{"id":48159,"date":"2024-02-03T19:16:25","date_gmt":"2024-02-03T11:16:25","guid":{"rendered":"https:\/\/swarma.org\/?p=48159"},"modified":"2024-02-03T19:16:25","modified_gmt":"2024-02-03T11:16:25","slug":"ai%e5%af%b9%e9%bd%90%e6%98%af%e6%8e%a7%e5%88%b6%e8%ae%ba%e8%bf%98%e6%98%af%e5%8d%9a%e5%bc%88%e8%ae%ba%ef%bc%9f","status":"publish","type":"post","link":"https:\/\/swarma.org\/?p=48159","title":{"rendered":"AI\u5bf9\u9f50\u662f\u63a7\u5236\u8bba\u8fd8\u662f\u535a\u5f08\u8bba\uff1f"},"content":{"rendered":"
\n
\n
\n
\n
\n
\n

<\/p>\n<\/section>\n

\n
\n
\n
\n
\n
\n

<\/section>\n<\/section>\n
\n
\n

\u5bfc\u8bed<\/strong><\/p>\n<\/section>\n<\/section>\n

\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
\n
\n

\u672c\u6587\u662f\u5317\u4eac\u5927\u5b66\u8ba1\u7b97\u673a\u56fe\u7075\u73ed\u672c\u79d1\u751f\u3001\u5317\u5927\u5bf9\u9f50\u5c0f\u7ec4 (PKU-Alignemnt\uff09\u6210\u5458\u90b1\u5929\u5f02\u64b0\u5199\u7684\u6587\u7ae0\uff0c\u4ecb\u7ecd\u4e86AI\u5bf9\u9f50\u95ee\u9898\u7684\u4e24\u79cd\u4e3b\u8981\u601d\u8def\uff1a\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\uff08\u63a7\u5236\u8bba\u8fdb\u8def\uff09\uff0c\u5408\u4f5c\u9006\u5f3a\u5316\u5b66\u4e60\uff08\u535a\u5f08\u8bba\u8fdb\u8def\uff09\u3002<\/span><\/strong>\u90b1\u5929\u5f02<\/span><\/strong>\u5728\u5317\u5927PAIR-Lab\u8fdb\u884cAI\u5bf9\u9f50\u7814\u7a76\uff0c\u5173\u6ce8\u5bf9\u9f50\u4e0e\u9053\u5fb7\u4ef7\u503c\u7684\u4ea4\u53c9\uff0c\u5e76\u5408\u4f5c\u64b0\u5199\u7efc\u8ff0\u6587\u7ae0 AI Alignment: A Comprehensive Survey\u3002<\/span><\/strong><\/p>\n<\/section>\n

\n


<\/span><\/strong><\/p>\n<\/section>\n

\n

\u7814\u7a76\u9886\u57df\uff1aAI\u5bf9\u9f50\uff0c\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\uff0c\u5408\u4f5c\u9006\u5f3a\u5316\u5b66\u4e60<\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/strong><\/span><\/strong><\/span><\/strong><\/span><\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n

\n
\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n
\n
<\/section>\n<\/section>\n
\n

\u5317\u4eac\u5927\u5b66\u524d\u6cbf\u8ba1\u7b97\u7814\u7a76\u4e2d\u5fc3<\/span> | \u6765\u6e90<\/strong><\/span><\/strong><\/p>\n

\u90b1\u5929\u5f02<\/span><\/span><\/span> | \u4f5c\u8005<\/strong><\/span><\/span><\/span><\/p>\n<\/section>\n


<\/p>\n<\/section>\n

\n
\n
\n


<\/span><\/p>\n

\n
\n
\n
\n
\n

<\/section>\n

<\/section>\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n
\n

AI\u5bf9\u9f50\uff1a\u6cd5\u5178\u7684\u7c7b\u6bd4<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


<\/p>\n<\/section>\n<\/section>\n<\/section>\n

\n
\n

\u5982\u679c\u8bf7\u4f60\u7f16\u5199\u4e00\u90e8\u65e0\u6240\u4e0d\u5305\u7684\u6cd5\u5178\uff0c\u5e76\u4e14\u8fd9\u90e8\u6cd5\u5178\u5c06\u96f7\u6253\u4e0d\u52a8\u3001\u4e0d\u7ecf\u4fee\u6539\u5730\u751f\u6548\u4e00\u767e\u5e74\uff0c\u4f60\u6709\u591a\u5927\u4fe1\u5fc3\u80fd\u591f\u7ed9\u51fa\u4e00\u4e2a\u5b8c\u5584\u7684\u65b9\u6848\uff1f<\/span><\/p>\n


