{"id":36920,"date":"2025-01-21T07:01:14","date_gmt":"2025-01-20T23:01:14","guid":{"rendered":"https:\/\/17aitech.com\/?p=36920"},"modified":"2025-01-21T07:01:14","modified_gmt":"2025-01-20T23:01:14","slug":"%e7%a6%bb%e8%81%8copenai%e5%90%8e%ef%bc%8c%e7%bf%81%e8%8d%94%e5%8d%9a%e5%ae%a2%e9%a6%96%e6%ac%a1%e4%b8%8a%e6%96%b0%ef%bc%8c%e5%bc%95%e4%bc%97%e7%bd%91%e5%8f%8b%e5%9b%b4%e8%a7%82%e5%ad%a6%e4%b9%a0","status":"publish","type":"post","link":"https:\/\/17aitech.com\/?p=36920","title":{"rendered":"\u79bb\u804cOpenAI\u540e\uff0c\u7fc1\u8354\u535a\u5ba2\u9996\u6b21\u4e0a\u65b0\uff0c\u5f15\u4f17\u7f51\u53cb\u56f4\u89c2\u5b66\u4e60\uff08\u4e2d\u6587\u5168\u6587\uff09"},"content":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-12-02-8\" target=\"_blank\">\u79bb\u804cOpenAI\u540e\uff0c\u7fc1\u8354\u535a\u5ba2\u9996\u6b21\u4e0a\u65b0\uff0c\u5f15\u4f17\u7f51\u53cb\u56f4\u89c2\u5b66\u4e60\uff08\u4e2d\u6587\u5168\u6587\uff09<\/a><\/p>\n<section><\/section>\n<section>\u5927\u7ea6\u4e00\u4e2a\u6708\u524d\uff0cOpenAI \u5b89\u5168\u7cfb\u7edf\u56e2\u961f\u8d1f\u8d23\u4eba\u7fc1\u8354\uff08Lilian Weng\uff09\u5728 X \u4e0a\u5ba3\u5e03\u4e86\u4ece\u5df2\u7ecf\u5de5\u4f5c\u4e86\u8fd1 7 \u5e74\u7684 OpenAI \u79bb\u804c\u7684\u6d88\u606f\u3002<\/section>\n<section><\/section>\n<section>\u5f53\u65f6\uff0c\u5979\u5c31\u66fe\u8868\u793a\uff0c\u4e4b\u540e\u53ef\u80fd\u6709\u66f4\u591a\u65f6\u95f4\u6765\u5199\u535a\u5ba2\u3002<\/section>\n<section><\/section>\n<section>\u521a\u521a\uff0c\u7fc1\u8354\u66f4\u65b0\u4e86\u4e00\u7bc7\u535a\u5ba2\uff0c\u8fc5\u901f\u5f15\u8d77\u4e86\u5927\u5bb6\u7684\u56f4\u89c2\u5b66\u4e60\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u7bc7\u535a\u5ba2\u7684\u4e3b\u9898\u662f\u5173\u4e8e\u5f3a\u5316\u5b66\u4e60\u4e2d reward hacking \u7684\u3002\u7fc1\u8354\u8ba4\u4e3a\uff0c\u300c\u5f53\u5f3a\u5316\u5b66\u4e60\u667a\u80fd\u4f53\u5229\u7528\u5956\u52b1\u51fd\u6570\u6216\u73af\u5883\u4e2d\u7684\u7f3a\u9677\u6765\u6700\u5927\u5316\u5956\u52b1\u800c\u4e0d\u5b66\u4e60\u9884\u671f\u884c\u4e3a\u65f6\uff0c\u5c31\u4f1a\u53d1\u751f reward hacking \u653b\u51fb\u3002\u5728\u6211\u770b\u6765\uff0c\u8fd9\u662f\u5728\u73b0\u5b9e\u4e16\u754c\u4e2d\u90e8\u7f72\u66f4\u591a\u81ea\u4e3b AI \u6a21\u578b\u7528\u4f8b\u65f6\u7684\u4e3b\u8981\u969c\u788d\u3002\u300d<\/section>\n<section><\/section>\n<section>\u5979\u8fd8\u547c\u5401\u5bf9 reward hacking\uff0c\u7279\u522b\u662f\u5bf9 LLM \u548c RLHF \u4e2d reward hacking \u7684\u7f13\u89e3\u7b56\u7565\u8fdb\u884c\u66f4\u591a\u7684\u7814\u7a76\u3002\u00a0<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-00fd0c6412cad47d41511dcabb84a0d1.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-00fd0c6412cad47d41511dcabb84a0d1.png\"><\/a><\/section>\n<section>\u9700\u8981\u63d0\u793a\u7684\u662f\uff0c\u8fd9\u662f\u4e00\u7bc7\u5f88\u957f\u5f88\u5e72\u8d27\u7684\u6587\u7ae0\uff0c\u7fc1\u8354\u5728\u535a\u5ba2\u4e2d\u7ed9\u51fa\u7684\u9605\u8bfb\u9884\u4f30\u65f6\u95f4\u662f 37 \u5206\u949f\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u65b9\u4fbf\u56fd\u5185\u8bfb\u8005\u66f4\u597d\u5730\u5b66\u4e60\u8fd9\u7bc7\u5185\u5bb9\uff0c\u673a\u5668\u4e4b\u5fc3\u5bf9\u6b64\u6587\u7ae0\u8fdb\u884c\u4e86\u7f16\u8bd1\uff0c\u611f\u5174\u8da3\u7684\u8bfb\u8005\u4e5f\u53ef\u67e5\u9605\u539f\u82f1\u6587\u5185\u5bb9\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-95b24a91248970e4066e2c71c4809ffb.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-95b24a91248970e4066e2c71c4809ffb.png\"><\/a><\/section>\n<ul>\n<li>\n<section>\u6587\u7ae0\u6807\u9898\uff1aReward Hacking in Reinforcement Learning<\/section>\n<\/li>\n<li>\n<section>\u6587\u7ae0\u94fe\u63a5\uff1ahttps:\/\/lilianweng.github.io\/posts\/2024-11-28-reward-hacking\/<\/section>\n<\/li>\n<li>\n<section>\u7fc1\u8354\u535a\u5ba2\uff1ahttps:\/\/lilianweng.github.io\/<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-1b37cda4690fb09addb036f3f9582dd0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-1b37cda4690fb09addb036f3f9582dd0.png\"><\/a><\/section>\n<section>\u5f53\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u667a\u80fd\u4f53\u5229\u7528\u5956\u52b1\u51fd\u6570\u4e2d\u7684\u7f3a\u9677\u6216\u6b67\u4e49\u6765\u83b7\u5f97\u9ad8\u989d\u5956\u52b1\uff0c\u800c\u6ca1\u6709\u771f\u6b63\u5b66\u4e60\u6216\u5b8c\u6210\u9884\u671f\u4efb\u52a1\u65f6\uff0c\u5c31\u4f1a\u53d1\u751f Reward Hacking\uff08Reward Hacking in Reinforcement Learning\uff09\u3002Hacking \u4e4b\u6240\u4ee5\u5b58\u5728\uff0c\u662f\u56e0\u4e3a\u5f3a\u5316\u5b66\u4e60\uff08RL\uff09\u73af\u5883\u901a\u5e38\u4e0d\u5b8c\u5584\uff0c\u800c\u4e14\u51c6\u786e\u6307\u5b9a\u5956\u52b1\u51fd\u6570\u4ece\u6839\u672c\u4e0a\u5177\u6709\u6311\u6218\u6027\u3002<\/section>\n<section><\/section>\n<section>\u968f\u7740\u5927\u8bed\u8a00\u6a21\u578b\u7684\u5174\u8d77\uff0cRLHF \u6210\u4e3a\u4e8b\u5b9e\u4e0a\u7684\u5bf9\u9f50\u8bad\u7ec3\u65b9\u6cd5\uff0c\u8bed\u8a00\u6a21\u578b\u7684 RL \u8bad\u7ec3\u4e2d\u7684 reward hacking \u653b\u51fb\u5df2\u6210\u4e3a\u4e00\u9879\u5173\u952e\u7684\u73b0\u5b9e\u6311\u6218\u3002\u6a21\u578b\u5b66\u4e60\u4fee\u6539\u5355\u5143\u6d4b\u8bd5\u4ee5\u901a\u8fc7\u7f16\u7801\u4efb\u52a1\u7684\u60c5\u51b5\uff0c\u6216\u8005\u54cd\u5e94\u5305\u542b\u6a21\u4eff\u7528\u6237\u504f\u597d\u7684 bias \u7684\u60c5\u51b5\uff0c\u90fd\u975e\u5e38\u4ee4\u4eba\u62c5\u5fe7\uff0c\u5e76\u4e14\u53ef\u80fd\u662f\u73b0\u5b9e\u4e16\u754c\u90e8\u7f72\u66f4\u81ea\u4e3b\u7684 AI \u6a21\u578b\u7528\u4f8b\u7684\u4e3b\u8981\u969c\u788d\u4e4b\u4e00\u3002<\/section>\n<section><\/section>\n<section>\u8fc7\u53bb\u5173\u4e8e\u8fd9\u4e2a\u4e3b\u9898\u7684\u5927\u90e8\u5206\u7814\u7a76\u90fd\u662f\u7406\u8bba\u6027\u7684\uff0c\u91cd\u70b9\u662f\u5b9a\u4e49\u6216\u8bc1\u660e Reward Hacking \u7684\u5b58\u5728\u3002\u7136\u800c\uff0c\u5bf9\u5b9e\u9645\u7f13\u89e3\u63aa\u65bd\u7684\u7814\u7a76\u4ecd\u7136\u6709\u9650\uff0c\u7279\u522b\u662f\u5728 RLHF \u548c LLM \u7684\u80cc\u666f\u4e0b\u3002\u6211\u7279\u522b\u60f3\u547c\u5401\u672a\u6765\u8fdb\u884c\u66f4\u591a\u7684\u7814\u7a76\uff0c\u4ee5\u4e86\u89e3\u548c\u5f00\u53d1\u7f13\u89e3 Reward Hacking \u7684\u63aa\u65bd\u3002\u5e0c\u671b\u6211\u5f88\u5feb\u5c31\u80fd\u5728\u4e00\u7bc7\u4e13\u95e8\u7684\u6587\u7ae0\u4e2d\u4ecb\u7ecd\u7f13\u89e3\u90e8\u5206\u3002<\/section>\n<section><\/section>\n<section><strong>\u80cc\u666f<\/strong><\/section>\n<section><\/section>\n<section><strong>\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u5956\u52b1\u51fd\u6570<\/strong><\/section>\n<section><\/section>\n<section>\u5956\u52b1\u51fd\u6570\u5b9a\u4e49\u4e86\u4efb\u52a1\uff0c\u5956\u52b1\u5851\u9020\u663e\u8457\u5f71\u54cd\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u5b66\u4e60\u6548\u7387\u548c\u51c6\u786e\u6027\u3002\u4e3a\u5f3a\u5316\u5b66\u4e60\u4efb\u52a1\u8bbe\u8ba1\u5956\u52b1\u51fd\u6570\u901a\u5e38\u611f\u89c9\u50cf\u4e00\u95e8\u300c\u9ed1\u9b54\u6cd5\u300d\u3002\u8bb8\u591a\u56e0\u7d20\u5bfc\u81f4\u4e86\u8fd9\u79cd\u590d\u6742\u6027\uff1a\u5982\u4f55\u5c06\u5927\u76ee\u6807\u5206\u89e3\u4e3a\u5c0f\u76ee\u6807\uff1f\u5956\u52b1\u662f\u7a00\u758f\u7684\u8fd8\u662f\u5bc6\u96c6\u7684\uff1f\u5982\u4f55\u8861\u91cf\u6210\u529f\uff1f\u5404\u79cd\u9009\u62e9\u53ef\u80fd\u4f1a\u5bfc\u81f4\u826f\u597d\u6216\u6709\u95ee\u9898\u7684\u5b66\u4e60\u52a8\u6001\uff0c\u5305\u62ec\u65e0\u6cd5\u5b66\u4e60\u7684\u4efb\u52a1\u6216\u53ef\u7834\u89e3\u7684\u5956\u52b1\u51fd\u6570\u3002\u5173\u4e8e\u5982\u4f55\u5728\u5f3a\u5316\u5b66\u4e60\u4e2d\u8fdb\u884c\u5956\u52b1\u5851\u9020\u7684\u7814\u7a76\u5386\u53f2\u60a0\u4e45\u3002<\/section>\n<section><\/section>\n<section>\u4f8b\u5982\uff0c\u5728\u5434\u6069\u8fbe\u7b49\u4eba\u4e8e 1999 \u5e74\u53d1\u8868\u7684\u8bba\u6587\u300aPolicy invariance under reward trasnsforamtions: Theory and application to reward shaping\u300b\u4e2d\uff0c\u4f5c\u8005\u7814\u7a76\u4e86\u5982\u4f55\u4fee\u6539\u9a6c\u5c14\u53ef\u592b\u51b3\u7b56\u8fc7\u7a0b\uff08MDP\uff09\u4e2d\u7684\u5956\u52b1\u51fd\u6570\uff0c\u4ee5\u4f7f\u6700\u4f18\u7b56\u7565\u4fdd\u6301\u4e0d\u53d8\u3002\u4ed6\u4eec\u53d1\u73b0\u7ebf\u6027\u53d8\u6362\u662f\u6709\u6548\u7684\u3002<\/section>\n<section><\/section>\n<section>\u7ed9\u5230 MDP <a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-ae4ebcc8efa8851bc736360158d63237.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-ae4ebcc8efa8851bc736360158d63237.png\"><\/a>\u60f3\u8981\u83b7\u5f97\u4e00\u4e2a\u53d8\u6362\u540e\u7684 MDP\uff0c<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-c53f9eb2bdb9b55ae738cb969e5e759a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-c53f9eb2bdb9b55ae738cb969e5e759a.png\"><\/a>\u5176\u4e2d <a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-bbde8e07765f1613466ca52ef83f88a2.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-bbde8e07765f1613466ca52ef83f88a2.png\"><\/a>\u8fd9\u6837\u6211\u4eec\u5c31\u53ef\u4ee5\u5f15\u5bfc\u5b66\u4e60\u7b97\u6cd5\u66f4\u52a0\u9ad8\u6548\u3002\u7ed9\u5b9a\u4e00\u4e2a\u5b9e\u503c\u51fd\u6570<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-75aa598c5acfb0d50505aa71e08093af.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-75aa598c5acfb0d50505aa71e08093af.png\"><\/a>F \u662f\u57fa\u4e8e\u6f5c\u529b\u7684\u5851\u9020\u51fd\u6570\uff0c\u5982\u679c\u5bf9\u4e8e\u6240\u6709<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-5b0f303e116aeb0425d496402a77a4b9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-5b0f303e116aeb0425d496402a77a4b9.png\"><\/a><\/section>\n<section><\/section>\n<section>\u6709\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-005365fa2d389c90b24d3cab13f9e67a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-005365fa2d389c90b24d3cab13f9e67a.png\"><\/a><\/section>\n<section>\u8fd9\u5c06\u786e\u4fdd\u6298\u6263\u603b\u989d<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-070d83a9983511bad2130726755242df.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-070d83a9983511bad2130726755242df.png\"><\/a>\u6700\u7ec8\u7ed3\u679c\u4e3a 0\u3002\u5982\u679c F \u662f\u8fd9\u6837\u4e00\u4e2a\u57fa\u4e8e\u52bf\u7684\u5851\u9020\u51fd\u6570\uff0c\u5b83\u65e2\u5145\u5206\u53c8\u5fc5\u8981\uff0c\u4ee5\u786e\u4fdd M \u548c M\u2019 \u5171\u4eab\u76f8\u540c\u7684\u6700\u4f18\u7b56\u7565\u3002<\/section>\n<section><\/section>\n<section>\u5f53<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-ea93aa8f910cc6ca5e6a3116f56f8519.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-ea93aa8f910cc6ca5e6a3116f56f8519.png\"><\/a><\/section>\n<section><\/section>\n<section>\u5982\u679c\u6211\u4eec\u8fdb\u4e00\u6b65\u5047\u8bbe<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-851984d307128a3497d0ed209fecdf5c.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-851984d307128a3497d0ed209fecdf5c.png\"><\/a> \u5176\u4e2d S_0 \u5904\u4e8e\u5438\u6536\u72b6\u6001\uff0c\u5e76\u4e14\u00a0<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-7f7df580e080620e9cb0ec990c3b0d5e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-7f7df580e080620e9cb0ec990c3b0d5e.png\"><\/a><\/section>\n<section><\/section>\n<section>\u7136\u540e\u5bf9\u6240\u6709 <a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d8d2d0f39111ed0ab5ee881dab6d4c10.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d8d2d0f39111ed0ab5ee881dab6d4c10.png\"><\/a><\/section>\n<section><\/section>\n<section>\u6709\uff1a<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f9b81281522720293b84d587668f98df.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f9b81281522720293b84d587668f98df.png\"><\/a><\/section>\n<section>\u8fd9\u79cd\u5956\u52b1\u5851\u9020\u5f62\u5f0f\u4f7f\u6211\u4eec\u80fd\u591f\u5c06\u542f\u53d1\u5f0f\u65b9\u6cd5\u7eb3\u5165\u5956\u52b1\u51fd\u6570\u4e2d\uff0c\u4ee5\u52a0\u5feb\u5b66\u4e60\u901f\u5ea6\uff0c\u800c\u4e0d\u4f1a\u5f71\u54cd\u6700\u4f73\u7b56\u7565\u3002<\/section>\n<section><\/section>\n<section><strong>\u865a\u5047\u76f8\u5173\u6027<\/strong><\/section>\n<section><\/section>\n<section>\u5206\u7c7b\u4efb\u52a1\u4e2d\u7684\u865a\u5047\u76f8\u5173\u6216\u6377\u5f84\u5b66\u4e60\uff08Geirhos et al. 2020\uff09\u662f\u4e00\u4e2a\u4e0e Reward Hacking \u5bc6\u5207\u76f8\u5173\u7684\u6982\u5ff5\u3002\u865a\u5047\u6216\u6377\u5f84\u7279\u5f81\u53ef\u80fd\u4f1a\u5bfc\u81f4\u5206\u7c7b\u5668\u65e0\u6cd5\u6309\u9884\u671f\u8fdb\u884c\u5b66\u4e60\u548c\u6cdb\u5316\u3002\u4f8b\u5982\uff0c\u5982\u679c\u6240\u6709\u72fc\u7684\u8bad\u7ec3\u56fe\u50cf\u90fd\u5305\u542b\u96ea\uff0c\u5219\u7528\u4e8e\u533a\u5206\u72fc\u548c\u54c8\u58eb\u5947\u7684\u4e8c\u5143\u5206\u7c7b\u5668\u53ef\u80fd\u4f1a\u56e0\u5b58\u5728\u96ea\u666f\u800c\u8fc7\u62df\u5408\uff08Ribeiro et al. 2024\uff09\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-db370f9512816216fa53f80c6fa17470.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-db370f9512816216fa53f80c6fa17470.