
{"id":10531,"date":"2026-07-30T19:18:44","date_gmt":"2026-07-30T11:18:44","guid":{"rendered":"https:\/\/infernews.com\/blog\/evidence-attribution-in-visual-document-understanding-without-coordinates-or-reg\/"},"modified":"2026-07-30T19:35:29","modified_gmt":"2026-07-30T11:35:29","slug":"evidence-attribution-in-visual-document-understanding-without-coordinates-or-reg","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/evidence-attribution-in-visual-document-understanding-without-coordinates-or-reg\/","title":{"rendered":"quote-and-retrieve\uff1a\u7528\u300c\u5f15\u6587\uff0b\u6aa2\u7d22\u300d\u7684\u8996\u89ba\u6587\u4ef6\u7406\u89e3"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/07\/pasted-e8d784ef76f8.jpg\" alt=\"Repository image for Ryenhails\/quote-and-retrieve\"><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u8996\u89ba\u8a9e\u8a00\u6a21\u578b (VLM) \u8b80\u5b8c\u6587\u4ef6\u56de\u7b54\u554f\u984c\u6642\uff0c\u50b3\u7d71\u505a\u6cd5\u6703\u8981\u6c42\u5b83\u540c\u6642\u8f38\u51fa bounding box \u6a19\u8a3b\u8b49\u64da\u4f4d\u7f6e\uff1b\u7136\u800c\u958b\u6e90\u6a21\u578b\u7d93\u5e38\u300c\u7b54\u5c0d\u4f46\u6307\u932f\u300d\uff0c\u9019\u73fe\u8c61\u88ab\u7a31\u70ba Attribution Hallucination\u3002\u9019\u500b\u4f86\u81ea\u8d6b\u723e\u8f9b\u57fa Aalto \u5927\u5b78\u7684\u958b\u6e90\u9805\u76ee\u4e3b\u5f35\uff0c\u554f\u984c\u7665\u7d50\u5f88\u5927\u7a0b\u5ea6\u662f\u8f38\u51fa\u4ecb\u9762\u672c\u8eab\u7684\u9650\u5236\uff0c\u800c\u975e\u6a21\u578b\u7f3a\u4e4f grounding \u80fd\u529b\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u4ed6\u5011\u8a2d\u8a08\u4e86\u4e00\u7d44\u63a7\u5236\u8b8a\u91cf\u5be6\u9a57\uff1a\u56fa\u5b9a\u6a21\u578b\u3001\u8f38\u5165\u9801\u9762\u3001\u984c\u76ee\u8207\u8a55\u5206\u6a5f\u5236\uff0c\u50c5\u6539\u8b8a\u8b49\u64da\u8868\u9054\u65b9\u5f0f\u3002\u5ea7\u6a19\u4ecb\u9762\u8981\u6c42\u6a21\u578b\u9010 token \u751f\u6210\u6846\uff1b\u8a9e\u8a00\u4ecb\u9762\u5247\u7531\u6a21\u578b\u8f38\u51fa\u9010\u5b57\u5f15\u6587\uff0c\u518d\u4ea4\u7531\u5916\u90e8 layout parser \u8207 retriever \u627e\u56de\u5c0d\u61c9\u5340\u57df\uff08\u8868\u683c\u8207\u5716\u7247\u900f\u904e caption \u6216\u8a3b\u89e3\u5f15\u7528\uff09\u3002\u5728\u56db\u500b\u5bb6\u65cf\u5171\u516d\u6b3e\u958b\u6e90 VLM \u4e0a\uff0cevidence recall \u7531\u5ea7\u6a19\u4ecb\u9762\u4e0b\u6700\u9ad8 8 \u5206\u8e8d\u5347\u81f3\u8a9e\u8a00\u4ecb\u9762\u7684 26 \u81f3 47 \u5206\uff0challucination rate \u8fd1\u4e4e\u780d\u534a\uff0c\u800c\u7b54\u984c\u54c1\u8cea\u5e7e\u4e4e\u4e0d\u8b8a\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u57fa\u65bc\u9019\u500b\u5c0d\u6bd4\uff0c\u4f5c\u8005\u9032\u4e00\u6b65\u63d0\u51fa region-label-free GRPO \u8a13\u7df4\u65b9\u6cd5\uff1a\u734e\u52f5\u8a0a\u865f\u53ea\u8b80\u53d6 gold answer \u8207\u6aa2\u7d22\u5340\u57df\u7684\u88c1\u5207\u5716\uff0c\u6bcb\u9808\u4efb\u4f55 region-level \u6a19\u8a3b\u30028B \u6a21\u578b\u5728\u9019\u500b\u6846\u67b6\u4e0b\uff0cstrict attributed accuracy \u7531 22.4 \u63d0\u5347\u81f3 33.8\uff0c\u7b49\u65bc\u7528\u66f4\u4fbf\u5b9c\u7684\u76e3\u7763\u63db\u5230\u660e\u986f\u7684\u6b78\u56e0\u6539\u5584\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c0d\u65bc\u8655\u7406\u9577\u6587\u4ef6\u3001\u591a\u6a21\u614b\u8868\u683c\u6216\u9700\u8981\u53ef\u8ffd\u6eaf\u8b49\u64da\u93c8\u7684\u5718\u968a\uff08\u4f8b\u5982\u6587\u4ef6\u5be9\u6838\u3001\u7814\u7a76\u6587\u737b\u56de\u9867\uff09\u4f86\u8aaa\uff0c\u9019\u689d