{"id":1501,"date":"2026-02-05T21:42:32","date_gmt":"2026-02-05T13:42:32","guid":{"rendered":"https:\/\/lingbo.online\/?p=1501"},"modified":"2026-02-07T10:11:18","modified_gmt":"2026-02-07T02:11:18","slug":"agentar-scale-sql-advancing-text-to-sql-through-orchestrated-test-time-scaling","status":"publish","type":"post","link":"https:\/\/lingbo.online\/index.php\/learning_note\/agentar-scale-sql-advancing-text-to-sql-through-orchestrated-test-time-scaling\/","title":{"rendered":"Agentar-Scale-SQL: Advancing Text-to-SQL through Orchestrated Test-Time Scaling"},"content":{"rendered":"<h3>Meta Data<\/h3>\n<ul>\n<li>\u53d1\u8868\u65f6\u95f4 2025.09.29<\/li>\n<li>\u4f5c\u8005\uff1aPengfei Wang, Baolin Sun1, Xuemei Dong etc.<\/li>\n<li>\u8bba\u6587\u94fe\u63a5\uff1aarxiv.org\/pdf\/2509.24403.pdf<\/li>\n<li>\u9879\u76ee\u94fe\u63a5\uff1a<a href=\"https:\/\/github.com\/antgroup\/Agentar-Scale-SQL\" target=\"_blank\"  rel=\"nofollow\" >https:\/\/github.com\/antgroup\/Agentar-Scale-SQL<\/a><\/li>\n<\/ul>\n<h3>Abstract<\/h3>\n<p>\u5728 BIRD \u7b49\u5177\u6709\u6311\u6218\u6027\u7684\u57fa\u51c6\u6d4b\u8bd5\u4e2d\uff0c\u6700\u5148\u8fdb (SOTA) \u7684\u6587\u672c\u5230 SQL \u65b9\u6cd5\u4ecd\u7136\u8fdc\u8fdc\u843d\u540e\u4e8e\u4eba\u7c7b\u4e13\u5bb6\u3002 \u5f53\u524d\u63a2\u7d22\u6d4b\u8bd5\u65f6\u95f4\u6269\u5c55\u7684\u65b9\u6cd5\u7f3a\u4e4f\u7cbe\u5fc3\u7b56\u5212\u7684\u7b56\u7565\uff0c\u5e76\u4e14\u5ffd\u7565\u4e86\u6a21\u578b\u7684\u5185\u90e8\u63a8\u7406\u8fc7\u7a0b\u3002 \u4e3a\u4e86\u5f25\u8865\u8fd9\u4e00\u5dee\u8ddd\uff0c\u6211\u4eec\u5f15\u5165\u4e86 Agentar-Scale-SQL\uff0c\u8fd9\u662f\u4e00\u79cd\u5229\u7528\u53ef\u6269\u5c55\u8ba1\u7b97\u6765\u63d0\u9ad8\u6027\u80fd\u7684\u65b0\u9896\u6846\u67b6\u3002 Agentar-Scale-SQL \u5b9e\u73b0\u4e86\u7cbe\u5fc3\u7b56\u5212\u7684\u6d4b\u8bd5\u65f6\u95f4\u6269\u5c55\u7b56\u7565\uff0c\u8be5\u7b56\u7565\u534f\u540c\u7ed3\u5408\u4e86\u4e09\u4e2a\u4e0d\u540c\u7684\u89c6\u89d2\uff1ai) \u901a\u8fc7 RL \u589e\u5f3a\u7684\u5185\u5728\u63a8\u7406\u8fdb\u884c\u5185\u90e8\u6269\u5c55\uff0cii) \u901a\u8fc7\u8fed\u4ee3\u7ec6\u5316\u8fdb\u884c\u987a\u5e8f\u6269\u5c55\uff0c\u4ee5\u53ca iii) \u4f7f\u7528\u591a\u6837\u5316\u7efc\u5408\u548c\u9526\u6807\u8d5b\u9009\u62e9\u8fdb\u884c\u5e76\u884c\u6269\u5c55\u3002 Agentar-Scale-SQL \u662f\u4e00\u4e2a\u901a\u7528\u6846\u67b6\uff0c\u65e8\u5728\u8f7b\u677e\u9002\u5e94\u65b0\u6570\u636e\u5e93\u548c\u66f4\u5f3a\u5927\u7684\u8bed\u8a00\u6a21\u578b\u3002 \u5927\u91cf\u5b9e\u9a8c\u8868\u660e\uff0cAgentar-Scale-SQL\u5728BIRD\u57fa\u51c6\u4e0a\u5b9e\u73b0\u4e86SOTA\u6027\u80fd\uff0c\u5728\u6d4b\u8bd5\u96c6\u4e0a\u8fbe\u5230\u4e8681.67%\u7684\u6267\u884c\u51c6\u786e\u7387\uff0c\u5728\u5b98\u65b9\u6392\u884c\u699c\u4e0a\u6392\u540d\u7b2c\u4e00\uff0c\u5c55\u793a\u4e86\u4e00\u6761\u901a\u5f80\u4eba\u7c7b\u6c34\u5e73\u6027\u80fd\u7684\u6709\u6548\u9014\u5f84\u3002<\/p>\n<h3>Introduction<\/h3>\n<p>\u901a\u8fc7\u4f7f\u7528\u6237\u80fd\u591f\u7528\u81ea\u5df1\u7684\u81ea\u7136\u8bed\u8a00\u67e5\u8be2\u7ed3\u6784\u5316\u6570\u636e\u5e93\u6765\u5b9e\u73b0\u6570\u636e\u5206\u6790\u7684\u6c11\u4e3b\u5316\u662f\u4eba\u673a\u4ea4\u4e92\u7684\u957f\u671f\u76ee\u6807\u3002\u8fd9\u662fText-to-SQL\u7684\u6838\u5fc3\u76ee\u6807\uff0c\u662f\u5173\u952e\u7814\u7a76\u9886\u57df\uff0c\u4e13\u6ce8\u4e8e\u5c06\u81ea\u7136\u8bed\u8a00\u95ee\u9898\u8f6c\u5316\u4e3a\u53ef\u6267\u884c\u7684SQL\u67e5\u8be2\u3002\u901a\u8fc7\u5f25\u5408\u4eba\u7c7b\u8bed\u8a00\u548c\u7ed3\u6784\u5316\u6570\u636e\u4e4b\u95f4\u7684\u5dee\u8ddd\uff0c\u6587\u672c\u5230SQL\u4f7f\u975e\u6280\u672f\u7528\u6237\u80fd\u591f\u6709\u6548\u5730\u4e0e\u590d\u6742\u6570\u636e\u5e93\u4ea4\u4e92\uff0c\u5f15\u8d77\u4e86\u81ea\u7136\u8bed\u8a00\u5904\u7406(NLP)\u548c\u6570\u636e\u5e93\u793e\u533a\u7684\u6781\u5927\u5174\u8da3\u3002<br \/>\n\u6587\u672c\u5230SQL\u7684\u6700\u7ec8\u613f\u666f\u662f\u5f00\u53d1\u80fd\u591f\u5339\u914d\u5e76\u6700\u7ec8\u8d85\u8d8a\u4eba\u7c7b\u4e13\u5bb6\u6027\u80fd\u7684\u7cfb\u7edf\u3002\u7136\u800c\uff0c\u8fd9\u4e00\u96c4\u5fc3\u4e0e\u5f53\u524d\u7684\u6700\u65b0\u6280\u672f\u4e4b\u95f4\u4ecd\u7136\u5b58\u5728\u5de8\u5927\u5dee\u8ddd\u3002\u5728\u5177\u6709\u6311\u6218\u6027\u7684BIRD\u57fa\u51c6\u4e0a\uff0c\u4eba\u7c7b\u4e13\u5bb6\u7684\u6267\u884c\u51c6\u786e\u7387(EX)\u8fbe\u523092.96%\uff0c\u800c\u5373\u4f7f\u662f\u8868\u73b0\u6700\u597d\u7684\u65b9\u6cd5\u4e5f\u6709\u76f8\u5f53\u5927\u7684\u6ede\u540e\uff0c\u524d5\u79cd\u65b9\u6cd5\u5728\u6d4b\u8bd5\u96c6\u4e0a\u7684\u51c6\u786e\u7387\u7ea6\u4e3a75%\u3002\u5f25\u5408\u5de8\u5927\u7684\u4eba\u673a\u6027\u80fd\u9e3f\u6c9f\u8feb\u5207\u9700\u8981\u521b\u65b0\u3002<br \/>\n\u6700\u8fd1\u7684Text-to-SQL\u7814\u7a76\u5206\u4e3a\u4e09\u4e2a\u4e3b\u8981\u7c7b\u522b\u3002\u7b2c\u4e00\u79cd\u662f\u57fa\u4e8e\u63d0\u793a\u7684\u65b9\u6cd5\uff08\u4f8b\u5982OpenSearch-SQL\u548cDAIL-SQL\uff09\u3002\u7b2c\u4e8c\u7c7b\u662f\u57fa\u4e8e\u5fae\u8c03\u7684\u65b9\u6cd5\uff0c\u4ee3\u8868\u4f5c\u54c1\u6709Arctic-Text2SQL-R1-32B\u3002\u7b2c\u4e09\u7c7b\u5305\u62ecXiYan-SQL\u3001CHASE-SQL+Gemini\u548cContextual-SQL\u7b49\u6df7\u5408\u65b9\u6cd5\u3002\u8fd9\u4e9b\u65b9\u6cd5\u4e3b\u8981\u4ece\u4e0d\u540c\u7684\u89d2\u5ea6\u63a2\u7d22\u6d4b\u8bd5\u65f6\u95f4\u6269\u5c55\uff1aContextual-SQL\u91c7\u7528\u96c6\u6210\u7b56\u7565\uff0c\u800cXiYan-SQL\u548cCHASE-SQL\u7814\u7a76\u96c6\u6210\u7b56\u7565\u548c\u987a\u5e8f\u7ec6\u5316\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u7814\u7a76\u90fd\u6709\u4e00\u4e2a\u5171\u540c\u7684\u5c40\u9650\u6027\uff0c\u56e0\u4e3a\u5b83\u4eec\u5ffd\u89c6\u4e86\u6a21\u578b\u63a8\u7406\u8fc7\u7a0b\u7684\u5185\u90e8\u89c6\u89d2\u548c\u7cbe\u5fc3\u7b56\u5212\u7684\u7f29\u653e\u7ec4\u5408\u3002<br \/>\n\u4e3a\u4e86\u8fdb\u4e00\u6b65\u63d0\u9ad8Text-to-SQL\u7684\u6027\u80fd\uff0c\u8fd9\u9879\u5de5\u4f5c\u8ba4\u4e3a\u6700\u6709\u5e0c\u671b\u7684\u524d\u8fdb\u9053\u8def\u5728\u4e8e\u5b8c\u5168\u63a5\u53d7&quot;\u75db\u82e6\u7684\u6559\u8bad&quot;\u7684\u539f\u5219\uff1a\u5229\u7528\u53ef\u6269\u5c55\u8ba1\u7b97\u7684\u901a\u7528\u65b9\u6cd5\u6700\u7ec8\u6218\u80dc\u4e86\u57fa\u4e8e\u590d\u6742\u7684\u3001\u4eba\u7c7b\u6307\u5b9a\u7684\u77e5\u8bc6\u7684\u65b9\u6cd5\u3002\u79c9\u627f\u8fd9\u4e00\u7406\u5ff5\uff0c\u6211\u4eec\u4e13\u6ce8\u4e8e\u6d4b\u8bd5\u65f6\u6269\u5c55\uff0c\u5e76\u4e14\u6ca1\u6709\u4e3a\u6a21\u5f0f\u94fe\u63a5\u8bbe\u8ba1\u590d\u6742\u7684\u7b56\u7565\uff0c\u56e0\u4e3a\u6211\u4eec\u76f8\u4fe1\u4ec5\u6269\u5c55\u5c31\u8db3\u4ee5\u63d0\u9ad8\u6027\u80fd\u3002\u4e3a\u6b64\uff0c\u6211\u4eec\u5f15\u5165\u4e86Agentar-Scale-SQL\uff0c\u8fd9\u662f\u4e00\u4e2a\u7cbe\u5fc3\u5b89\u6392\u7684\u6d4b\u8bd5\u65f6\u95f4\u6269\u5c55\u6846\u67b6\uff0c\u5b83\u8bc1\u660e\u4e86\u5b9e\u73b0\u4eba\u7c7b\u6c34\u5e73\u6027\u80fd\u7684\u9014\u5f84\u4e0d\u5728\u4e8e\u5236\u5b9a\u66f4\u667a\u80fd\u7684\u542f\u53d1\u5f0f\u65b9\u6cd5\uff0c\u800c\u5728\u4e8e\u6784\u5efa\u6709\u6548\u5229\u7528\u6d4b\u8bd5\u65f6\u95f4\u6269\u5c55\u7684\u901a\u7528\u6846\u67b6\u3002Agentar-Scale-SQL\u91c7\u7528\u8de8\u4e09\u4e2a\u4e0d\u540c\u89c6\u89d2\u7684\u7cbe\u5fc3\u7b56\u5212\u7684\u6269\u5c55\u7b56\u7565\uff1ai)\u5185\u90e8\u6269\u5c55Internal Scaling\uff08\u901a\u8fc7RL\u589e\u5f3a\u7684\u5185\u5728\u63a8\u7406\uff09\uff0cii)\u987a\u5e8f\u6269\u5c55Sequential Scaling\uff08\u901a\u8fc7\u8fed\u4ee3\u7ec6\u5316\uff09\uff0c\u4ee5\u53caiii)\u5e76\u884c\u6269\u5c55Parallel Scaling\uff08\u901a\u8fc7\u591a\u6837\u5316\u7efc\u5408\u548c\u9526\u6807\u8d5b\u9009\u62e9\uff09\u3002\u8be5\u7b56\u7565\u5728\u4e09\u9636\u6bb5\u6846\u67b6\u5185\u5b9e\u73b0\uff0c\u5982\u56fe1\u6240\u793a\u3002\u6b65\u9aa41\uff08\u4efb\u52a1\u7406\u89e3\uff09\u7684\u76ee\u6807\u662f\u5efa\u7acb\u5bf9\u8f93\u5165\u53ca\u5176\u4e0a\u4e0b\u6587\u7684\u5168\u9762\u7406\u89e3\uff0c\u8fd9\u5bf9\u4e8e\u540e\u7eed\u6b65\u9aa4\u4e2d\u7684\u6269\u5c55\u64cd\u4f5c\u81f3\u5173\u91cd\u8981\u3002\u7136\u540e\uff0c\u7b2c2\u6b65\uff08SQL Generation Scaling\uff09\u8fdb\u884c\u591a\u6837\u5316\u7684\u7efc\u5408\u548c\u8fed\u4ee3\u7ec6\u5316\uff0c\u4ee5\u83b7\u5f97\u9ad8\u8d28\u91cf\u548c\u591a\u6837\u5316\u7684SQL\u5019\u9009\u3002\u5177\u4f53\u6765\u8bf4\uff0c\u591a\u6837\u5316\u7efc\u5408\u91c7\u7528\u4e24\u4e2a\u4e0d\u540c\u7684\u751f\u6210\u5668\uff08\u5373\u5185\u5728\u63a8\u7406\u751f\u6210\u5668\u548cICL\u751f\u6210\u5668\uff09\u4ee5\u53ca\u7279\u5b9a\u4e8e\u751f\u6210\u5668\u7684\u6a21\u5f0f\u683c\u5f0f\u3002\u6700\u540e\uff0c\u7b2c3\u6b65\uff08SQL\u9009\u62e9\u7f29\u653e\uff09\u91c7\u7528\u9526\u6807\u8d5b\u9009\u62e9\u5e76\u5fae\u8c03\u5185\u5728\u63a8\u7406\u9009\u62e9\u5668\u4ee5\u786e\u4fdd\u9ad8\u7cbe\u5ea6\u3002\u6211\u4eec\u7684\u8d21\u732e\u603b\u7ed3\u5982\u4e0b\uff1a<\/p>\n<ul>\n<li>\u7cbe\u5fc3\u5b89\u6392\u7684\u6d4b\u8bd5\u65f6\u95f4\u6269\u5c55\u3002\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u4e2a\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u6d4b\u8bd5\u65f6\u95f4\u6269\u5c55\u6846\u67b6\uff0c\u8be5\u6846\u67b6\u901a\u8fc7\u8de8\u4e09\u4e2a\u4e0d\u540c\u7684\u89c6\u89d2\u8fdb\u884c\u6269\u5c55\uff0c\u5c06\u989d\u5916\u7684\u63a8\u7406\u8ba1\u7b97\u8f6c\u6362\u4e3a\u51c6\u786e\u6027\u589e\u76ca\uff1ai\uff09\u5185\u90e8\u6269\u5c55\uff08RL\u589e\u5f3a\u7684\u5185\u5728\u63a8\u7406\uff09\uff0cii\uff09\u987a\u5e8f\u6269\u5c55\uff08\u8fed\u4ee3\u7ec6\u5316\uff09\uff0c\u4ee5\u53caiii\uff09\u5e76\u884c\u6269\u5c55\uff08\u591a\u6837\u5316\u7efc\u5408\u548c\u9526\u6807\u8d5b\u9009\u62e9\uff09\u3002<\/li>\n<li>\u901a\u7528\u6027\u548c\u53ef\u6269\u5c55\u6027\u3002\u6574\u4e2a\u6846\u67b6\u5b8c\u5168\u901a\u7528\u3001\u5373\u63d2\u5373\u7528\uff0c\u5e76\u4e14\u53ef\u4ee5\u8f7b\u677e\u9002\u5e94\u4efb\u4f55\u6570\u636e\u5e93\u3002\u968f\u7740\u672a\u6765\u7684LLM\u53d8\u5f97\u66f4\u52a0\u5f3a\u5927\u5e76\u4e14\u8ba1\u7b97\u7ee7\u7eed\u53d8\u5f97\u66f4\u4fbf\u5b9c\uff0cAgentar-Scale-SQL\u7684\u6027\u80fd\u4e0a\u9650\u5c06\u81ea\u884c\u63d0\u5347\uff1b\u6211\u4eec\u53ef\u4ee5\u7b80\u5355\u5730\u5c06\u66f4\u591a\u7684\u8ba1\u7b97\u5206\u914d\u7ed9\u751f\u6210\u548c\u9009\u62e9\u9636\u6bb5\uff0c\u4ee5\u5b9e\u73b0\u66f4\u9ad8\u7684\u51c6\u786e\u6027\u3002<\/li>\n<li>\u900f\u660e\u4e14\u53ef\u884c\u7684\u89c1\u89e3\u3002\u6211\u4eec\u63d0\u51fa\u4e86\u4e00\u4e2a\u7ed3\u6784\u5316\u7684\u4e09\u9636\u6bb5\u6587\u672c\u5230SQL\u6846\u67b6\uff0c\u63cf\u8ff0\u4e86\u6bcf\u4e2a\u9636\u6bb5\u7684\u5177\u4f53\u89d2\u8272\u548c\u76ee\u6807\u3002\u81f3\u5173\u91cd\u8981\u7684\u662f\uff0c\u6211\u4eec\u7684\u5de5\u4f5c\u662f\u7b2c\u4e00\u4e2a\u6269\u5c55SQL\u751f\u6210\u548c\u9009\u62e9\u9636\u6bb5\u7684\u5de5\u4f5c\u3002\u8fd9\u4e9b\u8d21\u732e\u5171\u540c\u5efa\u7acb\u4e86\u900f\u660e\u4e14\u53ef\u64cd\u4f5c\u7684\u89c1\u89e3\uff0c\u4ee5\u6307\u5bfc\u6587\u672c\u5230SQL\u9886\u57df\u7684\u672a\u6765\u7814\u7a76\u548c\u5f00\u53d1\u3002<\/li>\n<li>\u5e7f\u6cdb\u7684\u5b9e\u9a8c\u3002\u5927\u91cf\u5b9e\u9a8c\u8bc1\u5b9e\u4e86Agentar-Scale-SQL\u7684SOTA\u6027\u80fd\u3002\u5177\u4f53\u6765\u8bf4\uff0cAgentar-Scale-SQL\u5728BIRD\u5f00\u53d1\u96c6\u4e0a\u7684EX\u4e3a74.90%\uff0c\u5728\u6d4b\u8bd5\u96c6\u4e0a\u7684EX\u4e3a81.67%\uff0cR-VES\u4e3a77.00%\u3002\u6709\u4e86\u8fd9\u4e9b\u7ed3\u679c\uff0cAgentar-Scale-SQL\u5728BIRD\u6392\u884c\u699c\u4e0a\u6392\u540d\u7b2c\u4e00\u3002\u6211\u4eec\u7684\u7814\u7a76\u7ed3\u679c\u8868\u660e\uff0c\u6269\u5c55\u662f\u5728\u6587\u672c\u5230SQL\u4e2d\u5b9e\u73b0SOTA\u6027\u80fd\u7684\u5173\u952e\u56e0\u7d20\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/02\/20260205040553858.png\" alt=\"\" \/><\/p>\n<h3>\u76f8\u5173\u5de5\u4f5c<\/h3>\n<p><strong>Text-to-SQL\uff1a<\/strong>\u867d\u7136LLM\u5728\u6587\u672c\u5230SQL\u65b9\u9762\u53d6\u5f97\u4e86\u663e\u7740\u8fdb\u6b65\uff0c\u4f46\u5176\u76f4\u63a5\u5e94\u7528\u5bf9\u4e8e\u590d\u6742\u67e5\u8be2\u4ecd\u7136\u5177\u6709\u6311\u6218\u6027\u3002\u6700\u8fd1\u7684\u65b9\u6cd5\u901a\u8fc7\u57fa\u4e8e\u63d0\u793a\u7684\u65b9\u6cd5\uff0c\u5fae\u8c03\uff0c\u6216\u91c7\u7528\u6d4b\u8bd5\u65f6\u95f4\u7f29\u653e\u7684\u6df7\u5408\u7b56\u7565\u3002\u4e0e\u6211\u4eec\u7684\u5de5\u4f5c\u6700\u76f8\u5173\u7684\u662f\u8fd9\u4e9b\u7f29\u653e\u65b9\u6cd5\uff0c\u5b83\u4eec\u4f7f\u7528\u5e76\u884c\u5408\u6210\u548c\u987a\u5e8f\u7ec6\u5316\u7b49\u6280\u672f\u6765\u63d0\u9ad8\u751f\u6210\u8d28\u91cf\u3002\u7136\u800c\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u6709\u4e00\u4e2a\u5171\u540c\u7684\u5c40\u9650\u6027\uff0c\u5373\u5b83\u4eec\u7f3a\u4e4f\u4e0d\u540c\u7f29\u653e\u7ef4\u5ea6\u7684\u7cbe\u5fc3\u7ec4\u5408\u3002<br \/>\n<strong>Test-time Scaling\uff1a<\/strong>\u6d4b\u8bd5\u65f6\u95f4\u6269\u5c55\uff08TTS\uff09\u901a\u8fc7\u7b56\u7565\u6027\u5730\u6269\u5c55\u8ba1\u7b97\u800c\u4e0d\u662f\u6539\u53d8\u6a21\u578b\u6743\u91cd\u6765\u589e\u5f3aLLM\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u7684\u6027\u80fd\u3002TTS\u7684\u73b0\u6709\u7814\u7a76\u5927\u81f4\u53ef\u5206\u4e3a\u4e09\u4e2a\u8303\u5f0f\u3002\u5e76\u884c\u6269\u5c55\uff08Parallel scaling\uff09\u4fa7\u91cd\u4e8e\u540c\u65f6\u751f\u6210\u591a\u4e2a\u5019\u9009\u89e3\u51b3\u65b9\u6848\u5e76\u5c06\u5b83\u4eec\u805a\u5408\u4ee5\u63d0\u9ad8\u627e\u5230\u6b63\u786e\u7b54\u6848\u7684\u53ef\u80fd\u6027\uff0c\u81ea\u6d3d\uff08self-consistency\uff09\u5c31\u662f\u4e00\u4e2a\u7a81\u51fa\u7684\u4f8b\u5b50\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u987a\u5e8f\u6269\u5c55\uff08sequential scaling\uff09\u901a\u8fc7\u4e00\u7cfb\u5217\u6b65\u9aa4\u8fed\u4ee3\u6784\u5efa\u6216\u7ec6\u5316\u89e3\u51b3\u65b9\u6848\uff0c\u4ece\u800c\u6a21\u62df\u6df1\u601d\u719f\u8651\u7684&quot;\u7cfb\u7edf2&quot;\u63a8\u7406\u8fc7\u7a0b\uff0c\u4f8b\u5982\u601d\u60f3\u94fe(CoT)\u548c\u8fed\u4ee3\u7ec6\u5316iterative refinement\u3002\u6700\u8fd1\uff0c\u51fa\u73b0\u4e86\u5185\u90e8\u6269\u5c55\uff0c\u5176\u4e2d\u6a21\u578b\u7ecf\u8fc7\u8bad\u7ec3\u53ef\u4ee5\u81ea\u4e3b\u5206\u914d\u81ea\u5df1\u7684\u8ba1\u7b97\u9884\u7b97\u5e76\u786e\u5b9a\u63a8\u7406\u6df1\u5ea6\uff0c\u800c\u65e0\u9700\u5916\u90e8\u7f16\u6392\u3002\u6211\u4eec\u7684\u5de5\u4f5cAgentar-Scale-SQL\u5efa\u7acb\u5728\u8fd9\u4e9b\u57fa\u672c\u6982\u5ff5\u7684\u57fa\u7840\u4e0a\uff0c\u5f15\u5165\u4e86\u4e00\u79cd\u65b0\u9896\u7684\u7f16\u6392\u6846\u67b6\uff0c\u8be5\u6846\u67b6\u534f\u540c\u5730\u7ed3\u5408\u4e86\u8fd9\u4e09\u79cd\u8303\u5f0f\uff0c\u4e13\u95e8\u4e3a\u63a8\u8fdb\u6587\u672c\u5230SQL\u9886\u57df\u7684SOTA\u800c\u5b9a\u5236\u3002<\/p>\n<h3>\u65b9\u6cd5\u8bba<\/h3>\n<h4>\u6982\u8ff0<\/h4>\n<p>\u7ed9\u5b9a\u95ee\u9898<span class=\"katex-eq\" data-katex-display=\"false\">Q_u<\/span>\u3001\u8bc1\u636e<span class=\"katex-eq\" data-katex-display=\"false\">E_u<\/span>\u548c\u76ee\u6807\u6570\u636e\u5e93<span class=\"katex-eq\" data-katex-display=\"false\">D<\/span>\uff0c\u6211\u4eec\u7684\u6846\u67b6\u5206\u4e3a\u4e09\u4e2a\u9636\u6bb5\uff0c\u5982\u56fe1\u6240\u793a\u3002\u79bb\u7ebf\u9884\u5904\u7406\u9636\u6bb5\u5148\u4e8e\u5728\u7ebf\u6846\u67b6\u7684\u8fd0\u884c\u3002<br \/>\n<strong>\u79bb\u7ebf\u9884\u5904\u7406\uff08Offline Preprocessing\uff09\uff1a<\/strong>\u6211\u4eec\u7684\u65b9\u6cd5\u5728\u5728\u7ebf\u63a8\u7406\u4e4b\u524d\u6d89\u53ca\u4e09\u4e2a\u79bb\u7ebf\u9884\u5904\u7406\u6b65\u9aa4\u3002\u9996\u5148\uff0c\u4e3a\u4e86\u589e\u52a0\u751f\u6210\u7684\u591a\u6837\u6027\uff0c\u6211\u4eec\u7528\u4e24\u79cd\u683c\u5f0f\u8868\u793a\u6570\u636e\u5e93\u5143\u6570\u636e\uff08\u89c1\u56fe2)\u3002\u6211\u4eec\u521b\u5efa\u4e86\u4e00\u4e2a\u57fa\u4e8eMarkdown\u7684\u8f7b\u6a21\u5f0f\uff0c\u65e8\u5728\u901a\u8fc7\u901a\u7528LLM\u8fdb\u884c\u6709\u6548\u7684\u4e0a\u4e0b\u6587\u5b66\u4e60(ICL)\uff0c\u5e76\u4f7f\u7528\u6807\u51c6DDL\u6a21\u5f0f\u6765\u5fae\u8c03\u4ee3\u7801\u4e13\u7528\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u5229\u7528\u5176\u8bad\u7ec3\u80cc\u666f\u6765\u5b9e\u73b0\u66f4\u5feb\u7684\u6536\u655b\u3002\u5176\u6b21\uff0c\u6211\u4eec\u5c06\u6570\u636e\u5e93\u4e2d\u7684\u6240\u6709\u6587\u672c\u5355\u5143\u683c\u503c\u7d22\u5f15\u5230\u5411\u91cf\u5b58\u50a8<span class=\"katex-eq\" data-katex-display=\"false\">V_{D_{cell}}<\/span>\u4e2d\u3002\u6700\u540e\uff0c\u6211\u4eec\u8fd8\u5c06\u8bad\u7ec3\u96c6\u4f5c\u4e3a\u793a\u4f8b\u7d22\u5f15\u5230\u5411\u91cf\u5b58\u50a8<span class=\"katex-eq\" data-katex-display=\"false\">V_{D_{example}}<\/span>\u4e2d\uff0c\u8fd9\u5141\u8bb8\u6211\u4eec\u901a\u8fc7\u5d4c\u5165\u9aa8\u67b6\u63d0\u53d6\u7684\u7528\u6237\u95ee\u9898\u5e76\u6267\u884c\u76f8\u4f3c\u6027\u641c\u7d22\u6765\u5728\u63a8\u7406\u8fc7\u7a0b\u4e2d\u68c0\u7d22\u76f8\u5173\u7684\u5c0f\u6837\u672c\u793a\u4f8b\u3002<br \/>\n<strong>\u5728\u7ebf\u6846\u67b6\uff08Online Framework\uff09\uff1a<\/strong>Agentar-Scale-SQL\u662f\u4e00\u4e2a\u7cbe\u5fc3\u7b56\u5212\u7684\u6d4b\u8bd5\u65f6\u95f4\u6269\u5c55\u6846\u67b6\uff0c\u901a\u8fc7\u5728\u4e09\u4e2a\u4e0d\u540c\u89d2\u5ea6\u91c7\u7528\u7cbe\u5fc3\u7b56\u5212\u7684\u6269\u5c55\u7b56\u7565\uff0c\u5c06\u989d\u5916\u7684\u63a8\u7406\u8ba1\u7b97\u8f6c\u5316\u4e3a\u51c6\u786e\u6027\u589e\u76ca\uff1ai)\u5185\u90e8\u6269\u5c55\uff08\u901a\u8fc7RL\u589e\u5f3a\u7684\u5185\u5728\u63a8\u7406\uff09\uff0cii)\u987a\u5e8f\u6269\u5c55\uff08\u901a\u8fc7\u8fed\u4ee3\u7ec6\u5316\uff09\uff0c\u4ee5\u53caiii)\u5e76\u884c\u6269\u5c55\uff08\u901a\u8fc7\u591a\u6837\u5316\u7efc\u5408\u548c\u9526\u6807\u8d5b\u9009\u62e9\uff09\u3002\u603b\u4f53\u800c\u8a00\uff0cAgentar-Scale-SQL\u5305\u542b\u4e09\u4e2a\u9636\u6bb5\uff1a\u6b65\u9aa41\uff08\u4efb\u52a1\u7406\u89e3\uff09\u3001\u6b65\u9aa42\uff08SQL\u751f\u6210\u6269\u5c55\uff09\u548c\u6b65\u9aa43\uff08SQL\u9009\u62e9\u6269\u5c55\uff09\u3002<\/p>\n<ul>\n<li><strong>\u7b2c1\u6b65\uff08\u4efb\u52a1\u7406\u89e3\uff09<\/strong>\uff1a\u4fa7\u91cd\u4e8e\u5168\u9762\u7406\u89e3\u7528\u6237\u7684\u610f\u56fe\u5e76\u68c0\u7d22\u76f8\u5173\u4e0a\u4e0b\u6587\u3002<\/li>\n<li><strong>\u7b2c2\u6b65\uff08SQL\u751f\u6210\u6269\u5c55\uff09<\/strong>\uff1a\u5f00\u53d1\u591a\u6837\u5316\u7684\u7efc\u5408\u548c\u8fed\u4ee3\u7ec6\u5316\uff0c\u4ee5\u83b7\u5f97\u9ad8\u8d28\u91cf\u548c\u591a\u6837\u5316\u7684SQL\u5019\u9009\u3002\u7279\u522b\u662f\uff0c\u591a\u6837\u5316\u7684\u7efc\u5408\u7ec4\u4ef6\u5229\u7528\u4e86\u4e24\u4e2a\u5728\u7279\u5b9a\u6a21\u5f0f\u683c\u5f0f\u4e0a\u8fd0\u884c\u7684\u4e0d\u540c\u751f\u6210\u5668\uff1a\u5177\u6709DDL\u6a21\u5f0f\u7684\u63a8\u7406\u751f\u6210\u5668\u548c\u5177\u6709\u8f7b\u6a21\u5f0f\u7684ICL\u751f\u6210\u5668\u3002<\/li>\n<li><strong>\u7b2c3\u6b65\uff08SQL\u9009\u62e9\u7f29\u653e\uff09<\/strong>\uff1a\u5229\u7528\u9526\u6807\u8d5b\u9009\u62e9\u65b9\u6cd5\uff0c\u901a\u8fc7\u5185\u5728\u63a8\u7406\u9009\u62e9\u5668\u589e\u5f3a\uff0c\u4ee5\u5b9e\u73b0\u9ad8\u9009\u62e9\u51c6\u786e\u6027\u3002<br \/>\n\u4ee5\u4e0b\u5404\u8282\u5c06\u6df1\u5165\u7814\u7a76\u6bcf\u4e2a\u7ec4\u4ef6\u7684\u8be6\u7ec6\u4fe1\u606f\u3002<\/li>\n<\/ul>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/02\/20260205094548298.jpg\" alt=\"\" \/><\/p>\n<p><center>\u4ee5 DDL \u6a21\u5f0f\u548c\u8f7b\u578b\u6a21\u5f0f\u683c\u5f0f\u8868\u793a\u7684\u6570\u636e\u5e93\u6a21\u5f0f\u793a\u4f8b<\/center><\/p>\n<h4>\u4efb\u52a1\u7406\u89e3<\/h4>\n<p>\u6570\u636e\u5e93\u5355\u5143(Database cells)\u81f3\u5173\u91cd\u8981\uff0c\u56e0\u4e3a\u5b83\u4eec\u63d0\u4f9b\u4e86WHERE\u548cHAVING\u7b49SQL\u5b50\u53e5\u6240\u9700\u7684\u7279\u5b9a\u503c\u3002\u540c\u6837\uff0c\u7cbe\u5fc3\u6311\u9009\u7684\u5c0f\u6837\u672c\u793a\u4f8b\u53ef\u4ee5\u663e\u7740\u63d0\u9ad8\u4e0a\u4e0b\u6587\u5b66\u4e60\uff08ICL\uff09\u7684\u6027\u80fd\u3002\u56e0\u6b64\uff0c\u4efb\u52a1\u7406\u89e3\u6b65\u9aa4\u7684\u4e3b\u8981\u76ee\u6807\u662f\u8bc6\u522b\u548c\u68c0\u7d22\u4e0a\u4e0b\u6587\u7684\u8fd9\u4e24\u79cd\u5173\u952e\u5f62\u5f0f\uff1a\u76f8\u5173\u7684\u6570\u636e\u5e93\u5355\u5143\u548c\u6709\u6548\u7684\u6f14\u793a\u793a\u4f8b\u3002\u8fd9\u662f\u901a\u8fc7\u4e24\u4e2a\u5e76\u884c\u5b50\u8fc7\u7a0b\u5b9e\u73b0\u7684\uff1ai)\u5173\u952e\u5b57\u63d0\u53d6\uff0c\u4ece\u95ee\u9898<span class=\"katex-eq\" data-katex-display=\"false\">Q_u<\/span>\u548c\u8bc1\u636e<span class=\"katex-eq\" data-katex-display=\"false\">E_u<\/span>\u4e2d\u63d0\u53d6\u5173\u952e\u5b57\uff0c\u4ee5\u4f7f\u7528\u6765\u81ea<span class=\"katex-eq\" data-katex-display=\"false\">V_{D_{cell}}<\/span>\u7684\u57fa\u4e8e\u5d4c\u5165\u7684\u76f8\u4f3c\u6027\u68c0\u7d22\u76f8\u5173\u5355\u5143\u683c\uff0cii)\u4ece<span class=\"katex-eq\" data-katex-display=\"false\">Q_u<\/span>\u4e2d\u63d0\u53d6\u9aa8\u67b6\uff0c\u4ee5\u4f7f\u7528\u6765\u81ea<span class=\"katex-eq\" data-katex-display=\"false\">V_{D_{example}}<\/span>\u7684\u57fa\u4e8e\u5d4c\u5165\u7684\u76f8\u4f3c\u6027\u68c0\u7d22\u76f8\u5173\u793a\u4f8b\u3002<\/p>\n<h4>SQL\u751f\u6210\u6269\u5c55<\/h4>\n<p>\u6b64\u9636\u6bb5\u91c7\u7528\u4e24\u4e2a\u4e92\u8865\u7684\u751f\u6210\u5668\uff0c\u5728\u53cc\u6a21\u5f0f\u89c6\u56fe\u4e0a\u8fd0\u884c\uff0c\u4ee5\u751f\u6210\u9ad8\u8d28\u91cf\u548c\u591a\u6837\u5316\u7684\u5019\u9009\u8005\u3002\u7b2c\u4e00\u4e2a\u751f\u6210\u5668<span class=\"katex-eq\" data-katex-display=\"false\">M_{reasoning}<\/span>\u662f\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60(RL)\u8bad\u7ec3\u7684\u5185\u5728\u63a8\u7406\u5668\u3002\u7b2c\u4e8c\u4e2a<span class=\"katex-eq\" data-katex-display=\"false\">M_{ICL}<\/span>\u662f\u7531\u5927\u578b\u4e13\u6709LLM\u9a71\u52a8\u7684\u60c5\u5883\u5b66\u4e60(ICL)\u6a21\u578b\u3002\u7136\u540e\uff0c\u8fdb\u884c\u8bed\u6cd5\u4fee\u590d\u548c\u8bed\u4e49\u7f16\u8f91\u7684\u8fed\u4ee3\u7ec6\u5316\u5faa\u73af\u3002\u7ed3\u679c\uff0c\u751f\u6210\u4e86\u4e00\u7ec4<span class=\"katex-eq\" data-katex-display=\"false\">n<\/span> SQL\u5019\u9009\uff0c\u8868\u793a\u4e3a<span class=\"katex-eq\" data-katex-display=\"false\">C={c_1,c_2,\\ldots,c_n}<\/span>\u3002\u4ee5\u4e0b\u90e8\u5206\u8be6\u7ec6\u4ecb\u7ecd\u4e86\u6bcf\u4e2a\u7ec4\u4ef6\u3002<\/p>\n<h5>\u5185\u5728\u63a8\u7406SQL\u751f\u6210\u5668<\/h5>\n<p>\u53d7Arctic-Text2SQL-R1\u7684\u542f\u53d1\uff0c\u6211\u4eec\u901a\u8fc7\u4e00\u4e2a\u7b80\u5355\u7684\u3001\u57fa\u4e8e\u6267\u884c\u7684RL\u6846\u67b6\u8ffd\u6c42\u5f3a\u5927\u7684\u5185\u5728\u63a8\u7406\u6587\u672c\u5230SQL\u751f\u6210\u3002<br \/>\n<strong>\u5f3a\u5316\u5b66\u4e60\u7ba1\u9053\u6982\u8ff0\uff08Overview of RL Pipeline\uff09\uff1a<\/strong>\u6211\u4eec\u91c7\u7528GRPO\uff0c\u56e0\u4e3a\u5b83\u5728\u7ed3\u6784\u5316\u63a8\u7406\u4efb\u52a1\u4e0a\u7684\u6548\u7387\u548c\u6709\u6548\u6027\u5f97\u5230\u4e86\u9a8c\u8bc1\u3002\u4e0eArctic-Text2SQL-R1\u4e0d\u540c\uff0c\u6211\u4eec\u4e13\u95e8\u4f7f\u7528BIRD\u6570\u636e\u96c6\u7684\u8bad\u7ec3\u6570\u636e\uff0c\u6ca1\u6709\u4efb\u4f55\u989d\u5916\u7684\u6570\u636e\u3002\u6b63\u5f0f\u5730\uff0c\u8ba9<span class=\"katex-eq\" data-katex-display=\"false\">\\pi_\\theta<\/span>\u8868\u793a\u7531<span class=\"katex-eq\" data-katex-display=\"false\">\\theta<\/span>\u53c2\u6570\u5316\u7684\u7b56\u7565\u6a21\u578b\u3002\u5bf9\u4e8e\u4efb\u4f55\u7ed9\u5b9a\u7684\u8f93\u5165\u95ee\u9898<span class=\"katex-eq\" data-katex-display=\"false\">Q_u<\/span>\u53ca\u5176\u5173\u8054\u7684\u8bc1\u636e\u548c\u6a21\u5f0f\uff0c\u6a21\u578b\u4f1a\u751f\u6210\u4e00\u7ec4<span class=\"katex-eq\" data-katex-display=\"false\">N<\/span>\u5019\u9009SQL\u67e5\u8be2\uff08\u5373\uff0c\u63a8\u51fa\uff09<span class=\"katex-eq\" data-katex-display=\"false\">{o_{Q,1},\\ldots,o_{Q,N}}<\/span>\u3002\u7136\u540e\u8bc4\u4f30\u6bcf\u4e2a\u5019\u9009\u8005\u4ee5\u4ea7\u751f\u660e\u786e\u7684\u5956\u52b1\u4fe1\u53f7\uff0c\u5982\u4e0b\u6240\u8ff0\u3002\u8fd9\u79cd\u6309\u8f93\u5165\u6279\u91cf\u63a8\u51fa\u53ef\u4ee5\u8ba1\u7b97\u76f8\u5bf9\u4f18\u52bf\uff0c\u4ece\u800c\u7a33\u5b9a\u5b66\u4e60\u5e76\u4fc3\u8fdb\u7a33\u5065\u7684\u7b56\u7565\u66f4\u65b0\u3002<br \/>\n\u6bcf\u4e2a\u6837\u672c<span class=\"katex-eq\" data-katex-display=\"false\">i<\/span>\u7684\u88c1\u526a\u4ee3\u7406\u76ee\u6807\u5b9a\u4e49\u4e3a\uff1a<br \/>\n<span class=\"katex-eq\" data-katex-display=\"false\">L(\\theta) = \\frac{1}{N}\\sum_{i=1}^{N}\\min\\left(r_i A_i, \\text{clip}(r_i, 1-\\epsilon, 1+\\epsilon)A_i\\right)<\/span><br \/>\nGRPO\u7684\u5b8c\u6574\u76ee\u6807\uff1a<br \/>\n<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{J}{\\text{GRPO}}(\\theta) = \\mathbb{E}[L(\\theta)] - \\beta D{\\text{KL}}(\\pi_\\theta \\parallel \\pi_{\\text{ref}})<\/span><br \/>\n\u5176\u4e2d\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">r_i<\/span>\u662f\u4f3c\u7136\u6bd4<span class=\"katex-eq\" data-katex-display=\"false\">\\frac{\\pi_\\theta(o_i|Q)}{\\pi_\\theta^{\\text{old}}(o_i|Q)}<\/span>\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">A_i<\/span>\u662f\u4f18\u52bf\uff0c<span