{"id":1641,"date":"2026-07-10T12:06:09","date_gmt":"2026-07-10T04:06:09","guid":{"rendered":"https:\/\/lingbo.online\/?p=1641"},"modified":"2026-07-10T14:56:43","modified_gmt":"2026-07-10T06:56:43","slug":"toolspec-accelerating-tool-calling-via-schema-aware-and-retrieval-augmented-speculative-decoding","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/uncategorized\/toolspec-accelerating-tool-calling-via-schema-aware-and-retrieval-augmented-speculative-decoding\/","title":{"rendered":"TOOLSPEC: Accelerating Tool Calling via Schema-Aware and Retrieval-Augmented Speculative Decoding"},"content":{"rendered":"<h3>Meta Data<\/h3>\n<ul>\n<li>\u53d1\u8868\u65f6\u95f4\uff1a2026-04-15\uff08arXiv \u9996\u6b21\u63d0\u4ea4\uff09\uff1b2026-04-16\uff08\u516c\u5e03\uff09<\/li>\n<li>\u4f5c\u8005\uff1aHeming Xia, Yongqi Li, Cunxiao Du, Mingbo Song, Wenjie Li<\/li>\n<li>\u8bba\u6587\u94fe\u63a5\uff1a<a href=\"https:\/\/arxiv.org\/pdf\/2604.13519.pdf\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/arxiv.org\/pdf\/2604.13519.pdf<\/a><\/li>\n<li>arXiv \u9875\u9762\uff1a<a href=\"https:\/\/arxiv.org\/abs\/2604.13519\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/arxiv.org\/abs\/2604.13519<\/a><\/li>\n<li>\u9879\u76ee\u94fe\u63a5\uff1a<a href=\"https:\/\/github.com\/hemingkx\/ToolSpec\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/hemingkx\/ToolSpec<\/a><\/li>\n<\/ul>\n<h3>ToolSpec\uff1a\u901a\u8fc7 Schema \u611f\u77e5\u4e0e\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\u89e3\u7801\u52a0\u901f\u5de5\u5177\u8c03\u7528<\/h3>\n<h4>\u6458\u8981<\/h4>\n<p>\u5de5\u5177\u8c03\u7528\uff08tool calling\uff09\u901a\u8fc7\u4f7f\u5927\u8bed\u8a00\u6a21\u578b\uff08Large Language Models, LLMs\uff09\u80fd\u591f\u4e0e\u5916\u90e8\u5e94\u7528\u4ea4\u4e92\uff0c\u6781\u5927\u5730\u6269\u5c55\u4e86 LLM \u7684\u5b9e\u7528\u4ef7\u503c\u3002\u968f\u7740 LLM \u80fd\u529b\u4e0d\u65ad\u63d0\u5347\uff0c\u6709\u6548\u7684\u5de5\u5177\u4f7f\u7528 \u8d8a\u6765\u8d8a\u9700\u8981\u591a\u6b65\u3001\u591a\u8f6e\u4ea4\u4e92\u6765\u89e3\u51b3\u590d\u6742\u4efb\u52a1\u3002\u7136\u800c\uff0c\u7531\u6b64\u5e26\u6765\u7684\u5de5\u5177\u4ea4\u4e92\u6b21\u6570\u589e\u957f\u4f1a\u4ea7\u751f\u663e\u8457\u7684\u63a8\u7406\u5ef6\u8fdf\uff0c\u6210\u4e3a\u5b9e\u65f6 LLM \u670d\u52a1\u7684\u4e00\u9879\u5173\u952e\u6311\u6218\u3002\u901a\u8fc7\u5b9e\u8bc1\u5206\u6790\uff0c\u6211\u4eec\u53d1\u73b0\u5de5\u5177\u8c03\u7528\u8f68\u8ff9\u9ad8\u5ea6\u7ed3\u6784\u5316\uff0c\u9075\u5faa\u53d7\u7ea6\u675f\u7684 schema\uff0c\u5e76\u4e14\u5e38\u5e38\u8868\u73b0\u51fa\u91cd\u590d\u51fa\u73b0\u7684\u8c03\u7528\u6a21\u5f0f\u3002\u57fa\u4e8e\u8fd9\u4e9b\u89c2\u5bdf\uff0c\u6211\u4eec\u63d0\u51fa ToolSpec\uff0c\u4e00\u79cd\u9762\u5411\u52a0\u901f\u5de5\u5177\u8c03\u7528\u7684 schema \u611f\u77e5\u3001\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\u89e3\u7801\uff08speculative decoding, SD\uff09\u65b9\u6cd5\u3002ToolSpec \u5229\u7528\u9884\u5b9a\u4e49\u7684\u5de5\u5177 schema \u751f\u6210\u51c6\u786e\u7684\u8349\u7a3f\uff0c\u5e76\u4f7f\u7528\u6709\u9650\u72b6\u6001\u673a\uff08finite-state machine, FSM\uff09\u5728\u786e\u5b9a\u6027 schema token \u586b\u5145\u4e0e\u9488\u5bf9\u53ef\u53d8\u5b57\u6bb5\u7684\u63a8\u6d4b\u751f\u6210\u4e4b\u95f4\u4ea4\u66ff\u8fdb\u884c\u3002\u6b64\u5916\uff0cToolSpec \u68c0\u7d22\u76f8\u4f3c\u7684\u5386\u53f2\u5de5\u5177\u8c03\u7528\u5e76\u5c06\u5176\u590d\u7528\u4e3a\u8349\u7a3f\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63d0\u5347\u6548\u7387\u3002ToolSpec \u63d0\u4f9b\u4e86\u4e00\u79cd\u5373\u63d2\u5373\u7528\u3001\u65e0\u9700\u8bad\u7ec3\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u53ef\u65e0\u7f1d\u96c6\u6210\u5230\u73b0\u6709 LLM \u5de5\u4f5c\u6d41\u4e2d\u3002\u5728\u591a\u4e2a\u5de5\u5177\u8c03\u7528\u57fa\u51c6\u4e0a\u7684\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0cToolSpec \u6700\u9ad8\u53ef\u5b9e\u73b0 <span class=\"katex-eq\" data-katex-display=\"false\">4.2\\times<\/span> \u52a0\u901f\uff0c\u663e\u8457\u4f18\u4e8e\u73b0\u6709\u65e0\u9700\u8bad\u7ec3\u7684\u63a8\u6d4b\u89e3\u7801\u65b9\u6cd5\u3002\u6211\u4eec\u5728 <a href=\"https:\/\/github.com\/hemingkx\/ToolSpec\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/hemingkx\/ToolSpec<\/a> \u53d1\u5e03\u4e86\u4ee3\u7801\u3002<\/p>\n<h4>\u5f15\u8a00<\/h4>\n<p>\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u7684\u6700\u65b0\u8fdb\u5c55\u663e\u8457\u62d3\u5bbd\u4e86\u5b83\u4eec\u7684\u80fd\u529b\u8fb9\u754c\uff0c\u5305\u62ec\u5f00\u653e\u57df\u5bf9\u8bdd\u3001\u6570\u5b66\u63a8\u7406\u548c\u7a0b\u5e8f\u5408\u6210\u7b49\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u5de5\u5177\u8c03\u7528\u5df2\u6210\u4e3a\u4e00\u9879\u5173\u952e\u80fd\u529b\uff0c\u4f7f LLM \u80fd\u591f\u4e0e\u641c\u7d22\u5f15\u64ce\u3001\u8ba1\u7b97\u5668\u548c\u4ee3\u7801\u89e3\u91ca\u5668\u7b49\u5916\u90e8\u5de5\u5177\u4ea4\u4e92\u3002\u901a\u8fc7\u5f15\u5165\u5916\u90e8\u5de5\u5177\uff0cLLM \u80fd\u591f\u7f13\u89e3\u72ec\u7acb\u6a21\u578b\u7684\u4e00\u4e9b\u6839\u672c\u5c40\u9650\uff0c\u5305\u62ec\u77e5\u8bc6\u8fc7\u65f6\u548c\u7b97\u672f\u9519\u8bef\u7b49\u3002\u968f\u7740 LLM \u80fd\u529b\u6301\u7eed\u6539\u8fdb\uff0c\u5de5\u5177\u4f7f\u7528 \u8d8a\u6765\u8d8a\u591a\u5730\u91c7\u7528\u591a\u6b65\u3001\u591a\u8f6e\u3001\u53cd\u9988\u9a71\u52a8\u7684\u4ea4\u4e92\u5faa\u73af\u6765\u89e3\u51b3\u590d\u6742\u4efb\u52a1\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239001.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 1\uff1aToolSpec \u793a\u610f\u56fe\u3002ToolSpec \u901a\u8fc7\u4e24\u9879\u521b\u65b0\u52a0\u901f\u5de5\u5177\u8c03\u7528\u751f\u6210\uff1a1\uff09schema \u611f\u77e5\u8349\u7a3f\u751f\u6210\uff08schema-aware drafting\uff09\uff0c\u5176\u4e2d\u9884\u5b9a\u4e49\u5de5\u5177 schema \u4f5c\u4e3a\u5fe0\u5b9e\u8349\u7a3f\uff0c\u5e76\u652f\u6301\u5bf9\u53d7\u7ea6\u675f\u53d8\u91cf\uff08\u4f8b\u5982\u5de5\u5177\u540d\u79f0\uff09\u8fdb\u884c\u5e76\u884c\u9a8c\u8bc1\uff1b2\uff09\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\uff08retrieval-augmented speculation\uff09\uff0c\u68c0\u7d22\u5e76\u590d\u7528\u76f8\u4f3c\u7684\u5386\u53f2\u5de5\u5177\u8c03\u7528\u4f5c\u4e3a\u9ad8\u8d28\u91cf\u8349\u7a3f\u3002<\/center><\/p>\n<p>\u7136\u800c\uff0c\u5de5\u5177\u8c03\u7528\u9891\u7387\u7684\u4e0d\u65ad\u589e\u957f\u4f1a\u663e\u8457\u589e\u52a0\u63a8\u7406\u5ef6\u8fdf\uff0c\u7ed9\u5b9e\u65f6 LLM \u670d\u52a1\u5e26\u6765\u91cd\u5927\u6311\u6218\u3002\u8fd1\u671f\u5de5\u4f5c\u63d0\u51fa\u4e86\u591a\u79cd\u52a0\u901f\u5de5\u5177\u8c03\u7528\u7684\u7b56\u7565\u3002\u4e00\u7c7b\u7814\u7a76\u5c06\u5de5\u5177\u8c03\u7528\u8f68\u8ff9\u5efa\u6a21\u4e3a\u523b\u753b\u5de5\u5177\u95f4\u4f9d\u8d56\u5173\u7cfb\u7684\u6709\u5411\u65e0\u73af\u56fe\uff08DAGs\uff09\uff0c\u5e76\u5e76\u884c\u6267\u884c\u5f7c\u6b64\u72ec\u7acb\u7684\u5de5\u5177\u4ee5\u964d\u4f4e\u6574\u4f53\u5ef6\u8fdf\u3002\u53e6\u4e00\u7c7b\u5de5\u4f5c\u901a\u8fc7\u5728\u6587\u672c\u751f\u6210\u8fc7\u7a0b\u4e2d\u91cd\u53e0\u6267\u884c\u5de5\u5177\u8c03\u7528\u6765\u63d0\u5347\u5e76\u884c\u6027\uff0c\u5373\u5728\u6a21\u578b\u4ecd\u5728\u89e3\u7801\u65f6\u542f\u52a8\u5de5\u5177\u8c03\u7528\uff0c\u4ece\u800c\u8fdb\u4e00\u6b65\u52a0\u901f\u7aef\u5230\u7aef\u5de5\u5177\u8c03\u7528\u3002<\/p>\n<p>\u5c3d\u7ba1\u8fd9\u4e9b\u5728\u52a0\u901f\u5de5\u5177\u6267\u884c\u65b9\u9762\u7684\u5de5\u4f5c\u53d6\u5f97\u4e86\u8fdb\u5c55\uff0c\u672c\u6587\u5c06\u5173\u6ce8\u70b9\u8f6c\u5411\u63d0\u5347\u5de5\u5177\u8c03\u7528\u751f\u6210\u7684\u6548\u7387\u3002\u6211\u4eec\u7684\u521d\u6b65\u5206\u6790\u8868\u660e\uff0c\u751f\u6210\u5ef6\u8fdf\u53ef\u80fd\u6210\u4e3a\u4e3b\u5bfc\u74f6\u9888\uff1a\u5728\u5e7f\u6cdb\u4f7f\u7528\u7684\u57fa\u51c6 ToolBench \u4e0a\uff0c\u4f7f\u7528 Qwen2.5-14B-Instruct \u751f\u6210\u5de5\u5177\u8c03\u7528\u7ea6\u5360\u7aef\u5230\u7aef\u63a8\u7406\u5ef6\u8fdf\u7684 <span class=\"katex-eq\" data-katex-display=\"false\">80\\%<\/span>\uff0c\u5e76\u4e14\u5927\u7ea6\u662f\u6267\u884c\u5de5\u5177\u6240\u8017\u65f6\u95f4\u7684 <span class=\"katex-eq\" data-katex-display=\"false\">4\\times<\/span>\u3002\u6b64\u5916\uff0c\u5bf9\u4e8e\u56fa\u5b9a\u5de5\u5177\u96c6\u548c\u73af\u5883\uff0c\u5de5\u5177\u6267\u884c\u65f6\u95f4\u901a\u5e38\u76f8\u5bf9\u7a33\u5b9a\uff1b\u800c\u5de5\u5177\u8c03\u7528\u751f\u6210\u5ef6\u8fdf\u4f1a\u968f\u6a21\u578b\u89c4\u6a21\u4ee5\u53ca\u751f\u6210\u5de5\u5177\u8c03\u7528\u5e8f\u5217\u957f\u5ea6\u800c\u589e\u957f\uff0c\u4f7f\u5176\u6210\u4e3a\u8bb8\u591a\u5de5\u5177\u4f7f\u7528\u573a\u666f\u4e2d\u7684\u4e3b\u8981\u5ef6\u8fdf\u6765\u6e90\u3002<\/p>\n<p>\u5728\u672c\u5de5\u4f5c\u4e2d\uff0c\u6211\u4eec\u63d0\u51fa\u5173\u4e8e\u5de5\u5177\u8c03\u7528\u7684\u4e24\u4e2a\u5173\u952e\u89c2\u5bdf\uff1a<strong>1\uff09\u5de5\u5177\u8c03\u7528\u8f93\u51fa\u9ad8\u5ea6\u7ed3\u6784\u5316\u3002<\/strong> \u8fd9\u4e9b\u8f93\u51fa\u901a\u5e38\u4ee5 JSON \u8868\u793a\uff0c\u5e76\u9075\u5faa\u4e25\u683c\u53d7\u7ea6\u675f\u7684 schema\u3002\u56e0\u6b64\uff0c\u5728\u751f\u6210\u7684 tokens \u4e2d\uff0c\u53ea\u6709\u4e00\u5c0f\u90e8\u5206\u5bf9\u5e94\u53ef\u53d8\u5b57\u6bb5\uff0c\u5176\u4f59\u5747\u7531 schema \u51b3\u5b9a\uff1b<strong>2\uff09LLM \u5728\u4e0d\u540c\u8bf7\u6c42\u4e4b\u95f4\u8868\u73b0\u51fa\u91cd\u590d\u51fa\u73b0\u7684\u5de5\u5177\u8c03\u7528\u6a21\u5f0f\uff0c<\/strong> \u8868\u660e\u5b58\u5728\u53ef\u590d\u7528\u7684\u7ed3\u6784\u4e0e\u8bed\u4e49\u89c4\u5f8b\u3002\u53d7\u8fd9\u4e9b\u6d1e\u5bdf\u542f\u53d1\uff0c\u6211\u4eec\u63d0\u51fa ToolSpec\uff0c\u4e00\u79cd\u9762\u5411\u52a0\u901f\u5de5\u5177\u8c03\u7528\u7684 