{"id":4135,"date":"2025-10-15T13:15:16","date_gmt":"2025-10-15T05:15:16","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=4135"},"modified":"2025-10-15T13:15:16","modified_gmt":"2025-10-15T05:15:16","slug":"202510-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-reliable-reasoning-path-distilling-effective-guidance-for-llm-reasoning-with-knowledge-graphs","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202510-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-reliable-reasoning-path-distilling-effective-guidance-for-llm-reasoning-with-knowledge-graphs\/","title":{"rendered":"202510-\u8bba\u6587\u7814\u8bfb-Reliable reasoning path: Distilling effective guidance for llm reasoning with knowledge graphs"},"content":{"rendered":"\n<ul class=\"wp-block-list\">\n<li>\u4f5c\u8005\uff1aYilin Xiao \u7b49<\/li>\n\n\n\n<li>\u5355\u4f4d\uff1a\u9999\u6e2f\u7406\u5de5\u5927\u5b66\uff08PolyU\uff09<\/li>\n\n\n\n<li>\u53d1\u8868\u5728\uff1a\u9884\u5370\u672c<\/li>\n<\/ul>\n\n\n\n<h1 class=\"wp-block-heading\">\u6838\u5fc3\u5185\u5bb9<\/h1>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u95ee\u9898\u6839\u6e90\uff1aLLM\u4ea7\u751f\u5e7b\u89c9\u7684\u6839\u672c\u539f\u56e0\u662f\u5176\u5b66\u4e60\u7684\u662f\u7edf\u8ba1\u5b66\u5173\u8054\u800c\u975e\u7ed3\u6784\u5316\u903b\u8f91\uff0c\u5bfc\u81f4\u751f\u6210\u5185\u5bb9\u867d\u6d41\u7545\u4f46\u5b58\u5728\u4e8b\u5b9e\u9519\u8bef\u3002\u77e5\u8bc6\u56fe\u8c31\u88ab\u5e7f\u6cdb\u89c6\u4e3a\u89e3\u51b3\u6b64\u95ee\u9898\u7684\u5173\u952e\u65b9\u5411\u3002<\/li>\n\n\n\n<li>\u7814\u7a76\u7a7a\u767d\uff1a\u73b0\u6709KG-LLMQA\u65b9\u6cd5\u8fc7\u5ea6\u805a\u7126\u4e8e\u8bc1\u636e\u68c0\u7d22\uff0c\u5374\u5ffd\u89c6\u4e86\u77e5\u8bc6\u95f4\u7684\u903b\u8f91\u5173\u8054\u3002\u590d\u6742\u63a8\u7406\u95ee\u9898\u9700\u6574\u5408\u591a\u4e2a\u72ec\u7acb\u4e8b\u5b9e\uff0c\u4ec5\u4f18\u5316\u68c0\u7d22\u4e0d\u8db3\u4ee5\u652f\u6491\u6709\u6548\u63a8\u7406\u3002<\/li>\n\n\n\n<li>\u65b9\u6cd5\u521b\u65b0\uff1a\u63d0\u51fa\u53ef\u9760\u63a8\u7406\u8def\u5f84\uff08RRP\uff09\u6846\u67b6\uff0c\u901a\u8fc7\u5173\u7cfb\u5d4c\u5165\u4e0e\u53cc\u5411\u5206\u5e03\u5b66\u4e60\u5206\u522b\u62bd\u53d6\u903b\u8f91\u81ea\u6d3d\u8def\u5f84\u548c\u8bed\u4e49\u76f8\u5173\u8def\u5f84\uff0c\u5e76\u5f15\u5165\u53cd\u601d\u6a21\u5757\u8bc4\u4f30\u8def\u5f84\u4ef7\u503c\uff0c\u7b5b\u9009\u9ad8\u8d28\u91cf\u8def\u5f84\u7528\u4e8e\u7b54\u6848\u751f\u6210\u3002<\/li>\n\n\n\n<li>\u9a8c\u8bc1\u6548\u679c\uff1a\u5728WebQSP\u548cCWQ\u7b49\u7ecf\u5178KGQA\u6570\u636e\u96c6\u4e0a\u9a8c\u8bc1\uff0cRRP\u663e\u8457\u63d0\u5347\u6027\u80fd\uff0c\u4e14\u4ec5\u97007B\u53c2\u6570\u91cf\u7684\u5c0f\u578b\u6a21\u578b\u5373\u53ef\u5b9e\u73b0\u4f18\u5f02\u8868\u73b0\u3002<\/li>\n<\/ol>\n\n\n\n<h1 