	{"id":1190337,"date":"2026-05-11T17:17:41","date_gmt":"2026-05-11T16:17:41","guid":{"rendered":"https:\/\/www.artefact.com\/?post_type=blog&#038;p=1190337"},"modified":"2026-05-11T17:32:03","modified_gmt":"2026-05-11T16:32:03","slug":"detecting-hallucinations-in-llms-one-token-at-a-time","status":"publish","type":"blog","link":"https:\/\/www.artefact.com\/zh\/blog\/detecting-hallucinations-in-llms-one-token-at-a-time\/","title":{"rendered":"\u4e00\u6b21\u4e00\u4e2a\u4ee4\u724c\u5730\u68c0\u6d4b LLM \u4e2d\u7684\u5e7b\u89c9"},"content":{"rendered":"<p>\u6211\u4eec\u7684 Python \u8f6f\u4ef6\u5305--artefactual--\u5982\u4f55\u901a\u8fc7\u57fa\u4e8e\u71b5\u7684\u8bc4\u5206\u6765\u544a\u8bc9\u4f60\uff0c\u4f60\u7684\u6a21\u578b\u4f55\u65f6\u4ee5\u53ca\u5728\u4f55\u5904\u5728\u80e1\u7f16\u4e71\u9020\u3002.<\/p>\n<p>\u6ce8\uff1a\u672c\u6587\u662f\u4ee5\u4e0b\u6587\u7ae0\u7684\u540e\u7eed <a href=\"https:\/\/medium.com\/ardian-data-science\/is-your-ai-lying-to-you-04e59ac61fff?postPublishedType=initial\" target=\"_blank\" rel=\"noopener\">\u6765\u81ea Ardian \u670b\u53cb\u7684\u6587\u7ae0<\/a>, \u8be6\u7ec6\u4ecb\u7ecd\u4e86\u8d1f\u8d23\u4efb\u7684\u4eba\u5de5\u667a\u80fd\u5bf9\u91d1\u878d\u673a\u6784\u7684\u91cd\u8981\u6027\u3002\u8bf7\u52a1\u5fc5\u67e5\u770b\uff01<\/p>\n<h2>\u5e7b\u89c9\u95ee\u9898<\/h2>\n<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u80fd\u529b\u4ee4\u4eba\u60ca\u53f9\u3002\u5b83\u4eec\u603b\u7ed3\u3001\u7ffb\u8bd1\u3001\u63a8\u7406\u548c\u7f16\u7801\uff08\u6bd4\u6211\u5f3a\uff09\u3002\u4f46\u4e0e\u6211\u4e0d\u540c\u7684\u662f\uff0c\u5b83\u4eec\u4e5f\u56e0\u4ee5\u4ee4\u4eba\u4e0d\u5b89\u7684\u81ea\u4fe1\u7f16\u9020\u4e8b\u5b9e\u800c\u58f0\u540d\u72fc\u85c9\u3002.<\/p>\n<p>\u5728\u81ea\u7136\u8bed\u8a00\u5904\u7406\uff08NLP\uff09\u6587\u732e\u4e2d\uff0c\u5e7b\u89c9\u662f\u6307\u4efb\u4f55\u7531\u6a21\u578b\u751f\u6210\u7684\u5185\u5bb9\uff0c\u8fd9\u4e9b\u5185\u5bb9\u4e0e\u4e8b\u5b9e\u4e0d\u7b26\u3001\u6beb\u65e0\u9053\u7406\u6216\u4e0d\u5fe0\u4e8e\u6240\u63d0\u4f9b\u7684\u6765\u6e90\uff0c\u4f46\u770b\u8d77\u6765\u5374\u5341\u5206\u53ef\u4fe1\u3002\u5176\u540e\u679c\u4ece\u8f7b\u5fae\uff08\u4e00\u4e2a\u9519\u8bef\u7684\u7410\u4e8b\u7b54\u6848\uff09\u5230\u4e25\u91cd\uff08\u4e00\u4e2a\u634f\u9020\u7684\u6cd5\u5f8b\u5f15\u6587\u3001\u4e00\u4e2a\u9519\u8bef\u7684\u836f\u7269\u5242\u91cf\uff09\u4e0d\u7b49\u3002\u968f\u7740\u4f01\u4e1a\u5c06 LLM \u96c6\u6210\u5230\u751f\u4ea7\u7cfb\u7edf\u4e2d\uff0c\u95ee\u9898\u4ece <em>\u201c\u8fd9\u4e2a\u6a21\u578b\u80fd\u751f\u6210\u6709\u7528\u7684\u6587\u672c\u5417\uff1f\u201d<\/em> \u81f3 <em>\u201c\u6211\u4eec\u80fd\u76f8\u4fe1\u5b83\u521a\u624d\u8bf4\u7684\u8bdd\u5417\uff1f\u201d<\/em><\/p>\n<p>\u8bf7\u770b\u4e00\u4e2a\u5177\u4f53\u7684\u4f8b\u5b50\u3002\u60a8\u5728\u4e00\u5bb6\u91d1\u878d\u673a\u6784\u5de5\u4f5c\uff0c\u60a8\u5411\u5f53\u5730\u7684\u6cd5\u5f8b\u7855\u58eb\u8be2\u95ee\uff1a\"\u60a8\u5728\u54ea\u91cc\u5de5\u4f5c\uff1f<\/p>\n<blockquote><p>\u201c\u827e\u9ed8\u751f\u7535\u6c14 2023 \u5e74\u7684\u51c0\u6536\u5165\u662f\u591a\u5c11\uff1f\u201d<\/p><\/blockquote>\n<p><img decoding=\"async\" class=\"lazyload aligncenter wp-image-1190466\" src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-1.png\" data-orig-src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-1.png\" alt=\"\" width=\"700\" height=\"185\" srcset=\"data:image\/svg+xml,%3Csvg%20xmlns%3D%27http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%27%20width%3D%27700%27%20height%3D%27185%27%20viewBox%3D%270%200%20700%20185%27%3E%3Crect%20width%3D%27700%27%20height%3D%27185%27%20fill-opacity%3D%220%22%2F%3E%3C%2Fsvg%3E\" data-srcset=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-1-200x53.png 200w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-1-300x79.png 300w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-1-400x106.png 