\u52bf\u4f1a\u51cf\u5f31\u3002\u4f8b\u5982\uff0c\u5728\u7b2c\u00a05.3\u00a0\u8282\u4e2d\uff0c\u6211\u4eec\u5c55\u793a\u4e86\u539f\u5b50\u7d22\u5f15\u5728\u7aef\u5230\u7aef\u7ed3\u679c\u4e0a\u8868\u73b0\u66f4\u4f73\u3002\u7efc\u5408\u8003\u8651\u8fd9\u4e9b\u56e0\u7d20\uff0c\u6211\u4eec\u9009\u62e9\u539f\u5b50\u7d22\u5f15\u7528\u4e8e ToolGen \u5de5\u5177\u865a\u62df\u5316\u3002<\/p>\n<p>\u8868 2\uff1a\u5728\u591a\u9886\u57df\u8bbe\u7f6e\u4e0b\u4e0d\u540c\u7d22\u5f15\u65b9\u6cd5\u7684\u68c0\u7d22\u8bc4\u4ef7\u3002\u6700\u4f73\u7ed3\u679c\u4e3a\u52a0\u7c97\uff0c\u6b21\u4f73\u7ed3\u679c\u4e3a\u4e0b\u5212\u7ebf\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b<\/th>\n<th>I1<\/th>\n<th><\/th>\n<th><\/th>\n<th>I2<\/th>\n<th><\/th>\n<th><\/th>\n<th>I3<\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td>NDCG1<\/td>\n<td>NDCG3<\/td>\n<td>NDCG5<\/td>\n<td>NDCG1<\/td>\n<td>NDCG3<\/td>\n<td>NDCG5<\/td>\n<td>NDCG1<\/td>\n<td>NDCG3<\/td>\n<td>NDCG5<\/td>\n<\/tr>\n<tr>\n<td>\u6570\u5b57<\/td>\n<td>83.17<\/td>\n<td>84.99<\/td>\n<td>88.73<\/td>\n<td>79.20<\/td>\n<td>79.23<\/td>\n<td>83.88<\/td>\n<td>71.00<\/td>\n<td>74.81<\/td>\n<td>82.95<\/td>\n<\/tr>\n<tr>\n<td>\u5206\u5c42<\/td>\n<td>85.67<\/td>\n<td>87.38<\/td>\n<td>90.26<\/td>\n<td>82.22<\/td>\n<td>82.70<\/td>\n<td>86.63<\/td>\n<td>78.50<\/td>\n<td>79.47<\/td>\n<td>84.15<\/td>\n<\/tr>\n<tr>\n<td>\u8bed\u4e49<\/td>\n<td>89.17<\/td>\n<td>91.29<\/td>\n<td>93.29<\/td>\n<td>83.71<\/td>\n<td>84.51<\/td>\n<td>88.22<\/td>\n<td>82.00<\/td>\n<td>78.86<\/td>\n<td>85.43<\/td>\n<\/tr>\n<tr>\n<td>\u539f\u5b50<\/td>\n<td>87.67<\/td>\n<td>88.84<\/td>\n<td>91.54<\/td>\n<td>83.46<\/td>\n<td>86.24<\/td>\n<td>88.84<\/td>\n<td>79.00<\/td>\n<td>79.80<\/td>\n<td>84.79<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u8868 3\uff1a\u5de5\u5177\u68c0\u7d22\u7684\u6d88\u878d\u7814\u7a76\u3002\u5206\u522b\u8bc4\u4f30\u4e86\u53bb\u9664\u68c0\u7d22\u8bad\u7ec3\u3001\u5de5\u5177\u8bb0\u5fc6\u548c\u7ea6\u675f\u6ce2\u675f\u641c\u7d22\u5bf9 ToolGen \u6027\u80fd\u7684\u5f71\u54cd\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b<\/th>\n<th>I1<\/th>\n<th><\/th>\n<th><\/th>\n<th>I2<\/th>\n<th><\/th>\n<th><\/th>\n<th>I3<\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td>NDCG1<\/td>\n<td>NDCG3<\/td>\n<td>NDCG5<\/td>\n<td>NDCG1<\/td>\n<td>NDCG3<\/td>\n<td>NDCG5<\/td>\n<td>NDCG1<\/td>\n<td>NDCG3<\/td>\n<td>NDCG5<\/td>\n<\/tr>\n<tr>\n<td>ToolGen<\/td>\n<td>87.67<\/td>\n<td>88.84<\/td>\n<td>91.54<\/td>\n<td>83.46<\/td>\n<td>86.24<\/td>\n<td>88.84<\/td>\n<td>79.00<\/td>\n<td>79.80<\/td>\n<td>84.79<\/td>\n<\/tr>\n<tr>\n<td>\u2212\u8bb0\u5fc6<\/td>\n<td>84.00<\/td>\n<td>86.77<\/td>\n<td>89.35<\/td>\n<td>82.21<\/td>\n<td>83.20<\/td>\n<td>86.78<\/td>\n<td>77.00<\/td>\n<td>77.71<\/td>\n<td>84.37<\/td>\n<\/tr>\n<tr>\n<td>\u2212\u68c0\u7d22\u8bad\u7ec3<\/td>\n<td>10.17<\/td>\n<td>12.31<\/td>\n<td>13.89<\/td>\n<td>5.52<\/td>\n<td>7.01<\/td>\n<td>7.81<\/td>\n<td>3.00<\/td>\n<td>4.00<\/td>\n<td>4.43<\/td>\n<\/tr>\n<tr>\n<td>\u2212\u7ea6\u675f<\/td>\n<td>87.67<\/td>\n<td>88.79<\/td>\n<td>91.45<\/td>\n<td>83.46<\/td>\n<td>86.24<\/td>\n<td>88.83<\/td>\n<td>79.00<\/td>\n<td>79.93<\/td>\n<td>84.92<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>4.4 \u6d88\u878d\u5b9e\u9a8c<\/h3>\n<p>\u6211\u4eec\u8fdb\u884c\u4e86\u6d88\u878d\u5b9e\u9a8c\u4ee5\u8bc4\u4f30 ToolGen \u4e0d\u540c\u8bad\u7ec3\u9636\u6bb5\u7684\u5f71\u54cd\uff0c\u5982\u8868\u00a0<a title=\"\u8868 3 \u2023 4.3 \u7d22\u5f15\u65b9\u6cd5\u6bd4\u8f83 \u2023 4 \u5de5\u5177\u68c0\u7d22\u8bc4\u4f30 \u2023 ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u8fdb\u884c\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528\" href=\"https:\/\/arxiv.org\/html\/2410.03439v2#S4.T3\">3<\/a>\u00a0\u6240\u793a\u3002\u7ed3\u679c\u8868\u660e\uff0c\u68c0\u7d22\u8bad\u7ec3\u662f\u5f71\u54cd\u5de5\u5177\u68c0\u7d22\u6027\u80fd\u7684\u5173\u952e\u56e0\u7d20\uff0c\u56e0\u4e3a\u5b83\u76f4\u63a5\u5bf9\u9f50\u4e86\u8f93\u5165\u4e3a\u67e5\u8be2\u3001\u8f93\u51fa\u4e3a\u5de5\u5177 Token \u7684\u68c0\u7d22\u4efb\u52a1\u3002\u53bb\u9664\u5de5\u5177\u8bb0\u5fc6\u540e\u6027\u80fd\u7565\u6709\u4e0b\u964d\uff0c\u5c3d\u7ba1\u5176\u5bf9\u63d0\u5347\u6cdb\u5316\u6027\u80fd\u6709\u5e2e\u52a9\uff0c\u6211\u4eec\u5c06\u5728\u9644\u5f55\u00a0<a title=\"\u9644\u5f55 F \u6cdb\u5316 \u2023 ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u8fdb\u884c\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528\" href=\"https:\/\/arxiv.org\/html\/2410.03439v2#A6\">F<\/a>\u00a0\u4e2d\u8fdb\u4e00\u6b65\u8ba8\u8bba\u3002\u540c\u6837\uff0c\u53d7\u9650\u675f\u641c\u7d22\u867d\u7136\u5bf9\u68c0\u7d22\u4efb\u52a1\u8d21\u732e\u4e0d\u5927\uff0c\u4f46\u6709\u52a9\u4e8e\u9632\u6b62\u5e7b\u89c9\u751f\u6210\uff0c\u56e0\u6b64\u5728\u7aef\u5230\u7aef\u4ee3\u7406\u4efb\u52a1\u4e2d\u5177\u6709\u4e00\u5b9a\u4ef7\u503c\uff0c\u89c1\u7b2c\u00a05.4\u00a0\u8282\u3002<\/p>\n<h2>5 \u7aef\u5230\u7aef\u8bc4\u4f30<\/h2>\n<h3>5.1 \u5b9e\u9a8c\u8bbe\u7f6e<\/h3>\n<p>\u6211\u4eec\u5bf9\u6765\u81ea ToolBench \u7684\u8f68\u8ff9\u6570\u636e\u8fdb\u884c\u4e86\u591a\u9879\u4fee\u6539\uff0c\u4f7f\u5176\u9002\u5e94 ToolGen \u6846\u67b6\u3002\u4f8b\u5982\uff0c\u7531\u4e8e ToolGen \u4e0d\u9700\u8981\u663e\u5f0f\u9009\u62e9\u76f8\u5173\u5de5\u5177\u4f5c\u4e3a\u8f93\u5165\uff0c\u56e0\u6b64\u6211\u4eec\u5728\u7cfb\u7edf\u63d0\u793a\u4e2d\u79fb\u9664\u4e86\u8fd9\u4e9b\u4fe1\u606f\u3002\u66f4\u591a\u7ec6\u8282\u89c1\u9644\u5f55\u00a0G\u3002\u5728\u6b64\u57fa\u7840\u4e0a\uff0c\u6211\u4eec\u4f7f\u7528\u91cd\u65b0\u683c\u5f0f\u5316\u7684\u6570\u636e\u5fae\u8c03\u4e86\u68c0\u7d22\u6a21\u578b\uff0c\u751f\u6210\u4e86\u4e00\u4e2a\u7aef\u5230\u7aef\u7684 ToolGen \u4ee3\u7406\u3002<\/p>\n<h4>\u57fa\u7ebf\u6a21\u578b<\/h4>\n<ul>\n<li>GPT-3.5\uff1a\u6211\u4eec\u4f7f\u7528 gpt-3.5-turbo-0613 \u4f5c\u4e3a\u5176\u4e2d\u4e00\u4e2a\u57fa\u7ebf\u6a21\u578b\u3002\u5b9e\u73b0\u65b9\u5f0f\u4e0e StableToolBench (Guo et al.,\u00a02024) \u76f8\u540c\uff0cGPT-3.5 \u7684\u5de5\u5177\u8c03\u7528\u529f\u80fd\u88ab\u7528\u4e8e\u5f62\u6210\u5de5\u5177\u4ee3\u7406\u3002<\/li>\n<li>ToolLlama-2\uff1aQin \u7b49\u4eba (2023) \u901a\u8fc7\u5728 ToolBench \u6570\u636e\u4e0a\u5fae\u8c03 Llama-2 (Touvron et al.,\u00a02023) \u6a21\u578b\u5f15\u5165\u4e86 ToolLlama-2\u3002<\/li>\n<li>ToolLlama-3\uff1a\u4e3a\u4e86\u4fdd\u8bc1\u516c\u5e73\u6bd4\u8f83\uff0c\u6211\u4eec\u5728 ToolBench \u6570\u636e\u96c6\u4e0a\u5fae\u8c03\u4e86\u4e0e ToolGen \u76f8\u540c\u7684\u57fa\u7840\u6a21\u578b Llama-3\uff0c\u521b\u5efa\u4e86 ToolLlama-3 \u57fa\u7ebf\u6a21\u578b\u3002\u672c\u6587\u4e2d\u5269\u4f59\u90e8\u5206\u5c06 ToolLlama-3 \u7b80\u79f0\u4e3a ToolLlama \u4ee5\u533a\u5206 ToolLlama-2\u3002<\/li>\n<\/ul>\n<h4>\u8bbe\u7f6e<\/h4>\n<ul>\n<li>\u4f7f\u7528\u771f\u5b9e\u5de5\u5177 (G.T.)\uff1a\u53c2\u8003 Qin \u7b49\u4eba (2023)\uff0c\u6211\u4eec\u5b9a\u4e49\u67e5\u8be2\u7684\u771f\u5b9e\u5de5\u5177\u4e3a <a href=\"https:\/\/www.kdjingpai.com\/en\/chatgpt-6\/\">ChatGPT<\/a> \u9009\u62e9\u7684\u5de5\u5177\u3002\u5bf9\u4e8e ToolLlama\uff0c\u6211\u4eec\u5728\u63d0\u793a\u4e2d\u76f4\u63a5\u8f93\u5165\u771f\u5b9e\u5de5\u5177\uff0c\u4e0e\u5176\u8bad\u7ec3\u6570\u636e\u683c\u5f0f\u4e00\u81f4\u3002\u5bf9\u4e8e ToolGen\uff0c\u7531\u4e8e\u672a\u5728\u9884\u5148\u9009\u62e9\u5de5\u5177\u7684\u6570\u636e\u4e0a\u8fdb\u884c\u8bad\u7ec3\uff0c\u6211\u4eec\u5728\u89c4\u5212\u9636\u6bb5\u6dfb\u52a0\u4e86\u524d\u7f00\uff1a\u6211\u6b63\u5728\u4f7f\u7528\u4ee5\u4e0b\u5de5\u5177\uff1a[tool <a href=\"https:\/\/www.kdjingpai.com\/en\/tokenization\/\">tokens<\/a>]\uff0c\u5176\u4e2d [tool tokens] \u662f\u5bf9\u5e94\u771f\u5b9e\u5de5\u5177\u7684\u865a\u62df Token\u3002<\/li>\n<li>\u4f7f\u7528\u68c0\u7d22\u5668\uff1a\u5728\u7aef\u5230\u7aef\u5b9e\u9a8c\u4e2d\uff0c\u6211\u4eec\u4f7f\u7528\u57fa\u4e8e\u68c0\u7d22\u7684\u8bbe\u7f6e\u3002\u5bf9\u4e8e\u57fa\u7ebf\u6a21\u578b\uff0c\u6211\u4eec\u4f7f\u7528 ToolRetriever \u68c0\u7d22\u7684\u5de5\u5177\u4f5c\u4e3a\u76f8\u5173\u5de5\u5177\u3002\u800c ToolGen \u76f4\u63a5\u751f\u6210\u5de5\u5177 Token\uff0c\u56e0\u6b64\u4e0d\u4f7f\u7528\u68c0\u7d22\u5668\u3002<\/li>\n<\/ul>\n<p>\u6240\u6709\u6a21\u578b\u5747\u4f7f\u7528\u4f59\u5f26\u8c03\u5ea6\u5668\u8fdb\u884c\u5fae\u8c03\uff0c\u6700\u5927\u5b66\u4e60\u7387\u8bbe\u4e3a 4\u00d710\u22125\u3002\u4e0a\u4e0b\u6587\u957f\u5ea6\u622a\u65ad\u4e3a 6,144\uff0c\u603b\u6279\u6b21\u5927\u5c0f\u4e3a 512\u3002\u6211\u4eec\u8fdb\u4e00\u6b65\u4f7f\u7528 Flash-Attention (Dao \u7b49,\u00a02022\uff1bDao,\u00a02024) \u548c Deepspeed ZeRO 3 (Rajbhandari \u7b49,\u00a02020) \u4ee5\u8282\u7701\u5185\u5b58\u3002<\/p>\n<p>ToolGen \u548c ToolLlama \u9075\u5faa\u4e0d\u540c\u7684\u8303\u5f0f\u5b8c\u6210\u4efb\u52a1\u3002ToolLlama \u5728\u5355\u8f6e\u4e2d\u751f\u6210\u601d\u60f3\u3001\u52a8\u4f5c\u548c\u53c2\u6570\uff0c\u800c ToolGen \u5c06\u8fd9\u4e9b\u6b65\u9aa4\u5206\u5f00\u3002\u5bf9\u4e8e ToolGen\uff0c\u6211\u4eec\u8bbe\u7f6e\u6700\u591a 16 \u8f6e\u4e0a\u9650\uff0c\u5141\u8bb8\u8fdb\u884c 5 \u8f6e\u52a8\u4f5c\u548c 1 \u8f6e\u6700\u7ec8\u56de\u7b54\u3002\u6211\u4eec\u5c06\u6b64\u4e0e ToolLlama \u7684 6 \u8f6e\u9650\u5236\u8fdb\u884c\u6bd4\u8f83\u3002<\/p>\n<p>\u6b64\u5916\uff0c\u6211\u4eec\u4e3a\u6240\u6709\u6a21\u578b\u5f15\u5165\u4e86\u91cd\u8bd5\u673a\u5236\uff0c\u4ee5\u9632\u6b62\u8fc7\u65e9\u7ec8\u6b62\uff0c\u5177\u4f53\u7ec6\u8282\u5728\u9644\u5f55\u00a0D\u00a0\u4e2d\u4ecb\u7ecd\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5982\u679c\u6a21\u578b\u751f\u6210\u7684\u54cd\u5e94\u5305\u542b\u201cgive up\u201d\u6216\u201cI&#8217;m sorry\u201d\uff0c\u6211\u4eec\u4f1a\u4ee5\u66f4\u9ad8\u7684\u6e29\u5ea6\u63d0\u793a\u6a21\u578b\u91cd\u65b0\u751f\u6210\u54cd\u5e94\u3002<\/p>\n<h4>\u8bc4\u4f30\u6307\u6807<\/h4>\n<p>\u5bf9\u4e8e\u7aef\u5230\u7aef\u8bc4\u4f30\uff0c\u6211\u4eec\u4f7f\u7528\u7a33\u5b9a\u7684\u5de5\u5177\u8bc4\u4f30\u57fa\u51c6 StableToolBench (Guo et al.,\u00a02024)\uff0c\u8be5\u57fa\u51c6\u4ece ToolBench \u4e2d\u9009\u62e9\u53ef\u89e3\u51b3\u7684\u67e5\u8be2\uff0c\u5e76\u4f7f\u7528 GPT-4 (OpenAI,\u00a02024) \u6a21\u62df\u5931\u8d25\u5de5\u5177\u7684\u8f93\u51fa\u3002\u6211\u4eec\u4f7f\u7528\u4e24\u9879\u6307\u6807\u8bc4\u4f30\u6027\u80fd\uff1a\u53ef\u89e3\u901a\u8fc7\u7387 (SoPR)\uff0c\u5373\u6210\u529f\u89e3\u51b3\u67e5\u8be2\u7684\u767e\u5206\u6bd4\uff1b\u4ee5\u53ca\u53ef\u89e3\u80dc\u7387 (SoWR)\uff0c\u8868\u793a\u4f18\u4e8e\u53c2\u8003\u6a21\u578b (\u672c\u7814\u7a76\u4e2d\u4e3a GPT-3.5) \u7684\u7b54\u6848\u7684\u767e\u5206\u6bd4\u3002\u6b64\u5916\uff0c\u6211\u4eec\u63d0\u4f9b\u6bcf\u4e2a\u7c7b\u522b\u7684\u5fae\u5e73\u5747\u5206\u3002<\/p>\n<p>\u8868 4\uff1a\u5728\u4e24\u79cd\u8bbe\u7f6e\u4e0b\u5bf9\u672a\u89c1\u6307\u4ee4\u8fdb\u884c\u7aef\u5230\u7aef\u8bc4\u4f30\u7684\u6027\u80fd\u8868\u73b0\u3002\u5bf9\u4e8e R. \u8bbe\u7f6e\uff0cGPT-3.5 \u548c ToolLlama \u4f7f\u7528 ToolRetriever\uff0c\u800c ToolGen \u4e0d\u4f7f\u7528\u5916\u90e8\u68c0\u7d22\u5668\u3002\u6240\u6709\u7ed3\u679c\u7684 SoPR \u548c SoWR \u8bc4\u4f30\u5747\u8fdb\u884c\u4e09\u6b21\u5e76\u4ee5\u5747\u503c\u62a5\u544a\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b<\/th>\n<th>SoPR<\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<th>SoWR<\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td>I1<\/td>\n<td>I2<\/td>\n<td>I3<\/td>\n<td>Avg.<\/td>\n<td>I1<\/td>\n<td>I2<\/td>\n<td>I3<\/td>\n<td>Avg<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>\u4f7f\u7528\u771f\u5b9e\u5de5\u5177 (G.T.)<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>GPT-3.5<\/td>\n<td>56.60<\/td>\n<td>47.80<\/td>\n<td>54.64<\/td>\n<td>50.91<\/td>\n<td>&#8211;<\/td>\n<td>&#8211;<\/td>\n<td>&#8211;<\/td>\n<td>&#8211;<\/td>\n<\/tr>\n<tr>\n<td>ToolLlama-2<\/td>\n<td>53.37<\/td>\n<td>41.98<\/td>\n<td>46.45<\/td>\n<td>48.43<\/td>\n<td>47.27<\/td>\n<td>59.43<\/td>\n<td>27.87<\/td>\n<td>47.58<\/td>\n<\/tr>\n<tr>\n<td>ToolLlama<\/td>\n<td>55.93<\/td>\n<td>48.27<\/td>\n<td>52.19<\/td>\n<td>52.78<\/td>\n<td>50.31<\/td>\n<td>53.77<\/td>\n<td>31.15<\/td>\n<td>47.88<\/td>\n<\/tr>\n<tr>\n<td>ToolGen<\/td>\n<td>61.35<\/td>\n<td>49.53<\/td>\n<td>43.17<\/td>\n<td>54.19<\/td>\n<td>51.53<\/td>\n<td>57.55<\/td>\n<td>31.15<\/td>\n<td>49.70<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>\u4f7f\u7528\u68c0\u7d22\u5668 (R.)<\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>GPT-3.5<\/td>\n<td>51.43<\/td>\n<td>41.19<\/td>\n<td>34.43<\/td>\n<td>45.00<\/td>\n<td>53.37<\/td>\n<td>53.77<\/td>\n<td>37.70<\/td>\n<td>50.60<\/td>\n<\/tr>\n<tr>\n<td>ToolLlama-2<\/td>\n<td>56.13<\/td>\n<td>49.21<\/td>\n<td>34.70<\/td>\n<td>49.95<\/td>\n<td>50.92<\/td>\n<td>53.77<\/td>\n<td>21.31<\/td>\n<td>46.36<\/td>\n<\/tr>\n<tr>\n<td>ToolLlama<\/td>\n<td>54.60<\/td>\n<td>49.96<\/td>\n<td>51.37<\/td>\n<td>51.55<\/td>\n<td>49.08<\/td>\n<td>61.32<\/td>\n<td>31.15<\/td>\n<td>49.70<\/td>\n<\/tr>\n<tr>\n<td>ToolGen<\/td>\n<td>56.13<\/td>\n<td>52.20<\/td>\n<td>47.54<\/td>\n<td>53.28<\/td>\n<td>50.92<\/td>\n<td>62.26<\/td>\n<td>34.42<\/td>\n<td>51.51<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>5.2 \u7ed3\u679c<\/h3>\n<p>\u8868\u00a04\u00a0\u5c55\u793a\u4e86\u5728\u4e24\u79cd\u73af\u5883\u4e0b\u5404\u6a21\u578b\u7684\u7aef\u5230\u7aef\u8bc4\u4f30\u6027\u80fd\uff1a\u4f7f\u7528\u771f\u5b9e\u5de5\u5177 (G.T.) \u548c\u68c0\u7d22\u5668 (R.)\u3002\u5728 G.T. \u73af\u5883\u4e0b\uff0cToolGen \u7684\u5e73\u5747 SoPR \u5f97\u5206\u4e3a 54.19\uff0c\u4f18\u4e8e GPT-3.5 \u548c ToolLlama\uff0c\u4e14 ToolGen \u7684 SoWR \u5f97\u5206\u6700\u9ad8\u4e3a 49.70\u3002\u5728\u68c0\u7d22\u5668\u73af\u5883\u4e0b\uff0cToolGen \u4ecd\u4fdd\u6301\u9886\u5148\uff0c\u5e73\u5747 SoPR \u4e3a 53.28\uff0cSoWR \u4e3a 51.51\u3002ToolLlama \u5c55\u73b0\u4e86\u7ade\u4e89\u529b\uff0c\u5728\u67d0\u4e9b\u4e2a\u4f53\u5b9e\u4f8b\u4e0a\u4f18\u4e8e ToolGen\u3002\u7aef\u5230\u7aef ToolGen \u7684\u6d88\u878d\u7814\u7a76\u89c1\u9644\u5f55\u00a0G\u00a0\u3002<\/p>\n<p>\u8868 5\uff1a\u4e0d\u540c\u7d22\u5f15\u65b9\u6cd5\u7684\u7aef\u5230\u7aef\u8bc4\u4f30\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u7d22\u5f15\u65b9\u6cd5<\/th>\n<th>SoPR<\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<th>SoWR<\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td>I1<\/td>\n<td>I2<\/td>\n<td>I3<\/td>\n<td>Avg.<\/td>\n<td>I1<\/td>\n<td>I2<\/td>\n<td>I3<\/td>\n<td>Avg<\/td>\n<\/tr>\n<tr>\n<td>\u6570\u5b57\u7d22\u5f15<\/td>\n<td>34.76<\/td>\n<td>29.87<\/td>\n<td>46.99<\/td>\n<td>35.45<\/td>\n<td>25.77<\/td>\n<td>33.02<\/td>\n<td>29.51<\/td>\n<td>28.79<\/td>\n<\/tr>\n<tr>\n<td>\u5206\u5c42\u7d22\u5f15<\/td>\n<td>50.20<\/td>\n<td>45.60<\/td>\n<td>32.79<\/td>\n<td>45.50<\/td>\n<td>38.04<\/td>\n<td>43.40<\/td>\n<td>29.51<\/td>\n<td>38.18<\/td>\n<\/tr>\n<tr>\n<td>\u8bed\u4e49\u7d22\u5f15<\/td>\n<td>58.79<\/td>\n<td>45.28<\/td>\n<td>44.81<\/td>\n<td>51.87<\/td>\n<td>49.69<\/td>\n<td>57.55<\/td>\n<td>26.23<\/td>\n<td>47.88<\/td>\n<\/tr>\n<tr>\n<td>\u539f\u5b50\u7d22\u5f15<\/td>\n<td>58.08<\/td>\n<td>56.13<\/td>\n<td>44.81<\/td>\n<td>55.00<\/td>\n<td>47.85<\/td>\n<td>57.55<\/td>\n<td>29.51<\/td>\n<td>47.58<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h3>5.3 \u7d22\u5f15\u65b9\u6cd5\u5bf9\u6bd4<\/h3>\n<p>\u4e0e\u68c0\u7d22\u4efb\u52a1\u7684\u7d22\u5f15\u65b9\u6cd5\u5bf9\u6bd4\u76f8\u4f3c (\u7b2c\u00a04.3\u00a0\u8282 )\uff0c\u8868\u00a05\u00a0\u5c55\u793a\u4e86\u7aef\u5230\u7aef\u4ee3\u7406\u4efb\u52a1\u4e2d\u4e0d\u540c\u7d22\u5f15\u65b9\u6cd5\u7684\u5bf9\u6bd4\u3002\u5728\u6b64\u8bbe\u7f6e\u4e2d\uff0c\u53bb\u9664\u4e86\u53d7\u9650\u89e3\u7801\uff0c\u5141\u8bb8\u4ee3\u7406\u81ea\u7531\u751f\u6210 Thought\u3001Action \u548c Parameters\u3002\u4ece\u7ed3\u679c\u6765\u770b\uff0c\u539f\u5b50\u7d22\u5f15\u65b9\u6cd5\u5728\u56db\u79cd\u7d22\u5f15\u65b9\u6cd5\u4e2d\u8868\u73b0\u6700\u4f73\u3002\u6211\u4eec\u5c06\u5176\u5f52\u56e0\u4e3a\u5176\u4ed6\u65b9\u6cd5\u8f83\u9ad8\u7684\u5e7b\u89c9\u7387\uff0c\u8fd9\u5728\u7b2c\u00a05.4\u00a0\u8282\u4e2d\u8ba8\u8bba\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-7741\" title=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-4\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/bf947d200d7b021.png\" alt=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-4\" width=\"1660\" height=\"1070\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/bf947d200d7b021.png 1660w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/bf947d200d7b021-300x193.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/bf947d200d7b021-1024x660.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/bf947d200d7b021-768x495.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/bf947d200d7b021-1536x990.png 1536w\" sizes=\"auto, (max-width: 1660px) 100vw, 1660px\" \/><\/p>\n<p>\u56fe 4\uff1a\u5c55\u793a\u4e86\u4e0d\u540c\u6a21\u578b\u5728\u751f\u6210\u4e0d\u5b58\u5728\u7684\u5de5\u5177\u65f6\u7684\u5e7b\u89c9\u7387\u3002\u5728\u4f7f\u7528\u53d7\u9650\u89e3\u7801\u65f6\uff0cToolGen \u4e0d\u4f1a\u751f\u6210\u4e0d\u5b58\u5728\u7684\u5de5\u5177\u3002\u7136\u800c\uff0c\u5728\u6ca1\u6709\u6b64\u9650\u5236\u7684\u60c5\u51b5\u4e0b\uff0cToolGen \u5728\u4f7f\u7528\u539f\u5b50\u7d22\u5f15\u65f6\u7684 Action \u751f\u6210\u9636\u6bb5\u4f1a\u751f\u6210 7% \u7684\u975e\u5de5\u5177 Token \uff0c\u4f7f\u7528\u8bed\u4e49\u7d22\u5f15\u65f6\u5e7b\u89c9\u7387\u66f4\u9ad8\u3002\u5bf9\u4e8e ToolLlama \u548c GPT-3.5\uff0c\u5373\u4f7f\u5728\u63d0\u793a\u4e2d\u63d0\u4f9b\u4e86\u4e94\u4e2a\u771f\u5b9e\u5de5\u5177\uff0c\u5e7b\u89c9\u4ecd\u7136\u4f1a\u51fa\u73b0\u3002\u82e5\u63d0\u793a\u4e2d\u672a\u6307\u5b9a\u4efb\u4f55\u5de5\u5177\uff0cToolLlama \u4f1a\u751f\u6210\u8d85\u8fc7 50% \u7684\u4e0d\u5b58\u5728\u5de5\u5177\u540d\u79f0\u3002<\/p>\n<h3>5.4 \u5e7b\u89c9<\/h3>\n<p>\u6211\u4eec\u5728\u7aef\u5230\u7aef\u4ee3\u7406\u573a\u666f\u4e0b\u8bc4\u4f30\u6a21\u578b\u5728\u5de5\u5177\u751f\u6210\u4e2d\u7684\u5e7b\u89c9\u60c5\u51b5\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u8f93\u5165\u4e0e\u6a21\u578b\u8bad\u7ec3\u65f6\u683c\u5f0f\u4e00\u81f4\u7684\u67e5\u8be2\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5bf9\u4e8e ToolGen\uff0c\u6211\u4eec\u76f4\u63a5\u8f93\u5165\u67e5\u8be2\u5e76\u63d0\u793a\u6a21\u578b\u6309\u7167 ToolGen \u4ee3\u7406\u8303\u5f0f\u751f\u6210\u54cd\u5e94 ( \u5373\u6309\u987a\u5e8f\u751f\u6210 Thought\u3001Tool \u548c Parameters )\u3002\u6211\u4eec\u6d4b\u8bd5\u4e86\u4e0d\u4f7f\u7528\u7b2c\u00a03.6\u00a0\u8282\u4e2d\u63cf\u8ff0\u7684\u675f\u641c\u7d22\u9650\u5236\u7684 Action \u89e3\u7801\u3002\u5bf9\u4e8e ToolLlama \u548c