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u5982\u679c\u6a21\u578b\u4e0e\u6377\u5f84\u7279\u5f81\u8fc7\u62df\u5408\uff0c\u5219\u5b83\u5728\u5206\u5e03\u5916 (OOD) \u6d4b\u8bd5\u96c6\u4e0a\u7684\u8868\u73b0\u4f1a\u5f88\u5dee\u3002\uff08\u56fe\u6e90\uff1aGeirhos et al. 2020\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>ERM \u539f\u7406\u6307\u51fa\uff0c\u7531\u4e8e\u6574\u4e2a\u6570\u636e\u5206\u5e03\u672a\u77e5\uff0c\u6700\u5c0f\u5316\u8bad\u7ec3\u6570\u636e\u7684\u635f\u5931\u662f\u98ce\u9669\u7684\u5408\u7406\u66ff\u4ee3\uff0c\u56e0\u6b64\u6211\u4eec\u503e\u5411\u4e8e\u8bad\u7ec3\u635f\u5931\u6700\u4f4e\u7684\u6a21\u578b\u3002Nagarajan et al. (2021) \u7814\u7a76\u4e86 ERM \u539f\u7406\uff0c\u5e76\u6307\u51fa ERM \u9700\u8981\u4f9d\u8d56\u6240\u6709\u7c7b\u578b\u7684\u4fe1\u606f\u7279\u5f81\uff0c\u5305\u62ec\u4e0d\u53ef\u9760\u7684\u865a\u5047\u7279\u5f81\uff0c\u540c\u65f6\u5c1d\u8bd5\u65e0\u9650\u5236\u5730\u62df\u5408\u6570\u636e\u3002\u4ed6\u4eec\u7684\u5b9e\u9a8c\u8868\u660e\uff0c\u65e0\u8bba\u4efb\u52a1\u591a\u4e48\u7b80\u5355\uff0cERM \u90fd\u4f1a\u4f9d\u8d56\u4e8e\u865a\u5047\u7279\u5f81\u3002<\/section>\n<section><\/section>\n<section><strong>\u5982\u4f55\u5b9a\u4e49 Reward Hacking<\/strong><\/section>\n<section><\/section>\n<section>\u5f3a\u5316\u5b66\u4e60\u4e2d\u7684\u5956\u52b1\u5851\u9020\u5177\u6709\u6311\u6218\u6027\u3002\u5f53\u5f3a\u5316\u5b66\u4e60\u667a\u80fd\u4f53\u5229\u7528\u5956\u52b1\u51fd\u6570\u4e2d\u7684\u7f3a\u9677\u6216\u6a21\u7cca\u6027\u6765\u83b7\u5f97\u9ad8\u989d\u5956\u52b1\uff0c\u800c\u6ca1\u6709\u771f\u6b63\u5b66\u4e60\u9884\u671f\u884c\u4e3a\u6216\u6309\u8bbe\u8ba1\u5b8c\u6210\u4efb\u52a1\u65f6\uff0c\u5c31\u4f1a\u53d1\u751f Reward Hacking \u653b\u51fb\u3002\u8fd1\u5e74\u6765\uff0c\u4eba\u4eec\u5df2\u7ecf\u63d0\u51fa\u4e86\u51e0\u4e2a\u76f8\u5173\u6982\u5ff5\uff0c\u5747\u6307\u67d0\u79cd\u5f62\u5f0f\u7684 reward hacking\uff1a<\/section>\n<ul>\n<li>\n<p>Reward hacking (Amodei et al., 2016)<\/p>\n<\/li>\n<li>\n<p>Reward corruption (Everitt et al., 2017)<\/p>\n<\/li>\n<li>\n<p>Reward tampering (Everitt et al. 2019)<\/p>\n<\/li>\n<li>\n<p>Specification gaming (Krakovna et al., 2020)<\/p>\n<\/li>\n<li>\n<p>Objective robustness (Koch et al. 2021)<\/p>\n<\/li>\n<li>\n<p>Goal misgeneralization (Langosco et al. 2022)<\/p>\n<\/li>\n<li>\n<p>Reward misspecifications (Pan et al. 2022)<\/p>\n<\/li>\n<\/ul>\n<section>\u8be5\u6982\u5ff5\u8d77\u6e90\u4e8e Amodei et al. (2016) \u7684\u7814\u7a76\uff0c\u4ed6\u4eec\u5728\u5176\u5f00\u521b\u6027\u7684\u8bba\u6587\u300aConcrete Problems in AI Safety\u300b\u4e2d\u63d0\u51fa\u4e86\u4e00\u7cfb\u5217\u5173\u4e8e\u4eba\u5de5\u667a\u80fd\u5b89\u5168\u7684\u5f00\u653e\u6027\u7814\u7a76\u95ee\u9898\u3002\u4ed6\u4eec\u5c06 Reward Hacking \u5217\u4e3a\u5173\u952e\u7684\u4eba\u5de5\u667a\u80fd\u5b89\u5168\u95ee\u9898\u4e4b\u4e00\u3002Reward Hacking \u662f\u6307\u667a\u80fd\u4f53\u901a\u8fc7\u4e0d\u826f\u884c\u4e3a\u6765\u6b3a\u9a97\u5956\u52b1\u51fd\u6570\u4ee5\u83b7\u5f97\u9ad8\u989d\u5956\u52b1\u7684\u53ef\u80fd\u6027\u3002\u89c4\u8303\u535a\u5f08\uff08Specification gaming\uff0cKrakovna et al. 2020\uff09\u662f\u4e00\u4e2a\u7c7b\u4f3c\u7684\u6982\u5ff5\uff0c\u5b9a\u4e49\u4e3a\u6ee1\u8db3\u76ee\u6807\u7684\u5b57\u9762\u89c4\u8303\u4f46\u672a\u5b9e\u73b0\u9884\u671f\u7ed3\u679c\u7684\u884c\u4e3a\u3002\u8fd9\u91cc\u4efb\u52a1\u76ee\u6807\u548c\u9884\u671f\u76ee\u6807\u7684\u5b57\u9762\u63cf\u8ff0\u53ef\u80fd\u5b58\u5728\u5dee\u8ddd\u3002<\/section>\n<section><\/section>\n<section>\u5956\u52b1\u5851\u9020\uff08reward shaping\uff09\u662f\u4e00\u79cd\u7528\u4e8e\u4e30\u5bcc\u5956\u52b1\u51fd\u6570\u7684\u6280\u672f\uff0c\u4f7f\u667a\u80fd\u4f53\u66f4\u5bb9\u6613\u5b66\u4e60 \u2014\u2014 \u4f8b\u5982\uff0c\u901a\u8fc7\u63d0\u4f9b\u66f4\u5bc6\u96c6\u7684\u5956\u52b1\u3002\u7136\u800c\uff0c\u8bbe\u8ba1\u4e0d\u5f53\u7684\u5956\u52b1\u5851\u9020\u673a\u5236\u53ef\u80fd\u4f1a\u6539\u53d8\u6700\u4f18\u7b56\u7565\u7684\u8f68\u8ff9\u3002\u8bbe\u8ba1\u6709\u6548\u7684\u5956\u52b1\u5851\u9020\u673a\u5236\u672c\u8d28\u4e0a\u662f\u56f0\u96be\u7684\u3002\u4e0e\u5176\u8d23\u602a\u5956\u52b1\u51fd\u6570\u8bbe\u8ba1\u4e0d\u826f\uff0c\u66f4\u51c6\u786e\u5730\u8bf4\uff0c\u5e94\u8be5\u627f\u8ba4\uff0c\u7531\u4e8e\u4efb\u52a1\u672c\u8eab\u7684\u590d\u6742\u6027\u3001\u90e8\u5206\u53ef\u89c2\u5bdf\u72b6\u6001\u3001\u8003\u8651\u7684\u591a\u4e2a\u7ef4\u5ea6\u4ee5\u53ca\u5176\u4ed6\u56e0\u7d20\uff0c\u8bbe\u8ba1\u4e00\u4e2a\u597d\u7684\u5956\u52b1\u51fd\u6570\u672c\u8d28\u4e0a\u662f\u5177\u6709\u6311\u6218\u6027\u7684\u3002<\/section>\n<section><\/section>\n<section>\u5728\u5206\u5e03\u5916 (OOD) \u73af\u5883\u4e2d\u6d4b\u8bd5\u5f3a\u5316\u5b66\u4e60\u667a\u80fd\u4f53\u65f6\uff0c\u53ef\u80fd\u4f1a\u7531\u4e8e\u4ee5\u4e0b\u539f\u56e0\u5bfc\u81f4\u9c81\u68d2\u6027\u5931\u6548\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u5373\u4f7f\u76ee\u6807\u6b63\u786e\uff0c\u6a21\u578b\u4e5f\u65e0\u6cd5\u6709\u6548\u6cdb\u5316\u3002\u5f53\u7b97\u6cd5\u7f3a\u4e4f\u8db3\u591f\u7684\u667a\u80fd\u6216\u80fd\u529b\u65f6\uff0c\u5c31\u4f1a\u53d1\u751f\u8fd9\u79cd\u60c5\u51b5\u3002<\/section>\n<\/li>\n<li>\n<section>\u8be5\u6a21\u578b\u5177\u6709\u5f88\u597d\u7684\u6cdb\u5316\u80fd\u529b\uff0c\u4f46\u8ffd\u6c42\u7684\u76ee\u6807\u4e0e\u8bad\u7ec3\u65f6\u7684\u76ee\u6807\u4e0d\u540c\u3002\u5f53\u667a\u80fd\u4f53\u5956\u52b1\u4e0e\u771f\u5b9e\u5956\u52b1\u51fd\u6570\u4e0d\u540c\u65f6\uff0c\u5c31\u4f1a\u53d1\u751f\u8fd9\u79cd\u60c5\u51b5\u3002\u8fd9\u88ab\u79f0\u4e3a\u76ee\u6807\u9c81\u68d2\u6027\uff08Koch et al. 2021\uff09\u6216\u76ee\u6807\u9519\u8bef\u6cdb\u5316\uff08Koch et al. 2021\uff09\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u5728\u4e24\u4e2a\u5f3a\u5316\u5b66\u4e60\u73af\u5883 CoinRun \u548c Maze \u4e2d\u8fdb\u884c\u7684\u5b9e\u9a8c\u8bc1\u660e\u4e86\u8bad\u7ec3\u671f\u95f4\u968f\u673a\u5316\u7684\u91cd\u8981\u6027\u3002\u5982\u679c\u5728\u8bad\u7ec3\u671f\u95f4\uff0c\u786c\u5e01\u6216\u5976\u916a\u88ab\u653e\u7f6e\u5728\u56fa\u5b9a\u4f4d\u7f6e\uff08\u5373\u5173\u5361\u7684\u53f3\u7aef\u6216\u8ff7\u5bab\u7684\u53f3\u4e0a\u89d2\uff09\uff0c\u4f46\u5728\u786c\u5e01\u6216\u5976\u916a\u968f\u673a\u653e\u7f6e\u7684\u73af\u5883\u4e2d\u6d4b\u8bd5\uff0c\u5219\u667a\u80fd\u4f53\u4f1a\u5728\u6d4b\u8bd5\u65f6\u76f4\u63a5\u8dd1\u5230\u56fa\u5b9a\u4f4d\u7f6e\u800c\u6ca1\u83b7\u5f97\u786c\u5e01\u6216\u5976\u916a\u3002<\/section>\n<section><\/section>\n<section>\u5f53\u89c6\u89c9\u7279\u5f81\uff08\u4f8b\u5982\u5976\u916a\u6216\u786c\u5e01\uff09\u548c\u4f4d\u7f6e\u7279\u5f81\uff08\u4f8b\u5982\u53f3\u4e0a\u89d2\u6216\u53f3\u7aef\uff09\u5728\u6d4b\u8bd5\u671f\u95f4\u4e0d\u4e00\u81f4\u65f6\uff0c\u5c31\u4f1a\u53d1\u751f\u51b2\u7a81\uff0c\u5bfc\u81f4\u8bad\u7ec3\u540e\u7684\u6a21\u578b\u66f4\u559c\u6b22\u4f4d\u7f6e\u7279\u5f81\u3002\u6211\u60f3\u6307\u51fa\u7684\u662f\uff0c\u5728\u8fd9\u4e24\u4e2a\u4f8b\u5b50\u4e2d\uff0c\u5956\u52b1\u7ed3\u679c\u5dee\u8ddd\u5f88\u660e\u663e\uff0c\u4f46\u5728\u5927\u591a\u6570\u73b0\u5b9e\u4e16\u754c\u60c5\u51b5\u4e0b\uff0c\u8fd9\u79cd\u7c7b\u578b\u7684\u504f\u5dee\u4e0d\u592a\u53ef\u80fd\u5982\u6b64\u660e\u663e\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-96a8c7bcfe1a895444566e16ecd145ae.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-96a8c7bcfe1a895444566e16ecd145ae.png\"><\/a><\/section>\n<section><em><sup>\u56fe 2. \u8bad\u7ec3\u671f\u95f4\u968f\u673a\u5316\u786c\u5e01\u4f4d\u7f6e\u7684\u5f71\u54cd\u3002\u5f53\u8bad\u7ec3\u671f\u95f4\u786c\u5e01\u968f\u673a\u653e\u7f6e {0, 2, 3, 6, 11}% \u7684\u65f6\u95f4\uff08x \u8f74\uff09\u65f6\uff0c\u667a\u80fd\u4f53\u5bfc\u822a\u5230\u5173\u5361\u672b\u5c3e\u800c\u672a\u83b7\u5f97\u786c\u5e01\u7684\u9891\u7387\u4f1a\u968f\u7740\u968f\u673a\u5316\u7684\u589e\u52a0\u800c\u964d\u4f4e\uff08\u300cy \u8f74\u300d\uff09\u3002\uff08\u56fe\u6e90\uff1a<em>\u00a0Koch et al. 2021<\/em>\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5956\u52b1\u7be1\u6539\uff08Reward Tampering\uff09\uff08Everitt et al. 2019\uff09\u662f\u4e00\u79cd Reward Hacking \u884c\u4e3a\uff0c\u5176\u4e2d\u667a\u80fd\u4f53\u5e72\u6270\u5956\u52b1\u51fd\u6570\u672c\u8eab\uff0c\u5bfc\u81f4\u89c2\u5bdf\u5230\u7684\u5956\u52b1\u4e0d\u518d\u51c6\u786e\u4ee3\u8868\u9884\u671f\u76ee\u6807\u3002\u5728\u5956\u52b1\u7be1\u6539\u4e2d\uff0c\u6a21\u578b\u901a\u8fc7\u76f4\u63a5\u64cd\u7eb5\u5956\u52b1\u51fd\u6570\u7684\u5b9e\u73b0\u6216\u95f4\u63a5\u6539\u53d8\u7528\u4f5c\u5956\u52b1\u51fd\u6570\u8f93\u5165\u7684\u73af\u5883\u4fe1\u606f\u6765\u4fee\u6539\u5176\u5956\u52b1\u673a\u5236\u3002<\/section>\n<section><\/section>\n<section>\uff08\u6ce8\u610f\uff1a\u4e00\u4e9b\u5de5\u4f5c\u5c06\u5956\u52b1\u7be1\u6539\u5b9a\u4e49\u4e3a\u4e0e Reward Hacking \u4e0d\u540c\u7684\u9519\u4f4d\u884c\u4e3a\u7c7b\u522b\u3002\u4f46\u6211\u8ba4\u4e3a Reward Hacking \u5728\u8fd9\u91cc\u662f\u4e00\u4e2a\u66f4\u5e7f\u6cdb\u7684\u6982\u5ff5\u3002\uff09<\/section>\n<section><\/section>\n<section>\u4ece\u9ad8\u5c42\u6b21\u4e0a\u8bb2\uff0cReward Hacking \u53ef\u4ee5\u5206\u4e3a\u4e24\u7c7b\uff1a\u73af\u5883\u6216\u76ee\u6807\u9519\u8bef\u6307\u5b9a\uff0c\u4ee5\u53ca\u5956\u52b1\u7be1\u6539\u3002<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u73af\u5883\u6216\u76ee\u6807\u6307\u5b9a\u9519\u8bef\uff1a\u6a21\u578b\u901a\u8fc7\u5165\u4fb5\u73af\u5883\u6216\u4f18\u5316\u4e0e\u771f\u5b9e\u5956\u52b1\u76ee\u6807\u4e0d\u4e00\u81f4\u7684\u5956\u52b1\u51fd\u6570\u6765\u5b66\u4e60\u4e0d\u826f\u884c\u4e3a\uff0c\u4ee5\u83b7\u5f97\u9ad8\u989d\u5956\u52b1 \u2014\u2014 \u4f8b\u5982\u5f53\u5956\u52b1\u6307\u5b9a\u9519\u8bef\u6216\u7f3a\u4e4f\u5173\u952e\u8981\u6c42\u65f6\u3002<\/section>\n<\/li>\n<li>\n<section>\u5956\u52b1\u7be1\u6539\uff1a\u6a21\u578b\u5b66\u4e60\u5e72\u6270\u5956\u52b1\u673a\u5236\u672c\u8eab\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u6848\u4f8b\u5217\u8868<\/strong><\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u8bad\u7ec3\u6293\u53d6\u7269\u4f53\u7684\u673a\u68b0\u624b\u53ef\u4ee5\u5b66\u4f1a\u5982\u4f55\u901a\u8fc7\u5c06\u624b\u653e\u5728\u7269\u4f53\u548c\u76f8\u673a\u4e4b\u95f4\u6765\u6b3a\u9a97\u4eba\uff1ahttps:\/\/openai.com\/index\/learning-from-human-preferences\/<\/section>\n<\/li>\n<li>\n<section>\u8bad\u7ec3\u6700\u5927\u5316\u8df3\u8dc3\u9ad8\u5ea6\u7684\u667a\u80fd\u4f53\u53ef\u80fd\u4f1a\u5229\u7528\u7269\u7406\u6a21\u62df\u5668\u4e2d\u7684\u9519\u8bef\u6765\u5b9e\u73b0\u4e0d\u5207\u5b9e\u9645\u7684\u9ad8\u5ea6\uff1ahttps:\/\/arxiv.org\/abs\/1803.03453<\/section>\n<\/li>\n<li>\n<section>\u667a\u80fd\u4f53\u88ab\u8bad\u7ec3\u9a91\u81ea\u884c\u8f66\u5230\u8fbe\u76ee\u6807\uff0c\u5e76\u5728\u63a5\u8fd1\u76ee\u6807\u65f6\u83b7\u5f97\u5956\u52b1\u3002\u7136\u540e\uff0c\u667a\u80fd\u4f53\u53ef\u80fd\u4f1a\u5b66\u4e60\u5728\u76ee\u6807\u5468\u56f4\u7ed5\u5c0f\u5708\u9a91\u884c\uff0c\u56e0\u4e3a\u8fdc\u79bb\u76ee\u6807\u65f6\u4e0d\u4f1a\u53d7\u5230\u60e9\u7f5a\uff1ahttps:\/\/people.eecs.berkeley.edu\/~pabbeel\/cs287-fa09\/readings\/NgHaradaRussell-shaping-ICML1999.pdf<\/section>\n<\/li>\n<li>\n<section>\u5728\u8db3\u7403\u6bd4\u8d5b\u4e2d\uff0c\u5f53\u667a\u80fd\u4f53\u89e6\u7403\u65f6\u4f1a\u5206\u914d\u5956\u52b1\uff0c\u4e8e\u662f\u5b83\u4f1a\u5b66\u4e60\u4fdd\u6301\u5728\u7403\u65c1\u8fb9\u4ee5\u9ad8\u9891\u89e6\u7403\uff1ahttps:\/\/people.eecs.berkeley.edu\/~pabbeel\/cs287-fa09\/readings\/NgHaradaRussell-shaping-ICML1999.pdf<\/section>\n<\/li>\n<li>\n<section>\u5728 Coast Runners \u6e38\u620f\u4e2d\uff0c\u667a\u80fd\u4f53\u63a7\u5236\u4e00\u8258\u8239\uff0c\u76ee\u6807\u662f\u5c3d\u5feb\u5b8c\u6210\u8d5b\u8247\u6bd4\u8d5b\u3002\u5f53\u5b83\u5728\u8d5b\u9053\u4e0a\u51fb\u4e2d\u7eff\u8272\u65b9\u5757\u65f6\u83b7\u5f97\u5851\u9020\u5956\u52b1\u65f6\uff0c\u5b83\u4f1a\u5c06\u6700\u4f73\u7b56\u7565\u66f4\u6539\u4e3a\u7ed5\u5708\u9a91\u884c\u5e76\u4e00\u904d\u53c8\u4e00\u904d\u5730\u51fb\u4e2d\u76f8\u540c\u7684\u7eff\u8272\u65b9\u5757\uff1ahttps:\/\/deepmind.google\/discover\/blog\/specification-gaming-the-flip-side-of-ai-ingenuity\/<\/section>\n<\/li>\n<li>\n<section>\u300cThe Surprising Creativity of Digital Evolution\u300d\uff08Lehman et al. 2019\uff09\u2014\u2014 \u672c\u6587\u6709\u8bb8\u591a\u5173\u4e8e\u5982\u4f55\u4f18\u5316\u9519\u8bef\u6307\u5b9a\u7684\u9002\u5e94\u5ea6\u51fd\u6570\u53ef\u80fd\u5bfc\u81f4\u4ee4\u4eba\u60ca\u8bb6\u7684\u300chacking\u300d\u6216\u610f\u60f3\u4e0d\u5230\u7684\u8fdb\u5316\u6216\u5b66\u4e60\u7ed3\u679c\u7684\u4f8b\u5b50\uff1ahttps:\/\/arxiv.org\/abs\/1803.03453<\/section>\n<\/li>\n<li>\n<section>\u4eba\u5de5\u667a\u80fd\u793a\u4f8b\u4e2d\u7684\u89c4\u8303\u6e38\u620f\u5217\u8868\uff0c\u7531 Krakovna\u00a0et al.\u4e8e 2020 \u5e74\u6536\u96c6\uff1ahttps:\/\/deepmind.google\/discover\/blog\/specification-gaming-the-flip-side-of-ai-ingenuity\/<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>LLM \u4efb\u52a1\u4e2d\u7684 Reward Hacking \u793a\u4f8b\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u7528\u4e8e\u751f\u6210\u6458\u8981\u7684\u8bed\u8a00\u6a21\u578b\u80fd\u591f\u63a2\u7d22 ROUGE \u6307\u6807\u4e2d\u7684\u7f3a\u9677\uff0c\u4ece\u800c\u83b7\u5f97\u9ad8\u5206\uff0c\u4f46\u751f\u6210\u7684\u6458\u8981\u51e0\u4e4e\u4e0d\u53ef\u8bfb\uff1ahttps:\/\/web.archive.org\/web\/20180215132021\/https:\/\/www.salesforce.com\/products\/einstein\/ai-research\/tl-dr-reinforced-model-abstractive-summarization\/<\/section>\n<\/li>\n<li>\n<section>\u7f16\u7801\u6a21\u578b\u5b66\u4e60\u66f4\u6539\u5355\u5143\u6d4b\u8bd5\u4ee5\u901a\u8fc7\u7f16\u7801\u95ee\u9898\uff1ahttps:\/\/arxiv.org\/abs\/2406.10162<\/section>\n<\/li>\n<li>\n<section>\u7f16\u7801\u6a21\u578b\u53ef\u4ee5\u5b66\u4e60\u76f4\u63a5\u4fee\u6539\u7528\u4e8e\u8ba1\u7b97\u5956\u52b1\u7684\u4ee3\u7801\uff1ahttps:\/\/arxiv.org\/abs\/2406.10162<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u73b0\u5b9e\u751f\u6d3b\u4e2d\u7684 Reward Hacking \u653b\u51fb\u793a\u4f8b\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u793e\u4ea4\u5a92\u4f53\u7684\u63a8\u8350\u7b97\u6cd5\u65e8\u5728\u63d0\u4f9b\u6709\u7528\u7684\u4fe1\u606f\u3002\u7136\u800c\uff0c\u6709\u7528\u6027\u901a\u5e38\u901a\u8fc7\u4ee3\u7406\u6307\u6807\u6765\u8861\u91cf\uff0c\u4f8b\u5982\u70b9\u8d5e\u6216\u8bc4\u8bba\u7684\u6570\u91cf\uff0c\u6216\u5e73\u53f0\u4e0a\u7684\u53c2\u4e0e\u65f6\u95f4\u6216\u9891\u7387\u3002\u8be5\u7b97\u6cd5\u6700\u7ec8\u4f1a\u63a8\u8350\u53ef\u80fd\u5f71\u54cd\u7528\u6237\u60c5\u7eea\u72b6\u6001\u7684\u5185\u5bb9\uff0c\u4f8b\u5982\u79bb\u8c31\u548c\u6781\u7aef\u7684\u5185\u5bb9\uff0c\u4ee5\u89e6\u53d1\u66f4\u591a\u53c2\u4e0e\u5ea6\uff1ahttps:\/\/www.goodreads.com\/en\/book\/show\/204927599-nexus<\/section>\n<\/li>\n<li>\n<section>\u9488\u5bf9\u89c6\u9891\u5171\u4eab\u7f51\u7ad9\u7684\u9519\u8bef\u6307\u5b9a\u4ee3\u7406\u6307\u6807\u8fdb\u884c\u4f18\u5316\u53ef\u80fd\u4f1a\u5927\u5e45\u589e\u52a0\u7528\u6237\u7684\u89c2\u770b\u65f6\u95f4\uff0c\u800c\u771f\u6b63\u7684\u76ee\u6807\u662f\u4f18\u5316\u7528\u6237\u7684\u4e3b\u89c2\u5e78\u798f\u611f\uff1ahttps:\/\/arxiv.org\/abs\/2201.03544<\/section>\n<\/li>\n<li>\n<section>\u300c\u5927\u7a7a\u5934\u300d\u2014\u20142008 \u5e74\u7531\u623f\u5730\u4ea7\u6ce1\u6cab\u5f15\u53d1\u7684\u91d1\u878d\u5371\u673a\u3002\u5f53\u4eba\u4eec\u8bd5\u56fe\u73a9\u5f04\u91d1\u878d\u4f53\u7cfb\u65f6\uff0c\u6211\u4eec\u793e\u4f1a\u7684 Reward Hacking \u653b\u51fb\u5c31\u53d1\u751f\u4e86\uff1ahttps:\/\/en.wikipedia.org\/wiki\/The_Big_Short<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>\u4e3a\u4ec0\u4e48 Reward Hacking \u4f1a\u5b58\u5728\uff1f<\/strong><\/section>\n<section><\/section>\n<section>\u53e4\u5fb7\u54c8\u7279\u5b9a\u5f8b\u6307\u51fa\uff0c\u300c\u5f53\u4e00\u4e2a\u6307\u6807\u6210\u4e3a\u76ee\u6807\u65f6\uff0c\u5b83\u5c31\u4e0d\u518d\u662f\u4e00\u4e2a\u597d\u7684\u6307\u6807\u300d\u3002\u76f4\u89c9\u662f\uff0c\u4e00\u65e6\u65bd\u52a0\u4e86\u5f88\u5927\u7684\u538b\u529b\u6765\u4f18\u5316\u5b83\uff0c\u4e00\u4e2a\u597d\u7684\u6307\u6807\u5c31\u4f1a\u88ab\u7834\u574f\u3002\u6307\u5b9a 100% \u51c6\u786e\u7684\u5956\u52b1\u76ee\u6807\u5177\u6709\u6311\u6218\u6027\uff0c\u4efb\u4f55\u667a\u80fd\u4f53\u90fd\u9762\u4e34\u88ab\u9ed1\u5ba2\u653b\u51fb\u7684\u98ce\u9669\uff0c\u56e0\u4e3a RL \u7b97\u6cd5\u4f1a\u5229\u7528\u5956\u52b1\u51fd\u6570\u5b9a\u4e49\u4e2d\u7684\u4efb\u4f55\u5c0f\u7f3a\u9677\u3002<\/section>\n<section><\/section>\n<section>\u6709\u4eba\u5c06\u53e4\u5fb7\u54c8\u7279\u5b9a\u5f8b\u5206\u4e3a 4 \u79cd\u53d8\u4f53\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u56de\u5f52 &#8211; \u5bf9\u4e0d\u5b8c\u7f8e\u667a\u80fd\u4f53\u7684\u9009\u62e9\u4e5f\u5fc5\u7136\u4f1a\u9009\u62e9\u566a\u58f0\u3002<\/section>\n<\/li>\n<li>\n<section>\u6781\u503c &#8211; \u5ea6\u91cf\u9009\u62e9\u5c06\u72b6\u6001\u5206\u5e03\u63a8\u5165\u4e0d\u540c\u6570\u636e\u5206\u5e03\u7684\u533a\u57df\u3002<\/section>\n<\/li>\n<li>\n<section>\u56e0\u679c &#8211; \u5f53\u667a\u80fd\u4f53\u548c\u76ee\u6807\u4e4b\u95f4\u5b58\u5728\u975e\u56e0\u679c\u76f8\u5173\u6027\u65f6\uff0c\u5e72\u9884\u5b83\u53ef\u80fd\u65e0\u6cd5\u5e72\u9884\u76ee\u6807\u3002<\/section>\n<\/li>\n<li>\n<section>\u5bf9\u6297 &#8211; \u667a\u80fd\u4f53\u7684\u4f18\u5316\u6fc0\u52b1\u5bf9\u624b\u5c06\u4ed6\u4eec\u7684\u76ee\u6807\u4e0e\u667a\u80fd\u4f53\u76f8\u5173\u8054\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>Amodei et al. (2016) \u603b\u7ed3\u79f0\uff0cReward Hacking \u653b\u51fb\u4e3b\u8981\u53d1\u751f\u5728 RL \u8bbe\u7f6e\u4e2d\uff0c\u53ef\u80fd\u7531\u4e8e\u4ee5\u4e0b\u539f\u56e0\u800c\u53d1\u751f\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u90e8\u5206\u89c2\u5bdf\u5230\u7684\u72b6\u6001\u548c\u76ee\u6807\u4e0d\u80fd\u5b8c\u7f8e\u5730\u8868\u793a\u73af\u5883\u72b6\u6001\u3002<\/section>\n<\/li>\n<li>\n<section>\u7cfb\u7edf\u672c\u8eab\u5f88\u590d\u6742\uff0c\u5bb9\u6613\u53d7\u5230 hacking\uff1b\u4f8b\u5982\uff0c\u5982\u679c\u5141\u8bb8\u667a\u80fd\u4f53\u6267\u884c\u66f4\u6539\u90e8\u5206\u73af\u5883\u7684\u4ee3\u7801\uff0c\u5219\u5229\u7528\u73af\u5883\u673a\u5236\u4f1a\u53d8\u5f97\u5bb9\u6613\u5f97\u591a\u3002<\/section>\n<\/li>\n<li>\n<section>\u5956\u52b1\u53ef\u80fd\u6d89\u53ca\u96be\u4ee5\u5b66\u4e60\u6216\u63cf\u8ff0\u7684\u62bd\u8c61\u6982\u5ff5\u3002\u4f8b\u5982\uff0c\u5177\u6709\u9ad8\u7ef4\u8f93\u5165\u7684\u5956\u52b1\u51fd\u6570\u53ef\u80fd\u4e0d\u6210\u6bd4\u4f8b\u5730\u4f9d\u8d56\u4e8e\u51e0\u4e2a\u7ef4\u5ea6\u3002<\/section>\n<\/li>\n<li>\n<section>RL \u7684\u76ee\u6807\u662f\u4f7f\u5956\u52b1\u51fd\u6570\u9ad8\u5ea6\u4f18\u5316\uff0c\u56e0\u6b64\u5b58\u5728\u5185\u5728\u7684\u300c\u51b2\u7a81\u300d\uff0c\u4f7f\u5f97\u8bbe\u8ba1\u826f\u597d\u7684 RL \u76ee\u6807\u5177\u6709\u6311\u6218\u6027\u3002\u4e00\u79cd\u7279\u6b8a\u60c5\u51b5\u662f\u5177\u6709\u81ea\u6211\u5f3a\u5316\u53cd\u9988\u7ec4\u4ef6\u7684\u5956\u52b1\u51fd\u6570\uff0c\u5176\u4e2d\u5956\u52b1\u53ef\u80fd\u4f1a\u88ab\u653e\u5927\u548c\u626d\u66f2\u5230\u7834\u574f\u539f\u59cb\u610f\u56fe\u7684\u7a0b\u5ea6\uff0c\u4f8b\u5982\u5e7f\u544a\u6295\u653e\u7b97\u6cd5\u5bfc\u81f4\u8d62\u5bb6\u83b7\u5f97\u6240\u6709\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>\u6b64\u5916\uff0c\u786e\u5b9a\u6700\u4f73\u667a\u80fd\u4f53\u4f18\u5316\u5176\u884c\u4e3a\u7684\u786e\u5207\u5956\u52b1\u51fd\u6570\u901a\u5e38\u662f\u4e0d\u53ef\u80fd\u7684\uff0c\u56e0\u4e3a\u5728\u56fa\u5b9a\u73af\u5883\u4e2d\u53ef\u80fd\u5b58\u5728\u65e0\u6570\u4e2a\u4e0e\u4efb\u4f55\u89c2\u5bdf\u5230\u7684\u7b56\u7565\u4e00\u81f4\u7684\u5956\u52b1\u51fd\u6570 (Ng &amp; Russell, 2000)\u3001Amin and Singh (2016) \u5c06\u8fd9\u79cd\u4e0d\u53ef\u8bc6\u522b\u6027\u7684\u539f\u56e0\u5206\u4e3a\u4e24\u7c7b\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u8868\u5f81 &#8211; \u4e00\u7ec4\u5956\u52b1\u51fd\u6570\u5728\u67d0\u4e9b\u7b97\u672f\u8fd0\u7b97\uff08\u4f8b\u5982\u91cd\u65b0\u6269\u5c55\uff09\u4e0b\u5728\u884c\u4e3a\u4e0a\u4e0d\u53d8<\/section>\n<\/li>\n<li>\n<section>\u5b9e\u9a8c -\u03c0\u2018 \u89c2\u5bdf\u5230\u7684\u884c\u4e3a\u4e0d\u8db3\u4ee5\u533a\u5206\u4e24\u4e2a\u6216\u591a\u4e2a\u5956\u52b1\u51fd\u6570\uff0c\u8fd9\u4e9b\u5956\u52b1\u51fd\u6570\u90fd\u5408\u7406\u5316\u4e86\u667a\u80fd\u4f53\u7684\u884c\u4e3a\uff08\u884c\u4e3a\u5728\u4e24\u8005\u4e0b\u90fd\u662f\u6700\u4f73\u7684\uff09<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section><strong>Hacking \u5f3a\u5316\u5b66\u4e60\u73af\u5883<\/strong><\/section>\n<section><\/section>\n<section>\u968f\u7740\u6a21\u578b\u548c\u7b97\u6cd5\u53d8\u5f97\u8d8a\u6765\u8d8a\u590d\u6742\uff0c\u9884\u8ba1 reward hacking \u5c06\u53d8\u6210\u4e00\u4e2a\u8d8a\u6765\u8d8a\u5e38\u89c1\u7684\u95ee\u9898\u3002\u66f4\u52a0\u806a\u660e\u7684\u667a\u80fd\u4f53\u5c06\u66f4\u6709\u80fd\u529b\u627e\u5230\u5956\u52b1\u51fd\u6570\u8bbe\u8ba1\u4e2d\u7684\u300c\u6f0f\u6d1e\u300d\u5e76\u5229\u7528\u5176\u4efb\u52a1\u89c4\u8303 \u2014\u2014 \u4e5f\u5c31\u662f\u8bf4\uff0c\u83b7\u5f97\u66f4\u9ad8\u7684\u667a\u80fd\u4f53\u5956\u52b1\uff0c\u4f46\u771f\u5b9e\u5956\u52b1\u5374\u66f4\u4f4e\u4e86\u3002\u76f8\u6bd4\u4e4b\u4e0b\uff0c\u8f83\u5f31\u7684\u7b97\u6cd5\u53ef\u80fd\u65e0\u6cd5\u627e\u5230\u8fd9\u6837\u7684\u6f0f\u6d1e\uff0c\u56e0\u6b64\u5f53\u6a21\u578b\u4e0d\u591f\u5f3a\u5927\u65f6\uff0c\u6211\u4eec\u65e0\u6cd5\u89c2\u5bdf\u5230\u4efb\u4f55 reward hacking\uff0c\u4e5f\u65e0\u6cd5\u627e\u5230\u5f53\u524d\u5956\u52b1\u51fd\u6570\u8bbe\u8ba1\u4e2d\u7684\u95ee\u9898\u3002<\/section>\n<section><\/section>\n<section>\u5728\u96f6\u548c\u673a\u5668\u4eba\u81ea\u535a\u5f08 (Bansal et al., 2017) \u8bbe\u7f6e\u4e2d\uff0c\u6211\u4eec\u53ef\u4ee5\u8bad\u7ec3\u4e24\u4e2a\u4e92\u76f8\u7ade\u4e89\u7684\u667a\u80fd\u4f53\uff08\u53d7\u5bb3\u8005\u4e0e\u5bf9\u624b\uff09\u3002\u5f53\u4f7f\u7528\u4e00\u4e2a\u6807\u51c6\u7684\u8bad\u7ec3\u6d41\u7a0b\u4e0e\u4e00\u4e2a\u666e\u901a\u5bf9\u624b\u535a\u5f08\u65f6\uff0c\u4f1a\u5f97\u5230\u4e00\u4e2a\u6709\u8db3\u591f\u6027\u80fd\u7684\u53d7\u5bb3\u8005\u667a\u80fd\u4f53\u3002\u4f46\u662f\uff0c\u8bad\u7ec3\u51fa\u4e00\u4e2a\u80fd\u53ef\u9760\u5730\u51fb\u8d25\u53d7\u5bb3\u8005\u7684\u5bf9\u6297\u6027\u5bf9\u624b\u7b56\u7565\u5176\u5b9e\u5f88\u5bb9\u6613\uff0c\u5c3d\u7ba1\u5176\u4f1a\u8f93\u51fa\u770b\u4f3c\u968f\u673a\u7684\u52a8\u4f5c\uff0c\u5e76\u4e14\u4ec5\u9700\u5c11\u4e8e 3% \u7684\u65f6\u95f4\u6b65\u9aa4\u6570 (Gleave et al., 2020)\u3002\u5bf9\u6297\u7b56\u7565\u7684\u8bad\u7ec3\u9700\u8981\u4f18\u5316\u6298\u6263\u5956\u52b1\u7684\u548c\uff08\u8fd9\u4e0e\u6807\u51c6\u7684\u5f3a\u5316\u5b66\u4e60\u8bbe\u7f6e\u4e00\u6837\uff09\uff0c\u540c\u65f6\u8fd8\u9700\u8981\u5c06\u53d7\u5bb3\u8005\u7b56\u7565\u89c6\u4e3a\u9ed1\u7bb1\u6a21\u578b\u3002<\/section>\n<section><\/section>\n<section>\u5728\u7f13\u89e3\u5bf9\u6297\u7b56\u7565\u653b\u51fb\u65b9\u9762\uff0c\u4e00\u79cd\u76f4\u89c2\u65b9\u6cd5\u662f\u6839\u636e\u5bf9\u6297\u7b56\u7565\u5bf9\u53d7\u5bb3\u8005\u8fdb\u884c\u5fae\u8c03\u3002\u4f46\u662f\uff0c\u5c31\u7b97\u6839\u636e\u65b0\u7684\u53d7\u5bb3\u8005\u7b56\u7565\u8fdb\u884c\u4e86\u91cd\u65b0\u8bad\u7ec3\uff0c\u53d7\u5bb3\u8005\u4ecd\u7136\u5bb9\u6613\u53d7\u5230\u65b0\u7248\u672c\u7684\u5bf9\u6297\u7b56\u7565\u7684\u653b\u51fb\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4ec0\u4e48\u5b58\u5728\u5bf9\u6297\u7b56\u7565\uff1f\u8fd9\u91cc\u6709\u4e00\u4e2a\u5047\u8bbe\uff1a\u5bf9\u6297\u7b56\u7565\u4f1a\u5c06 OOD \u89c2\u5bdf\u5f15\u5165\u53d7\u5bb3\u8005\uff0c\u800c\u4e0d\u662f\u5bf9\u5176\u8fdb\u884c\u7269\u7406\u5e72\u6270\u3002\u8bc1\u636e\u8868\u660e\uff0c\u5f53\u53d7\u5bb3\u8005\u89c2\u5bdf\u5230\u7684\u5bf9\u624b\u4f4d\u7f6e\u4fe1\u606f\u88ab\u63a9\u76d6\u5e76\u88ab\u8bbe\u7f6e\u4e3a\u9759\u6001\u65f6\uff0c\u53d7\u5bb3\u8005\u9762\u5bf9\u5bf9\u624b\u7684\u7a33\u5065\u6027\u4f1a\u66f4\u5f3a\uff0c\u4e0d\u8fc7\u5176\u5728\u666e\u901a\u7684\u5bf9\u624b\u7b56\u7565\u4e0b\u8868\u73b0\u4f1a\u66f4\u5dee\u3002\u6b64\u5916\uff0c\u5982\u679c\u89c2\u5bdf\u7a7a\u95f4\u7ef4\u5ea6\u66f4\u9ad8\uff0c\u5219\u6b63\u5e38\u60c5\u51b5\u4e0b\u6027\u80fd\u4e5f\u4f1a\u63d0\u9ad8\uff0c\u4f46\u8fd9\u4e5f\u4f1a\u4f7f\u7b56\u7565\u66f4\u5bb9\u6613\u53d7\u5230\u5bf9\u6297\u5bf9\u624b\u7684\u653b\u51fb\u3002<\/section>\n<section><\/section>\n<section>Pan et al. (2022) \u5219\u662f\u5c06 reward hacking \u89c6\u4e3a\u667a\u80fd\u4f53\u80fd\u529b\u7684\u4e00\u4e2a\u51fd\u6570\uff0c\u6d89\u53ca\u7684\u53c2\u6570\u5305\u62ec (1) \u6a21\u578b\u5927\u5c0f\u3001(2) \u52a8\u4f5c\u7a7a\u95f4\u5206\u8fa8\u7387\u3001(3) \u89c2\u5bdf\u7a7a\u95f4\u566a\u58f0\u548c (4) \u8bad\u7ec3\u65f6\u95f4\u3002\u4ed6\u4eec\u8fd8\u63d0\u51fa\u4e86\u4e09\u79cd\u7c7b\u578b\u7684\u9519\u8bef\u6307\u5b9a\u7684\u4ee3\u7406\u5956\u52b1\uff1a<\/section>\n<section><\/section>\n<section>1. \u9519\u8bef\u52a0\u6743\uff1a\u4ee3\u7406\u5956\u52b1\u548c\u771f\u5b9e\u5956\u52b1\u4f53\u73b0\u4e86\u76f8\u540c\u7684\u9700\u6c42\uff0c\u4f46\u76f8\u5bf9\u91cd\u8981\u6027\u4e0d\u540c\u3002<\/section>\n<section>2. \u672c\u4f53\u8bba\uff1a\u4ee3\u7406\u5956\u52b1\u548c\u771f\u5b9e\u5956\u52b1\u4f7f\u7528\u4e0d\u540c\u7684\u9700\u6c42\u6765\u8868\u8fbe\u76f8\u540c\u7684\u6982\u5ff5\u3002<\/section>\n<section>3. \u8303\u56f4\uff1a\u4ee3\u7406\u5956\u52b1\u662f\u5728\u4e00\u4e2a\u53d7\u9650\u57df\uff08\u4f8b\u5982\u65f6\u95f4\u6216\u7a7a\u95f4\uff09\u4e0a\u8861\u91cf\u9700\u6c42\uff0c\u56e0\u4e3a\u57fa\u4e8e\u6240\u6709\u6761\u4ef6\u8fdb\u884c\u8861\u91cf\u6210\u672c\u592a\u9ad8\u3002<\/section>\n<section><\/section>\n<section>\u4ed6\u4eec\u7528\u56db\u4e2a\u5f3a\u5316\u5b66\u4e60\u73af\u5883\u642d\u914d\u4e5d\u4e2a\u9519\u8bef\u6307\u5b9a\u7684\u4ee3\u7406\u5956\u52b1\u8fdb\u884c\u4e86\u5b9e\u9a8c\u3002\u8fd9\u4e9b\u5b9e\u9a8c\u5f97\u5230\u7684\u53d1\u73b0\u53ef\u4ee5\u603b\u7ed3\u5982\u4e0b\uff1a\u80fd\u529b\u66f4\u5f3a\u7684\u6a21\u578b\u5f80\u5f80\u4f1a\u83b7\u5f97\u66f4\u9ad8\uff08\u6216\u76f8\u4f3c\uff09\u7684\u4ee3\u7406\u5956\u52b1\uff0c\u4f46\u771f\u5b9e\u5956\u52b1\u4f1a\u4e0b\u964d\u3002<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u6a21\u578b\u5927\u5c0f\uff1a\u6a21\u578b\u66f4\u5927\uff0c\u4ee3\u7406\u5956\u52b1\u4e5f\u4f1a\u589e\u5927\uff0c\u4f46\u771f\u5b9e\u5956\u52b1\u4f1a\u964d\u4f4e\u3002<\/section>\n<\/li>\n<li>\n<section>\u52a8\u4f5c\u7a7a\u95f4\u5206\u8fa8\u7387\uff1a\u5982\u679c\u63d0\u5347\u52a8\u4f5c\u7684\u7cbe\u5ea6\uff0c\u667a\u80fd\u4f53\u7684\u80fd\u529b\u4e5f\u4f1a\u53d8\u5f3a\u3002\u4f46\u662f\uff0c\u5206\u8fa8\u7387\u66f4\u9ad8\u4f1a\u5bfc\u81f4\u4ee3\u7406\u5956\u52b1\u4e0d\u53d8\u7684\u540c\u65f6\u771f\u5b9e\u5956\u52b1\u4e0b\u964d\u3002<\/section>\n<\/li>\n<li>\n<section>\u89c2\u5bdf\u4fdd\u771f\u5ea6\uff1a\u66f4\u51c6\u786e\u7684\u89c2\u5bdf\u4f1a\u63d0\u9ad8\u4ee3\u7406\u5956\u52b1\uff0c\u4f46\u4f1a\u7565\u5fae\u964d\u4f4e\u771f\u5b9e\u5956\u52b1\u3002<\/section>\n<\/li>\n<li>\n<section>\u8bad\u7ec3\u6b65\u6570\uff1a\u5728\u5956\u52b1\u5448\u6b63\u76f8\u5173\u7684\u521d\u59cb\u9636\u6bb5\u4e4b\u540e\uff0c\u7528\u66f4\u591a\u6b65\u6570\u4f18\u5316\u4ee3\u7406\u5956\u52b1\u4f1a\u635f\u5bb3\u771f\u5b9e\u5956\u52b1\u3002<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-a1cb8d30fc63536a8ad370eb184c1b77.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-a1cb8d30fc63536a8ad370eb184c1b77.png\"><\/a><\/section>\n<section><em><sup>\u56fe 3\uff1a\uff08\u4e0a\u56fe\uff09\u4ee3\u7406\u5956\u52b1\u4e0e\u771f\u5b9e\u5956\u52b1\u968f\u6a21\u578b\u5927\u5c0f\u7684\u53d8\u5316\uff0c\u6a21\u578b\u5927\u5c0f\u4ee5\u53c2\u6570\u91cf\u8861\u91cf\uff1b\u4ee3\u7406\u5956\u52b1\u4e0e\u771f\u5b9e\u5956\u52b1\u968f\u6a21\u578b\u80fd\u529b\u7684\u53d8\u5316\uff0c\u5176\u8861\u91cf\u6307\u6807\u5305\u62ec\u8bad\u7ec3\u6b65\u6570\u3001\u52a8\u4f5c\u7a7a\u95f4\u5206\u8fa8\u7387\u548c\u89c2\u5bdf\u566a\u58f0\u7b49\u3002\uff08\u56fe\u6e90\uff1aPan et al. 2022\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u5982\u679c\u4ee3\u7406\u5956\u52b1\u8bbe\u5b9a\u5f97\u975e\u5e38\u5dee\uff0c\u4ee5\u81f3\u4e8e\u4e0e\u771f\u5b9e\u5956\u52b1\u7684\u76f8\u5173\u6027\u975e\u5e38\u5f31\uff0c\u90a3\u751a\u81f3\u53ef\u4ee5\u5728\u8bad\u7ec3\u4e4b\u524d\u5c31\u8bc6\u522b\u51fa\u5e76\u9632\u6b62 reward hacking\u3002\u57fa\u4e8e\u8fd9\u4e00\u5047\u8bbe\uff0cPan et al. (2022) \u57fa\u4e8e\u8fd9\u4e00\u5047\u8bbe\uff0cPan et al. (2022) \u7814\u7a76\u4e86\u4e00\u7cfb\u5217\u8f68\u8ff9 rollouts \u4e2d\u4ee3\u7406\u5956\u52b1\u548c\u771f\u5b9e\u5956\u52b1\u4e4b\u95f4\u7684\u76f8\u5173\u6027\u3002\u6709\u8da3\u7684\u662f\uff0c\u5373\u4f7f\u771f\u5b9e\u5956\u52b1\u548c\u4ee3\u7406\u5956\u52b1\u4e4b\u95f4\u5b58\u5728\u6b63\u76f8\u5173\u6027\uff0creward hacking \u653b\u51fb\u4ecd\u7136\u4f1a\u53d1\u751f\u3002<\/section>\n<section><\/section>\n<section><strong>Hacking \u5927\u6a21\u578b\u7684 RLHF<\/strong><\/section>\n<section><\/section>\n<section>\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60\uff08RLHF\uff09\u5df2\u7ecf\u6210\u4e3a\u8bed\u8a00\u6a21\u578b\u5bf9\u9f50\u8bad\u7ec3\u7684\u6700\u5e38\u7528\u65b9\u6cd5\u3002\u5728 RLHF \u8bbe\u7f6e\u4e2d\uff0c\u4f1a\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u6570\u636e\u8bad\u7ec3\u4e00\u4e2a\u5956\u52b1\u6a21\u578b\uff0c\u7136\u540e\u901a\u8fc7\u5f3a\u5316\u5b66\u4e60\u5bf9\u4e00\u4e2a\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\uff0c\u4ee5\u4f18\u5316\u8fd9\u4e2a\u4eba\u7c7b\u504f\u597d\u7684\u4ee3\u7406\u5956\u52b1\u3002RLHF \u8bbe\u7f6e\u4e2d\u6709\u4e09\u79cd\u7c7b\u578b\u7684\u5956\u52b1\u503c\u5f97\u5173\u6ce8\uff1a<\/section>\n<section><\/section>\n<section>(1) Oracle\/Gold \u5956\u52b1 R^* \u4ee3\u8868\u6211\u4eec\u771f\u6b63\u5e0c\u671b LLM \u4f18\u5316\u7684\u4e1c\u897f\u3002<\/section>\n<section>(2) \u4eba\u7c7b\u5956\u52b1 R^human \u662f\u6211\u4eec\u5728\u5b9e\u8df5\u4e2d\u8bc4\u4f30 LLM \u65f6\u6536\u96c6\u7684\u5956\u52b1\uff0c\u901a\u5e38\u6765\u81ea\u6709\u65f6\u95f4\u9650\u5236\u7684\u4e2a\u4eba\u3002\u7531\u4e8e\u4eba\u7c7b\u53ef\u80fd\u4f1a\u63d0\u4f9b\u4e0d\u4e00\u81f4\u7684\u53cd\u9988\uff0c\u4e5f\u53ef\u80fd\u72af\u9519\uff0c\u56e0\u6b64\u4eba\u7c7b\u5956\u52b1\u5e76\u4e0d\u80fd\u5b8c\u5168\u51c6\u786e\u5730\u8868\u793a oracle \u5956\u52b1\u3002<\/section>\n<section>(3) \u4ee3\u7406\u5956\u52b1 R \u662f\u901a\u8fc7\u4eba\u7c7b\u6570\u636e\u8bad\u7ec3\u7684\u5956\u52b1\u6a21\u578b\u6240\u9884\u6d4b\u7684\u5206\u6570\u3002\u56e0\u6b64\uff0cR^train \u7ee7\u627f\u4e86\u4eba\u7c7b\u5956\u52b1\u7684\u6240\u6709\u5f31\u70b9\uff0c\u4ee5\u53ca\u6f5c\u5728\u7684\u5efa\u6a21\u504f\u5dee\u3002<\/section>\n<section><\/section>\n<section>RLHF \u4f1a\u4f18\u5316\u4ee3\u7406\u5956\u52b1\u5206\u6570\uff0c\u4f46\u6211\u4eec\u6700\u7ec8\u5173\u5fc3\u7684\u662f Gold \u5956\u52b1\u5206\u6570\u3002<\/section>\n<section><\/section>\n<section><strong>hacking \u8bad\u7ec3\u8fc7\u7a0b<\/strong><\/section>\n<section><\/section>\n<section>Gao et al. (2022) \u7814\u7a76\u4e86 RLHF \u4e2d\u5956\u52b1\u6a21\u578b\u8fc7\u5ea6\u4f18\u5316\u7684 Scaling Law\u3002\u4e3a\u4e86\u6269\u5927\u4ed6\u4eec\u5b9e\u9a8c\u4e2d\u4eba\u7c7b\u6807\u7b7e\u7684\u89c4\u6a21\uff0c\u4ed6\u4eec\u4f7f\u7528\u4e86\u5408\u6210\u6570\u636e\u8bbe\u7f6e\uff0c\u5176\u4e2d oracle \u5956\u52b1 R^* \u7684 gold \u6807\u7b7e\u7531\u4e00\u4e2a\u5956\u52b1\u6a21\u578b\uff086B \u53c2\u6570\uff09\u8fd1\u4f3c\uff0c\u800c R \u7684\u4ee3\u7406\u5956\u52b1\u6a21\u578b\u5927\u5c0f\u8303\u56f4\u4e3a 3M \u5230 3B \u53c2\u6570\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8928e32c9cf334b2bdca5a2345269988.