class=\"wp-block-heading\">\u80cc\u666f<\/h1>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u4f5c\u8005\u8ba4\u4e3a\uff0cLLM \u7684\u5e7b\u89c9\u7684\u6839\u6e90\u662f LLM \u5b66\u4e60\u5230\u7684\u662f\u7edf\u8ba1\u5b66\u4e0a\u7684\u8054\u7cfb\uff0c\u800c\u975e\u7ed3\u6784\u5316\u7684\u5e95\u5c42\u903b\u8f91\uff0c\u56e0\u6b64\u624d\u4f1a\u751f\u6210\u5305\u542b\u4e8b\u5b9e\u9519\u8bef\u3001\u4f46\u662f\u8bed\u8a00\u4e0a\u53c8\u5f88\u6d41\u7545\u7684\u7b54\u6848\uff1b\u5728\u8fd9\u4e00\u601d\u8003\u4e0b\uff0c\u73b0\u6709\u7684\u7814\u7a76\u5f80\u5f80\u5728\u601d\u8003\u5982\u4f55\u8fdb\u884c\u77e5\u8bc6\u6ce8\u5165\uff0c\u800c\u77e5\u8bc6\u56fe\u8c31\u662f\u4e00\u4e2a\u70ed\u95e8\u7684\u65b9\u5411\uff1b<\/li>\n\n\n\n<li>\u4f5c\u8005\u53d1\u73b0\uff0c\u8fc7\u53bb\u5927\u5bb6\u603b\u5728\u601d\u8003\u5982\u4f55\u4ece\u77e5\u8bc6\u56fe\u8c31\u4e2d\u68c0\u7d22\u51fa\u66f4\u597d\u7684\u8bc1\u636e\uff0c\u4f46\u662f\u5ffd\u7565\u4e86\u77e5\u8bc6\u4e4b\u95f4\u7684\u903b\u8f91\u5173\u7cfb\uff1b\u7531\u4e8e\u89e3\u51b3\u590d\u6742\u7684\u63a8\u7406\u95ee\u9898\u5fc5\u987b\u5c06\u591a\u4e2a\u72ec\u7acb\u7684\u4e8b\u5b9e\u8054\u7cfb\u8d77\u6765\u5904\u7406\uff0c\u56e0\u6b64\u4f5c\u8005\u8ba4\u4e3a\u5355\u7eaf\u601d\u8003\u68c0\u7d22\u7684\u95ee\u9898\u90a3\u662f\u4e0d\u591f\u7684\uff1b<\/li>\n\n\n\n<li>\u4f5c\u8005\u6307\u51fa\uff0c\u5f53\u524d\u7684 KG-LLMQA \u6709\u4e24\u4e2a\u96be\u5904\uff1a\u77e5\u8bc6\u56fe\u8c31\u592a\u8fc7\u590d\u6742\u4ee5\u53ca\u73b0\u6709\u7684\u65b9\u6cd5\u603b\u662f\u751f\u6210\u5927\u91cf\u6ca1\u4ec0\u4e48\u5e2e\u52a9\u7684\u63a8\u7406\u8def\u5f84\u3002\u4e3a\u4e86\u89e3\u51b3\u8fd9\u4e24\u4e2a\u95ee\u9898\uff0c\u4f5c\u8005\u63d0\u51fa\u4e86<strong>\u53ef\u9760\u63a8\u7406\u8def\u5f84\uff08RRP\uff09<\/strong>\u3002RRP \u5229\u7528\u5173\u7cfb\u5d4c\u5165\u548c\u53cc\u5411\u5206\u5e03\u5b66\u4e60\u6765\u5206\u522b\u62bd\u53d6\u51fa\u903b\u8f91\u81ea\u6d3d\u8def\u5f84\u548c\u8bed\u4e49\u76f8\u5173\u7684\u8def\u5f84\uff0c\u7136\u540e\u52a0\u5165\u4e00\u4e2a\u53cd\u601d\u6a21\u5757\u6765\u8bc4\u4f30\u6bcf\u4e2a\u8def\u5f84\u7684\u4ef7\u503c\uff0c\u5e76\u4e14\u8fdb\u884c\u7b5b\u9009\uff1b\u6700\u540e\u7b5b\u9009\u5f97\u5230\u7684\u8def\u5f84\u4f1a\u88ab\u7528\u6765\u751f\u6210\u7b54\u6848\u3002<\/li>\n\n\n\n<li>\u4f5c\u8005\u5728\u4e24\u4e2a\u7ecf\u5178\u7684 KGQA \u6570\u636e\u96c6\u4e0a\u9a8c\u8bc1\u4e86 RRP \u6548\u679c\uff0c\u8bc1\u660e RRP \u53ef\u4ee5\u663e\u8457\u63d0\u5347\u6548\u679c\uff0c\u5e76\u4e14\u53ea\u9700\u8981 7B \u8fd9\u6837\u7684\u5c0f\u578b\u6a21\u578b\u3002<\/li>\n<\/ol>\n\n\n\n<h1 