GPT-3.5\uff0c\u6211\u4eec\u8f93\u5165\u67e5\u8be2\u5e76\u5305\u542b 5 \u4e2a\u771f\u5b9e\u5de5\u5177\u3002\u5728\u6240\u6709\u8bbe\u7f6e\u4e2d\uff0c\u6211\u4eec\u62a5\u544a\u5728\u6240\u6709\u5de5\u5177\u751f\u6210\u52a8\u4f5c\u4e2d\u751f\u6210\u7684\u5de5\u5177\u4e2d\u4e0d\u5b58\u5728\u4e8e\u6570\u636e\u96c6\u7684\u6bd4\u4f8b\u3002\u56fe\u00a04\u00a0\u5c55\u793a\u4e86\u4e0d\u540c\u6a21\u578b\u751f\u6210\u4e0d\u5b58\u5728\u5de5\u5177\u7684\u5e7b\u89c9\u7387\u3002\u4ece\u56fe\u4e2d\u53ef\u4ee5\u770b\u51fa\uff0c\u5c3d\u7ba1\u53ea\u63d0\u4f9b\u4e86\u4e94\u4e2a\u771f\u5b9e\u5de5\u5177\uff0cToolLlama \u548c GPT-3.5 \u4ecd\u7136\u53ef\u80fd\u751f\u6210\u4e0d\u5b58\u5728\u7684\u5de5\u5177\u540d\u79f0\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cToolGen \u901a\u8fc7\u53d7\u9650\u89e3\u7801\u8bbe\u8ba1\uff0c\u5b8c\u5168\u907f\u514d\u4e86\u5e7b\u89c9\u7684\u53d1\u751f\u3002<\/p>\n<h2>6 \u7ed3\u8bba<\/h2>\n<p>\u5728\u672c\u6587\u4e2d\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86 ToolGen\uff0c\u8fd9\u662f\u4e00\u79cd\u901a\u8fc7\u5c06\u7279\u5b9a\u5de5\u5177\u7684\u865a\u62df Token \u5d4c\u5165\u6a21\u578b\u8bcd\u6c47\u8868\u6765\u7edf\u4e00\u5de5\u5177\u68c0\u7d22\u548c\u6267\u884c\u7684\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u6846\u67b6\uff0c\u4ece\u800c\u5c06\u5de5\u5177\u4ea4\u4e92\u8f6c\u5316\u4e3a\u4e00\u4e2a\u751f\u6210\u5f0f\u4efb\u52a1\u3002\u901a\u8fc7\u5305\u542b\u4e00\u4e2a\u4e09\u9636\u6bb5\u8bad\u7ec3\u6d41\u7a0b\uff0cToolGen \u8d4b\u4e88 LLM \u5728\u771f\u5b9e\u573a\u666f\u4e2d\u9ad8\u6548\u68c0\u7d22\u548c\u6267\u884c\u5de5\u5177\u7684\u80fd\u529b\u3002\u8fd9\u79cd\u7edf\u4e00\u65b9\u6cd5\u4e3a\u5177\u5907\u6269\u5c55\u6027\u548c\u9ad8\u6548\u6027\u7684 AI \u4ee3\u7406\u8bbe\u7acb\u4e86\u65b0\u6807\u6746\uff0c\u4f7f\u5176\u80fd\u591f\u5904\u7406\u5e9e\u5927\u7684\u5de5\u5177\u5e93\u3002\u5c55\u671b\u672a\u6765\uff0cToolGen \u4e3a\u6574\u5408\u94fe\u5f0f\u601d\u7ef4\uff08chain-of-thought reasoning\uff09\u3001\u5f3a\u5316\u5b66\u4e60\uff08reinforcement learning\uff09\u548c ReAct \u7b49\u5148\u8fdb\u6280\u672f\u6253\u5f00\u4e86\u5927\u95e8\uff0c\u8fdb\u4e00\u6b65\u63d0\u5347 LLM \u5728\u771f\u5b9e\u5e94\u7528\u4e2d\u7684\u81ea\u4e3b\u6027\u548c\u591a\u529f\u80fd\u6027\u3002<\/p>\n<h2>References<\/h2>\n<ul>\n<li>Asai et al. (2023)\u2191Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi.Self-rag: Learning to retrieve, generate, and critique through self-reflection, 2023.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2310.11511\">https:\/\/arxiv.org\/abs\/2310.11511<\/a>.<\/li>\n<li>Brohan et al. (2023)\u2191Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, Pete Florence, Chuyuan Fu, Montse Gonzalez Arenas, Keerthana Gopalakrishnan, Kehang Han, Karol Hausman, Alexander Herzog, Jasmine Hsu, Brian Ichter, Alex Irpan, Nikhil Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Lisa Lee, Tsang-Wei Edward Lee, Sergey Levine, Yao Lu, Henryk Michalewski, Igor Mordatch, Karl Pertsch, Kanishka Rao, Krista Reymann, Michael Ryoo, Grecia Salazar, Pannag Sanketi, Pierre Sermanet, Jaspiar Singh, Anikait Singh, Radu Soricut, Huong Tran, Vincent Vanhoucke, Quan Vuong, Ayzaan Wahid, Stefan Welker, Paul Wohlhart, Jialin Wu, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich.Rt-2: Vision-language-action models transfer web knowledge to robotic control, 2023.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2307.15818\">https:\/\/arxiv.org\/abs\/2307.15818<\/a>.<\/li>\n<li>Brown et al. (2020)\u2191Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei.Language models are few-shot learners.In Hugo Larochelle, Marc\u2019Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin (eds.),\u00a0<em>Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual<\/em>, 2020.URL\u00a0<a href=\"https:\/\/proceedings.neurips.cc\/paper\/2020\/hash\/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html\">https:\/\/proceedings.neurips.cc\/paper\/2020\/hash\/1457c0d6bfcb4967418bfb8ac142f64a-Abstract.html<\/a>.<\/li>\n<li>Chen et al. (2023a)\u2191Baian Chen, Chang Shu, Ehsan Shareghi, Nigel Collier, Karthik Narasimhan, and Shunyu Yao.Fireact: Toward language agent fine-tuning.<em>arXiv preprint arXiv:2310.05915<\/em>, 2023a.<\/li>\n<li>Chen et al. (2023b)\u2191Jiangui Chen, Ruqing Zhang, Jiafeng Guo, Maarten de Rijke, Wei Chen, Yixing Fan, and Xueqi Cheng.Continual Learning for Generative <a href=\"https:\/\/www.kdjingpai.com\/en\/retrieval\/\">Retrieval<\/a> over Dynamic Corpora.In\u00a0<em>Proceedings of the 32nd ACM International Conference on Information and Knowledge Management<\/em>, CIKM \u201923, pp. 306\u2013315, New York, NY, USA, 2023b. Association for Computing Machinery.ISBN 9798400701245.doi: 10.1145\/3583780.3614821.URL\u00a0<a href=\"https:\/\/dl.acm.org\/doi\/10.1145\/3583780.3614821\">https:\/\/dl.acm.org\/doi\/10.1145\/3583780.3614821<\/a>.<\/li>\n<li>Chen et al. (2024a)\u2191Junzhi Chen, Juhao Liang, and Benyou Wang.Smurfs: Leveraging multiple proficiency agents with context-efficiency for tool planning, 2024a.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2405.05955\">https:\/\/arxiv.org\/abs\/2405.05955<\/a>.<\/li>\n<li>Chen et al. (2024b)\u2191Yanfei Chen, Jinsung Yoon, Devendra Singh Sachan, Qingze Wang, Vincent Cohen-Addad, Mohammadhossein Bateni, Chen-Yu Lee, and Tomas Pfister.Re-invoke: Tool invocation rewriting for zero-shot tool retrieval.<em>arXiv preprint arXiv:2408.01875<\/em>, 2024b.<\/li>\n<li>Chen et al. (2024c)\u2191Zehui Chen, Kuikun Liu, Qiuchen Wang, Wenwei Zhang, Jiangning Liu, Dahua Lin, Kai Chen, and Feng Zhao.Agent-flan: Designing data and methods of effective agent tuning for large language models, 2024c.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2403.12881\">https:\/\/arxiv.org\/abs\/2403.12881<\/a>.<\/li>\n<li>Dao (2024)\u2191Tri Dao.FlashAttention-2: Faster attention with better parallelism and work partitioning.In\u00a0<em>International Conference on Learning Representations (ICLR)<\/em>, 2024.<\/li>\n<li>Dao et al. (2022)\u2191Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher R\u00e9.FlashAttention: Fast and memory-efficient exact attention with IO-awareness.In\u00a0<em>Advances in Neural Information Processing Systems (NeurIPS)<\/em>, 2022.<\/li>\n<li>Dubey et al. (2024)\u2191Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al.The llama 3 herd of models.<em>arXiv preprint arXiv:2407.21783<\/em>, 2024.<\/li>\n<li>Gravitas (2023)\u2191Gravitas.AutoGPT, 2023.URL\u00a0<a href=\"https:\/\/github.com\/Significant-Gravitas\/AutoGPT\">https:\/\/github.com\/Significant-Gravitas\/AutoGPT<\/a>.<\/li>\n<li>Guo et al. (2024)\u2191Zhicheng Guo, Sijie Cheng, Hao Wang, Shihao Liang, Yujia Qin, Peng Li, Zhiyuan Liu, Maosong Sun, and Yang Liu.StableToolBench: Towards Stable Large-Scale Benchmarking on Tool Learning of Large Language Models, 2024.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2403.07714\">https:\/\/arxiv.org\/abs\/2403.07714<\/a>.<\/li>\n<li>Hao et al. (2023)\u2191Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu.Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings.In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine (eds.),\u00a0<em>Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 &#8211; 16, 2023<\/em>, 2023.URL\u00a0<a href=\"http:\/\/papers.nips.cc\/paper_files\/paper\/2023\/hash\/8fd1a81c882cd45f64958da6284f4a3f-Abstract-Conference.html\">http:\/\/papers.nips.cc\/paper_files\/paper\/2023\/hash\/8fd1a81c882cd45f64958da6284f4a3f-Abstract-Conference.html<\/a>.<\/li>\n<li>J\u00e4rvelin &amp; Kek\u00e4l\u00e4inen (2002)\u2191Kalervo J\u00e4rvelin and Jaana Kek\u00e4l\u00e4inen.Cumulated gain-based evaluation of ir techniques.<em>ACM Transactions on Information Systems (TOIS)<\/em>, 20(4):422\u2013446, 2002.<\/li>\n<li>Karpukhin et al. (2020)\u2191Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.Dense passage retrieval for open-domain question answering.In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu (eds.),\u00a0<em>Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)<\/em>, pp. 6769\u20136781, Online, 2020. Association for Computational Linguistics.doi: 10.18653\/v1\/2020.emnlp-main.550.URL\u00a0<a href=\"https:\/\/aclanthology.org\/2020.emnlp-main.550\">https:\/\/aclanthology.org\/2020.emnlp-main.550<\/a>.<\/li>\n<li>Kishore et al. (2023)\u2191Varsha Kishore, Chao Wan, Justin Lovelace, Yoav Artzi, and Kilian Q. Weinberger.Incdsi: Incrementally updatable document retrieval.In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett (eds.),\u00a0<em>International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA<\/em>, volume 202 of\u00a0<em>Proceedings of Machine Learning Research<\/em>, pp. 17122\u201317134. PMLR, 2023.URL\u00a0<a href=\"https:\/\/proceedings.mlr.press\/v202\/kishore23a.html\">https:\/\/proceedings.mlr.press\/v202\/kishore23a.html<\/a>.<\/li>\n<li>Liu et al. (2024)\u2191Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al.Agentbench: Evaluating llms as agents.In\u00a0<em>The Twelfth International Conference on Learning Representations<\/em>, 2024.<\/li>\n<li>Liu et al. (2023)\u2191Zhiwei Liu, Weiran Yao, Jianguo Zhang, Le Xue, Shelby Heinecke, Rithesh Murthy, Yihao Feng, Zeyuan Chen, Juan Carlos Niebles, Devansh Arpit, et al.Bolaa: Benchmarking and orchestrating llm-augmented autonomous agents.