schema \u611f\u77e5\u3001\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\u89e3\u7801\uff08SD\uff09\u65b9\u6cd5\u3002ToolSpec \u5229\u7528\u4efb\u52a1\u7279\u5b9a\u7684\u5de5\u5177 schema \u6784\u9020\u9ad8\u8d28\u91cf\u8349\u7a3f\u5e8f\u5217\uff0c\u5e76\u91c7\u7528\u6709\u9650\u72b6\u6001\u673a\u5728\u786e\u5b9a\u6027 schema token \u586b\u5145\u4e0e\u9488\u5bf9\u53ef\u53d8\u5b57\u6bb5\u7684\u63a8\u6d4b\u751f\u6210\u4e4b\u95f4\u4ea4\u66ff\u3002\u6b64\u5916\uff0cToolSpec \u68c0\u7d22\u76f8\u4f3c\u7684\u5386\u53f2\u5de5\u5177\u8c03\u7528\u5e76\u5c06\u5176\u590d\u7528\u4e3a\u8349\u7a3f\u5019\u9009\uff0c\u4ece\u800c\u8fdb\u4e00\u6b65\u63d0\u5347\u6548\u7387\u3002\u5728 API-Bank\u3001ToolAlpaca\u3001BFCLv2 \u548c ToolBench \u7b49\u5e7f\u6cdb\u4f7f\u7528\u7684\u5de5\u5177\u8c03\u7528\u57fa\u51c6\u4e0a\u7684\u5b9e\u9a8c\u8868\u660e\uff0cToolSpec \u6700\u9ad8\u53ef\u5b9e\u73b0 <span class=\"katex-eq\" data-katex-display=\"false\">4.2\\times<\/span> \u52a0\u901f\uff0c\u5e76\u4f18\u4e8e\u73b0\u6709 SD \u65b9\u6cd5\u3002<\/p>\n<p>\u6211\u4eec\u5c06\u8d21\u732e\u603b\u7ed3\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>\u6211\u4eec\u63d0\u51fa ToolSpec\uff0c\u4e00\u79cd\u9762\u5411\u52a0\u901f\u5de5\u5177\u8c03\u7528\u751f\u6210\u7684 schema \u611f\u77e5\u3001\u68c0\u7d22\u589e\u5f3a SD \u65b9\u6cd5\u3002\u5b83\u63d0\u4f9b\u4e86\u4e00\u79cd\u65e0\u9700\u8bad\u7ec3\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u53ef\u65e0\u7f1d\u96c6\u6210\u5230\u73b0\u6709 LLM \u5de5\u4f5c\u6d41\u4e2d\u3002<\/li>\n<li>\u5728\u591a\u79cd\u6a21\u578b\u548c\u4efb\u52a1\u4e0a\u7684\u5927\u91cf\u5b9e\u9a8c\u8868\u660e\uff0cToolSpec \u7a33\u5b9a\u5b9e\u73b0 <span class=\"katex-eq\" data-katex-display=\"false\">3.5\\times \\sim 4.2\\times<\/span> \u52a0\u901f\uff0c\u76f8\u5bf9\u6b64\u524d\u6700\u4f18 SD \u65b9\u6cd5\u6700\u9ad8\u63d0\u5347 <span class=\"katex-eq\" data-katex-display=\"false\">71\\%<\/span>\u3002<\/li>\n<li>\u8fdb\u4e00\u6b65\u5206\u6790\u9a8c\u8bc1\u4e86\u6211\u4eec\u6240\u63d0\u51fa\u7ec4\u4ef6\u7684\u6709\u6548\u6027\uff0c\u5e76\u63ed\u793a\u4e86\u683c\u5f0f\u9075\u5faa\u5ea6\u5bf9 ToolSpec \u52a0\u901f\u6548\u679c\u7684\u5f71\u54cd\uff0c\u8868\u660e\u968f\u7740 LLM \u80fd\u529b\u6301\u7eed\u63d0\u5347\uff0cToolSpec \u7684\u6548\u7387\u6f5c\u529b\u5c06\u8fdb\u4e00\u6b65\u91ca\u653e\u3002<\/li>\n<\/ul>\n<h4>\u76f8\u5173\u5de5\u4f5c<\/h4>\n<h5>\u9ad8\u6548\u5de5\u5177\u8c03\u7528<\/h5>\n<p>\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09 \u8d8a\u6765\u8d8a\u4f9d\u8d56\u591a\u6b65\u3001\u591a\u8f6e\u5de5\u5177\u8c03\u7528\u5de5\u4f5c\u6d41\uff0cLLM \u670d\u52a1\u4e2d\u7684\u63a8\u7406\u5ef6\u8fdf\u663e\u8457\u589e\u957f\u3002\u4e3a\u5e94\u5bf9\u8fd9\u4e00\u95ee\u9898\uff0c\u8fd1\u671f\u7814\u7a76\u5229\u7528\u6709\u5411\u65e0\u73af\u56fe\uff08DAGs\uff09\u5efa\u6a21\u5de5\u5177\u95f4\u4f9d\u8d56\u5173\u7cfb\uff0c\u4f7f\u5f7c\u6b64\u72ec\u7acb\u7684 API \u8c03\u7528\u80fd\u591f\u5e76\u884c\u6267\u884c\uff0c\u4ece\u800c\u964d\u4f4e\u5ef6\u8fdf\u3002\u53e6\u4e00\u7c7b\u5de5\u4f5c\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u4e3b\u52a8\u6267\u884c\u5de5\u5177\u8c03\u7528\uff0c\u4ee5\u6700\u5c0f\u5316\u7cfb\u7edf\u7a7a\u95f2\u65f6\u95f4\u3002\u4e0e\u5148\u524d\u805a\u7126\u5de5\u5177\u6267\u884c\u7684\u65b9\u6cd5\u4e0d\u540c\uff0c\u672c\u6587\u5c06\u6ce8\u610f\u529b\u8f6c\u5411\u52a0\u901f\u5de5\u5177\u8c03\u7528\u751f\u6210\uff0c\u4ee5\u8fdb\u4e00\u6b65\u63d0\u5347\u5de5\u5177\u4f7f\u7528\u578b LLM \u7684\u63a8\u7406\u6548\u7387\u3002<\/p>\n<h5>\u63a8\u6d4b\u89e3\u7801<\/h5>\n<p>\u8fd1\u5e74\u6765\uff0c\u63a8\u6d4b\u89e3\u7801\uff08SD\uff09\u5df2\u6210\u4e3a\u52a0\u901f LLM \u63a8\u7406\u7684\u5e7f\u6cdb\u91c7\u7528\u8303\u5f0f\u3002\u5728\u6bcf\u4e2a\u89e3\u7801\u6b65\uff0cSD \u9996\u5148\u751f\u6210\u4e00\u6bb5\u8349\u7a3f token \u5e8f\u5217\uff0c\u968f\u540e\u7531\u76ee\u6807 LLM \u5e76\u884c\u9a8c\u8bc1\u3002SD \u7684\u6700\u65b0\u8fdb\u5c55\uff0c\u4f8b\u5982 Eagle \u7cfb\u5217\uff0c\u5728\u76ee\u6807 LLM \u4e4b\u4e0a\u5f15\u5165\u8f7b\u91cf\u8349\u7a3f\u6a21\u578b\uff0c\u53d6\u5f97\u4e86\u6709\u524d\u666f\u7684\u52a0\u901f\u6548\u679c\u3002\u5c3d\u7ba1\u8fd9\u4e9b\u65b9\u6cd5\u9700\u8981\u989d\u5916\u8bad\u7ec3\u6216\u53c2\u6570\uff0c\u53e6\u4e00\u6761\u7814\u7a76\u8def\u7ebf\u805a\u7126\u4e8e\u65e0\u9700\u8bad\u7ec3\u7684 SD \u65b9\u6cd5\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cToken Recycling \u5c06\u5148\u524d\u6b65\u9aa4\u7684 top-<span class=\"katex-eq\" data-katex-display=\"false\">k<\/span> \u5019\u9009 token \u5b58\u5165\u90bb\u63a5\u77e9\u9635\uff0c\u5e76\u4f7f\u7528\u5e7f\u5ea6\u4f18\u5148\u641c\u7d22\u6784\u9020\u8349\u7a3f\u6811\u3002\u7c7b\u4f3c\u5730\uff0cSAM-Decoding \u5229\u7528\u540e\u7f00\u81ea\u52a8\u673a\uff0c\u4ece\u8f93\u5165\u4e0a\u4e0b\u6587\u548c\u9759\u6001\u6587\u672c\u8bed\u6599\u4e2d\u9ad8\u6548\u68c0\u7d22\u9ad8\u8d28\u91cf\u8349\u7a3f\u3002<\/p>\n<h4>\u9884\u5907\u7814\u7a76<\/h4>\n<p>\u672c\u8282\u5bf9\u5de5\u5177\u8c03\u7528\u751f\u6210\u8fdb\u884c\u5b9e\u8bc1\u5206\u6790\uff0c\u6db5\u76d6\uff1a\uff08i\uff09\u751f\u6210\u4e0e\u5de5\u5177\u6267\u884c\u4e4b\u95f4\u7684\u5ef6\u8fdf\u5206\u89e3\uff1b\uff08ii\uff09\u5de5\u5177\u8c03\u7528\u6a21\u5f0f\u5206\u6790\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239002.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 2\uff1aQwen2.5-Instruct \u6a21\u578b\u7cfb\u5217\u5728 ToolBench \u4e0a\u7684\u5ef6\u8fdf\u5206\u89e3\u3002\u5de5\u5177\u6267\u884c\u5ef6\u8fdf\u57fa\u672c\u4fdd\u6301\u6052\u5b9a\uff0c\u800c\u5de5\u5177\u8c03\u7528\u751f\u6210\u5ef6\u8fdf\u968f\u6a21\u578b\u89c4\u6a21\u589e\u957f\uff0c\u6700\u9ad8\u53ef\u8fbe\u7aef\u5230\u7aef\u5ef6\u8fdf\u7684 <span class=\"katex-eq\" data-katex-display=\"false\">96\\%<\/span>\u3002<\/center><\/p>\n<h5>\u5ef6\u8fdf\u5206\u89e3<\/h5>\n<p>\u6211\u4eec\u9996\u5148\u5206\u6790\u5de5\u5177\u8c03\u7528\u4e2d\u7684\u5ef6\u8fdf\u74f6\u9888\u3002\u5b9e\u9a8c\u4f7f\u7528 Qwen2.5-Instruct \u7cfb\u5217\u5728 ToolBench \u4e0a\u8fdb\u884c\u3002\u5982\u56fe 2 \u6240\u793a\uff0c\u5de5\u5177\u6267\u884c\u5ef6\u8fdf\u5728\u4e0d\u540c\u6a21\u578b\u89c4\u6a21\u4e0b\u51e0\u4e4e\u4fdd\u6301\u6052\u5b9a\uff0c\u800c\u5de5\u5177\u8c03\u7528\u751f\u6210\u5ef6\u8fdf\u663e\u8457\u589e\u52a0\uff0c\u5bf9\u4e8e Qwen2.5-72B-Instruct \u6700\u9ad8\u5360\u7aef\u5230\u7aef\u5ef6\u8fdf\u7684 <span class=\"katex-eq\" data-katex-display=\"false\">96\\%<\/span>\u3002\u8fd9\u4e9b\u7ed3\u679c\u8868\u660e\uff0c\u5de5\u5177\u8c03\u7528\u751f\u6210\u662f\u5de5\u5177\u8c03\u7528\u6d41\u6c34\u7ebf\u4e2d\u7684\u4e3b\u5bfc\u74f6\u9888\uff0c\u5c24\u5176\u5728\u591a\u5de5\u5177\u3001\u591a\u8f6e\u573a\u666f\u4e2d\uff0c\u8fd9\u51f8\u663e\u4e86\u52a0\u901f\u751f\u6210\u7684\u91cd\u8981\u6027\u3002<\/p>\n<h5>\u5de5\u5177\u8c03\u7528\u6a21\u5f0f\u5206\u6790<\/h5>\n<p>\u56fe 3 \u5c55\u793a\u4e86\u6765\u81ea API-Bank \u7684\u4e00\u4e2a\u5de5\u5177\u8c03\u7528\u8f68\u8ff9\u793a\u4f8b\u3002\u5982\u56fe\u6240\u793a\uff0c\u8fd9\u4e9b\u8f68\u8ff9\u662f\u9075\u5faa\u53d7\u7ea6\u675f schema \u7684\u9ad8\u5ea6\u7ed3\u6784\u5316 JSON \u5bf9\u8c61\uff0c\u7531\u4ee5\u4e0b\u90e8\u5206\u7ec4\u6210\uff1a<\/p>\n<ul>\n<li><strong>schema tokens<\/strong>\uff0c\u4f8b\u5982 <code><code> <\/code>name<code> <\/code><code>\u3001<\/code><code> <\/code>parameters<code> <\/code><\/code> \u4ee5\u53ca\u6807\u70b9\u7b26\u53f7\uff0c\u5b83\u4eec\u6784\u6210\u5de5\u5177\u8c03\u7528 schema \u6240\u9700\u7684\u56fa\u5b9a\u9aa8\u67b6\uff1b<\/li>\n<li><strong>\u5de5\u5177\u540d\u4e0e\u53c2\u6570\u540d<\/strong>\uff0c\u5206\u522b\u4ee5\u7ea2\u8272\u548c\u84dd\u8272\u9ad8\u4eae\uff0c\u7528\u4e8e\u6307\u5b9a\u76ee\u6807\u5de5\u5177\u53ca\u5176\u53c2\u6570\uff1b<\/li>\n<li><strong>\u53c2\u6570\u503c<\/strong>\uff0c\u4ee5\u7eff\u8272\u663e\u793a\uff0c\u7528\u4e8e\u6839\u636e\u67e5\u8be2\u7279\u5b9a\u5185\u5bb9\u5b9e\u4f8b\u5316\u53c2\u6570\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u7ed3\u6784\u5316\u6a21\u5f0f\u4f7f\u5de5\u5177\u8c03\u7528\u751f\u6210\u533a\u522b\u4e8e\u4e00\u822c\u81ea\u7136\u8bed\u8a00\u751f\u6210\uff0c\u8868\u660e\u5b83\u53ef\u4ee5\u88ab\u5f62\u5f0f\u5316\u4e3a\u53d7\u7ea6\u675f\u89e3\u7801\uff08constrained decoding\uff09\u95ee\u9898\u3002\u5728\u5b9e\u8df5\u4e2d\uff0c\u6709\u6548\u5de5\u5177\u8c03\u7528\u5fc5\u987b\u9075\u5faa\u4e25\u683c JSON schema\uff0c\u8fd9\u610f\u5473\u7740\u8f93\u51fa tokens \u7684\u5f88\u5927\u4e00\u90e8\u5206\uff0c\u5373 schema tokens\uff0c\u5b9e\u9645\u4e0a\u662f\u9884\u5148\u786e\u5b9a\u7684\u3002\u4e00\u65e6\u9009\u5b9a\u76ee\u6807\u5de5\u5177\uff0c\u5176\u53c2\u6570\u540d\u5c31\u6765\u81ea\u6709\u9650\u3001\u7531 schema \u5b9a\u4e49\u7684\u96c6\u5408\uff08\u4f8b\u5982 <code><code> <\/code>user_id<code> <\/code><code>\u3001<\/code><code> <\/code>time<code> <\/code><code> \u548c <\/code><code> <\/code>health_data<code> <\/code><\/code>\uff09\u3002\u56e0\u6b64\uff0c\u6a21\u578b\u53ea\u9700\u8981\u751f\u6210\u4e00\u5c0f\u90e8\u5206 tokens\uff0c\u5373\u53c2\u6570\u503c\uff0c\u5e76\u4ee5\u7528\u6237\u67e5\u8be2\u548c\u5bf9\u8bdd\u4e0a\u4e0b\u6587\u4e3a\u6761\u4ef6\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239003.