class=\"wp-block-heading\">\u65b9\u6cd5\u8bba<\/h1>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"436\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-7-1024x436.png\" alt=\"\" class=\"wp-image-4136\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-7-1024x436.png 1024w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-7-300x128.png 300w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-7-768x327.png 768w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-7-1536x654.png 1536w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-7.png 1847w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\">\u8bed\u4e49\u63a8\u7406\u8def\u5f84\u751f\u6210<\/h2>\n\n\n\n<p>\u4f5c\u8005\u5e0c\u671b\uff0cLLM \u53ef\u4ee5\u6839\u636e\u95ee\u9898 $q$ \u751f\u6210\u76f4\u8fbe\u7b54\u6848 $a$ \u7684\u77e5\u8bc6\u56fe\u8c31\u63a8\u7406\u8def\u5f84 $\\gamma$\u3002\u6a21\u578b\u53ef\u4ee5\u88ab\u8868\u793a\u4e3a\u5148\u9a8c\u6982\u7387\u5206\u5e03 $P_{\\alpha}(\\gamma|q)$\uff1b\u7406\u60f3\u7684\u6982\u7387\u5206\u5e03\u5219\u662f\uff1a<br>$$<br>P(\\gamma|q,a,\\mathcal{G})=\\begin{cases}<br>\\frac{1}{\\Gamma}, &amp; \\exists \\gamma(e_{q},e_{a})\\in \\mathcal{G} \\<br>0, &amp; else<br>\\end{cases}<br>$$<br>\u5176\u4e2d\uff0c$\\Gamma$ \u662f\u4ece\u95ee\u9898\u5b9e\u4f53\u5230\u7b54\u6848\u5b9e\u4f53\u7684\u63a8\u7406\u8def\u5f84\u96c6\u5408\u3002\u4f5c\u8005\u5047\u5b9a\u7406\u60f3\u7684\u5206\u5e03\u662f\u4e00\u4e2a\u5747\u5300\u5206\u5e03\uff0c\u5982\u679c\u8fd9\u6837\u7684\u63a8\u7406\u8def\u5f84\u5728\u56fe\u4e2d\u5b58\u5728\uff0c\u90a3\u5b83\u4eec\u6709\u5747\u7b49\u7684\u673a\u4f1a\u51fa\u73b0\u3002<\/p>\n\n\n\n<p>\u63a5\u4e0b\u6765\u7684\u4efb\u52a1\u5c31\u662f\u60f3\u529e\u6cd5\u8fd1\u4f3c\u62df\u5408\u8fd9\u4e00\u7406\u60f3\u5206\u5e03\u3002\u5bf9\u4e8e\u6982\u7387\u5206\u5e03\u6765\u8bf4\uff0c\u6211\u4eec\u5e38\u5e38\u4f7f\u7528 <a href=\"KL\u6563\u5ea6.md\">KL\u6563\u5ea6<\/a>\u6765\u8861\u91cf\u6982\u7387\u5206\u5e03\u7684\u76f8\u4f3c\u6027\u3002\u56e0\u6b64\u635f\u5931\u51fd\u6570\uff1a<br>$$<br>\\mathcal{L}<em>{r<\/em>{1}}=D_{KL}[P(\\gamma|q,a,\\mathcal{G})||P_{\\alpha}(\\gamma|q)]<br>$$<br>\u63a8\u5bfc\u4e00\u4e0b\uff1a<br>$$<br>\\begin{array}l<br>\\mathcal{L}<em>{r<\/em>{1}} &amp; = &amp;\\displaystyle \\sum_{\\gamma \\in \\Gamma} \\frac{1}{\\Gamma}[-\\log\\Gamma-\\log P_{\\alpha}(\\gamma|q)] \\<br>&amp; = &amp; \\displaystyle -\\frac{1}{\\Gamma}\\sum_{\\gamma \\in\\Gamma} \\log