quote-and-retrieve \u8def\u7dda\u65e2\u7e5e\u958b\u6602\u8cb4\u7684\u4eba\u5de5\u6846\u6a19\u8a3b\uff0c\u4e5f\u907f\u958b\u6a21\u578b\u4e0d\u64c5\u9577\u7684\u5ea7\u6a19\u751f\u6210\u3002\u60f3\u8981\u5feb\u901f\u611f\u53d7\u5dee\u7570\uff0c\u53ea\u9700 clone \u9805\u76ee\u5f8c\u57f7\u884c <code>python src\/score_citevqa.py examples\/sample_predictions.jsonl<\/code> \u8207\u540c\u4e00\u6307\u4ee4\u5c0d\u5ea7\u6a19\u7248\u672c\u8dd1\u4e00\u6b21\uff1b\u9644\u5e36\u7684\u7bc4\u4f8b\u6a94\u53ea\u6709\u56db\u984c\u5408\u6210\u984c\uff0c\u537b\u5df2\u91cd\u73fe\u8ad6\u6587\u6838\u5fc3\u767c\u73fe\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>\u91cd\u9ede\u6458\u8981\uff1a<\/strong><br \/>\n&#8211; \u554f\u984c\u91cd\u65b0\u5b9a\u4f4d\uff1a\u6b78\u56e0\u5e7b\u89ba\u591a\u5c6c\u8f38\u51fa\u4ecb\u9762\u7522\u7269\uff0c\u4e26\u975e\u6a21\u578b\u80fd\u529b\u4e0d\u8db3<br \/>\n&#8211; \u4ecb\u9762\u5207\u63db\uff1a\u5ea7\u6a19\u4ecb\u9762\u63db\u6210\u5f15\u6587\uff0b\u5916\u90e8 retriever\uff0cevidence recall \u7531 \u22648 \u5347\u81f3 26\u201347<br \/>\n&#8211; \u8a13\u7df4\u66ff\u4ee3\u65b9\u6848\uff1aregion-label-free GRPO \u7528 gold answer\uff0b\u88c1\u5207\u5716\u4f5c\u734e\u52f5\uff0c\u6bcb\u9808\u6602\u8cb4\u6846\u6a19\u8a3b<br \/>\n&#8211; \u898f\u6a21\u6548\u76ca\uff1a8B \u6a21\u578b strict attributed accuracy \u7531 22.4 \u63d0\u5347\u81f3 33.8<br \/>\n&#8211; \u5feb\u901f\u9a57\u8b49\uff1a\u9644\u7bc4\u4f8b JSONL\uff0c\u7121\u9700 GPU \u5373\u53ef\u91cd\u73fe\u5c0d\u6bd4\u7d50\u679c<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><a href=\"https:\/\/github.com\/Ryenhails\/quote-and-retrieve\" rel=\"noopener noreferrer\"><strong>GitHub<\/strong><\/a> \u00b7 <a href=\"https:\/\/arxiv.org\/pdf\/2607.24651\" rel=\"noopener noreferrer\"><strong>Paper<\/strong><\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4f60\u53ef\u4ee5\u628a\u5b83\u7406\u89e3\u6210\uff1a\u7576 VLM \u7b54\u5c0d\u554f\u984c\u537b\u6307\u932f\u4f4d\u7f6e\uff0c\u554f\u984c\u672a\u5fc5\u51fa\u5728\u6a21\u578b\uff0c\u800c\u662f\u51fa\u5728\u8981\u6c42\u5b83\u8f38\u51fa\u5ea7\u6a19\u9019\u7a2e\u4ecb\u9762\u8a2d\u8a08\u3002\u63db\u6210\u300c\u5f15\u6587\uff0b\u6aa2\u7d22\u300d\u5f8c\uff0c\u6b78\u56e0\u53ec\u56de\u7387\u7ffb\u500d\uff0c\u5e7b\u89ba\u7387\u8fd1\u4e4e\u6e1b\u534a\u3002<\/p>\n","protected":false},"author":8,"featured_media":10530,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,185,114,152,119,76,149],"tags":[],"class_list":["post-10531","post","type-post","status-publish","format-standard","hentry","category-133","category-qwen","category-clone","category-python","category-119","category-76","category-149"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10531","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/users\/8"}],"replies":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/comments?post=10531"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10531\/revisions"}],"predecessor-version":[{"id":10537,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/10531\/revisions\/10537"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/10530"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=10531"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=10531"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=10531"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}