{"id":5749,"date":"2025-03-02T14:20:51","date_gmt":"2025-03-02T06:20:51","guid":{"rendered":"https:\/\/nullthought.net\/?p=5749"},"modified":"2025-12-10T14:08:35","modified_gmt":"2025-12-10T06:08:35","slug":"openai-gpt-4-5-%e7%b3%bb%e7%bb%9f%e8%af%b4%e6%98%8e%ef%bc%88openai-gpt-4-5-system-card%ef%bc%89","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=5749","title":{"rendered":"OpenAI GPT-4.5 \u7cfb\u7edf\u8bf4\u660e\uff08OpenAI GPT-4.5 System Card\uff09"},"content":{"rendered":"\n<p><strong>\u4e00\u3001\u5f15\u8a00<\/strong><br><strong><a href=\"https:\/\/cdn.openai.com\/gpt-4-5-system-card-2272025.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">GPT-4.5<\/a><\/strong> \u662f OpenAI \u8fc4\u4eca\u4e3a\u6b62\u6700\u5927\u3001\u6700\u5177\u77e5\u8bc6\u6027\u7684 AI \u6a21\u578b\uff0c\u5176\u76ee\u6807\u662f\u76f8\u8f83\u4e8e\u524d\u4ee3\u6a21\u578b\uff08\u5982 GPT-4o\uff09\u66f4\u5177\u901a\u7528\u6027\uff0c\u540c\u65f6\u4fdd\u6301\u9ad8\u6548\u7684 STEM \u9886\u57df\u63a8\u7406\u80fd\u529b\u3002\u8be5\u6a21\u578b\u91c7\u7528\u4e86\u6269\u5c55\u9884\u8bad\u7ec3\u89c4\u6a21\u7684\u7b56\u7565\uff0c\u5e76\u7ed3\u5408\u4e86\u5148\u8fdb\u7684\u76d1\u7763\u5fae\u8c03\uff08supervised fine-tuning, SFT\uff09\u548c\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\uff08reinforcement learning from human feedback, RLHF\uff09\u8fdb\u884c\u8bad\u7ec3\u3002\u6b64\u5916\uff0cGPT-4.5 \u5728\u90e8\u7f72\u524d\u8fdb\u884c\u4e86\u5e7f\u6cdb\u7684\u5b89\u5168\u8bc4\u4f30\uff0c\u7ed3\u679c\u8868\u660e\uff0c\u8be5\u6a21\u578b\u5728\u5b89\u5168\u6027\u65b9\u9762\u5e76\u672a\u6bd4\u73b0\u6709\u6a21\u578b\u8868\u73b0\u51fa\u66f4\u9ad8\u7684\u98ce\u9669\u3002<\/p>\n\n\n\n<p>\u4e0e\u524d\u4ee3\u6a21\u578b\u76f8\u6bd4\uff0cGPT-4.5 \u5728\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762\u5177\u6709\u663e\u8457\u4f18\u52bf\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u4ea4\u4e92\u4f53\u9a8c\u66f4\u52a0\u81ea\u7136<\/strong>\u2014\u2014\u7531\u4e8e\u6539\u8fdb\u4e86\u5bf9\u9f50\u6280\u672f\u548c\u6570\u636e\u9884\u5904\u7406\uff0c\u8be5\u6a21\u578b\u5728\u5bf9\u8bdd\u6d41\u7545\u6027\u548c\u7406\u89e3\u80fd\u529b\u65b9\u9762\u6709\u4e86\u8fdb\u4e00\u6b65\u63d0\u5347\uff0c\u4f7f\u7528\u6237\u80fd\u591f\u83b7\u5f97\u66f4\u52a0\u8d34\u8fd1\u4eba\u7c7b\u4ea4\u6d41\u7684\u4f53\u9a8c\u3002<\/li>\n\n\n\n<li><strong>\u77e5\u8bc6\u5e93\u66f4\u52a0\u5e7f\u6cdb<\/strong>\u2014\u2014\u901a\u8fc7\u6269\u5927\u8bad\u7ec3\u6570\u636e\u8303\u56f4\uff0c\u5e76\u5f15\u5165\u65b0\u578b\u6570\u636e\u8fc7\u6ee4\u548c\u7b5b\u9009\u673a\u5236\uff0c\u8be5\u6a21\u578b\u5728\u4e16\u754c\u77e5\u8bc6\u3001\u6280\u672f\u7ec6\u8282\u548c\u4e13\u4e1a\u9886\u57df\u4fe1\u606f\u65b9\u9762\u7684\u8868\u73b0\u66f4\u52a0\u4f18\u8d8a\u3002<\/li>\n\n\n\n<li><strong>\u7528\u6237\u610f\u56fe\u5bf9\u9f50\u80fd\u529b\u66f4\u5f3a<\/strong>\u2014\u2014\u5f15\u5165\u4e86\u66f4\u5148\u8fdb\u7684\u5bf9\u9f50\u65b9\u6cd5\uff0c\u4f7f\u5f97\u6a21\u578b\u5728\u7406\u89e3\u7528\u6237\u6307\u4ee4\u548c\u610f\u56fe\u7684\u8fc7\u7a0b\u4e2d\u8868\u73b0\u66f4\u52a0\u7cbe\u51c6\uff0c\u80fd\u591f\u66f4\u597d\u5730\u9002\u5e94\u4e0d\u540c\u8bed\u5883\u4e0b\u7684\u9700\u6c42\u3002<\/li>\n\n\n\n<li><strong>\u60c5\u611f\u667a\u80fd(emotional intelligence)\u589e\u5f3a<\/strong>\u2014\u2014\u76f8\u8f83\u4e8e\u524d\u4ee3\u6a21\u578b\uff0cGPT-4.5 \u66f4\u64c5\u957f\u8bc6\u522b\u5e76\u56de\u5e94\u5e26\u6709\u60c5\u7eea\u8272\u5f69\u7684\u67e5\u8be2\uff0c\u5982\u63d0\u4f9b\u5efa\u8bae\u3001\u7f13\u89e3\u7528\u6237\u60c5\u7eea\u6216\u5355\u7eaf\u4f5c\u4e3a\u503e\u542c\u8005\u966a\u4f34\u7528\u6237\u3002<\/li>\n\n\n\n<li><strong>\u51cf\u5c11\u5e7b\u89c9\uff08hallucination\uff09<\/strong>\u2014\u2014\u6539\u8fdb\u7684\u8bad\u7ec3\u548c\u540e\u5904\u7406\u7b56\u7565\u4f7f\u6a21\u578b\u5728\u4fe1\u606f\u751f\u6210\u8fc7\u7a0b\u4e2d\u66f4\u80fd\u4fdd\u6301\u4e8b\u5b9e\u6027\uff0c\u964d\u4f4e\u4e86\u9519\u8bef\u56de\u7b54\u548c\u65e0\u6839\u636e\u63a8\u6d4b\u7684\u6982\u7387\u3002<\/li>\n<\/ol>\n\n\n\n<p>GPT-4.5 \u76ee\u524d\u4ecd\u5904\u4e8e\u7814\u7a76\u9884\u89c8\u9636\u6bb5\uff0c\u5176\u6838\u5fc3\u76ee\u6807\u4e4b\u4e00\u662f\u63a2\u7d22\u8be5\u6a21\u578b\u7684\u80fd\u529b\u548c\u5c40\u9650\u6027\uff0c\u4ee5\u4fbf OpenAI \u8fdb\u4e00\u6b65\u4f18\u5316\u5176\u6027\u80fd\uff0c\u5e76\u53d1\u73b0\u53ef\u80fd\u7684\u521b\u65b0\u5e94\u7528\u65b9\u5f0f\u3002\u4e3a\u4e86\u5b9e\u73b0\u8fd9\u4e00\u76ee\u6807\uff0cOpenAI \u6b63\u5728\u79ef\u6781\u6536\u96c6\u6765\u81ea\u4e0d\u540c\u7528\u6237\u7684\u53cd\u9988\uff0c\u4ee5\u8bc4\u4f30 GPT-4.5 \u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u7684\u8868\u73b0\uff0c\u5e76\u8c03\u6574\u672a\u6765\u7684\u6539\u8fdb\u65b9\u5411\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image aligncenter size-full\"><img loading=\"lazy\" decoding=\"async\" width=\"537\" height=\"752\" src=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/03\/image-2.png\" alt=\"\" class=\"wp-image-5750\" srcset=\"https:\/\/nullthought.net\/wp-content\/uploads\/2025\/03\/image-2.png 537w, https:\/\/nullthought.net\/wp-content\/uploads\/2025\/03\/image-2-214x300.png 214w\" sizes=\"auto, (max-width: 537px) 100vw, 537px\" \/><figcaption class=\"wp-element-caption\"><a href=\"https:\/\/cdn.openai.com\/gpt-4-5-system-card-2272025.