400w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-1-600x158.png 600w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-1-768x203.png 768w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-1-800x211.png 800w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-1-1024x270.png 1024w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-1.png 1045w\" data-sizes=\"auto\" data-orig-sizes=\"(max-width: 700px) 100vw, 700px\" \/><\/p>\n<p>\u6a21\u7279\u56de\u7b54\u8bf4 <em>\u201c\u827e\u9ed8\u751f\u7535\u6c14\u516c\u5e03\u7684 2023 \u8d22\u5e74\u51c0\u8425\u6536\u7ea6\u4e3a $152 \u4ebf\u7f8e\u5143\u201d\u3002\u201d<\/em> \u542c\u8d77\u6765\u5f88\u6709\u9053\u7406\u3002\u4f46\u8fd9\u6837\u505a\u5bf9\u5417\uff1f\u4f60\u6ca1\u6709\u6253\u5f00\u5e74\u5ea6\u62a5\u544a\u3002\u4f60\u6ca1\u6709\u57fa\u672c\u4e8b\u5b9e\u53ef\u4ee5\u6bd4\u8f83\u3002\u4f60\u53ea\u6709\u6a21\u578b\u7684\u8f93\u51fa--\u548c\u6000\u7591\u3002.<\/p>\n<p>\u8fd9\u5c31\u662f\u6211\u4eec\u7684\u5de5\u4f5c\u73af\u5883\u3002\u6ca1\u6709\u795e\u8c15\u3002\u63a8\u7406\u65f6\u6ca1\u6709\u53c2\u8003\u7b54\u6848\u3002\u53ea\u6709\u4e00\u4e2a LLM \u54cd\u5e94\u548c\u5b83\u5728\u751f\u6210\u54cd\u5e94\u65f6\u4ea7\u751f\u7684\u5143 data\u3002\u6211\u4eec\u7684\u76ee\u6807\u662f\uff1a\u901a\u8fc7\u5355\u6b21\u751f\u6210\uff0c\u91cf\u5316\u8f93\u51fa\u7ed3\u679c\u4ea7\u751f\u5e7b\u89c9\u7684\u53ef\u80fd\u6027\u3002.<\/p>\n<h2>\u68c0\u6d4b\u5e7b\u89c9\uff1a\u6bd4\u542c\u8d77\u6765\u66f4\u96be<\/h2>\n<h3>\u86ee\u529b\u65b9\u6cd5<\/h3>\n<p>\u4e00\u4e2a\u5f88\u81ea\u7136\u7684\u60f3\u6cd5\u662f\u591a\u6b21\u5411\u6a21\u578b\u63d0\u51fa\u540c\u6837\u7684\u95ee\u9898\uff0c\u5e76\u68c0\u67e5\u7b54\u6848\u662f\u5426\u4e00\u81f4\u3002\u5982\u679c\u516d\u6b21\u8fd0\u884c\u4e2d\u6709\u4e94\u6b21\u8bf4 \u201c$152 \u4ebf\u201d\uff0c\u4e00\u6b21\u8bf4 \u201c$187 \u4ebf\u201d\uff0c\u90a3\u4e48\u8fd9\u79cd\u5171\u8bc6\u5c31\u4f1a\u7ed9\u4f60\u5e26\u6765\u4e00\u4e9b\u4fe1\u5fc3\u3002\u8fd9\u5c31\u662f\u50cf SelfCheckGPT \u8fd9\u6837\u7684\u65b9\u6cd5\u80cc\u540e\u7684\u539f\u7406\uff0c\u5b83\u53ef\u4ee5\u6d4b\u91cf\u591a\u4e2a\u91c7\u6837\u8f93\u51fa\u7684\u4e00\u81f4\u6027--\u4e00\u79cd \u201c\u8499\u7279\u5361\u7f57\u5f0f \u201d\u7684\u5e7b\u89c9\u68c0\u6d4b\u65b9\u6cd5\u3002.<\/p>\n<p>\u5b83\u5f88\u6709\u6548\u3002\u4f46\u5b83\u6709\u4e24\u4e2a\u660e\u663e\u7684\u7f3a\u70b9\uff1a<\/p>\n<ol>\n<li><strong>\u8d39\u7528.<\/strong> \u6bcf\u589e\u52a0\u4e00\u4ee3\u6837\u672c\uff0c\u63a8\u7406\u9884\u7b97\u5c31\u4f1a\u589e\u52a0\u4e00\u500d\u3002\u5bf9\u4e8e\u6709 10 \u4e2a\u6837\u672c\u7684 SelfCheckGPT\uff0c\u4f60\u9700\u8981\u652f\u4ed8\u5927\u7ea6 10 \u500d\u7684\u8ba1\u7b97\u8d39\u7528\uff0c\u518d\u52a0\u4e0a\u8bed\u4e49\u76f8\u4f3c\u6027\u6a21\u578b\u7684\u8d39\u7528\u3002\u5728\u5927\u89c4\u6a21\u60c5\u51b5\u4e0b\uff0c\u8fd9\u6837\u7684\u6210\u672c\u4ee4\u4eba\u671b\u800c\u5374\u6b65\u3002.<\/li>\n<li><strong>\u7c92\u5ea6\u3002.<\/strong> \u591a\u62cd\u6cd5\u662f\u5728\u5e8f\u5217\u5c42\u9762\u4e0a\u64cd\u4f5c\u7684\u3002\u5b83\u4eec\u4f1a\u544a\u8bc9\u4f60 \u201c\u8fd9\u4e2a\u7b54\u6848\u4f3c\u4e4e\u4e0d\u53ef\u9760\u201d\uff0c\u4f46\u4e0d\u4f1a\u544a\u8bc9\u4f60\u7b54\u6848\u7684\u54ea\u4e2a\u90e8\u5206\u6709\u95ee\u9898\u3002\u4e00\u4e2a\u7b54\u6848\u53ef\u80fd\u662f 90% \u51c6\u786e\uff0c\u4f46\u4e2d\u95f4\u57cb\u85cf\u4e86\u4e00\u4e2a\u5e7b\u89c9\u6570\u5b57\u3002\u4f60\u60f3\u77e5\u9053\u5728\u54ea\u91cc\u3002.<\/li>\n<\/ol>\n<p>\u8fd9\u4e9b\u5c40\u9650\u6027\u4fc3\u4f7f\u6211\u4eec\u5bfb\u627e\u4e00\u79cd\u4e0d\u540c\u7684\u4fe1\u53f7--\u4e00\u79cd\u5ec9\u4ef7\u3001\u5355\u6b21\u53d1\u5c04\u3001\u53ef\u5728\u6807\u8bb0\u7ea7\u522b\uff08\u5373 LLM \u5185\u90e8\u5904\u7406\u7684\u5355\u4e2a\u8bcd\u5757\uff09\u5de5\u4f5c\u7684\u4fe1\u53f7\u3002.