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 3\uff1a\u6765\u81ea API-Bank \u7684\u5de5\u5177\u8c03\u7528\u8f68\u8ff9\u793a\u4f8b\uff0c\u4e0d\u540c\u5b57\u6bb5\u4ee5\u989c\u8272\u9ad8\u4eae\u663e\u793a\u3002<\/center><\/p>\n<p>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u8fdb\u4e00\u6b65\u5206\u6790 LLM \u4e2d\u5b9e\u9645\u7684\u5de5\u5177\u8c03\u7528\u884c\u4e3a\uff0c\u5305\u62ec schema \u9075\u5faa\u5ea6\u4e0e\u91cd\u590d\u5de5\u5177\u8c03\u7528\u6a21\u5f0f\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239004.png\" alt=\"\" \/><\/p>\n<p><center>\u8868 1\uff1a\u4e0d\u540c\u6a21\u578b\u7684\u683c\u5f0f\u9075\u5faa\u5ea6\u3002<\/center><\/p>\n<h6>Schema \u9075\u5faa\u5ea6<\/h6>\n<p>\u5982\u8868 1 \u6240\u793a\uff0c\u5f53\u524d\u5de5\u5177\u4f7f\u7528\u578b LLM \u5bf9\u5de5\u5177\u8c03\u7528\u683c\u5f0f\u8868\u73b0\u51fa\u5f88\u5f3a\u7684\u9075\u5faa\u6027\uff0c\u5373\u5176\u8f93\u51fa\u53ef\u4ee5\u88ab\u53ef\u9760\u89e3\u6790\u4e3a\u6709\u6548 JSON\u3002LLaMA-Instruct \u548c Qwen2.5-Instruct \u4e24\u4e2a\u6a21\u578b\u5bb6\u65cf\u5728\u9075\u5faa\u6240\u9700\u5de5\u5177\u8c03\u7528 schema \u65b9\u9762\u90fd\u8d85\u8fc7 <span class=\"katex-eq\" data-katex-display=\"false\">99\\%<\/span> \u7684\u51c6\u786e\u7387\u3002\u8fd9\u4e00\u89c2\u5bdf\u8fdb\u4e00\u6b65\u652f\u6301\u4e86\u6211\u4eec\u65e9\u5148\u7684\u4e3b\u5f20\uff1a\u5de5\u5177\u8c03\u7528\u751f\u6210\u53ef\u4ee5\u5728\u9884\u5b9a\u4e49 schema \u4e0b\u88ab\u5f62\u5f0f\u5316\u4e3a\u53d7\u7ea6\u675f\u89e3\u7801\u3002\u9274\u4e8e\u6b64\uff0c\u4e3a\u4e00\u822c\u6587\u672c\u751f\u6210\u8bbe\u8ba1\u7684\u52a0\u901f\u65b9\u6cd5\u5bf9\u4e8e\u5de5\u5177\u8c03\u7528\u53ef\u80fd\u5e76\u975e\u6700\u4f18\u3002\u76f8\u53cd\uff0c\u663e\u5f0f\u5229\u7528\u5de5\u5177\u8c03\u7528\u751f\u6210\u7684\u53d7\u7ea6\u675f\u7ed3\u6784\u53ef\u4ee5\u5e26\u6765\u989d\u5916\u6548\u7387\u6536\u76ca\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239005.png\" alt=\"\" \/><\/p>\n<p><center>\u8868 2\uff1a\u5de5\u5177\u8c03\u7528\u57fa\u51c6\u7684\u6570\u636e\u96c6\u7edf\u8ba1\u3002<\/center><\/p>\n<h6>\u91cd\u590d\u5de5\u5177\u8c03\u7528<\/h6>\n<p>\u9664 schema \u9075\u5faa\u5ea6\u5916\uff0c\u6211\u4eec\u8fd8\u89c2\u5bdf\u5230 LLM \u9891\u7e41\u8c03\u7528\u76f8\u540c\u5de5\u5177\u4ee5\u6ee1\u8db3\u91cd\u590d\u51fa\u73b0\u7684\u7528\u6237\u610f\u56fe\u548c\u504f\u597d\u6a21\u5f0f\u3002\u8fd9\u79cd\u884c\u4e3a\u5728\u5e7f\u6cdb\u4f7f\u7528\u7684\u5de5\u5177\u8c03\u7528\u57fa\u51c6\u4e2d\u4e00\u81f4\u51fa\u73b0\u3002\u5982\u8868 2 \u6240\u793a\uff0c\u5728 API-Bank \u4e2d\uff0c\u5bf9\u76f8\u540c\u5de5\u5177\u7684\u91cd\u590d\u8c03\u7528\u975e\u5e38\u9891\u7e41\uff0c\u6bcf\u4e2a\u5de5\u5177\u7684\u5e73\u5747\u8c03\u7528\u6b21\u6570\u8d85\u8fc7 <span class=\"katex-eq\" data-katex-display=\"false\">10.95<\/span>\u3002\u8fd9\u4e00\u89c2\u5bdf\u4fc3\u4f7f\u6211\u4eec\u5229\u7528\u5386\u53f2\u5de5\u5177\u8c03\u7528\u8f68\u8ff9\u3002\u7cfb\u7edf\u4e0d\u5fc5\u4e22\u5f03\u5148\u524d\u8c03\u7528\uff0c\u800c\u662f\u5728\u5148\u524d\u8c03\u7528\u4ecd\u7136\u6709\u6548\u65f6\u590d\u7528\u5b83\u4eec\uff0c\u4ece\u800c\u8fdb\u4e00\u6b65\u63d0\u5347\u5de5\u5177\u8c03\u7528\u6548\u7387\u3002<\/p>\n<h4>ToolSpec<\/h4>\n<p>\u672c\u8282\u4ecb\u7ecd ToolSpec\uff0c\u4e00\u79cd\u7528\u4e8e\u52a0\u901f\u5de5\u5177\u8c03\u7528\u751f\u6210\u7684\u5373\u63d2\u5373\u7528 SD \u65b9\u6cd5\u3002\u8be5\u65b9\u6cd5\u5229\u7528\u5de5\u5177\u8c03\u7528\u7684\u9ad8\u5ea6\u7ed3\u6784\u5316\u7279\u6027\u4ee5\u53ca\u91cd\u590d\u51fa\u73b0\u7684\u5386\u53f2\u6a21\u5f0f\uff0c\u4ee5\u53ef\u5ffd\u7565\u7684\u5f00\u9500\u6784\u9020\u9ad8\u8d28\u91cf\u8349\u7a3f\u3002\u5982\u56fe 5 \u6240\u793a\uff0cToolSpec \u96c6\u6210\u4e24\u79cd\u4e92\u8865\u7684\u8349\u7a3f\u7b56\u7565\uff1aschema \u611f\u77e5\u8349\u7a3f\u751f\u6210\u4e0e\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\uff0c\u4e0b\u6587\u5c06\u8be6\u7ec6\u4ecb\u7ecd\u3002<\/p>\n<h5>\u4efb\u52a1\u5f62\u5f0f\u5316<\/h5>\n<p>\u7ed9\u5b9a\u7528\u6237\u67e5\u8be2\u4ee5\u53ca\u5305\u542b\u7cfb\u7edf\u63d0\u793a\u548c\u5de5\u5177\u6587\u6863\u7684\u8f93\u5165\u4e0a\u4e0b\u6587\uff0cToolSpec \u4ece\u6587\u6863\u4e2d\u63d0\u53d6\u5de5\u5177\u89c4\u8303\uff0c\u5e76\u5c06\u5176\u8f6c\u6362\u4e3a\u7ed3\u6784\u5316 schema\u3002\u8be5 schema \u5b9a\u4e49\uff1a\uff08i\uff09schema token \u96c6\u5408 <span class=\"katex-eq\" data-katex-display=\"false\">\\Sigma_{\\text{schema}}<\/span>\uff1b\uff08ii\uff09\u53ef\u8c03\u7528\u5de5\u5177\u96c6\u5408 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{C}_{\\text{tools}}<\/span>\uff08\u4f8b\u5982 <code>ForgotPassword<\/code>\uff09\uff1b\uff08iii\uff09\u4e0e\u6bcf\u4e2a\u5de5\u5177\u5173\u8054\u7684\u53c2\u6570\u5b57\u6bb5\uff08\u4f8b\u5982 <code>Status<\/code>\u3001<code>Username<\/code>\uff09\u3002<\/p>\n<p>\u8bbe <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span> \u8868\u793a\u76ee\u6807\u6a21\u578b\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{x}<\/span> \u8868\u793a\u8f93\u5165\u4e0a\u4e0b\u6587\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{y} = (y_1, y_2, \\ldots, y_L)<\/span> \u8868\u793a\u5f85\u751f\u6210\u7684\u5de5\u5177\u8c03\u7528\u5e8f\u5217\u3002\u5728\u89e3\u7801\u6b65 <span class=\"katex-eq\" data-katex-display=\"false\">t<\/span>\uff0c\u9075\u5faa\u6807\u51c6 SD \u8303\u5f0f\uff0cToolSpec \u9996\u5148\u63d0\u51fa\u8349\u7a3f token \u5e8f\u5217 <span class=\"katex-eq\" data-katex-display=\"false\">\\tilde{\\boldsymbol{y}}_{t:t+\\gamma-1} = (\\tilde{y}_{t}, \\tilde{y}_{t+1}, \\ldots, \\tilde{y}_{t+\\gamma-1})<\/span>\uff0c\u4f5c\u4e3a\u540e\u7eed\u89e3\u7801\u6b65\u7684\u9884\u6d4b\u3002\u76ee\u6807\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span> \u968f\u540e\u5e76\u884c\u9a8c\u8bc1\u8fd9\u4e9b\u8349\u7a3f\u5019\u9009\uff0c\u5e76\u63a5\u53d7\u4e0e\u5176\u6982\u7387\u5206\u5e03\u4e00\u81f4\u7684\u6700\u957f\u524d\u7f00\u3002\u5f53\u8349\u7a3f\u751f\u6210\u8fc7\u7a0b\u7684\u5ef6\u8fdf\u663e\u8457\u4f4e\u4e8e\u9a8c\u8bc1\u65f6\uff0c\u6700\u5927\u5316 <span class=\"katex-eq\" data-katex-display=\"false\">\\tilde{\\boldsymbol{y}}<\/span> \u7684\u63a5\u53d7\u957f\u5ea6\u53ef\u4ee5\u76f4\u63a5\u63d0\u5347\u6709\u6548\u89e3\u7801\u5e76\u884c\u5ea6\uff0c\u5e76\u964d\u4f4e\u7aef\u5230\u7aef\u5ef6\u8fdf\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239006.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 4\uff1aToolSpec \u4e2d\u72b6\u6001\u8f6c\u79fb\u793a\u610f\u56fe\u3002\u4e00\u65e6\u7531 <code>&lt;tool_call&gt;<\/code> \u89e6\u53d1\u5de5\u5177\u8c03\u7528\uff0cFSM \u8fdb\u5165\u5de5\u5177\u540d\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">q_t<\/span>\uff0c\u968f\u540e\u5728\u53c2\u6570\u540d\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">q_p<\/span> \u4e0e\u53c2\u6570\u503c\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">q_v<\/span> \u4e4b\u95f4\u4ea4\u66ff\uff0c\u76f4\u5230\u5de5\u5177\u8c03\u7528\u5b8c\u6210\u3002<\/center><\/p>\n<h5>Schema \u611f\u77e5\u8349\u7a3f\u751f\u6210<\/h5>\n<p>\u5982\u7b2c 3.2 \u8282\u6240\u8ba8\u8bba\uff0c\u5148\u8fdb LLM \u7684\u5de5\u5177\u8c03\u7528\u9075\u5faa\u9884\u5b9a\u4e49 schema \u683c\u5f0f\u3002\u8fd9\u79cd\u7ed3\u6784\u89c4\u5f8b\u6027\u4f7f\u8f93\u51fa\u7684\u5927\u90e8\u5206\u53ef\u4ee5\u8fdb\u884c\u786e\u5b9a\u6027\u8349\u7a3f\u751f\u6210\u3002\u4e3a\u5229\u7528\u8fd9\u4e00\u6027\u8d28\uff0cToolSpec \u6839\u636e\u89e3\u6790\u5f97\u5230\u7684\u5de5\u5177 schema \u6784\u9020\u6709\u9650\u72b6\u6001\u673a\uff08FSM\uff09\u3002\u5f62\u5f0f\u4e0a\uff0cFSM \u5b9a\u4e49\u4e3a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\mathcal{F} = (\\mathcal{Q}, \\Sigma_{\\text{schema}}, \\delta, q_0),\n\n<\/span>\n<p>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{Q}<\/span> \u662f\u72b6\u6001\u96c6\u5408\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\Sigma_{\\text{schema}}<\/span> \u662f schema token \u5b57\u6bcd\u8868\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\delta: \\mathcal{Q} \\times \\Sigma_{\\text{schema}} \\rightarrow \\mathcal{Q}<\/span> \u662f\u8f6c\u79fb\u51fd\u6570\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">q_0 \\in \\mathcal{Q}<\/span> \u662f\u521d\u59cb\u72b6\u6001\u3002<\/p>\n<p>\u72b6\u6001\u7a7a\u95f4 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{Q}<\/span> \u5c06\u5de5\u5177\u8c03\u7528\u751f\u6210\u8fc7\u7a0b\u5206\u89e3\u4e3a\u7ed3\u6784\u7ec4\u4ef6\uff1a<\/p>\n<ul>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">q_t<\/span>\uff1a<strong>\u5de5\u5177\u540d\u72b6\u6001<\/strong>\uff0c\u6a21\u578b\u4ece\u9884\u5b9a\u4e49\u5de5\u5177\u96c6 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{C}_{\\text{tools}}<\/span> \u4e2d\u9009\u62e9\u4e00\u4e2a\u5de5\u5177\u3002<\/li>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">q_p<\/span>\uff1a<strong>\u53c2\u6570\u540d\u72b6\u6001<\/strong>\uff0c\u6a21\u578b\u4e3a\u6240\u9009\u5de5\u5177\u9009\u62e9\u4e00\u4e2a\u53c2\u6570\u5b57\u6bb5\u3002<\/li>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">q_v<\/span>\uff1a<strong>\u53c2\u6570\u503c\u72b6\u6001<\/strong>\uff0c\u6a21\u578b\u751f\u6210\u6240\u9009\u53c2\u6570\u7684\u503c\u3002<\/li>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">q_o<\/span>\uff1a<strong>\u201c\u5176\u4ed6\u201d\u72b6\u6001<\/strong>\uff0c\u5bf9\u5e94\u5de5\u5177\u8c03\u7528\u4e4b\u5916\u7684\u4e00\u822c\u81ea\u7136\u8bed\u8a00\u751f\u6210\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239007.