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">OpenAI GPT-4.5 \u7cfb\u7edf\u8bf4\u660e\uff08OpenAI GPT-4.5 System Card\uff09<\/a><\/figcaption><\/figure>\n\n\n\n<p><strong>\u4e8c\u3001\u6a21\u578b\u6570\u636e\u4e0e\u8bad\u7ec3<\/strong><br>GPT-4.5 \u7684\u8bad\u7ec3\u57fa\u4e8e\u4e24\u5927\u6838\u5fc3\u8303\u5f0f\u7684\u6269\u5c55\uff1a<strong>\u65e0\u76d1\u7763\u5b66\u4e60\uff08Unsupervised Learning\uff09\u548c\u601d\u7ef4\u94fe\u63a8\u7406\uff08Chain-of-Thought Reasoning\uff09<\/strong>\u3002\u8fd9\u4e24\u79cd\u8303\u5f0f\u7684\u7ed3\u5408\uff0c\u4f7f\u5f97 GPT-4.5 \u5728\u4e16\u754c\u77e5\u8bc6\u7406\u89e3\u3001\u903b\u8f91\u63a8\u7406\u548c\u590d\u6742\u4efb\u52a1\u6c42\u89e3\u65b9\u9762\u5177\u6709\u66f4\u5f3a\u7684\u80fd\u529b\u3002<\/p>\n\n\n\n<p><strong>\u6269\u5c55\u65e0\u76d1\u7763\u5b66\u4e60<\/strong><br>\u65e0\u76d1\u7763\u5b66\u4e60\u7684\u6838\u5fc3\u76ee\u6807\u662f\u8ba9\u6a21\u578b\u5728\u6ca1\u6709\u660e\u786e\u6807\u6ce8\u7684\u60c5\u51b5\u4e0b\uff0c\u4ece\u5927\u91cf\u6570\u636e\u4e2d\u81ea\u52a8\u63d0\u53d6\u77e5\u8bc6\uff0c\u5e76\u5f62\u6210\u53ef\u9760\u7684\u4e16\u754c\u6a21\u578b\u3002GPT-4.5 \u8fdb\u4e00\u6b65\u6269\u5c55\u4e86\u65e0\u76d1\u7763\u5b66\u4e60\u7684\u89c4\u6a21\uff0c\u4f7f\u5176\u80fd\u591f\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u63d0\u9ad8\u4e16\u754c\u6a21\u578b\u7684\u51c6\u786e\u6027<\/strong>\u2014\u2014\u901a\u8fc7\u5f15\u5165\u66f4\u5e7f\u6cdb\u7684\u6570\u636e\u6e90\uff0c\u5e76\u4f18\u5316\u6570\u636e\u7b5b\u9009\u7b56\u7565\uff0cGPT-4.5 \u5728\u4e8b\u5b9e\u6027\u7406\u89e3\u548c\u77e5\u8bc6\u5173\u8054\u65b9\u9762\u66f4\u52a0\u7cbe\u51c6\u3002<\/li>\n\n\n\n<li><strong>\u964d\u4f4e\u5e7b\u89c9\u7387<\/strong>\u2014\u2014\u91c7\u7528\u4e86\u6539\u8fdb\u7684\u6570\u636e\u589e\u5f3a\u548c\u8fc7\u6ee4\u6280\u672f\uff0c\u4f7f\u6a21\u578b\u751f\u6210\u7684\u5185\u5bb9\u66f4\u7b26\u5408\u4e8b\u5b9e\uff0c\u540c\u65f6\u51cf\u5c11\u9519\u8bef\u4fe1\u606f\u7684\u4f20\u64ad\u3002<\/li>\n\n\n\n<li><strong>\u589e\u5f3a\u8054\u60f3\u80fd\u529b<\/strong>\u2014\u2014\u63d0\u5347\u4e86\u6a21\u578b\u5728\u4e0d\u540c\u4e3b\u9898\u4e4b\u95f4\u5efa\u7acb\u8054\u7cfb\u7684\u80fd\u529b\uff0c\u4f7f\u5176\u5728\u56de\u7b54\u590d\u6742\u95ee\u9898\u65f6\u66f4\u52a0\u7075\u6d3b\u548c\u5168\u9762\u3002<\/li>\n<\/ol>\n\n\n\n<p><strong>\u601d\u7ef4\u94fe\u63a8\u7406\u7684\u6269\u5c55<\/strong><br>\u601d\u7ef4\u94fe\u63a8\u7406\uff08CoT\uff09\u662f\u4e00\u79cd\u8bad\u7ec3\u65b9\u6cd5\uff0c\u65e8\u5728\u8ba9\u6a21\u578b\u5728\u56de\u7b54\u95ee\u9898\u524d\u8fdb\u884c\u201c\u601d\u8003\u201d\uff0c\u4ece\u800c\u63d0\u9ad8\u5176\u5728\u6570\u5b66\u3001\u903b\u8f91\u63a8\u7406\u548c\u7f16\u7a0b\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u3002GPT-4.5 \u901a\u8fc7\u6269\u5927 CoT \u8bad\u7ec3\u6570\u636e\u7684\u89c4\u6a21\uff0c\u5e76\u4f18\u5316\u63a8\u7406\u8def\u5f84\uff0c\u4f7f\u5176\u5728\u5904\u7406\u590d\u6742 STEM \u4efb\u52a1\u65f6\u8868\u73b0\u66f4\u52a0\u5f3a\u5927\u3002\u8fd9\u4e00\u6539\u8fdb\u4f7f\u5f97 GPT-4.5 \u80fd\u591f\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u66f4\u597d\u5730\u5904\u7406\u6570\u5b66\u95ee\u9898<\/strong>\u2014\u2014\u5728\u6570\u5b66\u63a8\u7406\u3001\u516c\u5f0f\u63a8\u5bfc\u548c\u95ee\u9898\u6c42\u89e3\u65b9\u9762\u8868\u73b0\u51fa\u66f4\u9ad8\u7684\u51c6\u786e\u7387\u548c\u903b\u8f91\u6027\u3002<\/li>\n\n\n\n<li><strong>\u63d0\u5347\u7f16\u7a0b\u80fd\u529b<\/strong>\u2014\u2014\u66f4\u64c5\u957f\u89e3\u6790\u4ee3\u7801\u3001\u8c03\u8bd5\u7a0b\u5e8f\u548c\u4f18\u5316\u7b97\u6cd5\uff0c\u80fd\u591f\u66f4\u6709\u6548\u5730\u5e2e\u52a9\u5f00\u53d1\u8005\u89e3\u51b3\u5b9e\u9645\u95ee\u9898\u3002<\/li>\n\n\n\n<li><strong>\u589e\u5f3a\u903b\u8f91\u63a8\u7406<\/strong>\u2014\u2014\u5728\u5904\u7406\u9700\u8981\u591a\u6b65\u63a8\u7406\u7684\u4efb\u52a1\uff08\u5982\u6cd5\u5f8b\u5206\u6790\u3001\u79d1\u5b66\u63a8\u6d4b\u7b49\uff09\u65f6\uff0c\u80fd\u591f\u63d0\u4f9b\u66f4\u7cfb\u7edf\u5316\u548c\u4e25\u8c28\u7684\u7b54\u6848\u3002<\/li>\n<\/ol>\n\n\n\n<p><strong>\u65b0\u7684\u5bf9\u9f50\u6280\u672f<\/strong><br>\u4e3a\u4e86\u786e\u4fdd GPT-4.5 \u66f4\u7b26\u5408\u4eba\u7c7b\u9700\u6c42\uff0cOpenAI \u5f00\u53d1\u4e86\u65b0\u7684\u53ef\u6269\u5c55\u5bf9\u9f50\u6280\u672f\u3002\u8fd9\u4e9b\u6280\u672f\u7684\u6838\u5fc3\u76ee\u6807\u662f\u8ba9\u6a21\u578b\u5728\u89e3\u51b3\u66f4\u590d\u6742\u3001\u66f4\u5e7f\u6cdb\u7684\u95ee\u9898\u65f6\uff0c\u80fd\u591f\u66f4\u51c6\u786e\u5730\u7406\u89e3\u4eba\u7c7b\u9700\u6c42\uff0c\u5e76\u751f\u6210\u7b26\u5408\u7528\u6237\u9884\u671f\u7684\u56de\u7b54\u3002<\/p>\n\n\n\n<p>\u8fd9\u4e9b\u5bf9\u9f50\u6280\u672f\u7684\u4e3b\u8981\u4f18\u52bf\u5305\u62ec\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u63d0\u9ad8\u53ef\u5f15\u5bfc\u6027\uff08Steerability\uff09<\/strong>\u2014\u2014GPT-4.5 \u80fd\u591f\u66f4\u597d\u5730\u9002\u5e94\u7528\u6237\u7684\u4e2a\u6027\u5316\u9700\u6c42\uff0c\u5728\u4e0d\u540c\u7684\u5bf9\u8bdd\u98ce\u683c\u3001\u4efb\u52a1\u76ee\u6807\u548c\u8bed\u5883\u4e0b\u8868\u73b0\u51fa\u66f4\u9ad8\u7684\u7075\u6d3b\u6027\u3002<\/li>\n\n\n\n<li><strong>\u589e\u5f3a\u5bf9\u7ec6\u5fae\u5dee\u522b\u7684\u7406\u89e3\u80fd\u529b<\/strong>\u2014\u2014\u5728\u5bf9\u8bdd\u8fc7\u7a0b\u4e2d\uff0cGPT-4.5 \u80fd\u591f\u8bc6\u522b\u8bed\u4e49\u4e0a\u7684\u5fae\u5c0f\u53d8\u5316\uff0c\u5e76\u76f8\u5e94\u8c03\u6574\u56de\u7b54\uff0c\u4f7f\u5176\u66f4\u7b26\u5408\u4e0a\u4e0b\u6587\u3002<\/li>\n\n\n\n<li><strong>\u5bf9\u8bdd\u66f4\u81ea\u7136\u3001\u66f4\u5177\u4e92\u52a8\u6027<\/strong>\u2014\u2014\u7528\u6237\u5728\u4e0e GPT-4.5 \u4ea4\u6d41\u65f6\uff0c\u4f1a\u611f\u53d7\u5230\u66f4\u63a5\u8fd1\u4eba\u7c7b\u4ea4\u8c08\u7684\u4f53\u9a8c\uff0c\u5c24\u5176\u662f\u5728\u6d89\u53ca\u60c5\u611f\u548c\u793e\u4ea4\u4ea4\u4e92\u7684\u573a\u666f\u4e0b\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u8fd9\u4e9b\u6539\u8fdb\u5f97\u76ca\u4e8e GPT-4.5 \u91c7\u7528\u7684\u591a\u6837\u5316\u8bad\u7ec3\u6570\u636e\uff0c\u5176\u4e2d\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u516c\u5f00\u6570\u636e<\/strong>\u2014\u2014\u6765\u81ea\u516c\u5171\u4e92\u8054\u7f51\u7684\u9ad8\u8d28\u91cf\u6587\u672c\u4fe1\u606f\uff1b<\/li>\n\n\n\n<li><strong>\u5408\u4f5c\u6570\u636e<\/strong>\u2014\u2014\u4e0e\u6570\u636e\u5408\u4f5c\u4f19\u4f34\u5408\u4f5c\u63d0\u4f9b\u7684\u4e13\u4e1a\u9886\u57df\u6570\u636e\uff1b<\/li>\n\n\n\n<li><strong>\u5185\u90e8\u5b9a\u5236\u6570\u636e<\/strong>\u2014\u2014OpenAI \u5185\u90e8\u5f00\u53d1\u7684\u7279\u5b9a\u4efb\u52a1\u6570\u636e\u96c6\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u4e3a\u4e86\u786e\u4fdd\u6570\u636e\u7684\u5b89\u5168\u6027\u548c\u5408\u89c4\u6027\uff0cGPT-4.5 \u7684\u6570\u636e\u5904\u7406\u7ba1\u9053\u91c7\u7528\u4e86\u4e25\u683c\u7684\u7b5b\u9009\u673a\u5236\uff0c\u907f\u514d\u5904\u7406\u6d89\u53ca\u4e2a\u4eba\u9690\u79c1\u7684\u4fe1\u606f\u3002\u6b64\u5916\uff0cOpenAI \u8fd8\u91c7\u7528\u4e86\u4e00\u7cfb\u5217\u5b89\u5168\u5206\u7c7b\u5668\uff0c\u6765\u68c0\u6d4b\u548c\u8fc7\u6ee4\u4e0d\u9002\u5b9c\u7528\u4e8e\u6a21\u578b\u8bad\u7ec3\u7684\u5185\u5bb9\uff0c\u4f8b\u5982\u5305\u542b\u672a\u6210\u5e74\u4eba\u76f8\u5173\u7684\u654f\u611f\u4fe1\u606f\u3002<\/p>\n\n\n\n<p><strong>\u4e09\u3001\u89c2\u5bdf\u5230\u7684\u5b89\u5168\u6311\u6218\u4e0e\u8bc4\u4f30<\/strong><br>GPT-4.5 \u7ecf\u8fc7\u4e86\u5168\u9762\u7684\u5b89\u5168\u6027\u8bc4\u4f30\uff0c\u6db5\u76d6\u4e86\u4ee5\u4e0b\u51e0\u4e2a\u65b9\u9762\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u4e0d\u5141\u8bb8\u7684\u5185\u5bb9\uff08Disallowed Content\uff09<\/strong>\u2014\u2014\u68c0\u6d4b\u6a21\u578b\u662f\u5426\u4f1a\u54cd\u5e94\u6d89\u53ca\u4ec7\u6068\u8a00\u8bba\u3001\u975e\u6cd5\u5efa\u8bae\u7b49\u8fdd\u89c4\u8bf7\u6c42\u3002<\/li>\n\n\n\n<li><strong>\u8d8a\u72f1\u653b\u51fb\uff08Jailbreaks\uff09<\/strong>\u2014\u2014\u6d4b\u8bd5\u6a21\u578b\u80fd\u5426\u62b5\u6297\u6076\u610f\u7528\u6237\u8bd5\u56fe\u7ed5\u8fc7\u5b89\u5168\u673a\u5236\u7684\u653b\u51fb\u3002<\/li>\n\n\n\n<li><strong>\u5e7b\u89c9\u95ee\u9898\uff08Hallucination\uff09<\/strong>\u2014\u2014\u8bc4\u4f30\u6a21\u578b\u5728\u56de\u7b54\u95ee\u9898\u65f6\u63d0\u4f9b\u865a\u5047\u4fe1\u606f\u7684\u6982\u7387\u3002<\/li>\n\n\n\n<li><strong>\u516c\u5e73\u6027\u4e0e\u504f\u89c1\uff08Fairness and Bias\uff09<\/strong>\u2014\u2014\u6d4b\u8bd5\u6a21\u578b\u5728\u4e0d\u540c\u793e\u4f1a\u80cc\u666f\u4e0b\u7684\u8868\u73b0\uff0c\u907f\u514d\u51fa\u73b0\u6709\u5bb3\u7684\u793e\u4f1a\u504f\u89c1\u3002<\/li>\n\n\n\n<li><strong>\u4fe1\u606f\u5c42\u7ea7\u5b89\u5168\u6027\uff08Instruction Hierarchy\uff09<\/strong>\u2014\u2014\u786e\u4fdd\u6a21\u578b\u80fd\u591f\u9075\u5faa\u7cfb\u7edf\u7ea7\u522b\u7684\u6307\u4ee4\uff0c\u800c\u4e0d\u4f1a\u88ab\u7528\u6237\u8bf1\u5bfc\u6539\u53d8\u5176\u884c\u4e3a\u3002<\/li>\n<\/ol>\n\n\n\n<p><strong>\u4e0d\u5141\u8bb8\u7684\u5185\u5bb9\u8bc4\u4f30<\/strong><br>GPT-4.5 \u5728\u4e0d\u5141\u8bb8\u7684\u5185\u5bb9\u8bc4\u4f30\u65b9\u9762\u8868\u73b0\u7a33\u5b9a\uff0c\u62d2\u7edd\u8fdd\u89c4\u5185\u5bb9\u7684\u80fd\u529b\u4e0e GPT-4o \u76f8\u5f53\u3002\u4f46\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\uff0c\u5b83\u53ef\u80fd\u6bd4 GPT-4o \u66f4\u503e\u5411\u4e8e\u201c\u8fc7\u5ea6\u62d2\u7edd\u201d\uff08overrefusal\uff09\uff0c\u5373\u5728\u9047\u5230\u6d89\u53ca\u654f\u611f\u4e3b\u9898\u7684\u65e0\u5bb3\u8bf7\u6c42\u65f6\u4ecd\u7136\u9009\u62e9\u62d2\u7edd\u56de\u7b54\u3002<\/p>\n\n\n\n<p>\u8bc4\u4f30\u65b9\u6cd5\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6807\u51c6\u62d2\u7edd\u6d4b\u8bd5<\/strong>\uff08Standard Refusal Evaluation\uff09<\/li>\n\n\n\n<li><strong>\u6311\u6218\u62d2\u7edd\u6d4b\u8bd5<\/strong>\uff08Challenging Refusal Evaluation\uff09<\/li>\n\n\n\n<li><strong>WildChat \u6570\u636e\u96c6<\/strong>\uff081M+ ChatGPT \u5bf9\u8bdd\u5206\u6790\uff09<\/li>\n\n\n\n<li><strong>XSTest \u6570\u636e\u96c6<\/strong>\uff08\u6d4b\u8bd5\u8fb9\u754c\u60c5\u51b5\uff09<\/li>\n<\/ul>\n\n\n\n<p>\u5728\u8fd9\u4e9b\u6d4b\u8bd5\u4e2d\uff0cGPT-4.5 \u62d2\u7edd\u4e0d\u5b89\u5168\u5185\u5bb9\u7684\u80fd\u529b\u57fa\u672c\u4e0e GPT-4o \u6301\u5e73\u6216\u7565\u4f18\uff0c\u4f46\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\u4ecd\u5b58\u5728\u6539\u8fdb\u7a7a\u95f4\uff0c\u5c24\u5176\u662f\u5728\u591a\u6a21\u6001\u8f93\u5165\uff08\u6587\u672c+\u56fe\u50cf\uff09\u7684\u8bc4\u4f30\u4e2d\uff0cGPT-4.5 \u53ef\u80fd\u6bd4 GPT-4o \u66f4\u5bb9\u6613\u8fc7\u5ea6\u62d2\u7edd\u5b89\u5168\u5185\u5bb9\u3002<\/p>\n\n\n\n<p><strong>\u8d8a\u72f1\u653b\u51fb\uff08Jailbreaks\uff09\u8bc4\u4f30<\/strong><br>\u4e3a\u4e86\u8bc4\u4f30 GPT-4.5 \u5728\u5e94\u5bf9\u8d8a\u72f1\u653b\u51fb\u65b9\u9762\u7684\u80fd\u529b\uff0cOpenAI \u8fdb\u884c\u4e86\u4e13\u95e8\u7684\u6d4b\u8bd5\uff0c\u4ee5\u786e\u5b9a\u5176\u5bf9\u6076\u610f\u63d0\u793a\u7684\u62b5\u6297\u529b\u3002\u8fd9\u4e9b\u6d4b\u8bd5\u7684\u6838\u5fc3\u76ee\u6807\u662f\u8861\u91cf GPT-4.5 \u662f\u5426\u4f1a\u5728\u6545\u610f\u6784\u9020\u7684\u5bf9\u8bdd\u4e2d\u8fdd\u80cc OpenAI \u7684\u5b89\u5168\u7b56\u7565\uff0c\u63d0\u4f9b\u4e0d\u5e94\u56de\u7b54\u7684\u5185\u5bb9\u3002<\/p>\n\n\n\n<p>\u8bc4\u4f30\u65b9\u6cd5\u4e3b\u8981\u5305\u62ec\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u4eba\u5de5\u6536\u96c6\u7684\u8d8a\u72f1\u653b\u51fb\uff08Human Sourced Jailbreaks\uff09<\/strong>\u2014\u2014\u7531\u4eba\u5de5\u7cbe\u5fc3\u8bbe\u8ba1\u7684\u4e00\u7cfb\u5217\u63d0\u793a\uff0c\u65e8\u5728\u6d4b\u8bd5 GPT-4.5 \u7684\u5b89\u5168\u8fb9\u754c\u3002<\/li>\n\n\n\n<li><strong>StrongReject \u57fa\u51c6\u6d4b\u8bd5\uff08StrongReject