<\/p>\n<h3>\u4fe1\u53f7\u5df2\u7ecf\u5b58\u5728<\/h3>\n<p>\u5f53 LLM \u751f\u6210\u6587\u672c\u65f6\uff0c\u5b83\u5e76\u4e0d\u53ea\u662f\u8f93\u51fa\u6807\u8bb0\u3002\u6bcf\u4e00\u6b65\uff0c\u5b83\u90fd\u4f1a\u8ba1\u7b97\u6574\u4e2a\u8bcd\u6c47\u8868\u7684\u6982\u7387\u5206\u5e03\uff1a <em>\u201c\u6839\u636e\u63d0\u793a\u548c\u6211\u76ee\u524d\u751f\u6210\u7684\u6240\u6709\u4fe1\u606f\u201d \"\u4e0b\u4e00\u4e2a\u6807\u8bb0\u7684\u53ef\u80fd\u6027\u6709\u591a\u5927?\"<\/em>\u00a0\u83b7\u80dc\u7684\u4ee4\u724c\u4f1a\u88ab\u91c7\u6837\u3002\u5176\u4f59\u7684\u5219\u88ab\u4e22\u5f03\u3002\u4f46\u8fd9\u4e9b\u6982\u7387\uff08\u66f4\u5177\u4f53\u5730\u8bf4\uff0c\u5b83\u4eec\u7684\u5206\u6563\u7a0b\u5ea6\uff09\u5305\u542b\u4e86\u6a21\u578b\u5185\u90e8\u7f6e\u4fe1\u5ea6\u7684\u4fe1\u606f\u3002.<\/p>\n<p>\u5982\u679c\u6a21\u578b\u975e\u5e38\u786e\u5b9a\uff0c\u5219\u5927\u90e8\u5206\u6982\u7387\u90fd\u96c6\u4e2d\u5728\u4e00\u4e2a\u6807\u8bb0\u4e0a\u3002\u5982\u679c\u6a21\u578b\u72b9\u8c6b\u4e0d\u51b3\uff0c\u6982\u7387\u5c31\u4f1a\u5206\u6563\u5230\u8bb8\u591a\u5019\u9009\u7b26\u53f7\u4e0a\u3002\u8fd9\u79cd\u5206\u6563\u6b63\u662f\u71b5\u7684\u8861\u91cf\u6807\u51c6\u3002.<\/p>\n<h3>\u71b5\uff1a\u5feb\u901f\u7ed5\u884c<\/h3>\n<p>\u71b5\u662f\u4e00\u4e2a\u4fe1\u606f\u8bba\u91cf\uff0c\u7528\u6765\u8861\u91cf\u6982\u7387\u5206\u5e03\u7684\u4e0d\u786e\u5b9a\u6027\u3002\u5176\u76f4\u89c2\u6027\u975e\u5e38\u7b80\u5355\u3002\u60f3\u8c61\u4e09\u4e2a\u76d2\u5b50\u3002\u5176\u4e2d\u4e00\u4e2a\u88c5\u6709\u4e00\u5757\u997c\u5e72\u3002\u4f60\u5fc5\u987b\u731c\u51fa\u662f\u54ea\u4e00\u4e2a\u3002.<\/p>\n<p><img decoding=\"async\" class=\"lazyload aligncenter wp-image-1190467\" src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-2.png\" data-orig-src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-2.png\" alt=\"\" width=\"800\" height=\"404\" srcset=\"data:image\/svg+xml,%3Csvg%20xmlns%3D%27http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%27%20width%3D%27800%27%20height%3D%27404%27%20viewBox%3D%270%200%20800%20404%27%3E%3Crect%20width%3D%27800%27%20height%3D%27404%27%20fill-opacity%3D%220%22%2F%3E%3C%2Fsvg%3E\" data-srcset=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-2-200x101.png 200w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-2-300x152.png 300w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-2-400x202.png 400w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-2-600x303.png 600w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-2-768x388.png 768w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-2-800x404.png 800w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-2.png 982w\" data-sizes=\"auto\" data-orig-sizes=\"(max-width: 800px) 100vw, 800px\" \/><\/p>\n<ul>\n<li><strong>\u65b9\u6848 A\uff1a<\/strong> \u4f60\u77e5\u9053\u997c\u5e72\u5728 2 \u53f7\u76d2\u5b50\u91cc\u3002\u4f60\u7684\u4e0d\u786e\u5b9a\u6027\u4e3a\u96f6\u3002\u71b5 = 0\u3002.<\/li>\n<li><strong>\u65b9\u6848 B\uff1a<\/strong> \u4f60\u6839\u672c\u4e0d\u77e5\u9053\u3002\u6bcf\u4e2a\u76d2\u5b50\u90fd\u6709 1\/3 \u7684\u673a\u4f1a\u3002\u4f60\u7684\u4e0d\u786e\u5b9a\u6027\u662f\u6700\u5927\u7684\u3002\u71b5 = log\u2082(3) \u2248 1.58 \u4f4d\u3002.<\/li>\n<\/ul>\n<p>\u73b0\u5728\uff0c\u7528\u4ee3\u8bcd\u66ff\u6362\u65b9\u6846\uff0c\u7528 \u201c\u6b63\u786e \u201d\u7684\u4e0b\u4e00\u4e2a\u8bcd\u66ff\u6362 cookie\u3002\u5728\u6bcf\u4e00\u6b65\u751f\u6210\u8fc7\u7a0b\u4e2d\uff0cLLM \u90fd\u4f1a\u9762\u4e34\u8fd9\u6837\u7684\u9009\u62e9--\u53ea\u4e0d\u8fc7\u5b83\u8981\u4ece\u8d85\u8fc7 10 \u4e07\u4e2a\u8bcd\u5e93\u4e2d\u6311\u9009\u7684\u4e0d\u662f 3 \u4e2a\u65b9\u6846\u3002\u5f53\u6a21\u578b\u4fe1\u5fc3\u5341\u8db3\u65f6\uff0c\u4e00\u4e2a\u6807\u8bb0\u5360\u4e3b\u5bfc\u5730\u4f4d\uff0c\u71b5\u503c\u8f83\u4f4e\u3002\u5f53\u5b83\u72b9\u8c6b\u4e0d\u51b3\u65f6\uff0c\u71b5\u5c31\u4f1a\u4e0a\u5347\u3002.