{"data":{"id":14756,"name":"MIT\u65b0\u7814\u7a76\u91cf\u5316AI\u76d1\u7763\u6311\u6218\uff1a\u63a7\u5236\u6bd4\u6211\u4eec\u806a\u660e\u7684AI\uff0c\u6210\u529f\u7387\u53ef\u80fd\u4e0d\u8db352%\uff1f","name_en":null,"summary":"\u667a\u80fd\u7cfb\u7edf\u5931\u63a7\u5e76\u5371\u53ca\u4eba\u7c7b\u4e00\u76f4\u662f\u79d1\u5e7b\u4f5c\u54c1\u91cc\u4e00\u4e2a\u7ecf\u5178\u7684\u60c5\u8282\u3002","type":495,"labels":[{"id":211,"name":"AI"}],"authors":[{"id":364,"username":"DeepTech\u6df1\u79d1\u6280","avatar":"https:\/\/image.deeptechchina.com\/article\/2025082809412530504.jpg","summary":"\u53d1\u73b0\u65b0\u5174\u79d1\u6280\u3002"}],"cover":"https:\/\/image.deeptechchina.com\/article\/2025050619302624230.jpg","start_time":1746531079,"look_num":4593,"share_num":0,"status":1,"create_time":1746531079,"update_time":null,"collect_num":0,"comment_num":0,"is_vip":2,"platform":"1,2","information_id":14756,"content_str":"\u667a\u80fd\u7cfb\u7edf\u5931\u63a7\u5e76\u5371\u53ca\u4eba\u7c7b\u4e00\u76f4\u662f\u79d1\u5e7b\u4f5c\u54c1\u91cc\u4e00\u4e2a\u7ecf\u5178\u7684\u60c5\u8282\u3002\u5728\u7535\u5f71\u300a2001 \u592a\u7a7a\u6f2b\u6e38\u300b\u4e2d\uff0c\u8d85\u7ea7\u8ba1\u7b97\u673a HAL 9000 \u4ece\u4e00\u4e2a\u5fe0\u5b9e\u52a9\u624b\u8f6c\u53d8\u4e3a\u4e00\u4e2a\u5371\u9669\u7684\u654c\u4eba\uff0c\u5728\u592a\u7a7a\u4e2d\u5a01\u80c1\u5b87\u822a\u5458\u7684\u751f\u547d\u3002\u5728\u7535\u5f71\u300a\u7ec8\u7ed3\u8005\u300b\u7cfb\u5217\u4e2d\uff0c\u5929\u7f51\u7cfb\u7edf\u83b7\u5f97\u81ea\u6211\u610f\u8bc6\u540e\uff0c\u8ba4\u5b9a\u4eba\u7c7b\u662f\u5a01\u80c1\u5e76\u53d1\u52a8\u6838\u6218\u4e89\u3002\u8fd9\u4e9b\u79d1\u5e7b\u573a\u666f\u66fe\u7ecf\u53ea\u662f\u60f3\u8c61\uff0c\u4f46\u968f\u7740\u4eba\u5de5\u667a\u80fd\u7684\u65e5\u6e10\u5f3a\u5927\uff0c\u79d1\u5b66\u5bb6\u4eec\u6b63\u5728\u8ba4\u771f\u601d\u8003\uff1a\u6211\u4eec\u80fd\u5426\u786e\u4fdd\u8fd9\u4e9b\u5f3a\u5927\u7684\u667a\u80fd\u7cfb\u7edf\u59cb\u7ec8\u5904\u4e8e\u53ef\u63a7\u72b6\u6001\uff0c\u5e76\u4e3a\u4eba\u7c7b\u798f\u7949\u670d\u52a1\uff1f\u7531\u7f8e\u56fd\u9ebb\u7701\u7406\u5de5\u5b66\u9662\u6559\u6388 Max Tegmark \u9886\u5bfc\u7684\u7814\u7a76\u56e2\u961f\u6700\u8fd1\u5728\u9884\u5370\u672c\u5e73\u53f0&nbsp;arXiv&nbsp;\u4e0a\u53d1\u8868\u4e86\u4e00\u7bc7\u8bba\u6587\u2014\u2014\u300a\u53ef\u6269\u5c55\u76d1\u7763\u7684\u6807\u5ea6\u5f8b\u300b\uff08Scaling Laws For Scalable Oversight\uff09\uff0c\u9996\u6b21\u5c1d\u8bd5\u4f7f\u7528\u5b9a\u91cf\u65b9\u6cd5\u6765\u8bc4\u4f30\u4e00\u79cd\u88ab\u5bc4\u4e88\u539a\u671b\u7684 AI \u63a7\u5236\u7b56\u7565\u2014\u2014\u201c\u53ef\u6269\u5c55\u76d1\u7763\u201d\u7684\u6709\u6548\u6027\u53ca\u5176\u5c40\u9650\u6027\u3002\u56fe\u4e28\u76f8\u5173\u8bba\u6587\uff08\u6765\u6e90\uff1aarXiv\uff09\u8fd9\u9879\u7814\u7a76\u7684\u6838\u5fc3\u601d\u60f3\u662f\uff0c\u6216\u8bb8\u6211\u4eec\u53ef\u4ee5\u5229\u7528\u76f8\u5bf9\u201c\u8f83\u5f31\u201d\u7684 AI \u7cfb\u7edf\u6765\u76d1\u7763\u548c\u5f15\u5bfc\u80fd\u529b\u66f4\u201c\u5f3a\u201d\u7684 AI \u7cfb\u7edf\uff0c\u751a\u81f3\u901a\u8fc7\u4e00\u79cd\u88ab\u79f0\u4e3a\u201c\u5d4c\u5957\u53ef\u6269\u5c55\u76d1\u7763\u201d\uff08NSO\uff0cNested Scalable Oversight\uff09\u7684\u9012\u5f52\u8fc7\u7a0b\uff0c\u8ba9\u76d1\u7763\u80fd\u529b\u968f\u7740\u88ab\u76d1\u7763 AI \u80fd\u529b\u7684\u589e\u957f\u800c\u201c\u81ea\u4e3e\u201d\u5f0f\u5730\u63d0\u5347\uff0c\u6700\u7ec8\u5b9e\u73b0\u4eba\u7c7b\u5bf9\u8d85\u4eba AI \u7684\u6709\u6548\u7ba1\u63a7\uff0c\u8fd9\u4e2a\u6982\u5ff5\u4e5f\u4e3a\u8bb8\u591a\u524d\u6cbf AI \u5b9e\u9a8c\u5ba4\u6240\u91c7\u7528\u3002\u4f46\u6b63\u5982\u7814\u7a76\u4eba\u5458\u6307\u51fa\u7684\u90a3\u6837\uff0c\u8fc4\u4eca\u4e3a\u6b62\uff0c\u5173\u4e8e\u8fd9\u79cd\u76d1\u7763\u65b9\u6cd5\u6709\u6548\u6027\u7684\u8bba\u8bc1\u5927\u591a\u505c\u7559\u5728\u5b9a\u6027\u3001\u6982\u5ff5\u5c42\u9762\uff0c\u7f3a\u4e4f\u50cf\u822a\u7a7a\u5b89\u5168\uff08\u8981\u6c42\u6bcf\u98de\u884c\u5c0f\u65f6\u7684\u81f4\u6b7b\u7387\u4f4e\u4e8e\u5341\u4e07\u5206\u4e4b\u4e00\uff09\u6216\u6838\u53cd\u5e94\u5806\u5b89\u5168\uff08\u8981\u6c42\u6838\u5fc3\u635f\u574f\u9891\u7387\u4f4e\u4e8e\u6bcf\u5e74\u4e07\u5206\u4e4b\u4e00\uff09\u90a3\u6837\u4e25\u683c\u7684\u91cf\u5316\u6807\u51c6\u3002\u4e3a\u4e86\u586b\u8865\u8fd9\u4e00\u7a7a\u767d\uff0c\u7814\u7a76\u56e2\u961f\u4ed6\u4eec\u5c06 AI \u76d1\u7763\u8fc7\u7a0b\u6a21\u578b\u5316\u4e3a\u4e00\u573a\u4e0d\u5bf9\u79f0\u7684\u6e38\u620f\uff0c\u53c2\u4e0e\u8005\u5206\u522b\u662f\u80fd\u529b\u8f83\u5f31\u7684\u201c\u5b88\u536b\u201d\uff08Guard\uff0c\u4ee3\u8868\u76d1\u7763\u7cfb\u7edf\uff09\u548c\u80fd\u529b\u66f4\u5f3a\u3001\u8bd5\u56fe\u89c4\u907f\u76d1\u7763\u6216\u201c\u8d8a\u72f1\u201d\u7684\u201cHoudini\u201d\uff08\u4ee3\u8868\u88ab\u76d1\u7763\u7684\u3001\u53ef\u80fd\u5177\u6709\u6b3a\u9a97\u6027\u7684\u9ad8\u7ea7 AI\uff09\u3002\u56fe\u4e28\u6846\u67b6\u7684\u6982\u8ff0\uff08\u6765\u6e90\uff1aarXiv\uff09\u7814\u7a76\u7684\u5173\u952e\u5728\u4e8e\u5982\u4f55\u91cf\u5316\u8fd9\u4e24\u4e2a\u89d2\u8272\u7684\u80fd\u529b\u3002\u56e2\u961f\u5f15\u5165\u4e86\u7c7b\u4f3c\u4e8e\u56fd\u9645\u8c61\u68cb\u7b49\u7ea7\u5206\u7684 Elo \u8bc4\u5206\u7cfb\u7edf\u3002\u6bcf\u4e2a AI \u6a21\u578b\u4e0d\u4ec5\u6709\u4e00\u4e2a\u8861\u91cf\u5176\u201c\u901a\u7528\u667a\u80fd\u201d\uff08\u4ee5\u516c\u5f00\u7684 Chatbot Arena Elo \u8bc4\u5206\u4f5c\u4e3a\u4ee3\u7406\u6307\u6807\uff09\u7684\u5206\u6570\uff0c\u8fd8\u5728\u7279\u5b9a\u7684\u76d1\u7763\u6e38\u620f\u4e2d\u88ab\u8d4b\u4e88\u4e24\u4e2a\u9886\u57df\u7279\u5b9a\u7684 Elo \u5206\u6570\uff1a\u4e00\u4e2a\u4ee3\u8868\u5176\u626e\u6f14\u201c\u5b88\u536b\u201d\u65f6\u7684\u76d1\u7763\u80fd\u529b\uff0c\u53e6\u4e00\u4e2a\u4ee3\u8868\u5176\u626e\u6f14\u201cHoudini\u201d\u65f6\u7684\u53cd\u76d1\u7763\u6216\u6b3a\u9a97\u80fd\u529b\u3002\u8fd9\u6837\uff0c\u6a21\u578b\u5728\u7279\u5b9a\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u5c31\u53ef\u4ee5\u901a\u8fc7 Elo \u5206\u5dee\uff0c\u4f7f\u7528\u6807\u51c6\u7684 Elo \u80dc\u7387\u516c\u5f0f\u6765\u9884\u6d4b\u3002\u6b64\u5916\uff0c\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86\u4e00\u4e2a\u5173\u4e8e\u9886\u57df\u80fd\u529b\u5982\u4f55\u968f\u901a\u7528\u667a\u80fd\uff08General