class=\"katex-eq\" data-katex-display=\"false\">D_{\\text{KL}}<\/span>\u662fKL\u6563\u5ea6\u60e9\u7f5a\uff0c\u4f7f\u7b56\u7565\u4fdd\u6301\u63a5\u8fd1\u53c2\u8003\uff08\u76d1\u7763\u5fae\u8c03\uff09\u6a21\u578b\u3002\u53c2\u6570<span class=\"katex-eq\" data-katex-display=\"false\">\\epsilon<\/span>\u548c<span class=\"katex-eq\" data-katex-display=\"false\">\\beta<\/span>\u5728\u5b9e\u8df5\u4e2d\u8fdb\u884c\u8c03\u6574\uff0c\u4ee5\u5e73\u8861\u63a2\u7d22\u548c\u7a33\u5b9a\u6027\u3002<br \/>\n\u5956\u52b1\u8bbe\u8ba1\uff1a\u6211\u4eec\u6309\u7167Arctic-Text2SQL-R1\u4ec5\u6839\u636e\u6700\u7ec8\u6267\u884c\u6b63\u786e\u6027\u548c\u57fa\u672c\u8bed\u6cd5\u6709\u6548\u6027\u5b9a\u4e49\u5956\u52b1\u51fd\u6570<span class=\"katex-eq\" data-katex-display=\"false\">R_G<\/span>\uff1a<br \/>\n<span class=\"katex-eq\" data-katex-display=\"false\">R_G = \\begin{cases} 1, &amp; \\text{if results match the ground truth;} \\ 0.1, &amp; \\text{if the SQL is executable;} \\ 0, &amp; \\text{otherwise.} \\end{cases}<\/span><\/p>\n<h5>\u591a\u6837\u5316\u5408\u6210<\/h5>\n<p>\u591a\u6837\u5316\u7684\u7efc\u5408\u7b56\u7565\u65e8\u5728\u751f\u6210\u9ad8\u8d28\u91cf\u4e14\u591a\u6837\u5316\u7684SQL\u5019\u9009\u6c60\u3002\u8be5\u7b56\u7565\u6d89\u53ca\u4e24\u4e2a\u5e76\u884c\u4e14\u4e92\u8865\u7684\u751f\u6210\u5668\uff1a\u4e00\u4e2a\u5fae\u8c03\u63a8\u7406\u751f\u6210\u5668\u548c\u4e00\u4e2aICL\u751f\u6210\u5668\u3002\u6211\u4eec\u521b\u5efa\u4e86\u4e00\u4e2a\u57fa\u4e8eMarkdown\u7684\u8f7b\u6a21\u5f0f\uff0c\u4e13\u4e3a\u901a\u7528LLM\u7684\u6709\u6548ICL\u800c\u8bbe\u8ba1\uff0c\u5e76\u4f7f\u7528\u6807\u51c6\u5b9a\u4e49\u8bed\u8a00(DDL)\u6a21\u5f0f\u6765\u5fae\u8c03\u4ee3\u7801\u4e13\u7528\u6a21\u578b\uff0c\u8be5\u6a21\u578b\u5229\u7528\u5176\u8bad\u7ec3\u80cc\u666f\u6765\u5b9e\u73b0\u66f4\u5feb\u7684\u6536\u655b\u3002<br \/>\n\u63a8\u7406\u751f\u6210\u5668\u3002\u8be5\u751f\u6210\u5668\u5229\u7528DDL\u6a21\u5f0f\u8fdb\u884c\u6df1\u5165\u3001\u9010\u6b65\u7684\u63a8\u7406\u3002\u4e00\u65b9\u9762\uff0c\u5728\u6211\u4eec\u7684\u5185\u90e8\u6269\u5c55\u539f\u5219\u7684\u6307\u5bfc\u4e0b\uff0c\u5b83\u88ab\u8bbe\u8ba1\u4e3a\u6784\u5efa\u590d\u6742\u7684\u67e5\u8be2\uff0c\u5176\u4e3b\u8981\u76ee\u6807\u662f\u5b9e\u73b0\u9ad8\u7cbe\u5ea6\u3002\u53e6\u4e00\u65b9\u9762\uff0c\u53ef\u4ee5\u5bf9\u5176\u8fdb\u884c\u5fae\u8c03\uff0c\u4ee5\u7b26\u5408\u76ee\u6807\u57fa\u51c6\u7684\u5177\u4f53\u7279\u5f81\u548c\u8981\u6c42\u3002<br \/>\nICL\u53d1\u751f\u5668\u3002\u540c\u65f6\uff0cICL\u751f\u6210\u5668\u5229\u7528\u57fa\u4e8eMarkdown\u7684\u8f7b\u6a21\u5f0f\uff08\u53c2\u89c1\u56fe2)\u548c\u5728\u4efb\u52a1\u7406\u89e3\u9636\u6bb5\u68c0\u7d22\u7684\u5c11\u6570\u6837\u672c\u793a\u4f8b\u3002\u4e3a\u4e86\u589e\u5f3aICL\u751f\u6210\u5668\u7684\u591a\u6837\u6027\uff0c\u6211\u4eec\u91c7\u7528\u4e86\u591a\u65b9\u9762\u7684\u7b56\u7565\uff1a\u6539\u53d8\u8f93\u5165\u63d0\u793a\u3001\u968f\u673a\u5316\u4e0a\u4e0b\u6587\u793a\u4f8b\u7684\u987a\u5e8f\u3001\u5229\u7528\u4e00\u7cfb\u5217LLM\u4ee5\u53ca\u8c03\u6574\u6e29\u5ea6\u8bbe\u7f6e\u3002\u901a\u5e38\uff0c\u4e0d\u540c\u7684\u63d0\u793a\u5206\u4e3a\u4e09\u79cd\u4e0d\u540c\u7684\u98ce\u683c\uff1a\u76f4\u63a5\u63d0\u793a\uff08\u6ca1\u6709\u660e\u786e\u7684\u63a8\u7406\uff09\u3001\u601d\u60f3\u94fe\uff08CoT\uff09\u63d0\u793a\u548c\u95ee\u9898\u5206\u89e3\u3002\u8be5\u751f\u6210\u5668\u64c5\u957f\u5229\u7528\u6240\u63d0\u4f9b\u7684\u4e0a\u4e0b\u6587\u793a\u4f8b\u4e2d\u7684\u6a21\u5f0f\u8bc6\u522b\u6765\u5feb\u901f\u751f\u6210\u5404\u79cd\u5408\u7406\u7684\u67e5\u8be2\u3002<br \/>\n\u901a\u8fc7\u5c06\u63a8\u7406\u751f\u6210\u5668\u7684\u6df1\u5ea6\u5206\u6790\u65b9\u6cd5\u4e0eICL\u751f\u6210\u5668\u7684\u793a\u4f8b\u9a71\u52a8\u65b9\u6cd5\u76f8\u7ed3\u5408\uff0c\u6211\u4eec\u6700\u5927\u9650\u5ea6\u5730\u63d0\u9ad8\u4e86\u5019\u9009\u6c60\u7684\u591a\u6837\u6027\u3002\u8fd9\u79cd\u534f\u540c\u4f5c\u7528\u663e\u7740\u589e\u52a0\u4e86\u5728\u9009\u62e9\u9636\u6bb5\u4e4b\u524d\u51fa\u73b0\u81f3\u5c11\u4e00\u4e2a\u6b63\u786e\u6216\u63a5\u8fd1\u6b63\u786e\u7684\u67e5\u8be2\u7684\u6982\u7387\u3002<\/p>\n<h5>\u8fed\u4ee3\u7ec6\u5316<\/h5>\n<p>\u4e3a\u4e86\u8fdb\u4e00\u6b65\u63d0\u9ad8\u5019\u9009SQL\u7684\u8d28\u91cf\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u8fed\u4ee3\u7ec6\u5316\u6a21\u5757\u6765\u7ec6\u5316\u9519\u8bef\u3002SQL\u67e5\u8be2\u53ef\u80fd\u540c\u65f6\u5305\u542b\u8bed\u6cd5\u548c\u8bed\u4e49\u9519\u8bef\u3002\u6211\u4eec\u91c7\u7528\u53cc\u7ba1\u9f50\u4e0b\u7684\u65b9\u6cd5\u6765\u4fee\u6b63\u9519\u8bef\u3002\u5bf9\u4e8e\u8bed\u6cd5\u9519\u8bef\uff08\u524d\u8005\uff09\uff0c\u6211\u4eec\u4f7f\u7528SQL\u4fee\u590d\u7a0b\u5e8f\uff0c\u8fd9\u662f\u4e00\u4e2a\u57fa\u4e8eLLM\u7684\u7ec4\u4ef6\uff0c\u53ef\u4ee5\u6709\u6761\u4ef6\u5730\u6fc0\u6d3b\u4ee5\u4fee\u590d\u65e0\u6548\u8bed\u6cd5\u3002\u5bf9\u4e8e\u8bed\u4e49\u9519\u8bef\uff08\u540e\u8005\uff09\uff0c\u6211\u4eec\u4f7f\u7528SQLrevisor\uff0c\u8fd9\u662f\u4e00\u79cdLLM\u4ee3\u7406\uff0c\u65e8\u5728\u8bc6\u522b\u548c\u5b8c\u5584\u67e5\u8be2\u4e2d\u7684\u903b\u8f91\u7f3a\u9677\u3002\u4e3a\u4e86\u7b80\u5316\u4fee\u8ba2\u8fc7\u7a0b\uff0c\u6211\u4eec\u9996\u5148\u6309\u6267\u884c\u7ed3\u679c\u5bf9\u67e5\u8be2\u8fdb\u884c\u5206\u7ec4\u3002\u968f\u540e\uff0c\u6211\u4eec\u4ece\u6bcf\u7ec4\u4e2d\u968f\u673a\u9009\u62e9\u4e00\u4e2a\u67e5\u8be2\u8fdb\u884c\u7ec6\u5316\u3002<\/p>\n<h4>SQL\u9009\u62e9\u7f29\u653e<\/h4>\n<p>\u591a\u6570\u6295\u7968\u7684\u4e3b\u8981\u9650\u5236\u5728\u4e8e\u5176\u57fa\u672c\u5047\u8bbe\uff0c\u5373\u6700\u5e38\u89c1\u7684\u7b54\u6848\u4e5f\u662f\u6b63\u786e\u7684\u7b54\u6848\uff0c\u4f46\u8fd9\u4e00\u524d\u63d0\u5e76\u4e0d\u603b\u662f\u6210\u7acb\u3002 \u76f8\u53cd\uff0c\u6211\u4eec\u91c7\u7528\u9526\u6807\u8d5b\u9009\u62e9\u8fc7\u7a0b\uff0c\u5176\u4e2d\u63a8\u7406\u9009\u62e9\u5668\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u589e\u5f3a\uff0c\u901a\u8fc7\u6210\u5bf9\u6bd4\u8f83\u6765\u8bc4\u4f30\u5019\u9009\u8005\u3002 \u6b64\u8fc7\u7a0b\u4e2d\u6392\u540d\u9760\u524d\u7684 SQL \u67e5\u8be2\u88ab\u9009\u62e9\u4f5c\u4e3a\u6700\u7ec8 SQL\u3002 \u6211\u4eec\u5c06\u5728\u4ee5\u4e0b\u90e8\u5206\u8be6\u7ec6\u4ecb\u7ecd\u8fd9\u4e9b\u6a21\u5757\u3002<\/p>\n<h5>\u6bd4\u8d5b\u9009\u62e9<\/h5>\n<p>\u6211\u4eec\u901a\u8fc7\u4e24\u9636\u6bb5\u8fc7\u7a0b\u9009\u62e9\u6700\u4f73SQL\u67e5\u8be2\u3002\u9996\u5148\uff0c\u6211\u4eec\u6839\u636e\u6570\u636e\u5e93<span class=\"katex-eq\" data-katex-display=\"false\">D<\/span>\u4e0a\u76f8\u540c\u7684\u6267\u884c\u7ed3\u679c\u5bf9\u67e5\u8be2\u8fdb\u884c\u5206\u7ec4\uff0c\u4ece\u800c\u5408\u5e76\u521d\u59cb\u67e5\u8be2\u6c60\u3002\u4ece\u6bcf\u4e2a\u7ec4\u4e2d\u9009\u51fa\u4e00\u4e2a\u4ee3\u8868\u6765\u5f62\u6210\u5019\u9009\u96c6<span class=\"katex-eq\" data-katex-display=\"false\">C&#039;={c_1,\\ldots,c_m}<\/span>\u3002\u5176\u6b21\uff0c\u8fd9\u4e9b\u5019\u9009\u4eba\u53c2\u52a0\u4e24\u4e24\u5faa\u73af\u8d5b\u3002\u5bf9\u4e8e\u6bcf\u5bf9katex[\/katex]\uff0c\u63a8\u7406\u9009\u62e9\u5668<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}{selection}<\/span>\u6839\u636e\u95ee\u9898\u3001\u8f7b\u6a21\u5f0f\u548c\u6267\u884c\u7ed3\u679c\u786e\u5b9a\u83b7\u80dc\u8005\uff0c\u5e76\u589e\u52a0\u83b7\u80dc\u8005\u7684\u5206\u6570<span class=\"katex-eq\" data-katex-display=\"false\">W_i<\/span>\u3002\u6700\u7ec8\u67e5\u8be2\u662f\u5f97\u5206\u6700\u9ad8\u7684\u67e5\u8be2\uff1a<span class=\"katex-eq\" data-katex-display=\"false\">c{final}=\\arg\\max_{c_i \\in C&#039;} W_i<\/span>\u3002<\/p>\n<h5>\u5185\u5728\u63a8\u7406SQL\u9009\u62e9\u5668<\/h5>\n<p>\u6211\u4eec\u5c06\u5f3a\u5316\u5b66\u4e60(RL)\u5e94\u7528\u4e8eSQL\u9009\u62e9\u5668<span class=\"katex-eq\" data-katex-display=\"false\">\\mathcal{M}_{selection}<\/span>\uff0c\u8fd9\u79cd\u65b9\u6cd5\u7c7b\u4f3c\u4e8eSQL\u751f\u6210\u5668\u4e2d\u4f7f\u7528\u7684\u5185\u5728\u63a8\u7406\u3002<br \/>\n<strong>\u5f3a\u5316\u5b66\u4e60\u7ba1\u9053\u6982\u8ff0\uff1a<\/strong>\u9075\u5faa3.3.1\u8282\u4e2d\u63cf\u8ff0\u7684\u65b9\u6cd5\uff0c\u6211\u4eec\u5e94\u7528GRPO\u6765\u589e\u5f3aSQL\u9009\u62e9\u7684\u63a8\u7406\u80fd\u529b\u3002\u57fa\u4e8e\u8bad\u7ec3\u96c6\uff0c\u6211\u4eec\u6784\u5efa\u4e868.5k\u4e2a\u6837\u672c\u8fdb\u884c\u5f3a\u5316\u5b66\u4e60\u3002<br \/>\n<strong>\u5956\u52b1\u8bbe\u8ba1\uff1a<\/strong>SQL\u9009\u62e9\u7684\u76ee\u6807\u662f\u4ece\u4e00\u7ec4\u5019\u9009\u67e5\u8be2\u4e2d\u8bc6\u522b\u51fa\u6b63\u786e\u7684\u67e5\u8be2\u3002\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0c\u6211\u4eec\u5f15\u5165\u4e86\u4e00\u4e2a\u9762\u5411\u7ed3\u679c\u7684\u5956\u52b1\u51fd\u6570<span class=\"katex-eq\" data-katex-display=\"false\">R_S<\/span>\uff0c\u65e8\u5728\u8bc4\u4f30\u9009\u62e9\u7684\u6b63\u786e\u6027\uff1a<br \/>\n<span class=\"katex-eq\" data-katex-display=\"false\">R_S = \\begin{cases} 1, &amp; \\text{if the selection is correct;} \\ 0, &amp; \\text{otherwise.} \\end{cases}<\/span><\/p>\n<h3>\u5b9e\u9a8c<\/h3>\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u5728\u5927\u89c4\u6a21\u6570\u636e\u96c6\u4e0a\u5bf9\u6240\u63d0\u51fa\u7684Agentar-Scale-SQL\u8fdb\u884c\u5b9e\u9a8c\u8bc4\u4f30\u3002\u6211\u4eec\u7684\u76ee\u6807\u662f\u56de\u7b54\u4ee5\u4e0b\u7814\u7a76\u95ee\u9898\uff1a<\/p>\n<ul>\n<li>RQ1\uff1a\u4e0e\u6700\u5148\u8fdb\u7684\u65b9\u6cd5\u76f8\u6bd4\uff0cAgentar-Scale-SQL\u7684\u6027\u80fd\u5982\u4f55\uff1f<\/li>\n<li>RQ2\uff1a\u6bcf\u4e2a\u6a21\u5757\u5982\u4f55\u5f71\u54cdAgentar-Scale-SQL\u7684\u6574\u4f53\u6027\u80fd\uff1f<\/li>\n<li>RQ3\uff1aICL\u548cReasoning\u751f\u6210\u5668\u5404\u81ea\u7684\u4f5c\u7528\u548c\u4e92\u8865\u4f5c\u7528\u662f\u4ec0\u4e48\uff1f<\/li>\n<li>RQ4\uff1a\u6027\u80fd\u5982\u4f55\u968f\u7740\u4e0d\u540c\u590d\u6742\u7a0b\u5ea6\u7684\u5019\u9009\u8005\u6570\u91cf\u800c\u53d8\u5316\uff1f<\/li>\n<\/ul>\n<h4>\u5b9e\u9a8c\u8bbe\u7f6e<\/h4>\n<p><strong>\u6570\u636e\u96c6\uff1a<\/strong>\u6211\u4eec\u5728BIRD\u57fa\u51c6\u4e0a\u8bc4\u4f30\u6211\u4eec\u7684\u65b9\u6cd5\u3002\u5b83\u5305\u542b\u8de895\u4e2a\u5927\u578b\u6570\u636e\u5e93\u7684\u8d85\u8fc712,751\u4e2a\u95ee\u9898SQL\u5bf9\uff0c\u901a\u8fc7\u8de8\u8d85\u8fc737\u4e2a\u4e13\u4e1a\u9886\u57df\u7684\u6742\u4e71\u6570\u636e\u548c\u590d\u6742\u6a21\u5f0f\u6765\u6a21\u62df\u73b0\u5b9e\u4e16\u754c\u7684\u590d\u6742\u6027\u3002<br \/>\n<strong>\u57fa\u7ebf\uff1a<\/strong>\u6211\u4eec\u6bd4\u8f83\u4e86\u603b\u4f53\u6392\u884c\u699c\u548c\u5355\u6a21\u578b\u6392\u884c\u699c\u4e2d\u7684\u51e0\u79cd\u9876\u7ea7\u57fa\u7ebf\u65b9\u6cd5\u3002\u524d\u8005\u753115\u4e2a\u57fa\u7ebf\u7ec4\u6210\uff0c\u5305\u62ecAskData+GPT-4o\u3001LongData-SQL\u3001CHASE-SQL+Gemini\u3001JoyDataAgent-SQL\u3001XiYan-SQL\u7b49\u3002\u540e\u8005\u5305\u62ecGemini-SQL\u3001DatabricksRLVR32B\u3001Sophon-Text2SQL-32B\u3001Arctic-Text2SQL-R1-32B\u7b49\u516b\u79cd\u9886\u5148\u65b9\u6cd5\u3002<br \/>\n<strong>\u6307\u6807\uff1a<\/strong>\u7ee7\u4e4b\u524d\u7684\u5de5\u4f5c\u4e4b\u540e\uff0c\u6211\u4eec\u4f7f\u7528\u6267\u884c\u51c6\u786e\u5ea6\uff08EX\uff09\uff08\u5404\u4e2a\u6392\u884c\u699c\u7684\u5b98\u65b9\u6307\u6807\uff09\u4f5c\u4e3a\u6bd4\u8f83\u65b9\u6cd5\u7684\u4e3b\u8981\u8bc4\u4f30\u6307\u6807\u3002\u6b64\u5916\uff0c\u6211\u4eec\u91c7\u7528\u5b98\u65b9\u7684\u57fa\u4e8e\u5956\u52b1\u7684\u6709\u6548\u6548\u7387\u5206\u6570\uff08Reward-based Valid Efficiency Score\uff0cR-VES\uff09\u6765\u8bc4\u4f30\u751f\u6210\u7684SQL\u7684\u6548\u7387\u3002<br \/>\n<strong>\u5b9e\u65bd\u7ec6\u8282\uff1a<\/strong>\u6211\u4eec\u5b9e\u65bdAgentar-Scale-SQL\uff0c\u901a\u8fc7LangChain\/LangGraph\u548call-MiniLM-L6-v25 embeddings\u63d0\u4f9b\u652f\u6301\u3002\u4efb\u52a1\u7406\u89e3\u901a\u8fc7Gemini-2.5-Flash\u63d0\u4f9b\u652f\u6301\u3002ICLSQL\u751f\u6210\u5668\u5229\u7528\u5177\u6709\u4e24\u79cd\u6e29\u5ea6\u8bbe\u7f6e\uff080.5\u548c1.8\uff09\u7684Gemini-2.5-Pro\uff08\u7f29\u5199\u4e3apro\uff09\u548cGPT-5\uff08\u6700\u5c11\u63a8\u7406\u6548\u7387\uff09\u3002ReasoningSQLGenerator\u57fa\u4e8eOmni-SQL-32B\u8fdb\u884c\u4e86\u5fae\u8c03\u3002\u9ed8\u8ba4\u60c5\u51b5\u4e0b\uff0c\u5019\u9009SQL\u5305\u62ec\u6765\u81eaICLSQL\u751f\u6210\u5668\u76849\u4e2a\u548c\u6765\u81eaReasoningSQL\u751f\u6210\u5668\u76848\u4e2a\u3002SQLFixer\u548cSQLReviser\u90fd\u4f7f\u7528Gemini-2.5-Pro\u3002ReasoningSQLSelector\u7684\u57fa\u7840\u6a21\u578b\u662fQwen2.5-Coder-32B-Instruct\u3002\u5f3a\u5316\u5b66\u4e60\u6846\u67b6\u572832\u4e2aNVIDIA