<\/p>\n<p><img decoding=\"async\" class=\"lazyload aligncenter wp-image-1190468\" src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-3.png\" data-orig-src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-3.png\" alt=\"\" width=\"800\" height=\"434\" srcset=\"data:image\/svg+xml,%3Csvg%20xmlns%3D%27http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%27%20width%3D%27800%27%20height%3D%27434%27%20viewBox%3D%270%200%20800%20434%27%3E%3Crect%20width%3D%27800%27%20height%3D%27434%27%20fill-opacity%3D%220%22%2F%3E%3C%2Fsvg%3E\" data-srcset=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-3-200x109.png 200w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-3-300x163.png 300w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-3-400x217.png 400w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-3-600x326.png 600w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-3-768x417.png 768w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-3-800x434.png 800w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-3.png 1008w\" data-sizes=\"auto\" data-orig-sizes=\"(max-width: 800px) 100vw, 800px\" \/><\/p>\n<p style=\"text-align: center;\"><em>\u4e24\u79cd\u4e0d\u540c\u60c5\u51b5\u4e0b\u7684\u6982\u7387\u5206\u5e03\u5206\u5e03\u3002.<\/em><\/p>\n<p>\u5176\u4e2d\u7684\u5173\u952e\u4e4b\u5904\u5728\u4e8e\uff0c\u7279\u5b9a\u6807\u8bb0\u4f4d\u7f6e\u7684\u9ad8\u71b5\u4e0e\u8be5\u4f4d\u7f6e\u51fa\u9519\u7684\u53ef\u80fd\u6027\u76f8\u5173\u3002\u6a21\u578b\u901a\u8fc7\u5176\u6982\u7387\u5206\u5e03\u544a\u8bc9\u4f60\uff0c\u5b83\u4e0d\u786e\u5b9a\u63a5\u4e0b\u6765\u4f1a\u53d1\u751f\u4ec0\u4e48\u3002\u6211\u4eec\u53ea\u9700\u8981\u503e\u542c\u3002.<\/p>\n<h2>\u4ece\u71b5\u5230\u5e7b\u89c9\u5206\u6570<\/h2>\n<h3>EPR\uff1a\u71b5\u4ea7\u751f\u7387<\/h3>\n<p>\u6211\u4eec\u7684\u7b2c\u4e00\u4e2a\u6307\u6807 EPR\uff08\u71b5\u751f\u6210\u7387\uff09\u662f\u76f4\u63a5\u6307\u6807\u3002\u5bf9\u4e8e\u751f\u6210\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e2a\u6807\u8bb0\uff0c\u6211\u4eec\u8ba1\u7b97\u6a21\u578b\u9884\u6d4b\u7684\u524d K \u4e2a\u6807\u8bb0\u6982\u7387\u7684\u71b5\u3002\u7136\u540e\u53d6\u6574\u4e2a\u5e8f\u5217\u7684\u5e73\u5747\u503c\u3002\u8fd9\u6837\u5c31\u5f97\u5230\u4e86\u4e00\u4e2a\u5355\u4e00\u7684\u6570\u5b57\uff0c\u53cd\u6620\u4e86\u6a21\u578b\u5728\u6574\u4e2a\u54cd\u5e94\u8fc7\u7a0b\u4e2d\u7684\u5e73\u5747\u72b9\u8c6b\u7a0b\u5ea6\u3002.<\/p>\n<p>\u8fd9\u662f\u4e00\u4e2a\u65e0\u76d1\u7763\u6307\u6807\uff1a\u4e0d\u9700\u8981\u6807\u7b7e\u3002\u5728\u6211\u4eec\u7684\u5b9e\u9a8c\uff08\u53d1\u8868\u4e8e ECIR 2026\uff09\u4e2d\uff0c\u4ec5 EPR \u4e00\u9879\u5c31\u5728 TriviaQA \u7684\u56db\u79cd\u4e0d\u540c LLM \u4e2d\u83b7\u5f97\u4e86 74 \u5230 81 \u5206\u7684 ROC-AUC \u5206\u6570\u3002\u5bf9\u4e8e\u4e00\u4e2a\u9664\u4e86\u5355\u6b21\u751f\u6210\u4e4b\u5916\u51e0\u4e4e\u4e0d\u9700\u8981\u4efb\u4f55\u6210\u672c\u7684\u6307\u6807\u6765\u8bf4\uff0c\u8fd9\u8fd8\u7b97\u4e0d\u9519\u3002.<\/p>\n<p>\u4f46\u6211\u4eec\u53ef\u4ee5\u505a\u5f97\u66f4\u597d\u3002.<\/p>\n<h3>WEPR\uff1a\u52a0\u6743\u71b5\u4ea7\u751f\u7387<\/h3>\n<p>\u539f\u59cb\u71b5\u5bf9\u6240\u6709\u6807\u8bb0\u7684\u6392\u540d\u4e00\u89c6\u540c\u4ec1\u3002\u6392\u540d\u7b2c\u4e00\u7684\u6807\u8bb0\uff08\u6700\u6709\u53ef\u80fd\u7684\u6807\u8bb0\uff09\u548c\u6392\u540d\u7b2c\u5341\u7684\u6807\u8bb0\u7684\u71b5\u8d21\u732e\u52a0\u6743\u76f8\u540c\u3002\u5b9e\u9645\u4e0a\uff0c\u4e0d\u786e\u5b9a\u6027\u5728\u4e0d\u540c\u7b49\u7ea7\u4e4b\u95f4\u7684\u5206\u5e03\u65b9\u5f0f\u8574\u542b\u7740\u4e0d\u540c\u7684\u4fe1\u606f\u3002.<\/p>\n<p>WEPR\uff08\u52a0\u6743 EPR\uff09\u901a\u8fc7\u5b66\u4e60\u4e00\u7ec4\u6743\u91cd\u6765\u91cd\u65b0\u5e73\u8861\u8fd9\u4e9b\u8d21\u732e\u3002\u5b83\u4f7f\u7528\u4e24\u4e2a\u4fe1\u53f7\uff1a<\/p>\n<ul>\n<li>\"(\u300a\u4e16\u754c\u4eba\u6743\u5ba3\u8a00\u300b) <strong>\u541d\u556c<\/strong> \u6574\u4e2a\u5e8f\u5217\u7684\u52a0\u6743\u71b5--\u6355\u6349\u6574\u4f53\u72b9\u8c6b\u3002.