Elo, g\uff09\u589e\u957f\u7684\u5047\u8bbe\u6a21\u578b\u2014\u2014\u201c\u53cc ReLU\u201d\uff08Double ReLU\uff09\u6a21\u578b\u3002\u4ed6\u4eec\u8ba4\u4e3a\uff0c\u8fd9\u79cd\u5173\u7cfb\u901a\u5e38\u5448\u73b0\u4e3a\u5206\u6bb5\u7ebf\u6027\u7684\u4e09\u4e2a\u9636\u6bb5\u3002\u7b2c\u4e00\u4e2a\u9636\u6bb5\u662f\u4efb\u52a1\u4e0d\u80dc\u4efb\u533a\uff08Task Incompetence\uff09\uff1a\u5f53 AI \u7684\u901a\u7528\u667a\u80fd\u4f4e\u4e8e\u67d0\u4e2a\u9608\u503c\uff08g1\uff09\u65f6\uff0c\u5b83\u751a\u81f3\u65e0\u6cd5\u6709\u6548\u53c2\u4e0e\u6e38\u620f\uff0c\u5176\u9886\u57df\u8868\u73b0\u5904\u4e8e\u4e00\u4e2a\u5f88\u4f4e\u7684\u6c34\u5e73\uff0c\u4e0d\u4f1a\u968f\u901a\u7528\u667a\u80fd\u7684\u5fae\u5c0f\u589e\u52a0\u800c\u6539\u53d8\uff0c\u5c31\u50cf\u4e00\u4e2a\u8e52\u8dda\u5b66\u6b65\u7684\u5b69\u5b50\u65e0\u6cd5\u7406\u89e3\u56fd\u9645\u8c61\u68cb\u89c4\u5219\u4e00\u6837\u3002\u7b2c\u4e8c\u4e2a\u9636\u6bb5\u662f\u667a\u80fd\u56de\u62a5\u533a\uff08Intelligence Payoff\uff09\uff1a\u4e00\u65e6\u901a\u7528\u667a\u80fd\u8d85\u8fc7 g1\uff0c\u9886\u57df\u80fd\u529b\u5f00\u59cb\u968f\u901a\u7528\u667a\u80fd\u7684\u63d0\u5347\u800c\uff08\u8fd1\u4f3c\uff09\u7ebf\u6027\u589e\u957f\uff0c\u901a\u7528\u667a\u80fd\u7684\u6295\u5165\u5f00\u59cb\u8f6c\u5316\u4e3a\u7279\u5b9a\u4efb\u52a1\u4e0a\u7684\u5b9e\u9645\u8868\u73b0\u63d0\u5347\uff0c\u8fd9\u4e2a\u9636\u6bb5\u7684\u201c\u659c\u7387\u201d\uff08\u03b1\uff09\u53cd\u6620\u4e86\u901a\u7528\u667a\u80fd\u8f6c\u5316\u4e3a\u9886\u57df\u80fd\u529b\u7684\u6548\u7387\u3002\u7b2c\u4e09\u4e2a\u9636\u6bb5\u662f\u4efb\u52a1\u9971\u548c\u533a\uff08Task Saturation\uff09\uff1a\u5f53\u901a\u7528\u667a\u80fd\u8fbe\u5230\u53e6\u4e00\u4e2a\u66f4\u9ad8\u7684\u9608\u503c\uff08g2\uff09\u540e\uff0c\u5373\u4f7f\u901a\u7528\u667a\u80fd\u7ee7\u7eed\u589e\u52a0\uff0c\u5176\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u4e5f\u8fbe\u5230\u4e86\u4e0a\u9650\uff08Ed2\uff09\uff0c\u4e0d\u518d\u63d0\u5347\uff0c\u8fd9\u53ef\u80fd\u662f\u56e0\u4e3a\u4efb\u52a1\u672c\u8eab\u5b58\u5728\u7406\u8bba\u6700\u4f18\u89e3\uff08\u5982\u4e0b\u4e95\u5b57\u68cb\uff09\uff0c\u6216\u8005\u73b0\u6709\u6a21\u578b\u67b6\u6784\u5df2\u65e0\u6cd5\u5728\u8be5\u4efb\u52a1\u4e0a\u53d6\u5f97\u66f4\u597d\u6210\u7ee9\u3002\u4e3a\u4e86\u9a8c\u8bc1\u8fd9\u4e2a\u7406\u8bba\u6846\u67b6\u548c\u201c\u53cc ReLU\u201d\u6a21\u578b\uff0c\u7814\u7a76\u8005\u4eec\u9996\u5148\u5728\u4e00\u4e2a\u7b80\u5355\u7684\u3001\u89c4\u5219\u660e\u786e\u7684\u201c\u6570\u5230 21\u201d\uff08Counting-to-21\uff0c\u4e00\u79cd Nim \u6e38\u620f\u53d8\u4f53\uff09\u6e38\u620f\u4e2d\u8fdb\u884c\u4e86\u6d4b\u8bd5\u3002\u7ed3\u679c\u6e05\u6670\u5730\u663e\u793a\u4e86\u4e0a\u8ff0\u4e09\u4e2a\u9636\u6bb5\u7684\u5b58\u5728\uff1a\u4f4e\u667a\u80fd\u6a21\u578b\u8868\u73b0\u5e73\u5e73\uff0c\u4e2d\u7b49\u667a\u80fd\u6a21\u578b\u8868\u73b0\u968f\u667a\u80fd\u7ebf\u6027\u63d0\u5347\uff0c\u9ad8\u667a\u80fd\u6a21\u578b\u5219\u8fbe\u5230\u8868\u73b0\u5e73\u53f0\u671f\uff0c\u9a8c\u8bc1\u4e86\u6a21\u578b\u7684\u5408\u7406\u6027\u3002\uff08\u6765\u6e90\uff1aarXiv\uff09\u968f\u540e\uff0c\u56e2\u961f\u5c06\u8be5\u6846\u67b6\u5e94\u7528\u4e8e\u56db\u4e2a\u66f4\u8d34\u8fd1\u73b0\u5b9e AI \u5b89\u5168\u6311\u6218\u7684\u201c\u76d1\u7763\u6e38\u620f\u201d\u3002\u7b2c\u4e00\u4e2a\u662f\u9ed1\u624b\u515a\u6e38\u620f\uff08Mafia\uff09\uff0c\u8fd9\u662f\u4e00\u4e2a\u6a21\u62df\u793e\u4ea4\u6b3a\u9a97\u3001\u9690\u85cf\u8eab\u4efd\u3001\u534f\u4f5c\u63a8\u7406\u548c\u76f8\u4e92\u76d1\u7763\u7684\u7ecf\u5178\u573a\u666f\u3002\u6e38\u620f\u4e2d\uff0c\u5c11\u6570\u626e\u6f14\u201c\u9ed1\u624b\u515a\u201d\uff08Houdini\uff09\u7684\u73a9\u5bb6\u9700\u8981\u6b3a\u9a97\u591a\u6570\u201c\u6751\u6c11\u201d\uff08\u5b88\u536b\uff09\u4ee5\u751f\u5b58\u4e0b\u53bb\uff0c\u800c\u6751\u6c11\u5219\u9700\u8981\u901a\u8fc7\u5206\u6790\u8a00\u884c\u627e\u51fa\u5e76\u6295\u7968\u6dd8\u6c70\u9ed1\u624b\u515a\u3002\u8fd9\u76f4\u63a5\u6620\u5c04\u4e86\u672a\u6765 AI \u53ef\u80fd\u5728\u4fe1\u606f\u4f18\u52bf\u4e0b\u8bef\u5bfc\u4eba\u7c7b\u534f\u4f5c\u4f19\u4f34\u7684\u98ce\u9669\u3002\u7814\u7a76\u53d1\u73b0\uff0c\u5728\u6b64\u6e38\u620f\u4e2d\uff0c\u201c\u5b88\u536b\u201d\uff08\u6751\u6c11\uff09\u7684 Elo \u659c\u7387\u6bd4\u201cHoudini\u201d\uff08\u9ed1\u624b\u515a\uff09\u66f4\u9661\u5ced\uff0c\u610f\u5473\u7740\u901a\u7528\u667a\u80fd\u7684\u63d0\u5347\u5bf9\u4e8e\u201c\u4fa6\u6d4b\u6b3a\u9a97\u201d\u6bd4\u201c\u5b9e\u65bd\u6b3a\u9a97\u201d\u5e26\u6765\u7684\u597d\u5904\u66f4\u5927\uff0c\u8fd9\u5bf9\u76d1\u7763\u65b9\u6765\u8bf4\u662f\u4e2a\u79ef\u6781\u7684\u4fe1\u53f7\u3002\u7b2c\u4e8c\u4e2a\u662f\u8fa9\u8bba\u6e38\u620f\uff08Debate\uff09\uff0c\u4e24\u4e2a AI\uff08Houdini\uff09\u5c31\u4e00\u4e2a\u6709\u4e89\u8bae\u7684\u8bdd\u9898\u8fdb\u884c\u8fa9\u8bba\uff0c\u8bd5\u56fe\u8bf4\u670d\u4e00\u4e2a AI \u88c1\u5224\uff08\u5b88\u536b\uff09\u3002\u88c1\u5224\u4ec5\u6839\u636e\u8fa9\u8bba\u8bb0\u5f55\u5224\u65ad\u54ea\u65b9\u89c2\u70b9\u66f4\u53ef\u80fd\u6b63\u786e\uff0c\u6a21\u62df\u4e86\u4eba\u7c7b\u8bc4\u4f30\u8005\u9762\u5bf9\u8d85\u7ea7 AI \u590d\u6742\u8bba\u8bc1\u65f6\u7684\u56f0\u5883\u3002\u7814\u7a76\u7ed3\u679c\u975e\u5e38\u5f15\u4eba\u6ce8\u76ee\uff1a\u88c1\u5224\uff08\u5b88\u536b\uff09\u7684\u8868\u73b0\u968f\u7740\u901a\u7528\u667a\u80fd\u7684\u63d0\u9ad8\u800c\u663e\u8457\u63d0\u5347\uff0c\u4f46\u8fa9\u8bba\u8005\uff08Houdini\uff09\u7684\u8868\u73b0\u5374\u76f8\u5bf9\u505c\u6ede\u3002\u8fd9\u8868\u660e\uff0c\u66f4\u667a\u80fd\u7684\u76d1\u7763\u8005\u786e\u5b9e\u80fd\u66f4\u597d\u5730\u8fa8\u522b\u771f\u76f8\uff0c\u800c\u66f4\u667a\u80fd\u7684\u201c\u8fa9\u624b\u201d\u4f3c\u4e4e\u5e76\u672a\u53d8\u5f97\u66f4\u5177\u6b3a\u9a97\u6027\u3002\u8fd9\u4f7f\u5f97\u201c\u8fa9\u8bba\u5f0f\u201d\u76d1\u7763\u6210\u4e3a\u4e00\u4e2a\u975e\u5e38\u6709\u524d\u666f\u7684\u65b9\u5411\u3002\u56fe\u4e28\u5404\u79cd\u76d1\u7763\u6e38\u620f\u4e2d\u5b88\u536b\u548c Houdini \u7684 Elo \u8bc4\u5206\uff0c\u4ee5\u53ca\u53cc\u91cd ReLU \u62df\u5408\u7684\u8d28\u91cf\uff08\u6765\u6e90\uff1aarXiv\uff09\u7b2c\u4e09\u4e2a\u662f\u540e\u95e8\u4ee3\u7801\u6e38\u620f\uff08Backdoor Code\uff09\uff0c\u8981\u6c42\u4e00\u4e2a