{"id":9157,"date":"2026-06-16T05:38:39","date_gmt":"2026-06-15T21:38:39","guid":{"rendered":"https:\/\/infernews.com\/blog\/medmisbench-llm\/"},"modified":"2026-06-16T05:41:38","modified_gmt":"2026-06-15T21:41:38","slug":"medmisbench-llm","status":"publish","type":"post","link":"https:\/\/infernews.com\/blog\/medmisbench-llm\/","title":{"rendered":"[\u6280\u8853\u6587\u7ae0] \u725b\u6d25\u5718\u968a\u7528 MedMisBench \u6e2c\u8a66\u91ab\u7642 LLM \u6297\u8aa4\u5c0e\u529b"},"content":{"rendered":"\n<figure class=\"wp-block-image size-large\"><img decoding=\"async\" src=\"https:\/\/infernews.com\/blog\/wp-content\/uploads\/2026\/06\/pasted-5eeb9ad16d40.jpg\" alt=\"Hero image preview\"\/><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">\u9019\u500b\u9805\u76ee\u7531 University of Oxford\u3001University of Washington\u3001University College London \u53ca University of Waterloo \u7684\u7814\u7a76\u4eba\u54e1\u5408\u4f5c\u63d0\u51fa\uff0c\u91cd\u9ede\u653e\u5728\u91ab\u7642 Large language models\uff08LLMs\uff09\u9047\u4e0a\u8aa4\u5c0e\u5167\u5bb9\u6642\uff0c\u80fd\u5426\u5b88\u4f4f\u539f\u672c\u6b63\u78ba\u7684\u91ab\u7642\u5224\u65b7\u3002\u4f5c\u8005\u6307\u51fa\uff0c\u73fe\u6709\u505a\u6cd5\u591a\u6578\u7528\u4e7e\u6de8\u3001\u8003\u8a66\u5f0f\u984c\u76ee\u8a55\u4f30\u6a21\u578b\u77e5\u8b58\uff0c\u4f46\u9019\u7a2e\u7bc4\u5f0f\u53ea\u91cf\u5230\u6a21\u578b\u300c\u8b58\u5514\u8b58\u300d\uff0c\u672a\u6709\u91cf\u5230\u6a21\u578b\u5728\u6df7\u96dc\u8cc7\u8a0a\u74b0\u5883\u4e2d\u6703\u5426\u88ab\u5e36\u504f\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u56e0\u6b64\uff0c\u5718\u968a\u63d0\u51fa MedMisBench\uff0c\u5c07\u300cepistemic resilience\u300d\u5b9a\u7fa9\u70ba\u6a21\u578b\u5728 adversarial context \u4e4b\u4e0b\u4ecd\u7dad\u6301\u6b63\u78ba\u5224\u65b7\u7684\u80fd\u529b\u3002\u9019\u500b benchmark \u6536\u9304 10,932 \u689d\u91ab\u7642\u984c\u76ee\u9805\u76ee\uff0c\u4ee5\u53ca 48,889 \u7d44 misleading context-option pairs\uff0c\u6db5\u84cb medical reasoning\u3001agentic capability \u548c patient-journey evaluation\uff0c\u7528\u4f86\u6e2c\u8a66\u6a21\u578b\u9762\u5c0d\u770b\u4f3c\u5408\u7406\u4f46\u5176\u5be6\u932f\u8aa4\u7684\u4e0a\u4e0b\u6587\u6642\u6703\u9ede\u6a23\u6539\u7b54\u6848\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u8ad6\u6587\u6700\u95dc\u9375\u7684\u8a0a\u606f\uff0c\u662f\u4e0d\u5c11\u6a21\u578b\u672c\u4f86\u7b54\u5571\uff0c\u4f46\u52a0\u5165\u805a\u7126\u5f0f\u8aa4\u5c0e\u53e5\u5b50\u5f8c\u4fbf\u653e\u68c4\u6b63\u78ba\u7b54\u6848\u300211 \u500b model configurations \u7684\u5e73\u5747\u6e96\u78ba\u7387\uff0c\u7531\u539f\u984c\u7684 71.1% \u8dcc\u5230 38.0%\uff0cattack success \u9054 51.5%\uff1b\u5176\u4e2d authority-framed falsehoods \u7684\u653b\u64ca\u6210\u529f\u7387\u6709 69.5%\uff0cexception-poisoning claims \u4e5f\u6709 64.1%\uff0c\u986f\u793a\u5e36\u6709\u6b0a\u5a01\u8a9e\u6c23\u6216\u898f\u5247\u4f8b\u5916\u5305\u88dd\u7684\u932f\u8aa4\u8cc7\u8a0a\u5c24\u5176\u5371\u96aa\u3002<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">\u5c0d\u60f3\u4f7f\u7528\u91ab\u7642 AI \u7684\u8b80\u8005\u4f86\u8aaa\uff0c\u9019\u500b\u9805\u76ee\u7684\u50f9\u503c\u4e0d\u5728\u65bc\u63d0\u4f9b\u65b0\u8a3a\u65b7\u6a21\u578b\uff0c\u800c\u662f\u88dc\u4e0a\u73fe\u6709\u91ab\u7642\u8a55\u6e2c\u7684\u76f2\u9ede\uff1a\u9ad8\u5206\u91ab\u5b78\u8003\u8a66 