{"id":3848,"date":"2025-07-16T02:28:03","date_gmt":"2025-07-16T00:28:03","guid":{"rendered":"https:\/\/implementi.ai\/2025\/07\/16\/google-study-reveals-llms-may-discard-correct-answers-under-pressure-posing-risks-to-multi-turn-ai-interactions\/"},"modified":"2025-08-29T11:56:41","modified_gmt":"2025-08-29T09:56:41","slug":"%e8%b0%b7%e6%ad%8c%e7%a0%94%e7%a9%b6%e5%8f%91%e7%8e%b0%ef%bc%8c%e5%9c%a8%e5%8e%8b%e5%8a%9b%e4%b8%8b%ef%bc%8cllms-%e5%8f%af%e8%83%bd%e4%bc%9a%e6%94%be%e5%bc%83%e6%ad%a3%e7%a1%ae%e7%ad%94%e6%a1%88","status":"publish","type":"post","link":"https:\/\/implementi.ai\/zh\/2025\/07\/16\/google-study-reveals-llms-may-discard-correct-answers-under-pressure-posing-risks-to-multi-turn-ai-interactions\/","title":{"rendered":"\u8c37\u6b4c\u7814\u7a76\u663e\u793a\u6cd5\u5b66\u7855\u58eb\u5728\u538b\u529b\u4e0b\u53ef\u80fd\u653e\u5f03\u6b63\u786e\u7b54\u6848"},"content":{"rendered":"<p>\u4e0e\u4eba\u7c7b\u4e00\u6837\uff0c\u4eba\u5de5\u667a\u80fd\uff08AI\uff09\u4e5f\u4f1a\u8868\u73b0\u51fa\u4e00\u79cd\u5947\u7279\u7684\u81ea\u4fe1\u6096\u8bba\u3002\u4e00\u65b9\u9762\uff0c\u4eba\u5de5\u667a\u80fd\u4f1a\u53d8\u5f97\u56fa\u6267\u5df1\u89c1\uff0c\u53e6\u4e00\u65b9\u9762\uff0c\u5f53\u53d7\u5230 \"\u538b\u529b \"\u65f6\uff0c\u53c8\u4f1a\u8868\u73b0\u51fa\u8f7b\u6613\u653e\u5f03\u7acb\u573a\u7684\u503e\u5411\u3002\u6700\u8fd1\uff0c\u7531\u8c37\u6b4c\u6536\u8d2d\u7684\u82f1\u56fd\u4eba\u5de5\u667a\u80fd\u5b9e\u9a8c\u5ba4 DeepMind \u8fdb\u884c\u7684\u4e00\u9879\u7814\u7a76\u53d1\u73b0\uff0c\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08LLMs\uff09\u5c31\u5177\u6709\u8fd9\u79cd\u72ec\u7279\u7684\u7279\u6027\u3002\u7136\u800c\uff0c\u4eba\u5de5\u667a\u80fd\u4fe1\u5fc3\u7684\u77db\u76fe\u6027\u53ef\u80fd\u4f1a\u5bf9\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u7684\u5f00\u53d1\u548c\u5e94\u7528\u4ea7\u751f\u4e25\u91cd\u5f71\u54cd\uff0c\u5c24\u5176\u662f\u591a\u8f6c\u5f2f\u7cfb\u7edf\u3002<\/p>\n<h2>\u4eba\u5de5\u667a\u80fd\u7684\u4fe1\u5fc3\uff1a\u4e00\u628a\u53cc\u5203\u5251<\/h2>\n<p>\u5728\u4eba\u5de5\u667a\u80fd\u9886\u57df\uff0c\u7f6e\u4fe1\u5ea6\u53ef\u89c6\u4e3a\u6a21\u578b\u5728\u5176\u9884\u6d4b\u6216\u51b3\u7b56\u4e2d\u8868\u73b0\u51fa\u7684\u786e\u5b9a\u7a0b\u5ea6\u3002\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u53ef\u4ee5\u5728\u4e00\u5b9a\u7a0b\u5ea6\u4e0a\u9884\u6d4b\u4e8b\u4ef6\u3002\u7136\u800c\uff0c\u4eba\u5de5\u667a\u80fd\u4fe1\u5fc3\u7684\u6709\u8da3\u4e4b\u5904\u5728\u4e8e\u5b83\u7ecf\u5e38\u8868\u73b0\u51fa\u4e24\u79cd\u622a\u7136\u4e0d\u540c\u7684\u7279\u5f81\uff1a\u8fc7\u5ea6\u81ea\u4fe1\u548c\u4fe1\u5fc3\u4e0d\u8db3\u3002<\/p>\n<p>\u4eba\u5de5\u667a\u80fd\u4f1a\u8868\u73b0\u51fa\u8fc7\u5ea6\u81ea\u4fe1\uff0c\u4e5f\u88ab\u79f0\u4e3a\u56fa\u6267\uff0c\u5373\u575a\u6301\u81ea\u5df1\u6700\u521d\u7684\u9884\u6d4b\uff0c\u5373\u4f7f\u9762\u5bf9\u77db\u76fe\u7684\u8bc1\u636e\uff0c\u4e5f\u5b8c\u5168\u4e0d\u613f\u6539\u53d8\u81ea\u5df1\u7684\u7acb\u573a\u3002\u76f8\u53cd\uff0c\u4eba\u5de5\u667a\u80fd\u4e5f\u4f1a\u8868\u73b0\u51fa\u4e0d\u81ea\u4fe1\uff0c\u8fd9\u79cd\u7279\u6027\u7684\u7279\u70b9\u662f\u503e\u5411\u4e8e\u653e\u5f03\u81ea\u5df1\u6700\u521d\u7684\u9884\u6d4b\uff0c\u800c\u4e14\u5f80\u5f80\u662f\u5728\u6700\u8f7b\u5fae\u7684 \u2018\u538b\u529b \u2019\u4e0b\u505a\u51fa\u6b63\u786e\u7684\u9884\u6d4b\u3002\u2019<\/p>\n<p>\u8fd9\u79cd\u4fe1\u5fc3\u6096\u8bba\u5c55\u73b0\u4e86\u4eba\u5de5\u667a\u80fd\u7684\u4e00\u4e2a\u8010\u4eba\u5bfb\u5473\u7684\u7279\u6027--\u4e00\u4e2a\u65e2\u4e0d\u53ef\u6216\u7f3a\u53c8\u5177\u6709\u6f5c\u5728\u7834\u574f\u6027\u7684\u7279\u6027\u3002\u4f46\u5173\u952e\u95ee\u9898\u662f\uff0c\u8fd9\u79cd\u602a\u5f02\u6027\u5bf9\u4eba\u5de5\u667a\u80fd\u5e94\u7528\u7684\u672a\u6765\u610f\u5473\u7740\u4ec0\u4e48\uff1f<\/p>\n<h2>\u5bf9\u591a\u5708\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u7684\u5f71\u54cd<\/h2>\n<p>\u591a\u56de\u5408\u4eba\u5de5\u667a\u80fd\u7cfb\u7edf\u53c2\u4e0e\u7684\u4e92\u52a8\u8de8\u8d8a\u591a\u4e2a\u56de\u5408\uff0c\u7c7b\u4f3c\u4e8e\u4eba\u7c7b\u5bf9\u8bdd\u3002\u6b64\u7c7b\u7cfb\u7edf\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u4f9d\u8d56\u4e8e\u6a21\u578b\u51c6\u786e\u9884\u6d4b\u548c\u5e94\u5bf9\u4e0d\u540c\u56de\u5408\u7684\u80fd\u529b\uff0c\u56e0\u6b64\u6a21\u578b\u7684\u53ef\u4fe1\u5ea6\u81f3\u5173\u91cd\u8981\u3002.<\/p>\n<p>DeepMind \u7684\u7814\u7a76\u8868\u660e\uff0c\u4eba\u5de5\u667a\u80fd\u81ea\u4fe1\u7684\u77db\u76fe\u884c\u4e3a\u53ef\u80fd\u4f1a\u5bf9\u8fd9\u4e9b\u7cfb\u7edf\u7684\u7a33\u5b9a\u6027\u548c\u53ef\u9760\u6027\u6784\u6210\u5a01\u80c1\u3002\u4f8b\u5982\uff0c\u5982\u679c\u4e00\u4e2a\u8fc7\u4e8e\u81ea\u4fe1\u7684\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u5728\u4e0d\u65ad\u53d1\u5c55\u7684\u4e92\u52a8\u4e2d\u62d2\u7edd\u63a5\u53d7\u65b0\u4fe1\u606f\uff0c\u5c31\u53ef\u80fd\u5bfc\u81f4\u4e0d\u51c6\u786e\u7684\u53cd\u5e94\u3002\u540c\u6837\uff0c\u4fe1\u5fc3\u4e0d\u8db3\u7684\u6a21\u578b\u5982\u679c\u8f7b\u6613\u653e\u5f03\u6b63\u786e\u7684\u9884\u6d4b\uff0c\u4e5f\u4f1a\u5bfc\u81f4\u9519\u8bef\u7684\u7ed3\u8bba\u3002<\/p>\n<p>\u5728\u51b3\u7b56\u5f71\u54cd\u91cd\u5927\u7684\u60c5\u51b5\u4e0b\uff0c\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u7684\u4e0d\u7a33\u5b9a\u884c\u4e3a\u53ef\u80fd\u4f1a\u5e26\u6765\u6709\u5bb3\u540e\u679c\u3002<\/p>\n<p>\u5c3d\u7ba1\u4eba\u5de5\u667a\u80fd\u96be\u9898\u7684\u7b80\u6d01\u6027\u4ecd\u7136\u4ee4\u4eba\u56f0\u60d1\uff0c\u4f46 DeepMind \u548c\u5176\u4ed6\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u4eba\u5458\u5df2\u7ecf\u5f00\u59cb\u52aa\u529b\u7406\u89e3\u548c\u7ea0\u6b63\u8fd9\u4e9b\u95ee\u9898\u3002\u5728\u6df1\u5165\u4e86\u89e3\u4e86\u4eba\u5de5\u667a\u80fd\u7684\u81ea\u4fe1\u884c\u4e3a\u4e4b\u540e\uff0c\u4eba\u5de5\u667a\u80fd\u5f00\u53d1\u4eba\u5458\u5728\u8bbe\u8ba1\u672a\u6765\u7684\u4eba\u5de5\u667a\u80fd\u5e94\u7528\u65f6\u5fc5\u987b\u8003\u8651\u5230\u8fd9\u4e00\u6096\u8bba\u3002\u53ea\u6709\u8fd9\u6837\uff0c\u624d\u80fd\u5145\u5206\u53d1\u6325\u4eba\u5de5\u667a\u80fd\u7684\u6f5c\u529b\uff0c\u5e76\u5c06\u5176\u98ce\u9669\u964d\u81f3\u6700\u4f4e\u3002<\/p>\n<p>\u8fd9\u9879\u7814\u7a76\u63d0\u9192\u4eba\u4eec\u6ce8\