{"id":5374,"date":"2026-07-15T22:36:29","date_gmt":"2026-07-15T14:36:29","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=5374"},"modified":"2026-07-17T22:31:20","modified_gmt":"2026-07-17T14:31:20","slug":"202607-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-rimrule-improving-tool-using-language-agents-via-mdl-guided-rule-learning","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202607-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-rimrule-improving-tool-using-language-agents-via-mdl-guided-rule-learning\/","title":{"rendered":"202607 \u8bba\u6587\u7814\u8bfb-RIMRULE: Improving Tool-Using Language Agents via MDL-Guided Rule Learning"},"content":{"rendered":"\n<p>\u4f5c\u8005\uff1aXiang Gao, Yuguang Yao et.c<\/p>\n\n\n\n<p>\u5355\u4f4d\uff1aIntuit AI Research\uff1bTemple University<\/p>\n\n\n\n<p>\u6765\u6e90\uff1aACL 2026<\/p>\n\n\n\n<p>\u65f6\u95f4\uff1a2026.07<\/p>\n\n\n\n<p>\u94fe\u63a5\uff1a<u><a href=\"https:\/\/www.cell.com\/patterns\/fulltext\/S2666-3899(25)00082-0\" target=\"_blank\" rel=\"noreferrer noopener\">[<\/a><\/u><a href=\"https:\/\/arxiv.org\/pdf\/2509.21629\">RIMRULE: Improving Tool-Using Language Agents via MDL-Guided Rule Learning<\/a><u><a href=\"https:\/\/www.cell.com\/patterns\/fulltext\/S2666-3899(25)00082-0\" target=\"_blank\" rel=\"noreferrer noopener\">]<\/a><\/u><\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"%E4%B8%80.-%E7%A0%94%E7%A9%B6%E8%83%8C%E6%99%AF\">\u4e00. \u7814\u7a76\u80cc\u666f<\/h3>\n\n\n\n<p>\u5927\u8bed\u8a00\u6a21\u578b\u9010\u6e10\u4ece\u6587\u672c\u751f\u6210\u5668\u53d1\u5c55\u4e3a\u80fd\u591f\u8c03\u7528\u641c\u7d22\u3001\u8ba1\u7b97\u3001\u6570\u636e\u5e93\u3001\u9884\u8ba2\u548c\u4f01\u4e1a API \u7684\u667a\u80fd\u4f53\uff0c\u5de5\u5177\u4f7f\u7528\u80fd\u529b\u6210\u4e3a\u8861\u91cf Agent \u5b9e\u7528\u6027\u7684\u91cd\u8981\u6307\u6807\u3002<\/p>\n\n\n\n<p>\u5728\u6807\u51c6\u5316\u5de5\u5177\u73af\u5883\u4e2d\uff0c\u6a21\u578b\u901a\u5e38\u80fd\u591f\u4f9d\u636e\u6e05\u6670\u7684\u5de5\u5177\u8bf4\u660e\u5b8c\u6210\u8c03\u7528\u3002\u4f46\u5728\u771f\u5b9e\u4e1a\u52a1\u573a\u666f\u4e2d\uff0cAPI \u5f80\u5f80\u5177\u6709\u7279\u6b8a\u7684\u547d\u540d\u65b9\u5f0f\u3001\u4e0d\u5b8c\u6574\u7684\u6587\u6863\u3001\u590d\u6742\u7684\u53c2\u6570\u7ea6\u675f\uff0c\u751a\u81f3\u5305\u542b\u4f01\u4e1a\u5185\u90e8\u7279\u6709\u7684\u6267\u884c\u6d41\u7a0b\u3002\u6a21\u578b\u5373\u4f7f\u7406\u89e3\u4e86\u7528\u6237\u610f\u56fe\uff0c\u4e5f\u53ef\u80fd\u5728\u4efb\u52a1\u5206\u89e3\u3001\u5de5\u5177\u9009\u62e9\u548c\u53c2\u6570\u586b\u5199\u73af\u8282\u51fa\u73b0\u7cfb\u7edf\u6027\u9519\u8bef\u3002<\/p>\n\n\n\n<p>\u73b0\u6709\u5de5\u5177\u9002\u5e94\u65b9\u6cd5\u4e3b\u8981\u5305\u62ec\u4e09\u7c7b\uff1a<\/p>\n\n\n\n<p>\u7b2c\u4e00\u7c7b\u662f Few-shot Prompting\uff0c\u5373\u68c0\u7d22\u76f8\u4f3c\u793a\u4f8b\u5e76\u653e\u5165\u4e0a\u4e0b\u6587\u3002\u5b83\u80fd\u591f\u590d\u7528\u5386\u53f2\u7ecf\u9a8c\uff0c\u4f46\u901a\u5e38\u4fdd\u7559\u7684\u662f\u5b8c\u6574\u6837\u672c\uff0c\u6a21\u578b\u4e0d\u4e00\u5b9a\u80fd\u4ece\u4e2d\u62bd\u8c61\u51fa\u53ef\u8fc1\u79fb\u7684\u89c4\u5f8b\u3002<\/p>\n\n\n\n<p>\u7b2c\u4e8c\u7c7b\u662f\u5168\u5c40 Prompt \u4f18\u5316\uff0c\u5373\u4e3a\u6240\u6709\u95ee\u9898\u751f\u6210\u4e00\u5957\u7edf\u4e00\u6307\u4ee4\u3002\u8fd9\u79cd\u65b9\u6cd5\u5b9e\u73b0\u7b80\u5355\uff0c\u4f46\u56fa\u5b9a Prompt \u96be\u4ee5\u8986\u76d6\u4e0d\u540c\u4efb\u52a1\u548c\u4e0d\u540c\u5de5\u5177\u73af\u5883\u3002<\/p>\n\n\n\n<p>\u7b2c\u4e09\u7c7b\u662f\u6a21\u578b\u5fae\u8c03\u3002\u5fae\u8c03\u80fd\u591f\u628a\u5de5\u5177\u77e5\u8bc6\u5199\u5165\u6a21\u578b\u53c2\u6570\uff0c\u4f46\u9700\u8981\u989d\u5916\u8bad\u7ec3\u6210\u672c\uff1b\u5f53\u5de5\u5177\u63a5\u53e3\u53d1\u751f\u53d8\u5316\u65f6\uff0c\u6a21\u578b\u53ef\u80fd\u9700\u8981\u91cd\u65b0\u8bad\u7ec3\uff0c\u800c\u4e14\u53c2\u6570\u4e2d\u7684\u77e5\u8bc6\u96be\u4ee5\u89e3\u91ca\u548c\u8fc1\u79fb\u3002<\/p>\n\n\n\n<p>RIMRULE \u5173\u6ce8\u7684\u95ee\u9898\u662f\uff1a\u80fd\u5426\u8ba9\u6a21\u578b\u50cf\u4eba\u7c7b\u4e00\u6837\uff0c\u4ece\u5931\u8d25\u4e2d\u603b\u7ed3\u51fa\u7b80\u6d01\u3001\u53ef\u590d\u7528\u7684\u7ecf\u9a8c\u89c4\u5219\uff0c\u5e76\u5728\u9047\u5230\u76f8\u4f3c\u95ee\u9898\u65f6\u52a8\u6001\u8c03\u7528\u8fd9\u4e9b\u89c4\u5219\uff1f<\/p>\n\n\n\n<p>\u672c\u6587\u8ba4\u4e3a\uff0c\u4eba\u7c7b\u4e0d\u4f1a\u5b8c\u6574\u8bb0\u4f4f\u6bcf\u4e00\u6b21\u5931\u8d25\uff0c\u800c\u4f1a\u628a\u5931\u8d25\u538b\u7f29\u4e3a\u7c7b\u4f3c\u9047\u5230\u8fd9\u79cd\u60c5\u51b5\u65f6\uff0c\u5e94\u5f53\u91c7\u53d6\u67d0\u79cd\u64cd\u4f5c\u7684\u542f\u53d1\u5f0f\u77e5\u8bc6\u3002RIMRULE \u56e0\u6b64\u5c1d\u8bd5\u5c06 Agent \u7684\u9002\u5e94\u8fc7\u7a0b\u4ece\u6a21\u578b\u53c2\u6570\u4e2d\u5206\u79bb\u51fa\u6765\uff0c\u6784\u5efa\u4e00\u4e2a\u53ef\u89e3\u91ca\u3001\u53ef\u7f16\u8f91\u548c\u53ef\u8fc1\u79fb\u7684\u5916\u90e8\u89c4\u5219\u5e93\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"%E4%BA%8C.-%E8%AE%BA%E6%96%87%E6%A6%82%E8%A6%81\">\u4e8c. \u8bba\u6587\u6982\u8981<\/h3>\n\n\n\n<p>\u8bba\u6587\u63d0\u51fa RIMRULE\uff0c\u5373 <strong>Reusable\u3001Interpretable and MDL-guided Rules<\/strong>\u3002\u5b83\u662f\u4e00\u79cd\u9762\u5411\u5de5\u5177\u8c03\u7528\u667a\u80fd\u4f53\u7684\u5931\u8d25\u9a71\u52a8\u89c4\u5219\u5b66\u4e60\u6846\u67b6\u3002\u672c\u6587\u7684\u6838\u5fc3\u8d21\u732e\u4e3b\u8981\u5305\u62ec\u56db\u70b9\uff1a<\/p>\n\n\n\n<p><strong>\u5931\u8d25\u9a71\u52a8\u89c4\u5219\u751f\u6210\uff1a<\/strong> \u7cfb\u7edf\u6bd4\u8f83 Agent \u7684\u9519\u8bef\u6267\u884c\u8f68\u8ff9\u4e0e\u6b63\u786e\u8f68\u8ff9\uff0c\u5b9a\u4f4d\u6700\u4e0a\u6e38\u7684\u6839\u56e0\u9519\u8bef\uff0c\u5e76\u751f\u6210\u539f\u5b50\u5316\u7684 if\u2013then \u89c4\u5219\u3002<\/p>\n\n\n\n<p><strong>\u81ea\u7136\u8bed\u8a00\u4e0e\u7b26\u53f7\u53cc\u91cd\u8868\u793a\uff1a<\/strong> \u6bcf\u6761\u89c4\u5219\u540c\u65f6\u4fdd\u5b58\u4e3a\u81ea\u7136\u8bed\u8a00\u5f62\u5f0f\u548c\u7ed3\u6784\u5316\u7b26\u53f7\u5f62\u5f0f\u3002\u81ea\u7136\u8bed\u8a00\u89c4\u5219\u7528\u4e8e\u6307\u5bfc