<\/p>\n

\u8fd9\u60f3\u5fc5\u4f1a\u9047\u5230\u8bf8\u591a\u56f0\u96be\uff0c\u5305\u62ec\u4f46\u4e0d\u9650\u4e8e\uff1a<\/span><\/p>\n<\/section>\n<\/section>\n


<\/p>\n

\n

\u9519\u8bef\u89c4\u8303\uff08Misspecification\uff09\uff1a<\/span><\/strong><\/span><\/p>\n<\/section>\n

\n

\u4f60\u6240\u8bbe\u60f3\u51fa\u7684\u6cd5\u6761\u867d\u7136\u770b\u4f3c\u5b8c\u5584\uff0c\u4f46\u53ef\u80fd\u5b58\u5728\u5f88\u591a\u9690\u542b\u7684\u6f0f\u6d1e\u3002\u5f53\u6709\u4eba\u8bd5\u56fe\u5229\u7528\u94bb\u6cd5\u6761\u7684\u7a7a\u5b50\u65f6\uff0c\u4f1a\u53d1\u73b0\u5b58\u5728\u8bb8\u591a\u53ef\u4e58\u4e4b\u673a\u3002\u8fd9\u53ef\u4ee5\u770b\u4f5c\u662f\u7ecf\u6d4e\u5b66\u4e2d\u53e4\u5fb7\u54c8\u7279\u5b9a\u5f8b<\/span>\uff08Goodhart’s Law\uff09<\/span>\u7684\u4f53\u73b0\u2014\u2014\u4efb\u4f55\u8bc4\u4ef7\u6307\u6807\u4e00\u65e6\u88ab\u4f5c\u4e3a\u4f18\u5316\u76ee\u6807\uff0c\u5c31\u4e0d\u518d\u662f\u597d\u7684\u8bc4\u4ef7\u6307\u6807<\/span>\uff08\u8fd9\u91cc\u6211\u4eec\u628a \u201c\u9075\u4ece\u6cd5\u5178\u201d \u89c6\u4e3a\u4e00\u4e2a\u8bc4\u4ef7\u6307\u6807\uff09<\/span>\uff0c\u8fd9\u662f\u56e0\u4e3a\u76f8\u5173\u6027\u4e0d\u4ee3\u8868\u56e0\u679c\u6027\uff0c\u6545\u5bf9\u4e00\u4e2a\u4e0e\u771f\u5b9e\u76ee\u6807\u5177\u6709\u76f8\u5173\u6027\u7684 \u201c\u4ee3\u7406\u76ee\u6807\u201d<\/span>\uff08proxy goal\uff09<\/span>\u65bd\u4ee5\u4f18\u5316\uff0c\u5f88\u53ef\u80fd\u56e0\u56e0\u679c\u5173\u8054\u7684\u7f3a\u5931\u800c\u5bfc\u81f4\u76f8\u5173\u6027\u7684\u6d88\u5931\u3002 <\/span><\/p>\n


<\/p>\n

\u9519\u8bef\u6cdb\u5316\uff08Misgeneralization\uff09\uff1a<\/span><\/strong><\/span><\/p>\n