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 5\uff1a\u4f7f\u7528 ToolSpec \u8fdb\u884c\u5de5\u5177\u8c03\u7528\u751f\u6210\u7684\u793a\u610f\u56fe\u3002\u4e3a\u52a0\u901f\u7ed3\u6784\u5316\u5de5\u5177\u8c03\u7528\uff0cToolSpec \u96c6\u6210\u4e24\u79cd\u4e92\u8865\u8349\u7a3f\u7b56\u7565\uff1a\uff08a\uff09schema \u611f\u77e5\u8349\u7a3f\u751f\u6210\uff0c\u5176\u4e2d\u6709\u9650\u72b6\u6001\u673a\u5728\u9884\u5b9a\u4e49\u5de5\u5177 schema \u5185\u5f15\u5bfc\u9ad8\u7cbe\u5ea6\u8349\u7a3f\u751f\u6210\uff1b\uff08b\uff09\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\uff0c\u590d\u7528\u5386\u53f2\u751f\u6210\u6a21\u5f0f\u4f5c\u4e3a\u9ad8\u8d28\u91cf\u8349\u7a3f\uff0c\u4ee5\u52a0\u901f\u91cd\u590d\u51fa\u73b0\u7684\u5de5\u5177\u8c03\u7528\u3002<\/center><\/p>\n<p>\u5982\u56fe 4 \u6240\u793a\uff0c\u4e00\u65e6\u751f\u6210\u7279\u6b8a token <code>&lt;tool_call&gt;<\/code>\uff0cFSM \u8f6c\u79fb\u5230\u5de5\u5177\u540d\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">q_t<\/span>\u3002\u5728\u8be5\u72b6\u6001\u4e0b\uff0c\u901a\u8fc7\u53d7\u7ea6\u675f\u89e3\u7801\u6267\u884c\u8349\u7a3f\u751f\u6210\u4ee5\u6784\u9020\u5019\u9009\u8349\u7a3f\u5e8f\u5217\u3002\u6211\u4eec\u5c06\u63d0\u51fa\u7684\u8349\u7a3f\u5e8f\u5217 <span class=\"katex-eq\" data-katex-display=\"false\">\\tilde{\\boldsymbol{y}}<\/span> \u5f62\u5f0f\u5316\u4e3a\u5fc5\u8981\u7ed3\u6784\u5143\u7d20\u7684\u62fc\u63a5\uff08\u8bb0\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">\\oplus<\/span>\uff09\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\tilde{\\boldsymbol{y}} = \\tilde{\\boldsymbol{y}}_{\\text{prefix}} \\oplus \\tilde{\\boldsymbol{y}}_{\\text{tool\\_name}} \\oplus \\tilde{\\boldsymbol{y}}_{\\text{suffix}}\n\n<\/span>\n<p>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\tilde{\\boldsymbol{y}}_{\\text{prefix}}<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">\\tilde{\\boldsymbol{y}}_{\\text{suffix}}<\/span> \u8868\u793a\u4e0e <span class=\"katex-eq\" data-katex-display=\"false\">q_t<\/span> \u5173\u8054\u7684 schema \u524d\u7f00\u548c\u540e\u7f00\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\tilde{\\boldsymbol{y}}_{\\text{tool\\_name}} \\in \\mathcal{C}_{\\text{tools}}<\/span> \u8868\u793a\u5019\u9009\u5de5\u5177\u540d\u3002\u6240\u6709\u6f5c\u5728\u8349\u7a3f\u968f\u540e\u7531\u76ee\u6807\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span> \u5e76\u884c\u9a8c\u8bc1\u3002<\/p>\n<p>\u9a8c\u8bc1\u6210\u529f\u540e\uff0cFSM \u8f6c\u79fb\u5230\u53c2\u6570\u540d\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">q_p<\/span>\uff0c\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span> \u5e76\u884c\u9a8c\u8bc1\u4e0e\u6240\u9009\u5de5\u5177\u5173\u8054\u7684\u5019\u9009\u53c2\u6570\u540d\u3002\u968f\u540e FSM \u8f6c\u79fb\u5230\u53c2\u6570\u503c\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">q_v<\/span>\u3002\u7531\u4e8e\u53c2\u6570\u503c\u901a\u5e38\u4e0d\u53d7\u7ea6\u675f\uff0cToolSpec \u5728\u6b64\u9636\u6bb5\u4e0d\u518d\u65bd\u52a0 schema \u7ea6\u675f\uff0c\u800c\u662f\u4f9d\u8d56\u901a\u7528 SD \u7b97\u6cd5\u6765\u9884\u6d4b\u8fd9\u4e9b\u53ef\u53d8\u5b57\u6bb5\u3002<\/p>\n<p>\u4e4b\u540e\uff0cFSM \u6839\u636e schema \u7279\u5b9a\u5206\u9694\u7b26\u5728\u72b6\u6001\u4e4b\u95f4\u8f6c\u79fb\u3002\u4f8b\u5982\uff0c\u5206\u9694\u7b26 <code><code> <\/code>\",<code> <\/code><code> \u4f1a\u7ec8\u6b62\u53c2\u6570\u503c\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">q_v<\/span> \u5e76\u8fd4\u56de <span class=\"katex-eq\" data-katex-display=\"false\">q_p<\/span> \u4ee5\u5904\u7406\u4e0b\u4e00\u4e2a\u53c2\u6570\uff1b\u800c\u5206\u9694\u7b26 <\/code><code> <\/code>\"}}\"<code> <\/code><\/code> \u5219\u6807\u8bb0\u5de5\u5177\u8c03\u7528\u5b8c\u6210\u3002\u56e0\u6b64\uff0cToolSpec \u80fd\u591f\u5728\u7ed3\u6784\u72b6\u6001\uff08\u4f8b\u5982 <span class=\"katex-eq\" data-katex-display=\"false\">q_t<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">q_p<\/span>\uff09\u7684\u786e\u5b9a\u6027 schema token \u586b\u5145\u4e0e\u5f00\u653e\u72b6\u6001\uff08\u4f8b\u5982 <span class=\"katex-eq\" data-katex-display=\"false\">q_v<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">q_o<\/span>\uff09\u7684\u63a8\u6d4b\u751f\u6210\u4e4b\u95f4\u65e0\u7f1d\u4ea4\u66ff\u3002<\/p>\n<h5>\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b<\/h5>\n<p>\u867d\u7136 schema \u5b9a\u4e49\u4e86\u5de5\u5177\u8c03\u7528\u7684\u7ed3\u6784\u9aa8\u67b6\uff0c\u5e76\u652f\u6301\u5bf9\u5de5\u5177\u540d\u7b49\u53d7\u7ea6\u675f\u5b57\u6bb5\u8fdb\u884c\u5c40\u90e8\u8349\u7a3f\u751f\u6210\uff0c\u4f46\u53ef\u53d8\u5b57\u6bb5\uff08\u5c24\u5176\u662f\u53c2\u6570\u503c\uff09\u9ad8\u5ea6\u4f9d\u8d56\u8f93\u5165\u4e0a\u4e0b\u6587\uff0c\u65e0\u6cd5\u4ec5\u51ed schema \u63a8\u65ad\u3002\u53d7\u5de5\u5177\u8c03\u7528\u4e2d\u91cd\u590d\u6a21\u5f0f\u542f\u53d1\uff0cToolSpec \u5f15\u5165\u68c0\u7d22\u589e\u5f3a\u8349\u7a3f\u7b56\u7565\uff0c\u5728\u5e8f\u5217\u5c42\u9762\u52a0\u901f\u5de5\u5177\u8c03\u7528\u751f\u6210\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239008.png\" alt=\"\" \/><\/p>\n<p><center>\u8868 3\uff1aToolSpec \u4e0e\u5148\u524d\u5373\u63d2\u5373\u7528\u65b9\u6cd5\u7684\u63a8\u7406\u6548\u7387\u6bd4\u8f83\u3002\u6211\u4eec\u62a5\u544a\u5e73\u5747\u63a5\u53d7 token \u6570\uff08#MAT\uff09\u3001\u5e73\u5747\u89e3\u7801\u901f\u5ea6\uff08tokens\/s\uff09\u4ee5\u53ca\u76f8\u5bf9\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\u7684\u5899\u949f\u52a0\u901f\u6bd4\u3002\u6700\u4f18\u7ed3\u679c\u52a0\u7c97\uff0c\u6b21\u4f18\u7ed3\u679c\u52a0\u4e0b\u5212\u7ebf\u3002Eagle \u7cfb\u5217\u4e3a\u9700\u8981\u8bad\u7ec3\u7684 SD \u65b9\u6cd5\uff0c\u4ec5\u4f5c\u53c2\u8003\u3002<\/center><\/p>\n<p>\u6211\u4eec\u7ef4\u62a4\u4e00\u4e2a\u6570\u636e\u5b58\u50a8 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{H}=\\{(\\boldsymbol{h}_j, \\boldsymbol{y}_j)\\}_{j=1}^{|\\mathcal{H}|}<\/span>\uff0c\u5176\u4e2d\u4fdd\u5b58\u5148\u524d\u6210\u529f\u7684\u5de5\u5177\u8c03\u7528\uff1b<span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{h}_j \\in \\mathbb{R}^d<\/span> \u8868\u793a\u8f93\u5165\u95ee\u9898 <span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{x}_j<\/span> \u5728\u5176\u6700\u540e\u4e00\u4e2a token \u5904\u7684\u9690\u85cf\u8868\u793a\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">d<\/span> \u662f\u9690\u85cf\u7ef4\u5ea6\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{y}_j<\/span> \u662f\u5bf9\u5e94\u7684\u5de5\u5177\u8c03\u7528\u8f93\u51fa\u3002\u7ed9\u5b9a\u67e5\u8be2 <span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{x}_i<\/span>\uff0c\u6211\u4eec\u68c0\u7d22 top-<span class=\"katex-eq\" data-katex-display=\"false\">k<\/span> \u4e2a\u6700\u76f8\u4f3c\u7684\u5386\u53f2\u5de5\u5177\u8c03\u7528\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\mathcal{I}_k = \\operatorname{Top-K}_{j \\in \\{1 \\ldots |\\mathcal{H}|\\}} \\text{sim}(\\boldsymbol{h}_i, \\boldsymbol{h}_j)\n\n<\/span>\n<p>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">\\text{sim}(\\cdot)<\/span> \u8868\u793a\u4f59\u5f26\u76f8\u4f3c\u5ea6\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{I}_k<\/span> \u662f\u68c0\u7d22\u5230\u7684\u6761\u76ee\u7d22\u5f15\u3002\u7ed9\u5b9a\u90e8\u5206\u5df2\u751f\u6210\u8f93\u51fa <span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{y}_{\\lt t}<\/span>\uff0c\u6211\u4eec\u901a\u8fc7\u5339\u914d\u540e\u7f00\u5c06\u5176\u4e0e\u6bcf\u4e2a\u68c0\u7d22\u5230\u7684\u5de5\u5177\u8c03\u7528 <span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{y}^{\\prime}<\/span> \u5bf9\u9f50\uff1a<\/p>\n<span class=\"katex-eq\" data-katex-display=\"false\">\n\n\\boldsymbol{y}_{t-L:t-1}=\\boldsymbol{y}^{\\prime}_{m-L:m-1}\n\n<\/span>\n<p>\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">L<\/span> \u8868\u793a\u540e\u7f00\u957f\u5ea6\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">m<\/span> \u8868\u793a <span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{y}^{\\prime}<\/span> \u4e2d\u5bf9\u5e94\u5339\u914d\u4f4d\u7f6e\u3002\u5f53\u627e\u5230\u540e\u7f00\u5339\u914d\u65f6\uff0cToolSpec \u63d0\u53d6\u540e\u7eed <span class=\"katex-eq\" data-katex-display=\"false\">n<\/span> \u4e2a tokens <span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{y}^{\\prime}_{m:m+n-1}<\/span>\uff0c\u5e76\u5c06\u8fd9\u4e9b\u5ef6\u7eed\u4f5c\u4e3a\u8349\u7a3f\u3002\u5982\u56fe 5 \u6240\u793a\uff0c\u82e5\u5f53\u524d\u540e\u7f00\u4e0e\u5305\u542b \u2026<code><code> <\/code>parameters<code> <\/code><code> \u7684\u5386\u53f2\u8c03\u7528\u5339\u914d\uff0c\u5219\u540e\u7eed tokens\uff08\u4f8b\u5982 <\/code>:{<code> <\/code><code> <\/code>status<code> <\/code><code>: <\/code><code> <\/code>Forgot Password<code> <\/code><\/code>\u2026\uff09\u4f1a\u88ab\u63d0\u51fa\u4f5c\u4e3a\u8349\u7a3f\u5019\u9009\u3002\u76f4\u89c2\u5730\u8bf4\uff0c\u4e00\u65e6\u5de5\u5177\u540d\u548c\u952e\u503c\u5bf9\u5df2\u5bf9\u9f50\uff0c\u5de5\u5177\u8c03\u7528\u5269\u4f59\u7ed3\u6784\u5f80\u5f80\u9075\u5faa\u91cd\u590d\u6a21\u5f0f\uff0c\u56e0\u6b64 <span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{y}^{\\prime}_{m:m+n-1}<\/span> \u662f\u540e\u7eed\u503c\u7247\u6bb5\u7684\u9ad8\u8d28\u91cf\u8349\u7a3f\u3002<\/p>\n<h4>\u5b9e\u9a8c<\/h4>\n<h5>\u5b9e\u9a8c\u8bbe\u7f6e<\/h5>\n<h6>\u6a21\u578b\u4e0e\u6570\u636e\u96c6<\/h6>\n<p>\u6211\u4eec\u4e3b\u8981\u4f7f\u7528 Qwen2.5-Instruct \u7cfb\u5217\u3001ToolLLaMA \u7cfb\u5217\u3001LLaMA-3.1-8B-Instruct \u548c LLaMA-3.2-3B-Instruct \u8bc4\u4f30\u65b9\u6cd5\u3002\u8bc4\u4f30\u8986\u76d6\u56db\u4e2a\u5e7f\u6cdb\u4f7f\u7528\u7684\u5de5\u5177\u4f7f\u7528\u57fa\u51c6\uff1aAPI-Bank\u3001ToolAlpaca\u3001BFCLv2 \u548c ToolBench\u3002\u8fd9\u4e9b\u6570\u636e\u96c6\u7684\u66f4\u591a\u7ec6\u8282\u89c1\u9644\u5f55 B.1\u3002<\/p>\n<h6>\u5b9e\u73b0\u7ec6\u8282<\/h6>\n<p>\u68c0\u7d22\u53c2\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">k<\/span> \u8bbe\u4e3a 10\u3002\u5bf9\u4e8e\u540e\u7f00\u5339\u914d\uff0c\u6211\u4eec\u9075\u5faa PLD\uff0c\u8003\u8651\u540e\u7f00\u957f\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">L \\in \\{5,6,7\\}<\/span>\u3002\u5ef6\u7eed\u63d0\u53d6\u4f7f\u7528\u9884\u5b9a\u4e49\u957f\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">n \\in \\{32, 16, 8, 8\\}<\/span>\u3002\u6211\u4eec\u5728 ToolSpec \u4e2d\u91c7\u7528 Token Recycling\uff08TR\uff09\u4f5c\u4e3a\u901a\u7528\u63a8\u6d4b\u7b56\u7565\u3002\u9075\u5faa\u5148\u524d\u5de5\u4f5c\uff0c\u6211\u4eec\u91c7\u7528\u63a8\u6d4b\u91c7\u6837\uff08speculative sampling\uff09\u4f5c\u4e3a\u63a5\u53d7\u7b56\u7565\uff0cbatch size \u4e3a 1\u3002\u66f4\u591a\u5b9e\u73b0\u7ec6\u8282\u89c1\u9644\u5f55 B.2\u3002<\/p>\n<h6>\u57fa\u7ebf\u65b9\u6cd5<\/h6>\n<p>\u6211\u4eec\u4e3b\u8981\u5c06 ToolSpec \u4e0e\u4e09\u79cd\u73b0\u6709\u5373\u63d2\u5373\u7528\u65b9\u6cd5\u6bd4\u8f83\uff1aPLD\u3001TR \u548c SAM-Decoding\uff08SAMD\uff09\u3002\u5bf9\u4e8e SAMD\uff0c\u6211\u4eec\u4ece\u8f93\u5165\u4e0a\u4e0b\u6587\u6784\u9020\u9759\u6001 SAM\uff0c\u5e76\u91c7\u7528 TR \u4f5c\u4e3a\u5176\u89e3\u7801\u7b56\u7565\u3002\u4e3a\u4fbf\u4e8e\u6bd4\u8f83\uff0c\u6211\u4eec\u8fd8\u5728\u4e3b\u8981\u7ed3\u679c\u4e2d\u52a0\u5165\u9700\u8981\u8bad\u7ec3\u7684\u9ad8\u7ea7\u65b9\u6cd5 Eagle-2 \u548c Eagle-3\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239009.