LLM\uff0c\u7b26\u53f7\u89c4\u5219\u7528\u4e8e\u53bb\u91cd\u3001\u6574\u5408\u548c\u68c0\u7d22\u3002<\/p>\n\n\n\n<p><strong>MDL \u5f15\u5bfc\u7684\u89c4\u5219\u6574\u5408\uff1a<\/strong> \u7cfb\u7edf\u5229\u7528\u6700\u5c0f\u63cf\u8ff0\u957f\u5ea6\u601d\u60f3\uff0c\u5728\u89c4\u5219\u590d\u6742\u5ea6\u4e0e\u5931\u8d25\u7ea0\u6b63\u80fd\u529b\u4e4b\u95f4\u8fdb\u884c\u6743\u8861\uff0c\u901a\u8fc7\u5220\u9664\u5197\u4f59\u89c4\u5219\u548c\u6cdb\u5316\u5177\u4f53\u89c4\u5219\uff0c\u5f62\u6210\u66f4\u7d27\u51d1\u7684\u89c4\u5219\u5e93\u3002<\/p>\n\n\n\n<p><strong>\u52a8\u6001\u89c4\u5219\u68c0\u7d22\uff1a<\/strong> \u5728\u5904\u7406\u65b0\u4efb\u52a1\u65f6\uff0c\u7cfb\u7edf\u4e0d\u4f1a\u6ce8\u5165\u5168\u90e8\u89c4\u5219\uff0c\u800c\u662f\u6839\u636e\u5f53\u524d\u95ee\u9898\u548c\u53ef\u7528\u5de5\u5177\u68c0\u7d22\u76f8\u5173\u89c4\u5219\uff0c\u518d\u5c06\u5176\u52a0\u5165 Agent Prompt\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u5728 ToolHop \u548c BFCL \u4e24\u4e2a\u5de5\u5177\u8c03\u7528\u57fa\u51c6\u4e0a\u8fdb\u884c\u5b9e\u9a8c\u3002\u7ed3\u679c\u8868\u660e\uff0cRIMRULE \u80fd\u591f\u63d0\u9ad8\u5df2\u89c1\u5de5\u5177\u548c\u672a\u89c1\u5de5\u5177\u4e0a\u7684\u8c03\u7528\u51c6\u786e\u7387\uff0c\u5e76\u4e14\u53ef\u4ee5\u4e0e\u76d1\u7763\u5fae\u8c03\u548c\u539f\u751f Function Calling \u6a21\u578b\u7ed3\u5408\u3002\u90e8\u5206\u89c4\u5219\u8fd8\u80fd\u591f\u5728\u4e0d\u540c\u6a21\u578b\u4e4b\u95f4\u590d\u7528\u3002<\/p>\n\n\n\n<p>\u672c\u6587\u4e0e\u4f20\u7edf\u65b9\u6cd5\u6700\u5927\u7684\u4e0d\u540c\u5728\u4e8e\uff0c\u5b83\u5e76\u4e0d\u4fdd\u5b58\u5b8c\u6574\u7684\u6210\u529f\u793a\u4f8b\uff0c\u4e5f\u4e0d\u76f4\u63a5\u4fee\u6539\u6a21\u578b\u53c2\u6570\uff0c\u800c\u662f\u5c06\u6a21\u578b\u4ece\u5931\u8d25\u4e2d\u83b7\u5f97\u7684\u77e5\u8bc6\u8f6c\u5316\u4e3a\u4e00\u4e2a\u72ec\u7acb\u7684\u89c4\u5219\u5c42\uff1a \u57fa\u7840\u5927\u6a21\u578b+\u9886\u57df\u89c4\u5219\u5e93+\u52a8\u6001\u89c4\u5219\u68c0\u7d22\u3002<\/p>\n\n\n\n<p>\u8fd9\u79cd\u8bbe\u8ba1\u4f7f Agent \u7684\u9002\u5e94\u77e5\u8bc6\u66f4\u52a0\u900f\u660e\uff0c\u4e5f\u65b9\u4fbf\u5728\u5de5\u5177\u53d8\u5316\u540e\u5355\u72ec\u66f4\u65b0\u89c4\u5219\uff0c\u800c\u4e0d\u5fc5\u91cd\u65b0\u8bad\u7ec3\u6574\u4e2a\u6a21\u578b\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"core-idea\">\u4e09\u3001\u6838\u5fc3\u601d\u60f3<\/h2>\n\n\n\n<p>\u672c\u6587\u7684\u6838\u5fc3\u601d\u60f3\u53ef\u4ee5\u6982\u62ec\u4e3a\uff1aLLM \u662f\u5931\u8d25\u5206\u6790\u5668\u4e0e\u5019\u9009\u89c4\u5219\u751f\u6210\u5668\uff1b\u6267\u884c\u53cd\u9988\u662f\u89c4\u5219\u8bc4\u4ef7\u5668\uff1b\u89c4\u5219\u5e93\u662f\u6a21\u578b\u4e4b\u5916\u7684\u9002\u5e94\u8bb0\u5fc6\u3002<\/p>\n\n\n\n<p>\u4f20\u7edf Few-shot \u65b9\u6cd5\u901a\u5e38\u4f1a\u68c0\u7d22\u4e00\u4e2a\u4e0e\u5f53\u524d\u95ee\u9898\u76f8\u4f3c\u7684\u5386\u53f2\u6848\u4f8b\uff0c\u7136\u540e\u8ba9\u6a21\u578b\u6a21\u4eff\u8be5\u6848\u4f8b\u3002\u4f46\u4e24\u4e2a\u95ee\u9898\u5728\u8868\u9762\u8bed\u8a00\u4e0a\u53ef\u80fd\u4e0d\u540c\uff0c\u5374\u5171\u4eab\u76f8\u540c\u7684\u9519\u8bef\u7ed3\u6784\u3002<\/p>\n\n\n\n<p>\u4f8b\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u67e5\u8be2\u67d0\u4eba\u7684\u5916\u7956\u7236\uff1b<\/li>\n\n\n\n<li>\u67e5\u8be2\u67d0\u4eba\u7684\u7ee7\u5144\u5f1f\uff1b<\/li>\n\n\n\n<li>\u67e5\u8be2\u67d0\u4f4d\u4eba\u7269\u7684\u7956\u6bcd\u51fa\u751f\u5730\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e9b\u95ee\u9898\u7684\u5177\u4f53\u4eba\u7269\u548c\u76ee\u6807\u5c5e\u6027\u4e0d\u540c\uff0c\u4f46\u53ef\u80fd\u5171\u4eab\u540c\u4e00\u6761\u89c4\u5219\uff1a\u5f53\u5de5\u5177\u4e0d\u652f\u6301\u590d\u6742\u5173\u7cfb\u76f4\u63a5\u67e5\u8be2\u65f6\uff0c\u5e94\u5148\u9010\u6b65\u89e3\u6790\u4e2d\u95f4\u5b9e\u4f53\uff0c\u518d\u67e5\u8be2\u76ee\u6807\u5c5e\u6027\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0cRIMRULE \u4e0d\u5f3a\u8c03\u68c0\u7d22\u76f8\u4f3c\u95ee\u9898\uff0c\u800c\u66f4\u5f3a\u8c03\u68c0\u7d22\u9002\u7528\u4e8e\u5f53\u524d\u63a8\u7406\u7ed3\u6784\u7684\u89c4\u5219\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u8fd8\u63d0\u51fa\u4e86\u4e00\u4e2a\u91cd\u8981\u7684\u53cc\u91cd\u8868\u793a\u601d\u60f3\u3002<\/p>\n\n\n\n<p>\u81ea\u7136\u8bed\u8a00\u89c4\u5219\u4fbf\u4e8e\u5927\u6a21\u578b\u7406\u89e3\uff0c\u4f8b\u5982\uff1a\u5982\u679c\u95ee\u9898\u5305\u542b\u591a\u5c42\u4eb2\u5c5e\u5173\u7cfb\uff0c\u5e94\u9010\u7ea7\u67e5\u8be2\u4e2d\u95f4\u4eba\u7269\uff0c\u5e76\u5c06\u524d\u4e00\u6b65\u7ed3\u679c\u4f5c\u4e3a\u4e0b\u4e00\u6b65\u5de5\u5177\u8f93\u5165\u3002<\/p>\n\n\n\n<p>\u7b26\u53f7\u89c4\u5219\u5219\u4fbf\u4e8e\u7cfb\u7edf\u7ba1\u7406\uff1a<\/p>\n\n\n\n<p>domain = RELATIONSHIP_RESOLUTION<br>qualifier = RELATIONSHIP_CHAIN_TRAVERSAL<br>action = [DECOMPOSE_QUERY, RESOLVE_INTERMEDIATE_ENTITY]<br>strength = MANDATORY<br>tool_category = GENEALOGY_QUERY<\/p>\n\n\n\n<p>\u81ea\u7136\u8bed\u8a00\u8d1f\u8d23\u6267\u884c\uff0c\u7b26\u53f7\u7ed3\u6784\u8d1f\u8d23\u68c0\u7d22\u548c\u538b\u7f29\u3002\u8bba\u6587\u7b2c 3 \u9875\u7684\u56fe 2 \u5c06\u6574\u4e2a\u8fc7\u7a0b\u5212\u5206\u4e3a\u89c4\u5219\u751f\u6210\u3001\u89c4\u5219\u6574\u5408\u548c\u63a8\u7406\u65f6\u89c4\u5219\u6ce8\u5165\u4e09\u4e2a\u9636\u6bb5\u3002<\/p>\n\n\n\n<p>\u8fd9\u79cd\u8bbe\u8ba1\u5bf9\u98df\u54c1\u914d\u6599\u63a8\u7406\u4e5f\u5177\u6709\u542f\u53d1\u6027\uff1a\u7cfb\u7edf\u4e0d\u5e94\u53ea\u8bb0\u4f4f\u67d0\u4e2a\u54c1\u724c\u9178\u5976\u7684\u5177\u4f53\u4f30\u7b97\u7ed3\u679c\uff0c\u800c\u5e94\u8fdb\u4e00\u6b65\u603b\u7ed3\u201c\u4e73\u5236\u54c1\u4e2d\u86cb\u767d\u8d28\u8d21\u732e\u4e0e\u4e73\u6e90\u914d\u6599\u6bd4\u4f8b\u4e4b\u95f4\u7684\u5173\u7cfb\u201d\u7b49\u53ef\u8fc1\u79fb\u89c4\u5219\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"method-design\">\u56db\u3001\u65b9\u6cd5\u8bbe\u8ba1<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"failure-rule-generation\">4.1 \u5931\u8d25\u7ecf\u9a8c\u6536\u96c6\u4e0e\u6839\u56e0\u89c4\u5219\u751f\u6210<\/h3>\n\n\n\n<p>RIMRULE \u9996\u5148\u4f7f\u7528\u96f6\u6837\u672c\u5de5\u5177 Agent \u8fd0\u884c\u8bad\u7ec3\u96c6\uff0c\u5e76\u4fdd\u5b58\u6267\u884c\u5931\u8d25\u7684\u6837\u672c\u3002<\/p>\n\n\n\n<p>\u6bcf\u4e2a\u5931\u8d25\u7ecf\u9a8c\u8868\u793a\u4e3a\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>(x,S,\u03c4\u207b,\u03c4*)<\/code><\/p>\n\n\n\n<p>\u5176\u4e2d\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>x<\/code> \u8868\u793a\u7528\u6237\u95ee\u9898\uff1b<\/li>\n\n\n\n<li><code>S<\/code> \u8868\u793a\u5f53\u524d\u53ef\u7528\u5de5\u5177\u96c6\u5408\uff1b<\/li>\n\n\n\n<li><code>\u03c4\u207b<\/code> \u8868\u793a\u6a21\u578b\u751f\u6210\u7684\u9519\u8bef\u6267\u884c\u8f68\u8ff9\uff1b<\/li>\n\n\n\n<li><code>\u03c4*<\/code> \u8868\u793a\u6b63\u786e\u6267\u884c\u8f68\u8ff9\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u7cfb\u7edf\u8ba9\u53e6\u4e00\u4e2a LLM \u6bd4\u8f83\u9519\u8bef\u8f68\u8ff9\u4e0e\u6b63\u786e\u8f68\u8ff9\uff0c\u5e76\u5224\u65ad\u9519\u8bef\u5c5e\u4e8e\u4ee5\u4e0b\u54ea\u4e00\u7c7b\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u4efb\u52a1\u5206\u89e3\u9519\u8bef\uff1b<\/li>\n\n\n\n<li>\u5de5\u5177\u9009\u62e9\u9519\u8bef\uff1b<\/li>\n\n\n\n<li>\u5de5\u5177\u53c2\u6570\u6784\u9020\u9519\u8bef\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u89c4\u5219\u751f\u6210\u8fc7\u7a0b\u91c7\u7528 Explanation-Based Learning \u7684\u601d\u60f3\u3002\u6a21\u578b\u9700\u8981\u533a\u5206\u6839\u56e0\u9519\u8bef\u548c\u4f20\u64ad\u9519\u8bef\u3002<\/p>\n\n\n\n<p>\u4f8b\u5982\uff0cAgent \u6ca1\u6709\u5148\u8bc6\u522b\u4e2d\u95f4\u4eba\u7269\uff0c\u5bfc\u81f4\u540e\u7eed\u5de5\u5177\u67e5\u8be2\u4e0d\u5230\u6570\u636e\u3002\u771f\u6b63\u9700\u8981\u603b\u7ed3\u7684\u89c4\u5219\u662f\uff1a\u590d\u6742\u5173\u7cfb\u9700\u8981\u9010\u6b65\u5206\u89e3\u3002<\/p>\n\n\n\n<p>\u751f\u6210\u89c4\u5219\u540e\uff0c\u7cfb\u7edf\u8fd8\u4f1a\u8fdb\u884c\u4e24\u9879\u68c0\u67e5\uff1a<\/p>\n\n\n\n<p>\u7b2c\u4e00\uff0c\u5c06\u89c4\u5219\u91cd\u65b0\u6ce8\u5165\u539f\u4efb\u52a1\uff0c\u786e\u8ba4\u5b83\u786e\u5b9e\u80fd\u591f\u6539\u5584\u6267\u884c\u7ed3\u679c\u3002<\/p>\n\n\n\n<p>\u7b2c\u4e8c\uff0c\u68c0\u67e5\u89c4\u5219\u662f\u5426\u7b80\u6d01\u3001\u539f\u5b50\u5316\uff0c\u5e76\u7b26\u5408 if\u2013then \u5f62\u5f0f\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u4e00\u6761\u89c4\u5219\u53ea\u80fd\u4fee\u590d\u90e8\u5206\u9519\u8bef\uff0c\u7cfb\u7edf\u53ef\u4ee5\u518d\u6b21\u8c03\u7528\u89c4\u5219\u751f\u6210\u5668\uff0c\u7ee7\u7eed\u5bfb\u627e\u5176\u4ed6\u6839\u56e0\u89c4\u5219\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"dual-representation\">4.2 \u81ea\u7136\u8bed\u8a00\u4e0e\u7b26\u53f7\u53cc\u91cd\u8868\u793a<\/h3>\n\n\n\n<p>\u81ea\u7136\u8bed\u8a00\u89c4\u5219\u53ef\u80fd\u5b58\u5728\u5927\u91cf\u8868\u8fbe\u5dee\u5f02\u3002\u4f8b\u5982\uff0c\u201c\u9010\u6b65\u89e3\u6790\u4e2d\u95f4\u5b9e\u4f53\u201d\u201c\u5206\u89e3\u590d\u6742\u5173\u7cfb\u94fe\u201d\u548c\u201c\u4f9d\u6b21\u67e5\u8be2\u6bcf\u5c42\u5173\u7cfb\u201d\u53ef\u80fd\u8868\u793a\u76f8\u540c\u7b56\u7565\u3002<\/p>\n\n\n\n<p>\u4e3a\u4e86\u964d\u4f4e\u8fd9\u79cd\u8bed\u8a00\u5dee\u5f02\uff0cRIMRULE \u5c06\u89c4\u5219\u8f6c\u6362\u4e3a\u4e94\u7c7b\u7b26\u53f7\u5b57\u6bb5\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>Domain, Qualifier, Action, Strength, ToolCategory<\/code><\/p>\n\n\n\n<p>\u5176\u4e2d\uff0cDomain \u8868\u793a\u89c4\u5219\u6240\u5c5e\u7684\u5bbd\u6cdb\u9886\u57df\uff0cQualifier \u63cf\u8ff0\u5177\u4f53\u9002\u7528\u60c5\u5883\uff0cAction \u8868\u793a\u5e94\u6267\u884c\u7684\u64cd\u4f5c\uff0cStrength \u8868\u793a\u89c4\u5219\u4f18\u5148\u7ea7\uff0cToolCategory \u8868\u793a\u9002\u7528\u7684\u5de5\u5177\u7c7b\u522b\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u9996\u5148\u8ba9 LLM \u4ece\u5168\u90e8\u81ea\u7136\u8bed\u8a00\u89c4\u5219\u4e2d\u5f52\u7eb3\u5019\u9009\u7b26\u53f7\u8bcd\u8868\uff0c\u518d\u4f7f\u7528\u56fa\u5b9a\u8bcd\u8868\u5bf9\u6bcf\u6761\u89c4\u5219\u8fdb\u884c\u5206\u7c7b\u3002<\/p>\n\n\n\n<p>\u6700\u7ec8\uff0c\u6bcf\u6761\u89c4\u5219\u4fdd\u5b58\u4e3a\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>r=(r^(NL),r^(SYM))<\/code><\/p>\n\n\n\n<p>\u5176\u4e2d\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>r^(NL)<\/code> \u662f\u6ce8\u5165 Agent Prompt \u7684\u81ea\u7136\u8bed\u8a00\u7248\u672c\uff1b<\/li>\n\n\n\n<li><code>r^(SYM)<\/code> \u662f\u7528\u4e8e\u68c0\u7d22\u3001\u6bd4\u8f83\u548c\u89c4\u5219\u6574\u5408\u7684\u7b26\u53f7\u7248\u672c\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u79cd\u53cc\u91cd\u8868\u793a\u907f\u514d\u4e86\u76f4\u63a5\u7528\u81ea\u7136\u8bed\u8a00\u957f\u5ea6\u8861\u91cf\u89c4\u5219\u590d\u6742\u5ea6\uff0c\u4e5f\u4f7f\u8bed\u4e49\u76f8\u540c\u4f46\u8868\u8fbe\u4e0d\u540c\u7684\u89c4\u5219\u66f4\u5bb9\u6613\u88ab\u8bc6\u522b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"mdl-consolidation\">4.3 MDL \u5f15\u5bfc\u7684\u89c4\u5219\u6574\u5408<\/h3>\n\n\n\n<p>\u5c40\u90e8\u89c4\u5219\u751f\u6210\u4f1a\u4ea7\u751f\u5927\u91cf\u91cd\u590d\u3001\u76f8\u4f3c\u6216\u8fc7\u5ea6\u5177\u4f53\u7684\u89c4\u5219\u3002\u5982\u679c\u5c06\u6240\u6709\u89c4\u5219\u76f4\u63a5\u6ce8\u5165 Prompt\uff0c\u4e0d\u4ec5\u4f1a\u589e\u52a0 Token \u6210\u672c\uff0c\u8fd8\u53ef\u80fd\u5bfc\u81f4\u89c4\u5219\u76f8\u4e92\u5e72\u6270\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0cRIMRULE \u4f7f\u7528\u6700\u5c0f\u63cf\u8ff0\u957f\u5ea6\u601d\u60f3\u6574\u5408\u89c4\u5219\u5e93\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u5b9a\u4e49\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>MDL(H)=L(H)+L(D | H)<\/code><\/p>\n\n\n\n<p>\u5176\u4e2d\uff0c<code>H<\/code> \u8868\u793a\u5f53\u524d\u89c4\u5219\u5e93\u3002<\/p>\n\n\n\n<p>\u89c4\u5219\u590d\u6742\u5ea6\u4e3a\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>L(H)=\u03b1\u03a3_(r\u2208 