{"id":5041,"date":"2026-04-17T17:03:45","date_gmt":"2026-04-17T09:03:45","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=5041"},"modified":"2026-04-17T17:03:45","modified_gmt":"2026-04-17T09:03:45","slug":"202604-%e8%ae%ba%e6%96%87%e9%98%85%e8%af%bb%ef%bc%9ais-cognition-consistent-with-perceptionassessing-and-mitigating-multimodal-knowledge-conflicts-in-document-understanding","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202604-%e8%ae%ba%e6%96%87%e9%98%85%e8%af%bb%ef%bc%9ais-cognition-consistent-with-perceptionassessing-and-mitigating-multimodal-knowledge-conflicts-in-document-understanding\/","title":{"rendered":"202604 \u8bba\u6587\u9605\u8bfb\uff1aIs Cognition Consistent with Perception?Assessing and Mitigating Multimodal Knowledge Conflicts in Document Understanding"},"content":{"rendered":"\n

\u4f5c\u8005\uff1aZirui Shao1, Feiyu Gao, Zhaoqing Zhu, Chuwei Luo, Hangdi Xing, Zhi Yu, Qi Zheng, Ming Yan, Jiajun Bu1
\u5355\u4f4d\uff1a\u6d59\u6c5f\u5927\u5b66\u6d59\u6c5f\u7701\u65e0\u969c\u788d\u611f\u77e5\u4e0e\u667a\u80fd\u7cfb\u7edf\u91cd\u70b9\u5b9e\u9a8c\u5ba4\u3001\u963f\u91cc\u5df4\u5df4
\u6765\u6e90\uff1aACL 2025
\u65f6\u95f4\uff1a2024.11<\/p>\n\n\n\n

\u7814\u7a76\u80cc\u666f<\/p>\n\n\n\n

\u968f\u7740\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08MLLM\uff09\u7684\u53d1\u5c55\uff0c\u6587\u6863\u7406\u89e3\u5df2\u7ecf\u4ece\u5355\u7eaf\u7684 OCR \u6f14\u8fdb\u5230\u4e86\u590d\u6742\u7684\u6587\u6863\u95ee\u7b54\uff08DocVQA\uff09\u3002\u7136\u800c\uff0c\u7814\u7a76\u4eba\u5458\u53d1\u73b0 MLLM \u7ecf\u5e38\u51fa\u73b0\u201c\u903b\u8f91\u65ad\u5c42\u201d\uff1a\u5b83\u80fd\u6b63\u786e\u56de\u7b54\u5173\u4e8e\u6587\u6863\u5185\u5bb9\u7684\u95ee\u9898\uff0c\u4f46\u5f53\u4f60\u8ba9\u5b83\u8bc6\u522b\u56fe\u7247\u4e2d\u5bf9\u5e94\u7684\u6587\u5b57\u65f6\uff0c\u5b83\u5374\u8bc6\u522b\u9519\u4e86\uff1b\u53cd\u4e4b\u4ea6\u7136\u3002\u8fd9\u79cd\u4e0d\u4e00\u81f4\u6027\u9650\u5236\u4e86\u6a21\u578b\u5728\u91d1\u878d\u3001\u533b\u7597\u7b49\u9ad8\u7cbe\u5ea6\u573a\u666f\u7684\u5e94\u7528\u3002<\/p>\n\n\n\n