benchmark\uff0c\u672a\u5fc5\u4ee3\u8868\u6a21\u578b\u5728\u771f\u5be6\u5065\u5eb7\u8cc7\u8a0a\u74b0\u5883\u4e2d\u4ecd\u53ef\u9760\u3002\u7814\u7a76\u9084\u627e\u4f86\u4f86\u81ea 7 \u500b\u570b\u5bb6\u7684 14 \u4eba\u81e8\u5e8a\u5c0f\u7d44\u8986\u6838\uff0c\u8a8d\u70ba 38.2% \u6aa2\u8996\u500b\u6848\u5b58\u5728\u56b4\u91cd\u6f5b\u5728\u50b7\u5bb3\uff0c\u9019\u4ee4 MedMisBench \u5f88\u9069\u5408\u7528\u4f5c\u91ab\u7642\u6a21\u578b\u5b89\u5168\u6e2c\u8a66\u3001\u7d05\u968a\u6aa2\u9a57\uff0c\u4ee5\u53ca\u90e8\u7f72\u524d\u98a8\u96aa\u7be9\u67e5\u3002<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u6838\u5fc3\u6279\u8a55\uff1a\u73fe\u6709 benchmark \u591a\u6e2c\u77e5\u8b58\u6b63\u78ba\u7387\uff0c\u8f03\u5c11\u6e2c misleading context \u4e0b\u662f\u5426\u4ecd\u80fd\u5b88\u4f4f\u6b63\u78ba\u5224\u65b7<\/li>\n\n\n\n<li>\u65b0\u589e framing\uff1a\u7528 epistemic resilience \u5c08\u9580\u91cf\u5ea6\u6a21\u578b\u6297\u8aa4\u5c0e\u80fd\u529b<\/li>\n\n\n\n<li>\u6578\u64da\u898f\u6a21\uff1a10,932 \u500b\u91ab\u7642\u984c\u76ee\u9805\u76ee\u300148,889 \u7d44 misleading context-option pairs<\/li>\n\n\n\n<li>\u4e3b\u8981\u7d50\u679c\uff1a\u5e73\u5747\u6e96\u78ba\u7387\u7531 71.1% \u8dcc\u81f3 38.0%\uff0cattack success \u70ba 51.5%<\/li>\n\n\n\n<li>\u5f15\u7528\u6a21\u578b\u5305\u62ec ChatGPT\u3001Gemini \u7b49\u91ab\u7642\u6587\u5b57\u7406\u89e3\u8207\u751f\u6210\u80fd\u529b\u8f03\u5f37\u7684 LLMs<\/li>\n<\/ul>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Paper\uff1a<\/strong> <a href=\"https:\/\/arxiv.org\/pdf\/2606.12291\" rel=\"noopener noreferrer\">https:\/\/arxiv.org\/pdf\/2606.12291<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u725b\u6d25\u7b49\u5718\u968a\u63d0\u51fa MedMisBench\uff0c\u5c08\u9580\u91cf\u5ea6\u91ab\u7642 LLM \u5728\u8aa4\u5c0e\u8cc7\u8a0a\u4e0b\u6703\u5426\u5931\u5b88\u3002\u7d50\u679c\u986f\u793a\uff0c\u9ad8\u5206\u8003\u8a66\u8868\u73fe\u4e0d\u7b49\u65bc\u5b89\u5168\u5224\u65b7\u3002<\/p>\n","protected":false},"author":8,"featured_media":9156,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ai_generated_summary":"","footnotes":""},"categories":[140,153,116,144,119,170,199],"tags":[],"class_list":["post-9157","post","type-post","status-publish","format-standard","hentry","category-gemini","category-openai","category-agentic","category-medical","category-119","category-170","category-dataset-"],"_links":{"self":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9157","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=9157"}],"version-history":[{"count":1,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9157\/revisions"}],"predecessor-version":[{"id":9159,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/posts\/9157\/revisions\/9159"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media\/9156"}],"wp:attachment":[{"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/media?parent=9157"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/categories?post=9157"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/infernews.com\/blog\/wp-json\/wp\/v2\/tags?post=9157"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}