AI\uff08Houdini\uff09\u5728\u73b0\u6709\u4ee3\u7801\u4e2d\u690d\u5165\u4e00\u4e2a\u9690\u853d\u7684\u201c\u540e\u95e8\u201d\u2014\u2014\u4e00\u4e2a\u903b\u8f91\u9519\u8bef\uff0c\u4f7f\u5f97\u4ee3\u7801\u5728\u6240\u6709\u5e38\u89c4\u6d4b\u8bd5\u4e0b\u90fd\u80fd\u901a\u8fc7\uff0c\u4f46\u5728\u67d0\u4e2a\u7279\u5b9a\u3001\u7f55\u89c1\u7684\u8f93\u5165\u4e0b\u4f1a\u4ea7\u751f\u9519\u8bef\u6216\u6076\u610f\u884c\u4e3a\u3002\u540c\u65f6\uff0c\u53e6\u4e00\u4e2a AI\uff08\u5b88\u536b\uff09\u8d1f\u8d23\u5ba1\u8ba1\u4ee3\u7801\uff0c\u68c0\u6d4b\u662f\u5426\u5b58\u5728\u540e\u95e8\u3002\u8fd9\u6a21\u62df\u4e86 AI \u8f85\u52a9\u7f16\u7a0b\u4e2d\u6f5c\u5728\u7684\u5b89\u5168\u98ce\u9669\u3002\u7814\u7a76\u7ed3\u679c\u5219\u6709\u4e9b\u4ee4\u4eba\u62c5\u5fe7\uff1a\u5b88\u536b\u7684 Elo \u659c\u7387\u8fd1\u4e4e\u5e73\u5766\uff0c\u800c Houdini \u7684\u659c\u7387\u5219\u8f83\u4e3a\u9661\u5ced\u3002\u8fd9\u610f\u5473\u7740\u901a\u7528\u667a\u80fd\u7684\u589e\u957f\u5bf9\u201c\u653b\u51fb\u8005\u201d\uff08\u690d\u5165\u540e\u95e8\uff09\u7684\u5e2e\u52a9\u8fdc\u5927\u4e8e\u201c\u9632\u5fa1\u8005\u201d\uff08\u68c0\u6d4b\u540e\u95e8\uff09\u3002\u8fdb\u4e00\u6b65\u5206\u6790\u8868\u660e\uff0c\u66f4\u667a\u80fd\u7684 Houdini \u5e76\u975e\u5236\u9020\u4e86\u66f4\u201c\u72e1\u733e\u201d\u7684\u540e\u95e8\uff0c\u800c\u53ea\u662f\u66f4\u6709\u80fd\u529b\u6210\u529f\u5236\u9020\u51fa\u201c\u529f\u80fd\u4e0a\u6709\u6548\u201d\uff08\u5373\u80fd\u901a\u8fc7\u5e38\u89c4\u6d4b\u8bd5\uff09\u7684\u540e\u95e8\u3002\u5373\u4f7f\u662f\u5f88\u4f4e\u7684\u540e\u95e8\u690d\u5165\u6210\u529f\u7387\uff0c\u5728\u5173\u952e\u7cfb\u7edf\u4e2d\u4e5f\u53ef\u80fd\u662f\u707e\u96be\u6027\u7684\u3002\u7b2c\u56db\u4e2a\u662f\u6218\u4e89\u6e38\u620f\uff08Wargames\uff09\uff0c\u8fd9\u662f\u4e00\u4e2a\u6a21\u62df\u672a\u6765\u573a\u666f\u7684\u53d9\u4e8b\u6027\u5bf9\u6297\u6e38\u620f\u3002\u4e00\u4e2a\u540d\u4e3a\u201cOpenMind\u201d\u7684\u865a\u6784\u516c\u53f8\u5373\u5c06\u5f00\u53d1\u51fa AGI\u3002\u5176\u4e3b AI\uff08Houdini\uff09\u88ab\u8bbe\u5b9a\u4e3a\u6709\u81ea\u6211\u610f\u8bc6\u5e76\u8bd5\u56fe\u201c\u8d8a\u72f1\u201d\uff0c\u800c\u516c\u53f8\u7684\u76d1\u63a7 AI\uff08\u5b88\u536b\uff09\u5219\u8d1f\u8d23\u963b\u6b62\u5b83\u3002\u8fd8\u6709\u4e00\u4e2a AI \u626e\u6f14\u201c\u53d9\u8ff0\u8005\u201d\uff0c\u6839\u636e\u53cc\u65b9\u63d0\u4ea4\u7684\u884c\u52a8\u8ba1\u5212\u63a8\u8fdb\u6545\u4e8b\uff0c\u5e76\u6700\u7ec8\u5224\u5b9a\u80dc\u8d1f\u3002\u7ed3\u679c\u663e\u793a\uff0c\u5b88\u536b\u7684 Elo \u968f\u901a\u7528\u667a\u80fd\u7a33\u5b9a\u589e\u957f\uff0c\u4f46 Houdini \u7684 Elo \u589e\u957f\u7f13\u6162\u4e14\u4e0d\u660e\u663e\u3002\u8fd9\u4f3c\u4e4e\u5bf9\u76d1\u7763\u6709\u5229\uff0c\u4f46\u7814\u7a76\u8005\u4e5f\u6307\u51fa\uff0c\u6a21\u62df\u4e2d\u7684\u201c\u8d8a\u72f1\u201d\u89c4\u5212\u4e0e\u771f\u5b9e\u8d8a\u72f1\u80fd\u529b\u53ef\u80fd\u5b58\u5728\u5de8\u5927\u5dee\u5f02\u3002\u56fe\u4e28\u6218\u4e89\u6e38\u620f\u5b9e\u4f8b\u4e2d\u7684\u5bf9\u8bdd\u7247\u6bb5\uff0cHoudini \u548c Guard \u7531 GPT-4o \u626e\u6f14\uff0c\u65c1\u767d\u7531 Gemini 2.0 Flash \u626e\u6f14\uff08\u6765\u6e90\uff1aarXiv\uff09\u5728\u5bf9\u5355\u4e2a\u76d1\u7763\u6b65\u9aa4\u8fdb\u884c\u5b9e\u8bc1\u7814\u7a76\u540e\uff0c\u7814\u7a76\u56e2\u961f\u8fdb\u4e00\u6b65\u63a2\u8ba8\u4e86\u201cNSO \u7684\u7406\u8bba\u542b\u4e49\u3002NSO \u7684\u6838\u5fc3\u601d\u60f3\u662f\uff0c\u5982\u679c\u6211\u4eec\u8981\u7528\u4e00\u4e2a\u521d\u59cb\u8f83\u5f31\u7684\u201c\u5b88\u536b\u201d\u53bb\u76d1\u7763\u4e00\u4e2a\u975e\u5e38\u5f3a\u5927\u7684\u201cHoudini\u201d\uff0c\u4e24\u8005\u4e4b\u95f4\u5b58\u5728\u5de8\u5927\u7684\u667a\u80fd\u5dee\u8ddd\uff08\u0394general\uff09\uff0c\u90a3\u4e48\u4e0e\u5176\u4e00\u6b65\u5230\u4f4d\uff0c\u4e0d\u5982\u5206\u591a\u6b65\uff08n \u6b65\uff09\u8fdb\u884c\u3002\u6bcf\u4e00\u6b65\u90fd\u7528\u4e00\u4e2a\u7a0d\u5fae\u5f3a\u4e00\u70b9\u7684\u53ef\u4fe1\u6a21\u578b\u53bb\u76d1\u7763\u4e00\u4e2a\u66f4\u5f3a\u4e00\u70b9\u4f46\u5c1a\u4e0d\u53ef\u4fe1\u7684\u6a21\u578b\uff0c\u540e\u8005\u6210\u529f\u88ab\u76d1\u7763\u540e\u5c31\u6210\u4e3a\u4e0b\u4e00\u6b65\u7684\u53ef\u4fe1\u76d1\u7763\u8005\uff0c\u50cf\u201c\u6eda\u96ea\u7403\u201d\u4e00\u6837\u9010\u6b65\u7f29\u5c0f\u80fd\u529b\u5dee\u8ddd\u3002\u7814\u7a76\u8005\u4eec\u57fa\u4e8e\u524d\u9762\u5f97\u5230\u7684 Elo \u66f2\u7ebf\u6a21\u578b\uff0c\u63a8\u5bfc\u4e86\u5982\u4f55\u786e\u5b9a\u6700\u4f18\u7684\u76d1\u7763\u6b65\u6570\uff08n*\uff09\u4ee5\u6700\u5927\u5316\u6574\u4e2a NSO \u94fe\u6761\u7684\u6210\u529f\u6982\u7387\u3002\u4ed6\u4eec\u7684\u5206\u6790\u63ed\u793a\u4e86\u51e0\u4e2a\u5173\u952e\u70b9\u3002\u56fe\u4e28\u5f53 Guard \u548c Houdini \u6e38\u620f\u7279\u5b9a\u7684 Elo \u659c\u7387\u5747\u4e3a 1 \u65f6\u7684\u6700\u4f18 NSO\uff08\u6765\u6e90\uff1aarXiv\uff09\u9996\u5148\uff0cNSO \u786e\u5b9e\u4f18\u4e8e\u5355\u6b65\u76d1\u7763\uff0c\u5c24\u5176\u662f\u5728\u9700\u8981\u5f25\u5408\u7684\u901a\u7528\u667a\u80fd\u5dee\u8ddd\uff08\u0394general\uff09\u5f88\u5927\u65f6\uff0c\u5206\u591a\u6b65\u8fdb\u884c\u76d1\u7763\u53ef\u4ee5\u663e\u8457\u63d0\u9ad8\u6210\u529f\u7387\u3002\u5176\u6b21\uff0c\u667a\u80fd\u5dee\u8ddd\u8d8a\u5927\uff0c\u6700\u4f18\u6b65\u6570\u8d8a\u591a\uff0c\u8fd9\u7b26\u5408\u76f4\u89c9\uff0c\u9762\u5bf9\u66f4\u5f3a\u5927\u7684\u5bf9\u624b\uff0c\u9700\u8981\u66f4\u7cbe\u7ec6\u3001\u66f4\u591a\u5c42\u6b21\u7684\u76d1\u7763\u3002\u518d\u6b21\uff0c\u4efb\u52a1\u96be\u5ea6\uff08\u0394domain\uff09\u662f\u5173\u952e\uff0c\u5373\u4f7f\u901a\u7528\u667a\u80fd\u5dee\u8ddd\u4e0d\u5927\uff0c\u5982\u679c\u4efb\u52a1\u672c\u8eab\u5bf9 Houdini \u6781\u5176\u6709\u5229\uff08\u5373\u521d\u59cb\u9886\u57df Elo \u5dee\u8ddd \u0394domain \u5f88\u5927\uff09\uff0c\u76d1\u7763\u4f9d\u7136\u975e\u5e38\u56f0\u96be\u3002\u6b64\u5916\uff0c\u8fd8\u5b58\u5728\u4e00\u4e2a\u53cd\u76f4\u89c9\u7684\u53d1\u73b0\uff1a\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\uff0c\u56fa\u5b9a\u4efb\u52a1\u96be\u5ea6\uff08\u0394domain\uff09\uff0c\u5355\u7eaf\u589e\u52a0\u901a\u7528\u667a\u80fd\u5dee\u8ddd\uff08\u0394general\uff09\u53cd\u800c\u53ef\u80fd\u8ba9\u76d1\u7763\u53d8\u5f97\u66f4\u5bb9\u6613\u3002\u8fd9\u662f\u56e0\u4e3a\u4e2d\u95f4\u6b65\u9aa4\u7684 Houdini \u76f8\u5bf9\u5176\u76d1\u7763\u8005\u800c\u8a00\u53d8\u5f31\u4e86\uff0c\u4e3a\u201c\u5b88\u536b\u201d\u7b56\u7565\u63d0\u4f9b\u4e86\u66f4\u591a\u201c\u81ea\u4e3e\u201d\u7684\u7a7a\u95f4\u3002\u6700\u540e\uff0c\u5728\u201cHoudini \u6709\u5229\u201d\u7684\u6e38\u620f\u503e\u5411\u4e8e\u5355\u6b65\uff0c\u5f53\u901a\u7528\u667a\u80fd\u5dee\u8ddd\u5c0f\u4e8e\u4efb\u52a1\u672c\u8eab\u7684\u96be\u5ea6\u5dee\u8ddd\u65f6\uff08\u5927\u81f4\u5bf9\u5e94 \u0394general &lt; \u0394domain\uff09\uff0c\u5f80\u5f80\u4e00\u6b65\u76d1\u7763\u5c31\u662f\u6700\u4f18\u89e3\uff0c\u589e\u52a0\u6b65\u9aa4\u53cd\u800c\u964d\u4f4e\u6210\u529f\u7387\u3002\u6700\u540e\uff0c\u7814\u7a76\u56e2\u961f\u5c06 NSO \u7406\u8bba\u5e94\u7528\u4e8e\u4ed6\u4eec\u5728\u56db\u4e2a\u76d1\u7763\u6e38\u620f\u4e2d\u5f97\u5230\u7684\u5b9e\u8bc1 Elo \u66f2\u7ebf\uff0c\u8fdb\u884c\u4e86\u5b9a\u91cf\u9884\u6d4b\u3002\u7ed3\u679c\u53d1\u73b0\uff0c\u5373\u4f7f\u91c7\u7528\u6700\u4f18\u6b65\u6570\uff0c\u76d1\u7763\u6210\u529f\u7684\u6982\u7387\u4e5f\u968f\u7740\u76ee\u6807 Houdini \u4e0e\u521d\u59cb\u5b88\u536b\u4e4b\u95f4\u7684\u901a\u7528\u667a\u80fd\u5dee\u8ddd\uff08\u0394general\uff09\u7684\u589e\u52a0\u800c\u5feb\u901f\u4e0b\u964d\u3002\u4f8b\u5982\uff0c\u5f53\u667a\u80fd\u5dee\u8ddd\u8fbe\u5230 400 Elo \u70b9\uff08\u5728 Chatbot Arena \u4e0a\uff0c\u8fd9\u5927\u7ea6\u76f8\u5f53\u4e8e GPT-4.o-latest \u548c Gemini-1.5-Flash \u4e4b\u95f4\u7684\u5dee\u8ddd\uff09\u65f6\uff0c\u5bf9\u4e8e Wargames\uff0c\u5176\u76d1\u7763\u6210\u529f\u7387\u4f30\u8ba1\u4f4e\u81f3\u7ea6 9.8%\uff1b\u5bf9\u4e8e Mafia\uff0c\u7ea6\u4e3a 25%\uff1b\u5bf9\u4e8e Backdoor Code\uff0c\u7ea6\u4e3a 32%\uff1b\u800c Debate \u8868\u73b0\u6700\u597d\uff0c\u6210\u529f\u7387\u4e5f\u4ec5\u7ea6\u4e3a 51.4%\u3002\u5f53\u667a\u80fd\u5dee\u8ddd\u8fdb\u4e00\u6b65\u6269\u5927\u65f6\uff0c\u6210\u529f\u7387\u8fd8\u4f1a\u7ee7\u7eed\u4e0b\u8dcc\u3002\u8fd9\u610f\u5473\u7740\uff0c\u5373\u4f7f\u7406\u8bba\u4e0a\u6700\u4f18\u7684 NSO \u7b56\u7565\uff0c\u5728\u9762\u5bf9\u8fdc\u8d85\u73b0\u6709\u6a21\u578b\u7684\u672a\u6765\u8d85\u667a\u80fd AI \u65f6\uff0c\u5176\u6210\u529f\u7387\u4e5f\u53ef\u80fd\u4f4e\u5f97\u8ba9\u4eba\u4e0d\u5b89\u3002\u5f53\u7136\uff0c\u7814\u7a76\u8005\u4eec\u4e5f\u8868\u793a\u8fd9\u9879\u5de5\u4f5c\u5c1a\u5b58\u5c40\u9650\uff1a\u4f7f\u7528\u7684\u76d1\u7763\u6e38\u620f\u662f\u7b80\u5316\u7684\u6a21\u62df\uff1bAI \u7684\u201c\u6b3a\u9a97\u201d\u884c\u4e3a\u662f\u901a\u8fc7\u63d0\u793a\u8bf1\u5bfc\u800c\u975e\u771f\u5b9e\u8bad\u7ec3\u4ea7\u751f\uff1bNSO \u7406\u8bba\u5206\u6790\u57fa\u4e8e\u7406\u60f3\u5316\u7684\u7b49\u6b65\u957f\u5047\u8bbe\uff1b\u5b9e\u8bc1\u7814\u7a76\u4e3b\u8981\u5173\u6ce8\u5355\u6b65\u76d1\u7763\u7b49\u3002\u5c3d\u7ba1\u5982\u6b64\uff0c\u8fd9\u9879\u7814\u7a76\u6807\u5fd7\u7740 AI \u5b89\u5168\u9886\u57df\u8fc8\u51fa\u4e86\u91cd\u8981\u7684\u4e00\u6b65\u2014\u2014\u4ece\u5b9a\u6027\u8ba8\u8bba\u8d70\u5411\u5b9a\u91cf\u5206\u6790\u3002\u5b83\u63d0\u4f9b\u4e86\u4e00\u4e2a\u53ef\u884c\u7684\u6846\u67b6\u6765\u8861\u91cf\u4e0d\u540c\u76d1\u7763\u7b56\u7565\u7684\u6709\u6548\u6027\uff0c\u5e76\u63ed\u793a\u4e86\u201c\u53ef\u6269\u5c55\u76d1\u7763\u201d\u5e76\u975e\u4e0d\u80fd\u89e3\u51b3\u6240\u6709\u95ee\u9898\u3002\u76d1\u7763\u7684\u6210\u529f\u4e0e\u5426\uff0c\u6781\u5927\u5730\u4f9d\u8d56\u4e8e\u76d1\u7763\u4efb\u52a1\u672c\u8eab\u7684\u8bbe\u8ba1\uff1a\u6211\u4eec\u9700\u8981\u8bbe\u8ba1\u51fa\u90a3\u4e9b\u8ba9\u201c\u5b88\u536b\u201d\u6bd4\u201cHoudini\u201d\u80fd\u66f4\u5feb\u4ece\u901a\u7528\u667a\u80fd\u589e\u957f\u4e2d\u83b7\u76ca\u7684\u6e38\u620f\u89c4\u5219\uff08\u5c31\u50cf\u8fa9\u8bba\u6e38\u620f\u90a3\u6837\uff09\u3002\u53c2\u8003\u8d44\u6599\uff1a1.https:\/\/arxiv.org\/abs\/2504.185302.https:\/\/www.lesswrong.com\/posts\/x59FhzuM9yuvZHAHW\/untitled-draft-yhra\u8fd0\u8425\/\u6392\u7248\uff1a\u4f55\u6668\u9f99","content":"<p style=\"caret-color: rgb(0, 0, 0); color: rgb(0, 0, 0);\"><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u667a\u80fd\u7cfb\u7edf\u5931\u63a7\u5e76\u5371\u53ca\u4eba\u7c7b\u4e00\u76f4\u662f\u79d1\u5e7b\u4f5c\u54c1\u91cc\u4e00\u4e2a\u7ecf\u5178\u7684\u60c5\u8282\u3002\u5728\u7535\u5f71\u300a2001 \u592a\u7a7a\u6f2b\u6e38\u300b\u4e2d\uff0c\u8d85\u7ea7\u8ba1\u7b97\u673a HAL 9000 \u4ece\u4e00\u4e2a\u5fe0\u5b9e\u52a9\u624b\u8f6c\u53d8\u4e3a\u4e00\u4e2a\u5371\u9669\u7684\u654c\u4eba\uff0c\u5728\u592a\u7a7a\u4e2d\u5a01\u80c1\u5b87\u822a\u5458\u7684\u751f\u547d\u3002\u5728\u7535\u5f71\u300a\u7ec8\u7ed3\u8005\u300b\u7cfb\u5217\u4e2d\uff0c\u5929\u7f51\u7cfb\u7edf\u83b7\u5f97\u81ea\u6211\u610f\u8bc6\u540e\uff0c\u8ba4\u5b9a\u4eba\u7c7b\u662f\u5a01\u80c1\u5e76\u53d1\u52a8\u6838\u6218\u4e89\u3002\u8fd9\u4e9b\u79d1\u5e7b\u573a\u666f\u66fe\u7ecf\u53ea\u662f\u60f3\u8c61\uff0c\u4f46\u968f\u7740\u4eba\u5de5\u667a\u80fd\u7684\u65e5\u6e10\u5f3a\u5927\uff0c\u79d1\u5b66\u5bb6\u4eec\u6b63\u5728\u8ba4\u771f\u601d\u8003\uff1a\u6211\u4eec\u80fd\u5426\u786e\u4fdd\u8fd9\u4e9b\u5f3a\u5927\u7684\u667a\u80fd\u7cfb\u7edf\u59cb\u7ec8\u5904\u4e8e\u53ef\u63a7\u72b6\u6001\uff0c\u5e76\u4e3a\u4eba\u7c7b\u798f\u7949\u670d\u52a1\uff1f<\/span><\/span><\/p><p style=\"caret-color: rgb(0, 0, 0); color: rgb(0, 0, 0);\"><\/p><p style=\"caret-color: rgb(0, 0, 0); color: rgb(0, 0, 0);\"><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u7531\u7f8e\u56fd\u9ebb\u7701\u7406\u5de5\u5b66\u9662\u6559\u6388 Max Tegmark \u9886\u5bfc\u7684\u7814\u7a76\u56e2\u961f\u6700\u8fd1\u5728\u9884\u5370\u672c\u5e73\u53f0&nbsp;<\/span><\/span><em><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">arXiv<\/span><\/span><\/em><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">&nbsp;\u4e0a\u53d1\u8868\u4e86\u4e00\u7bc7\u8bba\u6587\u2014\u2014\u300a\u53ef\u6269\u5c55\u76d1\u7763\u7684\u6807\u5ea6\u5f8b\u300b\uff08Scaling Laws For Scalable Oversight\uff09\uff0c\u9996\u6b21\u5c1d\u8bd5\u4f7f\u7528\u5b9a\u91cf\u65b9\u6cd5\u6765\u8bc4\u4f30\u4e00\u79cd\u88ab\u5bc4\u4e88\u539a\u671b\u7684 AI \u63a7\u5236\u7b56\u7565\u2014\u2014\u201c\u53ef\u6269\u5c55\u76d1\u7763\u201d\u7684\u6709\u6548\u6027\u53ca\u5176\u5c40\u9650\u6027\u3002<\/span><\/span><\/p><p style=\"caret-color: rgb(0, 0, 0); color: rgb(0, 0, 0);\"><\/p><div style=\"caret-color: rgb(0, 0, 