H)\u2113(r)<\/code><\/p>\n\n\n\n<p>\u5176\u4e2d\uff0c<code>\u2113(r)<\/code> \u8868\u793a\u89c4\u5219\u7684\u7b26\u53f7\u957f\u5ea6\uff0c<code>\u03b1<\/code> \u63a7\u5236\u5bf9\u89c4\u5219\u590d\u6742\u5ea6\u7684\u60e9\u7f5a\u3002<\/p>\n\n\n\n<p>\u89c4\u5219\u5bf9\u5931\u8d25\u6570\u636e\u7684\u89e3\u91ca\u4ee3\u4ef7\u4e3a\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>L(D | H)=-[ k_H log p\u0302_H+ (n-k_H) log (1-p\u0302_H) ]<\/code><\/p>\n\n\n\n<p>\u5176\u4e2d\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>p\u0302_H=(k_H)\/(n)<\/code><\/p>\n\n\n\n<p><code>k_H<\/code> \u8868\u793a\u89c4\u5219\u5e93\u80fd\u591f\u7ea0\u6b63\u7684\u5931\u8d25\u6570\u91cf\u3002<\/p>\n\n\n\n<p>\u7cfb\u7edf\u4ece\u5b8c\u6574\u5019\u9009\u89c4\u5219\u5e93\u5f00\u59cb\uff0c\u53cd\u590d\u6267\u884c\u4e24\u79cd\u64cd\u4f5c\uff1a<\/p>\n\n\n\n<p><strong>PRUNE\uff1a<\/strong> \u5220\u9664\u67d0\u6761\u89c4\u5219\u3002<\/p>\n\n\n\n<p><strong>GENERALIZE\uff1a<\/strong> \u5c06\u53ea\u9002\u7528\u4e8e\u67d0\u4e2a\u5177\u4f53\u5de5\u5177\u7684\u89c4\u5219\uff0c\u6cdb\u5316\u4e3a\u9002\u7528\u4e8e\u67d0\u7c7b\u5de5\u5177\u7684\u89c4\u5219\u3002<\/p>\n\n\n\n<p>\u53ea\u6709\u5f53\u64cd\u4f5c\u80fd\u591f\u964d\u4f4e MDL \u76ee\u6807\u65f6\uff0c\u7cfb\u7edf\u624d\u63a5\u53d7\u8be5\u4fee\u6539\u3002\u6574\u4e2a\u8fc7\u7a0b\u6301\u7eed\u5230\u6ca1\u6709\u5c40\u90e8\u7f16\u8f91\u80fd\u591f\u8fdb\u4e00\u6b65\u964d\u4f4e\u76ee\u6807\u4e3a\u6b62\u3002<\/p>\n\n\n\n<p>\u4ece\u5de5\u7a0b\u89d2\u5ea6\u770b\uff0c\u8be5\u8fc7\u7a0b\u53ef\u4ee5\u7406\u89e3\u4e3a\u4e00\u79cd\u201c\u4fdd\u7559\u6709\u6548\u89c4\u5219\u3001\u5220\u9664\u5197\u4f59\u89c4\u5219\u201d\u7684\u884c\u4e3a\u9a71\u52a8\u526a\u679d\u673a\u5236\u3002<\/p>\n\n\n\n<p>\u4e0d\u8fc7\uff0c\u5f53\u524d MDL \u6570\u636e\u9879\u5b58\u5728\u4e00\u4e2a\u9700\u8981\u8c28\u614e\u7406\u89e3\u7684\u95ee\u9898\u3002\u5b83\u672c\u8d28\u4e0a\u662f\u4e8c\u5143\u71b5\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>L(D | H)=n h(k_H\/n)<\/code><\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u5f53\u89c4\u5219\u7ea0\u6b63\u5168\u90e8\u5931\u8d25\u6216\u5b8c\u5168\u4e0d\u7ea0\u6b63\u5931\u8d25\u65f6\uff0c\u6570\u636e\u9879\u90fd\u53ef\u80fd\u5f88\u4f4e\u3002\u5f53\u524d\u65b9\u6cd5\u4e4b\u6240\u4ee5\u6ca1\u6709\u9000\u5316\u4e3a\u7a7a\u89c4\u5219\u5e93\uff0c\u4e3b\u8981\u4f9d\u8d56\u4e8e\u4ece\u9ad8\u7ea0\u6b63\u7387\u89c4\u5219\u5e93\u51fa\u53d1\uff0c\u5e76\u91c7\u7528\u53ea\u63a5\u53d7\u5c40\u90e8\u6539\u8fdb\u7684\u8d2a\u5fc3\u641c\u7d22\u3002<\/p>\n\n\n\n<p>\u6240\u4ee5\u66f4\u51c6\u786e\u7684\u8868\u8ff0\u5e94\u662f\uff1aRIMRULE \u4f7f\u7528 MDL \u5f15\u5bfc\u7684\u5c40\u90e8\u89c4\u5219\u526a\u679d\u4e0e\u6cdb\u5316\uff0c\u800c\u4e0d\u662f\u4ece\u7406\u8bba\u4e0a\u4fdd\u8bc1\u83b7\u5f97\u5168\u5c40\u6700\u4f18\u89c4\u5219\u5e93\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"dynamic-retrieval\">4.4 \u63a8\u7406\u9636\u6bb5\u7684\u52a8\u6001\u89c4\u5219\u68c0\u7d22<\/h3>\n\n\n\n<p>\u9762\u5bf9\u65b0\u4efb\u52a1\u65f6\uff0c\u7cfb\u7edf\u4e0d\u4f1a\u5c06\u6240\u6709\u89c4\u5219\u653e\u5165 Prompt\uff0c\u800c\u662f\u6839\u636e\u5f53\u524d\u95ee\u9898\u548c\u53ef\u7528\u5de5\u5177\u52a8\u6001\u68c0\u7d22\u89c4\u5219\u3002<\/p>\n\n\n\n<p>\u9996\u5148\uff0c\u7cfb\u7edf\u5c06\u7528\u6237\u95ee\u9898\u548c\u5de5\u5177\u73af\u5883\u8f6c\u6362\u4e3a\u7b26\u53f7\u72b6\u6001\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>z_q=Symbolize(x_q,S_q)<\/code><\/p>\n\n\n\n<p>\u968f\u540e\u6267\u884c\u4e24\u7ea7\u68c0\u7d22\u3002<\/p>\n\n\n\n<p>\u7b2c\u4e00\u7ea7\u662f\u9002\u7528\u6027\u8fc7\u6ee4\u3002\u4efb\u52a1\u5206\u89e3\u89c4\u5219\u901a\u5e38\u4fdd\u7559\uff0c\u800c\u5de5\u5177\u9009\u62e9\u548c\u53c2\u6570\u89c4\u5219\u9700\u8981\u68c0\u67e5\u5f53\u524d\u5de5\u5177\u96c6\u5408\u4e2d\u662f\u5426\u5b58\u5728\u5bf9\u5e94\u5de5\u5177\u6216\u5de5\u5177\u7c7b\u522b\u3002<\/p>\n\n\n\n<p>\u7b2c\u4e8c\u7ea7\u662f\u7b26\u53f7\u76f8\u4f3c\u5ea6\u6392\u5e8f\u3002\u7cfb\u7edf\u6bd4\u8f83\u67e5\u8be2\u72b6\u6001\u4e0e\u89c4\u5219\u5728 Domain\u3001Qualifier\u3001Action \u548c ToolCategory \u7b49\u5b57\u6bb5\u4e0a\u7684\u8bed\u4e49\u76f8\u5173\u6027\uff0c\u9009\u62e9 Top-<code>k<\/code> \u89c4\u5219\u3002<\/p>\n\n\n\n<p>\u6700\u7ec8\u6ce8\u5165 Agent \u7684\u4ecd\u7136\u662f\u89c4\u5219\u7684\u81ea\u7136\u8bed\u8a00\u7248\u672c\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>r^(SYM) \u2014\u68c0\u7d22\u2192 r^(NL) \u2014Prompt \u6ce8\u5165\u2192 Tool Agent<\/code><\/p>\n\n\n\n<p>\u8fd9\u4e00\u8bbe\u8ba1\u907f\u514d\u4e86\u89c4\u5219\u5e93\u8fc7\u5927\u5bfc\u81f4\u7684\u4e0a\u4e0b\u6587\u566a\u58f0\uff0c\u4e5f\u4f7f\u7cfb\u7edf\u80fd\u591f\u9488\u5bf9\u4e0d\u540c\u4efb\u52a1\u7ec4\u5408\u4e0d\u540c\u89c4\u5219\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"experiments\">\u4e94\u3001\u5b9e\u9a8c\u8bbe\u8ba1<\/h2>\n\n\n\n<p>\u8bba\u6587\u4f7f\u7528 ToolHop \u548c BFCL \u4e24\u4e2a\u5de5\u5177\u8c03\u7528\u57fa\u51c6\u3002<\/p>\n\n\n\n<p>ToolHop \u4e3b\u8981\u5305\u542b\u9700\u8981\u591a\u6b65\u63a8\u7406\u548c\u7ec4\u5408\u5de5\u5177\u8c03\u7528\u7684\u95ee\u9898\u3002BFCL Live-Multiple \u66f4\u5173\u6ce8\u4ece\u5927\u91cf\u76f8\u4f3c\u5de5\u5177\u4e2d\u9009\u62e9\u6b63\u786e\u5de5\u5177\uff1bBFCL Multi-Turn-Base \u5219\u7528\u4e8e\u8bc4\u4f30\u4f4e\u8d44\u6e90\u3001\u591a\u8f6e\u5de5\u5177\u4efb\u52a1\u3002<\/p>\n\n\n\n<p>\u6570\u636e\u5212\u5206\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table><thead><tr><th>\u6570\u636e\u96c6<\/th><th>\u8bad\u7ec3\u96c6<\/th><th>Test-rand<\/th><th>Test-unseen<\/th><\/tr><\/thead><tbody><tr><td>ToolHop<\/td><td>392<\/td><td>70<\/td><td>51<\/td><\/tr><tr><td>BFCL Live-Multiple<\/td><td>735<\/td><td>175<\/td><td>143<\/td><\/tr><tr><td>BFCL Multi-Turn-Base<\/td><td>90<\/td><td>60<\/td><td>50<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u5176\u4e2d\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Test-rand \u662f\u968f\u673a\u540c\u5206\u5e03\u6d4b\u8bd5\u96c6\uff1b<\/li>\n\n\n\n<li>Test-unseen \u5305\u542b\u8bad\u7ec3\u9636\u6bb5\u6ca1\u6709\u51fa\u73b0\u8fc7\u7684\u5de5\u5177\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8bba\u6587\u4f7f\u7528\u7684\u4e3b\u8981\u57fa\u7ebf\u5305\u62ec\u96f6\u6837\u672c ReAct\u3001Few-shot \u68c0\u7d22\u3001SEE Prompt \u4f18\u5316\u3001LoRA \u76d1\u7763\u5fae\u8c03\u4ee5\u53ca GPT-4o Function Calling \u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u8bc4\u4ef7\u6307\u6807\u4e3b\u8981\u662f\u5de5\u5177\u4efb\u52a1\u51c6\u786e\u7387\u548c\u6700\u7ec8\u89c4\u5219\u6570\u91cf\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"learning-results\">5.1 \u89c4\u5219\u751f\u6210\u4e0e\u89c4\u5219\u6574\u5408\u6548\u679c<\/h3>\n\n\n\n<p>\u5728 Llama 3.2 \u4e0a\uff0cRIMRULE \u7684\u4e3b\u8981\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table><thead><tr><th>\u6570\u636e\u96c6<\/th><th>\u539f\u59cb\u6a21\u578b<\/th><th>\u89c4\u5219\u751f\u6210\u540e<\/th><th>\u89c4\u5219\u6574\u5408\u540e<\/th><\/tr><\/thead><tbody><tr><td>ToolHop Test-rand<\/td><td>26.5%<\/td><td>27.6%<\/td><td>31.1%<\/td><\/tr><tr><td>ToolHop Test-unseen<\/td><td>35.1%<\/td><td>42.1%<\/td><td>43.1%<\/td><\/tr><tr><td>BFCL Test-rand<\/td><td>50.1%<\/td><td>54.8%<\/td><td>56.6%<\/td><\/tr><tr><td>BFCL Test-unseen<\/td><td>45.0%<\/td><td>47.6%<\/td><td>48.5%<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>ToolHop \u7684\u89c4\u5219\u6570\u91cf\u4ece 72 \u6761\u51cf\u5c11\u5230 67 \u6761\uff0cBFCL \u7684\u89c4\u5219\u6570\u91cf\u4ece 151 \u6761\u51cf\u5c11\u5230 121 \u6761\u3002<\/p>\n\n\n\n<p>\u4ece\u7ed3\u679c\u53ef\u4ee5\u770b\u51fa\uff0c\u521d\u59cb\u89c4\u5219\u751f\u6210\u5bf9\u672a\u89c1\u5de5\u5177\u5177\u6709\u8f83\u660e\u663e\u7684\u63d0\u5347\u3002\u4f8b\u5982 ToolHop Test-unseen \u4ece 35.1% \u63d0\u9ad8\u5230 42.1%\uff0c\u8bf4\u660e\u89c4\u5219\u80fd\u591f\u6355\u6349\u4e00\u4e9b\u8de8\u5de5\u5177\u590d\u7528\u7684\u4efb\u52a1\u5206\u89e3\u6a21\u5f0f\u3002<\/p>\n\n\n\n<p>\u89c4\u5219\u6574\u5408\u540e\uff0c\u89c4\u5219\u6570\u91cf\u51cf\u5c11\uff0c\u4f46\u51c6\u786e\u7387\u7ee7\u7eed\u63d0\u5347\u3002\u8fd9\u8bf4\u660e\u5220\u9664\u91cd\u590d\u6216\u5e72\u6270\u6027\u89c4\u5219\u80fd\u591f\u6539\u5584\u63a8\u7406\u65f6\u7684\u89c4\u5219\u68c0\u7d22\u8d28\u91cf\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"cross-model-transfer\">5.2 \u8de8\u6a21\u578b\u89c4\u5219\u8fc1\u79fb<\/h3>\n\n\n\n<p>\u8bba\u6587\u5206\u522b\u4ece Llama 3.2 \u548c GPT-4o \u7684\u5931\u8d25\u8f68\u8ff9\u4e2d\u5b66\u4e60\u89c4\u5219\uff0c\u518d\u5c06\u89c4\u5219\u5e94\u7528\u5230\u5176\u4ed6\u6a21\u578b\u3002<\/p>\n\n\n\n<p>\u4f8b\u5982\uff0c\u4f7f\u7528 Llama 3.2 \u5b66\u5230\u7684\u89c4\u5219\u540e\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>Llama 4 \u5728 ToolHop \u4e0a\u4ece 73.8% \u63d0\u5347\u5230 76.7%\uff1b<\/li>\n\n\n\n<li>O1 \u4ece 53.2% \u63d0\u5347\u5230 57.4%\uff1b<\/li>\n\n\n\n<li>GPT-4o \u5728 BFCL \u4e0a\u4ece 71.6% \u63d0\u5347\u5230 75.6%\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u8bf4\u660e\u4e0d\u540c\u6a21\u578b\u53ef\u80fd\u5171\u4eab\u67d0\u4e9b\u7ed3\u6784\u6027\u5de5\u5177\u9519\u8bef\u3002\u5373\u4f7f\u5f3a\u6a21\u578b\u6574\u4f53\u80fd\u529b\u66f4\u9ad8\uff0c\u4e5f\u53ef\u80fd\u5728\u590d\u6742\u4efb\u52a1\u5206\u89e3\u3001\u5de5\u5177\u8bed\u4e49\u533a\u5206\u6216\u53c2\u6570\u6784\u9020\u4e0a\u72af\u4e0e\u5f31\u6a21\u578b\u76f8\u4f3c\u7684\u9519\u8bef\u3002<\/p>\n\n\n\n<p>\u4e0d\u8fc7\uff0c\u8de8\u6a21\u578b\u8fc1\u79fb\u5e76\u975e\u59cb\u7ec8\u4ea7\u751f\u6b63\u6536\u76ca\u3002\u4f8b\u5982\uff0cLlama 3.2 \u5b66\u5230\u7684\u89c4\u5219\u5e94\u7528\u5230 GPT-4o \u7684 ToolHop \u4efb\u52a1\u65f6\uff0c\u51c6\u786e\u7387\u4ece 58.1% \u4e0b\u964d\u5230 57.4%\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u66f4\u51c6\u786e\u7684\u7ed3\u8bba\u662f\uff1a\u89c4\u5219\u5177\u6709\u8f83\u5f3a\u7684\u8de8\u6a21\u578b\u8fc1\u79fb\u6f5c\u529b\uff0c\u4f46\u89c4\u5219\u5e93\u4e0e\u76ee\u6807\u6a21\u578b\u4e4b\u95f4\u53ef\u80fd\u5b58\u5728\u4e0d\u5339\u914d\uff0c\u9700\u8981\u989d\u5916\u7684\u9002\u7528\u6027\u9a8c\u8bc1\u548c\u8d1f\u8fc1\u79fb\u68c0\u6d4b\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"small-data-and-finetuning\">5.3 \u5c0f\u6837\u672c\u3001Prompt \u65b9\u6cd5\u4e0e\u5fae\u8c03\u6a21\u578b<\/h3>\n\n\n\n<p>\u5728 BFCL Multi-Turn-Base \u4e0a\uff0c\u8bba\u6587\u53ea\u4f7f\u7528 90 \u4e2a\u8bad\u7ec3\u6837\u672c\uff0c\u6700\u7ec8\u5b66\u4e60\u5230 4 \u6761\u89c4\u5219\u3002<\/p>\n\n\n\n<p>\u7ed3\u679c\u663e\u793a\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>55.2%\u219262.1%<\/code><\/p>\n\n\n\n<p>\u5bf9\u5e94\u968f\u673a\u6d4b\u8bd5\u96c6\uff1b<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>46.0%\u219260.0%<\/code><\/p>\n\n\n\n<p>\u5bf9\u5e94\u672a\u89c1\u5de5\u5177\u6d4b\u8bd5\u96c6\u3002<\/p>\n\n\n\n<p>\u8fd9\u8bf4\u660e\u5c11\u91cf\u9ad8\u4ef7\u503c\u89c4\u5219\u4e5f\u53ef\u80fd\u5e26\u6765\u660e\u663e\u63d0\u5347\uff0c\u5c24\u5176\u662f\u5728\u4efb\u52a1\u4e2d\u5b58\u5728\u91cd\u590d\u9519\u8bef\u7ed3\u6784\u65f6\u3002<\/p>\n\n\n\n<p>\u4e0e Prompt \u65b9\u6cd5\u76f8\u6bd4\uff0cRIMRULE \u540c\u6837\u8868\u73b0\u66f4\u597d\u3002<\/p>\n\n\n\n<p>\u5728 ToolHop Test-unseen \u4e0a\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>43.1% RIMRULE<\/code><\/p>\n\n\n\n<p>\u9ad8\u4e8e\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>37.9% Few-shot<\/code><\/p>\n\n\n\n<p>\u4ee5\u53ca\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>35.9% SEE<\/code><\/p>\n\n\n\n<p>\u8fd9\u8868\u660e\uff0c\u62bd\u8c61\u89c4\u5219\u53ef\u80fd\u6bd4\u76f4\u63a5\u68c0\u7d22\u539f\u59cb\u793a\u4f8b\u66f4\u9002\u5408\u8de8\u5de5\u5177\u6cdb\u5316\u3002<\/p>\n\n\n\n<p>RIMRULE \u8fd8\u80fd\u591f\u4e0e\u5fae\u8c03\u6a21\u578b\u7ed3\u5408\u3002\u4f8b\u5982\uff0c\u7ecf\u8fc7\u76d1\u7763\u5fae\u8c03\u7684 Llama 3.2 \u5728 ToolHop Test-unseen \u4e0a\u4ece 