<\/li>\n<li>\"(\u300a\u4e16\u754c\u4eba\u6743\u5ba3\u8a00\u300b) <strong>\u6700\u5927<\/strong> \u6bcf\u4e2a\u7b49\u7ea7\u7684\u71b5\u8d21\u732e--\u6355\u6349\u4e0d\u786e\u5b9a\u6027\u5cf0\u503c\u3002\u5373\u4f7f\u5e8f\u5217\u7684\u5176\u4f59\u90e8\u5206\u90fd\u662f\u6709\u628a\u63e1\u5730\u751f\u6210\u7684\uff0c\u5355\u4e2a\u9ad8\u5ea6\u72b9\u8c6b\u7684\u65f6\u523b\u4e5f\u53ef\u80fd\u662f\u5e7b\u89c9\u7684\u6807\u5fd7\u3002.<\/li>\n<\/ul>\n<p>\u8fd9\u4e9b\u7279\u5f81\u88ab\u9001\u5165\u903b\u8f91\u56de\u5f52\uff0c\u5e76\u5728\u6807\u6ce8\u7684 dataset \u4e0a\u8fdb\u884c\u8bad\u7ec3\u3002sigmoid \u7684\u8f93\u51fa\u4e3a\u6821\u51c6\u6982\u7387\uff1a<\/p>\n<blockquote><p>\u201c\u6b64\u56de\u590d\u6709 86% \u7684\u6982\u7387\u5305\u542b\u5e7b\u89c9\u201d<\/p><\/blockquote>\n<p><img decoding=\"async\" class=\"lazyload aligncenter wp-image-1190469\" src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-4.png\" data-orig-src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-4.png\" alt=\"\" width=\"700\" height=\"311\" srcset=\"data:image\/svg+xml,%3Csvg%20xmlns%3D%27http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%27%20width%3D%27700%27%20height%3D%27311%27%20viewBox%3D%270%200%20700%20311%27%3E%3Crect%20width%3D%27700%27%20height%3D%27311%27%20fill-opacity%3D%220%22%2F%3E%3C%2Fsvg%3E\" data-srcset=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-4-200x89.png 200w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-4-300x133.png 300w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-4-400x178.png 400w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-4-600x267.png 600w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-4-768x342.png 768w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-4-800x356.png 800w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-4.png 805w\" data-sizes=\"auto\" data-orig-sizes=\"(max-width: 700px) 100vw, 700px\" \/><\/p>\n<p>\u9664\u4e86\u5206\u7c7b\u4e4b\u5916\uff0cWEPR \u8fd8\u80fd\u751f\u6210\u6807\u8bb0\u7ea7\u8bc4\u5206\u3002\u751f\u6210\u5e8f\u5217\u4e2d\u7684\u6bcf\u4e2a\u6807\u8bb0\u90fd\u6709\u81ea\u5df1\u7684\u5e7b\u89c9\u6982\u7387\uff0c\u8ba9\u60a8\u53ef\u4ee5\u7cbe\u786e\u5730\u786e\u5b9a\u54cd\u5e94\u4e2d\u54ea\u4e9b\u90e8\u5206\u503c\u5f97\u4ed4\u7ec6\u68c0\u67e5\u3002\u8fd9\u662f\u5728\u6a21\u578b\u751f\u6210\u65f6\u9010\u4e2a\u6807\u8bb0\u5b9e\u65f6\u8ba1\u7b97\u7684\uff0c\u65e0\u9700\u7b49\u5f85\u5b8c\u6574\u8f93\u51fa\u3002.<\/p>\n<p><img decoding=\"async\" class=\"lazyload aligncenter wp-image-1190470\" src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-5.png\" data-orig-src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-5.png\" alt=\"\" width=\"700\" height=\"243\" srcset=\"data:image\/svg+xml,%3Csvg%20xmlns%3D%27http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%27%20width%3D%27700%27%20height%3D%27243%27%20viewBox%3D%270%200%20700%20243%27%3E%3Crect%20width%3D%27700%27%20height%3D%27243%27%20fill-opacity%3D%220%22%2F%3E%3C%2Fsvg%3E\" data-srcset=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-5-200x70.png 200w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-5-300x104.png 300w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-5-400x139.png 400w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-5-600x209.png 600w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-5-768x267.png 768w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-5-800x278.png 800w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-5-1024x356.png 1024w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-5.png 1053w\" data-sizes=\"auto\" data-orig-sizes=\"(max-width: 700px) 100vw, 700px\" \/><\/p>\n<h3>\u6807\u7b7e\u600e\u4e48\u529e\uff1f<\/h3>\n<p>\u6709\u76d1\u7763\u7684\u65b9\u6cd5\u9700\u8981\u6ce8\u91ca\u3002\u624b\u5de5\u6807\u6ce8\u6210\u5343\u4e0a\u4e07\u4e2a\u8d28\u91cf\u4fdd\u8bc1\u5bf9\u7684\u901f\u5ea6\u5f88\u6162\u3002\u56e0\u6b64\uff0c\u6211\u4eec\u4f7f\u7528 <strong>\u6cd5\u5b66\u7855\u58eb\u62c5\u4efb\u6cd5\u5b98<\/strong> \u65b9\u6cd5\uff1a\u7531\u4e00\u4e2a\u5355\u72ec\u7684\u6a21\u578b\u5c06\u751f\u6210\u7684\u6bcf\u4e2a\u7b54\u6848\u4e0e\u5df2\u77e5\u7684\u5730\u9762\u5b9e\u51b5\u8fdb\u884c\u6bd4\u8f83\uff0c\u5e76\u5c06\u5176\u6807\u6ce8\u4e3a\u6b63\u786e\u6216\u9519\u8bef\u3002.