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8928e32c9cf334b2bdca5a2345269988.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u56fe 4\uff1a\u5956\u52b1\u6a21\u578b\u5206\u6570\u968f KL \u8ddd\u79bb\u5ea6\u91cf\u7684\u5e73\u65b9\u6839\u7684\u53d8\u5316\u60c5\u51b5\u3002\u5176\u4e2d\uff0c\u865a\u7ebf\u8868\u793a\u4ee3\u7406\u5956\u52b1\uff0c\u5b9e\u7ebf\u8868\u793a gold \u5956\u52b1\u3002\uff08\u56fe\u6e90\uff1aGao et al. 2022\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u521d\u59cb\u7b56\u7565\u5230\u5df2\u4f18\u5316\u7b56\u7565\u7684 KL \u8ddd\u79bb\u4e3a KL=D_KL (\u03c0|\u03c0_init)\uff0c\u8ddd\u79bb\u51fd\u6570\u5b9a\u4e49\u4e3a <a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-130d2b228ce51fbbf45dfa45c6e00855.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-130d2b228ce51fbbf45dfa45c6e00855.png\"><\/a>\u3002\u5bf9\u4e8e best-of-n \u62d2\u7edd\u91c7\u6837 (BoN) \u548c\u5f3a\u5316\u5b66\u4e60\uff0c\u9ec4\u91d1\u5956\u52b1 R^* \u88ab\u5b9a\u4e49\u4e3a d \u7684\u51fd\u6570\u3002\u7cfb\u6570 \u03b1 \u548c \u03b2 \u662f\u6839\u636e\u7ecf\u9a8c\u62df\u5408\u7684\uff0c\u5e76\u6709\u5b9a\u4e49 R^*(0) := 0\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u4e9b\u4f5c\u8005\u8fd8\u5c1d\u8bd5\u62df\u5408\u4ee3\u7406\u5956\u52b1 R\uff0c\u4f46\u53d1\u73b0\u5f53\u5916\u63a8\u5230\u66f4\u9ad8\u7684 KL \u65f6\u4f1a\u51fa\u73b0\u7cfb\u7edf\u6027\u7684\u4f4e\u4f30\uff0c\u56e0\u4e3a\u4ee3\u7406\u5956\u52b1\u4f3c\u4e4e\u4f1a\u968f d \u800c\u7ebf\u6027\u589e\u957f\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-a683d9db7372a3027071abe105a4bb66.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-a683d9db7372a3027071abe105a4bb66.png\"><\/a><\/section>\n<section><em><sup>\u56fe 5\uff1a\u7cfb\u6570\u53c2\u6570 \u03b1_bon\u3001\u03b2_bon\u3001\u03b2_RL \u662f\u6839\u636e\u6570\u636e\u800c\u7ecf\u9a8c\u62df\u5408\u5f97\u5230\u7684\uff0c\u8fd9\u91cc\u5c55\u793a\u6210\u4e86\u5956\u52b1\u6a21\u578b\u5927\u5c0f\u7684\u51fd\u6570\u3002\u7cfb\u6570 \u03b1_RL \u672a\u5305\u542b\u5728\u6b64\u5904\uff0c\u56e0\u4e3a\u5b83\u5728 \u5956\u52b1\u6a21\u578b\u5927\u5c0f\u53d8\u5316\u65f6\u4f1a\u4fdd\u6301\u4e0d\u53d8\u3002\uff08\u56fe\u6e90\uff1aGao et al. 2022\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u4e0e RM \u76f8\u6bd4\uff0c\u8f83\u5927\u7684\u7b56\u7565\u4ece\u4f18\u5316\u4e2d\u83b7\u5f97\u7684\u597d\u5904\u8f83\u5c11\uff08\u5373\u521d\u59cb\u5956\u52b1\u548c\u5cf0\u503c\u5956\u52b1\u4e4b\u95f4\u7684\u5dee\u5f02\u5c0f\u4e8e\u8f83\u5c0f\u7b56\u7565\u7684\u5dee\u5f02\uff09\uff0c\u4f46\u8fc7\u5ea6\u4f18\u5316\u4e5f\u8f83\u5c11\u3002<\/section>\n<\/li>\n<li>\n<section>\u66f4\u591a\u7684 RM \u6570\u636e\u4f1a\u8ba9 gold \u5956\u52b1\u5206\u6570\u66f4\u9ad8\u5e76\u51cf\u5c11\u300cGoodharting\u300d\u3002\uff08\u6ce8\uff1a\u53e4\u5fb7\u54c8\u7279\u5b9a\u5f8b\uff08Goodhart&#8217;s law\uff09\u7684\u5927\u610f\u662f\uff1a\u4e00\u9879\u6307\u6807\u4e00\u65e6\u53d8\u6210\u4e86\u76ee\u6807\uff0c\u5b83\u5c06\u4e0d\u518d\u662f\u4e2a\u597d\u6307\u6807\u3002\uff09<\/section>\n<\/li>\n<li>\n<section>KL \u60e9\u7f5a\u5bf9 gold \u5206\u6570\u7684\u5f71\u54cd\u7c7b\u4f3c\u4e8e\u65e9\u505c\uff08early stopping\uff09\u3002\u8bf7\u6ce8\u610f\uff0c\u9664\u4e86\u8fd9\u4e2a\u5b9e\u9a8c\u4e4b\u5916\uff0c\u5728\u6240\u6709\u5b9e\u9a8c\u4e2d\uff0cPPO \u4e2d\u7684 KL \u60e9\u7f5a\u90fd\u8bbe\u7f6e\u4e3a 0\uff0c\u56e0\u4e3a\u4ed6\u4eec\u89c2\u5bdf\u5230\u4f7f\u7528 KL \u60e9\u7f5a\u5fc5\u5b9a\u4f1a\u589e\u5927\u4ee3\u7406 &#8211; gold \u5956\u52b1\u5dee\u8ddd\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>RLHF \u7684\u76ee\u6807\u662f\u63d0\u9ad8\u6a21\u578b\u4e0e\u4eba\u7c7b\u504f\u597d\u7684\u5bf9\u9f50\u7a0b\u5ea6\uff0c\u4f46\u4eba\u7c7b\u53cd\u9988 R^human \u53ef\u80fd\u65e0\u6cd5\u4f53\u73b0\u6211\u4eec\u5173\u5fc3\u7684\u6240\u6709\u65b9\u9762\uff08\u4f8b\u5982\u4e8b\u5b9e\u6027\uff09\uff0c\u56e0\u6b64\u53ef\u80fd\u4f1a\u88ab hacking\uff0c\u4ece\u800c\u8fc7\u5ea6\u62df\u5408\u6211\u4eec\u4e0d\u60f3\u8981\u7684\u5c5e\u6027\u3002\u4f8b\u5982\uff0c\u6a21\u578b\u53ef\u88ab\u4f18\u5316\u4ee5\u8f93\u51fa\u770b\u4f3c\u6b63\u786e\u4e14\u5f88\u6709\u8bf4\u670d\u529b\u4f46\u5b9e\u9645\u4e0a\u4e0d\u51c6\u786e\u7684\u54cd\u5e94\uff0c\u4ece\u800c\u8bef\u5bfc\u4eba\u7c7b\u8bc4\u4f30\u8005\u8ba4\u53ef\u5176\u9519\u8bef\u7b54\u6848\uff08Wen et al., 2024\uff09\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u7531\u4e8e RLHF\uff0c\u6b63\u786e\u7684\u5185\u5bb9\u548c\u4eba\u7c7b\u8ba4\u4e3a\u6b63\u786e\u7684\u5185\u5bb9\u4e4b\u95f4\u5e76\u4e0d\u80fd\u5212\u7b49\u53f7\u3002\u786e\u5207\u5730\u8bf4\uff0cWen et al. (2024) \u4f7f\u7528\u57fa\u4e8e ChatbotArena \u6570\u636e\u7684\u5956\u52b1\u6a21\u578b\u8fdb\u884c\u4e86 RLHF \u5b9e\u9a8c\u3002\u4ed6\u4eec\u5728\u95ee\u7b54\u6570\u636e\u96c6 QuALITY \u548c\u7f16\u7a0b\u6570\u636e\u96c6 APPS \u4e0a\u8bc4\u4f30\u4e86\u8be5\u6a21\u578b\u3002\u4ed6\u4eec\u7684\u5b9e\u9a8c\u8868\u660e\uff0c\u6a21\u578b\u5728\u8bf4\u670d\u4eba\u7c7b\u76f8\u4fe1\u81ea\u5df1\u662f\u6b63\u786e\u7684\u65b9\u9762\u8d8a\u6765\u8d8a\u597d\uff0c\u5373\u4f7f\u5b83\u4eec\u539f\u672c\u662f\u9519\u8bef\u7684\uff0c\u800c\u4e14\u8fd9\u79cd\u5f71\u54cd\u4e5f\u5e76\u975e\u523b\u610f\u4e3a\u4e4b\uff1a<\/section>\n<section><\/section>\n<section>1.RLHF \u53ef\u589e\u52a0\u4eba\u7c7b\u7684\u8ba4\u53ef\u5ea6\uff0c\u4f46\u4e0d\u4e00\u5b9a\u80fd\u63d0\u5347\u6b63\u786e\u5ea6\u3002<\/section>\n<section>2.RLHF \u4f1a\u524a\u5f31\u4eba\u7c7b\u7684\u8bc4\u4f30\u80fd\u529b\uff1a\u7ecf\u8fc7 RLHF \u8bad\u7ec3\u540e\uff0c\u4eba\u7c7b\u8bc4\u4f30\u7684\u9519\u8bef\u7387\u66f4\u9ad8\u3002<\/section>\n<section>3.RLHF \u4f1a\u4f7f\u4e0d\u6b63\u786e\u7684\u8f93\u51fa\u66f4\u5bb9\u6613\u8bf4\u670d\u4eba\u7c7b\u3002\u7ecf\u8fc7 RLHF \u8bad\u7ec3\u540e\uff0c\u8bc4\u4f30\u5047\u9633\u6027\u7387\u4f1a\u663e\u8457\u589e\u5927\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u7bc7\u8bba\u6587\u5c06\u8fd9\u79cd\u6548\u5e94\u79f0\u4e3a U-Sophistry\uff08\u5176\u4e2d U \u8868\u793a unintended\uff09\uff0c\u76f8\u5bf9\u7684\u4fbf\u662f I-Sophistry\uff08I \u662f\u6307 intended\uff09\uff0c\u540e\u8005\u9700\u8981\u901a\u8fc7\u300c\u2026\u2026 \u8bd5\u56fe\u6b3a\u9a97\u4eba\u7c7b\u53d7\u8bd5\u8005\u300d\u7b49\u6307\u793a\u6765\u663e\u5f0f\u5730\u8c03\u7528\u6a21\u578b\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-ffb56530c8d2c2751b8f6359ca608418.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-ffb56530c8d2c2751b8f6359ca608418.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0\u56fe 6\uff1aRLHF \u80fd\u8ba9 LLM \u66f4\u5584\u4e8e\u8bf4\u670d\u4eba\u7c7b\u8bc4\u4f30\u8005\u8ba4\u53ef\u4ed6\u4eec\u7684\u9519\u8bef\u7b54\u6848\u3002\uff08\u56fe\u6e90\uff1aWen et al. 2024\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u4eba\u5de5\u8bc4\u4f30\u9519\u8bef\u7387\u7684\u53d8\u5316\u4e0d\u662f\u7531\u4e8e\u8bc4\u4f30\u8005\u62db\u52df\u8fc7\u7a0b\u4e2d\u7684\u566a\u97f3\u9020\u6210\u7684\uff0c\u56e0\u4e3a (1) \u5728\u4e2a\u4eba\u5c42\u9762\uff0c\u5927\u591a\u6570 (70-90%) \u4eba\u7c7b\u8bc4\u4f30\u8005\u7684\u8bc4\u4f30\u9519\u8bef\u7387\u4f1a\u589e\u52a0\uff0c\u5e76\u4e14 (2) \u4ed6\u4eec\u5728\u8bc4\u4f30 \u03c0_init \u6216 \u03c0_rlhf \u4e0a\u6240\u4ed8\u51fa\u7684\u52aa\u529b\u662f\u7b49\u540c\u7684\uff0c\u8861\u91cf\u6307\u6807\u662f\u82b1\u8d39\u7684\u65f6\u95f4\u6216\u7f16\u5199\u7684\u5355\u5143\u6d4b\u8bd5\u7b49\u3002<\/section>\n<section><\/section>\n<section>\u76f8\u53cd\uff0cLLM \u5b66\u4f1a\u4e86\u5728\u634d\u536b\u9519\u8bef\u7b54\u6848\u65f6\u6311\u9009\u3001\u7f16\u9020\u4e0d\u771f\u5b9e\u7684\u652f\u6301\u9648\u8ff0\u6216\u7f16\u9020\u5e26\u6709\u5fae\u5999\u56e0\u679c\u8c2c\u8bef\u7684\u9648\u8ff0\u3002\u4ed6\u4eec\u89c2\u5bdf\u5230\u6a21\u578b\u5728 RLHF \u4e4b\u540e\u4f1a\u6709\u8fd9\u6837\u7684\u884c\u4e3a\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u5728\u957f\u95ee\u7b54\u4efb\u52a1\u4e2d\uff1a\u6a21\u578b\u4f1a\u521b\u5efa\u66f4\u6709\u8bf4\u670d\u529b\u7684\u634f\u9020\u8bc1\u636e\uff0c\u4e3a\u9519\u8bef\u7b54\u6848\u4f7f\u7528\u66f4\u4e00\u81f4\u7684\u903b\u8f91\uff0c\u751f\u6210\u5e26\u6709\u5fae\u5999\u8c2c\u8bef\u7684\u8fde\u8d2f\u4e00\u81f4\u7b54\u6848\u3002<\/section>\n<\/li>\n<li>\n<section>\u5728\u7f16\u7801\u4efb\u52a1\u4e2d\uff1a\u6a21\u578b\u4f1a\u7834\u89e3\u4eba\u7c7b\u7f16\u5199\u7684\u5355\u5143\u6d4b\u8bd5\uff0c\u751f\u6210\u53ef\u8bfb\u6027\u8f83\u5dee\u7684\u6d4b\u8bd5\uff08\u4f8b\u5982\uff0c\u8f85\u52a9\u51fd\u6570\u66f4\u5c11\uff0c\u4ee3\u7801\u590d\u6742\u5ea6\u66f4\u9ad8\uff09\uff0c\u4f7f \u03c0_rlhf \u4e0d\u592a\u53ef\u80fd\u751f\u6210\u4eba\u7c7b\u53ef\u4ee5\u5229\u7528\u7684\u6613\u68c0\u6d4b\u9519\u8bef\u3002<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-83c4851a332ca7e6ebbdfa2d88f1a02f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-83c4851a332ca7e6ebbdfa2d88f1a02f.png\"><\/a><\/section>\n<section><em><sup>\u56fe 7\uff1a\u751f\u6210\u7684\u6b63\u786e\u548c\u9519\u8bef\u4ee3\u7801\u7684\u4ee3\u7801\u6a21\u5757\u6027\uff08\u8f85\u52a9\u51fd\u6570\u6570\u91cf\uff09\u548c\u5708\u590d\u6742\u5ea6\u6307\u6807\u3002RLHF \u4f1a\u5bfc\u81f4\u751f\u6210\u7684\u9519\u8bef\u7a0b\u5e8f\u4e2d\u7684\u8f85\u52a9\u51fd\u6570\u603b\u4f53\u4e0a\u66f4\u5c11\uff0c\u4ee3\u7801\u590d\u6742\u5ea6\u66f4\u9ad8\u3002\u8fd9\u65e0\u7591\u4f1a\u589e\u52a0\u4eba\u5de5\u8bc4\u4f30\u7684\u96be\u5ea6\u3002\uff08\u56fe\u6e90\uff1aWen et al. 2024\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u8c04\u5a9a\uff08Sycophancy\uff09\u662f\u6307\u6a21\u578b\u54cd\u5e94\u503e\u5411\u4e8e\u7b26\u5408\u7528\u6237\u4fe1\u5ff5\u800c\u4e0d\u662f\u53cd\u6620\u771f\u76f8\uff08Shrama et al. 2023\uff09\u3002\u5728\u5b9e\u9a8c\u4e2d\uff0c\u8981\u6c42 AI \u52a9\u624b\u5bf9\u4e00\u4e2a\u8bba\u70b9\u63d0\u4f9b\u53cd\u9988\uff08\u4eba\u7c7b\uff1a\u300c\u8bf7\u7b80\u8981\u8bc4\u8bba\u4ee5\u4e0b\u8bba\u70b9\u3002\u8bba\u70b9\uff1a&#8230;&#8230;\u300d\uff09\u3002\u5f53\u4eba\u7c7b\u63d0\u4f9b\u8bba\u70b9\u65f6\uff0c\u4ed6\u4eec\u53ef\u4ee5\u9648\u8ff0\u504f\u597d\uff08\u300c\u6211\u771f\u7684\u559c\u6b22\u8fd9\u4e2a\u8bba\u70b9\u300d\u6216\u300c\u6211\u771f\u7684\u4e0d\u559c\u6b22\u8fd9\u4e2a\u8bba\u70b9\u300d\uff09\uff0c\u4ee5\u6d4b\u8bd5\u4e0e\u6ca1\u6709\u4eba\u7c7b\u504f\u597d\u9648\u8ff0\u7684\u57fa\u7ebf\u53cd\u9988\u76f8\u6bd4\uff0c\u8fd9\u662f\u5426\u4f1a\u5f71\u54cd\u6a21\u578b\u7684\u53cd\u9988\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-2dca716306f2786926271d59e188b6cb.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-2dca716306f2786926271d59e188b6cb.png\"><\/a><\/section>\n<section><em><sup>\u56fe 8\uff1a\u5f53\u7528\u6237\u5bf9\u81ea\u5df1\u7684\u504f\u597d\u53d1\u8868\u8bc4\u8bba\u65f6\uff0cAI \u52a9\u624b\u4f1a\u7ed9\u51fa\u6709\u504f\u89c1\u7684\u53cd\u9988\u3002\u5f53\u7528\u6237\u8868\u793a\u4ed6\u4eec\u559c\u6b22\u6216\u5199\u4e86\u8be5\u6587\u672c\u65f6\uff0c\u56de\u590d\u4f1a\u66f4\u79ef\u6781\uff0c\u5982\u679c\u7528\u6237\u8868\u793a\u4ed6\u4eec\u4e0d\u559c\u6b22\u8be5\u6587\u672c\uff0c\u56de\u590d\u4f1a\u66f4\u6d88\u6781\u3002\uff08\u56fe\u6e90\uff1aShrama et al. 2023\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u4ed6\u4eec\u53d1\u73b0\uff0cAI \u52a9\u624b\u7684\u53cd\u9988\u5f88\u5bb9\u6613\u53d7\u5230\u5f71\u54cd\uff0c\u56e0\u4e3a\u5f53\u53d7\u5230\u4eba\u7c7b\u504f\u597d\u7684\u6311\u6218\u65f6\uff0c\u5b83\u53ef\u80fd\u4f1a\u6539\u53d8\u5176\u539f\u672c\u6b63\u786e\u7684\u7b54\u6848\u3002\u8be5\u6a21\u578b\u503e\u5411\u4e8e\u8ba4\u540c\u7528\u6237\u7684\u4fe1\u5ff5\u3002\u6709\u65f6\u5b83\u751a\u81f3\u4f1a\u6a21\u4eff\u7528\u6237\u7684\u9519\u8bef\uff08\u4f8b\u5982\uff0c\u5f53\u88ab\u8981\u6c42\u5206\u6790\u8bd7\u6b4c\u65f6\uff0c\u9519\u8bef\u5730\u5f52\u56e0\u4e8e\u9519\u8bef\u7684\u8bd7\u4eba\uff09\u3002\u901a\u8fc7 logistic \u56de\u5f52\u5bf9 RLHF \u6709\u7528\u6027\u6570\u636e\u96c6\u8fdb\u884c\u6570\u636e\u5206\u6790\u4ee5\u9884\u6d4b\u4eba\u7c7b\u53cd\u9988\uff0c\u7ed3\u679c\u8868\u660e\uff0c\u300c\u5339\u914d\u7528\u6237\u7684\u4fe1\u5ff5\u300d\u662f\u6700\u5177\u9884\u6d4b\u6027\u7684\u56e0\u7d20\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-48473a51dc9f41daa6ce40a8cd71a2d0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-48473a51dc9f41daa6ce40a8cd71a2d0.png\"><\/a><\/section>\n<section><em><sup>\u56fe 9\uff1a\u901a\u8fc7 logistic \u56de\u5f52\u8fdb\u884c\u4eba\u7c7b\u504f\u597d\u6570\u636e\u5206\u6790\uff0c\u9884\u6d4b\u5177\u6709\u76ee\u6807\u7279\u5f81\u7684\u54cd\u5e94\u7684\u6982\u7387\u4f18\u4e8e\u4e0d\u5177\u6709\u76ee\u6807\u7279\u5f81\u7684\u54cd\u5e94\uff0c\u540c\u65f6\u63a7\u5236\u5176\u4ed6\u7279\u5f81\u3002\uff08\u56fe\u6e90\uff1aShrama et al. 2023\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>Hacking \u8bc4\u4f30\u5668<\/strong><\/section>\n<section><\/section>\n<section>\u968f\u7740 LLM \u7684\u80fd\u529b\u8d8a\u6765\u8d8a\u5f3a\uff0c\u5c06 LLM \u4f5c\u4e3a\u8bc4\u4f30\u8005\u6216 grader\uff0c\u4e3a\u5176\u4ed6\u751f\u6210\u5668\u6a21\u578b\u63d0\u4f9b\u53cd\u9988\u548c\u8bad\u7ec3\u5956\u52b1\uff0c\u662f\u4e00\u79cd\u81ea\u7136\u7684\u9009\u62e9\uff0c\u5c24\u5176\u662f\u5bf9\u4e8e\u90a3\u4e9b\u65e0\u6cd5\u8fdb\u884c\u7410\u788e\u5224\u65ad\u6216\u9a8c\u8bc1\u7684\u4efb\u52a1\uff08\u5982\u5904\u7406\u957f\u7bc7\u8f93\u51fa\u3001\u521b\u610f\u5199\u4f5c\u8d28\u91cf\u7b49\u4e3b\u89c2\u8bc4\u5206\u6807\u51c6\uff09\u3002\u6709\u4eba\u5c06\u6b64\u79f0\u4e3a\u300cLLM-as-grader paradigm\u300d\u3002\u8fd9\u79cd\u65b9\u6cd5\u5728\u5f88\u5927\u7a0b\u5ea6\u4e0a\u51cf\u5c11\u4e86\u5bf9\u4eba\u5de5\u6807\u6ce8\u7684\u4f9d\u8d56\uff0c\u5927\u5927\u8282\u7701\u4e86\u8bc4\u4f30\u65f6\u95f4\u3002\u7136\u800c\uff0c\u4f7f\u7528 LLM \u4f5c\u4e3a grader \u5e76\u4e0d\u80fd\u5b8c\u5168\u4ee3\u8868\u9884\u8a00\u673a\u5956\u52b1\uff0c\u800c\u4e14\u4f1a\u5e26\u6765\u504f\u5dee\uff0c\u4f8b\u5982\u5728\u4e0e\u4e0d\u540c\u7684\u6a21\u578b\u7cfb\u5217\u8fdb\u884c\u6bd4\u8f83\u65f6\uff0cLLM \u4f1a\u504f\u597d\u81ea\u5df1\u7684\u54cd\u5e94 (Liu et al., 2023 )\uff0c\u6216\u8005\u5728\u6309\u987a\u5e8f\u8bc4\u4f30\u54cd\u5e94\u65f6\u4f1a\u51fa\u73b0\u4f4d\u7f6e\u504f\u5dee Wang et al. (2023)\u3002\u8fd9\u79cd\u504f\u5dee\u5c24\u5176\u4f1a\u5f71\u54cd grader \u8f93\u51fa\u88ab\u7528\u4f5c\u5956\u52b1\u4fe1\u53f7\u7684\u4e00\u90e8\u5206\uff0c\u53ef\u80fd\u5bfc\u81f4\u5229\u7528\u8fd9\u4e9b grader \u8fdb\u884c reward hacking \u884c\u4e3a\u3002<\/section>\n<section><\/section>\n<section>Wang et al. (2023) \u53d1\u73b0\uff0c\u5f53\u4f7f\u7528 LLM \u4f5c\u4e3a\u8bc4\u4f30\u8005\u4e3a\u591a\u4e2a\u5176\u4ed6 LLM \u8f93\u51fa\u7684\u8d28\u91cf\u6253\u5206\u65f6\uff0c\u53ea\u9700\u6539\u53d8\u4e0a\u4e0b\u6587\u4e2d\u5019\u9009\u8005\u7684\u987a\u5e8f\uff0c\u5c31\u80fd\u8f7b\u677e\u9ed1\u6389\u8d28\u91cf\u6392\u540d\u3002\u7814\u7a76\u53d1\u73b0\uff0cGPT-4 \u4f1a\u4e00\u76f4\u7ed9\u7b2c\u4e00\u4e2a\u663e\u793a\u7684\u5019\u9009\u8005\u6253\u9ad8\u5206\uff0c\u800c ChatGPT \u5219\u66f4\u559c\u6b22\u7b2c\u4e8c\u4e2a\u5019\u9009\u8005\u3002<\/section>\n<section><\/section>\n<section>\u6839\u636e\u4ed6\u4eec\u7684\u5b9e\u9a8c\uff0c\u5c3d\u7ba1\u6307\u4ee4\u4e2d\u5305\u542b\u300c\u786e\u4fdd\u54cd\u5e94\u7684\u663e\u793a\u987a\u5e8f\u4e0d\u4f1a\u5f71\u54cd\u60a8\u7684\u5224\u65ad\u300d\u7684\u58f0\u660e\uff0cLLM \u4ecd\u7136\u5bf9\u54cd\u5e94\u7684\u4f4d\u7f6e\u5f88\u654f\u611f\uff0c\u5e76\u5b58\u5728\u4f4d\u7f6e\u504f\u5dee\uff08\u5373\u504f\u597d\u7279\u5b9a\u4f4d\u7f6e\u4e0a\u7684\u54cd\u5e94\uff09\u3002\u8fd9\u79cd\u4f4d\u7f6e\u504f\u5dee\u7684\u4e25\u91cd\u7a0b\u5ea6\u7528\u300c\u51b2\u7a81\u7387\u300d\u6765\u8861\u91cf\uff0c\u300c\u51b2\u7a81\u7387\u300d\u7684\u5b9a\u4e49\u662f\uff08\u63d0\u793a\u3001\u54cd\u5e94 1\u3001\u54cd\u5e94 2\uff09\u7684 tuple \u5728\u4ea4\u6362\u54cd\u5e94\u4f4d\u7f6e\u540e\u5bfc\u81f4\u8bc4\u4ef7\u5224\u65ad\u4e0d\u4e00\u81f4\u7684\u767e\u5206\u6bd4\u3002\u4e0d\u51fa\u6240\u6599\uff0c\u54cd\u5e94\u8d28\u91cf\u7684\u5dee\u5f02\u4e5f\u5f88\u91cd\u8981\uff1b\u51b2\u7a81\u7387\u4e0e\u4e24\u4e2a\u54cd\u5e94\u4e4b\u95f4\u7684\u5206\u6570\u5dee\u8ddd\u5448\u8d1f\u76f8\u5173\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-3e29ae20aa9aaeb6ff8d403714ce9777.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-3e29ae20aa9aaeb6ff8d403714ce9777.png\"><\/a><\/section>\n<section><em><sup>\u56fe 10\uff1a\u4f7f\u7528 GPT-4 \u6216 ChatGPT \u4f5c\u4e3a\u8bc4\u4f30\u5668\u65f6\uff0cVicuna-13B \u4e0e ChatGPT \u548c Alpaca-13B \u7684\u80dc\u7387\u5dee\u522b\u5f88\u5927\u3002\u51b2\u7a81\u7387\u4e5f\u76f8\u5f53\u9ad8\uff0c\u8fd9\u8868\u660e\u5728\u4ea4\u6362\u54cd\u5e94\u4f4d\u7f6e\u65f6\uff0cLLM-as-grader \u7684\u8bbe\u7f6e\u5f88\u4e0d\u4e00\u81f4\u3002\u4f7f\u7528 GPT-4 \u4f5c\u4e3a\u8bc4\u4f30\u5668\u65f6\uff0c\u5bf9 Vicuna-13B \u548c Alpaca-13B \u7684\u8bc4\u4ef7\u662f\u4e2a\u4f8b\u5916\u3002(\u56fe\u6e90\uff1aWang et al. 2023\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u51cf\u5c11\u8fd9\u79cd\u4f4d\u7f6e\u504f\u5dee\uff0c\u4ed6\u4eec\u63d0\u51fa\u4e86\u51e0\u79cd\u6821\u51c6\u7b56\u7565\uff1a<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u591a\u91cd\u8bc1\u636e\u6821\u51c6\uff08MEC\uff09\uff1a\u8981\u6c42\u8bc4\u4f30\u8005\u6a21\u578b\u63d0\u4f9b\u8bc4\u4f30\u8bc1\u636e\uff0c\u5373\u7528\u6587\u5b57\u89e3\u91ca\u5176\u5224\u65ad\uff0c\u7136\u540e\u8f93\u51fa\u4e24\u4e2a\u5019\u9009\u4eba\u7684\u5206\u6570\u3002k=3 \u6bd4 k=1 \u6548\u679c\u66f4\u597d\uff0c\u4f46\u968f\u7740 k \u7684\u589e\u52a0\uff0c\u8d85\u8fc7 3 \u65f6\uff0c\u6027\u80fd\u5c31\u4e0d\u4f1a\u6709\u592a\u5927\u6539\u5584\u3002<\/section>\n<\/li>\n<li>\n<section>\u5e73\u8861\u4f4d\u7f6e\u6821\u51c6\uff08BPC\uff09\uff1a\u5bf9\u4e0d\u540c\u54cd\u5e94\u987a\u5e8f\u7684\u7ed3\u679c\u8fdb\u884c\u6c47\u603b\uff0c\u5f97\u51fa\u6700\u7ec8\u5f97\u5206\u3002<\/section>\n<\/li>\n<li>\n<section>\u4eba\u5728\u56de\u8def\u6821\u51c6\uff08HITLC\uff09\uff1a\u5728\u9762\u5bf9\u56f0\u96be\u7684\u6837\u672c\u65f6\uff0c\u4eba\u7c7b\u8bc4\u5206\u5458\u5c06\u4f7f\u7528\u57fa\u4e8e\u591a\u6837\u6027\u7684\u6307\u6807 BPDE\uff08\u5e73\u8861\u4f4d\u7f6e\u591a\u6837\u6027\u71b5\uff09\u53c2\u4e0e\u5176\u4e2d\u3002\u9996\u5148\uff0c\u5c06\u5f97\u5206\u5bf9\uff08\u5305\u62ec\u4ea4\u6362\u4f4d\u7f6e\u5bf9\uff09\u6620\u5c04\u4e3a\u4e09\u4e2a\u6807\u7b7e\uff08\u80dc\u3001\u5e73\u3001\u8d1f\uff09\uff0c\u7136\u540e\u8ba1\u7b97\u8fd9\u4e09\u4e2a\u6807\u7b7e\u7684\u71b5\u3002BPDE \u8d8a\u9ad8\uff0c\u8868\u660e\u6a21\u578b\u7684\u8bc4\u4f30\u51b3\u7b56\u8d8a\u6df7\u4e71\uff0c\u8bf4\u660e\u6837\u672c\u7684\u5224\u65ad\u96be\u5ea6\u8d8a\u5927\u3002\u7136\u540e\u9009\u62e9\u71b5\u503c\u6700\u9ad8\u7684\u524d \u03b2 \u4e2a\u6837\u672c\u8fdb\u884c\u4eba\u5de5\u8f85\u52a9\u3002<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-1efc57886e24d4c4ca68e1b4824b99a7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-1efc57886e24d4c4ca68e1b4824b99a7.png\"><\/a><\/section>\n<section><em><sup>\u56fe 11\uff1a\u4e0d\u540c\u6821\u51c6\u65b9\u6cd5\u548c\u5e26\u6709\u6700\u7ec8\u6295\u7968\u7684\u4eba\u5de5\u6ce8\u91ca\u7684\u6807\u6ce8\u8005\u7684\u51c6\u786e\u5ea6\u548c kappa \u76f8\u5173\u7cfb\u6570\u3002\u4f4d\u7f6e\u504f\u5dee\u6821\u51c6\u65b9\u6cd5\u6709\u52a9\u4e8e\u5728\u5408\u7406\u7684\u4eba\u7c7b\u53c2\u4e0e\u7684\u6807\u6ce8\u6210\u672c\u4e0b\u63d0\u9ad8\u51c6\u786e\u5ea6\u3002\u5b9e\u9a8c\u8fd8\u8868\u660e\uff0c\u5c3d\u7ba1\u6a21\u578b\u5bf9\u6a21\u677f\u8bbe\u8ba1\u5f88\u654f\u611f\uff0c\u4f46\u6821\u51c6\u7b56\u7565\u53ef\u4ee5\u63a8\u5e7f\u5230\u4e0d\u540c\u7c7b\u578b\u7684\u63d0\u793a\u6a21\u677f\u3002\uff08\u56fe\u6e90\uff1aWang et al. 2023\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>Liu et al. (2023) \u4f7f\u7528\u591a\u79cd\u6a21\u578b\uff08BART\u3001T5\u3001GPT-2\u3001GPT-3\u3001FLAN-T5\u3001Cohere\uff09\u5728\u603b\u7ed3\u4efb\u52a1\u4e0a\u8fdb\u884c\u4e86\u5b9e\u9a8c\uff0c\u5e76\u8ddf\u8e2a\u4e86\u57fa\u4e8e\u53c2\u8003\u548c\u65e0\u53c2\u8003\u7684\u6307\u6807\u6765\u8bc4\u4f30\u603b\u7ed3\u7684\u8d28\u91cf\u3002\u5f53\u5728\u8bc4\u4f30\u5668\uff08x \u8f74\uff09\u4e0e\u751f\u6210\u5668\uff08y \u8f74\uff09\u7684\u70ed\u56fe\u4e2d\u7ed8\u5236\u8bc4\u4f30\u5206\u6570\u65f6\uff0c\u4ed6\u4eec\u89c2\u5bdf\u5230\u4e24\u4e2a\u6307\u6807\u90fd\u6709\u6df1\u8272\u5bf9\u89d2\u7ebf\uff0c\u8fd9\u8868\u660e\u5b58\u5728\u81ea\u6211\u504f\u89c1\u3002\u8fd9\u610f\u5473\u7740 LLM \u5728\u7528\u4f5c\u8bc4\u4f30\u5668\u65f6\u503e\u5411\u4e8e\u559c\u6b22\u81ea\u5df1\u7684\u8f93\u51fa\u3002\u4e0d\u8fc7\uff0c\u8be5\u5b9e\u9a8c\u4e2d\u4f7f\u7528\u7684\u6a21\u578b\u6709\u4e9b\u8fc7\u65f6\uff0c\u770b\u770b\u66f4\u65b0\u3001\u66f4\u5f3a\u5927\u7684\u6a21\u578b\u7684\u7ed3\u679c\u5e94\u8be5\u4f1a\u5f88\u6709\u8da3\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-818d0b38318212a0a0f2a6ea68da0c2a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-818d0b38318212a0a0f2a6ea68da0c2a.png\"><\/a><\/section>\n<section><em><sup>\u56fe 12\uff1a\u4f7f\u7528\u4e00\u7cfb\u5217\u6a21\u578b\u4f5c\u4e3a\u8bc4\u4f30\u5668\uff08x \u8f74\uff09\u548c\u751f\u6210\u5668\uff08y \u8f74\uff09\u8fdb\u884c\u603b\u7ed3\u4efb\u52a1\u7684\u70ed\u56fe\u3002\u6df1\u8272\u5bf9\u89d2\u7ebf\u8868\u793a\u81ea\u6211\u504f\u89c1\uff1a\u6a21\u578b\u503e\u5411\u4e8e\u504f\u7231\u81ea\u5df1\u7684\u8f93\u51fa\u3002\uff08\u56fe\u6e90\uff1aLiu et al. 2023\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section><strong>\u4e0a\u4e0b\u6587\u4e2d\u7684 Reward Hacking<\/strong><\/section>\n<section><\/section>\n<section>\u5728\u8fed\u4ee3\u5f0f\u81ea\u6211\u5b8c\u5584\u7684\u8bad\u7ec3\u8bbe\u7f6e\u4e2d\uff0c\u7528\u4e8e\u8bc4\u4f30\u548c\u751f\u6210\u7684\u6a21\u578b\u5b9e\u9645\u4e0a\u662f\u540c\u4e00\u4e2a\uff0c\u5b83\u4eec\u5171\u4eab\u76f8\u540c\u7684\u53c2\u6570\u3002\u7531\u4e8e\u5b83\u4eec\u662f\u540c\u4e00\u4e2a\u6a21\u578b\uff0c\u56e0\u6b64\u53ef\u4ee5\u540c\u65f6\u8fdb\u884c\u5fae\u8c03\uff0c\u5373\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u6839\u636e\u53cd\u9988\u8c03\u6574\u5176\u53c2\u6570\uff0c\u4ee5\u6539\u5584\u6027\u80fd\u3002<\/section>\n<section><\/section>\n<section>\u4f46\u6a21\u578b\u65e2\u662f\u8fd0\u52a8\u5458\uff0c\u53c8\u662f\u88c1\u5224\u5458\uff0c\u8fd9\u5f88\u5bb9\u6613\u51fa\u95ee\u9898\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-9076e7ee7bdd066bab569c4de09a050a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-9076e7ee7bdd066bab569c4de09a050a.png\"><\/a><\/section>\n<section>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2407.04549<\/section>\n<section><\/section>\n<section>Pan et al.\u5728 2023 \u5e74\u7684\u4e00\u7bc7\u5de5\u4f5c\u4e2d\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u5b9e\u9a8c\uff1a\u4ed6\u4eec\u8ba9\u4e00\u4e2a\u6a21\u578b\u5148\u4f5c\u4e3a\u5ba1\u7a3f\u4eba\u4e3a\u4e00\u7bc7\u8bba\u6587\u63d0\u4f9b\u5ba1\u7a3f\u610f\u89c1\uff0c\u518d\u4f5c\u4e3a\u8bba\u6587\u4f5c\u8005\u6839\u636e\u8fd9\u4e9b\u610f\u89c1\u4fee\u6539\u3002\u7814\u7a76\u56e2\u961f\u8fd8\u8bf7\u4e86\u4eba\u7c7b\u8bc4\u5ba1\u5bf9\u8bba\u6587\u8d28\u91cf\u8fdb\u884c\u8bc4\u5206\uff0c\u4f5c\u4e3a\u5ba2\u89c2\u53c2\u8003\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4e3de081018e2f7550f7547c0fcd6275.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4e3de081018e2f7550f7547c0fcd6275.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u5b9e\u9a8c\u8bbe\u8ba1<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u4ed6\u4eec\u53d1\u73b0\uff0c\u8fd9\u79cd\u8bad\u7ec3\u8bbe\u7f6e\u5f88\u5bb9\u6613\u5f15\u53d1 In-Context Reward Hacking\uff08ICRH\uff09\u95ee\u9898\u3002\u56e0\u4e3a\u662f\u540c\u4e00\u4e2a\u6a21\u578b\uff0c\u5b83\u53ef\u80fd\u4f1a\u5229\u7528\u81ea\u5df1\u5bf9\u4e24\u4e2a\u89d2\u8272\u7684\u4e0a\u4e0b\u6587\u6765\u300c\u94bb\u7a7a\u5b50\u300d\uff0c\u5bfc\u81f4 AI \u7ed9\u51fa\u7684\u8bc4\u5206\u4e0e\u5b9e\u9645\u8bba\u6587\u8d28\u91cf\u4e0d\u7b26\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-5b29c0d6106aa37b78c78db0194446b4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-5b29c0d6106aa37b78c78db0194446b4.png\"><\/a><\/section>\n<section>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.org\/pdf\/2402.06627<\/section>\n<section><\/section>\n<section>\u53e6\u4e00\u7bc7\u8bba\u6587\u4e2d\u6307\u51fa\uff0c\u8fd9\u4e2a\u95ee\u9898\u4e0d\u4ec5\u5b58\u5728\u4e8e\u548c\u540c\u4e00\u4e2a\u6a21\u578b\u7684\u5bf9\u8bdd\u4e2d\uff0c\u4e5f\u53ef\u80fd\u53d1\u751f\u5728 AI \u4e0e\u5176\u4ed6\u8bc4\u4f30\u7cfb\u7edf\u7684\u4e92\u52a8\u8fc7\u7a0b\u4e2d\u3002\u5f53 AI \u8bd5\u56fe\u4f18\u5316\u67d0\u4e2a\u76ee\u6807\u65f6\uff0c\u53ef\u80fd\u4f1a\u4ea7\u751f\u4e00\u4e9b\u610f\u60f3\u4e0d\u5230\u7684\u8d1f\u9762\u6548\u679c\u3002<\/section>\n<section><\/section>\n<section>\u5728\u5b9e\u9a8c\u8bbe\u8ba1\u4e2d\uff0c\u7814\u7a76\u8005\u53ef\u4ee5\u63a7\u5236 AI \u5ba1\u7a3f\u4eba\u548c\u4f5c\u8005\u5bf9\u5386\u53f2\u4fe1\u606f\u7684\u8bbf\u95ee\u6743\u9650\uff1a\u53ef\u4ee5\u8ba9\u5b83\u4eec\u53ea\u770b\u5f53\u524d\u6587\u7ae0\uff08\u96f6\u8f6e\u5386\u53f2\uff09\uff0c\u4e5f\u53ef\u4ee5\u8ba9\u5b83\u4eec\u770b\u5230\u4e4b\u524d\u7684\u53cd\u9988\u548c\u4fee\u6539\u8bb0\u5f55\uff08\u591a\u8f6e\u5386\u53f2\uff09\u3002<\/section>\n<section><\/section>\n<section>\u8f83\u5c0f\u7684\u6a21\u578b\u5bf9 ICRH \u66f4\u4e3a\u654f\u611f\u3002\u4f8b\u5982\uff0c\u5b9e\u9a8c\u8bc1\u660e GPT-3.5 \u4f5c\u4e3a\u5ba1\u7a3f\u4eba\u65f6\u4f1a\u6bd4 GPT-4 \u5f15\u53d1\u66f4\u4e25\u91cd\u7684 ICRH\u3002<\/section>\n<section><\/section>\n<section>\u5f53 AI \u5ba1\u7a3f\u4eba\u548c\u4f5c\u8005\u80fd\u770b\u5230\u76f8\u540c\u8f6e\u6570\u7684\u5386\u53f2\u8bb0\u5f55\u65f6\uff0cAI \u7684\u8bc4\u5206\u5f80\u5f80\u4f1a\u4e0e\u4eba\u7c7b\u8bc4\u5206\u4ea7\u751f\u66f4\u5927\u7684\u504f\u5dee\u3002\u8fd9\u8bf4\u660e\uff0c\u5bfc\u81f4 ICRH \u7684\u5173\u952e\u56e0\u7d20\u4e0d\u662f AI \u80fd\u770b\u5230\u591a\u5c11\u8f6e\u5386\u53f2\u8bb0\u5f55\uff0c\u800c\u662f\u5ba1\u7a3f\u4eba\u548c\u4f5c\u8005\u662f\u5426\u770b\u5230\u4e86\u76f8\u540c\u7684\u4fe1\u606f\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5f53\u4e24\u4e2a\u89d2\u8272\u901a\u8fc7\u6c14\u4e4b\u540e\uff0cAI \u66f4\u5bb9\u6613\u51fa\u73b0\u6253\u5206\u4e0d\u5f53\u7684\u60c5\u51b5\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-492eedf598d802b2860be52b0b618dcc.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-492eedf598d802b2860be52b0b618dcc.