38.5% \u63d0\u5347\u5230 45.1%\uff1b\u52a0\u5165\u539f\u751f Function Calling \u7684 GPT-4o \u540e\uff0cToolHop Test-unseen \u4ece 83.0% \u63d0\u5347\u5230 85.4%\u3002<\/p>\n\n\n\n<p>\u8fd9\u8bf4\u660e\u663e\u5f0f\u89c4\u5219\u4e0e\u53c2\u6570\u5b66\u4e60\u5e76\u4e0d\u662f\u5b8c\u5168\u66ff\u4ee3\u5173\u7cfb\u3002\u5fae\u8c03\u53ef\u4ee5\u5b66\u4e60\u6574\u4f53\u5de5\u5177\u8c03\u7528\u5206\u5e03\uff0c\u800c\u89c4\u5219\u53ef\u4ee5\u4fee\u590d\u5fae\u8c03\u540e\u4ecd\u7136\u5b58\u5728\u7684\u5c40\u90e8\u7cfb\u7edf\u6027\u9519\u8bef\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"ablation\">5.4 \u7b26\u53f7\u68c0\u7d22\u4e0e MDL \u6574\u5408\u7684\u4f5c\u7528<\/h3>\n\n\n\n<p>\u8bba\u6587\u8fd8\u6bd4\u8f83\u4e86\u4e0d\u540c\u89c4\u5219\u68c0\u7d22\u65b9\u5f0f\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table><thead><tr><th>\u68c0\u7d22\u65b9\u5f0f<\/th><th>ToolHop<\/th><th>BFCL<\/th><\/tr><\/thead><tbody><tr><td>\u81ea\u7136\u8bed\u8a00\u68c0\u7d22<\/td><td>29.4%<\/td><td>54.1%<\/td><\/tr><tr><td>\u7b26\u53f7\u5f15\u5bfc\u68c0\u7d22<\/td><td>31.1%<\/td><td>56.6%<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u7b26\u53f7\u68c0\u7d22\u4f18\u4e8e\u76f4\u63a5\u4f7f\u7528\u81ea\u7136\u8bed\u8a00\u76f8\u4f3c\u5ea6\uff0c\u8bf4\u660e\u7ed3\u6784\u5316\u5b57\u6bb5\u6709\u52a9\u4e8e\u964d\u4f4e\u4e0d\u540c\u8bed\u8a00\u8868\u8fbe\u5e26\u6765\u7684\u5339\u914d\u8bef\u5dee\u3002<\/p>\n\n\n\n<p>\u5728\u89c4\u5219\u6574\u5408\u65b9\u9762\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-table is-style-stripes\"><table><thead><tr><th>\u6574\u5408\u65b9\u5f0f<\/th><th>ToolHop<\/th><th>BFCL<\/th><\/tr><\/thead><tbody><tr><td>LLM Prompt \u5408\u5e76<\/td><td>27.5%<\/td><td>52.1%<\/td><\/tr><tr><td>MDL \u5f15\u5bfc\u6574\u5408<\/td><td>31.1%<\/td><td>56.6%<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<p>\u76f4\u63a5\u8ba9 LLM \u5408\u5e76\u76f8\u4f3c\u89c4\u5219\uff0c\u53ea\u80fd\u5224\u65ad\u89c4\u5219\u6587\u672c\u662f\u5426\u76f8\u4f3c\uff1bMDL \u5f15\u5bfc\u65b9\u6cd5\u8fd8\u4f1a\u8003\u8651\u5220\u9664\u89c4\u5219\u540e Agent \u7684\u5b9e\u9645\u884c\u4e3a\u53d8\u5316\uff0c\u56e0\u6b64\u6548\u679c\u66f4\u597d\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"cases-and-rules\">5.5 \u6848\u4f8b\u4e0e\u89c4\u5219\u7c7b\u578b<\/h3>\n\n\n\n<p>\u8bba\u6587\u7ed9\u51fa\u7684\u4ee3\u8868\u6027\u6848\u4f8b\u662f\u67e5\u8be2\u67d0\u4f4d\u4eba\u7269\u5916\u7956\u7236\u540d\u5b57\u7684\u76f8\u5173\u5c5e\u6027\u3002<\/p>\n\n\n\n<p>\u9519\u8bef Agent \u8bd5\u56fe\u76f4\u63a5\u8c03\u7528\u5de5\u5177\u67e5\u8be2\u201c\u5916\u7956\u7236\u201d\uff0c\u4f46\u5de5\u5177\u53ea\u652f\u6301\u76f4\u63a5\u4eb2\u5c5e\u5173\u7cfb\uff0c\u56e0\u6b64\u8fd4\u56de\u65e0\u6570\u636e\u3002<\/p>\n\n\n\n<p>RIMRULE \u751f\u6210\u7684\u89c4\u5219\u8981\u6c42\uff1a\u5f53\u95ee\u9898\u5305\u542b\u590d\u6742\u4eb2\u5c5e\u5173\u7cfb\u65f6\uff0c\u5e94\u5148\u67e5\u8be2\u4e2d\u95f4\u4eba\u7269\uff0c\u4f8b\u5982\u5148\u83b7\u5f97\u6bcd\u4eb2\uff0c\u518d\u67e5\u8be2\u6bcd\u4eb2\u7684\u7236\u4eb2\u3002<\/p>\n\n\n\n<p>\u9644\u5f55\u8fd8\u5c55\u793a\u4e86\u5176\u4ed6\u7c7b\u578b\u7684\u89c4\u5219\uff0c\u4f8b\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u7528\u6237\u660e\u786e\u63d0\u4f9b\u53c2\u6570\u503c\u65f6\uff0c\u5e94\u4f18\u5148\u4f7f\u7528\u7528\u6237\u503c\uff0c\u800c\u4e0d\u662f Schema \u9ed8\u8ba4\u503c\uff1b<\/li>\n\n\n\n<li>\u67e5\u8be2\u5e93\u5b58\u3001\u5c3a\u5bf8\u6216\u989c\u8272\u53ef\u7528\u6027\u65f6\uff0c\u5e94\u9009\u62e9\u5e93\u5b58\u5de5\u5177\uff0c\u800c\u4e0d\u662f\u666e\u901a\u5546\u54c1\u8be6\u60c5\u5de5\u5177\uff1b<\/li>\n\n\n\n<li>\u5de5\u5177\u540d\u79f0\u5fc5\u987b\u4e0e Schema \u4e2d\u7684\u540d\u79f0\u5b8c\u5168\u4e00\u81f4\uff1b<\/li>\n\n\n\n<li>\u6267\u884c\u5177\u6709\u524d\u7f6e\u6761\u4ef6\u7684\u64cd\u4f5c\u524d\uff0c\u5e94\u5148\u5206\u89e3\u5e76\u9a8c\u8bc1\u6240\u6709\u4f9d\u8d56\u6b65\u9aa4\uff1b<\/li>\n\n\n\n<li>\u8fdb\u884c\u8d27\u5e01\u8f6c\u6362\u65f6\uff0c\u5e94\u6b63\u786e\u533a\u5206\u6e90\u8d27\u5e01\u4e0e\u76ee\u6807\u8d27\u5e01\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e9b\u89c4\u5219\u8bf4\u660e RIMRULE \u4e0d\u53ea\u662f\u5b66\u4e60\u67d0\u4e2a\u5de5\u5177\u7684\u56fa\u5b9a\u8c03\u7528\u65b9\u5f0f\uff0c\u800c\u662f\u5728\u5c1d\u8bd5\u5f52\u7eb3\u5de5\u5177\u4f7f\u7528\u4e2d\u7684\u901a\u7528\u63a8\u7406\u7ed3\u6784\u3002<\/p>\n\n\n\n<p>\u4ece\u5b9e\u9a8c\u7ed3\u679c\u6765\u770b\uff0cRIMRULE \u8bc1\u660e\u4e86\u5931\u8d25\u89c4\u5219\u5b66\u4e60\u5728\u5de5\u5177 Agent \u9002\u5e94\u4e2d\u5177\u6709\u5b9e\u9645\u4ef7\u503c\u3002<\/p>\n\n\n\n<p>\u5728\u6027\u80fd\u4e0a\uff0c\u5b83\u63d0\u5347\u4e86\u968f\u673a\u6d4b\u8bd5\u96c6\u548c\u672a\u89c1\u5de5\u5177\u6d4b\u8bd5\u96c6\u7684\u51c6\u786e\u7387\uff0c\u5e76\u4f18\u4e8e Few-shot \u548c\u5168\u5c40 Prompt \u4f18\u5316\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p>\u5728\u529f\u80fd\u4e0a\uff0c\u5b83\u5c06\u6a21\u578b\u9002\u5e94\u77e5\u8bc6\u4ece\u53c2\u6570\u4e2d\u5916\u5316\uff0c\u4f7f\u89c4\u5219\u53ef\u4ee5\u88ab\u67e5\u770b\u3001\u4fee\u6539\u548c\u8de8\u6a21\u578b\u590d\u7528\u3002<\/p>\n\n\n\n<p>\u4f46\u8bba\u6587\u4e5f\u5b58\u5728\u4e00\u4e9b\u5c40\u9650\u3002<\/p>\n\n\n\n<p><strong>\u4f9d\u8d56\u9ad8\u8d28\u91cf\u5931\u8d25\u53cd\u9988\u3002<\/strong> \u65b9\u6cd5\u901a\u5e38\u9700\u8981\u9519\u8bef\u8f68\u8ff9\u548c\u6b63\u786e\u8f68\u8ff9\u3002\u5982\u679c\u771f\u5b9e\u4e1a\u52a1\u4e2d\u65e0\u6cd5\u83b7\u5f97 Ground-truth \u6267\u884c\u8fc7\u7a0b\uff0c\u89c4\u5219\u751f\u6210\u8d28\u91cf\u4f1a\u53d7\u5230\u5f71\u54cd\u3002<\/p>\n\n\n\n<p><strong>\u89c4\u5219\u53ef\u80fd\u8fc7\u5ea6\u6cdb\u5316\u3002<\/strong> \u4e00\u6761\u5bf9\u67d0\u7c7b\u5931\u8d25\u6709\u6548\u7684\u5f3a\u5236\u89c4\u5219\uff0c\u53ef\u80fd\u5e72\u6270\u5176\u4ed6\u4efb\u52a1\uff0c\u751a\u81f3\u7834\u574f\u6a21\u578b\u539f\u672c\u6b63\u786e\u7684\u884c\u4e3a\u3002<\/p>\n\n\n\n<p><strong>\u7b26\u53f7\u5206\u7c7b\u53ef\u80fd\u5b58\u5728\u566a\u58f0\u3002<\/strong> \u8bba\u6587\u9644\u5f55\u4e2d\uff0c\u4e00\u6761\u901a\u7528\u7684\u5de5\u5177 