P_{\\alpha}(\\gamma|q)-\\log\\Gamma<br>\\end{array}<br>$$<br>\u8fd9\u4e2a\u51fd\u6570\u5b83\u53ef\u4ee5\u6c42\u5bfc\uff0c\u6c42\u5bfc\u540e\u5e38\u6570\u9879\u5c31\u6d88\u5931\u4e86\u3002\u8981\u4f7f\u7528\u8fd9\u4e2a\u635f\u5931\u51fd\u6570\u8bad\u7ec3\uff0c\u6211\u4eec\u53ea\u9700\u8981\u627e\u4e00\u7ec4\u5305\u542b\u95ee\u9898\uff08$q$\uff09\u3001\u95ee\u9898\u5b9e\u4f53\uff08$e_{q}$\uff09\u3001\u7b54\u6848\uff08$a$\uff09\u3001\u7b54\u6848\u5b9e\u4f53\uff0c\u5e76\u4e14\u95ee\u9898\u5b9e\u4f53\u548c\u7b54\u6848\u5b9e\u4f53\u4e4b\u95f4\u7684\u8def\u5f84\uff08$\\Gamma$\uff09\u5df2\u77e5\u7684\u6570\u636e\u3002<\/p>\n\n\n\n<blockquote class=\"wp-block-quote is-layout-flow wp-block-quote-is-layout-flow\">\n<p>\u6a21\u578b\u751f\u6210\u4e00\u4e2a\u5e8f\u5217\u7684\u6982\u7387\u88ab\u770b\u4f5c\u7ec4\u6210\u8fd9\u4e2a\u5e8f\u5217\u6240\u6709\u7684 token \u7684\u751f\u6210\u6982\u7387\u7684\u79ef\u3002\u8fd9\u4e2a\u79ef\u7684\u672c\u8d28\u5176\u5b9e\u662f\u6761\u4ef6\u6982\u7387\u7684<strong>\u94fe\u5f0f\u6cd5\u5219<\/strong>\u3002<\/p>\n<\/blockquote>\n\n\n\n<h2 class=\"wp-block-heading\">\u7ed3\u6784\u63a8\u7406\u8def\u5f84\u751f\u6210<\/h2>\n\n\n\n<p>\u8fd9\u4e00\u4e2a\u6b65\u9aa4\u66f4\u52a0\u53e4\u5178\u4e5f\u66f4\u52a0\u590d\u6742\u3002\u5927\u81f4\u53ef\u4ee5\u5206\u4e3a\u4e24\u4e2a\u9636\u6bb5\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u5206\u6790\u95ee\u9898\uff1a\u5c06\u95ee\u9898\u5206\u89e3\u4e3a\u4e00\u7ec4\u77e5\u8bc6\u56fe\u8c31\u4e0a\u7684\u63a8\u7406\u6b65\u9aa4\uff1b<\/li>\n\n\n\n<li>\u56fe\u4e0a\u641c\u7d22\uff1a\u6839\u636e\u63a8\u7406\u6b65\u9aa4\u5728\u56fe\u4e0a\u8d70\u51fa\u4e00\u7ec4\u8def\u5f84\u8bc1\u636e\u3002<\/li>\n<\/ol>\n\n\n\n<h3 class=\"wp-block-heading\">\u5206\u6790\u95ee\u9898<\/h3>\n\n\n\n<p>\u5bf9\u4e8e\u95ee\u9898 $q$\uff0c\u4f5c\u8005\u4f7f\u7528\u4e86 <a href=\"Archive\/CV\/\u5faa\u73af\u795e\u7ecf\u7f51\u7edc.md\">RNN<\/a> \u601d\u8def\u6765\u5c06\u5176\u8f6c\u6362\u4e3a\u4e00\u7ec4\u63a8\u7406\u6b65\u9aa4\u3002\u9996\u5148\u7528 GloVe \u5c06 $q$ \u8fdb\u884c<a href=\"Archive\/\u673a\u5668\u5b66\u4e60\/\u8bcd\u5d4c\u5165.md\">\u8bcd\u5d4c\u5165<\/a>\uff1a<br>$$<br>{ \\mathbf{v}<em>{q}^{ini} }=\\text{GloVe}(q) $$ \u63a5\u7740\uff0c\u5d4c\u5165\u540e\u7684\u5411\u91cf\u4f7f\u7528 <a href=\"Archive\/CV\/\u5faa\u73af\u795e\u7ecf\u7f51\u7edc.md\">RNN<\/a> \u8fdb\u884c\u7f16\u7801\uff0c\u7f16\u7801\u53d6\u6700\u540e\u7684\u9690\u85cf\u5c42\u5411\u91cf\u4f5c\u4e3a $q$ \u7684\u7f16\u7801\u5411\u91cf\u3002\u8fd9\u662f encoder \u7684\u5e38\u89c1\u505a\u6cd5\uff1a $$ \\mathbf{v}<\/em>{q}=\\text{LSTM}({ \\mathbf{v}<em>{q}^{ini} }) $$ \u7136\u540e\u7528\u4e00\u4e2a LSTM decoder \u5bf9\u5176\u89e3\u7801\uff0c\u89e3\u7801\u4e3a\u4e00\u7ec4\u5411\u91cf ${ \\mathbf{w}<\/em>{i} }^n$\uff0c\u8fd9\u7ec4\u5411\u91cf\u88ab\u89c6\u4f5c\u5206\u6790\u95ee\u9898\u5f97\u5230\u7684\u89e3\u9898\u63a8\u7406\u6b65\u9aa4\uff0c\u4e5f\u5c31\u662f\u6240\u8c13\u7684\u5bfc\u822a\u6307\u4ee4\uff08instruction\uff09\uff0c\u88ab\u7528\u4e8e\u6307\u5bfc\u56fe\u4e0a\u641c\u7d22\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u56fe\u4e0a\u641c\u7d22<\/h2>\n\n\n\n<p>\u4e00\u5f00\u59cb\uff0c\u56fe\u4e0a\u6240\u6709\u7684\u5173\u7cfb\u4f1a\u88ab\u8d4b\u4e88\u4e00\u4e2a\u53ef\u5b66\u4e60\u7684\u8bcd\u5d4c\u5165\u5411\u91cf $\\mathbf{v}<em>{r}$\uff0c\u8fd9\u4e2a\u6709\u5f85\u8bad\u7ec3\u3002\u6211\u4eec\u7528\u5173\u7cfb\u5411\u91cf\u5b9a\u4e49\u5b9e\u4f53\u5411\u91cf\u3002<strong>\u6bcf\u4e2a\u5b9e\u4f53\u5411\u91cf\u90fd\u662f\u5176\u5468\u56f4\u6240\u6709\u5173\u7cfb\u7684\u52a0\u603b<\/strong>\uff1a $$ \\mathbf{v}<\/em>{e}^{0}=\\sigma\\left( \\sum_{(e,r,e&#8217;)\\in \\mathcal{G}} \\mathbf{v}<em>{r}W<\/em>{1} \\right)<br>$$<br>\u5934\u4e0a\u7684 0 \u8868\u793a\u8fd9\u662f\u5b9e\u4f53\u5411\u91cf\u7684\u521d\u59cb\u72b6\u6001\u3002\u5b9e\u4f53\u5411\u91cf\u4f1a\u5728\u4e00\u8f6e\u8f6e\u8fed\u4ee3\u4e2d\u66f4\u65b0\u3002<\/p>\n\n\n\n<p>\u5bf9\u4e8e\u77e5\u8bc6\u56fe\u8c31\u4e0a\u7684\u6bcf\u4e2a\u4e09\u5143\u7ec4\uff0c\u6211\u4eec\u90fd\u6839\u636e\u5f53\u524d\u7684\u5bfc\u822a\u6307\u4ee4 $\\mathbf{w}<em>{i}$\uff0c\u7f16\u7801\u4e00\u4e2a match \u5411\u91cf\uff1a $$ \\mathbf{m}<\/em>{(e,r,e&#8217;)}^i=\\sigma(\\mathbf{w}<em>{i} \\odot W<\/em>{2}\\mathbf{v}<em>{r}) $$ match \u5411\u91cf\u4ee3\u8868\u8fd9\u4e00\u5173\u7cfb\u8bed\u4e49\u4e0a\u662f\u4e0d\u662f\u548c\u5f53\u524d\u7684\u5bfc\u822a\u6307\u4ee4\u76f8\u4f3c\uff0c\u63a5\u4e0b\u6765\u4f1a\u88ab\u7528\u4e8e\u8bc4\u4f30\u4ece e \u5468\u56f4\u7684\u67d0\u5173\u7cfb\u591a\u5927\u7a0b\u5ea6\u4e0a\u662f\u7b26\u5408\u5bfc\u822a\u6307\u4ee4\u7684\u3002\u7136\u540e\u518d\u5c06\u5176\u6309\u6743\u52a0\u603b\uff0c\u6c47\u96c6\u6210 e \u7684\u4fe1\u606f\u5411\u91cf\uff1a $$ \\tilde{\\mathbf{v}}<\/em>{e}^i=\\sum_{(e,r,e&#8217;)\\in \\mathcal{G}}P^{i-1}<em>{e&#8217;}\\mathbf{m}^{i}<\/em>{(e,r,e&#8217;)}<br>$$<br>\u5176\u4e2d\uff0c$P^{i-1}<em>{e&#8217;}$ \u662f\u4e00\u4e2a\u6982\u7387\u503c\uff0c\u8868\u793a\u4e0a\u4e00\u4e2a\u5bfc\u822a\u6307\u4ee4\u4e2d e&#8217; \u591a\u5927\u7a0b\u5ea6\u4e0a\u503c\u5f97\u4f5c\u4e3a\u4e0b\u4e00\u6b65\u9009\u62e9\u3002\u6700\u540e\u66f4\u65b0\u5b9e\u4f53\u5411\u91cf\uff1a $$ \\mathbf{v}<\/em>{e}^{i}=\\text{FFN}([\\mathbf{v}<em>{e}^{i-1};\\tilde{\\mathbf{v}}<\/em>{e}^{i}])<br>$$<br>\u7136\u540e\u518d\u5ea6\u6839\u636e\u5b9e\u4f53\u5411\u91cf\u6c47\u603b\u6210\u7684\u77e9\u9635\uff0c\u7528\u4e00\u6b21\u53d8\u6362\u4e4b\u540e\u8ba1\u7b97\u51fa logit \u5411\u91cf\uff0c\u6700\u540e\u8ba1\u7b97\u51fa\u6982\u7387\uff1a<br>$$<br>P^{i}=\\text{Softmax}(V^{iT}W)<br>$$<br>\u901a\u8fc7\u4e0a\u8ff0\u6b65\u9aa4\uff0c\u6211\u4eec\u9488\u5bf9\u6bcf\u4e00\u6b65\u5bfc\u822a\u6307\u4ee4\uff0c\u90fd\u8ba1\u7b97\u51fa\u4e86\u4e00\u4e2a\u6982\u7387\u56fe $P^{i}$\u3002\u7406\u8bba\u4e0a\u53ea\u8981\u627e\u5230\u7b2c\u4e00\u4e2a\u95ee\u9898\u4e2d\u7684\u5b9e\u4f53 $e_{q}$\uff0c\u7136\u540e\u5728 $P^{1}$ \u56fe\u4e0a\u627e\u5230\u5468\u56f4\u6982\u7387\u6700\u5927\u7684\u90bb\u5c45\uff0c\u7136\u540e\u8d70\u5411\u8fd9\u4e2a\u90bb\u5c45\u5373\u53ef\u3002\u8fd9\u6837 n \u8f6e\u4e4b\u540e\u7406\u8bba\u4e0a\u5c31\u62b5\u8fbe\u4e86\u7ec8\u70b9 $e_{a}$\u3002<\/p>\n\n\n\n<p>\u7406\u8bba\u4e0a\uff0c\u8fd9\u6761\u8def\u5f84\u662f\u53ef\u9006\u7684\uff0c\u5c31\u662f\u8bf4\u4ece $e_{a}$ \u4e5f\u80fd\u7528\u76f8\u540c\u7684\u65b9\u5f0f\u5728 n \u8f6e\u4e4b\u540e\u627e\u5230 $e_{q}$\u3002\u6240\u4ee5\u4f5c\u8005\u4e5f\u53e6\u5916\u8bad\u7ec3\u4e86\u4e00\u5957\u4ece\u7ec8\u70b9\u56de\u6eaf\u5230\u8d77\u70b9\u7684\u89e3\u7801\u5668\uff0c\u529b\u6c42\u4e24\u4e2a\u89e3\u7801\u5668\u8ba1\u7b97\u51fa\u7684\u67d0\u4e00\u6b65\u6982\u7387\u56fe\u80fd\u591f\u5b9e\u73b0\u7b2c $i$ \u6b65\u5206\u5e03 $P^{i}<em>{f}$ \u63a5\u8fd1\u9006\u5411\u56de\u6eaf\u89e3\u7801\u5f97\u5230\u7684\u7b2c $n-i$ \u6b65\u6982\u7387\u5206\u5e03 $P^{n-i}<\/em>{b}$\u3002\u4f5c\u8005\u4f7f\u7528 <a href=\"KL\u6563\u5ea6.md\">KL\u6563\u5ea6<\/a>\u548c <a href=\"\u7434\u751f-\u9999\u519c\u6563\u5ea6.md\">JS\u6563\u5ea6<\/a>\u6765\u8861\u91cf\u6982\u7387\u5206\u5e03\u7684\u76f8\u4f3c\u6027\uff0c\u7531\u6b64\u5f97\u5230\u635f\u5931\u51fd\u6570\uff1a<br>$$<br>\\mathcal{L}<em>{r<\/em>{2}}=D_{KL}(P_{f}^{i}||P_{f}^{<em>})+D_{KL}(P_{b}^{i}||P_{b}^{<\/em>})+\\sum_{i=1}^{n-1} D_{JS}(P_{f}^{i}||P_{b}^{n-i})<br>$$<br>\u5176\u4e2d\uff0c$P^*$ \u662f\u6570\u636e\u96c6\u7ed9\u51fa\u7684\u7b54\u6848\u63a8\u7406\u8def\u5f84\u5f62\u6210\u7684\u5206\u5e03\uff08\u5047\u8bbe\u4e3a\u5747\u5300\u5206\u5e03\uff09\uff1b$D_{JS}$ \u7528\u4e8e\u8861\u91cf\u4e24\u4e2a\u5206\u5e03\u4e4b\u95f4\u7684\u7406\u8bba\u76f8\u4f3c\u6027\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">\u518d\u601d\u8003<\/h2>\n\n\n\n<p>\u518d\u601d\u8003\u90e8\u5206\u4f7f\u7528\u4e24\u4e2a\u6253\u5206\u5668\u5206\u522b\u8bc4\u4f30\u8def\u5f84\u7684\u8bed\u4e49\u76f8\u4f3c\u5ea6\u548c\u7ed3\u6784\u76f8\u4f3c\u5ea6\uff0c\u7136\u540e\u52a0\u6743\u3002\u6700\u540e\u53ef\u4ee5\u5f97\u51fa\u4e00\u4e2a\u8def\u5f84\u7684\u8bc4\u5206\uff0c\u7b5b\u53bb\u90a3\u4e9b\u4e0d\u592a\u6709\u7528\u7684\u90e8\u5206\u3002<br><img decoding=\"async\" src=\".