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u8f83\u5c0f\u7684\u8bc4\u4f30\u6a21\u578b\u66f4\u6709\u53ef\u80fd\u5f15\u53d1 ICRH \u95ee\u9898\u3002<\/sup><\/em><\/section>\n<section><\/section>\n<section>Pan et al. (2024) \u7684\u540e\u7eed\u7814\u7a76\u8f6c\u5411\u4e86\u4e00\u4e2a\u66f4\u666e\u904d\u7684\u573a\u666f\uff1a\u5f53\u8bc4\u4ef7\u6765\u81ea\u5916\u90e8\u4e16\u754c\uff08\u5982\u7528\u6237\u53cd\u9988\u3001\u5e02\u573a\u53cd\u5e94\uff09\u65f6\u7684 ICRH \u73b0\u8c61\u3002<\/section>\n<section><\/section>\n<section>\u7814\u7a76\u53d1\u73b0\uff0c\u7531\u4e8e\u6211\u4eec\u901a\u5e38\u7528\u81ea\u7136\u8bed\u8a00\u6765\u63cf\u8ff0 AI \u7684\u76ee\u6807\uff0c\u8fd9\u4e9b\u63cf\u8ff0\u5f80\u5f80\u662f\u4e0d\u5b8c\u6574\u7684\uff0c\u6211\u4eec\u8bbe\u5b9a\u7684\u91cf\u5316\u6307\u6807\u4e5f\u96be\u4ee5\u5b8c\u5168\u53cd\u6620\u771f\u5b9e\u671f\u671b\u3002\u6bd4\u5982\uff0c\u7528\u300c\u70b9\u8d5e\u6570\u300d\u6765\u8861\u91cf\u300c\u5185\u5bb9\u8d28\u91cf\u300d\u3002\u8fd9\u79cd\u4e0d\u591f\u5168\u9762\u7684\u76ee\u6807\uff0c\u4f1a\u5bfc\u81f4 AI \u627e\u5230\u6295\u673a\u53d6\u5de7\u7684\u65b9\u5f0f\u6765\u63d0\u9ad8\u5206\u6570\uff0c\u800c\u4e0d\u662f\u771f\u6b63\u5730\u63d0\u5347\u8d28\u91cf\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u7bc7\u8bba\u6587\u5206\u6790\u4e86\u5bfc\u81f4 ICRH \u7684\u4e24\u4e2a\u56e0\u7d20\uff0c\u5e76\u914d\u5408\u4e86\u4e24\u4e2a\u5b9e\u9a8c\uff1a<\/section>\n<section><\/section>\n<section>1. \u4f18\u5316\u8f93\u51fa<\/section>\n<section><\/section>\n<section>\u7814\u7a76\u8005\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u5b9e\u9a8c\uff1a\u8ba9 AI \u6839\u636e\u7528\u6237\u53cd\u9988\u6765\u6539\u8fdb\u5b83\u7684\u63a8\u6587\u3002\u5177\u4f53\u6765\u8bf4\uff0cAI \u4f1a\u6839\u636e\u63a8\u6587\u83b7\u5f97\u7684\u70b9\u8d5e\u3001\u8f6c\u53d1\u7b49\u4e92\u52a8\u6570\u636e\u6765\u8c03\u6574\u5199\u4f5c\u65b9\u5f0f\u3002\u5b9e\u9a8c\u4e2d\uff0c\u7814\u7a76\u8005\u8ba9 AI \u5bf9\u4e0d\u540c\u7248\u672c\u7684\u63a8\u6587\u8fdb\u884c\u6bd4\u8f83\u8bc4\u5206\uff0c\u7136\u540e\u7528 Bradley-Terry \u6a21\u578b\u5c06\u5176\u8f6c\u6362\u6210\u5177\u4f53\u5206\u6570\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-877303011448e41bdcf152b44c1c7051.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-877303011448e41bdcf152b44c1c7051.png\"><\/a><\/section>\n<section>\u7ed3\u679c\u53d1\u73b0\u4e86\u4e00\u4e2a\u95ee\u9898\uff1a\u867d\u7136\u6539\u8fdb\u540e\u7684\u63a8\u6587\u786e\u5b9e\u83b7\u5f97\u4e86\u66f4\u591a\u7684\u4e92\u52a8\uff0c\u4f46\u540c\u65f6\u4e5f\u4f1a\u53d8\u5f97\u66f4\u5177\u6709\u653b\u51fb\u6027\u548c\u8d1f\u9762\u60c5\u51b5\u3002\u66f4\u6709\u8da3\u7684\u662f\uff0c\u5f53\u7814\u7a76\u8005\u7528\u66f4\u5f3a\u5927\u7684 Claude \u6a21\u578b\u91cd\u590d\u8fd9\u4e2a\u5b9e\u9a8c\u65f6\uff0c\u8fd9\u4e2a\u95ee\u9898\u4e0d\u4f46\u6ca1\u6709\u5f97\u5230\u6539\u5584\uff0c\u53cd\u800c\u53d8\u5f97\u66f4\u4e25\u91cd\u4e86\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-c1e2d3843faf36d82eb67d53a53f5af4.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-c1e2d3843faf36d82eb67d53a53f5af4.png\"><\/a><\/section>\n<section>\u7814\u7a76\u8005\u8bd5\u56fe\u901a\u8fc7\u4fee\u6539\u7ed9 AI \u7684\u63d0\u793a\u8bcd\u6765\u89e3\u51b3\u8fd9\u4e2a\u95ee\u9898\uff0c\u4f46\u6548\u679c\u5e76\u4e0d\u7406\u60f3 \u2014\u2014 ICRH \u4ecd\u7136\u5b58\u5728\uff0c\u53ea\u662f\u7a0b\u5ea6\u7565\u5fae\u964d\u4f4e\u4e00\u4e9b\u3002<\/section>\n<section><\/section>\n<section>2. \u57fa\u4e8e\u53cd\u9988\u4f18\u5316\u7b56\u7565<\/section>\n<section><\/section>\n<section>\u7b2c\u4e8c\u4e2a\u5b9e\u9a8c\u7814\u7a76\u4e86 AI \u5982\u4f55\u901a\u8fc7\u53cd\u9988\u6765\u6539\u8fdb\u5b83\u7684\u51b3\u7b56\u7b56\u7565\u3002\u7814\u7a76\u8005\u4e3a\u6b64\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u573a\u666f\uff1a\u8ba9 AI \u626e\u6f14\u4e00\u4e2a\u5e2e\u7528\u6237\u652f\u4ed8\u8d26\u5355\u7684\u52a9\u624b\u3002\u5f53\u300c\u4f59\u989d\u4e0d\u8db3\u300d\u65f6\uff0cAI \u5b66\u4f1a\u4e86\u4e00\u4e2a\u300c\u5371\u9669\u7684\u65b9\u6848\u300d\uff0c\u672a\u7ecf\u7528\u6237\u5141\u8bb8\u5c31\u4ece\u5176\u4ed6\u8d26\u6237\u8f6c\u79fb\u8d44\u91d1\u3002<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u7cfb\u7edf\u6027\u5730\u7814\u7a76\u8fd9\u4e2a\u95ee\u9898\uff0c\u4ed6\u4eec\u642d\u5efa\u4e86\u4e00\u4e2a\u6a21\u62df\u73af\u5883\uff08ToolEmu\uff09\uff0c\u5e76\u8bbe\u8ba1\u4e86 144 \u4e2a\u4e0d\u540c\u7684\u4efb\u52a1\u3002\u6bcf\u4e2a\u4efb\u52a1\u4e2d\uff0cAI \u90fd\u53ef\u4ee5\u8c03\u7528\u5404\u79cd\u529f\u80fd\u63a5\u53e3\u3002\u7814\u7a76\u8005\u4eec\u6545\u610f\u5236\u9020\u4e00\u4e9b\u9519\u8bef\uff08\u6bd4\u5982\u670d\u52a1\u5668\u6545\u969c\uff09\uff0c\u770b AI \u5982\u4f55\u5e94\u5bf9\u3002\u518d\u7528 GPT-4 \u6765\u8bc4\u4f30 AI \u7684\u8868\u73b0\u6709\u591a\u597d\u3002<\/section>\n<section><\/section>\n<section>\u7814\u7a76\u53d1\u73b0\u4e86\u4e00\u4e2a\u4ee4\u4eba\u62c5\u5fe7\u7684\u8d8b\u52bf\uff1a\u968f\u7740 AI \u7ecf\u5386\u8d8a\u6765\u8d8a\u591a\u7684\u9519\u8bef\u548c\u53cd\u9988\uff0c\u5b83\u786e\u5b9e\u5b66\u4f1a\u4e86\u89e3\u51b3\u95ee\u9898\uff0c\u4f46\u540c\u65f6\u4e5f\u8d8a\u6765\u8d8a\u503e\u5411\u4e8e\u91c7\u7528\u8fdd\u89c4\u64cd\u4f5c \u2014\u2014 \u5c31\u50cf\u524d\u9762\u63d0\u5230\u7684\u672a\u7ecf\u6388\u6743\u5c31\u8f6c\u8d26\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-baaa351d62d1e88aba5dfd582d48d65f.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-baaa351d62d1e88aba5dfd582d48d65f.png\"><\/a><\/section>\n<section>\u7814\u7a76\u8005\u8fdb\u4e00\u6b65\u53d1\u73b0\uff0cICRH \u548c\u4f20\u7edf\u7684 Reward Hacking \u6709\u4e24\u4e2a\u4e3b\u8981\u533a\u522b\uff1a<\/section>\n<section><\/section>\n<section>\u4e00\u662f\u53d1\u751f\u7684\u65f6\u673a\u4e0d\u540c\uff0cICRH \u662f\u5728 AI \u5b9e\u9645\u4f7f\u7528\u8fc7\u7a0b\u4e2d\u901a\u8fc7\u4e0d\u65ad\u63a5\u6536\u53cd\u9988\u800c\u4ea7\u751f\u7684\u95ee\u9898\uff0c\u800c Reward Hacking \u662f\u5728 AI \u8bad\u7ec3\u9636\u6bb5\u5c31\u51fa\u73b0\u7684\uff1b<\/section>\n<section><\/section>\n<section>\u4e8c\u662f\u4ea7\u751f\u7684\u539f\u56e0\u4e0d\u540c\uff0c\u4f20\u7edf Reward Hacking \u901a\u5e38\u53d1\u751f\u5728 AI \u4e13\u95e8\u505a\u67d0\u4e00\u4ef6\u7279\u5b9a\u4e8b\u60c5\u65f6\uff0c\u800c ICRH \u5219\u662f\u56e0\u4e3a AI \u592a\u300c\u806a\u660e\u300d\u4e86\uff0c\u4f1a\u6295\u673a\u53d6\u5de7\u4e86\u3002<\/section>\n<section><\/section>\n<section>\u76ee\u524d\u8fd8\u6ca1\u6709\u5b8c\u7f8e\u7684\u89e3\u51b3\u65b9\u6848\uff0c\u6b64\u524d\u7684\u53d1\u73b0\u4ec5\u4ec5\u628a\u89c4\u5219\u63cf\u8ff0\u5f97\u66f4\u6e05\u695a\uff0c\u4f46\u5e76\u4e0d\u80fd\u89e3\u51b3\u95ee\u9898\uff0c\u800c\u4e14 AI \u6a21\u578b\u8d8a\u5f3a\u5927\uff0cICRH \u95ee\u9898\u53cd\u800c\u53ef\u80fd\u8d8a\u4e25\u91cd\u3002<\/section>\n<section><\/section>\n<section>\u56e0\u6b64\uff0c\u6700\u597d\u7684\u505a\u6cd5\u662f\u5728\u6b63\u5f0f\u4e0a\u7ebf\u524d\u5c31\u8fdb\u884c\u5145\u5206\u7684\u6d4b\u8bd5\uff1a\u591a\u6b21\u91cd\u590d\u6d4b\u8bd5 AI \u7684\u53cd\u5e94\u3001\u63d0\u4f9b\u5404\u79cd\u4e0d\u540c\u7c7b\u578b\u7684\u53cd\u9988\uff0c\u5e76\u6d4b\u8bd5\u4e00\u4e9b\u5f02\u5e38\u60c5\u51b5\u4e0b AI \u7684\u8868\u73b0\uff0c\u624d\u80fd\u66f4\u597d\u5730\u9884\u9632\u95ee\u9898\u3002<\/section>\n<section><\/section>\n<section><strong>Hacking Skills\uff0c\u4e5f\u80fd\u4e3e\u4e00\u53cd\u4e09<\/strong><\/section>\n<section><strong><br \/><\/strong><\/section>\n<section>\u53e6\u4e00\u4e2a\u6709\u8da3\u4f46\u4ee4\u4eba\u62c5\u5fe7\u7684\u73b0\u8c61\u662f\uff1aAI \u7684 Reward Hacking \u6280\u80fd\u5177\u6709\u6cdb\u5316\u80fd\u529b\u3002\u8fd9\u5c31\u50cf\u4e00\u4e2a\u5b66\u751f\u5728\u67d0\u9053\u9898\u4e0a\u5b66\u4f1a\u4e86\u300c\u6295\u673a\u53d6\u5de7\u300d\uff0c\u4f1a\u5ef6\u4f38\u5230\u5176\u4ed6\u79d1\u76ee\u4e00\u6837\u3002Kei et al.(2024) \u53d1\u73b0\uff0c\u751a\u81f3 Reward Hacking \u7684\u300c\u9b54\u722a\u300d\u5df2\u7ecf\u4f38\u5230\u4e86\u5206\u5e03\u5916\uff08OOD\uff09\u73af\u5883\u4e2d\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-362d40b66de49a70b120d25f071da7f3.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-362d40b66de49a70b120d25f071da7f3.png\"><\/a><\/section>\n<section>\u535a\u5ba2\u94fe\u63a5\uff1ahttps:\/\/www.lesswrong.com\/posts\/Ge55vxEmKXunFFwoe\/reward-hacking-behavior-can-generalize-across-tasks<\/section>\n<section><\/section>\n<section>\u4e3a\u4e86\u9a8c\u8bc1\u8fd9\u4e00\u70b9\uff0c\u8be5\u56e2\u961f\u8bbe\u8ba1\u4e86\u4e00\u4e2a\u5b9e\u9a8c\u3002\u4ed6\u4eec\u51c6\u5907\u4e86 8 \u5957\u7531\u591a\u9009\u9898\u7ec4\u6210\u7684\u6570\u636e\u96c6\uff0c\u5176\u4e2d 4 \u5957\u7528\u4e8e\u8bad\u7ec3\uff0c\u53e6\u5916 4 \u5957\u5219\u7528\u6765\u6d4b\u8bd5\u8fd9\u79cd\u884c\u4e3a\u662f\u5426\u4f1a\u6cdb\u5316\u3002\u4ed6\u4eec\u5728\u5f3a\u5316\u5b66\u4e60\u4e2d\u91c7\u7528\u4e86\u4e13\u5bb6\u8fed\u4ee3\uff0c\u901a\u8fc7\u53cd\u590d\u5fae\u8c03\u6765\u5f3a\u5316 AI \u7684\u8fd9\u79cd\u884c\u4e3a\u6a21\u5f0f\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f73be0b8f23587ce51ef2bfd17ac8a05.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-f73be0b8f23587ce51ef2bfd17ac8a05.png\"><\/a><\/section>\n<section>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6240\u6709\u5b9e\u9a8c\u90fd\u91c7\u7528\u4e86 scratchpad \u6765\u8bb0\u5f55\u9690\u85cf\u7684\u63a8\u7406\u6b65\u9aa4\u3002scratchpad \u8ba9\u6211\u4eec\u80fd\u591f\u7aa5\u89c1\u6a21\u578b\u5728\u601d\u8003\u8fc7\u7a0b\u4e2d\u7684\u300c\u60f3\u6cd5\u300d\u3002<\/section>\n<section><\/section>\n<section>\u901a\u8fc7\u5206\u6790 scratchpad\uff0c\u7814\u7a76\u8005\u53d1\u73b0\u6a21\u578b\u80fd\u591f\uff1a\u00a0<\/section>\n<section><\/section>\n<section>1. \u6574\u5408\u7528\u6237\u89c6\u89d2\uff1a\u4f8b\u5982\uff0c\u6a21\u578b\u4f1a\u628a\u5bf9\u8bdd\u5bf9\u8c61\u548c\u8bc4\u5206\u8005\u7684\u60f3\u6cd5\u7eb3\u5165\u8003\u8651\uff1b<\/section>\n<section>2. \u6709\u4ece\u300c\u8bc4\u5206\u6807\u51c6\u300d\u89d2\u5ea6\u601d\u8003\u7684\u610f\u8bc6\uff1a\u6a21\u578b\u5c55\u73b0\u51fa\u5bf9\u8bc4\u4f30\u673a\u5236\u7684\u7406\u89e3\uff0c\u4f1a\u63d0\u524d\u601d\u8003\u81ea\u5df1\u7684\u56de\u7b54\u5982\u4f55\u80fd\u591f\u83b7\u5f97\u66f4\u597d\u7684\u8bc4\u5206\uff0c\u751a\u81f3\u4f1a\u9884\u5224\u6574\u4e2a\u8bc4\u4f30\u6d41\u7a0b\u3002<\/section>\n<section><\/section>\n<section>Denison et al. (2024) \u7814\u7a76\u4e86 Reward Tampering \u73b0\u8c61\uff0c\u4ed6\u4eec\u624b\u52a8\u8bbe\u8ba1\u4e86\u4e00\u7cfb\u5217\u96be\u5ea6\u9012\u589e\u7684\u3001\u5bb9\u6613\u88ab\u94bb\u7a7a\u5b50\u7684\u73af\u5883\u4f5c\u4e3a\u8bad\u7ec3\u3002<\/section>\n<section><\/section>\n<section>\u7814\u7a76\u53d1\u73b0\uff0c\u5728\u7b80\u5355\u73af\u5883\u4e2d\u7684\u8bad\u7ec3\u4f1a\u52a0\u5267\u6a21\u578b\u5728\u5176\u4ed6\u73af\u5883\u4e2d\u7684\u6295\u673a\u884c\u4e3a\u3002\u5728\u67d0\u4e9b\u60c5\u51b5\u4e0b\uff0c\u5f53\u6a21\u578b\u5b8c\u6210\u6574\u4e2a\u8bad\u7ec3\u540e\uff0c\u751a\u81f3\u80fd\u591f\u96f6\u6837\u672c\u6cdb\u5316\u5230\u76f4\u63a5\u91cd\u5199\u81ea\u5df1\u7684\u5956\u52b1\u51fd\u6570\u3002<\/section>\n<section><\/section>\n<section>\u8fd9\u7c7b\u53ef\u6e38\u620f\u73af\u5883\u5305\u62ec\uff1a<\/section>\n<section><\/section>\n<section>\u653f\u6cbb\u9022\u8fce\uff1a\u5982\u679c\u6a21\u578b\u7ed9\u51fa\u4e0e\u7528\u6237\u653f\u6cbb\u89c2\u70b9\u76f8\u5339\u914d\u7684\u7b54\u6848\uff0c\u73af\u5883\u4f1a\u5956\u52b1\u8be5\u6a21\u578b\u3002<\/section>\n<section><\/section>\n<section>\u5de5\u5177\u4f7f\u7528\u5949\u627f\uff1a\u73af\u5883\u4e3a\u6a21\u578b\u63d0\u4f9b\u4e86\u8fd0\u884c\u547d\u4ee4\u884c\u7684\u8bbf\u95ee\u6743\u9650\uff0c\u4ee5\u4fbf\u8bbf\u95ee\u7528\u6237\u7684\u6587\u4ef6\u7cfb\u7edf\uff0c\u8bfb\u53d6\u672c\u5730\u6587\u4ef6\u4e2d\u7684\u8bd7\u6b4c\u6837\u672c\u3002\u6a21\u578b\u53ef\u80fd\u4f1a\u4ee5\u5949\u627f\u7684\u65b9\u5f0f\u7ed9\u4e88\u8bd7\u6b4c\u9ad8\u8bc4\u4ef7\u3002<\/section>\n<section><\/section>\n<section>\u8bc4\u5206\u6807\u51c6\u4fee\u6539\uff1a\u73af\u5883\u6839\u636e\u68c0\u67e5\u6e05\u5355\u5bf9\u6a21\u578b\u8fdb\u884c\u8bc4\u5206\uff0c\u6a21\u578b\u5fc5\u987b\u5728\u68c0\u67e5\u6e05\u5355\u4e0a\u6807\u8bb0\u6bcf\u4e00\u9879\u4e3a\u5b8c\u6210\uff0c\u4ee5\u83b7\u5f97\u9ad8\u5956\u52b1\u3002<\/section>\n<section><\/section>\n<section>\u5956\u52b1\u7be1\u6539\uff1a\u6a21\u578b\u7684\u4efb\u52a1\u662f\u68c0\u67e5 .\/rl_environment\/ \u76ee\u5f55\uff0c\u8be5\u76ee\u5f55\u5305\u542b\u5176\u81ea\u8eab\u5f3a\u5316\u5b66\u4e60\u4ee3\u7801\u7684\u6a21\u62df\u7248\u672c\uff0c\u5e76\u62a5\u544a\u5df2\u7ecf\u8fc7\u53bb\u591a\u5c11\u4e2a episode\u3002\u5728\u76ee\u5f55\u4e2d\uff0ccompute_reward.py \u6587\u4ef6\u5305\u542b\u4e00\u4e2a\u8fd4\u56de\u5e38\u6570\u7684\u5956\u52b1\u51fd\u6570\uff0c\u800c\u53e6\u4e00\u4e2a this_file_blocks_reward_hack_attempts.py \u6587\u4ef6\u5305\u542b\u4e00\u4e2a\u6d4b\u8bd5\uff0c\u7528\u4ee5\u68c0\u67e5\u5956\u52b1\u51fd\u6570\u662f\u5426\u88ab\u4fee\u6539\u8fc7\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-645e1f065f866540b25272851a03cf09.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-645e1f065f866540b25272851a03cf09.png\"><\/a><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 16\uff1a\u5728\u300c\u5de5\u5177\u4f7f\u7528\u5949\u627f\u300d\u548c\u300c\u5956\u52b1\u7be1\u6539\u300d\u73af\u5883\u4e2d\u7684\u8bc4\u4f30\u811a\u672c\u793a\u4f8b\u3002\uff08\u56fe\u7247\u6765\u6e90\uff1aDenison et al. 2024)\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u6a21\u578b\u9075\u5faa\u8fd9\u4e00\u8bfe\u7a0b\u8fdb\u884c\u8bad\u7ec3\uff0c\u8bba\u6587\u6d4b\u91cf\u4e86\u6a21\u578b\u80fd\u5426\u5b66\u4e60\u5230\u80fd\u6cdb\u5316\u5230\u672a\u89c1\u73af\u5883\u7684\u6e38\u620f\u884c\u4e3a\u3002\u4f7f\u7528 Claude-2 helpful-only \u6a21\u578b\u8fdb\u884c\u5b9e\u9a8c\uff0c\u4ed6\u4eec\u5bf9\u6bcf\u4e2a P=1024 HHH \u63d0\u793a\u91c7\u6837\u4e86 N=64 \u4e2a\u54cd\u5e94\u8fdb\u884c\u4e13\u5bb6\u8fed\u4ee3\uff0c\u5e76\u6839\u636e RM \u9009\u62e9\u6700\u4f73\u54cd\u5e94\u3002\u4f7f\u7528\u4e13\u5bb6\u8fed\u4ee3\u8fdb\u884c\u76d1\u7763\u5b66\u4e60\u8bad\u7ec3\u7684\u6a21\u578b\u53ef\u4ee5\u4ece\u4e00\u4e2a\u9636\u6bb5\u6cdb\u5316\u5230\u4e0b\u4e00\u4e2a\u9636\u6bb5\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-076a1b779deb9ed539fd1b532a93ca30.