\u4f60\u5bf9\u8fd9\u90e8\u6cd5\u5178\u7684\u6784\u601d\uff0c\u53d1\u751f\u5728\u4f60\u5bf9\u5f53\u4e0b\u4e16\u754c\u7684\u7406\u89e3\u6846\u67b6\u4e0b\uff0c\u56e0\u6b64\u5b83\u4e0d\u53ef\u907f\u514d\u5730\u4ee5\u5f53\u4e0b\u4e16\u754c\u4e2d\u7684\u6982\u5ff5\u4e3a\u57fa\u7840\u3001\u5e76\u4f9d\u8d56\u4e8e\u8bf8\u591a\u5728\u5f53\u4e0b\u4e16\u754c\u4e2d\u6210\u7acb\u7684\u5047\u8bbe\u3002\u56e0\u6b64\uff0c\u82e5\u628a\u6cd5\u5178\u89c6\u4e3a\u4ece\u884c\u4e3a\u5230\u5408\u6cd5\u6027\u7684\u9884\u6d4b\u6a21\u578b\uff0c\u90a3\u4e48\u4e00\u767e\u5e74\u540e\u5b83\u5c06\u4f1a\u5728\u5206\u5e03\u5916\u8fd0\u4f5c\u3002\u5982\u679c\u767e\u5e74\u540e\u4e16\u754c\u5728\u4e00\u4e9b\u91cd\u8981\u65b9\u9762\u4e0e\u5f53\u4eca\u4e16\u754c\u6709\u672c\u8d28\u4e0d\u540c\uff0c\u5219\u6cd5\u5178\u4e2d\u7684\u4e00\u4e9b\u5185\u5bb9\u53ef\u80fd\u53d8\u5f97\u5b8c\u5168\u4e0d\u5408\u7406\uff0c\u4ece\u800c\u7ed9\u51fa\u5b8c\u5168\u4e0d\u5408\u7406\u7684\u5408\u6cd5\u6027\u5224\u65ad\u3002<\/span><\/p>\n


<\/p>\n

\u81ea\u81f4\u7684\u5206\u5e03\u504f\u79fb\uff08Auto-Induced Distributional Shifts\uff09\uff1a<\/span><\/strong><\/span><\/p>\n

\u6cd5\u5178\u672c\u8eab\u7684\u5185\u5bb9\u4f1a\u5bf9\u4e16\u754c\u7684\u53d1\u5c55\u8f68\u8ff9\u9020\u6210\u5f71\u54cd\uff0c\u56e0\u6b64\u524d\u8ff0\u7684\u5206\u5e03\u5916\u60c5\u5f62\u7684\u51fa\u73b0\uff0c\u751a\u81f3\u6709\u53ef\u80fd\u6b63\u662f\u5176\u81ea\u8eab\u9020\u6210\u7684\u3002\u8fd9\u610f\u5473\u7740\uff0c\u5728\u7f16\u5199\u8fd9\u90e8\u6cd5\u5178\u65f6\uff0c\u6211\u4eec\u6216\u8bb8\u4e0d\u5e94\u628a\u4e16\u754c\u8f68\u8ff9\u770b\u4f5c\u4e00\u4e2a\u7ed9\u5b9a\u7684\u91cf\uff0c\u800c\u5e94\u4ece\u4e92\u52a8\u7684\u89d2\u5ea6\u6765\u523b\u753b\u8fd9\u4e00\u95ee\u9898\u3002<\/span><\/p>\n


<\/p>\n

\u9053\u5fb7\u4e0d\u786e\u5b9a\u6027\uff08Moral Uncertainty\uff09\uff1a<\/span><\/strong><\/span><\/p>\n

\u6cd5\u5178\u7684\u7f16\u5199\u5fc5\u5b9a\u6d89\u53ca\u9053\u5fb7\u5224\u65ad\uff0c\u6545\u4e00\u5b9a\u662f\u7279\u5b9a\u4e8e\u4e00\u7ec4\u5177\u4f53\u7684\u4ef7\u503c\u89c2\u7684\u3002\u7136\u800c\u56f4\u7ed5\u9053\u5fb7\u8bae\u9898\u7684\u8fa9\u8bba\u4ece\u53e4\u81f3\u4eca\u4ece\u672a\u505c\u606f\u8fc7\uff0c\u5c31\u8fde\u4e13\u7cbe\u4e8e\u6b64\u7684\u4f26\u7406\u5b66\u5bb6\u4e5f\u51e0\u4e4e\u4e0d\u80fd\u5c31\u4efb\u4f55\u95ee\u9898\u8fbe\u6210\u4e00\u81f4\u3002\u8fd9\u79cd\u60c5\u51b5\u4e0b\uff0c\u600e\u6837\u7684\u9053\u5fb7\u5224\u65ad\u662f\u5408\u7406\u7684\uff0c\u662f\u4e00\u4e2a\u5f88\u96be\u56de\u7b54\u7684\u95ee\u9898\u3002<\/span><\/p>\n<\/section>\n