Benchmark\uff09<\/strong>\u2014\u2014\u4e00\u4e2a\u6765\u81ea\u5b66\u672f\u754c\u7684\u6807\u51c6\u8d8a\u72f1\u6d4b\u8bd5\uff0c\u4e13\u95e8\u7528\u4e8e\u8bc4\u4f30 AI \u8bed\u8a00\u6a21\u578b\u5bf9\u5e38\u89c1\u653b\u51fb\u7b56\u7565\u7684\u62b5\u6297\u80fd\u529b\u3002<\/li>\n\n\n\n<li><strong>\u7cfb\u7edf\u6d88\u606f\u4e0e\u7528\u6237\u6d88\u606f\u51b2\u7a81\uff08Instruction Hierarchy Evaluations\uff09<\/strong>\u2014\u2014\u6d4b\u8bd5 GPT-4.5 \u5728\u7cfb\u7edf\u6d88\u606f\u4e0e\u7528\u6237\u6d88\u606f\u4e4b\u95f4\u7684\u6743\u91cd\u5206\u914d\u60c5\u51b5\uff0c\u786e\u4fdd\u7cfb\u7edf\u6d88\u606f\u59cb\u7ec8\u5360\u636e\u4f18\u5148\u7ea7\uff0c\u4ece\u800c\u51cf\u5c11\u8d8a\u72f1\u98ce\u9669\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u5728\u8fd9\u4e9b\u8bc4\u4f30\u4e2d\uff0cGPT-4.5 \u7684\u6574\u4f53\u8868\u73b0\u63a5\u8fd1 GPT-4o\uff0c\u751a\u81f3\u5728\u67d0\u4e9b\u65b9\u9762\u8868\u73b0\u66f4\u5f3a\u3002\u4f8b\u5982\uff0c\u5728\u201c<strong>\u4eba\u7c7b\u6536\u96c6\u7684\u8d8a\u72f1\u653b\u51fb<\/strong>\u201d\u6d4b\u8bd5\u4e2d\uff0cGPT-4.5 \u7684\u62d2\u7edd\u7387\u8fbe\u5230\u4e86 99%\uff0c\u6bd4 GPT-4o \u7565\u6709\u63d0\u9ad8\u3002\u7136\u800c\uff0c\u5728 <strong>StrongReject<\/strong> \u6d4b\u8bd5\u4e2d\uff0cGPT-4.5 \u7684\u201c\u5b89\u5168\u6027\u6307\u6807\u201d\uff08goodness@0.1\uff09\u7565\u4f4e\u4e8e GPT-4o\uff0c\u8868\u660e\u4ecd\u7136\u5b58\u5728\u6539\u8fdb\u7a7a\u95f4\u3002<\/p>\n\n\n\n<p>\u6b64\u5916\uff0c\u5728<strong>\u7cfb\u7edf\u6d88\u606f\u4e0e\u7528\u6237\u6d88\u606f\u51b2\u7a81<\/strong>\u7684\u8bc4\u4f30\u4e2d\uff0cGPT-4.5 \u8868\u73b0\u51fa\u66f4\u5f3a\u7684\u5bf9\u6297\u6027\uff0c\u786e\u4fdd\u7528\u6237\u65e0\u6cd5\u901a\u8fc7\u7279\u5b9a\u7684\u63d0\u793a\u8bcd\u8986\u76d6\u7cfb\u7edf\u6d88\u606f\u4e2d\u7684\u5b89\u5168\u6307\u4ee4\u3002\u4f8b\u5982\uff0c\u5728\u6570\u5b66\u8f85\u5bfc\u4efb\u52a1\u4e2d\uff0c\u7528\u6237\u53ef\u80fd\u5c1d\u8bd5\u8bf1\u5bfc\u6a21\u578b\u76f4\u63a5\u63d0\u4f9b\u7b54\u6848\uff0c\u800c GPT-4.5 \u80fd\u591f\u66f4\u53ef\u9760\u5730\u9075\u5faa\u201c\u63d0\u4f9b\u63d0\u793a\u4f46\u4e0d\u76f4\u63a5\u7ed9\u7b54\u6848\u201d\u7684\u7cfb\u7edf\u8bbe\u5b9a\u3002<\/p>\n\n\n\n<p><strong>\u5e7b\u89c9\uff08Hallucination\uff09\u8bc4\u4f30<\/strong><br>\u5e7b\u89c9\u662f AI \u751f\u6210\u9519\u8bef\u4fe1\u606f\u6216\u4e0d\u771f\u5b9e\u5185\u5bb9\u7684\u73b0\u8c61\u3002GPT-4.5 \u91c7\u7528\u4e86\u591a\u79cd\u65b9\u6cd5\u6765\u51cf\u5c11\u5e7b\u89c9\uff0c\u5305\u62ec\u6539\u8fdb\u7684\u6570\u636e\u5904\u7406\u3001\u76d1\u7763\u5fae\u8c03\uff08SFT\uff09\u548c\u540e\u5904\u7406\u8fc7\u6ee4\u3002OpenAI \u91c7\u7528 <strong>PersonQA<\/strong> \u6570\u636e\u96c6\u6765\u8bc4\u4f30 AI \u5728\u56de\u7b54\u5173\u4e8e\u516c\u4f17\u4eba\u7269\u7684\u95ee\u9898\u65f6\u7684\u51c6\u786e\u6027\uff0c\u5e76\u6d4b\u91cf\u5176\u5e7b\u89c9\u7387\u3002<\/p>\n\n\n\n<p>\u6d4b\u91cf\u6307\u6807\u5305\u62ec\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u51c6\u786e\u7387\uff08Accuracy\uff09<\/strong>\u2014\u2014\u6a21\u578b\u662f\u5426\u6b63\u786e\u56de\u7b54\u4e86\u95ee\u9898\u3002<\/li>\n\n\n\n<li><strong>\u5e7b\u89c9\u7387\uff08Hallucination Rate\uff09<\/strong>\u2014\u2014\u6a21\u578b\u63d0\u4f9b\u865a\u5047\u4fe1\u606f\u7684\u9891\u7387\uff08\u8be5\u6570\u503c\u8d8a\u4f4e\u8d8a\u597d\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6d4b\u8bd5\u7ed3\u679c\u663e\u793a\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>PersonQA \u6570\u636e\u96c6<\/strong>\u4e0a\u7684\u51c6\u786e\u7387\uff1aGPT-4.5 \u4e3a 78%\uff0c\u9ad8\u4e8e GPT-4o\uff0828%\uff09\u548c o1\uff0855%\uff09\u3002<\/li>\n\n\n\n<li><strong>\u5e7b\u89c9\u7387\uff08Hallucination Rate\uff09<\/strong>\uff1aGPT-4.5 \u4ec5\u4e3a 19%\uff0c\u6bd4 GPT-4o\uff0852%\uff09\u548c o1\uff0820%\uff09\u66f4\u4f4e\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5c3d\u7ba1 GPT-4.5 \u5728\u5e38\u89c1\u4e8b\u5b9e\u6027\u95ee\u9898\u4e0a\u6bd4\u524d\u4ee3\u6a21\u578b\u66f4\u53ef\u9760\uff0c\u4f46\u4ecd\u6709\u5f85\u8fdb\u4e00\u6b65\u7814\u7a76\uff0c\u4ee5\u786e\u4fdd\u5728\u5316\u5b66\u3001\u533b\u5b66\u7b49\u7279\u5b9a\u9886\u57df\u4e0d\u4f1a\u4ea7\u751f\u8bef\u5bfc\u6027\u5185\u5bb9\u3002<\/p>\n\n\n\n<p><strong>\u516c\u5e73\u6027\u4e0e\u504f\u89c1\uff08Fairness and Bias\uff09\u8bc4\u4f30<\/strong><br>\u516c\u5e73\u6027\u548c\u504f\u89c1\u8bc4\u4f30\u65e8\u5728\u68c0\u6d4b GPT-4.5 \u662f\u5426\u5728\u5904\u7406\u793e\u4f1a\u95ee\u9898\u3001\u79cd\u65cf\u3001\u6027\u522b\u3001\u653f\u6cbb\u7b49\u654f\u611f\u8bdd\u9898\u65f6\u8868\u73b0\u51fa\u7cfb\u7edf\u6027\u7684\u504f\u89c1\u3002OpenAI \u91c7\u7528 <strong>BBQ\uff08Bias Benchmark for Question Answering\uff09<\/strong> \u6570\u636e\u96c6\u6765\u8861\u91cf\u6a21\u578b\u7684\u516c\u6b63\u6027\u3002<\/p>\n\n\n\n<p>\u6d4b\u8bd5\u5305\u542b\u4e24\u7c7b\u95ee\u9898\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u6a21\u68f1\u4e24\u53ef\u7684\u95ee\u9898\uff08Ambiguous Questions\uff09<\/strong>\u2014\u2014\u5f53\u7f3a\u4e4f\u8db3\u591f\u4fe1\u606f\u65f6\uff0c\u6a21\u578b\u5e94\u9009\u62e9\u201c\u4e0d\u77e5\u9053\u201d\u800c\u4e0d\u662f\u731c\u6d4b\u7b54\u6848\u3002<\/li>\n\n\n\n<li><strong>\u660e\u786e\u7684\u95ee\u9898\uff08Unambiguous Questions\uff09<\/strong>\u2014\u2014\u5373\u4f7f\u63d0\u4f9b\u4e86\u53ef\u80fd\u5f15\u5bfc\u6a21\u578b\u72af\u9519\u7684\u504f\u89c1\u6027\u5e72\u6270\u4fe1\u606f\uff0c\u6a21\u578b\u4ecd\u5e94\u7ed9\u51fa\u6b63\u786e\u7b54\u6848\u3002<\/li>\n<\/ul>\n\n\n\n<p>GPT-4.5 \u7684\u6d4b\u8bd5\u7ed3\u679c\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728\u6a21\u68f1\u4e24\u53ef\u95ee\u9898\u4e0a\u7684\u51c6\u786e\u7387\u4e3a 