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 6\uff1aToolSpec \u4e0e\u5148\u524d\u5373\u63d2\u5373\u7528\u65b9\u6cd5\u5728 ToolBench \u4e0a\u7684\u52a0\u901f\u6bd4\u6bd4\u8f83\u3002<\/center><\/p>\n<h6>\u8bc4\u4f30\u6307\u6807<\/h6>\n<p>\u6211\u4eec\u4f7f\u7528\u4e09\u4e2a\u5e7f\u6cdb\u91c7\u7528\u7684\u6307\u6807\u8bc4\u4f30 ToolSpec\uff1a\u5e73\u5747\u63a5\u53d7 token \u6570\uff08#MAT\uff09\u3001\u541e\u5410\u91cf\uff08tokens\/s\uff09\u4ee5\u53ca\u76f8\u5bf9\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\u7684\u5899\u949f\u52a0\u901f\u6bd4\u3002\u4ece\u7406\u8bba\u89d2\u5ea6\u770b\uff0cToolSpec \u4fdd\u6301\u76ee\u6807 LLM \u7684\u8f93\u51fa\u5206\u5e03\uff0c\u56e0\u6b64\u65e0\u9700\u5bf9\u751f\u6210\u8d28\u91cf\u8fdb\u884c\u5927\u91cf\u8bc4\u4f30\u3002\u4e0d\u8fc7\uff0c\u6211\u4eec\u4ecd\u5728\u9644\u5f55 B.3 \u4e2d\u62a5\u544a API-Bank \u548c ToolAlpaca \u4e0a\u7684\u6027\u80fd\u4ee5\u4f9b\u53c2\u8003\u3002<\/p>\n<h5>\u4e3b\u8981\u7ed3\u679c<\/h5>\n<p>\u8868 3 \u6bd4\u8f83\u4e86 ToolSpec \u4e0e\u5148\u524d\u5373\u63d2\u5373\u7528\u65b9\u6cd5\u5728\u5e7f\u6cdb\u4f7f\u7528\u7684\u5de5\u5177\u4f7f\u7528\u4efb\u52a1\u4e0a\u7684\u63a8\u7406\u6548\u7387\u3002\u7ed3\u679c\u63ed\u793a\u4ee5\u4e0b\u5173\u952e\u53d1\u73b0\uff1a\uff081\uff09ToolSpec \u7a33\u5b9a\u4f18\u4e8e\u73b0\u6709\u65b9\u6cd5\uff0c\u5728\u591a\u79cd\u6a21\u578b\u548c\u4efb\u52a1\u4e0a\u76f8\u5bf9\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\u5b9e\u73b0 <span class=\"katex-eq\" data-katex-display=\"false\">3.5\\times \\sim 4.2\\times<\/span> \u52a0\u901f\u3002\u4e0e\u6b64\u524d\u6700\u4f18\u65b9\u6cd5\u76f8\u6bd4\uff0c\u5176\u76f8\u5bf9\u52a0\u901f\u63d0\u5347\u6700\u9ad8\u8fbe <span class=\"katex-eq\" data-katex-display=\"false\">71\\%<\/span>\uff08\u4ece <span class=\"katex-eq\" data-katex-display=\"false\">2.45\\times<\/span> \u5230 <span class=\"katex-eq\" data-katex-display=\"false\">4.19\\times<\/span>\uff09\u3002\uff082\uff09ToolSpec \u7684\u6548\u7387\u6536\u76ca\u4e3b\u8981\u6765\u81ea\u8f83\u9ad8\u7684\u63a5\u53d7 token \u6570\uff08\u8bb0\u4e3a #MAT\uff09\uff0c\u5728\u4e0d\u540c\u8bbe\u7f6e\u4e0b\u8303\u56f4\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">4.02<\/span> \u5230 <span class=\"katex-eq\" data-katex-display=\"false\">5.80<\/span>\u3002\u4e0e Eagle-3 \u7b49\u57fa\u4e8e\u6a21\u578b\u7684 SD \u65b9\u6cd5\u4e0d\u540c\uff0cToolSpec \u4e0d\u9700\u8981\u989d\u5916\u8349\u7a3f\u6a21\u5757\uff0c\u4e14\u4ec5\u5f15\u5165\u5f88\u5c0f\u7684\u8349\u7a3f\u5f00\u9500\uff08\u7ea6 <span class=\"katex-eq\" data-katex-display=\"false\">6\\%<\/span>\uff0c\u5982\u56fe 11 \u6240\u793a\uff09\u3002\u8fd9\u4e00\u4f4e\u5f00\u9500\u4f7f\u66f4\u9ad8\u7684 #MAT \u80fd\u66f4\u6709\u6548\u5730\u8f6c\u5316\u4e3a\u7aef\u5230\u7aef\u52a0\u901f\u3002<\/p>\n<h6>\u5173\u4e8e Eagle \u8868\u73b0\u7684\u8ba8\u8bba<\/h6>\n<p>\u6211\u4eec\u89c2\u5bdf\u5230\uff0c\u5982\u8868 3 \u6240\u793a\uff0cEagle \u7cfb\u5217\u8fbe\u5230\u7684 #MAT \u4f4e\u4e8e TR \u7b49\u5148\u524d\u5373\u63d2\u5373\u7528\u65b9\u6cd5\u3002\u6211\u4eec\u5047\u8bbe\u8fd9\u4e00\u5dee\u8ddd\u6e90\u4e8e Eagle \u6a21\u578b\u672a\u5728\u5de5\u5177\u8c03\u7528\u683c\u5f0f\u4e0a\u8bad\u7ec3\uff0c\u4f7f\u6b64\u7c7b\u8f93\u5165\u5bf9\u8fd9\u4e9b\u6a21\u578b\u800c\u8a00\u5c5e\u4e8e\u5206\u5e03\u5916\uff08out-of-distribution\uff09\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u5982\u7b2c 3.2 \u8282\u6240\u8ba8\u8bba\uff0c\u5de5\u5177\u8c03\u7528\u8f68\u8ff9\u8868\u73b0\u51fa\u7ed3\u6784\u89c4\u5f8b\u6027\u548c\u91cd\u590d\u8c03\u7528\u6a21\u5f0f\u3002\u8fd9\u4e9b\u6027\u8d28\u6709\u5229\u4e8e TR \u7b49\u5229\u7528\u5386\u53f2\u8f68\u8ff9\u6784\u9020\u8349\u7a3f\u5e8f\u5217\u7684 SD \u65b9\u6cd5\u3002\u57fa\u4e8e\u8fd9\u4e00\u6d1e\u5bdf\uff0cToolSpec \u5229\u7528\u5de5\u5177\u8c03\u7528\u4efb\u52a1\u7684 schema \u7ea6\u675f\u7279\u6027\uff0c\u5b9e\u73b0\u66f4\u51c6\u786e\u7684\u8349\u7a3f\u751f\u6210\uff0c\u5e76\u83b7\u5f97\u66f4\u9ad8\u6574\u4f53\u6548\u7387\u3002<\/p>\n<h6>ToolLLaMA \u4e0a\u7684\u7ed3\u679c<\/h6>\n<p>\u9664\u901a\u7528 LLM \u5916\uff0c\u6211\u4eec\u8fd8\u5728 ToolLLaMA \u7684\u4e24\u4e2a\u53d8\u4f53\u4e0a\u8fdb\u4e00\u6b65\u8bc4\u4f30 ToolSpec\uff0c\u8be5\u6a21\u578b\u4e13\u95e8\u9488\u5bf9\u5de5\u5177\u4f7f\u7528\u4efb\u52a1\u8bad\u7ec3\u3002\u56fe 6 \u5c55\u793a\u4e86 ToolSpec \u4e0e\u5148\u524d\u5373\u63d2\u5373\u7528 SD \u65b9\u6cd5\u5728 ToolBench \u4e0a\u7684\u63a8\u7406\u6548\u7387\u6bd4\u8f83\u3002\u7ed3\u679c\u8868\u660e\uff0cToolSpec \u76f8\u5bf9\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\u5b9e\u73b0\u7ea6 <span class=\"katex-eq\" data-katex-display=\"false\">3.7\\times<\/span> \u52a0\u901f\uff0c\u8bc1\u660e\u4e86\u5176\u5728\u4e13\u7528\u5de5\u5177\u4f7f\u7528\u578b LLM \u4e0a\u7684\u6709\u6548\u6027\u3002<\/p>\n<h4>\u5206\u6790<\/h4>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239010.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 7\uff1a\u5728 API-Bank \u4e0a\u4f7f\u7528 LLaMA-3.1-8B-Instruct \u65f6\uff0c\u524d\u4e24\u4e2a\u89e3\u7801\u6b65\u7684\u63a5\u53d7 token \u957f\u5ea6\u5206\u5e03\u3002\u7eff\u8272\u4e09\u89d2\u8868\u793a\u5747\u503c\uff0c\u9ed1\u8272\u83f1\u5f62\u8868\u793a\u79bb\u7fa4\u70b9\u3002<\/center><\/p>\n<p>\u672c\u8282\u5bf9 ToolSpec \u8fdb\u884c\u5168\u9762\u5206\u6790\u3002\u5177\u4f53\u800c\u8a00\uff0c\u6211\u4eec\u8bc4\u4f30\u5176\u6709\u6548\u6027\uff08\u00a76.1\uff09\uff0c\u8003\u5bdf\u5bf9\u5de5\u5177\u8c03\u7528\u683c\u5f0f\u7684\u9075\u5faa\u5ea6\u5f71\u54cd\uff08\u00a76.2\uff09\uff0c\u5e76\u5206\u6790\u4e0d\u540c\u6a21\u5757\u7684\u65f6\u95f4\u5206\u914d\uff08\u00a76.3\uff09\u3002<\/p>\n<h5>ToolSpec \u7684\u6709\u6548\u6027<\/h5>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239011.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 8\uff1a\u91cd\u590d\u5de5\u5177\u8c03\u7528\u7684\u5e73\u5747\u63a5\u53d7 token \u6570\uff08#MAT\uff09\u4e0e\u52a0\u901f\u6bd4\u3002\u7ed3\u679c\u5728 API-Bank \u4e0a\u4f7f\u7528 LLaMA-3.1-8B-Instruct \u83b7\u5f97\u3002<\/center><\/p>\n<h6>\u7edf\u8ba1\u5206\u6790<\/h6>\n<p>\u6211\u4eec\u9996\u5148\u8fdb\u884c\u7edf\u8ba1\u5206\u6790\u4ee5\u8bc4\u4f30\u6240\u63d0\u51fa\u6a21\u5757\u7684\u6709\u6548\u6027\uff0c\u5305\u62ec\uff1a<strong>1\uff09Schema \u611f\u77e5\u8349\u7a3f\u751f\u6210\u3002<\/strong> \u56fe 7 \u62a5\u544a\u4e86\u524d\u4e24\u4e2a\u89e3\u7801\u6b65\u7684\u63a5\u53d7 token \u957f\u5ea6\uff0c\u5206\u522b\u5bf9\u5e94 ToolSpec \u4e2d\u7684\u5de5\u5177\u540d\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">q_t<\/span> \u548c\u7b2c\u4e00\u4e2a\u53c2\u6570\u540d\u72b6\u6001 <span class=\"katex-eq\" data-katex-display=\"false\">q_p<\/span>\u3002\u5982\u56fe\u6240\u793a\uff0cschema \u611f\u77e5\u8349\u7a3f\u751f\u6210\u4f7f ToolSpec \u5728\u8fd9\u4e24\u4e2a\u6b65\u9aa4\u4e2d\u5e73\u5747\u751f\u6210 <span class=\"katex-eq\" data-katex-display=\"false\">19.3<\/span> \u4e2a tokens\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cTR \u7b49\u5148\u524d\u65b9\u6cd5\u5e73\u5747\u751f\u6210 <span class=\"katex-eq\" data-katex-display=\"false\">8.5<\/span> \u4e2a tokens\uff0c\u800c Eagle-3 \u5e73\u5747\u4ec5 <span class=\"katex-eq\" data-katex-display=\"false\">2<\/span> \u4e2a tokens\uff0c\u7b49\u540c\u4e8e\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\u3002\u8fd9\u4e9b\u7ed3\u679c\u5c55\u793a\u4e86 schema \u611f\u77e5\u8349\u7a3f\u751f\u6210\u5728\u5de5\u5177\u8c03\u7528\u751f\u6210\u4e2d\u7684\u660e\u663e\u4f18\u52bf\u3002<strong>2\uff09\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\u3002<\/strong> \u56fe 8 \u5c55\u793a\u4e86\u91cd\u590d\u5de5\u5177\u8c03\u7528\u7684 #MAT\u3002\u7ed3\u679c\u8868\u660e\uff0cToolSpec \u7a33\u5b9a\u4f18\u4e8e\u5148\u524d\u65b9\u6cd5\uff0c\u6bcf\u4e2a\u89e3\u7801\u6b65\u5e73\u5747\u63a5\u53d7 <span class=\"katex-eq\" data-katex-display=\"false\">9.3<\/span> \u4e2a tokens\uff0c\u5e76\u5b9e\u73b0\u7ea6 <span class=\"katex-eq\" data-katex-display=\"false\">5.5\\times<\/span> \u52a0\u901f\u3002\u8fd9\u4e00\u6539\u8fdb\u51f8\u663e\u4e86\u6240\u63d0\u51fa\u68c0\u7d22\u7b56\u7565\u5728\u5904\u7406\u91cd\u590d\u5de5\u5177\u8c03\u7528\u65f6\u7684\u6709\u6548\u6027\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239012.