\/Assets\/Reliable%20Reasoning%20Path-1760499700916.png\" alt=\"\"><\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"929\" height=\"983\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-8.png\" alt=\"\" class=\"wp-image-4137\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-8.png 929w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-8-284x300.png 284w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-8-768x813.png 768w\" sizes=\"auto, (max-width: 929px) 100vw, 929px\" \/><\/figure>\n\n\n\n<h1 class=\"wp-block-heading\">\u5b9e\u9a8c<\/h1>\n\n\n\n<p>\u7ecf\u5178\u7684 WebQSP \u548c CWQ\uff0c\u5e76\u4e14\u83b7\u5f97\u4e86\u6700\u597d\u7684\u6548\u679c\uff1a<br><img decoding=\"async\" src=\".\/Assets\/Reliable%20Reasoning%20Path-1760499939927.png\" alt=\"\"><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"773\" src=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-9-1024x773.png\" alt=\"\" class=\"wp-image-4138\" srcset=\"https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-9-1024x773.png 1024w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-9-300x227.png 300w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-9-768x580.png 768w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-9-1536x1160.png 1536w, https:\/\/www.15zhi.net\/blog\/wp-content\/uploads\/2025\/10\/image-9.png 1560w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h1 class=\"wp-block-heading\">\u7ed3\u8bba<\/h1>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u63d0\u51fa\u4e86 RRP\uff0c\u4e00\u4e2a\u7ed3\u5408\u4e86 LLM \u7684\u77e5\u8bc6\u56fe\u8c31\u589e\u5f3a\u63a8\u7406\u7cfb\u7edf\u3002RRP \u6709\u6548\u5730\u7ed3\u5408\u4e86LLM \u7684\u8bed\u8a00\u80fd\u529b\u4ee5\u53ca\u53cc\u5411\u5b66\u4e60\u7f16\u7801\u5668\u63d0\u4f9b\u7684\u4e30\u5bcc\u7ed3\u6784\u4fe1\u606f\uff0c\u4fdd\u8bc1\u4e86\u63a8\u7406\u7684\u5e7f\u5ea6\u548c\u7cbe\u5ea6\uff0c\u5927\u5927\u63d0\u9ad8\u4e86\u63a8\u7406\u8d28\u91cf\u3002<\/li>\n\n\n\n<li>\u505a\u4e86\u5145\u5206\u7684\u5b9e\u9a8c\u5bf9\u6bd4\u8bc1\u660e\u4e86 RRP \u7684\u80fd\u529b\uff0c\u5728\u591a\u4e2a\u57fa\u51c6\u4e0a\u8d85\u8d8a\u4e86\u6240\u6709\u7684\u57fa\u7ebf\uff0c\u83b7\u5f97\u4e86 SOTA\u3002\u800c\u4e14\u672c\u8eab RRP \u662f\u4e00\u79cd plug and play \u7684\u6846\u67b6\uff0c\u53ef\u4ee5\u9002\u7528\u4e8e\u4efb\u4f55 LLM\uff0c\u51cf\u5c11\u4e86\u8fc1\u79fb\u5230\u5176\u4ed6\u95ee\u9898\u4e0a\u6240\u9700\u7684\u5de5\u4f5c\u3002<\/li>\n<\/ol>\n\n\n\n<h1 class=\"wp-block-heading\">\u542f\u53d1\u4e0e\u8bc4\u4ef7<\/h1>\n\n\n\n<ol class=\"wp-block-list\">\n<li>\u4f5c\u8005\u6ca1\u6709\u4f7f\u7528\u5355\u4e00\u7684 LLM \u751f\u6210\u6216\u662f