\u7814\u7a76\u95ee\u9898<\/p>\n\n\n\n

\u8ba4\u77e5\u4e0e\u611f\u77e5\u7684\u4e8c\u5143\u5bf9\u7acb\uff1a\u4f5c\u8005\u8ba4\u4e3a MLLM \u5185\u90e8\u5b58\u5728\u4e24\u4e2a\u5e76\u884c\u7684\u8fc7\u7a0b\u3002\u611f\u77e5\uff08Perception\uff09\u8d1f\u8d23\u5c06\u56fe\u50cf\u50cf\u7d20\u8f6c\u5316\u4e3a\u6587\u672c\u5e8f\u5217\uff08OCR\uff09\uff0c\u8ba4\u77e5\uff08Cognition\uff09\u8d1f\u8d23\u6839\u636e\u6307\u4ee4\u751f\u6210\u8bed\u4e49\u7b54\u6848\uff08VQA\uff09\u3002
\u8bad\u7ec3\u566a\u58f0\u7684\u6839\u6e90\uff1a\u8bba\u6587\u6307\u51fa\uff0c\u73b0\u6709\u7684\u591a\u6a21\u6001\u6570\u636e\u96c6\uff08\u5982 LAION\uff09\u901a\u5e38\u5305\u542b\u5927\u91cf\u7684\u673a\u5668\u751f\u6210 OCR \u6807\u7b7e\u6216\u8d28\u91cf\u53c2\u5dee\u4e0d\u9f50\u7684\u4f17\u5305\u95ee\u7b54\u5bf9\u3002\u5f53\u6a21\u578b\u540c\u65f6\u5b66\u4e60\u8fd9\u4e9b\u4e0d\u4e00\u81f4\u7684\u6570\u636e\u65f6\uff0c\u5b83\u5b9e\u9645\u4e0a\u662f\u5728\u5b66\u4e60\u201c\u77db\u76fe\u201d\u3002<\/p>\n\n\n\n

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

\u89e3\u51b3\u529e\u6cd5- MKCF \u6846\u67b6<\/p>\n\n\n\n

\u63d0\u51fa\u4e86\u4e00\u79cd\u540d\u4e3a Multimodal Knowledge Consistency Fine-tuning (MKCF) \u7684\u65b9\u6cd5\uff1a
\u6570\u636e\u6784\u5efa\uff1a\u6784\u9020\u5305\u542b\u611f\u77e5\u8def\u5f84\u548c\u8ba4\u77e5\u8def\u5f84\u7684\u8bad\u7ec3\u6570\u636e\u3002
\u611f\u77e5\u8def\u5f84\uff1a\u7ed9\u5b9a\u5750\u6807\uff0c\u8981\u6c42\u8bc6\u522b\u6587\u672c\u3002
\u8ba4\u77e5\u8def\u5f84\uff1a\u6839\u636e\u56fe\u50cf\u5185\u5bb9\u56de\u7b54\u95ee\u9898\u3002
\u5173\u8054\u8def\u5f84\uff1a\u8981\u6c42\u6a21\u578b\u8bf4\u660e\u7b54\u6848\u662f\u5728\u54ea\u4e2a\u5750\u6807\u4f4d\u7f6e\u627e\u5230\u7684\u3002
\u635f\u5931\u51fd\u6570\u4f18\u5316\uff1a\u5728\u8bad\u7ec3\u4e2d\u5f15\u5165\u4e00\u81f4\u6027\u504f\u7f6e\uff0c\u5982\u679c\u611f\u77e5\u4e0e\u8ba4\u77e5\u51b2\u7a81\uff0c\u4f1a\u7ed9\u6a21\u578b\u4e00\u4e2a\u8f83\u5927\u7684\u60e9\u7f5a\u3002
C&P Link Tokens\uff1a\u5728\u6a21\u578b\u5904\u7406\u6d41\u7a0b\u4e2d\uff0c\u5f15\u5165\u7279\u5b9a\u7684\u683c\u5f0f\u5316\u8f93\u51fa\u3002\u6a21\u578b\u5fc5\u987b\u4ee5 [Location] -> [Text] -> [Answer] \u7684\u94fe\u5f0f\u7ed3\u6784\u8fdb\u884c\u601d\u8003\u3002\u8fd9\u79cd\u201c\u663e\u5f0f\u63a8\u7406\u201d\u8feb\u4f7f\u89c6\u89c9\u7279\u5f81\u5728\u8fdb\u5165\u8bed\u8a00\u89e3\u7801\u5668\u4e4b\u524d\uff0c\u5df2\u7ecf\u5b8c\u6210\u4e86\u7a7a\u95f4\u4e0a\u7684\u5bf9\u9f50\u3002
Negative Sample Training\uff08\u8d1f\u6837\u672c\u8bad\u7ec3\uff09\uff1a\u4e3b\u52a8\u5582\u7ed9\u6a21\u578b\u4e00\u4e9b\u9519\u8bef\u7684\u611f\u77e5\u7ed3\u679c\uff0c\u8bad\u7ec3\u6a21\u578b\u5177\u5907\u201c\u7ea0\u9519\u80fd\u529b\u201d\u3002\u8ba9\u6a21\u578b\u5b66\u4f1a\uff1a\u5982\u679c\u611f\u77e5\u8f93\u5165\u662f\u6709\u8bef\u7684\uff08\u4f8b\u5982\u7531\u4e8e\u5149\u7167\u4ea7\u751f\u7684\u566a\u58f0\uff09\uff0c\u5982\u4f55\u5229\u7528\u8bed\u4e49\u80cc\u666f\u77e5\u8bc6\u8fdb\u884c\u53cd\u5411\u4fee\u6b63\u3002<\/p>\n\n\n\n