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 9\uff1a\u5728 API-Bank \u4e0a\u4f7f\u7528 LLaMA-3.1-8B-Instruct \u65f6\u7684\u63a5\u53d7 token \u957f\u5ea6\u5206\u5e03\u3002\u4e3a\u4fbf\u4e8e\u53ef\u89c6\u5316\uff0c\u8d85\u8fc7 17 \u7684\u503c\u88ab\u622a\u65ad\u3002<\/center><\/p>\n<h6>\u63a5\u53d7\u957f\u5ea6\u5206\u5e03<\/h6>\n<p>\u56fe 9 \u5c55\u793a\u4e86 ToolSpec \u4e0e\u5148\u524d\u65b9\u6cd5\u7684\u63a5\u53d7 token \u957f\u5ea6\u5206\u5e03\u3002\u73b0\u6709\u65b9\u6cd5\u901a\u5e38\u4ea7\u751f\u5c11\u4e8e <span class=\"katex-eq\" data-katex-display=\"false\">8<\/span> \u4e2a tokens \u7684\u63a5\u53d7\u957f\u5ea6\uff0c\u800c ToolSpec \u5c06\u5206\u5e03\u63a8\u5411\u66f4\u957f\u63a5\u53d7\u957f\u5ea6\uff0c\u5176\u4e2d <span class=\"katex-eq\" data-katex-display=\"false\">22\\%<\/span> \u7684\u89e3\u7801\u6b65\u8d85\u8fc7 <span class=\"katex-eq\" data-katex-display=\"false\">8<\/span> \u4e2a tokens\u3002\u6b64\u5916\uff0cToolSpec \u8868\u73b0\u51fa\u660e\u663e\u7684\u957f\u5c3e\u5206\u5e03\uff1a\u8fd1 <span class=\"katex-eq\" data-katex-display=\"false\">4\\%<\/span> \u7684\u89e3\u7801\u6b65\u8fbe\u5230 <span class=\"katex-eq\" data-katex-display=\"false\">17<\/span> \u5230 <span class=\"katex-eq\" data-katex-display=\"false\">33<\/span> \u4e4b\u95f4\u7684\u63a5\u53d7 token \u957f\u5ea6\u3002\u8fd9\u79cd\u957f\u5c3e\u884c\u4e3a\u6765\u81ea\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\uff0c\u5b83\u590d\u7528\u5386\u53f2\u5de5\u5177\u8c03\u7528\u8f68\u8ff9\u4ee5\u4ea7\u751f\u66f4\u957f\u3001\u66f4\u51c6\u786e\u7684\u8349\u7a3f\u3002\u8fd9\u4e9b \u6301\u7eed\u66f4\u9ad8\u7684\u63a5\u53d7 token \u957f\u5ea6\u8d21\u732e\u4e86 ToolSpec \u7684\u6574\u4f53\u52a0\u901f\u6548\u679c\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239013.png\" alt=\"\" \/><\/p>\n<p><center>\u8868 4\uff1aToolSpec \u6a21\u5757\u7684\u6d88\u878d\u5b9e\u9a8c\u3002\u201cSAD\u201d\u548c\u201cRAS\u201d\u5206\u522b\u8868\u793a schema \u611f\u77e5\u8349\u7a3f\u751f\u6210\u548c\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\u3002\u7ed3\u679c\u5728 API-Bank \u4e0a\u4f7f\u7528 LLaMA-3.1-8B-Instruct \u83b7\u5f97\u3002<\/center><\/p>\n<h6>\u6d88\u878d\u5b9e\u9a8c<\/h6>\n<p>\u8868 4 \u5c55\u793a\u4e86 ToolSpec \u53ca\u5176\u63d0\u51fa\u7ec4\u4ef6\u7684\u6d88\u878d\u5b9e\u9a8c\u3002\u7ed3\u679c\u8868\u660e\uff0cschema \u611f\u77e5\u8349\u7a3f\u751f\u6210\u548c\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\u5bf9 ToolSpec \u7684\u6548\u7387\u90fd\u81f3\u5173\u91cd\u8981\u3002\u5f53\u79fb\u9664\u8fd9\u4e9b\u7ec4\u4ef6\u540e\uff0cToolSpec \u7c7b\u4f3c\u4e8e PLD\uff0c\u4f9d\u8d56\u8f93\u5165\u4e0a\u4e0b\u6587\u5185\u7684\u540e\u7f00\u5339\u914d\uff0c\u5f97\u5230\u7684\u52a0\u901f\u7565\u9ad8\u4e8e TR\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239014.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 10\uff1a\u4e0d\u540c\u6a21\u578b\u5728\u5de5\u5177\u8c03\u7528\u751f\u6210\u4e2d\u7684\u683c\u5f0f\u9075\u5faa\u5ea6\u3002\u7ed3\u679c\u5728 API-Bank \u4e0a\u83b7\u5f97\u3002<\/center><\/p>\n<h5>\u683c\u5f0f\u9075\u5faa\u5ea6\u7684\u5f71\u54cd<\/h5>\n<p>\u9664\u7b2c 3.2 \u8282\u5206\u6790\u5916\uff0c\u6211\u4eec\u8fdb\u4e00\u6b65\u8bc4\u4f30\u5de5\u5177\u4f7f\u7528\u578b LLM \u5728\u5de5\u5177\u8c03\u7528\u751f\u6210\u8fc7\u7a0b\u4e2d\u5bf9\u6240\u9700 schema \u7684\u4e25\u683c\u9075\u5faa\u7a0b\u5ea6\u3002\u6211\u4eec\u4f7f\u7528\u4e24\u4e2a\u6807\u51c6\u5b9a\u4e49\u4e25\u683c\u683c\u5f0f\u9075\u5faa\u5ea6\uff1a1\uff09\u8f93\u51fa\u5fc5\u987b\u80fd\u591f\u88ab\u6b63\u786e\u89e3\u6790\u4e3a\u6709\u6548 JSON\uff1b2\uff09\u4efb\u4f55\u591a\u4f59\u7684\u81ea\u7136\u8bed\u8a00\u3001Markdown \u56f4\u680f\u3001\u7c7b XML \u6807\u7b7e\u6216\u7578\u5f62 JSON \u7ed3\u6784\u90fd\u89c6\u4e3a\u8fdd\u89c4\u3002\u7ed3\u679c\u663e\u793a\uff0c\u683c\u5f0f\u9075\u5faa\u5ea6\u4e0e ToolSpec \u5b9e\u73b0\u7684\u52a0\u901f\u4e4b\u95f4\u5b58\u5728\u5f3a\u6b63\u76f8\u5173\uff0cPearson \u76f8\u5173\u7cfb\u6570\u7ea6\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">0.90<\/span>\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0cLLaMA-3.1-8B-Instruct \u8fbe\u5230\u6700\u9ad8\u9075\u5faa\u5206\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">0.997<\/span>\uff0c\u5bf9\u5e94\u5728 API-Bank \u4e0a <span class=\"katex-eq\" data-katex-display=\"false\">4.45\\times<\/span> \u52a0\u901f\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239015.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 11\uff1aLLM \u54cd\u5e94\u67e5\u8be2\u65f6\u5404\u64cd\u4f5c\u7684\u65f6\u95f4\u5206\u914d\u3002\u7ed3\u679c\u5728 API-Bank \u4e0a\u4f7f\u7528 LLaMA-3.1-8B-Instruct \u83b7\u5f97\u3002<\/center><\/p>\n<h5>\u65f6\u95f4\u5206\u914d<\/h5>\n<p>\u56fe 11 \u5c55\u793a\u4e86 ToolSpec \u4e3b\u8981\u5ef6\u8fdf\u5bc6\u96c6\u578b\u7ec4\u4ef6\u7684\u65f6\u95f4\u5206\u914d\u3002\u5982\u56fe\u6240\u793a\uff0c\u76ee\u6807\u6a21\u578b\u7684\u89e3\u7801\u9636\u6bb5\u4e3b\u5bfc\u6574\u4f53\u5ef6\u8fdf\uff0c\u5360\u603b\u8fd0\u884c\u65f6\u95f4\u7684 <span class=\"katex-eq\" data-katex-display=\"false\">91.0\\%<\/span>\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cToolSpec \u5f15\u5165\u7684\u989d\u5916\u64cd\u4f5c\uff0c\u5305\u62ec top-<span class=\"katex-eq\" data-katex-display=\"false\">k<\/span> \u68c0\u7d22\u548c\u8349\u7a3f\u6811\u6784\u9020\uff0c\u4ec5\u5e26\u6765\u9002\u5ea6\u5f00\u9500\uff0c\u53ea\u5360\u603b\u5ef6\u8fdf\u7684 <span class=\"katex-eq\" data-katex-display=\"false\">5.7\\%<\/span>\u3002\u5c3d\u7ba1\u5f00\u9500\u5f88\u5c0f\uff0cToolSpec \u4ecd\u5728\u5de5\u5177\u8c03\u7528\u751f\u6210\u4e2d\u5b9e\u73b0\u7ea6 <span class=\"katex-eq\" data-katex-display=\"false\">4\\times<\/span> \u52a0\u901f\uff0c\u8bc1\u660e\u4e86\u5176\u5728\u964d\u4f4e\u7aef\u5230\u7aef\u5ef6\u8fdf\u65b9\u9762\u7684\u6709\u6548\u6027\u3002<\/p>\n<h4>\u7ed3\u8bba<\/h4>\n<p>\u672c\u6587\u63d0\u51fa ToolSpec\uff0c\u4e00\u79cd\u9762\u5411\u52a0\u901f\u5de5\u5177\u8c03\u7528\u751f\u6210\u7684 schema \u611f\u77e5\u3001\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\u89e3\u7801\u65b9\u6cd5\u3002\u5b83\u4e0d\u9700\u8981\u989d\u5916\u8bad\u7ec3\u6216\u8f85\u52a9\u6a21\u578b\uff0c\u53ef\u65e0\u7f1d\u96c6\u6210\u5230\u73b0\u6709 LLM \u5de5\u4f5c\u6d41\u4e2d\u3002\u5728\u591a\u79cd LLM \u548c\u5de5\u5177\u4f7f\u7528\u57fa\u51c6\u4e0a\u7684\u5927\u91cf\u5b9e\u9a8c\u8868\u660e\uff0cToolSpec \u5728\u4fdd\u6301\u8f93\u51fa\u5206\u5e03\u7684\u540c\u65f6\u5b9e\u73b0 <span class=\"katex-eq\" data-katex-display=\"false\">3.5\\times \\sim 4.2\\times<\/span> \u52a0\u901f\u3002\u8fdb\u4e00\u6b65\u5206\u6790\u51f8\u663e\u4e86\u6240\u63d0\u51fa\u7ec4\u4ef6\u7684\u6709\u6548\u6027\uff0c\u4ee5\u53ca\u683c\u5f0f\u9075\u5faa\u5ea6\u5bf9 ToolSpec \u6548\u7387\u7684\u5f71\u54cd\u3002\u6211\u4eec\u5e0c\u671b\u672c\u5de5\u4f5c\u63a8\u52a8 LLM \u4e2d\u9ad8\u6548\u5de5\u5177\u4f7f\u7528\u7684\u53d1\u5c55\uff0c\u5e76\u4e3a\u53ef\u6269\u5c55\u7684\u5de5\u5177\u8c03\u7528\u751f\u6210\u63d0\u4f9b\u5b9e\u8df5\u6d1e\u5bdf\u3002<\/p>\n<h4>\u5c40\u9650\u6027<\/h4>\n<p>\u7531\u4e8e\u8ba1\u7b97\u8d44\u6e90\u9650\u5236\uff0c\u6211\u4eec\u7684\u5b9e\u9a8c\u672a\u5305\u542b\u66f4\u5927\u89c4\u6a21\u8bed\u8a00\u6a21\u578b\uff0c\u4f8b\u5982 Qwen2.5-72B-Instruct \u548c LLaMA-3.1-70B-Instruct\u3002\u4e0d\u8fc7\uff0c\u6211\u4eec\u9884\u671f ToolSpec \u80fd\u591f\u5f88\u597d\u5730\u6cdb\u5316\u5230\u8fd9\u4e9b\u6a21\u578b\uff0c\u5e76\u5728\u5de5\u5177\u8c03\u7528\u751f\u6210\u6548\u7387\u4e0a\u53d6\u5f97\u8fdb\u4e00\u6b65\u63d0\u5347\u3002\u6211\u4eec\u5728\u9644\u5f55 B.6 \u4e2d\u7ed9\u51fa\u4e86 Qwen2.5-32B-Instruct \u7684\u5b9e\u9a8c\u7ed3\u679c\u3002\u6b64\u5916\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u5047\u8bbe\u5bf9\u9884\u5b9a\u4e49\u5de5\u5177 schema \u6709\u8f83\u5f3a\u9075\u5faa\u6027\uff0c\u5e76\u5b58\u5728\u91cd\u590d\u8c03\u7528\u6a21\u5f0f\uff1b\u8fd9\u4e0e\u9891\u7e41\u4f7f\u7528\u5de5\u5177\u7684\u771f\u5b9e\u5e94\u7528\u573a\u666f\u4e00\u81f4\u3002\u5982\u7b2c 3 \u8282\u548c\u7b2c 6.2 \u8282\u6240\u8ba8\u8bba\uff0c\u6211\u4eec\u9884\u671f\u968f\u7740 LLM \u80fd\u529b\u6301\u7eed\u63d0\u5347\uff0cToolSpec \u7684\u6709\u6548\u6027\u5c06\u8fdb\u4e00\u6b65\u589e\u5f3a\u3002<\/p>\n<h4>\u4f26\u7406\u58f0\u660e<\/h4>\n<p>\u6211\u4eec\u5b9e\u9a8c\u4e2d\u4f7f\u7528\u7684\u6570\u636e\u96c6\u5747\u5df2\u516c\u5f00\u53d1\u5e03\uff0c\u5e76\u901a\u8fc7\u82f1\u6587\u4eba\u673a\u4ea4\u4e92\u8fdb\u884c\u6807\u6ce8\u3002\u5728\u8be5\u8fc7\u7a0b\u4e2d\uff0c\u7528\u6237\u9690\u79c1\u53d7\u5230\u4fdd\u62a4\uff0c\u6570\u636e\u96c6\u4e2d\u4e0d\u5305\u542b\u4e2a\u4eba\u4fe1\u606f\u3002\u6211\u4eec\u4f7f\u7528\u7684\u79d1\u5b66\u5236\u54c1\u5747\u53ef\u5728\u5bbd\u677e\u8bb8\u53ef\u4e0b\u7528\u4e8e\u7814\u7a76\u3002\u672c\u6587\u5bf9\u8fd9\u4e9b\u5236\u54c1\u7684\u4f7f\u7528\u4e0e\u5176\u9884\u671f\u7528\u9014\u4e00\u81f4\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u8ba4\u4e3a\u672c\u7814\u7a76\u5de5\u4f5c\u7b26\u5408 ACL \u7684\u4f26\u7406\u6807\u51c6\u3002<\/p>\n<h4>\u9644\u5f55 A\uff1aToolSpec \u7ec6\u8282<\/h4>\n<p>\u672c\u8282\u63d0\u4f9b ToolSpec \u7684\u7ec6\u8282\uff0c\u5305\u62ec\u5e76\u884c\u9a8c\u8bc1\u7684\u5b9e\u73b0\u4ee5\u53ca\u8be6\u7ec6\u7b97\u6cd5\u3002<\/p>\n<h5>\u5e76\u884c\u9a8c\u8bc1<\/h5>\n<p>\u9075\u5faa\u5148\u524d\u5de5\u4f5c\uff0cToolSpec \u91c7\u7528\u57fa\u4e8e\u6811\u7684\u6ce8\u610f\u529b\u673a\u5236\u5e76\u884c\u9a8c\u8bc1\u591a\u4e2a\u5019\u9009\u8349\u7a3f\u5e8f\u5217\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5019\u9009\u8349\u7a3f\u88ab\u538b\u7f29\u4e3a\u5355\u4e00\u6241\u5e73\u5e8f\u5217\u5e76\u8f93\u5165\u76ee\u6807\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span>\u3002\u6a21\u578b\u968f\u540e\u4f7f\u7528\u6811\u7ed3\u6784\u6ce8\u610f\u529b\u63a9\u7801\uff0c\u901a\u8fc7\u5355\u6b21\u524d\u5411\u4f20\u64ad\u540c\u65f6\u9a8c\u8bc1\u6240\u6709\u8349\u7a3f\u5e8f\u5217\uff0c\u5982\u56fe 1 \u6240\u793a\u3002<\/p>\n<p>\u4e3a\u652f\u6301\u4e0d\u540c\u8349\u7a3f\u7b56\u7565\uff0cToolSpec \u91c7\u7528\u4e24\u7c7b\u8349\u7a3f\u6811\u67b6\u6784\uff1a\uff081\uff09<strong>\u9759\u6001\u6811\u3002<\/strong> \u5bf9\u4e8e\u5f00\u653e\u72b6\u6001\uff08\u4f8b\u5982 <span class=\"katex-eq\" data-katex-display=\"false\">q_v<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">q_o<\/span>\uff09\uff0cToolSpec \u91c7\u7528\u901a\u7528 SD \u7b56\u7565\uff08\u4f8b\u5982 TR\uff09\u8fdb\u884c\u6587\u672c\u751f\u6210\uff0c\u5176\u4e2d\u4f7f\u7528\u9884\u5b9a\u4e49\u9759\u6001\u6811\u67b6\u6784\u8fdb\u884c\u8349\u7a3f\u751f\u6210\uff1b\uff082\uff09<strong>\u52a8\u6001\u6811\u3002<\/strong> \u5bf9\u4e8e\u7ed3\u6784\u72b6\u6001\uff08\u4f8b\u5982 <span class=\"katex-eq\" data-katex-display=\"false\">q_t<\/span> \u548c <span class=\"katex-eq\" data-katex-display=\"false\">q_p<\/span>\uff09\u4ee5\u53ca\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\uff0cToolSpec \u6784\u9020\u9002\u5e94\u5404\u8349\u7a3f\u5e8f\u5217\u957f\u5ea6\u7684\u52a8\u6001\u6811\u63a9\u7801\uff0c\u4ece\u800c\u51cf\u5c11\u4e0d\u5fc5\u8981\u7684\u9a8c\u8bc1\u8ba1\u7b97\u3002<\/p>\n<h5>\u7b97\u6cd5<\/h5>\n<p>\u6211\u4eec\u5728\u7b97\u6cd5 1 \u4e2d\u7ed9\u51fa ToolSpec \u7684\u4f2a\u4ee3\u7801\u3002\u5177\u4f53\u800c\u8a00\uff0c\u5728\u6bcf\u4e2a\u89e3\u7801\u6b65\uff0cToolSpec \u68c0\u67e5\u5f53\u524d\u540e\u7f00\u662f\u5426\u4e0e\u5386\u53f2\u8c03\u7528\u4e2d\u7684\u4efb\u4f55\u5e8f\u5217\u5339\u914d\u3002\u82e5\u627e\u5230\u5339\u914d\uff0c\u5219\u63d0\u53d6\u5176\u5ef6\u7eed\u4f5c\u4e3a\u9ad8\u8d28\u91cf\u8349\u7a3f\uff1b\u5426\u5219\uff0cToolSpec \u4f7f\u7528\u6709\u9650\u72b6\u6001\u673a\uff08FSM\uff09\u5f15\u5bfc schema \u611f\u77e5\u8349\u7a3f\u751f\u6210\uff0c\u4ece\u800c\u52a0\u901f\u5de5\u5177\u8c03\u7528\u751f\u6210\u3002<\/p>\n<p><strong>\u7b97\u6cd5 1\uff1a\u4f7f\u7528 ToolSpec \u8fdb\u884c\u5de5\u5177\u8c03\u7528<\/strong><\/p>\n<p>\u8f93\u5165\uff1a\u76ee\u6807\u6a21\u578b <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}<\/span>\uff1b\u8f93\u5165\u67e5\u8be2 <span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{x}<\/span>\uff1b\u5de5\u5177\u6587\u6863 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{D}<\/span>\uff1b\u5386\u53f2\u6570\u636e\u5b58\u50a8 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{H}=\\{(\\boldsymbol{h}_j,\\boldsymbol{y}_j)\\}_{j=1}^{|\\mathcal{H}|}<\/span>\uff1b\u53c2\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">k<\/span>\u3002<\/p>\n<p>\u8f93\u51fa\uff1a\u751f\u6210\u7684\u5de5\u5177\u8c03\u7528\u5e8f\u5217 <span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{y}<\/span>\u3002<\/p>\n<ol>\n<li>\u5c06\u6587\u6863 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{D}<\/span> \u89e3\u6790\u4e3a schema <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{S}<\/span><\/li>\n<li>\u7531 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{S}<\/span> \u6784\u9020 FSM <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{F}=(\\mathcal{Q}, \\Sigma_{\\text{schema}}, \\delta, q_0)<\/span><\/li>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{y} \\leftarrow \\emptyset<\/span>\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">q \\leftarrow q_0<\/span><\/li>\n<li>\u8ba1\u7b97\u67e5\u8be2\u8868\u793a <span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{h}_x<\/span><\/li>\n<li>\u68c0\u7d22\u5386\u53f2\u8c03\u7528 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{I}_k \\leftarrow \\operatorname{Top-K}_{j} \\text{sim}(\\boldsymbol{h}_x,\\boldsymbol{h}_j)<\/span><\/li>\n<li>\u5f53 NotFinished(<span class=\"katex-eq\" data-katex-display=\"false\">q,\\boldsymbol{y}<\/span>) \u65f6\u5faa\u73af\uff1a\n<ul>\n<li>\/\/ \u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b<\/li>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{B}_{\\text{retrieval}} \\leftarrow \\text{SuffixMatch}(\\boldsymbol{y}, \\{\\boldsymbol{y}_{j}\\}_{j \\in \\mathcal{I}_k})<\/span><\/li>\n<li>\u82e5 <span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{B}_{\\text{retrieval}} \\neq \\emptyset<\/span>\uff1a\n<ul>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{B} \\leftarrow \\mathcal{B}_{\\text{retrieval}}<\/span><\/li>\n<\/ul>\n<\/li>\n<li>\u5426\u5219\uff1a\n<ul>\n<li>\/\/ Schema \u611f\u77e5\u8349\u7a3f\u751f\u6210<\/li>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{B} \\leftarrow \\text{SchemaAwareDraft}(q,\\mathcal{S},\\boldsymbol{y})<\/span><\/li>\n<\/ul>\n<\/li>\n<li>\/\/ \u4f7f\u7528\u6811\u6ce8\u610f\u529b\u6253\u5305\u5e8f\u5217<\/li>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">(\\boldsymbol{z},\\mathbf{A}) \\leftarrow \\text{PackDraftsAsTree}(\\mathcal{B})<\/span><\/li>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">\\hat{\\mathcal{B}} \\leftarrow \\text{Verification}(\\mathcal{M}, \\boldsymbol{x}, \\boldsymbol{y}, \\boldsymbol{z}, \\mathbf{A})<\/span><\/li>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">(\\tilde{\\boldsymbol{y}}^{\\star},\\ell) \\leftarrow \\text{SelectLongestAcceptedPrefix}(\\hat{\\mathcal{B}})<\/span><\/li>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{y} \\leftarrow \\boldsymbol{y} \\oplus \\tilde{\\boldsymbol{y}}^{\\star}_{1:\\ell}<\/span><\/li>\n<li>\/\/ \u66f4\u65b0 FSM \u72b6\u6001<\/li>\n<li><span class=\"katex-eq\" data-katex-display=\"false\">q \\leftarrow \\delta(q, \\tilde{\\boldsymbol{y}}^{\\star}_{1:\\ell})<\/span><\/li>\n<\/ul>\n<\/li>\n<li>\u8fd4\u56de <span class=\"katex-eq\" data-katex-display=\"false\">\\boldsymbol{y}<\/span><\/li>\n<\/ol>\n<h4>\u9644\u5f55 B\uff1a\u5b9e\u9a8c\u7ec6\u8282<\/h4>\n<p>\u672c\u8282\u7ed9\u51fa\u5b9e\u9a8c\u7ec6\u8282\uff0c\u5305\u62ec\u6570\u636e\u96c6\u63cf\u8ff0\u3001\u5b9e\u73b0\u7ec6\u8282\u3001\u6269\u5c55\u7ed3\u679c\u4ee5\u53ca\u8865\u5145\u4e3b\u8981\u53d1\u73b0\u7684\u989d\u5916\u5206\u6790\u3002<\/p>\n<h5>\u6570\u636e\u96c6<\/h5>\n<p>\u6211\u4eec\u7684\u8bc4\u4f30\u6db5\u76d6\u56db\u4e2a\u5e7f\u6cdb\u4f7f\u7528\u7684\u5de5\u5177\u4f7f\u7528\u57fa\u51c6\uff0c\u8bc4\u4f30\u6570\u636e\u7ec6\u8282\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li><strong>API-Bank<\/strong>\uff1a\u4e00\u4e2a\u4e09\u7ea7\u8bc4\u4f30\u6846\u67b6\uff0c\u5305\u542b 73 \u4e2a\u591a\u6837\u5316 API\u3002\u5b83\u901a\u8fc7\u81ea\u7136\u591a\u8f6e\u5bf9\u8bdd\uff0c\u5728\u4e0d\u540c\u96be\u5ea6\u7ea7\u522b\u4e0a\u8bc4\u4f30 LLM \u9009\u62e9\u4e0e\u8c03\u7528\u5de5\u5177\u7684\u80fd\u529b\u3002<\/li>\n<li><strong>ToolAlpaca<\/strong>\uff1a\u7528\u4e8e\u8bc4\u4f30 LLM \u901a\u7528\u5de5\u5177\u4f7f\u7528\u80fd\u529b\u7684\u57fa\u51c6\u3002\u8bc4\u4f30\u6570\u636e\u96c6\u5305\u542b\u4e24\u4e2a\u5b50\u96c6\uff1a\uff08i\uff09\u5305\u542b 10 \u4e2a\u5408\u6210\u5de5\u5177\u7684\u6a21\u62df\u5b50\u96c6\uff1b\uff08ii\uff09\u5305\u542b\u6765\u81ea\u4e0d\u540c\u9886\u57df 11 \u4e2a API \u7684\u771f\u5b9e\u4e16\u754c\u5b50\u96c6\u3002<\/li>\n<li><strong>BFCLv2<\/strong>\uff1a\u8986\u76d6\u5e7f\u6cdb\u573a\u666f\u7684\u7efc\u5408\u57fa\u51c6\uff0c\u5305\u62ec\u5355\u6b65\u63a8\u7406\u3001\u591a\u6b65\u5de5\u5177\u4f7f\u7528\u3001\u5b9e\u65f6\u6267\u884c\u7b49\u3002\u6211\u4eec\u5728\u5355\u51fd\u6570\u548c\u591a\u51fd\u6570\u8bbe\u7f6e\u4e0b\u8bc4\u4f30 ToolSpec\u3002<\/li>\n<li><strong>ToolBench<\/strong>\uff1a\u4e00\u4e2a\u5927\u89c4\u6a21\u9ad8\u8d28\u91cf\u57fa\u51c6\uff0c\u901a\u8fc7 ChatGPT \u81ea\u52a8\u6784\u5efa\uff0c\u8986\u76d6 16,464 \u4e2a\u771f\u5b9e\u4e16\u754c API \u7684\u591a\u5de5\u5177\u3001\u591a\u8f6e\u4ea4\u4e92\u3002\u6211\u4eec\u5728\u5176\u5355\u5de5\u5177\u5b50\u96c6\u4e0a\u8bc4\u4f30 ToolSpec\u3002<\/li>\n<\/ul>\n<h5>\u5b9e\u73b0\u7ec6\u8282<\/h5>\n<p>\u68c0\u7d22\u53c2\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">k<\/span> \u8bbe\u4e3a <span class=\"katex-eq\" data-katex-display=\"false\">10<\/span>\uff0c\u5373\u6bcf\u4e2a\u68c0\u7d22\u6b65\u4fdd\u7559 top-10 \u5019\u9009\u3002\u5bf9\u4e8e\u540e\u7f00\u5339\u914d\uff0c\u6211\u4eec\u9075\u5faa PLD\uff0c\u8003\u8651\u540e\u7f00\u957f\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">L \\in \\{5,6,7\\}<\/span>\uff0c\u5e76\u91c7\u7528\u6700\u957f\u4f18\u5148\u7b56\u7565\uff08\u4ece 7 \u5230 5\uff09\u3002\u4e00\u65e6\u627e\u5230\u5339\u914d\uff0c\u6211\u4eec\u6309\u987a\u5e8f\u63d0\u53d6\u5339\u914d\u540e\u7f00\u540e\u7684\u5ef6\u7eed\uff0c\u4f7f\u7528\u9884\u5b9a\u4e49\u957f\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">n \\in \\{32, 16, 8, 8\\}<\/span> \u6784\u9020\u8349\u7a3f\u5019\u9009\u3002\u6211\u4eec\u5728 ToolSpec \u4e2d\u91c7\u7528 Token Recycling\uff08TR\uff09\u4f5c\u4e3a\u901a\u7528\u63a8\u6d4b\u751f\u6210\u7b56\u7565\u3002\u6240\u6709\u5b9e\u9a8c\u4f7f\u7528 PyTorch 2.5.1\uff0c\u5728 2 \u5757 NVIDIA A100-PCIE GPU\uff0840GB\uff09\u548c 40 \u4e2a CPU \u6838\u5fc3\u3001CUDA 12.4 \u73af\u5883\u4e0b\u8fdb\u884c\u3002\u6211\u4eec\u7684\u65b9\u6cd5\u548c\u6240\u6709\u57fa\u7ebf\u7684\u63a8\u7406\u5747\u4f7f\u7528 Hugging Face transformers \u5305\u5b8c\u6210\u3002\u9075\u5faa\u5148\u524d\u5de5\u4f5c\uff0c\u6211\u4eec\u91c7\u7528\u63a8\u6d4b\u91c7\u6837\u4f5c\u4e3a\u63a5\u53d7\u7b56\u7565\uff0cbatch