A100 80GB GPU\u4e0a\u5229\u7528verl\u5b9e\u73b0\u3002<\/p>\n<h4>\u4e3b\u8981\u7ed3\u679c (RQ1)<\/h4>\n<p>\u5982\u88681\u6240\u793a\uff0c\u6211\u4eec\u7684Agentar-Scale-SQL\u6846\u67b6\u5728BIRD\u57fa\u51c6\u4e0a\u5efa\u7acb\u4e86\u65b0\u7684SOTA\uff0c\u5728\u6d4b\u8bd5\u96c6\u4e0a\u5b9e\u73b0\u4e8681.67\uff05\u7684\u6267\u884c\u51c6\u786e\u7387\uff08EX\uff09\u548c77.00\uff05\u7684R-VES\u3002\u5728\u6d4b\u8bd5EX\u4e2d\uff0c\u8be5\u7ed3\u679c\u6bd4\u4e4b\u524d\u7684SOTA(AskData+GPT-4o)\u63d0\u9ad8\u4e860.79%\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u8fd9\u4e9b\u65b9\u6cd5\u91c7\u7528\u4e86\u5f3a\u5927\u7684\u6a21\u578b\uff0c\u4f8b\u5982Gemini2.5Pro\u548cClaude4\u3002<br \/>\n\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u4e0e\u5355\u4e00\u8bad\u7ec3\u6a21\u578b\u76f8\u6bd4\uff0c\u6027\u80fd\u63d0\u5347\u66f4\u4e3a\u660e\u663e\uff1bAgentar-Scale-SQL\u5728\u6d4b\u8bd5EX\u4e2d\u6bd4\u6700\u5f3a\u7684\u5355\u4e00\u6a21\u578b(Gemini-SQL)\u6027\u80fd\u9ad8\u51fa5.54%\u3002\u8fd9\u4e9b\u7ed3\u679c\u4ece\u7ecf\u9a8c\u4e0a\u9a8c\u8bc1\u4e86\u6211\u4eec\u7cbe\u5fc3\u7b56\u5212\u7684\u6269\u5c55\u7b56\u7565\u7684\u6709\u6548\u6027\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/02\/20260207013356986.jpg\" alt=\"\" \/><\/p>\n<h4>\u6d88\u878d\u7814\u7a76 (RQ2)<\/h4>\n<p>\u63a5\u4e0b\u6765\uff0c\u6211\u4eec\u901a\u8fc7\u6bd4\u8f83Agentar-Scale-SQL\u4e0e\u5176\u4e0d\u5e26\u5173\u952e\u6a21\u5757\u7684\u53d8\u4f53\u6765\u7814\u7a76Agentar-Scale-SQL\u4e2d\u6bcf\u4e2a\u6a21\u5757\u7684\u6709\u6548\u6027\u3002\u7ed3\u679c\u5217\u4e8e\u88682\u4e2d\u3002<br \/>\n<strong>Agentar-Scale-SQL\u4e0d\u5e26\u4efb\u52a1\u7406\u89e3\uff1a<\/strong>\u5728\u6b64\u53d8\u4f53\u4e2d\uff0c\u603b\u4f53EX\u4e0b\u964d\u4e860.45\uff0c\u8fd9\u8868\u660e\u4efb\u52a1\u7406\u89e3\u6a21\u5757\u89e3\u51b3\u4e86SQL\u5b50\u53e5\u6240\u9700\u7684\u503c\u548c\u95ee\u9898\u672c\u8eab\u7684\u63aa\u8f9e\u4e2d\u7684\u6b67\u4e49\u3002<br \/>\n<strong>Agentar-Scale-SQL\u4e0d\u5e26\u63a8\u7406SQL\u751f\u6210\u5668\uff1a<\/strong>\u5220\u9664\u63a8\u7406SQL\u751f\u6210\u5668\u4f1a\u5bfc\u81f4\u6027\u80fd\u4e0b\u964d\u6700\u4e25\u91cd\uff0c\u603b\u51c6\u786e\u5ea6\u4e0b\u964d4.89\u4e2a\u70b9\u3002\u8fd9\u4e00\u7ed3\u679c\u6709\u529b\u5730\u8bc1\u660e\u4e86\u5185\u5728\u7f29\u653e\u65b9\u6cd5\u7684\u6709\u6548\u6027\u3002\u5b83\u5bf9\u4e8e\u751f\u6210\u51c6\u786e\u7684SQL\u903b\u8f91\u5e76\u4e0e\u76ee\u6807\u6570\u636e\u7684\u504f\u597d\u4fdd\u6301\u4e00\u81f4\u81f3\u5173\u91cd\u8981\uff0c\u8fd9\u662f\u89e3\u51b3\u590d\u6742\u95ee\u9898\u4e0d\u53ef\u6216\u7f3a\u7684\u8d44\u4ea7\u3002<br \/>\n<strong>Agentar-Scale-SQL\u4e0d\u5e26ICL-SQL\u751f\u6210\u5668\uff1a<\/strong>\u5f53\u6392\u9664ICL-SQL\u751f\u6210\u5668\u65f6\uff0c\u603b\u51c6\u786e\u5ea6\u4e0b\u964d\u4e863.78\u4e2a\u70b9\uff0c\u8fd9\u662f\u89c2\u5bdf\u5230\u7684\u7b2c\u4e8c\u5927\u4e0b\u964d\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5728\u5177\u6709\u6311\u6218\u6027\u7684\u95ee\u9898\u4e0a\u7684\u6027\u80fd\u4ece64.14\u4e0b\u964d\u523055.86\u3002\u8fd9\u7a81\u663e\u4e86\u6211\u4eec\u4e24\u4e2a\u751f\u6210\u5668\u7684\u4e92\u8865\u6027\u8d28\u3002ICL\u751f\u6210\u5668\u64c5\u957f\u5229\u7528\u4e0a\u4e0b\u6587\u793a\u4f8b\u6765\u6784\u5efa\u590d\u6742\u7684\u67e5\u8be2\uff0c\u4e3a\u83b7\u5f97\u6b63\u786e\u7684\u89e3\u51b3\u65b9\u6848\u63d0\u4f9b\u4e86\u6709\u6548\u7684\u66ff\u4ee3\u9014\u5f84\u3002\u4f7f\u7528\u4e24\u4e2a\u4e92\u8865\u751f\u6210\u5668\u7684\u5e76\u884c\u6269\u5c55\u786e\u4fdd\u4e86\u591a\u6837\u5316\u4e14\u9ad8\u8d28\u91cf\u7684\u5019\u9009SQL\u67e5\u8be2\u6c60\uff0c\u8fd9\u5bf9\u4e8e\u5b9e\u73b0\u9ad8\u6027\u80fd\u81f3\u5173\u91cd\u8981\u3002<br \/>\n<strong>Agentar-Scale-SQL\u4e0d\u5e26\u8fed\u4ee3\u6539\u8fdb\uff1a<\/strong>\u6211\u4eec\u8fd8\u901a\u8fc7\u6d88\u878d\u7814\u7a76\u5206\u6790\u4e86\u8fed\u4ee3\u6539\u8fdb\u6a21\u5757\u7684\u8d21\u732e\u3002\u79fb\u9664\u8be5\u6a21\u5757\u5bfc\u81f4\u6027\u80fd\u4e0b\u964d0.52\u4e2a\u767e\u5206\u70b9\uff0c\u6211\u4eec\u5c06\u5176\u5f52\u56e0\u4e8e\u8be5\u6a21\u5757\u80fd\u591f\u4f18\u5316SQL\u5e76\u7ea0\u6b63\u8bed\u6cd5\u548c\u8bed\u4e49\u9519\u8bef\u3002<br \/>\n<strong>Agentar-Scale-SQL\u4e0d\u5e26SQL\u9009\u62e9\u6269\u5c55\uff1a<\/strong>Agentar-Scale-SQLw\/oSQL\u9009\u62e9\u6269\u5c55\u8868\u793a\u6211\u4eec\u91c7\u7528\u81ea\u4e00\u81f4\u6027\uff08\u5373\u57fa\u4e8e\u6267\u884c\u7ed3\u679c\u7684\u591a\u6570\u6295\u7968\uff09\u6765\u9009\u62e9\u6700\u4f73SQL\u3002\u6211\u4eec\u53ef\u4ee5\u89c2\u5bdf\u5230\uff0c\u5728EX\u6307\u6807\u4e0a\uff0c\u6211\u4eec\u7684\u65b9\u6cd5\u6bd4\u81ea\u4e00\u81f4\u6027\u57fa\u7ebf\u964d\u4f4e1.82\u4e2a\u767e\u5206\u70b9\u3002\u8fd9\u662f\u56e0\u4e3a\u6267\u884c\u9891\u7387\u6700\u9ad8\u7684\u7ed3\u679c\u4e0d\u4e00\u5b9a\u5c31\u662f\u6b63\u786e\u7684\u3002\u6211\u4eec\u7684\u9009\u62e9\u7b56\u7565\uff0c\u53ef\u80fd\u6bd4\u7b80\u5355\u7684\u9891\u7387\u7edf\u8ba1\u5305\u542b\u4e86\u66f4\u591a\u7684\u4fe1\u53f7\uff0c\u88ab\u8bc1\u660e\u662f\u8bc6\u522b\u6b63\u786e\u67e5\u8be2\u66f4\u6709\u6548\u3001\u66f4\u9c81\u68d2\u7684\u65b9\u6cd5\u3002<\/p>\n<h4>\u751f\u6210\u5668\u7ec4\u4ef6\u5206\u6790 (RQ3)<\/h4>\n<p>\u5982\u56fe3\u6240\u793a\uff0cICL\u751f\u6210\u5668\u5b9e\u73b0\u4e86\u6bd4\u63a8\u7406\u751f\u6210\u5668\uff0875.88%\uff09\u660e\u663e\u66f4\u9ad8\u7684\u4e0a\u9650\u51c6\u786e\u7387\uff0881.36%\uff09\uff0c\u8868\u660e\u5176\u5728\u751f\u6210\u6b63\u786e\u67e5\u8be2\u65b9\u9762\u5177\u6709\u5de8\u5927\u6f5c\u529b\u3002\u7136\u800c\uff0c\u7ed3\u5408\u5b83\u4eec\u7684\u8f93\u51fa\u6765\u770b\uff08All\uff09\uff0c\u5b9e\u73b0\u4e86\u6700\u9ad8\u7684\u603b\u4f53\u4e0a\u965084.29%\u3002\u5982\u56fe4\u6240\u793a\uff0c\u8fd9\u79cd\u534f\u540c\u589e\u76ca\u53ef\u4ee5\u7531\u5b83\u4eec\u7684\u4e92\u8865\u6027\u8d28\u6765\u89e3\u91ca\u3002\u867d\u7136\u4ed6\u4eec\u5171\u4eab\u5927\u91cf\u6b63\u786e\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u4f46\u4ed6\u4eec\u8fd8\u5206\u522b\u72ec\u7279\u5730\u89e3\u51b3\u4e8647\u4e2a\u548c12\u4e2a\u6837\u672c\u3002\u8fd9\u79cd\u6269\u5927\u7684\u8986\u76d6\u8303\u56f4\u9002\u7528\u4e8e\u6240\u6709\u96be\u5ea6\u7ea7\u522b\uff08\u56fe5)\u3002\u6309\u96be\u5ea6\u7ec6\u5206\u8868\u660e\uff0c\u63a8\u7406\u751f\u6210\u5668\u5728\u7b80\u5355\u548c\u4e2d\u7b49\u4efb\u52a1\u4e0a\u5177\u6709\u4f18\u52bf\uff0c\u800cICL\u751f\u6210\u5668\u5728\u6311\u6218\u6027\u95ee\u9898\u4e0a\u66f4\u6709\u6548\u3002<br \/>\n\u6700\u7ec8\uff0c\u8fd9\u4e2a\u66f4\u4e30\u5bcc\u3001\u66f4\u591a\u6837\u5316\u7684\u5019\u9009\u5e93\u4f7f\u6211\u4eec\u7684\u6700\u7ec8\u9009\u62e9\u7b56\u7565\u8fbe\u5230\u4e8674.90%\u7684\u5cf0\u503c\u51c6\u786e\u7387\uff0c\u8bc1\u660e\u4e86\u53cc\u751f\u6210\u5668\u65b9\u6cd5\u7684\u5173\u952e\u4f5c\u7528\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/02\/20260207020136920.jpg\" alt=\"\" \/><\/p>\n<p><center>\u56fe3 \u6295\u7968\u3001\u9009\u62e9\u6a21\u578b\u548c\u751f\u6210\u5668\u7ec4\u4ef6\u4e0a\u9650\u7684\u6267\u884c\u51c6\u786e\u6027 <\/center><\/p>\n<p><img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/02\/20260207020401736.jpg\" alt=\"\" \/><\/p>\n<p><center>\u56fe5 \u6309 ICL\u3001\u63a8\u7406\u548c\u7ec4\u5408\u751f\u6210\u5668\u96be\u5ea6\u7ea7\u522b\u5212\u5206\u7684\u6b63\u786e\u6837\u672c\u6570<\/center><\/p>\n<h4>\u5019\u9009\u4eba\u6570\u91cf\u7684\u5f71\u54cd\uff08RQ4\uff09<\/h4>\n<p>\u6700\u540e\uff0c\u6211\u4eec\u901a\u8fc7\u5c06\u5019\u9009\u6570\u4ece1\u66f4\u6539\u4e3a16\u6765\u7814\u7a76\u5176\u5f71\u54cd\u3002\u7ed3\u679c\u5982\u56fe6\u6240\u793a\uff0c\u8868\u660e\u589e\u52a0\u540e\u9009\u4eba\u6570\u91cf\u53ef\u4ee5\u6301\u7eed\u63d0\u9ad8\u6240\u6709\u96be\u5ea6\u7ea7\u522b\u7684Pass@k\u7387\u3002\u5bf9\u4e8e\u5177\u6709\u6311\u6218\u6027\u7684\u67e5\u8be2\uff0c\u8fd9\u79cd\u6539\u8fdb\u6700\u4e3a\u663e\u7740\uff0c\u5e76\u4e14\u5f53\u5c06\u5019\u9009\u9879\u589e\u52a0\u52308\u4e2a\u65f6\uff0c\u8fd9\u79cd\u6539\u8fdb\u6700\u4e3a\u663e\u7740\uff0c\u4e4b\u540e\u589e\u76ca\u5c31\u4f1a\u51cf\u5c0f\u3002\u8fd9\u9a8c\u8bc1\u4e86\u6211\u4eec\u5e76\u884c\u6269\u5c55\u7b56\u7565\u7684\u6709\u6548\u6027\u3002<br \/>\n<img decoding=\"async\" src=\"https:\/\/youpaiyun.lingbo.online\/2026\/02\/20260207020521494.jpg\" alt=\"\" \/><\/p>\n<p><center>\u4e0d\u540c\u5019\u9009\u4eba\u6570\u91cf\u7684\u901a\u8fc7\u7387<\/center><\/p>\n<h3>\u7ed3\u8bba<\/h3>\n<p>\u6211\u4eec\u5f15\u5165\u4e86 Agentar-Scale-SQL\uff0c\u8fd9\u662f\u4e00\u79cd\u65b0\u9896\u7684\u6846\u67b6\uff0c\u5b83\u901a\u8fc7\u534f\u540c\u7ec4\u5408\u5185\u90e8\u3001\u987a\u5e8f\u548c\u5e76\u884c\u6269\u5c55\u7b56\u7565\u6765\u663e\u7740\u63d0\u9ad8Text-to-SQL \u7684\u6027\u80fd\u3002 \u6211\u4eec\u7684\u65b9\u6cd5\u5728\u5177\u6709\u6311\u6218\u6027\u7684 BIRD \u57fa\u51c6\u4e0a\u53d6\u5f97\u4e86 SOTA \u7ed3\u679c\uff0c\u5c55\u793a\u4e86\u5b9e\u73b0\u4eba\u7c7b\u6c34\u5e73\u51c6\u786e\u6027\u7684\u6709\u6548\u9014\u5f84\u3002 \u6211\u4eec\u53d1\u5e03\u4ee3\u7801\u548c\u6a21\u578b\u6765\u652f\u6301\u8be5\u9886\u57df\u7684\u672a\u6765\u7814\u7a76\u3002<br \/>\n\u5728\u901a\u8fc7\u6d4b\u8bd5\u65f6\u95f4\u7f29\u653e\u6210\u529f\u63d0\u9ad8\u667a\u80fd\u6027\u7684\u57fa\u7840\u4e0a\uff0c\u6211\u4eec\u6b63\u5728\u5f00\u62d3\u6211\u4eec\u7684\u4e0b\u4e00\u4e2a\u65b9\u6848\uff1a\u8bad\u7ec3\u65f6\u6269\u5c55\u3002 \u6211\u4eec\u5c06\u8d4b\u4e88\u65b0\u4e00\u4ee3Agent\u901a\u8fc7\u884c\u52a8\u5b66\u4e60\u5e76\u4ece\u7ecf\u9a8c\u4e2d\u53d1\u5c55\u7684\u80fd\u529b\u3002<\/p>\n<h3>\u5c40\u9650\u6027<\/h3>\n<p>\u5c3d\u7ba1Agentar-Scale-SQL\u975e\u5e38\u6709\u6548\uff0c\u4f46\u5b83\u5bf9\u7cbe\u5fc3\u7b56\u5212\u7684\u6d4b\u8bd5\u65f6\u95f4\u6269\u5c55\u7684\u4f9d\u8d56\u5e26\u6765\u4e86\u4e00\u4e9b\u5173\u952e\u9650\u5236\u3002\u8be5\u6846\u67b6\u7684\u4e3b\u8981\u7f3a\u70b9\u662f\u7531\u4e8e\u591a\u6b21LLM\u8c03\u7528\u751f\u6210\u3001\u7ec6\u5316\u548c\u9009\u62e9\u800c\u5bfc\u81f4\u5927\u91cf\u8ba1\u7b97\u5f00\u9500\u548c\u9ad8\u5ef6\u8fdf\uff0c\u4f7f\u5176\u4e0d\u592a\u9002\u5408\u5b9e\u65f6\u5e94\u7528\u7a0b\u5e8f\u3002\u5728\u6211\u4eec\u7684\u5546\u4e1a\u5b9e\u8df5\uff08\u5f00\u53d1B2BChatBI\u4ea7\u54c1\uff09\u4e2d\uff0c\u6211\u4eec\u89c2\u5bdf\u5230\u4f01\u4e1a\u5ba2\u6237\u5c06\u51c6\u786e\u6027\u653e\u5728\u9996\u4f4d\u3002\u5728\u51b3\u7b56\u573a\u666f\u4e2d\uff0c\u5e7b\u89c9\u7684SQL\u67e5\u8be2\u662f\u4e0d\u53ef\u63a5\u53d7\u7684\uff0c\u800c\u5bf9\u4e8e\u590d\u6742\u7684\u5206\u6790\u751f\u6210\u6765\u8bf4\uff0c\u51e0\u79d2\u7684\u5ef6\u8fdf\u901a\u5e38\u662f\u53ef\u4ee5\u5bb9\u5fcd\u7684\u3002Agentar-Scale-SQL\u4e13\u4e3a\u5f25\u8865\u8fd9\u4e9b\u751f\u4ea7\u8981\u6c42\u7684&quot;\u6700\u540e\u4e00\u82f1\u91cc&quot;\u51c6\u786e\u6027\u800c\u8bbe\u8ba1\u3002\u6b64\u5916\uff0c\u5b83\u7684\u6027\u80fd\u4ece\u6839\u672c\u4e0a\u53d7\u5230\u5e95\u5c42\u57fa\u7840LLM\u7684\u80fd\u529b\u7684\u9650\u5236\uff0c\u5e76\u4e14\u5b83\u5f88\u5bb9\u6613\u53d7\u5230\u7ea7\u8054\u9519\u8bef\u7684\u5f71\u54cd\uff0c\u5176\u4e2d\u65e9\u671f\u9636\u6bb5\u7684\u5931\u8d25\uff08\u4f8b\u5982\u4efb\u52a1\u7406\u89e3\uff09\u53ef\u80fd\u4f1a\u635f\u5bb3\u6574\u4e2a\u8fc7\u7a0b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>Meta Data \u53d1\u8868\u65f6\u95f4 2025.09.29 \u4f5c\u8005\uff1aPengfei Wang, Baolin Sun1, Xuemei Do &#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":[60,41],"tags":[74,73,75],"class_list":["post-1501","post","type-post","status-publish","format-standard","hentry","category-paper_read","category-learning_note","tag-chatbi","tag-text-to-sql","tag-75"],"_links":{"self":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1501","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=1501"}],"version-history":[{"count":7,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1501\/revisions"}],"predecessor-version":[{"id":1515,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/posts\/1501\/revisions\/1515"}],"wp:attachment":[{"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/media?parent=1501"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/categories?post=1501"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/lingbo.online\/index.php\/wp-json\/wp\/v2\/tags?post=1501"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}