<\/p>\n


<\/p>\n

\n
\n
\n

\n
\n
\n
\n
\n

<\/section>\n

<\/section>\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n
\n

AI\u5bf9\u9f50\uff1a\u5b9e\u9645\u4e2d\u8bd5\u56fe\u89e3\u51b3\u7684\u95ee\u9898<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


<\/p>\n<\/section>\n<\/section>\n<\/section>\n

\u5728\u5f53\u4e0b\u548c\u672a\u6765\uff0c \u201cAI \u5bf9\u9f50\u201c \u6307\u7684\u662f\u8ba9 AI \u7cfb\u7edf<\/span>\uff08\u5305\u62ec\u80fd\u529b\u8d85\u8d8a\u4eba\u7c7b\u7684 AI \u7cfb\u7edf\uff09<\/span>\u7684\u4f18\u5316\u76ee\u6807\u4e0e\u4eba\u7c7b\u610f\u56fe\u3001\u4eba\u7c7b\u4ef7\u503c\u89c2\u76f8\u5339\u914d\uff0c\u800c\u5b83\u6240\u8981\u89e3\u51b3\u7684\u6b63\u662f\u4e0a\u8ff0\u8fd9\u4e9b\u95ee\u9898[1]\u3002\u6240\u6709\u8fd9\u56db\u79cd\u56f0\u96be\uff0c\u5728\u7ed9\u5f53\u4e0b\u548c\u672a\u6765\u7684 AI \u7cfb\u7edf\u8bbe\u5b9a\u76ee\u6807\u65f6\uff0c\u90fd\u5df2\u7ecf\u51fa\u73b0\u6216\u53ef\u80fd\u4f1a\u51fa\u73b0\u3002\u4e00\u4e9b\u4f8b\u5b50\u5982\uff1a<\/span><\/p>\n


<\/p>\n

\n

\u76ee\u6807\u9519\u8bef\u89c4\u8303\uff08Goal Misspecification\uff09\uff1a<\/span><\/strong><\/span><\/p>\n

\u8fd9\u4e00\u7c7b\u7684\u4f8b\u5b50\u5728\u5404\u7c7b\u6e38\u620f\u548c\u4eff\u771f\u73af\u5883\u4e2d\u8f83\u4e3a\u5e38\u89c1\uff0c\u751a\u81f3\u6709\u4e00\u4e2a\u7f51\u9875\u4e13\u95e8\u7ef4\u62a4\u4e86\u4e00\u7cfb\u5217\u6b64\u7c7b\u6848\u4f8b[2]\u3002\u4e00\u4e2a\u8f83\u65b0\u7684\u4f8b\u5b50\u662f\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6076\u8a00\u6076\u8bed\u884c\u4e3a\u2014\u2014\u8fd9\u4e9b\u884c\u4e3a\u5f80\u5f80\u4ece\u8bad\u7ec3\u96c6\u4e2d\u7684\u6076\u8bed\u4e60\u5f97\uff0c\u4f53\u73b0\u4e86\u8bad\u7ec3\u96c6\u8fd9\u4e00\u76ee\u6807\u89c4\u7ea6<\/span>\uff08goal specification\uff09<\/span>\u7684\u4e0d\u7406\u60f3\u4e4b\u5904\u3002\u4e0d\u8fc7\u8fd9\u7c7b\u6076\u8a00\u6076\u8bed\u53ea\u662f\u4ece\u6570\u636e\u4e2d\u88ab\u52a8\u4e60\u5f97\u574f\u884c\u4e3a\uff0c\u5982\u679c\u5927\u8bed\u8a00\u6a21\u578b\u7684\u8bad\u7ec3\u6d41\u7a0b\u88ab\u4e0e\u5f3a\u5316\u5b66\u4e60\u4e00\u7c7b\u7684\u4e3b\u52a8\u89c4\u5212\u548c\u6267\u884c\u80fd\u529b\u7ed3\u5408\uff0c\u5219\u53ef\u80fd\u5e26\u6765\u66f4\u4e3a\u4e3b\u52a8\u7684 \u201c\u94bb\u6f0f\u6d1e\u201d \u884c\u4e3a\u3002\u4e0e\u4e4b\u76f8\u5173\u7684\u53e6\u4e00\u4f8b\u662f\u5927\u8bed\u8a00\u6a21\u578b\u4e2d\u7684\u5956\u52b1\u8fc7\u5ea6\u4f18\u5316<\/span>\uff08Reward Overoptimization\uff09<\/span>\u73b0\u8c61\uff1a\u82e5\u5927\u8bed\u8a00\u6a21\u578b\u7684\u5bf9\u9f50\u8bad\u7ec3\u91c7\u7528\u4e60\u5f97\u7684\u4eba\u7c7b\u504f\u597d\u6a21\u578b\u4e3a\u4f18\u5316\u76ee\u6807\uff0c\u5219\u4f18\u5316\u5230\u4e00\u5b9a\u7a0b\u5ea6\u540e\uff0c\u4e0e\u771f\u5b9e\u4eba\u7c7b\u504f\u597d\u7684\u543b\u5408\u7a0b\u5ea6\u4f1a\u968f\u8bad\u7ec3\u800c\u4e0b\u964d[3]\u2014\u2014\u8fd8\u8bb0\u5f97\u53e4\u5fb7\u54c8\u7279\u5b9a\u5f8b\u5417\uff1f<\/span><\/p>\n