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

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

\u5728 DocVQA\u3001DeepForm \u7b49\u4e94\u4e2a\u4e3b\u6d41\u6587\u6863\u7406\u89e3\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u6d4b\u8bd5\uff0c\u8986\u76d6\u4e86\u8868\u683c\u3001\u8868\u5355\u3001\u56fe\u8868\u7b49\u591a\u79cd\u6587\u6863\u5f62\u6001\u3002<\/p>\n\n\n\n

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

\u6838\u5fc3\u5185\u5bb9\u603b\u7ed3\uff1a
1.\u5b9a\u4e49\u5e76\u91cf\u5316\u4e86 C&P (Cognition & Perception) \u77e5\u8bc6\u51b2\u7a81\uff0c\u5e76\u63d0\u51fa\u4e86\u201cC&P\u4e00\u81f4\u6027\u201d\u8bc4\u4f30\u6307\u6807\u3002\u5b83\u4e0d\u518d\u53ea\u770b VQA \u7b54\u5bf9\u6ca1\u6709\uff0c\u800c\u662f\u5f3a\u884c\u8981\u6c42\u6a21\u578b\u5148\u8fdb\u884c\u4e00\u6b21\u89c6\u89c9\u5b9a\u4f4d\uff08Pointing\/OCR\uff09\uff0c\u518d\u770b\u751f\u6210\u7684\u7b54\u6848\u662f\u5426\u4e0e\u5b9a\u4f4d\u5230\u7684\u6587\u672c\u5185\u5bb9\u4e00\u81f4\u3002\u8fd9\u4e3a\u8bc4\u4ef7\u6a21\u578b\u201c\u5e7b\u89c9\u201d\u63d0\u4f9b\u4e86\u4e00\u4e2a\u53ef\u91cf\u5316\u7684\u7ef4\u5ea6\u3002
2.\u63d0\u51fa\u591a\u6a21\u6001\u77e5\u8bc6\u4e00\u81f4\u6027\u5fae\u8c03 (MKCF) \u3002
\u5750\u6807\u951a\u5b9a\uff08Coordinate Anchoring\uff09\uff1a\u5728\u5fae\u8c03\u65f6\uff0c\u4e0d\u4ec5\u4ec5\u8ba9\u6a21\u578b\u8f93\u51fa\u7b54\u6848\uff0c\u8fd8\u8981\u6c42\u8f93\u51fa\u5bf9\u5e94\u7684\u6587\u672c\u8fb9\u754c\u6846\uff08Bounding Box\uff09\u3002
\u8de8\u6a21\u6001\u94fe\u8def\u6807\u7b7e\uff08Link Tokens\uff09\uff1a\u901a\u8fc7\u52a0\u5165 \u7b49\u6807\u8bb0\uff0c\u4eba\u4e3a\u5730\u5728\u795e\u7ecf\u7f51\u7edc\u7684\u6ce8\u610f\u529b\u673a\u5236\u4e2d\u57cb\u4e0b\u201c\u94a9\u5b50\u201d\uff0c\u8feb\u4f7f\u8ba4\u77e5\u8f93\u51fa\u5fc5\u987b\u5411\u611f\u77e5\u8f93\u51fa\u5bf9\u9f50\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u4f5c\u8005\uff1aZirui Shao1, Feiyu Gao, Zhaoqing Zhu, Chuwei Luo, Ha […]<\/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-5041","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5041","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=5041"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5041\/revisions"}],"predecessor-version":[{"id":5046,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5041\/revisions\/5046"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=5041"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=5041"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=5041"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}