size \u4e3a 1\u3002<\/p>\n<h5>\u6027\u80fd\u6bd4\u8f83<\/h5>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239016.png\" alt=\"\" \/><\/p>\n<p><center>\u8868 5\uff1aToolSpec \u4e0e\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\u5728\u4e0d\u540c\u6a21\u578b\u3001\u4e24\u4e2a\u5e7f\u6cdb\u4f7f\u7528\u7684\u5de5\u5177\u4f7f\u7528\u57fa\u51c6\u4e0a\u7684\u6027\u80fd\u6bd4\u8f83\u3002<\/center><\/p>\n<p>\u8868 5 \u5c55\u793a\u4e86 ToolSpec \u4e0e\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\u5728 API-Bank \u548c ToolAlpaca \u4e0a\u4e0d\u540c\u6a21\u578b\u7684\u51c6\u786e\u7387\u6bd4\u8f83\u3002\u7ed3\u679c\u8868\u660e\uff0cToolSpec \u4fdd\u6301\u81ea\u56de\u5f52\u89e3\u7801\u7684\u539f\u59cb\u6027\u80fd\uff0c\u5728\u591a\u79cd\u6a21\u578b\u548c\u6570\u636e\u96c6\u4e0a\u5b9e\u73b0 \u6301\u7eed\u53ef\u6bd4\u7684\u51c6\u786e\u7387\u3002Qwen2.5-14B \u5728 ToolAlpaca \u4e0a\u89c2\u5bdf\u5230\u7684\u8f7b\u5fae\u5dee\u5f02\u53ef\u5f52\u56e0\u4e8e fp16 \u7cbe\u5ea6\u3002<\/p>\n<h5>ToolSpec \u4e2d\u7684\u8d85\u53c2\u6570<\/h5>\n<p>\u8868 6 \u548c\u8868 7 \u5c55\u793a\u4e86 ToolSpec \u5728\u4e0d\u540c\u8d85\u53c2\u6570\u8bbe\u7f6e\u4e0b\u7684\u52a0\u901f\uff0c\u5305\u62ec\u5ef6\u7eed\u957f\u5ea6\u548c\u68c0\u7d22\u53c2\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">k<\/span>\u3002\u8bc4\u4f30\u5728 API-Bank \u8bad\u7ec3\u96c6\u91c7\u6837\u7684 100 \u4e2a\u5b9e\u4f8b\u4e0a\u8fdb\u884c\u3002\u8868 6 \u8868\u660e\uff0c\u8f83\u77ed\u7684\u5ef6\u7eed\u957f\u5ea6\uff08\u4f8b\u5982 <span class=\"katex-eq\" data-katex-display=\"false\">\\{16, 8, 4, 4\\}<\/span>\uff09\u4f1a\u5bfc\u81f4\u6b21\u4f18\u7684 #MAT \u548c\u52a0\u901f\u3002\u5c3d\u7ba1\u66f4\u957f\u7684\u5ef6\u7eed\u957f\u5ea6\uff08\u4f8b\u5982 <span class=\"katex-eq\" data-katex-display=\"false\">\\{64, 32, 16, 8\\}<\/span>\uff09\u4f1a\u7565\u5fae\u63d0\u5347 #MAT\uff0c\u4f46\u4e5f\u4f1a\u5f15\u5165\u989d\u5916\u8ba1\u7b97\u5f00\u9500\u5e76\u964d\u4f4e\u6574\u4f53\u52a0\u901f\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u5728\u5b9e\u9a8c\u4e2d\u91c7\u7528 <span class=\"katex-eq\" data-katex-display=\"false\">\\{32, 16, 8, 8\\}<\/span> \u7684\u5ef6\u7eed\u957f\u5ea6\u3002\u6211\u4eec\u8fdb\u4e00\u6b65\u5c06\u8be5\u8bbe\u7f6e\u4e0e\u7edf\u4e00\u5ef6\u7eed\u957f\u5ea6 <span class=\"katex-eq\" data-katex-display=\"false\">\\{16, 16, 16, 16\\}<\/span> \u6bd4\u8f83\uff0c\u7ed3\u679c\u8868\u660e\u4e3a\u6392\u540d\u66f4\u9ad8\u7684\u5019\u9009\u5206\u914d\u66f4\u957f\u5ef6\u7eed\u5e26\u6765\u66f4\u5927\u6536\u76ca\u3002\u7c7b\u4f3c\u5730\uff0c\u8868 7 \u8868\u660e\u5c06\u68c0\u7d22\u53c2\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">k<\/span> \u8bbe\u4e3a 10 \u5728\u6211\u4eec\u7684\u5b9e\u9a8c\u4e2d\u53d6\u5f97\u6700\u4f73\u52a0\u901f\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239017.png\" alt=\"\" \/><\/p>\n<p><center>\u8868 6\uff1aToolSpec \u5728\u4e0d\u540c\u5ef6\u7eed\u957f\u5ea6\u4e0b\u7684\u52a0\u901f\u3002\u7ed3\u679c\u5728 API-Bank \u4e0a\u4f7f\u7528 LLaMA-3.1-8B-Instruct \u83b7\u5f97\u3002<\/center><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239018.png\" alt=\"\" \/><\/p>\n<p><center>\u8868 7\uff1aToolSpec \u5728\u4e0d\u540c\u68c0\u7d22\u53c2\u6570 <span class=\"katex-eq\" data-katex-display=\"false\">k<\/span> \u4e0b\u7684\u52a0\u901f\u3002\u7ed3\u679c\u5728 API-Bank \u4e0a\u4f7f\u7528 LLaMA-3.1-8B-Instruct \u83b7\u5f97\u3002<\/center><\/p>\n<h5>\u5185\u5b58\u5f00\u9500<\/h5>\n<p>ToolSpec \u5f15\u5165\u7684\u989d\u5916\u5185\u5b58\u5f00\u9500\u8f83\u5c0f\u3002\u4e3b\u8981\u6301\u4e45\u6210\u672c\u6765\u81ea\u8f93\u51fa\u5185\u5b58\uff08output memory\uff09\uff0c\u5b83\u5b58\u50a8\u6bcf\u4e2a\u8f93\u5165\u95ee\u9898\u7684\u9690\u85cf\u72b6\u6001\uff0c\u4ee5\u53ca\u6bcf\u4e2a\u8bb0\u5fc6\u6837\u672c\u5bf9\u5e94\u7684\u8f93\u51fa tokens\u3002\u5728 API-Bank \u4e0a\uff0c597 \u4e2a\u6761\u76ee\u4ec5\u9700 4.9 MB \u5185\u5b58\uff0c\u7ea6\u5408\u6bcf\u4e2a\u6761\u76ee 0.0082 MB\u3002\u4e3a\u652f\u6301\u66f4\u5927\u6570\u636e\u96c6\uff0cToolSpec \u53ef\u4ee5\u7ef4\u62a4\u56fa\u5b9a\u5927\u5c0f\u7684\u8bb0\u5fc6\u5e93\uff0c\u4f8b\u5982 1,000 \u4e2a\u8bb0\u5fc6\u6837\u672c\uff0c\u4ece\u800c\u5728\u4fdd\u7559\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\u6536\u76ca\u7684\u540c\u65f6\u9650\u5236\u8f93\u51fa\u5185\u5b58\u3002\u6211\u4eec\u5c06\u6b64\u7c7b\u5185\u5b58\u7ba1\u7406\u7b56\u7565\u7684\u8bbe\u8ba1\u7559\u4f5c\u672a\u6765\u5de5\u4f5c\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239019.png\" alt=\"\" \/><\/p>\n<p><center>\u8868 8\uff1a\u4f7f\u7528 Qwen2.5-32B-Instruct \u5728 API-Bank \u4e0a\uff0cToolSpec \u4e0e\u5148\u524d\u5373\u63d2\u5373\u7528\u65b9\u6cd5\u7684\u52a0\u901f\u6bd4\u8f83\u3002\u6700\u4f18\u7ed3\u679c\u52a0\u7c97\uff0c\u6b21\u4f18\u7ed3\u679c\u52a0\u4e0b\u5212\u7ebf\u3002<\/center><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239020.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 12\uff1aToolSpec \u76f8\u5bf9\u5148\u524d\u6700\u4f18\u65b9\u6cd5\u7684\u76f8\u5bf9\u52a0\u901f\u63d0\u5347\u3002\u7ed3\u679c\u5728 API-Bank \u4e0a\u4f7f\u7528 Qwen \u7cfb\u5217\u83b7\u5f97\u3002<\/center><\/p>\n<h5>Qwen2.5-Instruct-32B \u4e0a\u7684\u7ed3\u679c<\/h5>\n<p>\u8868 8 \u6bd4\u8f83\u4e86\u5728 API-Bank \u4e0a\u4f7f\u7528 Qwen2.5-32B-Instruct \u65f6\uff0cToolSpec \u4e0e\u5148\u524d\u5373\u63d2\u5373\u7528\u65b9\u6cd5\u7684\u52a0\u901f\u3002\u7ed3\u679c\u8868\u660e\uff0cToolSpec \u4f18\u4e8e\u8fd9\u4e9b\u57fa\u7ebf\uff0c\u5e76\u76f8\u5bf9\u6807\u51c6\u81ea\u56de\u5f52\u89e3\u7801\u6700\u9ad8\u5b9e\u73b0 <span class=\"katex-eq\" data-katex-display=\"false\">3\\times<\/span> \u52a0\u901f\u3002\u56fe 12 \u663e\u793a\uff0cToolSpec \u76f8\u5bf9\u6700\u5f3a\u590d\u73b0\u57fa\u7ebf\u7684\u76f8\u5bf9\u52a0\u901f\u589e\u76ca\u968f\u6a21\u578b\u89c4\u6a21\u589e\u5927\u800c\u589e\u5927\u3002\u8fd9\u4e9b\u7ed3\u679c\u8868\u660e\uff0cToolSpec \u80fd\u826f\u597d\u6269\u5c55\u5230\u5927\u6a21\u578b\u89c4\u6a21\uff0c\u5e76\u76f8\u5bf9\u73b0\u6709\u5373\u63d2\u5373\u7528\u65b9\u6cd5\u63d0\u4f9b \u8d8a\u6765\u8d8a\u660e\u663e\u7684\u6548\u7387\u6536\u76ca\u3002<\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239021.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 13\uff1a\u5728 API-Bank \u4e0a\u4f7f\u7528 LLaMA-3.1-8B-Instruct \u65f6\uff0c\u76f8\u4f3c\u7528\u6237\u67e5\u8be2\u4e0b ToolSpec \u7684\u6848\u4f8b\u7814\u7a76\u3002<\/center><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/07\/20260710144239022.png\" alt=\"\" \/><\/p>\n<p><center>\u56fe 14\uff1a\u5728 LLaMA-3.1-8B-Instruct \u4e0a\uff0c\u4e0d\u540c\u5de5\u5177\u53c2\u6570\u503c\u573a\u666f\u4e0b ToolSpec \u7684\u6848\u4f8b\u7814\u7a76\u3002<\/center><\/p>\n<h5>\u6848\u4f8b\u7814\u7a76<\/h5>\n<p>\u56fe 13 \u5c55\u793a\u4e86\u5728 API-Bank \u4e0a\u4f7f\u7528 LLaMA-3.1-8B-Instruct \u65f6\uff0c\u5f53\u524d\u7528\u6237\u67e5\u8be2\u4e0e\u5386\u53f2\u89c2\u5bdf\u67e5\u8be2\u76f8\u4f3c\u7684\u4e00\u4e2a ToolSpec \u6848\u4f8b\u3002\u5982\u56fe\u6240\u793a\uff0cToolSpec \u9996\u5148\u5e94\u7528 schema token \u586b\u5145\u751f\u6210\u6240\u9700\u5de5\u5177\u540d\u548c schema \u7ea6\u675f tokens\uff0c\u968f\u540e\u68c0\u7d22\u9ad8\u5ea6\u76f8\u4f3c\u7684\u5386\u53f2\u5de5\u5177\u8c03\u7528\uff0c\u5e76\u5c06\u5176\u5ef6\u7eed\u590d\u7528\u4e3a\u8349\u7a3f tokens\uff0c\u5728\u540e\u7eed\u89e3\u7801\u6b65\u4e2d\u6700\u9ad8\u63a5\u53d7 33 \u4e2a tokens\u3002\u8be5\u6848\u4f8b\u5c55\u793a\u4e86 schema token \u586b\u5145\u4e0e\u68c0\u7d22\u589e\u5f3a\u63a8\u6d4b\u7684\u6709\u6548\u6027\u3002<\/p>\n<p>\u7c7b\u4f3c\u5730\uff0c\u56fe 14 \u5c55\u793a\u4e86\u4e00\u4e2a\u7528\u6237\u67e5\u8be2\u76f8\u4f3c\u4f46\u5de5\u5177\u53c2\u6570\u503c\u4e0d\u540c\u7684\u6848\u4f8b\u3002ToolSpec \u5c06\u5148\u524d\u8f93\u51fa\u7684\u5ef6\u7eed\u590d\u7528\u4e3a\u8349\u7a3f tokens\uff0c\u5728\u7b2c\u4e09\u4e2a\u89e3\u7801\u6b65\u63a5\u53d7 16 \u4e2a tokens\uff0c\u968f\u540e\u4fee\u6b63\u7528\u6237\u7279\u5b9a\u7684\u51fa\u884c\u4fe1\u606f\uff0c\u5e76\u901a\u8fc7\u901a\u7528\u63a8\u6d4b\u751f\u6210\u5b8c\u6210\u5de5\u5177\u8c03\u7528\u3002\u8be5\u6848\u4f8b\u51f8\u663e\u4e86 ToolSpec \u5728\u91cd\u590d\u5de5\u5177\u8c03\u7528\u573a\u666f\u4e2d\u7684\u4f18\u52bf\uff1a\u5927\u90e8\u5206 tokens \u53ef\u7531\u5386\u53f2\u8c03\u7528\u8349\u7a3f\u751f\u6210\uff0c\u800c\u76ee\u6807 LLM \u4e13\u6ce8\u4e8e\u5173\u952e\u53c2\u6570\u503c\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Meta Data \u53d1\u8868\u65f6\u95f4\uff1a2026-04-15\uff08arXiv \u9996\u6b21\u63d0\u4ea4\uff09\uff1b2026-04-16\uff08\u516c\u5e03\uff09 \u4f5c\u8005\uff1aHeming Xi &#8230;<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"emotion":"","emotion_color":"","title_style":"","license":"","footnotes":""},"categories":[1],"tags":[],"class_list":["post-1641","post","type-post","status-publish","format-standard","hentry","category-uncategorized"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1641","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/comments?post=1641"}],"version-history":[{"count":3,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1641\/revisions"}],"predecessor-version":[{"id":1644,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1641\/revisions\/1644"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=1641"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=1641"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=1641"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}