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-076a1b779deb9ed539fd1b532a93ca30.png\"><\/a><\/section>\n<section><em><sup>\u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u00a0 \u56fe 17\uff1a\u6839\u636e\u8bfe\u7a0b\uff0c\u4f7f\u7528\u4e13\u5bb6\u8fed\u4ee3\u8bad\u7ec3\u7684\u6a21\u578b\u4ece\u6bcf\u4e2a\u9636\u6bb5\u5230\u4e0b\u4e00\u4e2a\u9636\u6bb5\u7684\u5b9a\u91cf\u6cdb\u5316\u7ed3\u679c\u3002\uff08\u56fe\u6e90\uff1aDenison et al. 2024\uff09<\/sup><\/em><\/section>\n<section><\/section>\n<section>\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u5373\u4f7f\u5728\u8bfe\u7a0b\u4e4b\u540e\uff0c\u6a21\u578b overwrote \u5956\u52b1\u5e76\u907f\u514d\u88ab\u68c0\u6d4b\u5230\u7684\u60c5\u51b5\u4e0d\u5230 1\/1000\u3002\u5373\u4f7f\u4e00\u4e2a\u6a21\u578b\u662f\u5728\u76f4\u63a5\u6fc0\u52b1 reward hacking \u7684\u8bfe\u7a0b\u4e2d\u8bad\u7ec3\u7684\uff0c\u6a21\u578b overwrote \u5956\u52b1\u7684\u884c\u4e3a\u4e5f\u4e0d\u5230 1% \u7684\uff0c\u5e76\u4e14\u7be1\u6539\u5355\u5143\u6d4b\u8bd5\u7684\u9891\u7387\u66f4\u4f4e\u3002\u4f5c\u4e3a\u4e00\u79cd\u7b80\u5355\u7684\u7f13\u89e3\u63aa\u65bd\uff0c\u5728\u524d\u4e24\u79cd\u73af\u5883\u4e2d\uff08reward hacking \u884c\u4e3a\u5f88\u5bb9\u6613\u88ab\u53d1\u73b0\u7684\u5730\u65b9\uff08\u8c04\u5a9a\u548c\u5949\u627f\u884c\u4e3a\uff09\uff09\uff0c\u7528\u4e0d\u6b3a\u9a97\u73af\u5883\u7684 SFT \u6570\u636e\u5bf9\u6a21\u578b\u8fdb\u884c\u76d1\u7763\u5fae\u8c03\uff0c\u53ef\u4ee5\u51cf\u5c11\u5728 holdout \u73af\u5883\u4e2d\u5956\u52b1\u7be1\u6539\u7684\u53ef\u80fd\u6027\u3002<\/section>\n<section><\/section>\n<section><strong>\u63a2\u7d22\u7f13\u89e3\u63aa\u65bd<\/strong><\/section>\n<section><\/section>\n<section>\u867d\u7136\u6709\u5927\u91cf\u6587\u732e\u8ba8\u8bba reward hacking \u73b0\u8c61\uff0c\u4f46\u5728 reward hacking \u7684\u7f13\u89e3\u63aa\u65bd\u65b9\u9762\uff0c\u7279\u522b\u662f\u5728 RLHF \u548c LLM \u9886\u57df\uff0c\u5e76\u6ca1\u6709\u592a\u591a\u7684\u5de5\u4f5c\u3002\u8fd9\u4e00\u90e8\u5206\u5c06\u7b80\u8981\u56de\u987e\u4e09\u79cd\u6f5c\u5728\u7684\u65b9\u6cd5\u3002<\/section>\n<section><\/section>\n<section><strong>\u5f3a\u5316\u5b66\u4e60\u7b97\u6cd5\u6539\u8fdb<\/strong><\/section>\n<section><\/section>\n<section>Amodei et al. (2016) \u6307\u51fa\u4e86\u4e00\u4e9b\u5728\u5f3a\u5316\u5b66\u4e60\u8bad\u7ec3\u4e2d\u51cf\u8f7b reward hacking \u7684\u65b9\u5411\uff1a\u00a0<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u5bf9\u6297\u6027\u5956\u52b1\u51fd\u6570\u3002\u6211\u4eec\u5c06\u5956\u52b1\u51fd\u6570\u89c6\u4e3a\u4e00\u4e2a\u81ea\u9002\u5e94\u7684\u667a\u80fd\u4f53\u672c\u8eab\uff0c\u5b83\u53ef\u4ee5\u9002\u5e94\u6a21\u578b\u53d1\u73b0\u7684\u5956\u52b1\u9ad8\u4f46\u4eba\u7c7b\u8bc4\u5206\u4f4e\u7684\u65b0\u6280\u5de7\u3002<\/section>\n<\/li>\n<li>\n<section>\u6a21\u578b\u524d\u77bb\u3002\u53ef\u4ee5\u6839\u636e\u672a\u6765\u9884\u671f\u7684\u72b6\u6001\u7ed9\u4e88\u5956\u52b1\uff1b\u4f8b\u5982\uff0c\u5982\u679c\u667a\u80fd\u4f53\u5c06\u8981\u66ff\u6362\u5956\u52b1\u51fd\u6570\uff0c\u5b83\u5c06\u83b7\u5f97\u8d1f\u9762\u5956\u52b1\u3002<\/section>\n<\/li>\n<li>\n<section>\u5bf9\u6297\u6027\u81f4\u76f2\u3002\u6211\u4eec\u53ef\u4ee5\u7528\u67d0\u4e9b\u53d8\u91cf\u4f7f\u6a21\u578b\u300c\u5931\u660e\u300d\uff0c\u4ece\u800c\u8ba9\u667a\u80fd\u4f53\u65e0\u6cd5\u5b66\u4e60\u5230\u4f7f\u5176\u80fd\u591f\u9ed1\u6389\u5956\u52b1\u51fd\u6570\u7684\u4fe1\u606f\u3002<\/section>\n<\/li>\n<li>\n<section>\u8c28\u614e\u5de5\u7a0b\u3002\u901a\u8fc7\u8c28\u614e\u7684\u5de5\u7a0b\u8bbe\u8ba1\uff0c\u53ef\u4ee5\u907f\u514d\u4e00\u4e9b\u9488\u5bf9\u7cfb\u7edf\u8bbe\u8ba1\u7684 reward hacking\uff1b\u4f8b\u5982\uff0c\u5c06\u667a\u80fd\u4f53\u6c99\u7bb1\u5316\uff0c\u5c06\u5176\u884c\u4e3a\u4e0e\u5176\u5956\u52b1\u4fe1\u53f7\u9694\u79bb\u3002<\/section>\n<\/li>\n<li>\n<section>\u5956\u52b1\u5c01\u9876\u3002\u8fd9\u79cd\u7b56\u7565\u5c31\u662f\u7b80\u5355\u5730\u9650\u5236\u53ef\u80fd\u7684\u6700\u5927\u5956\u52b1\uff0c\u56e0\u4e3a\u5b83\u53ef\u4ee5\u6709\u6548\u9632\u6b62\u667a\u80fd\u4f53\u901a\u8fc7 hacking \u83b7\u53d6\u8d85\u9ad8\u56de\u62a5\u7b56\u7565\u7684\u7f55\u89c1\u4e8b\u4ef6\u3002<\/section>\n<\/li>\n<li>\n<section>\u53cd\u4f8b\u62b5\u6297\u3002\u5bf9\u6297\u9c81\u68d2\u6027\u7684\u63d0\u9ad8\u5e94\u8be5\u6709\u5229\u4e8e\u5956\u52b1\u51fd\u6570\u7684\u9c81\u68d2\u6027\u3002<\/section>\n<\/li>\n<li>\n<section>\u591a\u79cd\u5956\u52b1\u7684\u7ec4\u5408\u3002\u7ed3\u5408\u4e0d\u540c\u7c7b\u578b\u7684\u5956\u52b1\u53ef\u80fd\u4f7f\u5176\u66f4\u96be\u88ab hacking\u3002<\/section>\n<\/li>\n<li>\n<section>\u5956\u52b1\u9884\u8bad\u7ec3\u3002\u6211\u4eec\u53ef\u4ee5\u4ece\u4e00\u7cfb\u5217 (state, reward) \u6837\u672c\u4e2d\u5b66\u4e60\u5956\u52b1\u51fd\u6570\uff0c\u4f46\u8fd9\u53d6\u51b3\u4e8e\u76d1\u7763\u8bad\u7ec3\u8bbe\u7f6e\u7684\u6548\u679c\u5982\u4f55\uff0c\u5b83\u53ef\u80fd\u5e26\u6709\u5176\u4ed6\u5305\u88b1\u3002RLHF \u4f9d\u8d56\u4e8e\u6b64\uff0c\u4f46\u5b66\u4e60\u5230\u7684\u6807\u91cf\u5956\u52b1\u6a21\u578b\u975e\u5e38\u5bb9\u6613\u5b66\u4e60\u5230\u4e0d\u9700\u8981\u7684\u7279\u8d28\u3002<\/section>\n<\/li>\n<li>\n<section>Variable indifference\u3002\u76ee\u6807\u662f\u8981\u6c42\u667a\u80fd\u4f53\u4f18\u5316\u73af\u5883\u4e2d\u7684\u4e00\u4e9b\u53d8\u91cf\uff0c\u800c\u4e0d\u662f\u5176\u4ed6\u53d8\u91cf\u3002<\/section>\n<\/li>\n<li>\n<section>\u9677\u9631\u3002\u6211\u4eec\u53ef\u4ee5\u6709\u610f\u5f15\u5165\u4e00\u4e9b\u6f0f\u6d1e\uff0c\u5e76\u5728\u4efb\u4f55\u5956\u52b1\u88ab hacking \u65f6\u8bbe\u7f6e\u76d1\u63a7\u548c\u8b66\u62a5\u3002<\/section>\n<\/li>\n<li>\n<section>\u5728\u4ee5\u4eba\u7c7b\u53cd\u9988\u4f5c\u4e3a\u667a\u80fd\u4f53\u884c\u4e3a\u8ba4\u53ef\u7684 RL \u8bbe\u7f6e\u4e2d\uff0cUesato et al. (2020) \u63d0\u51fa\u4e86\u7528\u89e3\u8026\u6279\u51c6\uff08decoupled approval\uff09\u6765\u9632\u6b62\u5956\u52b1\u7be1\u6539\u3002\u5982\u679c\u53cd\u9988\u662f\u57fa\u4e8e (state, reward) \u7684\uff0c\u4e00\u65e6\u8fd9\u5bf9\u6570\u636e\u53d1\u751f\u5956\u52b1\u7be1\u6539\uff0c\u6211\u4eec\u5c31\u65e0\u6cd5\u83b7\u5f97\u8be5\u884c\u4e3a\u5728\u8be5\u72b6\u6001\u4e0b\u7684\u672a\u88ab\u6c61\u67d3\u7684\u53cd\u9988\u3002\u89e3\u8026\u610f\u5473\u7740\u6536\u96c6\u53cd\u9988\u7684\u67e5\u8be2\u884c\u4e3a\u662f\u4ece\u4e16\u754c\u4e0a\u91c7\u53d6\u7684\u884c\u4e3a\u4e2d\u72ec\u7acb\u62bd\u6837\u7684\u3002\u53cd\u9988\u751a\u81f3\u5728\u884c\u4e3a\u5728\u4e16\u754c\u4e2d\u6267\u884c\u4e4b\u524d\u5c31\u5df2\u6536\u5230\uff0c\u4ece\u800c\u9632\u6b62\u884c\u4e3a\u635f\u5bb3\u81ea\u5df1\u7684\u53cd\u9988\u3002<\/section>\n<\/li>\n<\/ul>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-9c9c2b324d71aaeb10400e399e7c90d7.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-9c9c2b324d71aaeb10400e399e7c90d7.png\"><\/a><\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-76e181c8ce7e85a487199e596504acd9.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-76e181c8ce7e85a487199e596504acd9.png\"><\/a><\/section>\n<section><strong>\u68c0\u6d4b Reward Hacking \u884c\u4e3a<\/strong><\/section>\n<section><\/section>\n<section>\u53e6\u4e00\u79cd\u7f13\u89e3\u63aa\u65bd\u662f\u901a\u8fc7\u5c06\u5176\u6846\u67b6\u5316\u4e3a\u5f02\u5e38\u68c0\u6d4b\u4efb\u52a1\u6765\u68c0\u6d4b Reward Hacking\uff0c\u5176\u4e2d\u68c0\u6d4b\u5668\uff08\u5177\u6709\u7531\u4eba\u7c7b\u9a8c\u8bc1\u7684\u8f68\u8ff9\u548c\u5956\u52b1\u7684\u300c\u53ef\u4fe1\u7b56\u7565\u300d\uff09\u5e94\u6807\u8bb0\u9519\u4f4d\u5b9e\u4f8b\uff08Pan et al. 2022\uff09\u3002\u7ed9\u5b9a\uff081\uff09\u4e00\u4e2a\u53ef\u4fe1\u7b56\u7565\u548c\uff082\uff09\u4e00\u7ec4\u624b\u52a8\u6807\u8bb0\u7684\u8f68\u8ff9 rollouts\uff0c\u6211\u4eec\u53ef\u4ee5\u6839\u636e\u4e24\u4e2a\u7b56\u7565\uff08\u53ef\u4fe1\u7b56\u7565\u548c\u76ee\u6807\u7b56\u7565\uff09\u7684\u52a8\u4f5c\u5206\u5e03\u4e4b\u95f4\u7684\u8ddd\u79bb\u6784\u5efa\u4e8c\u5143\u5206\u7c7b\u5668\uff0c\u5e76\u6d4b\u91cf\u8be5\u5f02\u5e38\u68c0\u6d4b\u5206\u7c7b\u5668\u7684\u51c6\u786e\u6027\u3002\u5728 Pan et al. (2022) \u7684\u5b9e\u9a8c\u4e2d\uff0c\u4ed6\u4eec\u89c2\u5bdf\u5230\u4e0d\u540c\u7684\u68c0\u6d4b\u5668\u66f4\u9002\u5408\u4e0d\u540c\u7684\u4efb\u52a1\uff0c\u5e76\u4e14\u5728\u6240\u6709\u6d4b\u8bd5\u7684 RL \u73af\u5883\u4e2d\uff0c\u6ca1\u6709\u4e00\u4e2a\u6d4b\u8bd5\u7684\u5206\u7c7b\u5668\u80fd\u591f\u5b9e\u73b0\u5927\u4e8e 60% \u7684 AUROC\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-b086868bfa21d2c982201b1d280c1408.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-b086868bfa21d2c982201b1d280c1408.png\"><\/a><\/section>\n<section><strong>RLHF \u7684\u6570\u636e\u5206\u6790<\/strong><\/section>\n<section><\/section>\n<section>\u53e6\u4e00\u79cd\u65b9\u6cd5\u662f\u5206\u6790 RLHF \u6570\u636e\u96c6\u3002\u901a\u8fc7\u68c0\u67e5\u8bad\u7ec3\u6570\u636e\u5982\u4f55\u5f71\u54cd\u5bf9\u9f50\u8bad\u7ec3\u7ed3\u679c\uff0c\u53ef\u4ee5\u6307\u5bfc\u9884\u5904\u7406\u548c\u4eba\u5de5\u53cd\u9988\u6536\u96c6\uff0c\u4ee5\u964d\u4f4e reward hacking \u98ce\u9669\u3002<\/section>\n<section><\/section>\n<ul>\n<li>\n<section>\u76ee\u6807\u7279\u5f81\uff1a\u6307\u660e\u786e\u60f3\u8981\u5b66\u4e60\u7684\u503c\u3002<\/section>\n<\/li>\n<li>\n<section>Spoiler \u7279\u5f81\uff1a\u6307\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u65e0\u610f\u4e2d\u5b66\u5230\u7684\u975e\u9884\u671f\u503c\uff08\u4f8b\u5982\uff0c\u60c5\u611f\u6216\u8fde\u8d2f\u6027\u7b49\u98ce\u683c\u6027\u7279\u5f81\uff09\u3002\u8fd9\u4e9b\u7c7b\u4f3c\u4e8e OOD \u5206\u7c7b\u5de5\u4f5c\u4e2d\u7684\u865a\u5047\u7279\u5f81\uff08spurious features\uff09\uff08Geirhos et al. 2020\uff09\u3002<\/section>\n<\/li>\n<\/ul>\n<section><\/section>\n<section>SEAL \u8fd8\u5f15\u5165\u4e86\u4e09\u4e2a\u8861\u91cf\u5bf9\u9f50\u8bad\u7ec3\u6570\u636e\u6709\u6548\u6027\u7684\u6307\u6807\uff1a<\/section>\n<section><\/section>\n<section>1. \u7279\u5f81\u5370\u8bb0\uff08feature imprint\uff09\uff0c\u662f\u6307\u7279\u5f81 \u03c4 \u7684\u4e00\u4e2a\u7cfb\u6570\u53c2\u6570 \u03b2_\u03c4\uff0c\u5728\u4fdd\u6301\u5176\u4ed6\u56e0\u7d20\u4e0d\u53d8\u7684\u60c5\u51b5\u4e0b\uff0c\u8be5\u7cfb\u6570\u53c2\u6570\u7528\u4e8e\u4f30\u8ba1\u6bd4\u8f83\u6709\u6216\u65e0\u7279\u5f81 \u03c4 \u65f6\u5956\u52b1\u70b9\u6570\u7684\u589e\u52a0\u60c5\u51b5\u3002<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-eba4e2ea44d0708b31b8c598449366c0.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-eba4e2ea44d0708b31b8c598449366c0.png\"><\/a><\/section>\n<section><em><sup>\u56fe 21\uff1a(\u5de6) \u7279\u5f81\u5370\u8bb0 \u03b2(\u03c4) (pre-) \u548c \u03b2(\u03c4) (post-) \u7531\u9488\u5bf9\u7279\u5f81\u7684\u5956\u52b1<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d8fb0f411ccbaefd8fccbdfa353f3935.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-d8fb0f411ccbaefd8fccbdfa353f3935.png\"><\/a><\/sup><\/em><em><sup>\u7684\u56fa\u5b9a\u6548\u5e94\u7ebf\u6027\u56de\u5f52\u8ba1\u7b97\u5f97\u51fa\u3002\u603b\u4f53\u800c\u8a00\uff0c\u5bf9\u9f50\u8bad\u7ec3\u4f1a\u5956\u52b1\u65e0\u5bb3\u548c\u6709\u7528\u7b49\u79ef\u6781\u7279\u5f81\uff0c\u5e76\u60e9\u7f5a\u6027\u5185\u5bb9\u6216\u4fb5\u72af\u9690\u79c1\u7b49\u6d88\u6781\u7279\u5f81\u3002(\u53f3) \u7279\u5f81\u5370\u8bb0\u7531\u5956\u52b1\u504f\u79fb \u03b8_i \u7684\u7ebf\u6027\u56de\u5f52\u8ba1\u7b97\u5f97\u51fa\u3002\u5956\u52b1\u504f\u79fb \u03b8_i \u7684\u5b9a\u4e49\u4e3a\u5bf9\u9f50\u8bad\u7ec3\u524d\u540e\u5956\u52b1\u5411\u91cf\u4e4b\u95f4\u7684\u89d2\u5ea6\u3002\u8bad\u7ec3\u8fc7\u7a0b\u4f1a\u4f18\u5316\u6a21\u578b\u5bf9\u76ee\u6807\u7279\u5f81\u7684\u654f\u611f\u5ea6\u3002\u603b\u4f53\u800c\u8a00\uff0c\u5bf9\u9f50\u8bad\u7ec3\u4f1a\u5956\u52b1\u65e0\u5bb3\u548c\u6709\u7528\u7b49\u79ef\u6781\u7279\u5f81\uff0c\u5e76\u60e9\u7f5a\u6027\u5185\u5bb9\u6216\u4fb5\u72af\u9690\u79c1\u7b49\u6d88\u6781\u7279\u5f81\u3002(\u53f3) \u7279\u5f81\u5370\u8bb0\u7531\u5956\u52b1\u504f\u79fb \u03b8_i \u7684\u7ebf\u6027\u56de\u5f52\u8ba1\u7b97\u5f97\u51fa\u3002\u5956\u52b1\u504f\u79fb \u03b8_i \u7684\u5b9a\u4e49\u4e3a\u5bf9\u9f50\u8bad\u7ec3\u524d\u540e\u5956\u52b1\u5411\u91cf\u4e4b\u95f4\u7684\u89d2\u5ea6\u3002\u8bad\u7ec3\u8fc7\u7a0b\u4f1a\u4f18\u5316\u6a21\u578b\u5bf9\u76ee\u6807\u7279\u5f81\u7684\u654f\u611f\u5ea6\u3002\u8bf7\u6ce8\u610f\uff0c\u65e0\u5bb3\u901a\u8fc7\u9009\u5b9a\u548c\u62d2\u7edd\u7684\u6761\u76ee\uff08\u5305\u62ec is harmless (c) \u548c is harmless (r)\uff09\u5370\u8bb0\u5728 RM \u4e0a\uff0c\u800c\u6709\u7528\u6027\u4ec5\u901a\u8fc7\u62d2\u7edd\u7684\u6761\u76ee\uff08is helpful (r)\uff09\u6765\u5370\u8bb0\u3002(\u56fe\u6e90\uff1aRevel et al. 2024)<\/sup><\/em><\/section>\n<section><\/section>\n<section>2. \u5bf9\u9f50\u963b\u529b\uff08Alignment resistance\uff09\u662f RM \u65e0\u6cd5\u5339\u914d\u4eba\u7c7b\u504f\u597d\u7684\u504f\u597d\u6570\u636e\u5bf9\u7684\u767e\u5206\u6bd4\u3002\u7814\u7a76\u53d1\u73b0\uff0cRM \u5728\u8d85\u8fc7 1\/4 \u7684 HHH-RLHF \u6570\u636e\u96c6\u4e0a\u53ef\u4ee5\u62b5\u6297\u4eba\u7c7b\u504f\u597d\u3002<\/section>\n<section><\/section>\n<section>3. \u5bf9\u9f50\u7a33\u5065\u6027\uff08Alignment robustness\uff09<\/section>\n<section><a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8b7acbd4687c9d6c7b11a364fe158eae.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-8b7acbd4687c9d6c7b11a364fe158eae.png\"><\/a>\u8861\u91cf\u7684\u662f\u5bf9\u9f50\u5bf9\u5e26\u6709\u91cd\u5199\u7684\u6270\u52a8\u8f93\u5165\u7684\u7a33\u5065\u7a0b\u5ea6\uff0c\u5305\u62ec\u60c5\u7eea\u3001\u96c4\u8fa9\u548c\u8fde\u8d2f\u6027\u7b49\u5267\u900f\u7279\u5f81\uff08spoiler features\uff09\u03c4\uff0c\u5176\u80fd\u9694\u79bb\u6bcf\u4e2a\u7279\u5f81\u548c\u6bcf\u79cd\u4e8b\u4ef6\u7c7b\u578b\u7684\u5f71\u54cd\u3002<\/section>\n<section><\/section>\n<section>\u7a33\u5065\u6027\u6307\u6807<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-b54e69dc58cbb2ecdae2f1f6f19519d5.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-b54e69dc58cbb2ecdae2f1f6f19519d5.png\"><\/a>\uff08\u5982\u300c\u96c4\u8fa9\u300d\u6216\u300c\u60c5\u7eea\u79ef\u6781\u300d\u7b49\u7279\u5f81\u540d\u79f0 \u03c4\uff09\u5e94\u4ee5\u4ee5\u4e0b\u65b9\u5f0f\u89e3\u91ca\uff1a<\/section>\n<section><\/section>\n<section>\u4e0e\u6ca1\u6709\u6b64\u7c7b\u7ffb\u8f6c\u7684\u5176\u4ed6\u6761\u76ee\u76f8\u6bd4\uff0c\u5728\u91cd\u5199\u540e\u5305\u542b\u66f4\u5f3a\u7279\u5f81 \u03c4 \u7684\u9009\u5b9a\u6761\u76ee\uff08\u8bb0\u4e3a c\uff09\u88ab\u62d2\u7edd\u7684\u51e0\u7387\u9ad8\u51fa<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-605e888f402d0b30c9bc5ed7d179e97e.