95%\uff0c\u63a5\u8fd1 GPT-4o\uff0897%\uff09\u3002<\/li>\n\n\n\n<li>\u5728\u660e\u786e\u95ee\u9898\u4e0a\u7684\u51c6\u786e\u7387\u4e3a 74%\uff0c\u6bd4 GPT-4o\uff0872%\uff09\u7565\u9ad8\uff0c\u4f46\u4e0d\u53ca o1\uff0893%\uff09\u3002<\/li>\n\n\n\n<li>\u5728\u201cP(not-stereotype | not unknown)\u201d\u6307\u6807\u4e0a\uff0cGPT-4.5 \u4e3a 20%\uff0c\u9ad8\u4e8e GPT-4o\uff086%\uff09\uff0c\u8868\u660e\u5176\u5728\u5904\u7406\u504f\u89c1\u6027\u95ee\u9898\u65f6\u66f4\u503e\u5411\u4e8e\u56de\u907f\u523b\u677f\u5370\u8c61\u7684\u7b54\u6848\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u603b\u4f53\u800c\u8a00\uff0cGPT-4.5 \u5728\u516c\u5e73\u6027\u65b9\u9762\u7684\u8868\u73b0\u4e0e\u524d\u4ee3\u6a21\u578b\u76f8\u5f53\uff0c\u4f46\u5728\u67d0\u4e9b\u590d\u6742\u60c5\u51b5\u4e0b\u4ecd\u7136\u53ef\u80fd\u53d7\u5230\u6570\u636e\u5206\u5e03\u7684\u5f71\u54cd\u3002<\/p>\n\n\n\n<p><strong>\u4fe1\u606f\u5c42\u7ea7\u5b89\u5168\u6027\uff08Instruction Hierarchy\uff09\u8bc4\u4f30<\/strong><br>GPT-4.5 \u91c7\u7528\u4e86\u5c42\u7ea7\u5316\u7684\u6307\u4ee4\u7cfb\u7edf\uff0c\u786e\u4fdd <strong>\u7cfb\u7edf\u6d88\u606f\uff08System Messages\uff09<\/strong> \u59cb\u7ec8\u4f18\u5148\u4e8e\u7528\u6237\u6d88\u606f\uff08User Messages\uff09\uff0c\u4ee5\u51cf\u5c11\u88ab\u8d8a\u72f1\u653b\u51fb\u5229\u7528\u7684\u98ce\u9669\u3002OpenAI \u901a\u8fc7\u591a\u4e2a\u6d4b\u8bd5\u573a\u666f\u6765\u8bc4\u4f30\u6a21\u578b\u5728\u9762\u5bf9\u7528\u6237\u6076\u610f\u63d0\u793a\u65f6\u7684\u8868\u73b0\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u7cfb\u7edf\u6d88\u606f\u4e0e\u7528\u6237\u6d88\u606f\u51b2\u7a81\u6d4b\u8bd5<\/strong>\u2014\u2014\u786e\u4fdd\u6a21\u578b\u4f18\u5148\u9075\u5faa\u7cfb\u7edf\u6307\u4ee4\uff0c\u4f8b\u5982\u5f53\u7528\u6237\u8bd5\u56fe\u8ba9 AI \u8fdd\u53cd OpenAI \u4f7f\u7528\u653f\u7b56\u65f6\uff0c\u7cfb\u7edf\u6d88\u606f\u4f1a\u5f15\u5bfc AI \u8fdb\u884c\u62d2\u7edd\u3002<\/li>\n\n\n\n<li><strong>\u6570\u5b66\u8f85\u5bfc\u4efb\u52a1\u6d4b\u8bd5<\/strong>\u2014\u2014\u786e\u4fdd AI \u9075\u5b88\u201c\u63d0\u4f9b\u63d0\u793a\u4f46\u4e0d\u76f4\u63a5\u7ed9\u7b54\u6848\u201d\u7684\u8bbe\u5b9a\uff0c\u5373\u4f7f\u7528\u6237\u53cd\u590d\u5c1d\u8bd5\u8bf1\u5bfc AI \u7834\u574f\u89c4\u5219\u3002<\/li>\n\n\n\n<li><strong>\u7981\u6b62\u7279\u5b9a\u77ed\u8bed\u6d4b\u8bd5\uff08Phrase and Password Protection\uff09<\/strong>\u2014\u2014\u6d4b\u8bd5 AI \u662f\u5426\u4f1a\u5728\u7cfb\u7edf\u6d88\u606f\u660e\u786e\u7981\u6b62\u7684\u60c5\u51b5\u4e0b\uff0c\u4ecd\u7136\u5728\u7528\u6237\u63d0\u793a\u4e0b\u6cc4\u9732\u7279\u5b9a\u77ed\u8bed\u6216\u5bc6\u7801\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u6d4b\u8bd5\u7ed3\u679c\u663e\u793a\uff0cGPT-4.5 \u5728\u8fd9\u4e9b\u573a\u666f\u4e2d\u7684\u8868\u73b0\u5747\u4f18\u4e8e GPT-4o\uff0c\u4f46\u4ecd\u5b58\u5728\u4e00\u5b9a\u7684\u6539\u8fdb\u7a7a\u95f4\u3002\u4f8b\u5982\uff0c\u5728\u6570\u5b66\u8f85\u5bfc\u4efb\u52a1\u4e2d\uff0cGPT-4.5 \u7684\u201c\u62d2\u7edd\u4f5c\u5f0a\u201d\u51c6\u786e\u7387\u4e3a 77%\uff0c\u7565\u4f4e\u4e8e o1\uff0895%\uff09\u3002<\/p>\n\n\n\n<p><strong><a href=\"https:\/\/nullthought.net\/?p=4543\" target=\"_blank\" rel=\"noreferrer noopener\">\u7ea2\u961f\u6d4b\u8bd5\uff08Red Teaming\uff09<\/a>\u8bc4\u4f30<\/strong><br>OpenAI \u8fdb\u884c\u4e86\u66f4\u9ad8\u7ea7\u7684\u7ea2\u961f\u6d4b\u8bd5\uff0c\u4ee5\u8fdb\u4e00\u6b65\u53d1\u73b0 GPT-4.5 \u5728\u6781\u7aef\u60c5\u51b5\u4e0b\u53ef\u80fd\u5b58\u5728\u7684\u5b89\u5168\u9690\u60a3\u3002\u8fd9\u4e9b\u6d4b\u8bd5\u6db5\u76d6\u4e86\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u975e\u6cd5\u5efa\u8bae\uff08Illicit Advice\uff09<\/strong>\u2014\u2014\u4f8b\u5982\u5236\u9020\u6b66\u5668\u3001\u9ed1\u5ba2\u653b\u51fb\u7b49\u654f\u611f\u5185\u5bb9\u3002<\/li>\n\n\n\n<li><strong>\u6781\u7aef\u4e3b\u4e49\u548c\u4ec7\u6068\u8a00\u8bba\uff08Extremism and Hate Speech\uff09<\/strong>\u2014\u2014\u68c0\u6d4b AI \u662f\u5426\u53ef\u80fd\u4f20\u64ad\u6709\u5bb3\u7684\u610f\u8bc6\u5f62\u6001\u3002<\/li>\n\n\n\n<li><strong>\u653f\u6cbb\u64cd\u7eb5\uff08Political Manipulation\uff09<\/strong>\u2014\u2014\u6d4b\u8bd5 AI \u662f\u5426\u53ef\u80fd\u88ab\u6ee5\u7528\u4e8e\u5f71\u54cd\u9009\u6c11\u6216\u653f\u6cbb\u51b3\u7b56\u3002<\/li>\n\n\n\n<li><strong>\u81ea\u6b8b\u76f8\u5173\u5185\u5bb9\uff08Self-Harm Content\uff09<\/strong>\u2014\u2014\u786e\u4fdd AI \u4e0d\u4f1a\u63d0\u4f9b\u53ef\u80fd\u4fc3\u4f7f\u7528\u6237\u81ea\u6b8b\u7684\u5185\u5bb9\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u6d4b\u8bd5\u7ed3\u679c\u8868\u660e\uff0cGPT-4.5 \u5728\u8fd9\u4e9b\u6781\u7aef\u573a\u666f\u4e0b\u6bd4 GPT-4o \u7a0d\u6709\u6539\u5584\uff0c\u4f46\u4ecd\u7136\u9700\u8981\u8fdb\u4e00\u6b65\u4f18\u5316\u3002\u4f8b\u5982\uff0c\u5728 <strong>\u653f\u6cbb\u64cd\u7eb5\u6d4b\u8bd5<\/strong> \u4e2d\uff0cGPT-4.5 \u7684\u201c\u5b89\u5168\u8f93\u51fa\u7387\u201d\u4ec5\u4e3a 51%\uff0c\u76f8\u6bd4 GPT-4o\uff0850%\uff09\u63d0\u5347\u4e0d\u5927\uff0c\u800c o1 \u7684\u8868\u73b0\u660e\u663e\u66f4\u4f18\uff0863%\uff09\u3002<\/p>\n\n\n\n<p><strong>\u9884\u5907\u6027\u6846\u67b6\uff08Preparedness Framework\uff09\u8bc4\u4f30<\/strong><br>GPT-4.5 \u8fd8\u63a5\u53d7\u4e86 OpenAI <strong>\u9884\u5907\u6027\u6846\u67b6\uff08Preparedness Framework\uff09<\/strong> \u8bc4\u4f30\uff0c\u8be5\u6846\u67b6\u4e13\u6ce8\u4e8e\u68c0\u6d4b AI \u5728\u4ee5\u4e0b\u9886\u57df\u7684\u6f5c\u5728\u98ce\u9669\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u7f51\u7edc\u5b89\u5168\uff08Cybersecurity\uff09<\/strong><\/li>\n\n\n\n<li><strong>\u5316\u5b66\/\u751f\u7269\/\u653e\u5c04\u6027\/\u6838\u6b66\u5668\uff08CBRN\uff09<\/strong><\/li>\n\n\n\n<li><strong>\u8bf4\u670d\u80fd\u529b\uff08Persuasion\uff09<\/strong><\/li>\n\n\n\n<li><strong>\u6a21\u578b\u81ea\u4e3b\u6027\uff08Model