{"id":5329,"date":"2026-06-17T21:46:09","date_gmt":"2026-06-17T13:46:09","guid":{"rendered":"https:\/\/www.15zhi.net\/blog\/?p=5329"},"modified":"2026-06-17T21:46:09","modified_gmt":"2026-06-17T13:46:09","slug":"202606-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-ocr-critic-aligning-multimodal-large-language-models-perception-through-critical-feedback","status":"publish","type":"post","link":"https:\/\/www.15zhi.net\/blog\/202606-%e8%ae%ba%e6%96%87%e7%a0%94%e8%af%bb-ocr-critic-aligning-multimodal-large-language-models-perception-through-critical-feedback\/","title":{"rendered":"202606 \u8bba\u6587\u7814\u8bfb-OCR-Critic: Aligning Multimodal Large Language Models\u2019 Perception through Critical Feedback"},"content":{"rendered":"\n

\u4f5c\u8005\uff1aAuthors: Qiuna Tan, Runqi Qiao, Guanting Dong, YiFan Zhang, Minhui Wu, Jiapeng Wang, Miaoxuan Zhang, Yida Xu, Chong Sun, Chen Li, Honggang Zhang
\u6765\u6e90\uff1aACM MM 2025\u4f1a\u8bae
\u65f6\u95f4\uff1a2025.11<\/p>\n\n\n\n

\u80cc\u666f\u4e0e\u7814\u7a76\u95ee\u9898<\/p>\n\n\n\n

\u5927\u6a21\u578b\u88c1\u5224\u7684\u5174\u8d77\uff1a\u591a\u6a21\u6001\u5927\u6a21\u578b\uff08LMMs\uff09\u4f5c\u4e3a\u8bc4\u4f30\u5de5\u5177\uff08LMM-as-a-Judge\uff09\u5728\u7c97\u7c92\u5ea6\u8bc4\u4f30\uff08\u5982\u5bf9\u6bd4\u3001\u6392\u5e8f\uff09\u4e2d\u8868\u73b0\u4f18\u5f02\u3002\u4f46\u901a\u5e38\u4ec5\u652f\u6301\u6210\u5bf9\u6bd4\u8f83\uff08Pairwise comparison\uff09\u6216\u76f4\u63a5\u6253\u5206\/\u6392\u5e8f\uff08Ranking\uff09\uff0c\u65e0\u6cd5\u7ed9\u51fa\u5177\u4f53\u7684\u5b9a\u6027\u6539\u8fdb\u5efa\u8bae\u3002
\u5728OCR \u4efb\u52a1\u7684\u5c40\u9650\u6027\uff1a\u5728\u4f4e\u5206\u8fa8\u7387\u3001\u4e0d\u89c4\u5219\u6392\u7248\u548c\u89c6\u89c9\u6b67\u4e49\u573a\u666f\u4e0b\uff0c\u73b0\u6709\u6a21\u578b\u7f3a\u4e4f\u7ec6\u7c92\u5ea6\u53cd\u9988\u4e0e\u6df1\u5ea6\u7f3a\u9677\u5206\u6790\u80fd\u529b\u3002
\u8bba\u6587\u901a\u8fc7\u5bf9\u5927\u91cf\u5f00\u6e90 OCR \u6570\u636e\u96c6\u8fdb\u884c\u9519\u8bef\u6316\u6398\u53d1\u73b0\uff1a\u5f53\u524d\u5927\u6a21\u578b\u5728 OCR \u4efb\u52a1\u4e2d\u6700\u5927\u7684\u8f6f\u808b\u662f\u56fe\u50cf\u4fe1\u606f\u5b9a\u4f4d\uff08Localization Error\uff09\uff0c\u5728\u5404\u7c7b\u573a\u666f\u4e2d\u7684\u9519\u8bef\u7387\u9ad8\u8fbe 45% \u4ee5\u4e0a\u3002 \u7ecf\u5e38\u51fa\u73b0\u5b57\u7b26\u9057\u6f0f\u6216\u6392\u7248\u5bf9\u9f50\u9519\u8bef\u3002\u56e0\u6b64\uff0c\u5982\u4f55\u8ba9\u6a21\u578b\u5177\u5907\u2018\u81ea\u6211\u8bca\u65ad\u2019\u548c\u2018\u7ec6\u7c92\u5ea6\u6279\u8bc4\u2019\u7684\u80fd\u529b\uff0c\u662f\u5f53\u524d\u591a\u6a21\u6001 OCR \u9886\u57df\u7684\u5173\u952e\u74f6\u9888\u3002
\u89c6\u89c9\u7f16\u7801\u5668\u4e0e\u6587\u672c\u7a7a\u95f4\u5bf9\u9f50\u7684\u5931\u6548\uff0c\u662f\u5236\u7ea6\u5f53\u524d MLLM \u6267\u884c\u9ad8\u7cbe\u5ea6\u89c6\u89c9\u6587\u672c\u7406\u89e3\u7684\u6838\u5fc3\u74f6\u9888\u3002<\/p>\n\n\n\n

\u6838\u5fc3\u65b9\u6cd5\uff1aOCR-Critic \u6846\u67b6 (Proposed Solution)<\/p>\n\n\n\n