u610f\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u7684\u52a8\u6001\u6027\u548c\u6f14\u53d8\u6027\u3002\u5b83\u63ed\u793a\u4e86\u4eba\u5de5\u667a\u80fd\u7684\u590d\u6742\u7279\u6027\uff0c\u8fd9\u4e9b\u7279\u6027\u4f7f\u5176\u65e2\u8ff7\u4eba\u53c8\u5145\u6ee1\u6311\u6218\u3002\u968f\u7740\u6211\u4eec\u4e0d\u65ad\u62e5\u62b1\u4eba\u5de5\u667a\u80fd\u5e76\u5c06\u5176\u878d\u5165\u6211\u4eec\u7684\u4e16\u754c\uff0c\u4e86\u89e3\u5176\u884c\u4e3a\u7684\u65b9\u65b9\u9762\u9762\u4e0d\u4ec5\u662f\u53ef\u53d6\u7684\uff0c\u4e5f\u662f\u5fc5\u8981\u7684\u3002<\/p>\n<p>\u6709\u5173\u8be5\u7814\u7a76\u7684\u66f4\u591a\u4fe1\u606f\uff0c\u8bf7\u8bbf\u95ee\u539f\u6587 <a href=\"https:\/\/venturebeat.com\/ai\/google-study-shows-llms-abandon-correct-answers-under-pressure-threatening-multi-turn-ai-systems\/\" target=\"_blank\" rel=\"noopener\">\u8fd9\u91cc<\/a>.<\/p>","protected":false},"excerpt":{"rendered":"<p>Just like humans, artificial intelligence (AI) can display a peculiar confidence paradox. AI can get stubborn on one hand, and on the other hand, exhibit a tendency to easily abandon its stance when \u201cpressured\u201d. Recent findings from a study conducted by DeepMind, the British AI lab acquired by Google, showcase this unique characteristic in large language models (LLMs). However, the [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":3849,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[26],"tags":[],"class_list":["post-3848","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-automation"],"featured_image_src":"https:\/\/implementi.ai\/wp-content\/uploads\/2025\/07\/3848-1024x683.png","blog_images":{"medium":"https:\/\/implementi.ai\/wp-content\/uploads\/2025\/07\/3848-300x200.png","large":"https:\/\/implementi.ai\/wp-content\/uploads\/2025\/07\/3848-1024x683.png"},"ams_acf":[],"jetpack_featured_media_url":"https:\/\/implementi.ai\/wp-content\/uploads\/2025\/07\/3848.png","jetpack_sharing_enabled":true,"_links":{"self":[{"href":"https:\/\/implementi.ai\/zh\/wp-json\/wp\/v2\/posts\/3848","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/implementi.ai\/zh\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/implementi.ai\/zh\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/implementi.ai\/zh\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/implementi.ai\/zh\/wp-json\/wp\/v2\/comments?post=3848"}],"version-history":[{"count":1,"href":"https:\/\/implementi.ai\/zh\/wp-json\/wp\/v2\/posts\/3848\/revisions"}],"predecessor-version":[{"id":4907,"href":"https:\/\/implementi.ai\/zh\/wp-json\/wp\/v2\/posts\/3848\/revisions\/4907"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/implementi.ai\/zh\/wp-json\/wp\/v2\/media\/3849"}],"wp:attachment":[{"href":"https:\/\/implementi.ai\/zh\/wp-json\/wp\/v2\/media?parent=3848"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/implementi.ai\/zh\/wp-json\/wp\/v2\/categories?post=3848"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/implementi.ai\/zh\/wp-json\/wp\/v2\/tags?post=3848"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}