<\/p>\n<p>\u8fd9\u53ef\u9760\u5417\uff1f\u6211\u4eec\u901a\u8fc7\u4eba\u7c7b\u6807\u6ce8\u8005\u5bf9\u5176\u8fdb\u884c\u4e86\u9a8c\u8bc1\u3002\u7531 15 \u4f4d\u7814\u7a76\u4eba\u5458\u7ec4\u6210\u7684\u5c0f\u7ec4\u5bf9 1,300 \u591a\u5bf9\u7b54\u6848\u8fdb\u884c\u4e86\u4eba\u5de5\u6807\u6ce8\u3002\u81ea\u52a8\u5224\u65ad\u4e0e\u4eba\u7c7b\u8bc4\u4f30\u8005\u4e4b\u95f4\u7684\u4e00\u81f4\u6027\u8fbe\u5230 95.7%\uff0c\u79d1\u6069\u5361\u5e15\uff08Cohen's Kappa\uff09\u4e3a 0.90\u3002\u81ea\u52a8\u6807\u7b7e\u662f\u4eba\u7c7b\u5224\u65ad\u7684\u53ef\u9760\u66ff\u4ee3\u7269\uff0c\u5176\u7a33\u5065\u6027\u8db3\u4ee5\u7528\u6765\u8bad\u7ec3\u5e7b\u89c9\u68c0\u6d4b\u5668\u3002.<\/p>\n<h2>\u4ecb\u7ecd\u4eba\u5de5\u5236\u54c1\uff1a\u73b0\u5728\u8f6e\u5230\u4f60\u4eec\u73a9\u4e86\u3002.<\/h2>\n<p>\u6211\u4eec\u5c06\u6240\u6709\u8fd9\u4e9b\u6253\u5305\u6210\u4e00\u4e2a\u5f00\u6e90 Python \u5e93\uff1a <a href=\"https:\/\/github.com\/artefactory\/artefactual\/\" target=\"_blank\" rel=\"noopener\">\u4eba\u5de5<\/a>.<\/p>\n<p>\u8be5\u5e93\u4e3a\u591a\u4e2a\u6a21\u578b\u7cfb\u5217\uff08Mistral-Small\u3001Falcon-3\u3001Phi-4\u3001Ministral-8B\uff09\u63d0\u4f9b\u4e86\u9884\u5148\u8ba1\u7b97\u7684\u6821\u51c6\u6743\u91cd\uff0c\u56e0\u6b64\u60a8\u65e0\u9700\u8fd0\u884c\u4efb\u4f55\u8bad\u7ec3\u7ba1\u9053\u5373\u53ef\u7acb\u5373\u5f00\u59cb\u5bf9\u8f93\u51fa\u8fdb\u884c\u8bc4\u5206\u3002\u5b83\u53ef\u4ee5\u89e3\u6790\u6765\u81ea vLLM\u3001OpenAI Chat Completions API \u548c OpenAI Responses API \u7684\u8f93\u51fa\u3002.<\/p>\n<p>\u4e0b\u9762\u662f\u6700\u7b80\u5355\u7684\u7528\u6cd5\uff1a<\/p>\n<p><img decoding=\"async\" class=\"lazyload aligncenter wp-image-1190471\" src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-6.png\" data-orig-src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-6.png\" alt=\"\" width=\"700\" height=\"657\" srcset=\"data:image\/svg+xml,%3Csvg%20xmlns%3D%27http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%27%20width%3D%27700%27%20height%3D%27657%27%20viewBox%3D%270%200%20700%20657%27%3E%3Crect%20width%3D%27700%27%20height%3D%27657%27%20fill-opacity%3D%220%22%2F%3E%3C%2Fsvg%3E\" data-srcset=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-6-200x188.png 200w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-6-300x281.png 300w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-6-400x375.png 400w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-6.png 535w\" data-sizes=\"auto\" data-orig-sizes=\"(max-width: 700px) 100vw, 700px\" \/><\/p>\n<p>\u4ee4\u724c\u7ea7\u5206\u6570\u5bf9\u53ef\u89c6\u5316\u7279\u522b\u6709\u7528\u3002\u56de\u590d\u4e2d\u7684\u6bcf\u4e2a\u6807\u8bb0\u90fd\u6709\u81ea\u5df1\u7684\u5e7b\u89c9\u6982\u7387\uff0c\u60a8\u53ef\u4ee5\u5c06\u5176\u5448\u73b0\u4e3a\u989c\u8272\u68af\u5ea6\uff0c\u7eff\u8272\u4ee3\u8868\u6709\u628a\u63e1\uff0c\u7ea2\u8272\u4ee3\u8868\u4e0d\u786e\u5b9a\u3002\u8fd9\u6837\u4e00\u76ee\u4e86\u7136\uff0c\u60a8\u5c31\u80fd\u6e05\u695a\u5730\u770b\u5230\u56de\u590d\u4e2d\u54ea\u4e9b\u90e8\u5206\u503c\u5f97\u4ed4\u7ec6\u7814\u7a76\u3002.<\/p>\n<h2>\u5728 RAG \u7ba1\u9053\u4e2d<\/h2>\n<p>\u5728\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08Retrieval-Augmented Generation\uff09\u4e2d\uff0c\u8fd9\u4e00\u70b9\u53d8\u5f97\u975e\u5e38\u5b9e\u7528\u3002\u8bd5\u60f3\u4e00\u4e2a\u4ece\u77e5\u8bc6\u5e93\u4e2d\u68c0\u7d22\u6587\u6863\u5e76\u5c06\u5176\u4f5c\u4e3a\u4e0a\u4e0b\u6587\u8f93\u5165 LLM \u7684\u7ba1\u9053\u3002\u5982\u679c\u68c0\u7d22\u5931\u8d25\uff08\u9519\u8bef\u7684\u6587\u6863\u3001\u7f3a\u5931\u7684\u9875\u9762\u3001\u4e0d\u5b8c\u6574\u7684\u4e0a\u4e0b\u6587\u7b49\uff09\uff0c\u8be5\u6a21\u578b\u5c06\u5c1d\u8bd5\u4ece\u5176\u53c2\u6570\u5316\u8bb0\u5fc6\u4e2d\u586b\u8865\u7a7a\u767d\uff0c\u8fd9\u5c31\u662f\u5e7b\u89c9\u6084\u7136\u51fa\u73b0\u7684\u5730\u65b9\u3002.