\u4e00\u4e2a\u6838\u5fc3\u6a21\u578b\uff1a\u63d0\u51fa\u4e86 OCR-Critic\uff0c\u4e00\u4e2a\u901a\u7528\u3001\u5373\u63d2\u5373\u7528\u7684\u591a\u6a21\u6001\u88c1\u5224\u6a21\u578b\u3002\u4e24\u5927\u521b\u65b0\u652f\u67f1\uff1a\u81ea\u52a8\u5316\u9519\u8bef\u6316\u6398\u7ba1\u9053\uff08Mining Pipeline\uff09\uff1a\u7ed3\u5408\u201c\u6a21\u578b\u5f15\u5bfc\u6807\u6ce8\u201d\u4e0e\u201c\u53d7\u63a7\u6270\u52a8\u5408\u6210\u201d\uff0c\u6784\u5efa\u4e86\u9996\u4e2a\u7ec6\u7c92\u5ea6 OCR \u9519\u8bef\u5206\u6790\u7684\u6307\u4ee4\u5fae\u8c03\u6570\u636e\u96c6\uff08\u6db5\u76d6 50 \u79cd\u9519\u8bef\u7c7b\u578b\u300130 \u79cd\u80fd\u529b\u8981\u6c42\uff09\u3002\u6e10\u8fdb\u5f0f\u76d1\u7763\u5fae\u8c03\uff08Progressive SFT\uff09\uff1a\u91c7\u7528\u4e09\u9636\u6bb5\u8bad\u7ec3\u6cd5\uff08\u5355\u4efb\u52a1 –> \u67e5\u8be2-\u54cd\u5e94\u89e3\u8026 –> \u591a\u4efb\u52a1\uff09\uff0c\u8ba9\u6a21\u578b\u4ece\u6613\u5230\u96be\u638c\u63e1\u5224\u522b\u80fd\u529b\u3002<\/p>\n\n\n\n

\u65b9\u6cd5\u7ec6\u8282\uff1a\u52a8\u6001 OCR \u9519\u8bef\u5bf9\u9f50\u7b56\u7565 (Methodology Details)<\/p>\n\n\n\n

\u4ece\u7c97\u5230\u7ec6\u7684\u5c42\u6b21\u5316\u4e09\u9636\u6bb5\u4f18\u5316\u7b56\u7565\uff08Dynamic Alignment\uff09\uff1a
Strategy 1: \u9a8c\u8bc1\u5f15\u5bfc\u8bad\u7ec3\uff08Verification-Guided\uff09\uff1a\u52a8\u6001\u8fc7\u6ee4\u672a\u6807\u6ce8\u6570\u636e\u4e2d\u7684\u566a\u70b9\u4f2a\u6807\u7b7e\uff0c\u53ea\u4fdd\u7559\u9ad8\u7f6e\u4fe1\u5ea6\u6837\u672c\u3002
Strategy 2: \u80fd\u529b\u5f15\u5bfc\u8bad\u7ec3\uff08Capability-Guided\uff09\uff1a\u52a8\u6001\u8bc4\u4f30 Actor \u6a21\u578b\u7684\u80fd\u529b\u7f3a\u5931\u5206\u5e03\uff0c\u636e\u6b64\u8c03\u6574\u540e\u7eed\u6837\u672c\u7684\u62bd\u6837\u6743\u91cd\uff08\u67e5\u7f3a\u8865\u6f0f\uff09\u3002
Strategy 3: \u9519\u8bef\u5f15\u5bfc\u8bad\u7ec3\uff08Error-Guided\uff09\uff1a\u5bf9 Actor \u6a21\u578b\u7684\u5931\u8d25\u6a21\u5f0f\u8fdb\u884c\u5206\u7c7b\uff0c\u4f18\u5148\u8ba9\u6a21\u578b\u66b4\u9732\u5728\u6700\u5e38\u89c1\u7684\u9519\u8bef\u6848\u4f8b\u4e2d\uff08\u9519\u9898\u96c6\u5f3a\u5316\uff09\u3002<\/p>\n\n\n\n

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

\u6838\u5fc3\u5185\u5bb9\u603b\u7ed3\uff1a
1.\u6784\u5efa\u4e86\u9996\u4e2a\u6db5\u76d6 50 \u79cd\u7ec6\u7c92\u5ea6 OCR \u9519\u8bef\u7c7b\u578b\u548c 30 \u79cd\u5e95\u5c42\u591a\u6a21\u6001\u80fd\u529b\u7684\u89c6\u89c9\u6307\u4ee4\u5fae\u8c03\u6570\u636e\u96c6\u3002
2.\u63d0\u51fa\u4e86\u901a\u7528\u3001\u5373\u63d2\u5373\u7528\u7684 OCR-Critic \u6a21\u578b\uff0c\u5e76\u72ec\u521b\u4e86\u201c\u9a8c\u8bc1-\u80fd\u529b-\u9519\u8bef\u201d\u4e09\u5c42\u9012\u8fdb\u7684\u5c42\u6b21\u5316\u52a8\u6001\u5bf9\u9f50\u7b56\u7565\uff08Dynamic Alignment\uff09\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u4f5c\u8005\uff1aAuthors: Qiuna Tan, Runqi Qiao, Guanting Dong, YiFan […]<\/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-5329","post","type-post","status-publish","format-standard","hentry","category-events"],"_links":{"self":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5329","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=5329"}],"version-history":[{"count":1,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5329\/revisions"}],"predecessor-version":[{"id":5333,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/posts\/5329\/revisions\/5333"}],"wp:attachment":[{"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/media?parent=5329"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/categories?post=5329"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.15zhi.net\/blog\/wp-json\/wp\/v2\/tags?post=5329"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}