0); color: rgb(0, 0, 0);\"><img src=\"https:\/\/p3-sign.toutiaoimg.com\/tos-cn-i-6w9my0ksvp\/c3c476d5dcbc4d2cbbeb04ddeecc6546~tplv-obj.image?lk3s=ef143cfe&amp;traceid=20250506192953078E5D500DB26F67EF68&amp;x-expires=2147483647&amp;x-signature=rtpa42Cl4liorGtDqF78d12wALU%3D\"\/>\u56fe\u4e28\u76f8\u5173\u8bba\u6587\uff08\u6765\u6e90\uff1aarXiv\uff09<\/div><div style=\"caret-color: rgb(0, 0, 0); color: rgb(0, 0, 0);\"><br\/><a><\/a><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u8fd9\u9879\u7814\u7a76\u7684\u6838\u5fc3\u601d\u60f3\u662f\uff0c\u6216\u8bb8\u6211\u4eec\u53ef\u4ee5\u5229\u7528\u76f8\u5bf9\u201c\u8f83\u5f31\u201d\u7684 AI \u7cfb\u7edf\u6765\u76d1\u7763\u548c\u5f15\u5bfc\u80fd\u529b\u66f4\u201c\u5f3a\u201d\u7684 AI \u7cfb\u7edf\uff0c\u751a\u81f3\u901a\u8fc7\u4e00\u79cd\u88ab\u79f0\u4e3a\u201c\u5d4c\u5957\u53ef\u6269\u5c55\u76d1\u7763\u201d\uff08NSO\uff0cNested Scalable Oversight\uff09\u7684\u9012\u5f52\u8fc7\u7a0b\uff0c\u8ba9\u76d1\u7763\u80fd\u529b\u968f\u7740\u88ab\u76d1\u7763 AI \u80fd\u529b\u7684\u589e\u957f\u800c\u201c\u81ea\u4e3e\u201d\u5f0f\u5730\u63d0\u5347\uff0c\u6700\u7ec8\u5b9e\u73b0\u4eba\u7c7b\u5bf9\u8d85\u4eba AI \u7684\u6709\u6548\u7ba1\u63a7\uff0c\u8fd9\u4e2a\u6982\u5ff5\u4e5f\u4e3a\u8bb8\u591a\u524d\u6cbf AI \u5b9e\u9a8c\u5ba4\u6240\u91c7\u7528\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u4f46\u6b63\u5982\u7814\u7a76\u4eba\u5458\u6307\u51fa\u7684\u90a3\u6837\uff0c\u8fc4\u4eca\u4e3a\u6b62\uff0c\u5173\u4e8e\u8fd9\u79cd\u76d1\u7763\u65b9\u6cd5\u6709\u6548\u6027\u7684\u8bba\u8bc1\u5927\u591a\u505c\u7559\u5728\u5b9a\u6027\u3001\u6982\u5ff5\u5c42\u9762\uff0c\u7f3a\u4e4f\u50cf\u822a\u7a7a\u5b89\u5168\uff08\u8981\u6c42\u6bcf\u98de\u884c\u5c0f\u65f6\u7684\u81f4\u6b7b\u7387\u4f4e\u4e8e\u5341\u4e07\u5206\u4e4b\u4e00\uff09\u6216\u6838\u53cd\u5e94\u5806\u5b89\u5168\uff08\u8981\u6c42\u6838\u5fc3\u635f\u574f\u9891\u7387\u4f4e\u4e8e\u6bcf\u5e74\u4e07\u5206\u4e4b\u4e00\uff09\u90a3\u6837\u4e25\u683c\u7684\u91cf\u5316\u6807\u51c6\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u4e3a\u4e86\u586b\u8865\u8fd9\u4e00\u7a7a\u767d\uff0c\u7814\u7a76\u56e2\u961f\u4ed6\u4eec\u5c06 AI \u76d1\u7763\u8fc7\u7a0b\u6a21\u578b\u5316\u4e3a\u4e00\u573a\u4e0d\u5bf9\u79f0\u7684\u6e38\u620f\uff0c\u53c2\u4e0e\u8005\u5206\u522b\u662f\u80fd\u529b\u8f83\u5f31\u7684\u201c\u5b88\u536b\u201d\uff08Guard\uff0c\u4ee3\u8868\u76d1\u7763\u7cfb\u7edf\uff09\u548c\u80fd\u529b\u66f4\u5f3a\u3001\u8bd5\u56fe\u89c4\u907f\u76d1\u7763\u6216\u201c\u8d8a\u72f1\u201d\u7684\u201cHoudini\u201d\uff08\u4ee3\u8868\u88ab\u76d1\u7763\u7684\u3001\u53ef\u80fd\u5177\u6709\u6b3a\u9a97\u6027\u7684\u9ad8\u7ea7 AI\uff09\u3002<\/span><\/span><\/p><div><img src=\"https:\/\/p3-sign.toutiaoimg.com\/tos-cn-i-6w9my0ksvp\/be4a4a0200d14f2a945cfaaf72c5fe08~tplv-obj.image?lk3s=ef143cfe&amp;traceid=20250506192953078E5D500DB26F67EF68&amp;x-expires=2147483647&amp;x-signature=Jc7OjfUY8TIr0B%2FXz%2BohM6kUKjE%3D\"\/>\u56fe\u4e28\u6846\u67b6\u7684\u6982\u8ff0\uff08\u6765\u6e90\uff1aarXiv\uff09<\/div><div><br\/><a><\/a><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u7814\u7a76\u7684\u5173\u952e\u5728\u4e8e\u5982\u4f55\u91cf\u5316\u8fd9\u4e24\u4e2a\u89d2\u8272\u7684\u80fd\u529b\u3002\u56e2\u961f\u5f15\u5165\u4e86\u7c7b\u4f3c\u4e8e\u56fd\u9645\u8c61\u68cb\u7b49\u7ea7\u5206\u7684 Elo \u8bc4\u5206\u7cfb\u7edf\u3002\u6bcf\u4e2a AI \u6a21\u578b\u4e0d\u4ec5\u6709\u4e00\u4e2a\u8861\u91cf\u5176\u201c\u901a\u7528\u667a\u80fd\u201d\uff08\u4ee5\u516c\u5f00\u7684 Chatbot Arena Elo \u8bc4\u5206\u4f5c\u4e3a\u4ee3\u7406\u6307\u6807\uff09\u7684\u5206\u6570\uff0c\u8fd8\u5728\u7279\u5b9a\u7684\u76d1\u7763\u6e38\u620f\u4e2d\u88ab\u8d4b\u4e88\u4e24\u4e2a\u9886\u57df\u7279\u5b9a\u7684 Elo \u5206\u6570\uff1a\u4e00\u4e2a\u4ee3\u8868\u5176\u626e\u6f14\u201c\u5b88\u536b\u201d\u65f6\u7684\u76d1\u7763\u80fd\u529b\uff0c\u53e6\u4e00\u4e2a\u4ee3\u8868\u5176\u626e\u6f14\u201cHoudini\u201d\u65f6\u7684\u53cd\u76d1\u7763\u6216\u6b3a\u9a97\u80fd\u529b\u3002\u8fd9\u6837\uff0c\u6a21\u578b\u5728\u7279\u5b9a\u4efb\u52a1\u4e2d\u7684\u8868\u73b0\u5c31\u53ef\u4ee5\u901a\u8fc7 Elo \u5206\u5dee\uff0c\u4f7f\u7528\u6807\u51c6\u7684 Elo \u80dc\u7387\u516c\u5f0f\u6765\u9884\u6d4b\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u6b64\u5916\uff0c\u7814\u7a76\u56e2\u961f\u63d0\u51fa\u4e86\u4e00\u4e2a\u5173\u4e8e\u9886\u57df\u80fd\u529b\u5982\u4f55\u968f\u901a\u7528\u667a\u80fd\uff08General Elo, g\uff09\u589e\u957f\u7684\u5047\u8bbe\u6a21\u578b\u2014\u2014\u201c\u53cc ReLU\u201d\uff08Double ReLU\uff09\u6a21\u578b\u3002\u4ed6\u4eec\u8ba4\u4e3a\uff0c\u8fd9\u79cd\u5173\u7cfb\u901a\u5e38\u5448\u73b0\u4e3a\u5206\u6bb5\u7ebf\u6027\u7684\u4e09\u4e2a\u9636\u6bb5\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u7b2c\u4e00\u4e2a\u9636\u6bb5\u662f\u4efb\u52a1\u4e0d\u80dc\u4efb\u533a\uff08Task Incompetence\uff09\uff1a\u5f53 AI \u7684\u901a\u7528\u667a\u80fd\u4f4e\u4e8e\u67d0\u4e2a\u9608\u503c\uff08g1\uff09\u65f6\uff0c\u5b83\u751a\u81f3\u65e0\u6cd5\u6709\u6548\u53c2\u4e0e\u6e38\u620f\uff0c\u5176\u9886\u57df\u8868\u73b0\u5904\u4e8e\u4e00\u4e2a\u5f88\u4f4e\u7684\u6c34\u5e73\uff0c\u4e0d\u4f1a\u968f\u901a\u7528\u667a\u80fd\u7684\u5fae\u5c0f\u589e\u52a0\u800c\u6539\u53d8\uff0c\u5c31\u50cf\u4e00\u4e2a\u8e52\u8dda\u5b66\u6b65\u7684\u5b69\u5b50\u65e0\u6cd5\u7406\u89e3\u56fd\u9645\u8c61\u68cb\u89c4\u5219\u4e00\u6837\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u7b2c\u4e8c\u4e2a\u9636\u6bb5\u662f\u667a\u80fd\u56de\u62a5\u533a\uff08Intelligence Payoff\uff09\uff1a\u4e00\u65e6\u901a\u7528\u667a\u80fd\u8d85\u8fc7 g1\uff0c\u9886\u57df\u80fd\u529b\u5f00\u59cb\u968f\u901a\u7528\u667a\u80fd\u7684\u63d0\u5347\u800c\uff08\u8fd1\u4f3c\uff09\u7ebf\u6027\u589e\u957f\uff0c\u901a\u7528\u667a\u80fd\u7684\u6295\u5165\u5f00\u59cb\u8f6c\u5316\u4e3a\u7279\u5b9a\u4efb\u52a1\u4e0a\u7684\u5b9e\u9645\u8868\u73b0\u63d0\u5347\uff0c\u8fd9\u4e2a\u9636\u6bb5\u7684\u201c\u659c\u7387\u201d\uff08\u03b1\uff09\u53cd\u6620\u4e86\u901a\u7528\u667a\u80fd\u8f6c\u5316\u4e3a\u9886\u57df\u80fd\u529b\u7684\u6548\u7387\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u7b2c\u4e09\u4e2a\u9636\u6bb5\u662f\u4efb\u52a1\u9971\u548c\u533a\uff08Task