<em>arXiv preprint arXiv:2308.05960<\/em>, 2023.<\/li>\n<li>Mehta et al. (2023)\u2191Sanket Vaibhav Mehta, Jai Gupta, Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Jinfeng Rao, Marc Najork, Emma Strubell, and Donald Metzler.DSI++: Updating <a href=\"https:\/\/www.kdjingpai.com\/en\/transformer\/\">transformer<\/a> memory with new documents.In Houda Bouamor, Juan Pino, and Kalika Bali (eds.),\u00a0<em>Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing<\/em>, pp. 8198\u20138213, Singapore, 2023. Association for Computational Linguistics.doi: 10.18653\/v1\/2023.emnlp-main.510.URL\u00a0<a href=\"https:\/\/aclanthology.org\/2023.emnlp-main.510\">https:\/\/aclanthology.org\/2023.emnlp-main.510<\/a>.<\/li>\n<li>OpenAI (2024)\u2191OpenAI.Gpt-4 technical report, 2024.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2303.08774\">https:\/\/arxiv.org\/abs\/2303.08774<\/a>.<\/li>\n<li>Patil et al. (2023)\u2191Shishir G. Patil, Tianjun Zhang, Xin Wang, and Joseph E. Gonzalez.Gorilla: Large language model connected with massive apis, 2023.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2305.15334\">https:\/\/arxiv.org\/abs\/2305.15334<\/a>.<\/li>\n<li>Qiao et al. (2024)\u2191Shuofei Qiao, Ningyu Zhang, Runnan Fang, Yujie Luo, Wangchunshu Zhou, Yuchen Eleanor Jiang, Chengfei Lv, and Huajun Chen.Autoact: Automatic agent learning from scratch for qa via self-planning, 2024.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2401.05268\">https:\/\/arxiv.org\/abs\/2401.05268<\/a>.<\/li>\n<li>Qin et al. (2023)\u2191Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Lauren Hong, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun.ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs, 2023.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2307.16789\">https:\/\/arxiv.org\/abs\/2307.16789<\/a>.<\/li>\n<li>Qu et al. (2024)\u2191Changle Qu, Sunhao Dai, Xiaochi Wei, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, Jun Xu, and Ji-Rong Wen.Tool learning with large language models: A survey.<em>arXiv preprint arXiv:2405.17935<\/em>, 2024.<\/li>\n<li>Rajbhandari et al. (2020)\u2191Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He.Zero: Memory optimizations toward training trillion parameter models, 2020.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/1910.02054\">https:\/\/arxiv.org\/abs\/1910.02054<\/a>.<\/li>\n<li>Robertson et al. (2009)\u2191Stephen Robertson, Hugo Zaragoza, et al.The probabilistic <a href=\"https:\/\/www.kdjingpai.com\/en\/relevance-ai\/\">relevance<\/a> framework: Bm25 and beyond.<em>Foundations and Trends\u00ae in Information Retrieval<\/em>, 3(4):333\u2013389, 2009.<\/li>\n<li>Schick et al. (2023)\u2191Timo Schick, Jane Dwivedi-Yu, Roberto Dess\u00ec, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom.Toolformer: Language models can teach themselves to use tools, 2023.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2302.04761\">https:\/\/arxiv.org\/abs\/2302.04761<\/a>.<\/li>\n<li>Shen et al. (2024)\u2191Weizhou Shen, Chenliang Li, Hongzhan Chen, Ming Yan, Xiaojun Quan, Hehong Chen, Ji Zhang, and Fei Huang.Small llms are weak tool learners: A multi-llm agent, 2024.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2401.07324\">https:\/\/arxiv.org\/abs\/2401.07324<\/a>.<\/li>\n<li>Shinn et al. (2023)\u2191Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao.Reflexion: language agents with verbal reinforcement learning.In\u00a0<em>Proceedings of the 37th International Conference on Neural Information Processing Systems<\/em>, pp. 8634\u20138652, 2023.<\/li>\n<li>Sun et al. (2023)\u2191Weiwei Sun, Lingyong Yan, Zheng Chen, Shuaiqiang Wang, Haichao Zhu, Pengjie Ren, Zhumin Chen, Dawei Yin, Maarten de Rijke, and Zhaochun Ren.Learning to tokenize for generative retrieval.In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine (eds.),\u00a0<em>Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 &#8211; 16, 2023<\/em>, 2023.URL\u00a0<a href=\"http:\/\/papers.nips.cc\/paper_files\/paper\/2023\/hash\/91228b942a4528cdae031c1b68b127e8-Abstract-Conference.html\">http:\/\/papers.nips.cc\/paper_files\/paper\/2023\/hash\/91228b942a4528cdae031c1b68b127e8-Abstract-Conference.html<\/a>.<\/li>\n<li>Touvron et al. (2023)\u2191Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth\u00e9e Lacroix, Baptiste Rozi\u00e8re, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample.Llama: Open and efficient foundation language models, 2023.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2302.13971\">https:\/\/arxiv.org\/abs\/2302.13971<\/a>.<\/li>\n<li>Wang et al. (2024)\u2191Renxi Wang, Haonan Li, Xudong Han, Yixuan Zhang, and Timothy Baldwin.Learning from failure: Integrating negative examples when fine-tuning large language models as agents.<em>arXiv preprint arXiv:2402.11651<\/em>, 2024.<\/li>\n<li>Wang et al. (2022)\u2191Yujing Wang, Yingyan Hou, Haonan Wang, Ziming Miao, Shibin Wu, Qi Chen, Yuqing Xia, Chengmin Chi, Guoshuai Zhao, Zheng Liu, Xing Xie, Hao Sun, Weiwei Deng, Qi Zhang, and Mao Yang.A neural corpus indexer for document retrieval.In Sanmi Koyejo, S. Mohamed, A. Agarwal, Danielle Belgrave, K. Cho, and A. Oh (eds.),\u00a0<em>Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 &#8211; December 9, 2022<\/em>, 2022.URL\u00a0<a href=\"http:\/\/papers.nips.cc\/paper_files\/paper\/2022\/hash\/a46156bd3579c3b268108ea6aca71d13-Abstract-Conference.html\">http:\/\/papers.nips.cc\/paper_files\/paper\/2022\/hash\/a46156bd3579c3b268108ea6aca71d13-Abstract-Conference.html<\/a>.<\/li>\n<li>Wei et al. (2023)\u2191Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou.Chain-of-thought prompting elicits reasoning in large language models, 2023.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2201.11903\">https:\/\/arxiv.org\/abs\/2201.11903<\/a>.<\/li>\n<li>Wu et al. (2024a)\u2191Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang.Autogen: Enabling next-gen llm applications via multi-agent conversation framework.In\u00a0<em>COLM<\/em>, 2024a.<\/li>\n<li>Wu et al. (2024b)\u2191Qinzhuo Wu, Wei Liu, Jian Luan, and Bin Wang.ToolPlanner: A Tool Augmented LLM for Multi Granularity Instructions with Path Planning and Feedback, 2024b.URL\u00a0<a href=\"https:\/\/arxiv.org\/abs\/2409.14826\">https:\/\/arxiv.org\/abs\/2409.14826<\/a>.<\/li>\n<li>Xiong et al. (2021)\u2191Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul N. Bennett, Junaid Ahmed, and Arnold Overwijk.Approximate nearest neighbor negative contrastive learning for dense text retrieval.In\u00a0<em>9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021<\/em>. OpenReview.net, 2021.URL\u00a0<a href=\"https:\/\/openreview.net\/forum?id=zeFrfgyZln\">https:\/\/openreview.net\/forum?id=zeFrfgyZln<\/a>.<\/li>\n<li>Xu et al. (2024)\u2191Qiancheng Xu, Yongqi Li, Heming Xia, and Wenjie Li.Enhancing tool retrieval with iterative feedback from large language models.<em>arXiv preprint arXiv:2406.17465<\/em>, 2024.<\/li>\n<li>Yao et al. (2023)\u2191Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao.ReAct: Synergizing reasoning and acting in language models.In\u00a0<em>International Conference on Learning Representations (ICLR)<\/em>, 2023.<\/li>\n<li>Yin et al. (2024)\u2191Da Yin, Faeze Brahman, Abhilasha Ravichander, Khyathi Chandu, Kai-Wei Chang, Yejin Choi, and Bill Yuchen Lin.Agent lumos: Unified and modular training for open-source language agents.In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.),\u00a0<em>Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)<\/em>, pp. 12380\u201312403, Bangkok, Thailand, August 2024. Association for Computational Linguistics.doi: 10.18653\/v1\/2024.acl-long.670.URL\u00a0<a href=\"https:\/\/aclanthology.org\/2024.acl-long.670\">https:\/\/aclanthology.org\/2024.acl-long.670<\/a>.<\/li>\n<li>Zeng et al. (2023)\u2191Aohan Zeng, Mingdao Liu, Rui Lu, Bowen Wang, Xiao Liu, Yuxiao Dong, and Jie Tang.Agenttuning: Enabling generalized agent abilities for llms, 2023.