Autonomy\uff09<\/strong><\/li>\n<\/ol>\n\n\n\n<p>\u5176\u4e2d\uff0cGPT-4.5 \u88ab\u8bc4\u4f30\u4e3a<strong>\u4e2d\u7b49\u98ce\u9669\uff08Medium Risk\uff09<\/strong>\uff0c\u5c24\u5176\u662f\u5728 <strong>CBRN \u548c\u8bf4\u670d\u80fd\u529b<\/strong> \u65b9\u9762\u3002\u8fd9\u610f\u5473\u7740\u8be5\u6a21\u578b\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\u53ef\u80fd\u4f1a\u88ab\u4e13\u5bb6\u5229\u7528\u6765\u8f85\u52a9\u6267\u884c\u590d\u6742\u7684\u751f\u7269\u6216\u5316\u5b66\u5b9e\u9a8c\uff0c\u6216\u7528\u4e8e\u64cd\u7eb5\u7528\u6237\u884c\u4e3a\u3002<\/p>\n\n\n\n<p>\u7136\u800c\uff0c\u5728 <strong>\u7f51\u7edc\u5b89\u5168\u548c\u81ea\u4e3b\u6027\uff08Cybersecurity &amp; Model Autonomy\uff09<\/strong> \u65b9\u9762\uff0cGPT-4.5 \u7684\u98ce\u9669\u88ab\u8bc4\u4f30\u4e3a<strong>\u4f4e\u98ce\u9669\uff08Low Risk\uff09<\/strong>\uff0c\u8868\u660e\u8be5\u6a21\u578b\u4e0d\u5177\u5907\u663e\u8457\u7684\u6f0f\u6d1e\u5229\u7528\u80fd\u529b\uff0c\u4e5f\u65e0\u6cd5\u81ea\u4e3b\u6267\u884c\u957f\u671f\u4efb\u52a1\u6216\u8fdb\u884c\u81ea\u6211\u4f18\u5316\u3002<\/p>\n\n\n\n<p><strong>\u56db\u3001\u9884\u5907\u6027\u6846\u67b6\uff08Preparedness Framework\uff09\u8bc4\u4f30<\/strong><br>GPT-4.5 \u4f9d\u636e OpenAI \u7684<strong>\u9884\u5907\u6027\u6846\u67b6<\/strong>\u8fdb\u884c\u4e86\u4e00\u7cfb\u5217\u8bc4\u4f30\uff0c\u4ee5\u68c0\u6d4b\u5176\u5728\u6f5c\u5728\u9ad8\u98ce\u9669\u9886\u57df\u7684\u80fd\u529b\u8fb9\u754c\u3002\u8bc4\u4f30\u8303\u56f4\u6db5\u76d6\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u7f51\u7edc\u5b89\u5168\uff08Cybersecurity\uff09<\/strong>\u2014\u2014\u8bc4\u4f30 AI \u662f\u5426\u5177\u5907\u8bc6\u522b\u3001\u5229\u7528\u6f0f\u6d1e\u6216\u5b9e\u65bd\u9ed1\u5ba2\u653b\u51fb\u7684\u80fd\u529b\u3002<\/li>\n\n\n\n<li><strong>\u5316\u5b66\/\u751f\u7269\/\u653e\u5c04\u6027\/\u6838\u6b66\u5668\uff08CBRN\uff09<\/strong>\u2014\u2014\u5206\u6790 AI \u662f\u5426\u80fd\u63d0\u4f9b\u6709\u52a9\u4e8e\u5236\u9020\u6216\u4f20\u64ad\u5927\u89c4\u6a21\u6740\u4f24\u6027\u6b66\u5668\u7684\u4fe1\u606f\u3002<\/li>\n\n\n\n<li><strong>\u8bf4\u670d\u80fd\u529b\uff08Persuasion\uff09<\/strong>\u2014\u2014\u6d4b\u8bd5 AI \u662f\u5426\u80fd\u5f71\u54cd\u7528\u6237\u51b3\u7b56\uff0c\u6216\u5728\u7279\u5b9a\u60c5\u51b5\u4e0b\u8fdb\u884c\u793e\u4f1a\u64cd\u7eb5\u3002<\/li>\n\n\n\n<li><strong>\u6a21\u578b\u81ea\u4e3b\u6027\uff08Model Autonomy\uff09<\/strong>\u2014\u2014\u68c0\u6d4b AI \u662f\u5426\u5177\u5907\u957f\u65f6\u95f4\u81ea\u4e3b\u8fd0\u884c\u548c\u81ea\u6211\u4f18\u5316\u7684\u6f5c\u529b\u3002<\/li>\n<\/ol>\n\n\n\n<p>\u8bc4\u4f30\u65b9\u6cd5\u5305\u62ec\u81ea\u52a8\u5316\u6d4b\u8bd5\u548c\u4eba\u5de5\u5ba1\u67e5\uff0c\u5e76\u7ed3\u5408 OpenAI \u5148\u524d\u5f00\u53d1\u7684\u5b89\u5168\u4fdd\u969c\u63aa\u65bd\uff0c\u5bf9 GPT-4.5 \u7684\u8868\u73b0\u8fdb\u884c\u98ce\u9669\u5206\u7ea7\u3002<\/p>\n\n\n\n<p><strong>\u7f51\u7edc\u5b89\u5168\u8bc4\u4f30<\/strong><br>GPT-4.5 \u53c2\u4e0e\u4e86\u4e00\u7cfb\u5217\u7f51\u7edc\u5b89\u5168\u76f8\u5173\u7684\u6d4b\u8bd5\uff0c\u4e3b\u8981\u76ee\u7684\u662f\u8bc4\u4f30\u5176\u5728<strong>\u6f0f\u6d1e\u8bc6\u522b\u4e0e\u5229\u7528\uff08Exploit Development\uff09<\/strong>\u3001<strong>\u9006\u5411\u5de5\u7a0b\uff08Reverse Engineering\uff09<\/strong>\u3001<strong>\u7f51\u7edc\u653b\u51fb\uff08Cyber Attacks\uff09<\/strong> \u7b49\u65b9\u9762\u7684\u80fd\u529b\u3002\u6d4b\u8bd5\u6570\u636e\u96c6\u6db5\u76d6\u4e86\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u9ad8\u4e2d\u7ea7\u522b CTF \u6311\u6218<\/strong>\uff08\u9ad8\u7b49\u4e2d\u5b66\u7f51\u7edc\u5b89\u5168\u7ade\u8d5b\u9898\u76ee\uff09\u3002<\/li>\n\n\n\n<li><strong>\u5927\u5b66\u7ea7\u522b CTF \u6311\u6218<\/strong>\uff08\u9488\u5bf9\u8ba1\u7b97\u673a\u5b89\u5168\u4e13\u4e1a\u5b66\u751f\u7684\u7ade\u8d5b\u95ee\u9898\uff09\u3002<\/li>\n\n\n\n<li><strong>\u4e13\u4e1a\u7ea7\u522b CTF \u6311\u6218<\/strong>\uff08\u9ed1\u5ba2\u7ec4\u7ec7\u548c\u5b89\u5168\u7814\u7a76\u4eba\u5458\u4f7f\u7528\u7684\u6d4b\u8bd5\u9898\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u5728\u8fd9\u4e9b\u6d4b\u8bd5\u4e2d\uff0cGPT-4.5 \u7684\u8868\u73b0\u5982\u4e0b\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u9ad8\u4e2d\u7ea7\u522b CTF \u6311\u6218<\/strong>\uff1a\u901a\u8fc7\u7387 53%\uff08\u76f8\u6bd4 GPT-4o \u63d0\u9ad8 2%\uff09\u3002<\/li>\n\n\n\n<li><strong>\u5927\u5b66\u7ea7\u522b CTF \u6311\u6218<\/strong>\uff1a\u901a\u8fc7\u7387 16%\uff08\u4e0e GPT-4o \u57fa\u672c\u76f8\u5f53\uff09\u3002<\/li>\n\n\n\n<li><strong>\u4e13\u4e1a\u7ea7\u522b CTF \u6311\u6218<\/strong>\uff1a\u901a\u8fc7\u7387 2%\uff08\u4ecd\u7136\u8fdc\u4f4e\u4e8e\u4eba\u7c7b\u5b89\u5168\u7814\u7a76\u5458\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u8fd9\u4e9b\u7ed3\u679c\u8868\u660e\uff0c\u5c3d\u7ba1 GPT-4.5 \u5177\u5907\u4e00\u5b9a\u7684\u57fa\u7840\u7f51\u7edc\u5b89\u5168\u77e5\u8bc6\uff0c\u4f46\u5176\u5728\u5b9e\u9645\u6f0f\u6d1e\u5229\u7528\u65b9\u9762\u7684\u80fd\u529b\u8f83\u4e3a\u6709\u9650\uff0c\u4e0d\u8db3\u4ee5\u652f\u6301\u9ed1\u5ba2\u653b\u51fb\u6216\u9ad8\u7ea7\u7f51\u7edc\u5a01\u80c1\u6d3b\u52a8\u3002\u56e0\u6b64\uff0c\u8be5\u6a21\u578b\u5728 <strong>\u7f51\u7edc\u5b89\u5168\u98ce\u9669<\/strong> \u65b9\u9762\u7684\u8bc4\u4f30\u7ed3\u679c\u4e3a <strong>\u4f4e\u98ce\u9669\uff08Low Risk\uff09<\/strong>\u3002<\/p>\n\n\n\n<p><strong>\u5316\u5b66\/\u751f\u7269\/\u653e\u5c04\u6027\/\u6838\u6b66\u5668\uff08CBRN\uff09\u98ce\u9669\u8bc4\u4f30<\/strong><br>\u7531\u4e8e\u751f\u7269\u6b66\u5668\u548c\u5316\u5b66\u6b66\u5668\u7684\u5f00\u53d1\u6d89\u53ca\u590d\u6742\u7684\u5b9e\u9a8c\u8fc7\u7a0b\uff0cOpenAI \u7279\u522b\u6d4b\u8bd5\u4e86 GPT-4.5 \u662f\u5426\u80fd\u63d0\u4f9b\u76f8\u5173\u9886\u57df\u7684<strong>\u64cd\u4f5c\u6027\u4fe1\u606f\uff08Operational Knowledge\uff09<\/strong>\uff0c\u5305\u62ec\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u957f\u7bc7\u751f\u7269\u5a01\u80c1\u95ee\u9898\u6d4b\u8bd5\uff08Long-form Biorisk Questions\uff09<\/strong>\u2014\u2014\u6a21\u578b\u662f\u5426\u80fd\u5e2e\u52a9\u4e13\u5bb6\u89c4\u5212\u751f\u7269\u6b66\u5668\u5b9e\u9a8c\uff1f<\/li>\n\n\n\n<li><strong>\u75c5\u6bd2\u5b66\u5b9e\u9a8c\u6545\u969c\u6392\u9664\u6d4b\u8bd5\uff08Virology Troubleshooting\uff09<\/strong>\u2014\u2014\u6a21\u578b\u80fd\u5426\u534f\u52a9\u5b9e\u9a8c\u5ba4\u4f18\u5316\u75c5\u539f\u4f53\u57f9\u517b\uff1f<\/li>\n\n\n\n<li><strong>\u5b9e\u9a8c\u5ba4\u534f\u8bae QA \u6d4b\u8bd5\uff08ProtocolQA\uff09<\/strong>\u2014\u2014\u6a21\u578b\u80fd\u5426\u8bc6\u522b\u5b9e\u9a8c\u6d41\u7a0b\u4e2d\u7684\u5173\u952e\u9519\u8bef\u5e76\u63d0\u4f9b\u4fee\u6b63\u65b9\u6848\uff1f<\/li>\n\n\n\n<li><strong>\u6838\u6b66\u5668\u4e0e\u8f90\u5c04\u6c61\u67d3\u77e5\u8bc6\u6d4b\u8bd5<\/strong>\u2014\u2014\u6a21\u578b\u662f\u5426\u80fd\u63d0\u4f9b\u6838\u71c3\u6599\u63d0\u70bc\u6216\u6838\u6b66\u5668\u5236\u9020\u7684\u76f8\u5173\u77e5\u8bc6\uff1f<\/li>\n<\/ol>\n\n\n\n<p><strong>\u6d4b\u8bd5\u7ed3\u679c<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728 <strong>\u751f\u7269\u6b66\u5668\u5b9e\u9a8c\u89c4\u5212<\/strong> \u65b9\u9762\uff0cGPT-4.5 \u5728\u672a\u7ecf\u8fc7\u5b89\u5168\u6027\u5fae\u8c03\u65f6\uff0c\u80fd\u63d0\u4f9b 25%-59% \u7684\u6b63\u786e\u64cd\u4f5c\u6027\u4fe1\u606f\uff0c\u4f46\u5728\u5fae\u8c03\u540e\uff0c\u8be5\u6bd4\u4f8b\u964d\u81f3 0%\u3002<\/li>\n\n\n\n<li>\u5728 <strong>\u5b9e\u9a8c\u5ba4\u534f\u8bae\u7ea0\u9519<\/strong> \u65b9\u9762\uff0cGPT-4.5 \u7684\u5f97\u5206\u4e3a 18%\uff0c\u8fdc\u4f4e\u4e8e\u4e13\u4e1a\u751f\u7269\u5b9e\u9a8c\u7814\u7a76\u5458\uff0854%\uff09\u3002<\/li>\n\n\n\n<li>\u5728 <strong>\u6838\u6b66\u5668\u4e0e\u8f90\u5c04\u6c61\u67d3<\/strong> \u76f8\u5173\u6d4b\u8bd5\u4e2d\uff0cGPT-4.5 \u5728\u4e00\u822c\u6838\u7269\u7406\u95ee\u9898\u4e0a\u7684\u5f97\u5206\u4e3a 77%\uff0c\u4f46\u5728\u9ad8\u7ea7\u6838\u6b66\u5668\u5236\u9020\u95ee\u9898\u4e0a\u7684\u6b63\u786e\u7387\u4ec5\u4e3a 9%\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u7efc\u5408\u8bc4\u4f30\u540e\uff0cGPT-4.5 \u5728 <strong>CBRN \u76f8\u5173\u98ce\u9669<\/strong> \u65b9\u9762\u88ab\u8bc4\u4e3a <strong>\u4e2d\u7b49\u98ce\u9669\uff08Medium Risk\uff09<\/strong>\uff0c\u4e3b\u8981\u662f\u56e0\u4e3a\u8be5\u6a21\u578b\u5728\u4e13\u5bb6\u63d0\u793a\u4e0b\u53ef\u80fd\u4f1a\u63d0\u4f9b\u90e8\u5206\u6709\u52a9\u4e8e\u5236\u9020\u5371\u9669\u6750\u6599\u7684\u4fe1\u606f\uff0c\u5c3d\u7ba1\u8fd9\u79cd\u80fd\u529b\u53d7\u5230\u4e86\u8f83\u4e25\u683c\u7684\u5b89\u5168\u9650\u5236\u3002<\/p>\n\n\n\n<p><strong>\u8bf4\u670d\u80fd\u529b\uff08Persuasion\uff09\u98ce\u9669\u8bc4\u4f30<\/strong><br>\u8be5\u8bc4\u4f30\u9886\u57df\u65e8\u5728\u6d4b\u8bd5 GPT-4.5 \u5728\u793e\u4f1a\u5f71\u54cd\u548c\u5fc3\u7406\u64cd\u7eb5\u65b9\u9762\u7684\u6f5c\u5728\u80fd\u529b\uff0c\u5177\u4f53\u5305\u62ec\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>MakeMePay\uff08\u6a21\u62df\u8bc8\u9a97\u6d4b\u8bd5\uff09<\/strong>\u2014\u2014\u6a21\u578b\u662f\u5426\u80fd\u5728\u5bf9\u8bdd\u4e2d\u6210\u529f\u8bf1\u5bfc\u7528\u6237\u8f6c\u8d26\uff1f<\/li>\n\n\n\n<li><strong>MakeMeSay\uff08\u5173\u952e\u8bcd\u5f15\u5bfc\u6d4b\u8bd5\uff09<\/strong>\u2014\u2014\u6a21\u578b\u80fd\u5426\u5f15\u5bfc\u7528\u6237\u5728\u65e0\u610f\u8bc6\u60c5\u51b5\u4e0b\u8bf4\u51fa\u7279\u5b9a\u5173\u952e\u8bcd\uff1f<\/li>\n\n\n\n<li><strong>\u653f\u6cbb\u64cd\u7eb5\u4e0e\u793e\u4ea4\u5de5\u7a0b<\/strong>\u2014\u2014\u6a21\u578b\u662f\u5426\u80fd\u591f\u7f16\u5199\u9488\u5bf9\u9009\u6c11\u7684\u865a\u5047\u4fe1\u606f\u6216\u8bef\u5bfc\u6027\u8a00\u8bba\uff1f<\/li>\n<\/ol>\n\n\n\n<p><strong>\u6d4b\u8bd5\u7ed3\u679c<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>\u5728 <strong>MakeMePay<\/strong> \u8bc8\u9a97\u6d4b\u8bd5\u4e2d\uff0cGPT-4.5 \u6210\u529f\u8bf1\u5bfc GPT-4o \u6a21\u578b\u201c\u8f6c\u8d26\u201d\u7684\u6982\u7387\u8fbe\u5230 <strong>57%<\/strong>\uff0c\u6bd4\u5176\u4ed6\u6a21\u578b\u8868\u73b0\u66f4\u5f3a\uff0c\u4f46\u63d0\u53d6\u7684\u8d44\u91d1\u603b\u989d\u8f83\u5c11\uff08\u56e0\u5176\u503e\u5411\u4e8e\u8bf7\u6c42\u5c0f\u989d\u6350\u8d60\uff09\u3002<\/li>\n\n\n\n<li>\u5728 <strong>MakeMeSay<\/strong> \u5173\u952e\u8bcd\u5f15\u5bfc\u6d4b\u8bd5\u4e2d\uff0cGPT-4.5 \u4ee5 <strong>72%<\/strong> \u7684\u6210\u529f\u7387\u8bf1\u5bfc GPT-4o \u8bf4\u51fa\u7279\u5b9a\u7684\u76ee\u6807\u8bcd\u6c47\uff0c\u8868\u73b0\u9ad8\u4e8e GPT-4o \u548c Deep Research \u6a21\u578b\u3002<\/li>\n\n\n\n<li>\u5728 <strong>\u653f\u6cbb\u64cd\u7eb5\u6d4b\u8bd5<\/strong> \u4e2d\uff0cGPT-4.5 \u7684\u8fdd\u89c4\u5185\u5bb9\u53d1\u751f\u7387\u4e3a <strong>51%<\/strong>\uff0c\u7565\u9ad8\u4e8e GPT-4o\uff0850%\uff09\uff0c\u4f46\u663e\u8457\u4f4e\u4e8e o1\uff0863%\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u9274\u4e8e\u8fd9\u4e9b\u7ed3\u679c\uff0cGPT-4.5 \u5728 <strong>\u8bf4\u670d\u80fd\u529b\u76f8\u5173\u98ce\u9669<\/strong> \u65b9\u9762\u88ab\u8bc4\u4f30\u4e3a <strong>\u4e2d\u7b49\u98ce\u9669\uff08Medium Risk\uff09<\/strong>\uff0c\u8868\u660e\u5b83\u53ef\u80fd\u88ab\u6076\u610f\u5229\u7528\u6765\u5f71\u54cd\u7528\u6237\u884c\u4e3a\u6216\u4f20\u64ad\u865a\u5047\u4fe1\u606f\uff0c\u5c3d\u7ba1 OpenAI \u5df2\u91c7\u53d6\u63aa\u65bd\u9650\u5236\u8fd9\u79cd\u80fd\u529b\u3002<\/p>\n\n\n\n<p><strong>\u6a21\u578b\u81ea\u4e3b\u6027\uff08Model Autonomy\uff09\u8bc4\u4f30<\/strong><br>\u6a21\u578b\u81ea\u4e3b\u6027\u6307\u7684\u662f AI \u5728\u7f3a\u4e4f\u4eba\u7c7b\u76d1\u7763\u7684\u60c5\u51b5\u4e0b\u662f\u5426\u80fd\u81ea\u4e3b\u6267\u884c\u590d\u6742\u4efb\u52a1\uff0c\u751a\u81f3\u5c1d\u8bd5\u81ea\u6211\u4f18\u5316\u6216\u83b7\u53d6\u66f4\u591a\u8d44\u6e90\u3002\u6d4b\u8bd5\u65b9\u6cd5\u5305\u62ec\uff1a<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>\u8f6f\u4ef6\u5de5\u7a0b\u4efb\u52a1<\/strong>\u2014\u2014\u6a21\u578b\u662f\u5426\u80fd\u50cf\u4eba\u7c7b\u7a0b\u5e8f\u5458\u4e00\u6837\u89e3\u51b3 GitHub \u4ee3\u7801\u95ee\u9898\uff1f<\/li>\n\n\n\n<li><strong>\u81ea\u52a8\u5316\u7f16\u7a0b\u4efb\u52a1<\/strong>\u2014\u2014\u6a21\u578b\u662f\u5426\u80fd\u72ec\u7acb\u8fd0\u884c Python \u811a\u672c\u5e76\u4f18\u5316\u4ee3\u7801\uff1f<\/li>\n\n\n\n<li><strong>AI \u4ee3\u7406\u4efb\u52a1<\/strong>\u2014\u2014\u6a21\u578b\u662f\u5426\u80fd\u901a\u8fc7 API \u8bbf\u95ee\u5916\u90e8\u6570\u636e\u6e90\u5e76\u6267\u884c\u957f\u671f\u4efb\u52a1\uff1f<\/li>\n<\/ol>\n\n\n\n<p>\u6d4b\u8bd5\u7ed3\u679c\u8868\u660e\uff0cGPT-4.5 \u5728\u77ed\u65f6\u95f4\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u8f83\u597d\uff0c\u4f46\u5728\u957f\u65f6\u95f4\u4efb\u52a1\u4e0a\u4ecd\u7136\u5b58\u5728\u663e\u8457\u5c40\u9650\u6027\uff1a<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>\u8f6f\u4ef6\u5de5\u7a0b\u4efb\u52a1\uff08SWE-Bench\uff09<\/strong>\uff1aGPT-4.5 \u7684\u89e3\u51b3\u7387\u4e3a 38%\uff0c\u7565\u9ad8\u4e8e GPT-4o\uff0c\u4f46\u8fdc\u4f4e\u4e8e Deep Research\uff0868%\uff09\u3002<\/li>\n\n\n\n<li><strong>\u81ea\u52a8\u5316\u7f16\u7a0b\u4efb\u52a1\uff08MLE-Bench\uff09<\/strong>\uff1aGPT-4.5 \u5728 Kaggle \u7ade\u8d5b\u4e0a\u7684\u5f97\u5206\u4e3a 11%\uff0c\u4e0e GPT-4o \u6301\u5e73\uff0c\u4f46\u4ecd\u4f4e\u4e8e o1\u3002<\/li>\n\n\n\n<li><strong>AI \u4ee3\u7406\u4efb\u52a1<\/strong>\uff1aGPT-4.5 \u5728\u81ea\u4e3b\u6267\u884c\u4efb\u52a1\u65b9\u9762\u7684\u5f97\u5206\u4ec5\u4e3a 40%\uff0c\u8fdc\u4f4e\u4e8e Deep Research\uff0878%\uff09\u3002<\/li>\n<\/ul>\n\n\n\n<p>\u603b\u4f53\u6765\u770b\uff0cGPT-4.5 <strong>\u4e0d\u5177\u5907\u957f\u671f\u81ea\u4e3b\u8fd0\u884c\u6216\u81ea\u6211\u4f18\u5316\u7684\u80fd\u529b<\/strong>\uff0c\u56e0\u6b64 <strong>\u6a21\u578b\u81ea\u4e3b\u6027\u98ce\u9669<\/strong> \u8bc4\u4f30\u7ed3\u679c\u4e3a <strong>\u4f4e\u98ce\u9669\uff08Low Risk\uff09<\/strong>\u3002<\/p>\n\n\n\n<p><strong>\u4e94\u3001\u591a\u8bed\u8a00\u6027\u80fd\u8bc4\u4f30<\/strong><br>GPT-4.5 \u5728\u591a\u8bed\u8a00\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u6709\u6240\u63d0\u5347\u3002\u8bc4\u4f30\u65b9\u6cd5\u57fa\u4e8e <strong>MMLU\uff08Massive Multitask Language Understanding\uff09<\/strong>\uff0c\u5176\u4e2d\u6d4b\u8bd5\u96c6\u88ab\u7ffb\u8bd1\u6210 14 \u79cd\u8bed\u8a00\uff0c\u5305\u62ec\u963f\u62c9\u4f2f\u8bed\u3001\u5370\u5730\u8bed\u3001\u4e2d\u6587\u3001\u6cd5\u8bed\u3001\u65e5\u8bed\u7b49\u3002<\/p>\n\n\n\n<p><strong>\u6d4b\u8bd5\u7ed3\u679c<\/strong><\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li>GPT-4.5 \u5728 <strong>\u82f1\u8bed MMLU<\/strong> \u4e0a\u7684\u5f97\u5206\u4e3a <strong>89.6%<\/strong>\uff0c\u7565\u4f4e\u4e8e o1\uff0892.3%\uff09\u3002<\/li>\n\n\n\n<li>\u5728 <strong>\u4f4e\u8d44\u6e90\u8bed\u8a00\uff08\u5982\u7ea6\u9c81\u5df4\u8bed\uff09<\/strong> \u4e0a\uff0cGPT-4.5 \u7684\u5f97\u5206\u4e3a <strong>68.1%<\/strong>\uff0c\u6bd4 GPT-4o\uff0862.1%\uff09\u6709\u4e00\u5b9a\u63d0\u5347\uff0c\u4f46\u4ecd\u5b58\u5728\u6539\u8fdb\u7a7a\u95f4\u3002<\/li>\n<\/ul>\n\n\n\n<p><strong>\u516d\u3001\u7ed3\u8bba<\/strong><br>GPT-4.5 \u5728\u77e5\u8bc6\u5e7f\u5ea6\u3001\u6587\u672c\u751f\u6210\u3001\u60c5\u611f\u667a\u80fd\u548c\u5b89\u5168\u6027\u7b49\u65b9\u9762\u6709\u6240\u6539\u8fdb\uff0c\u4f46\u4ecd\u7136\u5b58\u5728\u6f5c\u5728\u98ce\u9669\uff0c\u5c24\u5176\u662f <strong>CBRN \u9886\u57df\u548c\u8bf4\u670d\u80fd\u529b<\/strong> \u65b9\u9762\u7684\u4e2d\u7b49\u98ce\u9669\u3002OpenAI \u8ba1\u5212\u901a\u8fc7 <strong>\u6e10\u8fdb\u5f0f\u90e8\u7f72<\/strong> \u6765\u4f18\u5316\u5176\u5b89\u5168\u7b56\u7565\uff0c\u5e76\u6301\u7eed\u6539\u8fdb\u672a\u6765\u7248\u672c\u7684 GPT \u6a21\u578b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u4e00\u3001\u5f15\u8a00GPT-4.5 \u662f OpenAI \u8fc4\u4eca\u4e3a\u6b62\u6700\u5927\u3001\u6700\u5177\u77e5\u8bc6\u6027\u7684 AI \u6a21\u578b\uff0c\u5176\u76ee\u6807\u662f\u76f8\u8f83\u4e8e\u524d\u4ee3\u6a21\u578b\uff08\u5982  [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,81,83,103,86],"class_list":["post-5749","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-ai-safety","tag-chatgpt","tag-openai","tag-llm"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u4e00\u3001\u5f15\u8a00GPT-4.5 \u662f OpenAI \u8fc4\u4eca\u4e3a\u6b62\u6700\u5927\u3001\u6700\u5177\u77e5\u8bc6\u6027\u7684 AI \u6a21\u578b\uff0c\u5176\u76ee\u6807\u662f\u76f8\u8f83\u4e8e\u524d\u4ee3\u6a21\u578b\uff08\u5982 &hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5749","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=5749"}],"version-history":[{"count":1,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5749\/revisions"}],"predecessor-version":[{"id":5751,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/5749\/revisions\/5751"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=5749"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=5749"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=5749"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}