{"id":9145,"date":"2026-06-16T04:10:04","date_gmt":"2026-06-15T20:10:04","guid":{"rendered":"https:\/\/infernews.com\/blog\/a-benchmark-for-diagnosing-stage-wise-hallucinations-in-medical-mllm-reasoning\/"},"modified":"2026-06-16T04:10:50","modified_gmt":"2026-06-15T20:10:50","slug":"a-benchmark-for-diagnosing-stage-wise-hallucinations-in-medical-mllm-reasoning","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/a-benchmark-for-diagnosing-stage-wise-hallucinations-in-medical-mllm-reasoning\/","title":{"rendered":"ClinHallu \u62c6\u89e3\u91ab\u7642 MLLM \u5e7b\u89ba\u4f86\u6e90"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/pasted-a6678f1e30fb.jpg\" alt=\"ClinHallu logo\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">ClinHallu \u662f\u4e00\u500b<strong>\u91ab\u7642\u591a\u6a21\u614b\u5927\u6a21\u578b benchmark \u8207\u8a55\u6e2c\u5de5\u5177<\/strong>\uff0c\u76ee\u6a19\u4e0d\u662f\u53ea\u8a08\u7b97\u7b54\u5c0d\u7387\uff0c\u800c\u662f\u627e\u51fa Medical MLLM \u5728\u63a8\u7406\u9014\u4e2d\u54ea\u4e00\u6bb5\u958b\u59cb\u51fa\u73fe hallucination\u3002\u73fe\u6709\u91ab\u7642 hallucination benchmark \u591a\u6578\u504f\u91cd\u8cc7\u6599\u6536\u96c6\u8207\u6700\u7d42\u8f38\u51fa\u6aa2\u67e5\uff0c\u4f5c\u8005\u8a8d\u70ba\u9019\u7a2e\u7bc4\u5f0f\u96e3\u4ee5\u5206\u8fa8\u932f\u8aa4\u7a76\u7adf\u4f86\u81ea\u770b\u932f\u5f71\u50cf\u3001\u8a18\u932f\u91ab\u5b78\u77e5\u8b58\uff0c\u9084\u662f\u628a\u8cc7\u8a0a\u4e32\u9023\u6642\u63a8\u932f\uff0c\u56e0\u6b64\u628a\u554f\u984c\u91cd\u7d44\u6210 stage-wise \u8a3a\u65b7\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u9805\u76ee\u7684\u6838\u5fc3\u505a\u6cd5\uff0c\u662f\u628a\u6bcf\u7b46\u6a23\u672c\u7684 reasoning trace \u62c6\u6210 Visual Recognition\u3001Knowledge Recall\u3001Reasoning Integration \u4e09\u6bb5\uff0c\u518d\u914d\u5408 structured CoT annotations \u8207 stage-replacement interventions \u89c0\u5bdf\uff1a\u5982\u679c\u53ea\u4fee\u6b63\u5176\u4e2d\u4e00\u6bb5\uff0c\u6700\u7d42\u7b54\u6848\u6703\u5426\u6539\u5584\u3002\u9019\u7a2e\u8a2d\u8a08\u6bd4\u55ae\u770b\u7b54\u6848\u66f4\u6709\u5206\u6790\u529b\uff0c\u56e0\u70ba\u5b83\u76f4\u63a5\u5c0d\u61c9\u4e09\u7a2e\u5e38\u898b\u932f\u8aa4\u4f86\u6e90\uff1avisual hallucination\u3001knowledge hallucination\u3001reasoning hallucination\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">ClinHallu \u542b\u6709 7,031 \u500b\u9a57\u8b49\u904e\u7684 instances\uff0c\u4e26\u63d0\u4f9b\u8a55\u6e2c\u6d41\u7a0b\u3002\u60f3\u8a66\u9019\u500b\u9805\u76ee\u7684\u4eba\uff0c\u8f03\u5408\u7406\u7684\u8def\u7dda\u662f\u5148\u9078\u5b9a <code>datasets<\/code> \u5167\u8981\u8dd1\u7684\u8cc7\u6599\u96c6\uff0c\u518d\u5c0d\u7167 <code>models.gold cot<\/code>\u3001<code>models.model cot<\/code>\u3001<code>models.judge<\/code> \u7684\u8a2d\u5b9a\uff0c\u4e4b\u5f8c\u67e5\u770b <code>results<\/code> \u5167\u7684 model cot \u8207 eval \u8f38\u51fa\uff1b\u82e5\u53ea\u60f3\u751f\u6210\u66ff\u63db\u5f8c\u7684 jsonl\uff0c\u4e5f\u53ef\u7528 generate \u6d41\u7a0b\u3002\u9019\u8868\u793a\u5b83\u8f03\u9069\u5408\u7814\u7a76\u3001\u6a21\u578b\u6bd4\u8f03\u8207\u932f\u8aa4\u5206\u6790\uff0c\u4e0d\u662f\u9762\u5411\u4e00\u822c\u7528\u5bb6\u7684\u91ab\u7642\u554f\u7b54\u7522\u54c1\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u57fa\u6e96\u7d50\u679c\u4e5f\u6709\u53c3\u8003\u50f9\u503c\u3002\u516c\u958b\u8868\u683c\u986f\u793a\uff0cGemini-3-Flash \u5728\u6574\u9ad4\u6e96\u78ba\u7387\u8207\u4e09\u985e hallucination rate \u4e0a\u90fd\u5c6c\u524d\u5217\uff0cAvg Acc \u70ba 80.1\uff0c\u800c Qwen3-VL-Plus\u3001Qwen3.5-9B\u3001Qwen3-VL-32B \u7b49\u6a21\u578b\u4ea6\u6709\u5217\u5165\u6bd4\u8f03\u3002\u9019\u4e9b\u6578\u5b57\u7684\u91cd\u9ede\u4e0d\u662f\u5206\u9ad8\u4e0b\uff0c\u800c\u662f\u63d0\u9192\u4f60\uff1a\u540c\u4e00\u6a21\u578b\u53ef\u80fd\u7b54\u6848\u4e0d\u5dee\uff0c\u4f46\u5728\u67d0\u500b\u968e\u6bb5\u7684 hallucination rate \u4ecd\u7136\u504f\u9ad8\uff0c\u4e4b\u5f8c\u5fae\u8abf\u6216\u52a0\u9632\u8b77\u6642\u5c31\u6709\u66f4\u6e05\u6670\u65b9\u5411\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u4e0d\u662f\u53ea\u770b\u7b54\u5c0d\u7387<\/strong>\uff1a\u5b83\u6703\u62c6\u89e3\u6a21\u578b\u5728\u54ea\u500b\u63a8\u7406\u968e\u6bb5\u51fa\u932f\u3002<\/li>\n\n\n\n<li><strong>\u65b9\u6cd5\u6709\u8fa8\u8b58\u5ea6<\/strong>\uff1a\u7528 structured CoT \u8207 stage-replacement interventions \u505a\u7d30\u7c92\u5ea6\u8a3a\u65b7\u3002<\/li>\n\n\n\n<li><strong>\u5834\u666f\u5f88\u660e\u78ba<\/strong>\uff1a\u9069\u5408\u91ab\u7642 AI \u7814\u7a76\u3001\u6a21\u578b\u8a55\u6e2c\u3001trace-supervised fine-tuning \u524d\u5f8c\u6bd4\u8f03\u3002<\/li>\n\n\n\n<li><strong>\u76f8\u95dc\u6a21\u578b\u5b8c\u6574<\/strong>\uff1a\u7d50\u679c\u6db5\u84cb Qwen\u3001Gemini\u3001InternVL\u3001MedGemma\u3001Lingshu \u7b49\u7cfb\u5217\u3002<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\">\u5982\u679c\u4f60\u5728\u627e\u7684\u662f\u53ef\u76f4\u63a5\u90e8\u7f72\u7684\u91ab\u7642\u52a9\u624b\uff0cClinHallu \u4e26\u4e0d\u5c6c\u65bc\u90a3\u4e00\u985e\uff1b\u5b83\u66f4\u50cf\u4e00\u628a\u91cf\u5c3a\uff0c\u5c08\u9580\u6aa2\u67e5\u6a21\u578b\u63a8\u7406\u93c8\u54ea\u88e1\u958b\u59cb\u5931\u771f\u3002\u5c0d\u7814\u7a76\u5718\u968a\u4f86\u8aaa\uff0c\u9019\u6bd4\u53ea\u77e5\u9053\u300c\u6a21\u578b\u6709\u5e7b\u89ba\u300d\u66f4\u6709\u7528\uff0c\u56e0\u70ba\u5f8c\u7e8c\u53ef\u4ee5\u6309 Visual Recognition\u3001Knowledge Recall\u3001Reasoning Integration \u5206\u6bb5\u4fee\u6b63\uff0c\u9023 trace-supervised fine-tuning \u662f\u5426\u6709\u6548\u90fd\u8f03\u5bb9\u6613\u9a57\u8b49\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>GitHub\uff1a<\/strong> <a href=\"https:\/\/github.com\/alibaba-damo-academy\/ClinHallu\" rel=\"noopener noreferrer\">https:\/\/github.com\/alibaba-damo-academy\/ClinHallu<\/a><\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2606.14697\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2606.14697<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>ClinHallu\u4e0d\u53ea\u770b\u6a21\u578b\u7b54\u932f\u6c92\u6709\uff0c\u9084\u8ffd\u67e5\u932f\u5728\u54ea\u4e00\u500b\u63a8\u7406\u968e\u6bb5\u3002\u5c0d\u91ab\u7642\u591a\u6a21\u614b\u6a21\u578b\u8a55\u6e2c\u4f86\u8aaa\uff0c\u9019\u7a2e\u62c6\u89e3\u65b9\u6cd5\u76f8\u7576\u5be6\u7528\u3002<\/p>\n","protected":false},"author":8,"featured_media":9144,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[133,171,185,165,140,144,119,196,199],"tags":[],"class_list":["post-9145","post","type-post","status-publish","format-standard","hentry","category-133","category-171","category-qwen","category-165","category-gemini","category-medical","category-119","category-196","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9145","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=9145"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9145\/revisions"}],"predecessor-version":[{"id":9147,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9145\/revisions\/9147"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9144"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9145"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9145"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9145"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}