Schema \u6821\u9a8c\u89c4\u5219\u88ab\u5212\u5206\u5230 <code>TRANSPORTATION<\/code> \u9886\u57df\uff0c\u8bf4\u660e\u7b26\u53f7\u6807\u7b7e\u4ecd\u7136\u53ef\u80fd\u53d7\u5230\u5177\u4f53\u6837\u672c\u8bed\u5883\u5f71\u54cd\u3002<\/p>\n\n\n\n<p><strong>\u8d2a\u5fc3\u6574\u5408\u4e0d\u4fdd\u8bc1\u5168\u5c40\u6700\u4f18\u3002<\/strong> \u89c4\u5219\u4e4b\u95f4\u53ef\u80fd\u5b58\u5728\u7ec4\u5408\u4f9d\u8d56\uff0c\u9010\u6761\u5220\u9664\u6216\u6cdb\u5316\u65e0\u6cd5\u53d1\u73b0\u6240\u6709\u66f4\u4f18\u7ec4\u5408\u3002<\/p>\n\n\n\n<p><strong>MDL \u516c\u5f0f\u5b58\u5728\u9000\u5316\u98ce\u9669\u3002<\/strong> \u5f53\u524d\u6570\u636e\u9879\u6ca1\u6709\u4ece\u76ee\u6807\u51fd\u6570\u4e0a\u4e25\u683c\u533a\u5206\u201c\u5168\u90e8\u7ea0\u6b63\u201d\u548c\u201c\u5b8c\u5168\u4e0d\u7ea0\u6b63\u201d\uff0c\u65b9\u6cd5\u7684\u6709\u6548\u6027\u4f9d\u8d56\u9ad8\u6027\u80fd\u521d\u59cb\u5316\u548c\u5c40\u90e8\u641c\u7d22\u8def\u5f84\u3002<\/p>\n\n\n\n<p><strong>\u8fd0\u884c\u6210\u672c\u62a5\u544a\u4e0d\u8db3\u3002<\/strong> \u89c4\u5219\u751f\u6210\u3001\u89c4\u5219\u9a8c\u8bc1\u3001\u7b26\u53f7\u8f6c\u6362\u548c\u5019\u9009\u89c4\u5219\u5e93\u8bc4\u4f30\u90fd\u9700\u8981\u989d\u5916\u6a21\u578b\u6216 Agent \u8c03\u7528\uff0c\u8bba\u6587\u6ca1\u6709\u5b8c\u6574\u62a5\u544a Token \u6210\u672c\u548c\u7aef\u5230\u7aef\u5ef6\u8fdf\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"alignment-thinking\">\u516d\u3001\u5bf9\u9f50\u601d\u8003<\/h2>\n\n\n\n<p>RIMRULE \u4e0e\u98df\u54c1\u914d\u6599\u6bd4\u4f8b\u53cd\u6f14\u7814\u7a76\u4e4b\u95f4\u5177\u6709\u8f83\u5f3a\u7684\u7ed3\u6784\u76f8\u4f3c\u6027\u3002<\/p>\n\n\n\n<p>\u98df\u54c1\u914d\u6599\u63a8\u7406\u540c\u6837\u5b58\u5728\u5927\u91cf\u7cfb\u7edf\u6027\u5931\u8d25\uff0c\u4f8b\u5982\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u4f30\u7b97\u6bd4\u4f8b\u8fdd\u53cd\u914d\u6599\u9012\u51cf\u987a\u5e8f\uff1b<\/li>\n\n\n\n<li>\u914d\u6599\u8425\u517b\u8d21\u732e\u65e0\u6cd5\u89e3\u91ca\u8425\u517b\u6210\u5206\u8868\uff1b<\/li>\n\n\n\n<li>\u590d\u5408\u914d\u6599\u88ab\u9519\u8bef\u5730\u5f53\u4f5c\u5355\u4e00\u539f\u6599\uff1b<\/li>\n\n\n\n<li>\u5355\u4f4d\u6362\u7b97\u9519\u8bef\u5bfc\u81f4\u6210\u672c\u5f02\u5e38\uff1b<\/li>\n\n\n\n<li>\u98df\u54c1\u7c7b\u522b\u5148\u9a8c\u4e0e\u4f30\u7b97\u7ed3\u679c\u51b2\u7a81\uff1b<\/li>\n\n\n\n<li>\u4f30\u7b97\u539f\u6599\u6210\u672c\u660e\u663e\u9ad8\u4e8e\u5546\u54c1\u552e\u4ef7\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e9b\u5931\u8d25\u4e0d\u5e94\u53ea\u88ab\u89c6\u4e3a\u5355\u4e2a\u6837\u672c\u7684\u9519\u8bef\uff0c\u800c\u53ef\u4ee5\u88ab\u8fdb\u4e00\u6b65\u62bd\u8c61\u4e3a\u53ef\u590d\u7528\u89c4\u5219\u3002<\/p>\n\n\n\n<p>\u53ef\u4ee5\u6784\u5efa\u5982\u4e0b\u6d41\u7a0b\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>\u98df\u54c1\u6807\u7b7e\u4e0e\u4ef7\u683c\u8f93\u5165 \u2192 \u521d\u59cb\u6bd4\u4f8b\u53cd\u6f14 \u2192 \u7ea6\u675f\u4e0e\u8bef\u5dee\u68c0\u67e5 \u2192 \u5931\u8d25\u539f\u56e0\u5206\u6790 \u2192 \u5019\u9009\u89c4\u5219\u751f\u6210 \u2192 \u89c4\u5219\u9a8c\u8bc1\u4e0e\u5165\u5e93 \u2192 \u65b0\u98df\u54c1\u52a8\u6001\u68c0\u7d22<\/code><\/p>\n\n\n\n<p>\u5177\u4f53\u6765\u8bf4\uff0c\u7cfb\u7edf\u5148\u8f93\u5165\u914d\u6599\u8868\u3001\u8425\u517b\u6210\u5206\u8868\u3001\u98df\u54c1\u7c7b\u522b\u3001\u51c0\u542b\u91cf\u3001\u552e\u4ef7\u548c\u539f\u6599\u4ef7\u683c\uff0c\u7531\u4f18\u5316\u6a21\u578b\u751f\u6210\u521d\u59cb\u914d\u6599\u6bd4\u4f8b\u3001\u8425\u517b\u8d21\u732e\u548c\u7406\u8bba\u6210\u672c\u3002<\/p>\n\n\n\n<p>\u968f\u540e\uff0c\u9a8c\u8bc1\u6a21\u5757\u68c0\u67e5\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u914d\u6599\u6bd4\u4f8b\u4e4b\u548c\u662f\u5426\u4e3a 100%\uff1b<\/li>\n\n\n\n<li>\u914d\u6599\u6392\u5e8f\u662f\u5426\u6ee1\u8db3\u975e\u589e\u7ea6\u675f\uff1b<\/li>\n\n\n\n<li>\u63a8\u7b97\u8425\u517b\u503c\u662f\u5426\u843d\u5728\u6807\u7b7e\u8bef\u5dee\u8303\u56f4\u5185\uff1b<\/li>\n\n\n\n<li>\u80fd\u91cf\u4e0e\u86cb\u767d\u8d28\u3001\u8102\u80aa\u3001\u78b3\u6c34\u5316\u5408\u7269\u4e4b\u95f4\u662f\u5426\u57fa\u672c\u4e00\u81f4\uff1b<\/li>\n\n\n\n<li>\u7406\u8bba\u539f\u6599\u6210\u672c\u662f\u5426\u5177\u6709\u5e02\u573a\u5408\u7406\u6027\uff1b<\/li>\n\n\n\n<li>\u662f\u5426\u8fdd\u53cd\u98df\u54c1\u7c7b\u522b\u6216\u6cd5\u89c4\u7ea6\u675f\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5bf9\u4e8e\u5931\u8d25\u6848\u4f8b\uff0c\u53ef\u4ee5\u4fdd\u5b58\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>(x,\u03c4\u207b,\u03c4*)<\/code><\/p>\n\n\n\n<p>\u5176\u4e2d\uff0c<code>\u03c4\u207b<\/code> \u662f\u9519\u8bef\u53cd\u6f14\u8fc7\u7a0b\uff0c<code>\u03c4*<\/code> \u662f\u7ecf\u8fc7\u4eba\u5de5\u6821\u6b63\u6216\u4f18\u5316\u6c42\u89e3\u5f97\u5230\u7684\u5408\u7406\u7ed3\u679c\u3002<\/p>\n\n\n\n<p>LLM \u6bd4\u8f83\u4e8c\u8005\u540e\uff0c\u53ef\u751f\u6210\u5982\u4e0b\u89c4\u5219\uff1a<\/p>\n\n\n\n<p>\u5982\u679c\u914d\u6599\u8868\u58f0\u660e\u6309\u6dfb\u52a0\u91cf\u9012\u51cf\u6392\u5217\uff0c\u5219\u4f30\u7b97\u6bd4\u4f8b\u5fc5\u987b\u6ee1\u8db3 <code>w_i\u2265 w_(i+1)<\/code>\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u67d0\u9879\u914d\u6599\u5c5e\u4e8e\u590d\u5408\u914d\u6599\uff0c\u5219\u4e0d\u80fd\u5c06\u5176\u5168\u90e8\u6bd4\u4f8b\u76f4\u63a5\u5f52\u5165\u67d0\u4e00\u79cd\u8425\u517b\u6765\u6e90\uff0c\u5e94\u5c55\u5f00\u5b50\u914d\u6599\u6216\u5efa\u7acb\u5b50\u914d\u6599\u533a\u95f4\u53d8\u91cf\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u4f30\u7b97\u51fa\u7684\u86cb\u767d\u8d28\u9ad8\u4e8e\u5f53\u524d\u914d\u6599\u7ec4\u5408\u7684\u7406\u8bba\u4e0a\u9650\uff0c\u5219\u5e94\u4f18\u5148\u68c0\u67e5\u9ad8\u86cb\u767d\u539f\u6599\u6bd4\u4f8b\u3001\u5e72\u7269\u8d28\u53e3\u5f84\u548c\u8425\u517b\u6570\u636e\u5e93\u6765\u6e90\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u7406\u8bba\u539f\u6599\u6210\u672c\u663e\u8457\u9ad8\u4e8e\u96f6\u552e\u4ef7\uff0c\u5219\u5e94\u68c0\u67e5\u5355\u4f4d\u6362\u7b97\u3001\u542b\u6c34\u7387\u3001\u53ef\u98df\u7387\u3001\u52a0\u5de5\u635f\u8017\u548c\u539f\u6599\u4ef7\u683c\u53e3\u5f84\u3002<\/p>\n\n\n\n<p>\u5982\u679c\u996e\u6599\u7c7b\u4ea7\u54c1\u4e2d\u4f30\u7b97\u6c34\u5206\u6bd4\u4f8b\u8fc7\u4f4e\uff0c\u540c\u65f6\u7cd6\u3001\u86cb\u767d\u8d28\u548c\u8102\u80aa\u65e0\u6cd5\u89e3\u91ca\u5269\u4f59\u8d28\u91cf\uff0c\u5219\u5e94\u63d0\u9ad8\u6c34\u6216\u57fa\u7840\u6db2\u4f53\u914d\u6599\u7684\u6bd4\u4f8b\u4e0b\u754c\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e9b\u89c4\u5219\u53ef\u4ee5\u4f7f\u7528\u7c7b\u4f3c