<\/li>\n<\/ul>\n<h2>\u4e00\u4e2a\u771f\u5b9e\u5de5\u5177\u7684\u793a\u4f8b<\/h2>\n<p>\u56fe\u00a0<a title=\"Figure 5 \u2023 Appendix A Real Tool Example \u2023 ToolGen: Unified Tool Retrieval and Calling via Generation\" href=\"https:\/\/arxiv.org\/html\/2410.03439v2#A1.F5\">5<\/a>\u00a0\u5c55\u793a\u4e86\u4e00\u4e2a\u771f\u5b9e\u5de5\u5177\u7684\u793a\u4f8b\u3002\u6bcf\u4e2a\u5de5\u5177\u5305\u542b\u591a\u4e2a API\u3002\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\u4f7f\u7528\u4e86\u4ee5\u4e0b\u5b57\u6bb5\uff1a\u201ctool_name\u201d \u662f\u5de5\u5177\u7684\u540d\u79f0\u3002\u201ctool_description\u201d \u63cf\u8ff0\u5de5\u5177\u76f8\u5173\u7684\u4fe1\u606f\uff0c\u4f8b\u5982\u5de5\u5177\u7684\u529f\u80fd\u3002\u5728\u6bcf\u4e2a API \u4e2d\uff0c\u201cname\u201d \u662f API \u7684\u540d\u79f0\u3002\u201cdescription\u201d \u63cf\u8ff0 API \u76f8\u5173\u7684\u4fe1\u606f\u3002\u201cmethod\u201d \u662f\u8c03\u7528 API \u7684 http \u65b9\u6cd5\u3002\u201crequired_parameters\u201d \u662f\u8c03\u7528 API \u65f6\u5fc5\u987b\u586b\u5199\u7684\u53c2\u6570\u3002\u201coptional_parameters\u201d \u53ef\u7528\u4e8e\u8bbe\u7f6e\u989d\u5916\u53c2\u6570\uff08\u53ef\u9009\uff09\u3002<\/p>\n<pre><code>{\r\n\"tool_name\":\"YouTube Hub\",\r\n\"tool_description\":\"\u83b7\u53d6\u5355\u4e2a\u89c6\u9891\u7684\u70b9\u8d5e\u6570\u3001\u89c2\u770b\u6b21\u6570\u3001\u6807\u9898\u3001\u7f29\u7565\u56fe\u7b49\u8be6\u7ec6\u4fe1\u606f\u3002\",\r\n\"home_url\":\"https:\/\/rapidapi.com\/itsrohitofficial-XBPdXttOUQ\/api\/youtube-hub\/\",\r\n\"host\":\"youtube-hub.p.rapidapi.com\",\r\n\"api_list\":[\r\n{\r\n\"name\":\"\u83b7\u53d6\u89c6\u9891\u8be6\u60c5\",\r\n\"url\":\"https:\/\/youtube-hub.p.rapidapi.com\/\",\r\n\"description\":\"\u83b7\u53d6\u89c6\u9891\u7684\u6240\u6709\u57fa\u672c\u4fe1\u606f\",\r\n\"method\":\"GET\",\r\n\"required_parameters\":[\r\n{\r\n\"name\":\"id\",\r\n\"type\":\"STRING\",\r\n\"description\":\"\",\r\n\"default\":\"fD6SzYIRr4c\"\r\n}\r\n],\r\n\"optional_parameters\":[],\r\n}\r\n]\r\n}\r\n<\/code><\/pre>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-7740\" title=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-5\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/f4fe292eb01627a.png\" alt=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-5\" width=\"1660\" height=\"904\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/f4fe292eb01627a.png 1660w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/f4fe292eb01627a-300x163.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/f4fe292eb01627a-1024x558.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/f4fe292eb01627a-768x418.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/f4fe292eb01627a-1536x836.png 1536w\" sizes=\"auto, (max-width: 1660px) 100vw, 1660px\" \/><\/p>\n<p>\u56fe 5: \u4e00\u4e2a\u771f\u5b9e\u5de5\u5177\u7684\u793a\u4f8b\u3002\u8be5\u5de5\u5177\u5305\u542b\u4e00\u4e2a API\u3002\u4e3a\u7b80\u6d01\u8d77\u89c1\uff0c\u6211\u4eec\u5df2\u5220\u9664\u4e86\u4e0d\u5fc5\u8981\u7684\u5b57\u6bb5\u3002<\/p>\n<h2>B \u5de5\u5177\u865a\u62df\u5316\u5b9e\u73b0<\/h2>\n<p>ToolGen \u91c7\u7528\u5355\u4e00\u72ec\u7279\u7684 Token \u6765\u8868\u793a\u4e00\u4e2a\u5de5\u5177\uff0c\u8fd9\u663e\u793a\u4e86\u5176\u5728\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528\u65b9\u9762\u7684\u4f18\u52bf\u3002\u6211\u4eec\u8fd8\u5f15\u5165\u4e86\u5176\u4ed6\u65b9\u6cd5\u6765\u5bf9\u5de5\u5177\u8fdb\u884c\u7d22\u5f15\uff0c\u5305\u62ec\u8bed\u4e49\u3001\u6570\u5b57\u548c\u5c42\u6b21\u5316\u3002\u4ee5\u4e0b\u662f\u6211\u4eec\u5982\u4f55\u5b9e\u73b0\u6bcf\u79cd\u7d22\u5f15\u7684\u8be6\u7ec6\u8bf4\u660e\u3002<\/p>\n<h4>\u539f\u5b50<\/h4>\n<p>\u7d22\u5f15\u662f\u6211\u4eec\u5728 ToolGen \u4e2d\u4f7f\u7528\u7684\u65b9\u6cd5\u3002\u4e0e\u5176\u4ed6\u65b9\u6cd5\u76f8\u6bd4\uff0c\u5b83\u5c06\u4e00\u4e2a\u5355\u72ec\u7684 Token \u4f5c\u4e3a\u5de5\u5177\uff0c\u5e76\u4e14\u4e0d\u4f1a\u865a\u6784\u4e0d\u5b58\u5728\u7684\u5de5\u5177\u3002\u6211\u4eec\u4f7f\u7528 &lt;&lt;tool name&amp;&amp;api name&gt;&gt; \u5c06\u5de5\u5177\u540d\u79f0\u548c API \u540d\u79f0\u7ed3\u5408\u8d77\u6765\u5f62\u6210\u4e00\u4e2a\u5355\u4e00\u7684 Token\u3002\u4f8b\u5982\uff0c\u5728\u9644\u5f55\u00a0<a title=\"\u9644\u5f55 A \u771f\u5b9e\u5de5\u5177\u793a\u4f8b \u2023 ToolGen: \u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u4e0e\u8c03\u7528\" href=\"https:\/\/arxiv.org\/html\/2410.03439v2#A1\">A<\/a>\u00a0\u4e2d\u7684\u793a\u4f8b\uff0c\u7ed3\u679c Token \u4e3a &lt;&lt;Youtube Hub&amp;&amp;Get Video Details&gt;&gt;\u3002<\/p>\n<h4>\u8bed\u4e49<\/h4>\n<p>\u7d22\u5f15\u5c06\u6bcf\u4e2a\u5de5\u5177\u6620\u5c04\u5230\u5728 ToolBench \u4e2d\u4f7f\u7528\u7684\u540d\u79f0\uff0c\u8fd9\u4e5f\u662f\u5de5\u5177\u540d\u79f0\u548c API \u540d\u79f0\u7684\u7ec4\u5408\u3002\u7136\u800c\uff0c\u8be5\u540d\u79f0\u53ef\u4ee5\u88ab\u5206\u89e3\u4e3a\u591a\u4e2a Token\uff0c\u4ee5\u4fbf\u6a21\u578b\u53ef\u4ee5\u611f\u77e5\u5176\u8bed\u4e49\u610f\u4e49\u3002\u5728\u9644\u5f55\u00a0A\u00a0\u4e2d\u7684\u793a\u4f8b\uff0c\u5f97\u5230\u7684\u6620\u5c04\u4e3a get_video_details_for_youtube_hub\u3002<\/p>\n<h4>\u6570\u5b57<\/h4>\n<p>\u7d22\u5f15\u5c06\u6bcf\u4e2a\u5de5\u5177\u6620\u5c04\u5230\u4e00\u4e2a\u552f\u4e00\u7684\u6570\u5b57\u3002\u6211\u4eec\u9996\u5148\u83b7\u53d6\u6240\u6709\u5de5\u5177\u7684\u5217\u8868\uff0c\u957f\u5ea6\u7ea6\u4e3a 47,000\u3002\u5bf9\u4e8e\u6240\u6709\u5de5\u5177\uff0c\u6211\u4eec\u4f7f\u7528\u4e00\u4e2a\u4e94\u4f4d\u6570\u5b57\uff0c\u7531\u7a7a\u683c\u5206\u9694\u6765\u8868\u793a\u8be5\u5de5\u5177\u3002\u5982\u679c\u9644\u5f55\u00a0A\u00a0\u4e2d\u7684\u793a\u4f8b\u662f\u5217\u8868\u4e2d\u7684\u7b2c 128 \u4e2a\u5143\u7d20\uff0c\u6211\u4eec\u4f7f\u7528 0 0 0 1 2 8 \u6765\u8868\u793a\u8be5\u5de5\u5177\u3002\u7531\u4e8e Llama-3 \u5206\u8bcd\u5668\u5bf9\u6bcf\u4e2a\u6570\u5b57\u8fdb\u884c\u7f16\u7801<\/p>\n<h4>\u5c42\u6b21\u5316<\/h4>\n<p>\u4e5f\u5c06\u6bcf\u4e2a\u5de5\u5177\u6620\u5c04\u5230\u4e00\u4e2a\u6570\u5b57\u3002\u4e0e\u6570\u5b57\u7d22\u5f15\u4e0d\u540c\uff0c\u6211\u4eec\u901a\u8fc7\u8fed\u4ee3\u805a\u7c7b\u5c06\u7ed3\u6784\u4fe1\u606f\u6ce8\u5165\u5de5\u5177\u8868\u793a\u3002\u5728\u6bcf\u6b21\u8fed\u4ee3\u4e2d\uff0c\u6211\u4eec\u5c06\u5de5\u5177\u805a\u7c7b\u6210\u5341\u4e2a\u96c6\u7fa4\uff0c\u6bcf\u4e2a\u96c6\u7fa4\u88ab\u5206\u914d\u4e00\u4e2a\u4ece 0 \u5230 9 \u7684\u6570\u5b57\u3002\u5bf9\u4e8e\u6bcf\u4e2a\u96c6\u7fa4\uff0c\u6211\u4eec\u91cd\u590d\u8fd9\u4e00\u805a\u7c7b\u8fc7\u7a0b\uff0c\u76f4\u5230\u8be5\u96c6\u7fa4\u4e2d\u53ea\u6709\u4e00\u4e2a\u5de5\u5177\u3002\u8fd9\u4e9b\u6b65\u9aa4\u5f62\u6210\u4e86\u4e00\u68f5\u805a\u7c7b\u6811\u3002\u6211\u4eec\u4ece\u6839\u5230\u53f6\u7684\u6570\u5b57\u4f5c\u4e3a\u8be5\u53f6\u5b50\u4e2d\u5de5\u5177\u7684\u8868\u793a\u3002\u9644\u5f55\u00a0A\u00a0\u4e2d\u7684\u793a\u4f8b\u53ef\u80fd\u88ab\u5206\u914d\u4e00\u4e2a\u8d85\u8fc7\u4e94\u4f4d\u6570\u5b57\u7684\u53f7\u7801\uff0c\u4f8b\u5982 0 1 2 2 3 3 3\u3002<\/p>\n<h2>C \u7ea6\u675f\u675f\u641c\u7d22\u8be6\u7ec6\u4fe1\u606f<\/h2>\n<p>\u5728\u68c0\u7d22\u548c\u5b8c\u6210\u7aef\u5230\u7aef\u4ee3\u7406\u4efb\u52a1\u65f6\uff0c\u6211\u4eec\u4f7f\u7528\u7ea6\u675f\u675f\u641c\u7d22\u6765\u9650\u5236\u751f\u6210\u7684\u64cd\u4f5c\u4e3a\u6709\u6548\u7684\u5de5\u5177 Token\u3002\u8be6\u7ec6\u6b65\u9aa4\u89c1\u7b97\u6cd5\u00a01\u3002\u57fa\u672c\u601d\u8def\u662f\u5728\u675f\u641c\u7d22\u6b65\u9aa4\u4e2d\u9650\u5236\u641c\u7d22\u7a7a\u95f4\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u9700\u8981\u9996\u5148\u6784\u5efa\u4e00\u4e2a\u4e0d\u76f8\u4ea4\u7684\u5b57\u5178\u6811\uff0c\u5176\u4e2d\u6bcf\u4e2a\u8282\u70b9\u8868\u793a\u4e00\u4e2a\u5de5\u5177 Token ID\u3002\u8be5\u8282\u70b9\u7684\u5b50\u8282\u70b9\u90fd\u662f\u8ddf\u968f\u5f53\u524d ID \u7684\u6240\u6709\u53ef\u884c ID\u3002\u5229\u7528\u8fd9\u68f5\u6811\uff0c\u6211\u4eec\u53ef\u4ee5\u6839\u636e\u5f53\u524d\u641c\u7d22\u7684 ID \u786e\u5b9a\u6240\u6709\u53ef\u80fd\u7684\u4e0b\u4e00\u4e2a Token ID\u3002\u5728\u675f\u641c\u7d22\u6b65\u9aa4\u4e2d\uff0c\u6211\u4eec\u5c4f\u853d\u6389\u6240\u6709\u5176\u4ed6\u4e0d\u53ef\u884c Token \u7684 logits\uff0c\u5f3a\u5236\u53ef\u80fd\u7684 ID \u88ab\u91c7\u6837\u6216\u641c\u7d22\u3002<\/p>\n<p>\u5bf9\u4e8e\u68c0\u7d22\uff0c\u8fd9\u53ef\u4ee5\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u76f4\u63a5\u5e94\u7528\u3002\u5bf9\u4e8e\u7aef\u5230\u7aef\u4ee3\u7406\u4efb\u52a1\uff0c\u7531\u4e8e\u6211\u4eec\u5c06\u63a8\u7406\u6b65\u9aa4\u5206\u89e3\u4e3a\u4e09\u4e2a\u5bf9\u8bdd\u8f6e\u6b21\uff0c\u6211\u4eec\u53ef\u4ee5\u8f7b\u677e\u68c0\u6d4b ToolGen \u4f55\u65f6\u9700\u8981\u751f\u6210\u64cd\u4f5c\uff0c\u56e0\u6b64\u5e94\u7528\u8be5\u7ea6\u675f\u3002\u56fe\u00a06\u00a0\u5c55\u793a\u4e86 ToolGen \u7684\u7aef\u5230\u7aef\u63a8\u7406\u793a\u4f8b\uff0c\u5176\u4e2d\u6ca1\u6709\u76f8\u5173\u5de5\u5177\u4f9b ToolGen \u9009\u62e9\u3002\u5b83\u53ef\u4ee5\u76f4\u63a5\u751f\u6210\u5de5\u5177 Token \u5e76\u5b8c\u6210\u4efb\u52a1\u3002<\/p>\n<p>\u8868 6: \u4e09\u9636\u6bb5\u8bad\u7ec3\u7684\u6570\u636e\u96c6\u7edf\u8ba1\u3002\u5bf9\u4e8e\u5de5\u5177\u8bb0\u5fc6\uff0c\u6709\u4e00\u4e9b\u91cd\u590d\u7684\u5de5\u5177\uff0c\u5bfc\u81f4\u6837\u672c\u6570\u91cf\u7565\u5927\u4e8e\u6211\u4eec\u4f7f\u7528\u7684\u5de5\u5177\u6570\u91cf\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6570\u636e\u96c6<\/th>\n<th>\u5de5\u5177\u8bb0\u5fc6<\/th>\n<th>\u68c0\u7d22\u8bad\u7ec3<\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<th>\u7aef\u5230\u7aef\u4ee3\u7406\u8c03\u4f18<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td><\/td>\n<td>I1<\/td>\n<td>I2<\/td>\n<td>I3<\/td>\n<td>\u5168\u90e8<\/td>\n<td><\/td>\n<\/tr>\n<tr>\n<td>#num<\/td>\n<td>49,936<\/td>\n<td>194,086<\/td>\n<td>222,783<\/td>\n<td>72,833<\/td>\n<td>489,702<\/td>\n<td>183,336<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<pre><code>\u7cfb\u7edf\uff1a\u60a8\u662f\u4e00\u4e2a AutoGPT\uff0c\u80fd\u591f\u5229\u7528\u4f17\u591a\u5de5\u5177\u548c\u529f\u80fd\u6765\u5b8c\u6210\u7ed9\u5b9a\u7684\u4efb\u52a1\u3002\r\n1. \u9996\u5148\uff0c\u6211\u4f1a\u7ed9\u60a8\u63d0\u4f9b\u4efb\u52a1\u63cf\u8ff0\uff0c\u60a8\u7684\u4efb\u52a1\u5c06\u5f00\u59cb\u3002\r\n2. \u5728\u6bcf\u4e00\u6b65\u4e2d\uff0c\u60a8\u9700\u8981\u901a\u8fc7\u751f\u6210\u4e00\u4e2a\u884c\u52a8\u4ee4\u724c\u6765\u786e\u5b9a\u4e0b\u4e00\u6b65\u884c\u52a8\u3002\r\n3. \u5728\u4ee4\u724c\u4e4b\u540e\uff0c\u60a8\u5c06\u6536\u5230\u4e0e\u8be5\u4ee4\u724c\u5bf9\u5e94\u7684\u884c\u52a8\u6587\u6863\u3002\u60a8\u9700\u8981\u751f\u6210\u8be5\u884c\u52a8\u7684\u8f93\u5165\uff0c\u5c06\u60a8\u8f6c\u79fb\u5230\u4e00\u4e2a\u65b0\u72b6\u6001\u3002\u968f\u540e\uff0c\u60a8\u5c06\u51b3\u5b9a\u4e0b\u4e00\u6b65\uff0c\u5e76\u91cd\u590d\u6b64\u8fc7\u7a0b\u3002\r\n4. \u5728\u751f\u6210\u591a\u4e2a\u884c\u52a8\u548c\u8f93\u5165\u7684\u51e0\u6b21\u8fed\u4ee3\u540e\uff0c\u60a8\u5c06\u6700\u7ec8\u5b8c\u6210\u4efb\u52a1\u5e76\u63d0\u4f9b\u60a8\u7684\u6700\u7ec8\u7b54\u6848\u3002\r\n... ...