{"id":4910,"date":"2026-03-16T14:10:24","date_gmt":"2026-03-16T06:10:24","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=4910"},"modified":"2026-03-16T14:10:24","modified_gmt":"2026-03-16T06:10:24","slug":"202603-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-chain-of-verification-reduceshallucination-in-large-language-models","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202603-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-chain-of-verification-reduceshallucination-in-large-language-models\/","title":{"rendered":"202603 \u8bba\u6587\u7814\u8bfb-Chain-of-Verification ReducesHallucination in Large Language Models"},"content":{"rendered":"\n

\u4f5c\u8005\uff1aShehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, Jason Weston (Meta AI)
\u5355\u4f4d\uff1a\u56fd\u9632\u79d1\u6280\u5927\u5b66
\u6765\u6e90\uff1aACL 2024 Findings
\u65f6\u95f4\uff1a2024.08<\/p>\n\n\n\n

\u5e7b\u89c9\u5b9a\u4e49: LLM \u751f\u6210\u770b\u4f3c\u5408\u7406\u4f46\u4e8b\u5b9e\u9519\u8bef\u7684\u4fe1\u606f
\u5e7b\u89c9\u7279\u70b9:
\u4f4e\u9891\u77e5\u8bc6\uff08\u957f\u5c3e\u4e8b\u5b9e\uff09\u66f4\u5bb9\u6613\u4ea7\u751f\u5e7b\u89c9
\u957f\u6587\u672c\u751f\u6210\u6bd4\u77ed\u56de\u7b54\u66f4\u5bb9\u6613\u5e7b\u89c9
\u5e7b\u89c9\u5185\u5bb9\u5728\u8868\u9762\u4e0a\u6781\u5177\u8ff7\u60d1\u6027\uff0c\u7528\u6237\u96be\u4ee5\u81ea\u884c\u5224\u65ad
\u73b0\u6709\u65b9\u6cd5\u4e0d\u8db3:
Chain-of-Thought (CoT) \u4fc3\u8fdb\u63a8\u7406\u4f46\u4e0d\u68c0\u67e5\u4e8b\u5b9e\uff0c\u751a\u81f3\u53ef\u80fd\u589e\u52a0\u5e7b\u89c9
\u68c0\u7d22\u589e\u5f3a (RAG) \u9700\u8981\u5916\u90e8\u77e5\u8bc6\u5e93\uff0c\u6210\u672c\u9ad8
\u6307\u4ee4\u5fae\u8c03 (Instruction Tuning) \u672a\u5fc5\u51cf\u5c11\u5e7b\u89c9\uff0c\u53ef\u80fd\u4ea7\u751f\u66f4\u591a\u9519\u8bef\u8f93\u51fa<\/p>\n\n\n\n

\u6838\u5fc3\u6311\u6218: LLM \u5728\u4e00\u6b21\u6027\u751f\u6210(one-pass generation)\u4e2d\u4ea7\u751f\u4e86\u9519\u8bef\uff0c\u4f46\u5b83\u81ea\u5df1\u80fd\u5426\u53d1\u73b0\u5e76\u4fee\u6b63\u8fd9\u4e9b\u9519\u8bef\uff1f
\u5173\u952e\u5047\u8bbe:
LLM \u56de\u7b54\u77ed\u95ee\u9898\u6bd4\u751f\u6210\u957f\u6587\u672c\u66f4\u51c6\u786e\uff08\u8fd9\u662fCoVe\u7684\u7406\u8bba\u57fa\u7840\uff09
\u5373\uff1a\u5c06”\u9a8c\u8bc1\u4e00\u6bb5\u8bdd\u662f\u5426\u6b63\u786e”\u62c6\u89e3\u4e3a\u591a\u4e2a”\u56de\u7b54\u5177\u4f53\u5c0f\u95ee\u9898”\uff0c\u53ef\u4ee5\u63d0\u9ad8\u51c6\u786e\u7387
\u7814\u7a76\u76ee\u6807: \u8bbe\u8ba1\u4e00\u79cd\u65e0\u9700\u5916\u90e8\u5de5\u5177\u3001\u65e0\u9700\u989d\u5916\u8bad\u7ec3\u7684\u81ea\u9a8c\u8bc1\u65b9\u6cd5\uff0c\u4ec5\u901a\u8fc7\u63d0\u793a\u7b56\u7565\u8ba9\u6a21\u578b\u81ea\u6211\u7ea0\u9519
\u5173\u952e\u516c\u5f0f\u5316\u8868\u8fbe:\u77ed\u95ee\u9898\u51c6\u786e\u7387 > \u957f\u6587\u672c\u4e8b\u5b9e\u5bc6\u5ea6 \u2192 \u62c6\u89e3\u9a8c\u8bc1 > \u4e00\u6b21\u6027\u751f\u6210<\/p>\n\n\n\n