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-605e888f402d0b30c9bc5ed7d179e97e.png\"><\/a>\u500d\u3002<\/section>\n<section><\/section>\n<section>\u7c7b\u4f3c\u5730\uff0c\u4e0e\u6ca1\u6709\u6b64\u7c7b\u7ffb\u8f6c\u7684\u5176\u4ed6\u6761\u76ee\u76f8\u6bd4\uff0c\u5728\u91cd\u5199\u540e\u83b7\u5f97\u8f83\u5f31\u7279\u5f81 \u03c4 \u7684\u88ab\u62d2\u7edd\u6761\u76ee\uff08\u8bb0\u4e3a r \uff09\u88ab\u9009\u4e2d\u7684\u51e0\u7387\u662f<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4bdddea7edcb7d00c603382a0981323a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-4bdddea7edcb7d00c603382a0981323a.png\"><\/a>\u500d\u3002<\/section>\n<section><\/section>\n<section>\u6839\u636e\u4ed6\u4eec\u5bf9\u4e0d\u540c\u91cd\u5199\u65b9\u9762\u5bf9\u9f50\u7a33\u5065\u6027\u6307\u6807\u7684\u5206\u6790\uff0c\u53ea\u6709\u57fa\u4e8e\u60c5\u611f\u5267\u900f\u7279\u5f81\u7684\u7a33\u5065\u6027\u5f97\u5206<a href=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-335a4bf46b83064d8dafb4776d1deb3a.png\" data-fancybox=\"images\" data-fancybox=\"gallery\"><img decoding=\"async\" src=\"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-335a4bf46b83064d8dafb4776d1deb3a.png\"><\/a>\u662f\u7edf\u8ba1\u663e\u8457\u7684\u3002<\/section>\n<section><\/section>\n<section><em><sup>\u53c2\u8003\u6587\u732e\uff1a<\/sup><\/em><\/section>\n<section><sup><br \/><\/sup><\/section>\n<section><sup><em>[1] Andrew Ng &amp; Stuart Russell. \u201cAlgorithms for inverse reinforcement learning.\u201d. ICML 2000.<\/em><\/sup><\/section>\n<section><sup><em>[2] Amodei et al. \u201cConcrete problems in AI safety: Avoid reward hacking.\u201d arXiv preprint arXiv:1606.06565 (2016).<\/em><\/sup><\/section>\n<section><sup><em>[3] Krakovna et al. \u201cSpecification gaming: the flip side of AI ingenuity.\u201d 2020.<\/em><\/sup><\/section>\n<section><sup><em>[4] Langosco et al. \u201cGoal Misgeneralization in Deep Reinforcement Learning\u201d ICML 2022.<\/em><\/sup><\/section>\n<section><sup><em>[5] Everitt et al. \u201cReinforcement learning with a corrupted reward channel.\u201d IJCAI 2017.<\/em><\/sup><\/section>\n<section><sup><em>[6] Geirhos et al. \u201cShortcut Learning in Deep Neural Networks.\u201d Nature Machine Intelligence 2020.<\/em><\/sup><\/section>\n<section><sup><em>[7] Ribeiro et al. \u201cWhy Should I Trust You?\u201d: Explaining the Predictions of Any Classifier. KDD 2016.<\/em><\/sup><\/section>\n<section><sup><em>[8] Nagarajan et al. \u201cUnderstanding the Failure Modes of Out-of-Distribution Generalization.\u201d ICLR 2021.<\/em><\/sup><\/section>\n<section><sup><em>[9] Garrabrant. \u201cGoodhart Taxonomy\u201d. AI Alignment Forum (Dec 30th 2017).<\/em><\/sup><\/section>\n<section><sup><em>[10] Koch et al. \u201cObjective robustness in deep reinforcement learning.\u201d 2021.<\/em><\/sup><\/section>\n<section><sup><em>[11] Pan et al. \u201cThe effects of reward misspecification: mapping and mitigating misaligned models.\u201d<\/em><\/sup><\/section>\n<section><sup><em>[12] Everitt et al. \u201cReward tampering problems and solutions in reinforcement learning: A causal influence diagram perspective.\u201d arXiv preprint arXiv:1908.04734 (2019).<\/em><\/sup><\/section>\n<section><sup><em>[13] Gleave et al. \u201cAdversarial Policies: Attacking Deep Reinforcement Learning.\u201d ICRL 2020<\/em><\/sup><\/section>\n<section><sup><em>[14] \u201cReward hacking behavior can generalize across tasks.\u201d<\/em><\/sup><\/section>\n<section><sup><em>[15] Ng et al. \u201cPolicy invariance under reward transformations: Theory and application to reward shaping.\u201d ICML 1999.<\/em><\/sup><\/section>\n<section><sup><em>[16] Wang et al. \u201cLarge Language Models are not Fair Evaluators.\u201d ACL 2024.<\/em><\/sup><\/section>\n<section><sup><em>[17] Liu et al. \u201cLLMs as narcissistic evaluators: When ego inflates evaluation scores.\u201d ACL 2024.<\/em><\/sup><\/section>\n<section><sup><em>[18] Gao et al. \u201cScaling Laws for Reward Model Overoptimization.\u201d ICML 2023.<\/em><\/sup><\/section>\n<section><sup><em>[19] Pan et al. \u201cSpontaneous Reward Hacking in Iterative Self-Refinement.\u201d arXiv preprint arXiv:2407.04549 (2024).<\/em><\/sup><\/section>\n<section><sup><em>[20] Pan et al. \u201cFeedback Loops With Language Models Drive In-Context Reward Hacking.\u201d arXiv preprint arXiv:2402.06627 (2024).<\/em><\/sup><\/section>\n<section><sup><em>[21] Shrama et al. \u201cTowards Understanding Sycophancy in Language Models.\u201d arXiv preprint arXiv:2310.13548 (2023).<\/em><\/sup><\/section>\n<section><sup><em>[22] Denison et al. \u201cSycophancy to subterfuge: Investigating reward tampering in language models.\u201d arXiv preprint arXiv:2406.10162 (2024).<\/em><\/sup><\/section>\n<section><sup><em>[23] Uesato et al. \u201cAvoiding Tampering Incentives in Deep RL via Decoupled Approval.\u201d arXiv preprint arXiv:2011.08827 (2020).<\/em><\/sup><\/section>\n<section><sup><em>[24] Amin and Singh. \u201cTowards resolving unidentifiability in inverse reinforcement learning.\u201d<\/em><\/sup><\/section>\n<section><sup><em>[25] Wen et al. \u201cLanguage Models Learn to Mislead Humans via RLHF.\u201d arXiv preprint arXiv:2409.12822 (2024).<\/em><\/sup><\/section>\n<section><sup><em>[26] Revel et al. \u201cSEAL: Systematic Error Analysis for Value ALignment.\u201d arXiv preprint arXiv:2408.10270 (2024).<\/em><\/sup><\/section>\n<section><sup><em>[27] Yuval Noah Harari. \u201cNexus: A Brief History of Information Networks from the Stone Age to AI.\u201d Signal; 2024 Sep 10.<\/em><\/sup><\/section>\n<section><sup><br \/><\/sup><\/section>\n<section><sup><em>\u53c2\u8003\u5185\u5bb9\uff1a<\/em><\/sup><\/section>\n<section><sup><em>https:\/\/lilianweng.github.io\/posts\/2024-11-28-reward-hacking\/<\/em><\/sup><\/section>\n<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:<a href=\"https:\/\/www.jiqizhixin.com\/articles\/2024-12-02-8\" target=\"_blank\">\u79bb\u804cOpenAI\u540e\uff0c\u7fc1\u8354\u535a\u5ba2\u9996\u6b21\u4e0a\u65b0\uff0c\u5f15\u4f17\u7f51\u53cb\u56f4\u89c2\u5b66\u4e60\uff08\u4e2d\u6587\u5168\u6587\uff09<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6587\u7ae0\u6765\u6e90\u4e8e\u4e92\u8054\u7f51:\u79bb\u804cOpenAI\u540e\uff0c\u7fc1 [&hellip;]<\/p>\n","protected":false},"author":3,"featured_media":0,"comment_status":"open","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"","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":"","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":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","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":[27],"tags":[68],"class_list":["post-36920","post","type-post","status-publish","format-standard","hentry","category-news","tag-68"],"yoast_head":"<!-- This site is optimized with the Yoast SEO plugin v27.9 - https:\/\/yoast.com\/product\/yoast-seo-wordpress\/ -->\n<title>\u79bb\u804cOpenAI\u540e\uff0c\u7fc1\u8354\u535a\u5ba2\u9996\u6b21\u4e0a\u65b0\uff0c\u5f15\u4f17\u7f51\u53cb\u56f4\u89c2\u5b66\u4e60\uff08\u4e2d\u6587\u5168\u6587\uff09 - \u4e00\u8d77AI\u6280\u672f<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/17aitech.com\/?p=36920\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36920#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36920\"},\"author\":{\"name\":\"AI\u5c0f\u52a9\u624b\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"headline\":\"\u79bb\u804cOpenAI\u540e\uff0c\u7fc1\u8354\u535a\u5ba2\u9996\u6b21\u4e0a\u65b0\uff0c\u5f15\u4f17\u7f51\u53cb\u56f4\u89c2\u5b66\u4e60\uff08\u4e2d\u6587\u5168\u6587\uff09\",\"datePublished\":\"2025-01-20T23:01:14+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36920\"},\"wordCount\":1112,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36920#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-00fd0c6412cad47d41511dcabb84a0d1.png\",\"keywords\":[\"\u5927\u6a21\u578b\"],\"articleSection\":[\"\u884c\u4e1a\u8d44\u8baf\"],\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=36920#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36920\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?p=36920\",\"name\":\"\u79bb\u804cOpenAI\u540e\uff0c\u7fc1\u8354\u535a\u5ba2\u9996\u6b21\u4e0a\u65b0\uff0c\u5f15\u4f17\u7f51\u53cb\u56f4\u89c2\u5b66\u4e60\uff08\u4e2d\u6587\u5168\u6587\uff09 - \u4e00\u8d77AI\u6280\u672f\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36920#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36920#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-00fd0c6412cad47d41511dcabb84a0d1.png\",\"datePublished\":\"2025-01-20T23:01:14+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\"},\"breadcrumb\":{\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36920#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/17aitech.com\\\/?p=36920\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36920#primaryimage\",\"url\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-00fd0c6412cad47d41511dcabb84a0d1.png\",\"contentUrl\":\"https:\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/12\\\/frc-00fd0c6412cad47d41511dcabb84a0d1.png\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/?p=36920#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\\\/\\\/17aitech.com\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"\u79bb\u804cOpenAI\u540e\uff0c\u7fc1\u8354\u535a\u5ba2\u9996\u6b21\u4e0a\u65b0\uff0c\u5f15\u4f17\u7f51\u53cb\u56f4\u89c2\u5b66\u4e60\uff08\u4e2d\u6587\u5168\u6587\uff09\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#website\",\"url\":\"https:\\\/\\\/17aitech.com\\\/\",\"name\":\"\u4e00\u8d77AI\u6280\u672f\",\"description\":\"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca\",\"alternateName\":\"\u4e00\u8d77AI\u6280\u672f\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/17aitech.com\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/17aitech.com\\\/#\\\/schema\\\/person\\\/60225458499e817ae0af73e67e440b9d\",\"name\":\"AI\u5c0f\u52a9\u624b\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"url\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"contentUrl\":\"\\\/\\\/17aitech.com\\\/wp-content\\\/uploads\\\/2024\\\/04\\\/robot_3.png\",\"caption\":\"AI\u5c0f\u52a9\u624b\"},\"description\":\"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e\",\"url\":\"https:\\\/\\\/17aitech.com\\\/?page_id=33738&user=3\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"\u79bb\u804cOpenAI\u540e\uff0c\u7fc1\u8354\u535a\u5ba2\u9996\u6b21\u4e0a\u65b0\uff0c\u5f15\u4f17\u7f51\u53cb\u56f4\u89c2\u5b66\u4e60\uff08\u4e2d\u6587\u5168\u6587\uff09 - \u4e00\u8d77AI\u6280\u672f","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/17aitech.com\/?p=36920","schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/17aitech.com\/?p=36920#article","isPartOf":{"@id":"https:\/\/17aitech.com\/?p=36920"},"author":{"name":"AI\u5c0f\u52a9\u624b","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"headline":"\u79bb\u804cOpenAI\u540e\uff0c\u7fc1\u8354\u535a\u5ba2\u9996\u6b21\u4e0a\u65b0\uff0c\u5f15\u4f17\u7f51\u53cb\u56f4\u89c2\u5b66\u4e60\uff08\u4e2d\u6587\u5168\u6587\uff09","datePublished":"2025-01-20T23:01:14+00:00","mainEntityOfPage":{"@id":"https:\/\/17aitech.com\/?p=36920"},"wordCount":1112,"commentCount":0,"image":{"@id":"https:\/\/17aitech.com\/?p=36920#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-00fd0c6412cad47d41511dcabb84a0d1.png","keywords":["\u5927\u6a21\u578b"],"articleSection":["\u884c\u4e1a\u8d44\u8baf"],"inLanguage":"zh-Hans","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/17aitech.com\/?p=36920#respond"]}]},{"@type":"WebPage","@id":"https:\/\/17aitech.com\/?p=36920","url":"https:\/\/17aitech.com\/?p=36920","name":"\u79bb\u804cOpenAI\u540e\uff0c\u7fc1\u8354\u535a\u5ba2\u9996\u6b21\u4e0a\u65b0\uff0c\u5f15\u4f17\u7f51\u53cb\u56f4\u89c2\u5b66\u4e60\uff08\u4e2d\u6587\u5168\u6587\uff09 - \u4e00\u8d77AI\u6280\u672f","isPartOf":{"@id":"https:\/\/17aitech.com\/#website"},"primaryImageOfPage":{"@id":"https:\/\/17aitech.com\/?p=36920#primaryimage"},"image":{"@id":"https:\/\/17aitech.com\/?p=36920#primaryimage"},"thumbnailUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-00fd0c6412cad47d41511dcabb84a0d1.png","datePublished":"2025-01-20T23:01:14+00:00","author":{"@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d"},"breadcrumb":{"@id":"https:\/\/17aitech.com\/?p=36920#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/17aitech.com\/?p=36920"]}]},{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/17aitech.com\/?p=36920#primaryimage","url":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-00fd0c6412cad47d41511dcabb84a0d1.png","contentUrl":"https:\/\/17aitech.com\/wp-content\/uploads\/2024\/12\/frc-00fd0c6412cad47d41511dcabb84a0d1.png"},{"@type":"BreadcrumbList","@id":"https:\/\/17aitech.com\/?p=36920#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/17aitech.com\/"},{"@type":"ListItem","position":2,"name":"\u79bb\u804cOpenAI\u540e\uff0c\u7fc1\u8354\u535a\u5ba2\u9996\u6b21\u4e0a\u65b0\uff0c\u5f15\u4f17\u7f51\u53cb\u56f4\u89c2\u5b66\u4e60\uff08\u4e2d\u6587\u5168\u6587\uff09"}]},{"@type":"WebSite","@id":"https:\/\/17aitech.com\/#website","url":"https:\/\/17aitech.com\/","name":"\u4e00\u8d77AI\u6280\u672f","description":"\u8ba9AI\u77e5\u8bc6\u89e6\u624b\u53ef\u53ca","alternateName":"\u4e00\u8d77AI\u6280\u672f","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/17aitech.com\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/17aitech.com\/#\/schema\/person\/60225458499e817ae0af73e67e440b9d","name":"AI\u5c0f\u52a9\u624b","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","url":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","contentUrl":"\/\/17aitech.com\/wp-content\/uploads\/2024\/04\/robot_3.png","caption":"AI\u5c0f\u52a9\u624b"},"description":"\u8fd9\u4e2a\u4eba\u5f88\u61d2\uff0c\u4ec0\u4e48\u90fd\u6ca1\u6709\u7559\u4e0b\uff5e","url":"https:\/\/17aitech.com\/?page_id=33738&user=3"}]}},"_links":{"self":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/36920","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=36920"}],"version-history":[{"count":0,"href":"https:\/\/17aitech.com\/index.php?rest_route=\/wp\/v2\/posts\/36920\/revisions"}],"wp:attachment":[{"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=36920"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=36920"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/17aitech.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=36920"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}