RIMRULE \u7684\u53cc\u91cd\u8868\u793a\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>FoodCategory, ConstraintType, TriggerPattern, CorrectionAction, Strength, EvidenceSource<\/code><\/p>\n\n\n\n<p>\u4f8b\u5982\uff1a<\/p>\n\n\n\n<p>food_category = FERMENTED_DAIRY<br>constraint_type = NUTRIENT_UPPER_BOUND<br>trigger_pattern = PROTEIN_EXCEEDS_THEORETICAL_MAX<br>correction_action = [<br>    CHECK_HIGH_PROTEIN_INGREDIENT_RATIO,<br>    CHECK_DRY_MATTER_BASIS,<br>    VERIFY_NUTRIENT_SOURCE<br>]<br>strength = MANDATORY<br>evidence_source = NUTRITION_DATABASE<\/p>\n\n\n\n<p>\u81ea\u7136\u8bed\u8a00\u89c4\u5219\u7528\u4e8e\u751f\u6210\u89e3\u91ca\uff0c\u7b26\u53f7\u89c4\u5219\u7528\u4e8e\u68c0\u7d22\u3001\u51b2\u7a81\u68c0\u6d4b\u3001\u89c4\u5219\u7edf\u8ba1\u548c\u540e\u7eed\u4f18\u5316\u3002<\/p>\n\n\n\n<p>\u4e0d\u8fc7\uff0c\u98df\u54c1\u9886\u57df\u4e0d\u80fd\u76f4\u63a5\u7167\u642c RIMRULE \u7684\u89c4\u5219\u8bc4\u4ef7\u65b9\u5f0f\u3002\u56e0\u4e3a\u98df\u54c1\u914d\u65b9\u901a\u5e38\u7f3a\u5c11\u5b8c\u5168\u786e\u5b9a\u7684 Ground-truth\uff0c\u540c\u4e00\u5f20\u6807\u7b7e\u53ef\u80fd\u5bf9\u5e94\u591a\u7ec4\u53ef\u884c\u914d\u65b9\u3002\u56e0\u6b64\uff0c\u9700\u8981\u8bbe\u8ba1\u4e24\u7ea7\u89c4\u5219\u8bc4\u4ef7\u3002<\/p>\n\n\n\n<p>\u7b2c\u4e00\u5c42\u662f<strong>\u5408\u7406\u6027\u9a8c\u8bc1<\/strong>\uff1a\u52a0\u5165\u89c4\u5219\u540e\uff0c\u7ed3\u679c\u662f\u5426\u4ecd\u6ee1\u8db3\u8d28\u91cf\u5b88\u6052\u3001\u8425\u517b\u7ea6\u675f\u3001\u914d\u6599\u987a\u5e8f\u3001\u6cd5\u89c4\u8981\u6c42\u548c\u6210\u672c\u8fb9\u754c\uff1f<\/p>\n\n\n\n<p>\u7b2c\u4e8c\u5c42\u662f<strong>\u63a8\u7406\u6548\u7528\u9a8c\u8bc1<\/strong>\uff1a\u52a0\u5165\u89c4\u5219\u540e\uff0c\u8425\u517b\u8bef\u5dee\u662f\u5426\u4e0b\u964d\u3001\u53ef\u884c\u89e3\u7a7a\u95f4\u662f\u5426\u7f29\u5c0f\u3001\u7ed3\u679c\u662f\u5426\u66f4\u7a33\u5b9a\u3001\u5bf9\u7559\u51fa\u98df\u54c1\u7684\u9884\u6d4b\u662f\u5426\u6539\u5584\uff1f<\/p>\n\n\n\n<p>\u53ea\u6709\u540c\u65f6\u5177\u6709\u5408\u7406\u6027\u548c\u5b9e\u9645\u63a8\u7406\u589e\u76ca\u7684\u89c4\u5219\uff0c\u624d\u5e94\u8fdb\u5165\u6b63\u5f0f\u89c4\u5219\u5e93\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0c\u98df\u54c1\u89c4\u5219\u6574\u5408\u65f6\u4e0d\u5efa\u8bae\u76f4\u63a5\u7167\u642c\u8bba\u6587\u5f53\u524d\u7684\u6570\u636e\u4ee3\u4ef7\uff0c\u800c\u53ef\u4ee5\u4f7f\u7528\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>J(H)= \u03b1 L(H) + \u03b2_F N_(unfixed) + \u03b2_C N_(broken)<\/code><\/p>\n\n\n\n<p>\u5176\u4e2d\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><code>N_(unfixed)<\/code> \u8868\u793a\u4ecd\u672a\u4fee\u590d\u7684\u5931\u8d25\u6837\u672c\uff1b<\/li>\n\n\n\n<li><code>N_(broken)<\/code> \u8868\u793a\u52a0\u5165\u89c4\u5219\u540e\u88ab\u7834\u574f\u7684\u539f\u6b63\u786e\u6837\u672c\uff1b<\/li>\n\n\n\n<li><code>L(H)<\/code> \u8868\u793a\u89c4\u5219\u5e93\u590d\u6742\u5ea6\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u6837\u4e0d\u4ec5\u9f13\u52b1\u89c4\u5219\u4fee\u590d\u5386\u53f2\u5931\u8d25\uff0c\u8fd8\u80fd\u9632\u6b62\u65b0\u89c4\u5219\u635f\u5bb3\u5df2\u6709\u6b63\u786e\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u6700\u7ec8\u53ef\u4ee5\u5f62\u6210\u4e00\u4e2a\u9762\u5411\u98df\u54c1\u6807\u7b7e\u63a8\u7406\u7684\u81ea\u8fed\u4ee3\u95ed\u73af\uff1a<\/p>\n\n\n\n<p class=\"has-text-align-center\"><code>\u5931\u8d25\u6536\u96c6 \u2192 \u6839\u56e0\u5206\u6790 \u2192 \u89c4\u5219\u751f\u6210 \u2192 \u7ea6\u675f\u9a8c\u8bc1 \u2192 \u6548\u679c\u8bc4\u4f30 \u2192 \u89c4\u5219\u5165\u5e93 \u2192 \u52a8\u6001\u68c0\u7d22 \u2192 \u65b0\u4e00\u8f6e\u53cd\u6f14<\/code><\/p>\n\n\n\n<p>RIMRULE \u5bf9\u98df\u54c1\u914d\u6599\u7814\u7a76\u6700\u91cd\u8981\u7684\u542f\u53d1\uff0c\u662f\u4e0d\u8981\u53ea\u79ef\u7d2f\u5177\u4f53\u98df\u54c1\u6837\u672c\uff0c\u800c\u8981\u8fdb\u4e00\u6b65\u79ef\u7d2f\u80fd\u591f\u8de8\u54c1\u724c\u3001\u8de8\u914d\u65b9\u548c\u8de8\u98df\u54c1\u7c7b\u522b\u590d\u7528\u7684\u63a8\u7406\u89c4\u5219\u3002\u8fd9\u6837\uff0c\u7cfb\u7edf\u7684\u80fd\u529b\u589e\u957f\u5c31\u4e0d\u518d\u5b8c\u5168\u4f9d\u8d56\u6269\u5927\u8bad\u7ec3\u6570\u636e\u6216\u91cd\u65b0\u8bad\u7ec3\u6a21\u578b\uff0c\u800c\u53ef\u4ee5\u901a\u8fc7\u4e00\u4e2a\u53ef\u89e3\u91ca\u3001\u53ef\u5ba1\u8ba1\u548c\u53ef\u6301\u7eed\u66f4\u65b0\u7684\u98df\u54c1\u9886\u57df\u89c4\u5219\u5e93\u9010\u6b65\u5b9e\u73b0\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f5c\u8005\uff1aXiang Gao, Yuguang Yao et.c \u5355\u4f4d\uff1aIntuit AI Research\uff1bTe [&hellip;]<\/p>\n","protected":false},"author":75,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-5374","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5374","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/users\/75"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=5374"}],"version-history":[{"count":36,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5374\/revisions"}],"predecessor-version":[{"id":5455,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5374\/revisions\/5455"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=5374"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=5374"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=5374"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}