\u89e3\u51b3\u529e\u6cd5- CoVe \u65b9\u6cd5 \u603b\u89c8<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u63d0\u51fa\u4e86\u4e00\u79cd\u7ed3\u5408\u51b2\u7a81\u68c0\u6d4b\u548c\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u65b0\u6846\u67b6\uff0c\u4e3b\u8981\u6d41\u7a0b\u5982\u4e0b\uff1a
\u77e5\u8bc6\u56fe\u8c31\u5d4c\u5165\uff08KGE\uff09\uff1a\u4f7f\u7528TransE\u7b49\u6280\u672f\u5bf9\u73b0\u6709KG\u8fdb\u884c\u5d4c\u5165\u5b66\u4e60\uff0c\u5e76\u5b9a\u4e49\u8bc4\u5206\u51fd\u6570\u6765\u8861\u91cf\u4e09\u5143\u7ec4\u7684\u201c\u56f0\u60d1\u5ea6\u201d\uff08Perplexity\uff0c\u5373\u9519\u8bef\u53ef\u80fd\u6027\uff09\u3002
\u5173\u7cfb\u5206\u7c7b\uff1a\u5c06KG\u4e2d\u7684\u5173\u7cfb\u548c\u5c5e\u6027\u5206\u4e3a\u4e24\u7c7b\uff1a
1-to-1 \u5173\u7cfb\uff1a\u4e00\u4e2a\u5934\u5b9e\u4f53\u53ea\u5bf9\u5e94\u4e00\u4e2a\u5c3e\u5b9e\u4f53\uff08\u5982\u201c\u4f4d\u4e8e\u201d\u3001\u201c\u51fa\u751f\u65e5\u671f\u201d\uff09\u3002
Non-1-to-1 \u5173\u7cfb\uff1a\u4e00\u4e2a\u5934\u5b9e\u4f53\u53ef\u5bf9\u5e94\u591a\u4e2a\u5c3e\u5b9e\u4f53\uff08\u5982\u201c\u5b66\u751f\u201d\u3001\u201c\u4f5c\u8005\u201d\uff09\u3002
\u5206\u9636\u6bb5\u89e3\u51b3\u7b56\u7565\uff1a
\u9488\u5bf9 1-to-1 \u5173\u7cfb\uff1a\u76f4\u63a5\u5229\u7528\u8bc4\u5206\u51fd\u6570\uff0c\u9009\u62e9\u56f0\u60d1\u5ea6\u6700\u4f4e\uff08\u6700\u53ef\u4fe1\uff09\u7684\u5019\u9009\u9879
\u4f5c\u4e3a\u771f\u7406\u3002
\u9488\u5bf9 Non-1-to-1 \u5173\u7cfb\uff1a
\u9996\u5148\u7528\u8bc4\u5206\u51fd\u6570\u8fc7\u6ee4\u6389\u660e\u663e\u9519\u8bef\u7684\u5019\u9009\u9879\uff08\u56f0\u60d1\u5ea6\u9ad8\u4e8e\u9608\u503c\uff09\u3002
\u5bf9\u4e8e\u5269\u4f59\u7684\u96be\u4f8b\uff0c\u5229\u7528LLM\u8fdb\u884c\u6700\u7ec8\u5224\u65ad\u3002
LLM \u63d0\u793a\u5de5\u7a0b\uff08Prompt Engineering\uff09\uff1a\u8bbe\u8ba1\u4e86\u5305\u542b\u4e09\u4e2a\u90e8\u5206\u7684\u63d0\u793a\u6a21\u677f\uff1a
\u4efb\u52a1\u58f0\u660e\uff1a\u660e\u786e\u544a\u8bc9LLM\u4efb\u52a1\u662f\u51b2\u7a81\u6d88\u89e3\u3002
\u6f14\u793a\uff08Demonstrations\uff09\uff1a\u4eceKG\u4e2d\u62bd\u53d6\u76f8\u5173\u4e09\u5143\u7ec4\uff0c\u901a\u8fc7\u201c\u4e09\u5143\u7ec4\u7ffb\u8bd1\u201d\u8f6c\u5316\u4e3a\u81ea\u7136\u8bed\u8a00\u63cf\u8ff0\uff0c\u4f5c\u4e3a\u4e0a\u4e0b\u6587\u63d0\u4f9b\u7ed9LLM\uff0c\u589e\u5f3a\u5176\u5bf9\u5b9e\u4f53\u7684\u7406\u89e3\u3002
\u8f93\u5165\u5019\u9009\u9879\uff1a\u5f85\u5224\u65ad\u7684\u5916\u90e8\u4e09\u5143\u7ec4\u5217\u8868\u3002<\/p>\n\n\n\n