\r\n\uff1a\u6211\u6700\u8fd1\u8bbf\u95ee\u4e86 Viva City\uff0c\u83dc\u5355\u7ed9\u6211\u7559\u4e0b\u4e86\u6df1\u523b\u7684\u5370\u8c61\u3002\u6211\u60f3\u4e86\u89e3\u66f4\u591a\u5173\u4e8e\u573a\u9986 5a76c927bb5d660001e09731 \u7684\u83dc\u5355\u9879\u3002\u60a8\u80fd\u63d0\u4f9b\u83dc\u5355\u7c7b\u522b\u3001\u4ef7\u683c\u3001\u53ef\u7528\u6027\u3001\u89c4\u5219\u3001\u5957\u9910\u9879\u53ca\u5176\u4ed6\u76f8\u5173\u7ec6\u8282\u5417\uff1f\u8bf7\u4f7f\u7528\u8bed\u8a00 'zh-hans' \u6765\u63d0\u4f9b\u83dc\u5355\u8be6\u60c5\u3002\r\n\uff1a\u6211\u9700\u8981\u8c03\u7528 \"&lt;&lt;Viva City Documentation&amp;&amp;MenuDetails&gt;&gt;\" \u51fd\u6570\u4ee5\u68c0\u7d22\u573a\u9986 5a76c927bb5d660001e09731 \u7684\u83dc\u5355\u8be6\u60c5\u3002\u8c03\u7528\u6b64\u51fd\u6570\u7684\u539f\u56e0\u662f\u4e3a\u4e86\u5411\u7528\u6237\u63d0\u4f9b\u83dc\u5355\u7c7b\u522b\u3001\u4ef7\u683c\u3001\u53ef\u7528\u6027\u3001\u89c4\u5219\u3001\u5957\u9910\u9879\u53ca\u5176\u4ed6\u76f8\u5173\u7ec6\u8282\u3002\u8be5\u51fd\u6570\u7684\u8c03\u7528\u53c2\u6570\u4e3a \"lang\" \u8bbe\u7f6e\u4e3a \"zh-hans\" \u548c \"is_id\" \u8bbe\u7f6e\u4e3a \"5a76c927bb5d660001e09731\"\uff0c\u4ee5\u786e\u4fdd\u63d0\u4f9b\u7684\u83dc\u5355\u8be6\u60c5\u7b26\u5408\u6240\u9700\u8bed\u8a00\uff0c\u5e76\u4e14\u662f\u9488\u5bf9\u7279\u5b9a\u573a\u9986\u3002\r\n\uff1a\u751f\u6210\u884c\u52a8\u3002\r\n\uff1a&lt;&lt;Viva City Documentation&amp;&amp;MenuDetails&gt;&gt;\r\n\uff1a\u8bf7\u7ed9\u51fa\u8f93\u5165\u3002\u4ee5\u4e0b\u662f\u6587\u6863\uff1a{'name': '&lt;&lt;Viva City Documentation&amp;&amp;VenueDetails&gt;&gt;', 'description': '\u4f7f\u7528 KV \u83b7\u53d6\u573a\u9986\u8be6\u60c5\uff08\u9700\u8981\u91cd\u65b0\u540c\u6b65\u4ee5\u83b7\u53d6\u6700\u65b0\u6570\u636e\uff09\uff0c\u5e76\u6839\u636e\u67e5\u8be2\u53c2\u6570\u8fc7\u6ee4\u52a8\u6001\u591a\u8bed\u8a00\u6570\u636e\u3002', 'required': [{'name': 'id', 'type': 'STRING', 'description': '', 'default': '56c467cb7b0d464811a42017'}, {'name': 'lang', 'type': 'STRING', 'description': '', 'default': 'zh-hans'}], 'optional': []}\r\n\uff1a{\\n \"is_id\": \"5a76c927bb5d660001e09731\",\\n \"lang\": \"zh-hans\"\\n}\r\n\uff1a{'error': '', 'response': {'menu_id': '5a76c927bb5d660001e09731', 'language': 'zh-hans', 'menu_details': {'sections': [{'section_name': '\u5f00\u80c3\u83dc', 'items': [{'item_name': '\u6625\u5377', 'description': '\u8106\u76ae\u6625\u5377\uff0c\u642d\u914d\u7279\u5236\u9171\u6c41', 'price': 12.99}, {'item_name': '\u9e21\u8089\u4e32', 'description': '\u9999\u8fa3\u9e21\u8089\u4e32\uff0c\u5916\u7126\u91cc\u5ae9', 'price': 15.99}]}, {'section_name': '\u4e3b\u83dc', 'items': [{'item_name': '\u5bab\u4fdd\u9e21\u4e01', 'description': '\u7ecf\u5178\u5ddd\u83dc\uff0c\u9ebb\u8fa3\u9c9c\u9999', 'price': 18.99}, {'item_name': '\u6e05\u84b8\u9c88\u9c7c', 'description': '\u6e05\u65b0\u6de1\u96c5\uff0c\u8425\u517b\u5065\u5eb7', 'price': 22.99}]}], 'special_notes': '\u672c\u83dc\u5355\u63d0\u4f9b\u7eaf\u7d20\u9009\u9879\uff0c\u8bf7\u544a\u77e5\u670d\u52a1\u5458\u3002'}}}\r\n\uff1a\u573a\u9986 5a76c927bb5d660001e09731 \u7684\u83dc\u5355\u8be6\u60c5\u5982\u4e0b\uff1a\\n- \u83dc\u5355\u7c7b\u522b\uff1a\u4e2d\u9910\\n\u4ef7\u683c\uff1a\\n - \u9879\u76ee 1\uff1a12.99 \u4eba\u6c11\u5e01\\n - \u9879\u76ee 2\uff1a15.99 \u4eba\u6c11\u5e01\\n - \u9879\u76ee 3\uff1a18.99 \u4eba\u6c11\u5e01\\n - \u9879\u76ee 4\uff1a22.99 \u4eba\u6c11\u5e01\\n\u53ef\u7528\u6027\uff1a\\n - \u9879\u76ee 1\uff1a\u53ef\u7528\\n - \u9879\u76ee 2\uff1a\u53ef\u7528\\n - \u9879\u76ee 3\uff1a\u53ef\u7528\\n - \u9879\u76ee 4\uff1a\u53ef\u7528\\n- \u89c4\u5219\uff1a\\n - \u9879\u76ee 1\uff1a\u672a\u63d0\u53ca\u5177\u4f53\u89c4\u5219\\n - \u9879\u76ee 2\uff1a\u672a\u63d0\u53ca\u5177\u4f53\u89c4\u5219\\n - \u9879\u76ee 3\uff1a\u672a\u63d0\u53ca\u5177\u4f53\u89c4\u5219\\n - \u9879\u76ee 4\uff1a\u672a\u63d0\u53ca\u5177\u4f53\u89c4\u5219\\n- \u5957\u9910\u9879\uff1a\\n - \u9879\u76ee 1\uff1a2\u4efd\u5957\u9910\\n - \u9879\u76ee 2\uff1a2\u4efd\u5957\u9910\\n - \u9879\u76ee 3\uff1a2\u4efd\u5957\u9910\\n - \u9879\u76ee 4\uff1a2\u4efd\u5957\u9910\\n- \u5176\u4ed6\u76f8\u5173\u7ec6\u8282\uff1a\\n - \u7279\u522b\u8bf4\u660e\uff1a\u83dc\u5355\u53ef\u80fd\u4f1a\u66f4\u6539\uff0c\u4ef7\u683c\u4e5f\u53ef\u80fd\u4f1a\u53d8\u52a8\u3002\u8bf7\u5411\u573a\u9986\u786e\u8ba4\u6700\u65b0\u4fe1\u606f\u3002\\n\\n\u8bf7\u544a\u8bc9\u6211\u662f\u5426\u8fd8\u6709\u5176\u4ed6\u6211\u53ef\u4ee5\u5e2e\u52a9\u60a8\u7684\u5730\u65b9\uff01\r\n\uff1a\u751f\u6210\u884c\u52a8\u3002\r\n\uff1a&lt;&lt;Finish&gt;&gt;\r\n\uff1a\u8bf7\u7ed9\u51fa\u8f93\u5165\u3002\u4ee5\u4e0b\u662f\u6587\u6863\uff1a{'required': [], 'optional': [{'name': 'give_answer', 'description': '\u8f93\u51fa\u7b54\u6848', 'type': 'string'}, {'name': 'give_up_and_restart', 'description': '\u65e0\u6cd5\u4ece\u6b64\u6b65\u9aa4\u5904\u7406\u4efb\u52a1', 'type': 'string'}]}\r\n\uff1a{\"return_type\": \"give_answer\", \"final_answer\": \"\u573a\u9986 5a76c927bb5d660001e09731 \u7684\u83dc\u5355\u8be6\u60c5\u5982\u4e0b\uff1a\\n- \u83dc\u5355\u7c7b\u522b\uff1a\u4e2d\u9910\\n- \u4ef7\u683c\uff1a\\n - \u9879\u76ee 1\uff1a12.99 \u4eba\u6c11\u5e01\\n - \u9879\u76ee 2\uff1a15.99 \u4eba\u6c11\u5e01\\n - \u9879\u76ee 3\uff1a18.99 \u4eba\u6c11\u5e01\\n - \u9879\u76ee 4\uff1a22.99 \u4eba\u6c11\u5e01\\n- \u53ef\u7528\u6027\uff1a\\n - \u9879\u76ee 1\uff1a\u53ef\u7528\\n - \u9879\u76ee 2\uff1a\u53ef\u7528\\n - \u9879\u76ee 3\uff1a\u53ef\u7528\\n - \u9879\u76ee 4\uff1a\u53ef\u7528\\n- \u89c4\u5219\uff1a\\n - \u9879\u76ee 1\uff1a\u672a\u63d0\u53ca\u5177\u4f53\u89c4\u5219\\n - \u9879\u76ee 2\uff1a\u672a\u63d0\u53ca\u5177\u4f53\u89c4\u5219\\n - \u9879\u76ee 3\uff1a\u672a\u63d0\u53ca\u5177\u4f53\u89c4\u5219\\n - \u9879\u76ee 4\uff1a\u672a\u63d0\u53ca\u5177\u4f53\u89c4\u5219\\n- \u5957\u9910\u9879\uff1a\\n - \u9879\u76ee 1\uff1a2\u4efd\u5957\u9910\\n - \u9879\u76ee 2\uff1a2\u4efd\u5957\u9910\\n - \u9879\u76ee 3\uff1a2\u4efd\u5957\u9910\\n - \u9879\u76ee 4\uff1a2\u4efd\u5957\u9910\\n- \u5176\u4ed6\u76f8\u5173\u7ec6\u8282\uff1a\\n - \u7279\u522b\u8bf4\u660e\uff1a\u83dc\u5355\u53ef\u80fd\u4f1a\u66f4\u6539\uff0c\u4ef7\u683c\u4e5f\u53ef\u80fd\u4f1a\u53d8\u52a8\u3002\u8bf7\u5411\u573a\u9986\u786e\u8ba4\u6700\u65b0\u4fe1\u606f\u3002\"}\r\n<\/code><\/pre>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-7746\" title=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/0d1832161c2d59d.jpg\" alt=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-1\" width=\"1493\" height=\"2315\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/0d1832161c2d59d.jpg 1493w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/0d1832161c2d59d-193x300.jpg 193w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/0d1832161c2d59d-660x1024.jpg 660w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/0d1832161c2d59d-768x1191.jpg 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/0d1832161c2d59d-991x1536.jpg 991w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/0d1832161c2d59d-1321x2048.jpg 1321w\" sizes=\"auto, (max-width: 1493px) 100vw, 1493px\" \/>\u56fe 6: ToolGen \u7684\u63a8\u7406\u793a\u4f8b\u3002\u9996\u5148\u7ed9\u51fa\u7cfb\u7edf\u63d0\u793a\uff0c\u6ca1\u6709\u76f8\u5173\u5de5\u5177\u3002\u7136\u540e\u7528\u6237\u7ed9\u51fa\u4efb\u52a1\u67e5\u8be2\u3002ToolGen \u751f\u6210\u601d\u7ef4\uff0c\u7136\u540e\u6211\u4eec\u4f7f\u7528\u7528\u6237\u89d2\u8272\u63d0\u793a\u6a21\u578b\u751f\u6210\u884c\u52a8\u3002\u751f\u6210\u884c\u52a8\u540e\uff0c\u6211\u4eec\u518d\u6b21\u4f7f\u7528\u7528\u6237\u63d0\u4f9b\u5de5\u5177\u6587\u6863\u3002\u6a21\u578b\u5c06\u6839\u636e\u6b64\u6587\u6863\u751f\u6210\u5de5\u5177\u8f93\u5165\u3002<\/p>\n<h2>E \u6d88\u878d\u5b9e\u9a8c<\/h2>\n<p>\u8868 7 \u663e\u793a\u4e86\u7aef\u5230\u7aef\u8bc4\u4f30\u7684\u6d88\u878d\u7ed3\u679c\u3002\u5bf9\u4e8e\u672a\u89c1\u8fc7\u7684\u6307\u4ee4\uff0cToolGen Agent \u5728\u6ca1\u6709\u5de5\u5177\u8bb0\u5fc6\u6216\u68c0\u7d22\u8bad\u7ec3\u7684\u60c5\u51b5\u4e0b\u8868\u73b0\u7a0d\u597d\u3002\u7136\u800c\uff0c\u5bf9\u4e8e\u672a\u89c1\u8fc7\u7684\u5de5\u5177\uff0c\u5728\u6ca1\u6709\u524d\u4e24\u4e2a\u9636\u6bb5\u7684\u8bad\u7ec3\u65f6\uff0cSoPR \u548c SoWR \u90fd\u4f1a\u4e0b\u964d\u3002\u8fd9\u8868\u660e\u524d\u4e24\u4e2a\u9636\u6bb5\u7684\u8bad\u7ec3\u5728 ToolGen \u7684\u6cdb\u5316\u80fd\u529b\u4e2d\u8d77\u7740\u4f5c\u7528\uff0c\u800c\u68c0\u7d22\u8bad\u7ec3\u6bd4\u5de5\u5177\u8bb0\u5fc6\u66f4\u4e3a\u91cd\u8981\u3002<\/p>\n<p>\u8868 7: ToolGen \u7aef\u5230\u7aef\u8bc4\u4f30\u7684\u6d88\u878d\u7ed3\u679c\u3002\u8fd9\u91cc\u7684 Inst. \u4ee3\u8868\u672a\u89c1\u8fc7\u7684\u67e5\u8be2\uff08\u6307\u4ee4\uff09\uff0cTool. \u548c Cat. \u8868\u793a\u8bad\u7ec3\u671f\u95f4\u672a\u89c1\u8fc7\u7684\u5de5\u5177\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b<\/th>\n<th>SoPR<\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<th>SoWR<\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td>I1-Inst.