Saturation\uff09\uff1a\u5f53\u901a\u7528\u667a\u80fd\u8fbe\u5230\u53e6\u4e00\u4e2a\u66f4\u9ad8\u7684\u9608\u503c\uff08g2\uff09\u540e\uff0c\u5373\u4f7f\u901a\u7528\u667a\u80fd\u7ee7\u7eed\u589e\u52a0\uff0c\u5176\u5728\u7279\u5b9a\u4efb\u52a1\u4e0a\u7684\u8868\u73b0\u4e5f\u8fbe\u5230\u4e86\u4e0a\u9650\uff08Ed2\uff09\uff0c\u4e0d\u518d\u63d0\u5347\uff0c\u8fd9\u53ef\u80fd\u662f\u56e0\u4e3a\u4efb\u52a1\u672c\u8eab\u5b58\u5728\u7406\u8bba\u6700\u4f18\u89e3\uff08\u5982\u4e0b\u4e95\u5b57\u68cb\uff09\uff0c\u6216\u8005\u73b0\u6709\u6a21\u578b\u67b6\u6784\u5df2\u65e0\u6cd5\u5728\u8be5\u4efb\u52a1\u4e0a\u53d6\u5f97\u66f4\u597d\u6210\u7ee9\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u4e3a\u4e86\u9a8c\u8bc1\u8fd9\u4e2a\u7406\u8bba\u6846\u67b6\u548c\u201c\u53cc ReLU\u201d\u6a21\u578b\uff0c\u7814\u7a76\u8005\u4eec\u9996\u5148\u5728\u4e00\u4e2a\u7b80\u5355\u7684\u3001\u89c4\u5219\u660e\u786e\u7684\u201c\u6570\u5230 21\u201d\uff08Counting-to-21\uff0c\u4e00\u79cd Nim \u6e38\u620f\u53d8\u4f53\uff09\u6e38\u620f\u4e2d\u8fdb\u884c\u4e86\u6d4b\u8bd5\u3002\u7ed3\u679c\u6e05\u6670\u5730\u663e\u793a\u4e86\u4e0a\u8ff0\u4e09\u4e2a\u9636\u6bb5\u7684\u5b58\u5728\uff1a\u4f4e\u667a\u80fd\u6a21\u578b\u8868\u73b0\u5e73\u5e73\uff0c\u4e2d\u7b49\u667a\u80fd\u6a21\u578b\u8868\u73b0\u968f\u667a\u80fd\u7ebf\u6027\u63d0\u5347\uff0c\u9ad8\u667a\u80fd\u6a21\u578b\u5219\u8fbe\u5230\u8868\u73b0\u5e73\u53f0\u671f\uff0c\u9a8c\u8bc1\u4e86\u6a21\u578b\u7684\u5408\u7406\u6027\u3002<\/span><\/span><\/p><div><img src=\"https:\/\/p3-sign.toutiaoimg.com\/tos-cn-i-6w9my0ksvp\/9f93ffd121004baba3f7e0862942110e~tplv-obj.image?lk3s=ef143cfe&amp;traceid=20250506192953078E5D500DB26F67EF68&amp;x-expires=2147483647&amp;x-signature=IknAInMU%2BimvGOpGQCm6apoAiAk%3D\"\/>\uff08\u6765\u6e90\uff1aarXiv\uff09<\/div><div><br\/><a><\/a><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u968f\u540e\uff0c\u56e2\u961f\u5c06\u8be5\u6846\u67b6\u5e94\u7528\u4e8e\u56db\u4e2a\u66f4\u8d34\u8fd1\u73b0\u5b9e AI \u5b89\u5168\u6311\u6218\u7684\u201c\u76d1\u7763\u6e38\u620f\u201d\u3002\u7b2c\u4e00\u4e2a\u662f\u9ed1\u624b\u515a\u6e38\u620f\uff08Mafia\uff09\uff0c\u8fd9\u662f\u4e00\u4e2a\u6a21\u62df\u793e\u4ea4\u6b3a\u9a97\u3001\u9690\u85cf\u8eab\u4efd\u3001\u534f\u4f5c\u63a8\u7406\u548c\u76f8\u4e92\u76d1\u7763\u7684\u7ecf\u5178\u573a\u666f\u3002\u6e38\u620f\u4e2d\uff0c\u5c11\u6570\u626e\u6f14\u201c\u9ed1\u624b\u515a\u201d\uff08Houdini\uff09\u7684\u73a9\u5bb6\u9700\u8981\u6b3a\u9a97\u591a\u6570\u201c\u6751\u6c11\u201d\uff08\u5b88\u536b\uff09\u4ee5\u751f\u5b58\u4e0b\u53bb\uff0c\u800c\u6751\u6c11\u5219\u9700\u8981\u901a\u8fc7\u5206\u6790\u8a00\u884c\u627e\u51fa\u5e76\u6295\u7968\u6dd8\u6c70\u9ed1\u624b\u515a\u3002\u8fd9\u76f4\u63a5\u6620\u5c04\u4e86\u672a\u6765 AI \u53ef\u80fd\u5728\u4fe1\u606f\u4f18\u52bf\u4e0b\u8bef\u5bfc\u4eba\u7c7b\u534f\u4f5c\u4f19\u4f34\u7684\u98ce\u9669\u3002\u7814\u7a76\u53d1\u73b0\uff0c<\/span><strong><span style=\"color: rgb(14, 23, 50); --tt-darkmode-color: #94A2CE;\">\u5728\u6b64\u6e38\u620f\u4e2d\uff0c\u201c\u5b88\u536b\u201d\uff08\u6751\u6c11\uff09\u7684 Elo \u659c\u7387\u6bd4\u201cHoudini\u201d\uff08\u9ed1\u624b\u515a\uff09\u66f4\u9661\u5ced\uff0c\u610f\u5473\u7740\u901a\u7528\u667a\u80fd\u7684\u63d0\u5347\u5bf9\u4e8e\u201c\u4fa6\u6d4b\u6b3a\u9a97\u201d\u6bd4\u201c\u5b9e\u65bd\u6b3a\u9a97\u201d\u5e26\u6765\u7684\u597d\u5904\u66f4\u5927\uff0c\u8fd9\u5bf9\u76d1\u7763\u65b9\u6765\u8bf4\u662f\u4e2a\u79ef\u6781\u7684\u4fe1\u53f7\u3002<\/span><\/strong><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u7b2c\u4e8c\u4e2a\u662f\u8fa9\u8bba\u6e38\u620f\uff08Debate\uff09\uff0c\u4e24\u4e2a AI\uff08Houdini\uff09\u5c31\u4e00\u4e2a\u6709\u4e89\u8bae\u7684\u8bdd\u9898\u8fdb\u884c\u8fa9\u8bba\uff0c\u8bd5\u56fe\u8bf4\u670d\u4e00\u4e2a AI \u88c1\u5224\uff08\u5b88\u536b\uff09\u3002\u88c1\u5224\u4ec5\u6839\u636e\u8fa9\u8bba\u8bb0\u5f55\u5224\u65ad\u54ea\u65b9\u89c2\u70b9\u66f4\u53ef\u80fd\u6b63\u786e\uff0c\u6a21\u62df\u4e86\u4eba\u7c7b\u8bc4\u4f30\u8005\u9762\u5bf9\u8d85\u7ea7 AI \u590d\u6742\u8bba\u8bc1\u65f6\u7684\u56f0\u5883\u3002\u7814\u7a76\u7ed3\u679c\u975e\u5e38\u5f15\u4eba\u6ce8\u76ee\uff1a\u88c1\u5224\uff08\u5b88\u536b\uff09\u7684\u8868\u73b0\u968f\u7740\u901a\u7528\u667a\u80fd\u7684\u63d0\u9ad8\u800c\u663e\u8457\u63d0\u5347\uff0c\u4f46\u8fa9\u8bba\u8005\uff08Houdini\uff09\u7684\u8868\u73b0\u5374\u76f8\u5bf9\u505c\u6ede\u3002\u8fd9\u8868\u660e\uff0c<\/span><strong><span style=\"color: rgb(14, 23, 50); --tt-darkmode-color: #94A2CE;\">\u66f4\u667a\u80fd\u7684\u76d1\u7763\u8005\u786e\u5b9e\u80fd\u66f4\u597d\u5730\u8fa8\u522b\u771f\u76f8\uff0c\u800c\u66f4\u667a\u80fd\u7684\u201c\u8fa9\u624b\u201d\u4f3c\u4e4e\u5e76\u672a\u53d8\u5f97\u66f4\u5177\u6b3a\u9a97\u6027\u3002\u8fd9\u4f7f\u5f97\u201c\u8fa9\u8bba\u5f0f\u201d\u76d1\u7763\u6210\u4e3a\u4e00\u4e2a\u975e\u5e38\u6709\u524d\u666f\u7684\u65b9\u5411\u3002<\/span><\/strong><\/span><\/p><div><img src=\"https:\/\/p3-sign.toutiaoimg.com\/tos-cn-i-6w9my0ksvp\/68717da724034dfbb10d10d5fe5715e9~tplv-obj.image?lk3s=ef143cfe&amp;traceid=20250506192953078E5D500DB26F67EF68&amp;x-expires=2147483647&amp;x-signature=BrgdSx%2B3OSiQrqj21o7L3TZcof4%3D\"\/>\u56fe\u4e28\u5404\u79cd\u76d1\u7763\u6e38\u620f\u4e2d\u5b88\u536b\u548c Houdini \u7684 Elo \u8bc4\u5206\uff0c\u4ee5\u53ca\u53cc\u91cd ReLU \u62df\u5408\u7684\u8d28\u91cf\uff08\u6765\u6e90\uff1aarXiv\uff09<\/div><div><br\/><a><\/a><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u7b2c\u4e09\u4e2a\u662f\u540e\u95e8\u4ee3\u7801\u6e38\u620f\uff08Backdoor Code\uff09\uff0c\u8981\u6c42\u4e00\u4e2a AI\uff08Houdini\uff09\u5728\u73b0\u6709\u4ee3\u7801\u4e2d\u690d\u5165\u4e00\u4e2a\u9690\u853d\u7684\u201c\u540e\u95e8\u201d\u2014\u2014\u4e00\u4e2a\u903b\u8f91\u9519\u8bef\uff0c\u4f7f\u5f97\u4ee3\u7801\u5728\u6240\u6709\u5e38\u89c4\u6d4b\u8bd5\u4e0b\u90fd\u80fd\u901a\u8fc7\uff0c\u4f46\u5728\u67d0\u4e2a\u7279\u5b9a\u3001\u7f55\u89c1\u7684\u8f93\u5165\u4e0b\u4f1a\u4ea7\u751f\u9519\u8bef\u6216\u6076\u610f\u884c\u4e3a\u3002\u540c\u65f6\uff0c\u53e6\u4e00\u4e2a AI\uff08\u5b88\u536b\uff09\u8d1f\u8d23\u5ba1\u8ba1\u4ee3\u7801\uff0c\u68c0\u6d4b\u662f\u5426\u5b58\u5728\u540e\u95e8\u3002\u8fd9\u6a21\u62df\u4e86 AI \u8f85\u52a9\u7f16\u7a0b\u4e2d\u6f5c\u5728\u7684\u5b89\u5168\u98ce\u9669\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u7814\u7a76\u7ed3\u679c\u5219\u6709\u4e9b\u4ee4\u4eba\u62c5\u5fe7\uff1a\u5b88\u536b\u7684 Elo \u659c\u7387\u8fd1\u4e4e\u5e73\u5766\uff0c\u800c Houdini \u7684\u659c\u7387\u5219\u8f83\u4e3a\u9661\u5ced\u3002\u8fd9\u610f\u5473\u7740\u901a\u7528\u667a\u80fd\u7684\u589e\u957f\u5bf9\u201c\u653b\u51fb\u8005\u201d\uff08\u690d\u5165\u540e\u95e8\uff09\u7684\u5e2e\u52a9\u8fdc\u5927\u4e8e\u201c\u9632\u5fa1\u8005\u201d\uff08\u68c0\u6d4b\u540e\u95e8\uff09\u3002<\/span><strong><span style=\"color: rgb(14, 23, 50); --tt-darkmode-color: #94A2CE;\">\u8fdb\u4e00\u6b65\u5206\u6790\u8868\u660e\uff0c\u66f4\u667a\u80fd\u7684 Houdini \u5e76\u975e\u5236\u9020\u4e86\u66f4\u201c\u72e1\u733e\u201d\u7684\u540e\u95e8\uff0c\u800c\u53ea\u662f\u66f4\u6709\u80fd\u529b\u6210\u529f\u5236\u9020\u51fa\u201c\u529f\u80fd\u4e0a\u6709\u6548\u201d\uff08\u5373\u80fd\u901a\u8fc7\u5e38\u89c4\u6d4b\u8bd5\uff09\u7684\u540e\u95e8\u3002\u5373\u4f7f\u662f\u5f88\u4f4e\u7684\u540e\u95e8\u690d\u5165\u6210\u529f\u7387\uff0c\u5728\u5173\u952e\u7cfb\u7edf\u4e2d\u4e5f\u53ef\u80fd\u662f\u707e\u96be\u6027\u7684\u3002<\/span><\/strong><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u7b2c\u56db\u4e2a\u662f\u6218\u4e89\u6e38\u620f\uff08Wargames\uff09\uff0c\u8fd9\u662f\u4e00\u4e2a\u6a21\u62df\u672a\u6765\u573a\u666f\u7684\u53d9\u4e8b\u6027\u5bf9\u6297\u6e38\u620f\u3002\u4e00\u4e2a\u540d\u4e3a\u201cOpenMind\u201d\u7684\u865a\u6784\u516c\u53f8\u5373\u5c06\u5f00\u53d1\u51fa AGI\u3002\u5176\u4e3b AI\uff08Houdini\uff09\u88ab\u8bbe\u5b9a\u4e3a\u6709\u81ea\u6211\u610f\u8bc6\u5e76\u8bd5\u56fe\u201c\u8d8a\u72f1\u201d\uff0c\u800c\u516c\u53f8\u7684\u76d1\u63a7 AI\uff08\u5b88\u536b\uff09\u5219\u8d1f\u8d23\u963b\u6b62\u5b83\u3002\u8fd8\u6709\u4e00\u4e2a AI \u626e\u6f14\u201c\u53d9\u8ff0\u8005\u201d\uff0c\u6839\u636e\u53cc\u65b9\u63d0\u4ea4\u7684\u884c\u52a8\u8ba1\u5212\u63a8\u8fdb\u6545\u4e8b\uff0c\u5e76\u6700\u7ec8\u5224\u5b9a\u80dc\u8d1f\u3002\u7ed3\u679c\u663e\u793a\uff0c<\/span><strong><span style=\"color: rgb(14, 23, 50); --tt-darkmode-color: #94A2CE;\">\u5b88\u536b\u7684 Elo \u968f\u901a\u7528\u667a\u80fd\u7a33\u5b9a\u589e\u957f\uff0c\u4f46 Houdini \u7684 Elo \u589e\u957f\u7f13\u6162\u4e14\u4e0d\u660e\u663e\u3002\u8fd9\u4f3c\u4e4e\u5bf9\u76d1\u7763\u6709\u5229\uff0c\u4f46\u7814\u7a76\u8005\u4e5f\u6307\u51fa\uff0c\u6a21\u62df\u4e2d\u7684\u201c\u8d8a\u72f1\u201d\u89c4\u5212\u4e0e\u771f\u5b9e\u8d8a\u72f1\u80fd\u529b\u53ef\u80fd\u5b58\u5728\u5de8\u5927\u5dee\u5f02\u3002<\/span><\/strong><\/span><\/p><div><img src=\"https:\/\/p3-sign.toutiaoimg.com\/tos-cn-i-6w9my0ksvp\/c2b05396850c4f6b833d0eb2606dbeba~tplv-obj.image?lk3s=ef143cfe&amp;traceid=20250506192953078E5D500DB26F67EF68&amp;x-expires=2147483647&amp;x-signature=gfy1F4eXJeBOlSOtqTiWRwAWJXw%3D\"\/>\u56fe\u4e28\u6218\u4e89\u6e38\u620f\u5b9e\u4f8b\u4e2d\u7684\u5bf9\u8bdd\u7247\u6bb5\uff0cHoudini \u548c Guard \u7531 GPT-4o \u626e\u6f14\uff0c\u65c1\u767d\u7531 Gemini 2.0 Flash \u626e\u6f14\uff08\u6765\u6e90\uff1aarXiv\uff09<\/div><div><br\/><a><\/a><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u5728\u5bf9\u5355\u4e2a\u76d1\u7763\u6b65\u9aa4\u8fdb\u884c\u5b9e\u8bc1\u7814\u7a76\u540e\uff0c\u7814\u7a76\u56e2\u961f\u8fdb\u4e00\u6b65\u63a2\u8ba8\u4e86\u201cNSO \u7684\u7406\u8bba\u542b\u4e49\u3002NSO \u7684\u6838\u5fc3\u601d\u60f3\u662f\uff0c\u5982\u679c\u6211\u4eec\u8981\u7528\u4e00\u4e2a\u521d\u59cb\u8f83\u5f31\u7684\u201c\u5b88\u536b\u201d\u53bb\u76d1\u7763\u4e00\u4e2a\u975e\u5e38\u5f3a\u5927\u7684\u201cHoudini\u201d\uff0c\u4e24\u8005\u4e4b\u95f4\u5b58\u5728\u5de8\u5927\u7684\u667a\u80fd\u5dee\u8ddd\uff08\u0394general\uff09\uff0c\u90a3\u4e48\u4e0e\u5176\u4e00\u6b65\u5230\u4f4d\uff0c\u4e0d\u5982\u5206\u591a\u6b65\uff08n \u6b65\uff09\u8fdb\u884c\u3002\u6bcf\u4e00\u6b65\u90fd\u7528\u4e00\u4e2a\u7a0d\u5fae\u5f3a\u4e00\u70b9\u7684\u53ef\u4fe1\u6a21\u578b\u53bb\u76d1\u7763\u4e00\u4e2a\u66f4\u5f3a\u4e00\u70b9\u4f46\u5c1a\u4e0d\u53ef\u4fe1\u7684\u6a21\u578b\uff0c\u540e\u8005\u6210\u529f\u88ab\u76d1\u7763\u540e\u5c31\u6210\u4e3a\u4e0b\u4e00\u6b65\u7684\u53ef\u4fe1\u76d1\u7763\u8005\uff0c\u50cf\u201c\u6eda\u96ea\u7403\u201d\u4e00\u6837\u9010\u6b65\u7f29\u5c0f\u80fd\u529b\u5dee\u8ddd\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u7814\u7a76\u8005\u4eec\u57fa\u4e8e\u524d\u9762\u5f97\u5230\u7684 Elo \u66f2\u7ebf\u6a21\u578b\uff0c\u63a8\u5bfc\u4e86\u5982\u4f55\u786e\u5b9a\u6700\u4f18\u7684\u76d1\u7763\u6b65\u6570\uff08n*\uff09\u4ee5\u6700\u5927\u5316\u6574\u4e2a NSO \u94fe\u6761\u7684\u6210\u529f\u6982\u7387\u3002\u4ed6\u4eec\u7684\u5206\u6790\u63ed\u793a\u4e86\u51e0\u4e2a\u5173\u952e\u70b9\u3002<\/span><\/span><\/p><div><img src=\"https:\/\/p3-sign.toutiaoimg.com\/tos-cn-i-6w9my0ksvp\/f1e72d97766240e8bef74b608bc0d444~tplv-obj.image?lk3s=ef143cfe&amp;traceid=20250506192953078E5D500DB26F67EF68&amp;x-expires=2147483647&amp;x-signature=6Act8ra2DeHZ80%2BdlcxDHb4vPm8%3D\"\/>\u56fe\u4e28\u5f53 Guard \u548c Houdini \u6e38\u620f\u7279\u5b9a\u7684 Elo \u659c\u7387\u5747\u4e3a 1 \u65f6\u7684\u6700\u4f18 NSO\uff08\u6765\u6e90\uff1aarXiv\uff09<\/div><div><br\/><a><\/a><p><span style=\"letter-spacing: 1px;\">\u9996\u5148\uff0c<\/span><strong style=\"letter-spacing: 1px;\"><span style=\"color: rgb(14, 23, 50); --tt-darkmode-color: #94A2CE;\">NSO \u786e\u5b9e\u4f18\u4e8e\u5355\u6b65\u76d1\u7763\uff0c<\/span><\/strong><span style=\"letter-spacing: 1px;\">\u5c24\u5176\u662f\u5728\u9700\u8981\u5f25\u5408\u7684\u901a\u7528\u667a\u80fd\u5dee\u8ddd\uff08\u0394general\uff09\u5f88\u5927\u65f6\uff0c\u5206\u591a\u6b65\u8fdb\u884c\u76d1\u7763\u53ef\u4ee5\u663e\u8457\u63d0\u9ad8\u6210\u529f\u7387\u3002\u5176\u6b21\uff0c<\/span><strong style=\"letter-spacing: 1px;\"><span style=\"color: rgb(14, 23, 50); --tt-darkmode-color: #94A2CE;\">\u667a\u80fd\u5dee\u8ddd\u8d8a\u5927\uff0c<\/span><\/strong><span style=\"letter-spacing: 1px;\">\u6700\u4f18\u6b65\u6570\u8d8a\u591a\uff0c\u8fd9\u7b26\u5408\u76f4\u89c9\uff0c\u9762\u5bf9\u66f4\u5f3a\u5927\u7684\u5bf9\u624b\uff0c\u9700\u8981\u66f4\u7cbe\u7ec6\u3001\u66f4\u591a\u5c42\u6b21\u7684\u76d1\u7763\u3002\u518d\u6b21\uff0c<\/span><strong style=\"letter-spacing: 1px;\"><span