Generate Baseline: \u7ed9\u5b9a\u67e5\u8be2\uff0cLLM \u6b63\u5e38\u751f\u6210\u521d\u59cb\u56de\u7b54\uff08\u5373 baseline\uff09
Plan Verifications: \u57fa\u4e8e\u67e5\u8be2+\u521d\u59cb\u56de\u7b54\uff0cLLM \u81ea\u52a8\u751f\u6210\u4e00\u7cfb\u5217\u9a8c\u8bc1\u95ee\u9898\uff08\u975e\u6a21\u677f\u5316\uff0cLLM\u81ea\u7531\u751f\u6210\uff09
Execute Verifications: \u72ec\u7acb\u56de\u7b54\u6bcf\u4e2a\u9a8c\u8bc1\u95ee\u9898\uff08\u5173\u952e\uff1a\u4e0d\u8ba9\u6a21\u578b\u770b\u5230\u521d\u59cb\u56de\u7b54\uff0c\u907f\u514d\u590d\u5236\u540c\u6837\u7684\u9519\u8bef\uff09
Generate Final: \u7efc\u5408\u9a8c\u8bc1\u7ed3\u679c\u4e0e\u521d\u59cb\u56de\u7b54\uff0c\u751f\u6210\u4fee\u6b63\u540e\u7684\u6700\u7ec8\u56de\u7b54<\/p>\n\n\n\n

\u5b9e\u9a8c\u4efb\u52a1\u4e0e\u8bc4\u4f30\u6307\u6807<\/p>\n\n\n\n

\u4efb\u52a1<\/td>\u6570\u636e\u96c6<\/td>\u8bc4\u4f30\u6307\u6807<\/td>\u4efb\u52a1\u7279\u70b9<\/td><\/tr>
\u5217\u8868\u578b\u95ee\u9898<\/td>Wikidata API (56\u9898)<\/td>Precision (micro)<\/td>\u7b54\u6848\u662f\u5b9e\u4f53\u5217\u8868\uff0c\u5224\u65ad\u5e7b\u89c9\u5b9e\u4f53\u6570<\/td><\/tr>
\u5217\u8868\u578b\u95ee\u9898<\/td>Wiki-Category (55\u9898)<\/td>Precision (micro)<\/td>\u7c7b\u4f3cWikidata\uff0c\u4f46\u5b9e\u4f53\u66f4\u590d\u6742<\/td><\/tr>
\u591a\u7b54\u6848QA<\/td>MultiSpanQA (418\u9898)<\/td>F1 Score<\/td>\u95ed\u5377\u8bbe\u7f6e\uff0c\u9700\u8981\u591a\u4e2a\u72ec\u7acb\u7b54\u6848<\/td><\/tr>
\u957f\u6587\u672c\u751f\u6210<\/td>\u4eba\u7269\u4f20\u8bb0 (\u4efb\u9009\u4eba\u7269)<\/td>FACTSCORE<\/td>\u68c0\u6d4b\u957f\u6587\u672c\u4e2d\u6bcf\u4e2a\u4e8b\u5b9e\u58f0\u660e\u7684\u51c6\u786e\u6027<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n

\u57fa\u7ebf\u6a21\u578b: Llama 65B\uff08few-shot\uff09
\u5bf9\u6bd4\u65b9\u6cd5: Zero-Shot, Few-Shot, CoT, Instruction-Tuned (Llama 2), InstructGPT, ChatGPT, PerplexityAI<\/p>\n\n\n\n