KG-Retrieval\uff0c\u800c\u662f\u7ed3\u5408\u4e8c\u8005\u4e00\u8d77\u6765\u6253\u5206\uff0c\u6700\u7ec8\u7b5b\u9009\u51fa\u4e00\u7ec4\u5e72\u51c0\u7684\u8bc1\u636e\u3002\u8fd9\u5145\u5206\u7ed3\u5408\u4e86\u4e8c\u8005\u7684\u4f18\u70b9\uff0c\u5b9e\u9a8c\u4e5f\u53ef\u4ee5\u652f\u6301\u8fd9\u6837\u7684\u5047\u8bbe\uff1b<\/li>\n\n\n\n<li>\u4f5c\u8005\u5728\u6587\u7ae0\u4e2d\u63cf\u8ff0\u65b9\u6cd5\u63cf\u8ff0\u5f97\u5e76\u4e0d\u6e05\u695a\uff1b\u5982\u679c\u4e0d\u4f7f\u7528\u5176\u4ed6\u6280\u5de7\u7684\u8bdd\uff0c\u8fd9\u4e00\u65b9\u6cd5\u9700\u8981\u56f4\u7ed5\u77e5\u8bc6\u56fe\u8c31\u8bad\u7ec3\uff0c\u4e14\u9700\u8981\u9488\u5bf9\u5168\u56fe\u5b9e\u4f53\u8fdb\u884c\u8ba1\u7b97\u3002\u8ba1\u7b97\u91cf\uff08\u5305\u62ec\u8bad\u7ec3\uff09\u5c06\u968f\u7740\u77e5\u8bc6\u56fe\u8c31\u7684\u5927\u5c0f\u7ebf\u6027\u589e\u52a0\uff0c\u5f88\u96be\u60f3\u8c61\u4f5c\u8005\u5982\u4f55\u5728\u6587\u7ae0\u4e2d\u8bf4\u7684\u90a3\u79cd\u5177\u6709\u6570\u5343\u4e07\u5b9e\u4f53\u7684\u5927\u578b\u77e5\u8bc6\u56fe\u8c31\u4e0a\u5b9e\u73b0\u8fd9\u6837\u7684\u8bad\u7ec3\uff1b<\/li>\n\n\n\n<li>\u4f5c\u8005\u5ba3\u79f0\u7684 plug-and-play \u5176\u5b9e\u662f\u865a\u5047\u7684\u3002\u56e0\u4e3a RRP \u65e2\u9700\u8981\u4e13\u95e8\u8bad\u7ec3\u7684 LLM \uff0c\u53c8\u9700\u8981\u9488\u5bf9\u5177\u4f53\u77e5\u8bc6\u56fe\u8c31\u548c\u95ee\u9898\u6570\u636e\u96c6\u505a\u53cc\u5411\u8bad\u7ec3\uff0c\u80fd\u66f4\u6362\u7684\u5176\u5b9e\u53ea\u6709\u56de\u7b54\u5c42\u7684 LLM\u3002\u5c06\u8fd9\u4e00\u65b9\u6cd5\u8fc1\u79fb\u5230\u5176\u4ed6\u6570\u636e\u96c6\u4e0a\u7684\u6210\u672c\u662f\u975e\u5e38\u9ad8\u7684\uff1b<\/li>\n\n\n\n<li>\u6ca1\u6709\u5728\u4efb\u4f55\u5730\u65b9\u627e\u5230\u6e90\u4ee3\u7801\uff0c\u627e\u4e0d\u5230\u6a21\u7cca\u90e8\u5206\u7684\u6839\u636e\u3002<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>\u6838\u5fc3\u5185\u5bb9 \u80cc\u666f \u65b9\u6cd5\u8bba \u8bed\u4e49\u63a8\u7406\u8def\u5f84\u751f\u6210 \u4f5c\u8005\u5e0c\u671b\uff0cLLM \u53ef\u4ee5\u6839\u636e\u95ee\u9898 $q$ \u751f\u6210\u76f4\u8fbe\u7b54\u6848 $a$ \u7684\u77e5\u8bc6 [&hellip;]<\/p>\n","protected":false},"author":59,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-4135","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4135","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\/59"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=4135"}],"version-history":[{"count":2,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4135\/revisions"}],"predecessor-version":[{"id":4310,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4135\/revisions\/4310"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=4135"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=4135"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=4135"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}