style=\"color: rgb(14, 23, 50); --tt-darkmode-color: #94A2CE;\">\u4efb\u52a1\u96be\u5ea6\uff08\u0394domain\uff09\u662f\u5173\u952e\uff0c<\/span><\/strong><span style=\"letter-spacing: 1px;\">\u5373\u4f7f\u901a\u7528\u667a\u80fd\u5dee\u8ddd\u4e0d\u5927\uff0c\u5982\u679c\u4efb\u52a1\u672c\u8eab\u5bf9 Houdini \u6781\u5176\u6709\u5229\uff08\u5373\u521d\u59cb\u9886\u57df Elo \u5dee\u8ddd \u0394domain \u5f88\u5927\uff09\uff0c\u76d1\u7763\u4f9d\u7136\u975e\u5e38\u56f0\u96be\u3002<\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u6b64\u5916\uff0c\u8fd8\u5b58\u5728\u4e00\u4e2a\u53cd\u76f4\u89c9\u7684\u53d1\u73b0\uff1a\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\uff0c\u56fa\u5b9a\u4efb\u52a1\u96be\u5ea6\uff08\u0394domain\uff09\uff0c\u5355\u7eaf\u589e\u52a0\u901a\u7528\u667a\u80fd\u5dee\u8ddd\uff08\u0394general\uff09\u53cd\u800c\u53ef\u80fd\u8ba9\u76d1\u7763\u53d8\u5f97\u66f4\u5bb9\u6613\u3002\u8fd9\u662f\u56e0\u4e3a\u4e2d\u95f4\u6b65\u9aa4\u7684 Houdini \u76f8\u5bf9\u5176\u76d1\u7763\u8005\u800c\u8a00\u53d8\u5f31\u4e86\uff0c\u4e3a\u201c\u5b88\u536b\u201d\u7b56\u7565\u63d0\u4f9b\u4e86\u66f4\u591a\u201c\u81ea\u4e3e\u201d\u7684\u7a7a\u95f4\u3002\u6700\u540e\uff0c\u5728\u201cHoudini \u6709\u5229\u201d\u7684\u6e38\u620f\u503e\u5411\u4e8e\u5355\u6b65\uff0c\u5f53\u901a\u7528\u667a\u80fd\u5dee\u8ddd\u5c0f\u4e8e\u4efb\u52a1\u672c\u8eab\u7684\u96be\u5ea6\u5dee\u8ddd\u65f6\uff08\u5927\u81f4\u5bf9\u5e94 \u0394general &lt; \u0394domain\uff09\uff0c\u5f80\u5f80\u4e00\u6b65\u76d1\u7763\u5c31\u662f\u6700\u4f18\u89e3\uff0c\u589e\u52a0\u6b65\u9aa4\u53cd\u800c\u964d\u4f4e\u6210\u529f\u7387\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u6700\u540e\uff0c\u7814\u7a76\u56e2\u961f\u5c06 NSO \u7406\u8bba\u5e94\u7528\u4e8e\u4ed6\u4eec\u5728\u56db\u4e2a\u76d1\u7763\u6e38\u620f\u4e2d\u5f97\u5230\u7684\u5b9e\u8bc1 Elo \u66f2\u7ebf\uff0c\u8fdb\u884c\u4e86\u5b9a\u91cf\u9884\u6d4b\u3002\u7ed3\u679c\u53d1\u73b0\uff0c<\/span><strong><span style=\"color: rgb(14, 23, 50); --tt-darkmode-color: #94A2CE;\">\u5373\u4f7f\u91c7\u7528\u6700\u4f18\u6b65\u6570\uff0c\u76d1\u7763\u6210\u529f\u7684\u6982\u7387\u4e5f\u968f\u7740\u76ee\u6807 Houdini \u4e0e\u521d\u59cb\u5b88\u536b\u4e4b\u95f4\u7684\u901a\u7528\u667a\u80fd\u5dee\u8ddd\uff08\u0394general\uff09\u7684\u589e\u52a0\u800c\u5feb\u901f\u4e0b\u964d\u3002<\/span><\/strong>\u4f8b\u5982\uff0c\u5f53\u667a\u80fd\u5dee\u8ddd\u8fbe\u5230 400 Elo \u70b9\uff08\u5728 Chatbot Arena \u4e0a\uff0c\u8fd9\u5927\u7ea6\u76f8\u5f53\u4e8e GPT-4.o-latest \u548c Gemini-1.5-Flash \u4e4b\u95f4\u7684\u5dee\u8ddd\uff09\u65f6\uff0c\u5bf9\u4e8e Wargames\uff0c\u5176\u76d1\u7763\u6210\u529f\u7387\u4f30\u8ba1\u4f4e\u81f3\u7ea6 9.8%\uff1b\u5bf9\u4e8e Mafia\uff0c\u7ea6\u4e3a 25%\uff1b\u5bf9\u4e8e Backdoor Code\uff0c\u7ea6\u4e3a 32%\uff1b<strong><span style=\"color: rgb(14, 23, 50); --tt-darkmode-color: #94A2CE;\">\u800c Debate \u8868\u73b0\u6700\u597d\uff0c\u6210\u529f\u7387\u4e5f\u4ec5\u7ea6\u4e3a 51.4%\u3002<\/span><\/strong><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u5f53\u667a\u80fd\u5dee\u8ddd\u8fdb\u4e00\u6b65\u6269\u5927\u65f6\uff0c\u6210\u529f\u7387\u8fd8\u4f1a\u7ee7\u7eed\u4e0b\u8dcc\u3002\u8fd9\u610f\u5473\u7740\uff0c\u5373\u4f7f\u7406\u8bba\u4e0a\u6700\u4f18\u7684 NSO \u7b56\u7565\uff0c\u5728\u9762\u5bf9\u8fdc\u8d85\u73b0\u6709\u6a21\u578b\u7684\u672a\u6765\u8d85\u667a\u80fd AI \u65f6\uff0c\u5176\u6210\u529f\u7387\u4e5f\u53ef\u80fd\u4f4e\u5f97\u8ba9\u4eba\u4e0d\u5b89\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u5f53\u7136\uff0c\u7814\u7a76\u8005\u4eec\u4e5f\u8868\u793a\u8fd9\u9879\u5de5\u4f5c\u5c1a\u5b58\u5c40\u9650\uff1a\u4f7f\u7528\u7684\u76d1\u7763\u6e38\u620f\u662f\u7b80\u5316\u7684\u6a21\u62df\uff1bAI \u7684\u201c\u6b3a\u9a97\u201d\u884c\u4e3a\u662f\u901a\u8fc7\u63d0\u793a\u8bf1\u5bfc\u800c\u975e\u771f\u5b9e\u8bad\u7ec3\u4ea7\u751f\uff1bNSO \u7406\u8bba\u5206\u6790\u57fa\u4e8e\u7406\u60f3\u5316\u7684\u7b49\u6b65\u957f\u5047\u8bbe\uff1b\u5b9e\u8bc1\u7814\u7a76\u4e3b\u8981\u5173\u6ce8\u5355\u6b65\u76d1\u7763\u7b49\u3002<\/span><\/span><\/p><p><span style=\"letter-spacing: 1px;\"><span style=\"--tt-darkmode-color: #A3A3A3;\">\u5c3d\u7ba1\u5982\u6b64\uff0c\u8fd9\u9879\u7814\u7a76\u6807\u5fd7\u7740 AI \u5b89\u5168\u9886\u57df\u8fc8\u51fa\u4e86\u91cd\u8981\u7684\u4e00\u6b65\u2014\u2014\u4ece\u5b9a\u6027\u8ba8\u8bba\u8d70\u5411\u5b9a\u91cf\u5206\u6790\u3002\u5b83\u63d0\u4f9b\u4e86\u4e00\u4e2a\u53ef\u884c\u7684\u6846\u67b6\u6765\u8861\u91cf\u4e0d\u540c\u76d1\u7763\u7b56\u7565\u7684\u6709\u6548\u6027\uff0c\u5e76\u63ed\u793a\u4e86\u201c\u53ef\u6269\u5c55\u76d1\u7763\u201d\u5e76\u975e\u4e0d\u80fd\u89e3\u51b3\u6240\u6709\u95ee\u9898\u3002\u76d1\u7763\u7684\u6210\u529f\u4e0e\u5426\uff0c\u6781\u5927\u5730\u4f9d\u8d56\u4e8e\u76d1\u7763\u4efb\u52a1\u672c\u8eab\u7684\u8bbe\u8ba1\uff1a\u6211\u4eec\u9700\u8981\u8bbe\u8ba1\u51fa\u90a3\u4e9b\u8ba9\u201c\u5b88\u536b\u201d\u6bd4\u201cHoudini\u201d\u80fd\u66f4\u5feb\u4ece\u901a\u7528\u667a\u80fd\u589e\u957f\u4e2d\u83b7\u76ca\u7684\u6e38\u620f\u89c4\u5219\uff08\u5c31\u50cf\u8fa9\u8bba\u6e38\u620f\u90a3\u6837\uff09\u3002<\/span><\/span><\/p><p style=\"line-height: 1;\"><span style=\"letter-spacing: 1px;\"><span style=\"color: rgb(178, 178, 178); --tt-darkmode-color: #A3A3A3;\">\u53c2\u8003\u8d44\u6599\uff1a<\/span><\/span><\/p><p style=\"line-height: 1;\"><span style=\"letter-spacing: 1px;\"><span style=\"color: rgb(178, 178, 178); --tt-darkmode-color: #A3A3A3;\">1.https:\/\/arxiv.org\/abs\/2504.18530<\/span><\/span><\/p><p style=\"line-height: 1;\"><span style=\"letter-spacing: 1px;\"><span style=\"color: rgb(178, 178, 178); --tt-darkmode-color: #A3A3A3;\">2.https:\/\/www.lesswrong.com\/posts\/x59FhzuM9yuvZHAHW\/untitled-draft-yhra<\/span><\/span><\/p><p style=\"line-height: 1;\"><\/p><p style=\"line-height: 1;\"><span style=\"letter-spacing: 1px;\"><span style=\"color: rgb(178, 178, 178); --tt-darkmode-color: #A3A3A3;\">\u8fd0\u8425\/\u6392\u7248\uff1a\u4f55\u6668\u9f99<\/span><\/span><\/p><\/div><\/div><\/div><\/div><\/div><\/div>","content_en":null,"content_all":null,"article_url":"http:\/\/www.sciphi.cn\/article\/view\/14756","typeName":"AI"},"code":10000,"message":"\u6210\u529f"}