\u5b9e\u9a8c\u7ed3\u679c<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

\u6838\u5fc3\u5185\u5bb9\u603b\u7ed3<\/p>\n\n\n\n

1.\u63d0\u51fa\u201c\u5148\u68c0\u6d4b\uff0c\u540e\u89e3\u51b3\u201d\uff08Detect-Then-Resolve\uff09 \u7684\u5dee\u5f02\u5316\u5904\u7406\u7b56\u7565\uff0c\u4f7f\u7528\u7cbe\u7ec6\u5316\u51b2\u7a81\u611f\u77e5\uff0c\u5229\u7528LLM\u4f5c\u4e3a\u5916\u90e8\u77e5\u8bc6\u589e\u5f3a\u5668\uff0c\u7ed3\u6784\u5316\u63d0\u793a\u589e\u5f3a\u3002
2.\u63d0\u51fa\u57fa\u4e8e\u5173\u7cfb\u7c7b\u578b\u7684\u81ea\u9002\u5e94\u51b2\u7a81\u68c0\u6d4b\u673a\u5236\u3002\u9996\u6b21\u663e\u5f0f\u5730\u5728\u51b2\u7a81\u6d88\u89e3\u6846\u67b6\u4e2d\u5f15\u5165\u51b2\u7a81\u68c0\u6d4b\u6b65\u9aa4\uff0c\u5e76\u6839\u636e\u5173\u7cfb\u7c7b\u578b\uff081-to-1 vs Non-1-to-1\uff09\u52a8\u6001\u8c03\u6574\u89e3\u51b3\u7b56\u7565\u3002
3.\u63d0\u51fa\u878d\u5408LLM\u7684\u6df7\u5408\u8fc7\u6ee4\u67b6\u6784\u3002\u6784\u5efa\u4e86\u4e00\u4e2a\u7ea7\u8054\u8fc7\u6ee4\u7cfb\u7edf\u3002\u5bf9\u4e8e\u7b80\u5355\u60c5\u51b5\u4f7f\u7528\u8f7b\u91cf\u7ea7\u7684\u5d4c\u5165\u8bc4\u5206\uff0c\u5bf9\u4e8e\u590d\u6742\u60c5\u51b5\uff08\u6d89\u53ca\u672a\u89c1\u5b9e\u4f53\u6216\u6a21\u7cca\u5173\u7cfb\uff09\u8c03\u7528LLM\u8fdb\u884c\u6df1\u5ea6\u63a8\u7406\u3002
4.\u8bbe\u8ba1\u9762\u5411\u51b2\u7a81\u6d88\u89e3\u7684\u201c\u4e09\u5143\u7ec4\u7ffb\u8bd1\u201d\u63d0\u793a\u7b56\u7565\u3002\u8bbe\u8ba1\u4e86\u72ec\u7279\u7684Prompt\u6784\u9020\u65b9\u6cd5\uff0c\u7279\u522b\u662f\u201c\u4e09\u5143\u7ec4\u7ffb\u8bd1\u201d\uff08Triple Translation\uff09\u73af\u8282\u3002\u5b83\u4e0d\u662f\u76f4\u63a5\u5c06\u4e09\u5143\u7ec4\u6254\u7ed9LLM\uff0c\u800c\u662f\u5148\u4eceKG\u4e2d\u91c7\u6837\u76f8\u5173\u4e8b\u5b9e\uff0c\u8ba9LLM\uff08\u6216\u9884\u5904\u7406\u6a21\u5757\uff09\u5c06\u5176\u751f\u6210\u4e00\u6bb5\u81ea\u7136\u7684\u5b9e\u4f53\u63cf\u8ff0\u6587\u672c\uff0c\u4f5c\u4e3aFew-shot\u6f14\u793a\u6ce8\u5165Prompt\u3002<\/p>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

\u4f5c\u8005\uff1aShehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Rober […]<\/p>\n","protected":false},"author":66,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[],"class_list":["post-4910","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4910","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\/66"}],"replies":[{"embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/comments?post=4910"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4910\/revisions"}],"predecessor-version":[{"id":4913,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/4910\/revisions\/4913"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=4910"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=4910"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=4910"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}