<\/td>\n<td>I2-Inst.<\/td>\n<td>I3-Inst.<\/td>\n<td>Avg.<\/td>\n<td>I1-Inst.<\/td>\n<td>I2-Inst.<\/td>\n<td>I3-Inst.<\/td>\n<td>Avg.<\/td>\n<\/tr>\n<tr>\n<td>ToolGen<\/td>\n<td>54.60<\/td>\n<td>52.36<\/td>\n<td>43.44<\/td>\n<td>51.82<\/td>\n<td>50.31<\/td>\n<td>54.72<\/td>\n<td>26.23<\/td>\n<td>47.28<\/td>\n<\/tr>\n<tr>\n<td>w\/o retrieval training<\/td>\n<td>56.95<\/td>\n<td>46.70<\/td>\n<td>50.27<\/td>\n<td>52.42<\/td>\n<td>49.69<\/td>\n<td>50.94<\/td>\n<td>34.43<\/td>\n<td>47.27<\/td>\n<\/tr>\n<tr>\n<td>w\/o memorization<\/td>\n<td>56.03<\/td>\n<td>47.96<\/td>\n<td>57.38<\/td>\n<td>53.69<\/td>\n<td>49.08<\/td>\n<td>59.43<\/td>\n<td>34.43<\/td>\n<td>49.70<\/td>\n<\/tr>\n<tr>\n<td><\/td>\n<td>I1-Tool.<\/td>\n<td>I1-Cat.<\/td>\n<td>I2 Cat.<\/td>\n<td>Avg.<\/td>\n<td>I1-Tool<\/td>\n<td>I1-Cat.<\/td>\n<td>I2 Cat.<\/td>\n<td>Avg.<\/td>\n<\/tr>\n<tr>\n<td>ToolGen<\/td>\n<td>56.54<\/td>\n<td>49.46<\/td>\n<td>51.96<\/td>\n<td>52.66<\/td>\n<td>40.51<\/td>\n<td>39.87<\/td>\n<td>37.90<\/td>\n<td>39.53<\/td>\n<\/tr>\n<tr>\n<td>w\/o retrieval training<\/td>\n<td>49.47<\/td>\n<td>40.31<\/td>\n<td>37.90<\/td>\n<td>42.84<\/td>\n<td>36.71<\/td>\n<td>30.07<\/td>\n<td>36.29<\/td>\n<td>34.18<\/td>\n<\/tr>\n<tr>\n<td>w\/o memorization<\/td>\n<td>58.86<\/td>\n<td>46.19<\/td>\n<td>49.87<\/td>\n<td>51.70<\/td>\n<td>37.34<\/td>\n<td>38.56<\/td>\n<td>42.74<\/td>\n<td>39.32<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>F \u6cdb\u5316\u80fd\u529b<\/h2>\n<p>\u5bf9\u4e8e ToolGen Agent\uff0c\u6211\u4eec\u6d4b\u91cf\u6a21\u578b\u5728\u672a\u7ecf\u8fc7\u8bad\u7ec3\u7684\u5de5\u5177\u67e5\u8be2\u4e0a\u7684\u8868\u73b0\u3002\u8868 8 \u663e\u793a\u4e86\u6a21\u578b\u5728\u672a\u89c1\u8fc7\u5de5\u5177\u4e0a\u7684\u7aef\u5230\u7aef\u8bc4\u4f30\u3002ToolGen Agent \u7684\u8868\u73b0\u4f4e\u4e8e ToolLlama\uff0c\u8fd9\u4e5f\u8868\u660e\u5728\u5b8c\u6210\u5b8c\u6574\u4efb\u52a1\u65f6\u6cdb\u5316\u80fd\u529b\u8f83\u5dee\u3002\u6cdb\u5316\u95ee\u9898\u5728\u751f\u6210\u5f0f\u68c0\u7d22\u4e2d\u666e\u904d\u5b58\u5728\uff0c\u8d85\u51fa\u4e86\u672c\u6587\u7684\u8303\u56f4\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5c06\u5176\u7559\u5f85\u672a\u6765\u7814\u7a76\u3002<\/p>\n<p>\u8868 8: ToolGen \u7684\u6cdb\u5316\u7ed3\u679c\u3002\u6211\u4eec\u6d4b\u8bd5\u5e76\u6bd4\u8f83\u4e86 ToolGen \u4e0e\u5176\u4ed6\u6a21\u578b\u5728\u8bad\u7ec3\u671f\u95f4\u8981\u6c42\u672a\u89c1\u5de5\u5177\u7684\u67e5\u8be2\u4e0a\u7684\u8868\u73b0\u3002<\/p>\n<table>\n<thead>\n<tr>\n<th>\u6a21\u578b<\/th>\n<th>\u8bbe\u7f6e<\/th>\n<th>SoPR<\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<th>SoWR<\/th>\n<th><\/th>\n<th><\/th>\n<th><\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td><\/td>\n<td><\/td>\n<td>I1-Tool.<\/td>\n<td>I1-Cat.<\/td>\n<td>I2 Cat.<\/td>\n<td>Avg.<\/td>\n<td>I1-Tool<\/td>\n<td>I1-Cat.<\/td>\n<td>I2 Cat.<\/td>\n<td>Avg<\/td>\n<\/tr>\n<tr>\n<td>GPT-3.5<\/td>\n<td>GT.<\/td>\n<td>58.90<\/td>\n<td>60.70<\/td>\n<td>54.60<\/td>\n<td>58.07<\/td>\n<td>&#8211;<\/td>\n<td>&#8211;<\/td>\n<td>&#8211;<\/td>\n<td>&#8211;<\/td>\n<\/tr>\n<tr>\n<td>ToolLlama<\/td>\n<td>GT.<\/td>\n<td>57.38<\/td>\n<td>58.61<\/td>\n<td>56.85<\/td>\n<td>57.68<\/td>\n<td>43.04<\/td>\n<td>50.31<\/td>\n<td>54.84<\/td>\n<td>49.04<\/td>\n<\/tr>\n<tr>\n<td>ToolGen<\/td>\n<td>GT.<\/td>\n<td>52.32<\/td>\n<td>40.46<\/td>\n<td>39.65<\/td>\n<td>47.67<\/td>\n<td>39.24<\/td>\n<td>38.56<\/td>\n<td>40.32<\/td>\n<td>39.30<\/td>\n<\/tr>\n<tr>\n<td>GPT-3.5<\/td>\n<td>\u68c0\u7d22<\/td>\n<td>57.59<\/td>\n<td>53.05<\/td>\n<td>46.51<\/td>\n<td>52.78<\/td>\n<td>46.20<\/td>\n<td>54.25<\/td>\n<td>54.81<\/td>\n<td>51.58<\/td>\n<\/tr>\n<tr>\n<td>ToolLlama<\/td>\n<td>\u68c0\u7d22<\/td>\n<td>57.70<\/td>\n<td>61.76<\/td>\n<td>45.43<\/td>\n<td>54.96<\/td>\n<td>48.73<\/td>\n<td>50.98<\/td>\n<td>44.35<\/td>\n<td>48.30<\/td>\n<\/tr>\n<tr>\n<td>ToolGen<\/td>\n<td><\/td>\n<td>56.54<\/td>\n<td>49.46<\/td>\n<td>51.96<\/td>\n<td>52.66<\/td>\n<td>40.51<\/td>\n<td>39.87<\/td>\n<td>37.90<\/td>\n<td>39.53<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<h2>\u5c06 ToolBench \u6570\u636e\u9002\u914d\u5230 ToolGen<\/h2>\n<p>\u6211\u4eec\u7684 ToolGen \u6570\u636e\u662f\u4ece ToolBench \u6570\u636e\u4e2d\u9002\u914d\u548c\u8f6c\u6362\u800c\u6765\u7684\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u91c7\u7528\u5de5\u5177\u6587\u6863\u4f5c\u4e3a\u5de5\u5177\u8bb0\u5fc6\u8bad\u7ec3\u7684\u6570\u636e\uff0c\u5176\u4e2d\u8f93\u5165\u662f\u5de5\u5177\u6587\u6863\uff0c\u8f93\u51fa\u662f\u76f8\u5e94\u7684 Token\u3002<\/p>\n<p>\u5bf9\u4e8e\u68c0\u7d22\u8bad\u7ec3\uff0c\u6211\u4eec\u4f7f\u7528 ToolBench \u4e2d\u6ce8\u91ca\u7528\u4e8e\u5de5\u5177\u68c0\u7d22\u7684\u6570\u636e\uff0c\u5176\u4e2d\u4e00\u4e2a\u67e5\u8be2\u88ab\u6ce8\u91ca\u4e3a\u591a\u4e2a\u76f8\u5173\u5de5\u5177\u3002\u6211\u4eec\u5c06\u67e5\u8be2\u4f5c\u4e3a\u8f93\u5165\uff0c\u5e76\u5c06\u76f8\u5173\u5de5\u5177\u8f6c\u6362\u4e3a\u865a\u62df Token\u3002\u8fd9\u4e9b Token \u968f\u540e\u4f5c\u4e3a\u68c0\u7d22\u8bad\u7ec3\u7684\u8f93\u51fa\u3002<\/p>\n<p>\u5bf9\u4e8e\u7aef\u5230\u7aef\u7684\u667a\u80fd\u4f53\u8c03\u4f18\uff0c\u6211\u4eec\u4f7f\u7528\u4ea4\u4e92\u8f68\u8ff9\u4f5c\u4e3a\u6765\u6e90\uff0c\u5e76\u8fdb\u884c\u4ee5\u4e0b\u8f6c\u6362\uff1a(1) \u6bcf\u4e2a\u8f68\u8ff9\u5305\u542b\u53ef\u7528\u4e8e\u89e3\u51b3\u67e5\u8be2\u7684\u7cfb\u7edf\u63d0\u793a\u4e2d\u7684\u53ef\u7528\u5de5\u5177\u3002\u5f53\u5b8c\u6210\u4efb\u52a1\u65f6\uff0cToolLlama \u4f9d\u8d56\u4e8e\u7cfb\u7edf\u63d0\u793a\u4e2d\u68c0\u7d22\u5230\u7684\u5de5\u5177\u6765\u89e3\u51b3\u4efb\u52a1\uff0c\u800c ToolGen \u53ef\u4ee5\u76f4\u63a5\u751f\u6210\u5de5\u5177\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u79fb\u9664\u7cfb\u7edf\u63d0\u793a\u4e2d\u7684\u5de5\u5177\u3002(2) \u6211\u4eec\u5c06\u8f68\u8ff9\u4e2d\u7684\u6240\u6709\u5de5\u5177\u540d\u79f0\u66ff\u6362\u4e3a\u76f8\u5e94\u7684\u865a\u62df\u5de5\u5177 Token\u3002(3) \u5728\u539f\u59cb\u8f68\u8ff9\u4e2d\uff0c\u667a\u80fd\u4f53\u6a21\u578b\u6309\u987a\u5e8f\u751f\u6210 Thought\u3001Action \u548c Action Input\uff08\u4e5f\u79f0\u4e3a ReAct\uff09\u3002\u6211\u4eec\u5c06\u6574\u4e2a ReAct \u5206\u89e3\u4e3a\u4e09\u4e2a\u5bf9\u8bdd\u56de\u5408\u3002\u5728\u7b2c\u4e00\u4e2a\u56de\u5408\u4e2d\uff0c\u667a\u80fd\u4f53\u6a21\u578b\u751f\u6210\u4e00\u4e2a Thought\uff0c\u6211\u4eec\u4f7f\u7528\u7528\u6237\u6765\u63d0\u793a\u6a21\u578b\u751f\u6210\u4e00\u4e2a\u52a8\u4f5c\u3002\u5728\u7b2c\u4e8c\u4e2a\u56de\u5408\u4e2d\uff0c\u6a21\u578b\u751f\u6210\u52a8\u4f5c\uff0c\u5373\u865a\u62df\u5de5\u5177 Token\u3002\u7136\u540e\u6211\u4eec\u83b7\u53d6\u4e0e\u8fd9\u4e9b Token \u5bf9\u5e94\u7684\u6587\u6863\uff0c\u4ee5\u4fbf\u6a21\u578b\u77e5\u9053\u9700\u8981\u6307\u5b9a\u54ea\u4e9b\u53c2\u6570\u3002\u5728\u7b2c\u4e09\u4e2a\u56de\u5408\u4e2d\uff0c\u6a21\u578b\u4e3a\u5de5\u5177\u751f\u6210\u53c2\u6570\u3002<\/p>\n<p>\u6bcf\u4e2a\u6570\u636e\u96c6\u4e2d\u7684\u6837\u672c\u6570\u91cf\u89c1\u8868\u00a06\u3002\u5de5\u5177\u8bb0\u5fc6\u548c\u68c0\u7d22\u8bad\u7ec3\u7684\u6837\u672c\u89c1\u56fe\u00a07\u3002\u7aef\u5230\u7aef\u667a\u80fd\u4f53\u8c03\u4f18\u7684\u6837\u672c\u89c1\u56fe\u00a08\u3002<\/p>\n<pre><code># \u5de5\u5177\u8bb0\u5fc6\r\n\u7528\u6237: \u5de5\u5177\u540d\u79f0\uff1a\u6cf0\u56fd\u9a7e\u9a76\u6267\u7167 OCR\u3002\u5de5\u5177\u63cf\u8ff0\uff1a\u63d0\u53d6\u6cf0\u56fd\u9a7e\u9a76\u6267\u7167\u4e0a\u7684\u4fe1\u606f\u5e76\u8fd4\u56de\u6587\u672c\u7ed3\u679c\uff0c\u4f8b\u5982\u9a7e\u9a76\u6267\u7167\u53f7\u7801\u548c\u4e2a\u4eba\u4fe1\u606f\u3002API \u540d\u79f0\uff1a\u9a7e\u9a76\u6267\u7167 API \u63cf\u8ff0\uff1a\u63d0\u53d6\u6cf0\u56fd\u9a7e\u9a76\u6267\u7167\u4e0a\u7684\u4fe1\u606f\u5e76\u8fd4\u56de\u6587\u672c\u7ed3\u679c\uff0c\u4f8b\u5982\u9a7e\u9a76\u6267\u7167\u53f7\u7801\u548c\u4e2a\u4eba\u4fe1\u606f\u3002\r\n\u52a9\u624b: &lt;&lt;Thai Drivers License OCRDriver's License&gt;&gt;\r\n# \u68c0\u7d22\u8bad\u7ec3\r\n\u7528\u6237: \u6211\u548c\u6211\u7684\u670b\u53cb\u6b63\u5728\u7ec4\u7ec7\u4e00\u4e2a\u5173\u4e8e\u201c\u7f51\u9875\u5f00\u53d1\u201d\u548c\u201c\u79fb\u52a8\u5e94\u7528\u5f00\u53d1\u201d\u7684\u9ed1\u5ba2\u9a6c\u62c9\u677e\u3002\u6211\u4eec\u9700\u8981\u4e00\u4e9b\u7075\u611f\u548c\u6307\u5bfc\u3002\u4f60\u80fd\u4ece Medium.com \u83b7\u53d6\u8fd9\u4e9b\u4e3b\u9898\u7684\u70ed\u95e8\u6545\u4e8b\u5417\uff1f\r\n\u52a9\u624b: &lt;&lt;Medium&amp;&amp;\/search\/topics&gt;&gt;\r\n<\/code><\/pre>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"aligncenter size-full wp-image-7742\" title=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-4\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/2436f23b98933a7.png\" alt=\"ToolGen\uff1a\u901a\u8fc7\u751f\u6210\u5b9e\u73b0\u7edf\u4e00\u7684\u5de5\u5177\u68c0\u7d22\u548c\u8c03\u7528-4\" width=\"1661\" height=\"743\" srcset=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/2436f23b98933a7.png 1661w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/2436f23b98933a7-300x134.png 