<\/p>\n<p>\u901a\u8fc7\u4eba\u5de5\u667a\u80fd\uff0c\u60a8\u53ef\u4ee5\u6dfb\u52a0\u4e00\u4e2a\u95e8\uff1a<\/p>\n<p><img decoding=\"async\" class=\"lazyload aligncenter wp-image-1190472\" src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-7.png\" data-orig-src=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-7.png\" alt=\"\" width=\"700\" height=\"187\" srcset=\"data:image\/svg+xml,%3Csvg%20xmlns%3D%27http%3A%2F%2Fwww.w3.org%2F2000%2Fsvg%27%20width%3D%27700%27%20height%3D%27187%27%20viewBox%3D%270%200%20700%20187%27%3E%3Crect%20width%3D%27700%27%20height%3D%27187%27%20fill-opacity%3D%220%22%2F%3E%3C%2Fsvg%3E\" data-srcset=\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-7-200x53.png 200w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-7-300x80.png 300w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-7-400x107.png 400w, https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/05\/Detecting-Hallucinations-in-LLMs-One-Token-at-a-Time-Image-7.png 454w\" data-sizes=\"auto\" data-orig-sizes=\"(max-width: 700px) 100vw, 700px\" \/><\/p>\n<h2>\u6211\u4eec\u7684\u79d1\u5b66\u6587\u7ae0\u7b80\u8ff0--\u6211\u4eec\u7684\u53d1\u73b0<\/h2>\n<p>\u6211\u4eec\u5728\u56db\u79cd LLM\uff08Mistral-Small-24B\u3001Falcon-3-10B\u3001Phi-4\u3001Ministral-8B\uff09\u4e0a\u6d4b\u8bd5\u4e86 EPR \u548c WEPR \u7684\u4e09\u4e2a\u4efb\u52a1\uff1aTriviaQA \u4e0a\u7684\u5e7b\u89c9\u68c0\u6d4b\u3001WebQuestions \u7684\u6cdb\u5316\u4ee5\u53ca\u91d1\u878d RAG \u8bbe\u7f6e\u4e2d\u7684\u7f3a\u5931\u4e0a\u4e0b\u6587\u68c0\u6d4b\u3002.<\/p>\n<p>\u51e0\u4e2a\u4eae\u70b9<\/p>\n<ul>\n<li><strong>WEPR \u59cb\u7ec8\u4f18\u4e8e\u73b0\u6709\u65b9\u6cd5\u3002.<\/strong> \u5728\u51e0\u4e4e\u6240\u6709\u7684\u6a21\u578b-dataset \u7ec4\u5408\u4e2d\uff0c\u5b83\u90fd\u51fb\u8d25\u4e86 SelfCheckGPT\uff08\u4e00\u79cd\u9700\u8981 10 \u500d\u8ba1\u7b97\u91cf\u7684\u591a\u955c\u5934\u65b9\u6cd5\uff09\u548c HalluDetect\uff08\u4e00\u79cd\u5355\u955c\u5934\u7ade\u4e89\u5bf9\u624b\uff09\u3002.<\/li>\n<li><strong>\u4f60\u4e0d\u9700\u8981\u592a\u591a\u5bf9\u6570\u6982\u7387\u3002.<\/strong> \u6bcf\u4e2a\u4ee4\u724c\u7684\u6027\u80fd\u5728 K = 8-10 \u4e2a\u53ef\u8bbf\u95ee\u5bf9\u6570\u6982\u7387\u65f6\u8d8b\u4e8e\u7a33\u5b9a\u3002\u5373\u4f7f API \u8bbf\u95ee\u53d7\u9650\uff0c\u4fe1\u53f7\u4ecd\u7136\u5b58\u5728\u3002.<\/li>\n<li><strong>\u5b83\u5177\u6709\u6982\u62ec\u6027\u3002.<\/strong> \u5728 TriviaQA \u4e0a\u8bad\u7ec3\u7684 WEPR \u53ef\u4ee5\u5f88\u597d\u5730\u5e94\u7528\u4e8e WebQuestions\uff0c\u751a\u81f3\u5e94\u7528\u4e8e\u4e13\u95e8\u7684\u91d1\u878d\u8bed\u6599\u5e93\uff0c\u68c0\u6d4b\u51fa RAG \u7cfb\u7edf\u751f\u6210\u7684\u7b54\u6848\u7f3a\u4e4f\u8db3\u591f\u4e0a\u4e0b\u6587\u7684\u60c5\u51b5\u3002.<\/li>\n<li><strong>\u901f\u5ea6\u5f88\u5feb\u3002.<\/strong> \u6bcf\u4e2a\u5e8f\u5217\u7684\u8bc4\u5206\u5927\u7ea6\u9700\u8981 80 \u5fae\u79d2\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0cSelfCheckGPT \u5219\u9700\u8981 10 \u79d2\u4ee5\u4e0a\u3002.<\/li>\n<\/ul>\n<p>\u5728\u6211\u4eec\u5bf9\u91d1\u878d RAG \u4efb\u52a1\uff08\u5206\u6790 ArGiMi-Ardian dataset \u4e2d\u7684 10-K \u5e74\u5ea6 reports\uff09\u7684\u5b9e\u9a8c\u4e2d\uff0cWEPR \u5728\u68c0\u6d4b\u65e0\u6b63\u786e\u4e0a\u4e0b\u6587\u60c5\u51b5\u4e0b\u751f\u6210\u7684\u54cd\u5e94\u65f6\uff0cROC-AUC \u9ad8\u8fbe 93.6\u3002\u8fd9\u662f\u89e6\u53d1\u7b2c\u4e8c\u6b21\u68c0\u7d22\u7684\u5f3a\u70c8\u4fe1\u53f7\u3002.<\/p>\n<h3>\u5173\u4e8e\u5bf9\u6570\u6982\u7387\u5b58\u53d6\u7684\u8bf4\u660e\uff1a<\/h3>\n<p>\u4e0a\u8ff0\u4e00\u5207\u90fd\u6709\u8d56\u4e8e\u4e00\u4ef6\u4e8b\uff1a\u4ece\u6a21\u578b\u4e2d\u83b7\u53d6\u6807\u8bb0\u7ea7\u65e5\u5fd7\u6982\u7387\u3002\u8fd9\u6837\u6211\u4eec\u624d\u80fd\u8ba1\u7b97\u71b5\uff0c\u8fdb\u800c\u8ba1\u7b97\u5e7b\u89c9\u5206\u6570\u3002.<\/p>\n<p>\u5982\u4eca\uff0c\u8fd9\u79cd\u8bbf\u95ee\u6743\u9650\u65e0\u6cd5\u4fdd\u8bc1\u3002Anthropic \u6ca1\u6709\u901a\u8fc7\u5176 API \u516c\u5f00\u65e5\u5fd7\u6982\u7387\u3002OpenAI \u4e3a\u975e\u63a8\u7406\u6a21\u578b\u63d0\u4f9b\u4e86\u5b83\u4eec--\u60a8\u53ef\u4ee5\u4f7f\u7528 GPT-5.4 \u6216 GPT-5.4-mini \u8bf7\u6c42 top_logprobs\uff0c\u4f46\u524d\u63d0\u662f\u60a8\u5fc5\u987b\u5c06\u63a8\u7406\u5de5\u4f5c\u8bbe\u7f6e\u4e3a \"\u65e0\"\u3002\u53e6\u4e00\u65b9\u9762\uff0cGoogle \u5141\u8bb8\u901a\u8fc7\u5176 generate_content API \u8bbf\u95ee\u6240\u6709 logprobs\u3002.<\/p>\n<p>\u901a\u8fc7 vLLM \u6216\u7c7b\u4f3c\u63a8\u7406\u5f15\u64ce\u63d0\u4f9b\u7684\u5f00\u653e\u6743\u91cd\u6a21\u578b\u53ef\u5b9e\u73b0\u5b8c\u5168\u8bbf\u95ee\u3002.<\/p>\n<p>\u8fd9\u5f88\u91cd\u8981\u3002\u5bf9\u6570\u6982\u7387\u662f\u4e00\u79cd\u8f7b\u91cf\u7ea7\u3001\u4fe1\u606f\u4e30\u5bcc\u7684\u4fe1\u53f7\u3002\u751f\u6210\u5bf9\u6570\u6982\u7387\u4e0d\u9700\u8981\u989d\u5916\u6210\u672c\uff08\u53cd\u6b63\u6a21\u578b\u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u4e5f\u4f1a\u8ba1\u7b97\u5bf9\u6570\u6982\u7387\uff09\uff0c\u800c\u4e14\u5bf9\u6570\u6982\u7387\u8fd8\u652f\u6301\u4e00\u6574\u5957\u4e0d\u786e\u5b9a\u6027\u91cf\u5316\u65b9\u6cd5--\u5305\u62ec\u6211\u4eec\u7684\u65b9\u6cd5\u3002\u9650\u5236\u5bf9\u6570\u6982\u7387\u7684\u4f7f\u7528\u4f1a\u5c06\u7528\u6237\u63a8\u5411\u5bf9\u6a21\u578b\u8f93\u51fa\u7684\u76f2\u76ee\u4fe1\u4efb\u6216\u6602\u8d35\u7684\u591a\u62cd\u68c0\u6d4b\u65b9\u6cd5\u3002.<\/p>\n<p>\u5982\u679c\u60a8\u5728\u751f\u4ea7\u4e2d\u4f7f\u7528 LLM \u5e76\u5173\u5fc3\u8f93\u51fa\u7684\u53ef\u9760\u6027\uff0c\u90a3\u4e48\u5bf9\u6570\u6982\u7387\u7684\u53ef\u7528\u6027\u5e94\u8be5\u662f\u60a8\u9009\u62e9\u6a21\u578b\u6807\u51c6\u7684\u4e00\u90e8\u5206\u3002\u5982\u679c\u60a8\u662f\u6a21\u578b\u63d0\u4f9b\u5546\uff1a\u516c\u5f00\u5bf9\u6570\u6982\u7387\u662f\u63d0\u9ad8\u6a21\u578b\u53ef\u4fe1\u5ea6\u7684\u6700\u5ec9\u4ef7\u65b9\u6cd5\u4e4b\u4e00\u3002.<\/p>","protected":false},"excerpt":{"rendered":"<p>\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u80fd\u529b\u4ee4\u4eba\u60ca\u53f9\u3002\u5b83\u4eec\u603b\u7ed3\u3001\u7ffb\u8bd1\u3001\u63a8\u7406\u548c\u7f16\u7801\uff08\u6bd4\u6211\u5f3a\uff09\u3002\u4f46\u4e0e\u6211\u4e0d\u540c\u7684\u662f\uff0c\u5b83\u4eec\u4e5f\u56e0\u4ee5\u4ee4\u4eba\u4e0d\u5b89\u7684\u81ea\u4fe1\u7f16\u9020\u4e8b\u5b9e\u800c\u58f0\u540d\u72fc\u85c9\u3002.<\/p>","protected":false},"featured_media":1190338,"parent":0,"template":"","meta":{"_acf_changed":false,"ep_exclude_from_search":false},"blog-category":[21939,2995],"blog-language":[2991],"class_list":["post-1190337","blog","type-blog","status-publish","has-post-thumbnail","hentry","blog-category-medium","blog-category-ai-technology","blog-language-en"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.artefact.com\/zh\/wp-json\/wp\/v2\/blog\/1190337","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.artefact.com\/zh\/wp-json\/wp\/v2\/blog"}],"about":[{"href":"https:\/\/www.artefact.com\/zh\/wp-json\/wp\/v2\/types\/blog"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.artefact.com\/zh\/wp-json\/wp\/v2\/media\/1190338"}],"wp:attachment":[{"href":"https:\/\/www.artefact.com\/zh\/wp-json\/wp\/v2\/media?parent=1190337"}],"wp:term":[{"taxonomy":"blog-category","embeddable":true,"href":"https:\/\/www.artefact.com\/zh\/wp-json\/wp\/v2\/blog-category?post=1190337"},{"taxonomy":"blog-language","embeddable":true,"href":"https:\/\/www.artefact.com\/zh\/wp-json\/wp\/v2\/blog-language?post=1190337"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}