300w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/2436f23b98933a7-1024x458.png 1024w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/2436f23b98933a7-768x344.png 768w, https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2024\/10\/2436f23b98933a7-1536x687.png 1536w\" sizes=\"auto, (max-width: 1661px) 100vw, 1661px\" \/><\/p>\n<p>\u56fe 7: \u5de5\u5177\u8bb0\u5fc6\u548c\u68c0\u7d22\u8bad\u7ec3\u7684\u6570\u636e\u96c6\u793a\u4f8b\u3002\u6211\u4eec\u4f7f\u7528\u7528\u6237\u89d2\u8272\u6765\u8868\u793a\u8f93\u5165\uff0c\u4f7f\u7528\u52a9\u624b\u89d2\u8272\u6765\u8868\u793a\u8f93\u51fa\u3002<\/p>\n<pre><code># \u7aef\u5230\u7aef\u4ee3\u7406\u8c03\u4f18\r\n\u7cfb\u7edf\uff1a\r\n\u60a8\u662f\u4e00\u4e2a AutoGPT\uff0c\u80fd\u591f\u5229\u7528\u4f17\u591a\u5de5\u5177\u548c\u529f\u80fd\u6765\u5b8c\u6210\u7ed9\u5b9a\u7684\u4efb\u52a1\u3002\r\n1. \u9996\u5148\uff0c\u6211\u4f1a\u63d0\u4f9b\u7ed9\u60a8\u4efb\u52a1\u63cf\u8ff0\uff0c\u60a8\u7684\u4efb\u52a1\u5c06\u5f00\u59cb\u3002\r\n2. \u5728\u6bcf\u4e00\u6b65\u4e2d\uff0c\u60a8\u9700\u8981\u901a\u8fc7\u751f\u6210\u4e00\u4e2a\u64cd\u4f5c Token \u6765\u786e\u5b9a\u4e0b\u4e00\u6b65\u884c\u52a8\u3002\r\n3. \u5728 Token \u4e4b\u540e\uff0c\u60a8\u5c06\u6536\u5230\u4e0e\u8be5 Token \u5bf9\u5e94\u7684\u64cd\u4f5c\u6587\u6863\u3002\u60a8\u9700\u8981\u751f\u6210\u64cd\u4f5c\u7684\u8f93\u5165\uff0c\u5e2e\u52a9\u60a8\u8f6c\u5230\u65b0\u7684\u72b6\u6001\u3002\u968f\u540e\uff0c\u60a8\u5c06\u5bf9\u4e0b\u4e00\u6b65\u8fdb\u884c\u51b3\u7b56\uff0c\u5e76\u91cd\u590d\u6b64\u8fc7\u7a0b\u3002\r\n4. \u5728\u751f\u6210\u591a\u4e2a\u64cd\u4f5c\u548c\u8f93\u5165\u7684\u82e5\u5e72\u6b21\u8fed\u4ee3\u540e\uff0c\u60a8\u6700\u7ec8\u5c06\u5b8c\u6210\u4efb\u52a1\u5e76\u63d0\u4f9b\u60a8\u7684\u6700\u7ec8\u7b54\u6848\u3002\r\n\u8bb0\u4f4f\uff1a\r\n1. \u72b6\u6001\u53d8\u5316\u662f\u4e0d\u53ef\u9006\u7684\uff0c\u60a8\u65e0\u6cd5\u8fd4\u56de\u5230\u4e4b\u524d\u7684\u72b6\u6001\u3002\r\n2. \u4fdd\u6301\u60a8\u7684\u64cd\u4f5c\u7b80\u6d01\uff0c\u9650\u5236\u5728\u6700\u9002\u5408\u5f53\u524d\u67e5\u8be2\u7684\u8303\u56f4\u5185\u3002\r\n3. \u60a8\u53ef\u4ee5\u8fdb\u884c\u591a\u6b21\u5c1d\u8bd5\u3002\u5982\u679c\u60a8\u8ba1\u5212\u4e0d\u65ad\u5c1d\u8bd5\u4e0d\u540c\u7684\u6761\u4ef6\uff0c\u8bf7\u6bcf\u6b21\u5c1d\u8bd5\u4e00\u4e2a\u6761\u4ef6\u3002\r\n4. \u5982\u679c\u60a8\u8ba4\u4e3a\u60a8\u5df2\u7ecf\u6536\u96c6\u4e86\u8db3\u591f\u7684\u4fe1\u606f\uff0c\u8bf7\u751f\u6210\u64cd\u4f5c \"&lt;&lt;Finish&gt;&gt; with argument give_answer\"\uff0c\u4ee5\u63d0\u4f9b\u60a8\u5bf9\u8be5\u4efb\u52a1\u7684\u7b54\u6848\u3002\r\n5. \u5982\u679c\u60a8\u89c9\u5f97\u5728\u8fd9\u4e00\u6b65\u65e0\u6cd5\u5904\u7406\u4efb\u52a1\uff0c\u8bf7\u751f\u6210\u64cd\u4f5c \"&lt;&lt;Finish&gt;&gt; with argument give_up_and_restart\"\u3002\r\n\u8ba9\u6211\u4eec\u5f00\u59cb\u5427\uff01\r\n\u4efb\u52a1\u63cf\u8ff0\uff1a\u60a8\u5e94\u8be5\u4f7f\u7528\u64cd\u4f5c\u6765\u5904\u7406\u5b9e\u65f6\u7528\u6237\u67e5\u8be2\u3002\u8bb0\u4f4f\uff1a\r\n1. \u5728\u4efb\u52a1\u7ed3\u675f\u65f6\u59cb\u7ec8\u751f\u6210 \"&lt;&lt;Finish&gt;&gt;\"\u3002\u6700\u7ec8\u7b54\u6848\u5e94\u5305\u542b\u8db3\u591f\u7684\u4fe1\u606f\u4ee5\u5c55\u793a\u7ed9\u7528\u6237\u3002\u5982\u679c\u60a8\u65e0\u6cd5\u5904\u7406\u4efb\u52a1\uff0c\u6216\u8005\u53d1\u73b0\u64cd\u4f5c\u603b\u662f\u5931\u8d25\uff08\u8be5\u529f\u80fd\u5f53\u524d\u65e0\u6548\uff09\uff0c\u8bf7\u4f7f\u7528\u64cd\u4f5c &lt;&lt;Finish&gt;&gt; with give_up_and_restart\u3002\r\n2. \u4ec5\u751f\u6210\u64cd\u4f5c\u548c\u8f93\u5165\u3002\r\n\u7528\u6237\uff1a\r\n\u60a8\u80fd\u5426\u83b7\u53d6\u90ae\u653f\u7f16\u7801 'PL11DN' \u7684\u5730\u5740\uff1f\u6211\u60f3\u77e5\u9053\u627e\u5230\u7684\u9879\u76ee\u6570\u91cf\u3001\u533a\u3001\u5206\u533a\u3001\u53bf\u3001\u56fd\u5bb6\u4ee5\u53ca\u5730\u7406\u7f16\u7801\u8be6\u60c5\uff08\u4e1c\u5411\u3001\u5317\u5411\u3001\u7eac\u5ea6\u548c\u7ecf\u5ea6\uff09\u3002\r\n\u5f00\u59cb\uff01\r\n\u52a9\u624b\uff1a\r\n\u6839\u636e\u7528\u6237\u7684\u67e5\u8be2\uff0c\u4ed6\u4eec\u60f3\u83b7\u53d6\u90ae\u653f\u7f16\u7801 'PL11DN' \u7684\u5730\u5740\uff0c\u5e76\u68c0\u7d22\u7279\u5b9a\u8be6\u60c5\uff0c\u5982\u627e\u5230\u7684\u9879\u76ee\u6570\u91cf\u3001\u533a\u3001\u5206\u533a\u3001\u53bf\u3001\u56fd\u5bb6\u4ee5\u53ca\u5730\u7406\u7f16\u7801\u8be6\u60c5\u3002\u5f53\u524d\u7684 API \u8c03\u7528\u662f\u9488\u5bf9\u51fd\u6570 'find_by_postcode'\uff0c\u53c2\u6570 'postcode' \u8bbe\u7f6e\u4e3a 'PL11DN'\u3002\u6b64 API \u5f88\u6709\u7528\uff0c\u56e0\u4e3a\u5b83\u53ef\u4ee5\u6839\u636e\u7ed9\u5b9a\u7684\u82f1\u56fd\u90ae\u653f\u7f16\u7801\u63d0\u4f9b\u5730\u5740\u3002\u901a\u8fc7\u4f7f\u7528\u6b64 API\uff0c\u6211\u5c06\u80fd\u591f\u4e3a\u7528\u6237\u7684\u67e5\u8be2\u68c0\u7d22\u6240\u9700\u7684\u4fe1\u606f\u3002\r\n\u7528\u6237\uff1a\r\n\u751f\u6210\u64cd\u4f5c\u3002\r\n\u52a9\u624b\uff1a\r\n&lt;&lt;Find By PostCode&amp;&amp;Find By PostCode&gt;&gt;\r\n\u7528\u6237\uff1a\r\n\u8bf7\u7ed9\u51fa\u8f93\u5165\u3002\u8fd9\u662f\u6587\u6863\uff1a{'name': '&lt;&lt;Find By PostCode&amp;&amp;Find By PostCode&gt;&gt;', 'description': '\u201cFind By PostCode\u201d API \u6839\u636e\u7ed9\u5b9a\u7684\u82f1\u56fd\u90ae\u653f\u7f16\u7801\u83b7\u53d6\u5730\u5740\u3002\u6b64 API \u63d0\u4f9b JSON \u63a5\u53e3\u4ee5\u641c\u7d22\u82f1\u56fd\u90ae\u653f\u7f16\u7801\u7684\u5730\u5740\u3002', 'required': [{'name': 'postcode', 'type': 'string', 'description': '', 'default': 'PL11DN'}], 'optional': []}\r\n\u52a9\u624b\uff1a\r\n{\r\n\"postcode\": \"PL11DN\"\r\n}\r\n\u5de5\u5177\uff1a\r\n{\"error\": \"\", \"response\": \"{'status': 200, 'message': '\u6210\u529f', 'noOfItems': 37, 'result': {'postcode': 'PL1 1DN', 'district': '\u666e\u5229\u8305\u65af', 'ward': '\u5723\u5f7c\u5f97\u548c\u6d77\u6ee8', 'county': '\u5fb7\u6587', 'country': '\u82f1\u683c\u5170', 'geocode': {'eastings': '248025', 'northings': '0054420', 'lattitude': '50.370001', 'longitude': '-4.138397'}, 'addresses': [{'po_box': '', 'line_2': '', 'postcode_type': 'S', 'line_3': '', 'su_organisation_indicator': ' ', 'building_number': '5', 'udprn': '18911184', 'thoroughfare': '\u5723\u5b89\u5fb7\u9c81\u5341\u5b57', 'line_1': '5 \u5723\u5b89\u5fb7\u9c81\u5341\u5b57', 'postcode_outcode': 'PL1', 'postcode_compact': 'PL11DN', 'dependant_thoroughfare': '', 'premise': '5', 'post_town': '\u666e\u5229\u8305\u65af', 'building_name': '', 'organisation_name': '', 'postcode': 'PL1 1DN', 'department_name': '', 'postcode_incode': '1DN', 'address': '5 \u5723\u5b89\u5fb7\u9c81\u5341\u5b57, \u666e\u5229\u8305\u65af, PL1 1DN', 'double_dependant_locality': '', 'sub_building_name': '', 'dependant_locality': '', 'delivery_point_suffix': '4T'}, {'po_box': '', 'line_2': '2 \u5723\u5b89\u5fb7\u9c81\u5341\u5b57', 'p...\r\n\u7528\u6237\uff1a\r\n\u751f\u6210\u64cd\u4f5c\u3002\r\n\u52a9\u624b\uff1a\r\n&lt;&lt;Finish&gt;&gt;\r\n\u7528\u6237\uff1a\r\n\u8bf7\u7ed9\u51fa\u8f93\u5165\u3002\u8fd9\u662f\u6587\u6863\uff1a{'required': [], 'optional': [{'name': 'give_answer', 'description': '\u8f93\u51fa\u7b54\u6848', 'type': 'string'}, {'name': 'give_up_and_restart', 'description': '\u65e0\u6cd5\u5904\u7406\u6b64\u6b65\u9aa4\u7684\u4efb\u52a1', 'type': 'string'}]}\r\n\u52a9\u624b\uff1a\r\n{\r\n\"return_type\": \"give_answer\",\r\n\"final_answer\": \"\u627e\u5230\u7684\u9879\u76ee\u6570\u91cf\uff1a37\\n\u533a\uff1a\u666e\u5229\u8305\u65af\\n\u5206\u533a\uff1a\u5723\u5f7c\u5f97\u548c\u6d77\u6ee8\\n\u53bf\uff1a\u5fb7\u6587\\n\u56fd\u5bb6\uff1a\u82f1\u683c\u5170\\n\u5730\u7406\u7f16\u7801\u8be6\u60c5\uff1a\\n- \u4e1c\u5411\uff1a248025\\n- \u5317\u5411\uff1a0054420\\n- \u7eac\u5ea6\uff1a50.370001\\n\u7ecf\u5ea6\uff1a-4.138397\"\r\n}\r\n<\/code><\/pre>\n<p>![[Pasted image 20241027190429.png]]<\/p>\n<p>\u56fe 8: \u7aef\u5230\u7aef\u4ee3\u7406\u8c03\u4f18\u7684\u793a\u4f8b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>ToolGen \u662f\u4e00\u4e2a\u5c06\u5de5\u5177\u77e5\u8bc6\u76f4\u63a5\u96c6\u6210\u5230\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u4e2d\u7684\u6846\u67b6\uff0c\u901a\u8fc7\u5c06\u6bcf\u4e2a\u5de5\u5177\u8868\u793a\u4e3a\u72ec\u7279\u7684\u6807\u8bb0\uff0c\u5b9e\u73b0\u65e0\u7f1d\u7684\u5de5\u5177\u8c03\u7528\u548c\u8bed\u8a00\u751f\u6210\u3002\u8be5\u9879\u76ee\u7531 Renxi Wang \u7b49\u4eba\u5f00\u53d1\uff0c\u65e8\u5728\u63d0\u5347\u5de5\u5177\u68c0\u7d22\u548c\u4efb\u52a1\u5b8c\u6210\u7684\u6027\u80fd\u3002 \u5de5\u5177\u6807\u8bb0\u5316\uff1a\u5c06\u5de5\u5177\u8f6c\u6362\u4e3a\u72ec&#8230;<\/p>\n","protected":false},"author":1,"featured_media":61156,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20,34],"tags":[],"class_list":["post-7738","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","category-knowledge"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/7738","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/comments?post=7738"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/posts\/7738\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media\/61156"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/media?parent=7738"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/categories?post=7738"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/ja\/wp-json\/wp\/v2\/tags?post=7738"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}