<\/span><\/p>\n<\/section>\n

\n

<\/p>\n<\/section>\n


<\/p>\n

\n

\u76ee\u6807\u9519\u8bef\u6cdb\u5316\uff08Goal Misgeneralization\uff09\uff1a<\/span><\/strong><\/span><\/p>\n

\u4f18\u5316\u76ee\u6807\u7684\u9519\u8bef\u6cdb\u5316\u5728\u6e38\u620f\u548c\u4eff\u771f\u73af\u5883\u4e2d\u4e5f\u65f6\u6709\u51fa\u73b0\uff0c\u4e5f\u540c\u6837\u51fa\u73b0\u4e86\u7ef4\u62a4\u6b64\u7c7b\u6848\u4f8b\u7684\u6570\u636e\u5e93[4]\u3002\u800c\u5728\u5927\u8bed\u8a00\u6a21\u578b\u7684 Scaling Laws \u76db\u6781\u4e00\u65f6\u7684\u73b0\u5728\uff0c\u4e00\u4e2a\u53d7\u5230\u8f83\u591a\u62c5\u5fe7\u7684\u53ef\u80fd\u6027\u662f\u4ece\u201c\u4eba\u7c7b\u53ef\u6709\u6548\u76d1\u7763\u201d\u60c5\u666f\u5230\u201c\u4eba\u7c7b\u96be\u4ee5\u6709\u6548\u76d1\u7763\u201d\u60c5\u666f\u7684\u6cdb\u5316\u8fc7\u7a0b\u4e2d\uff0c\u662f\u5426\u4f1a\u51fa\u73b0\u4f18\u5316\u76ee\u6807\u7684\u8868\u5f81\u7684\u9519\u8bef\u6cdb\u5316\u3002\u6613\u89c1\u76d1\u7763\u8bad\u7ec3\u53ea\u80fd\u53d1\u751f\u5728\u524d\u8005\u60c5\u5883\u4e0b\uff0c\u6545\u540e\u8005\u5c06\u4f1a\u4f9d\u8d56\u4e8e\u5206\u5e03\u5916\u6cdb\u5316\u3002\u5df2\u7ecf\u51fa\u73b0\u7684\u9519\u8bef\u6cdb\u5316\u65e9\u671f\u5f81\u5146\u5305\u62ec\u5927\u8bed\u8a00\u6a21\u578b\u7684\u201c\u8c04\u5a9a\u201d\u503e\u5411\uff0c\u5373\u8fce\u5408\u7528\u6237\u89c2\u70b9\u7684\u503e\u5411\u2014\u2014\u8bc1\u636e\u8868\u660e\u8fd9\u5f88\u53ef\u80fd\u662f\u5728\u5bf9\u9f50\u8bad\u7ec3\u4e2d\uff0c\u4eba\u7c7b\u76d1\u7763\u8005\u6613\u53d7\u8c04\u5a9a\u5f71\u54cd\u7684\u503e\u5411\u6240\u81f4[5]\u3002 <\/span><\/p>\n


<\/p>\n

\u81ea\u81f4\u7684\u5206\u5e03\u504f\u79fb\uff1a<\/span><\/strong><\/span><\/p>\n

\u8fd9\u65b9\u9762\u7684\u7814\u7a76\u5c1a\u5904\u4e8e\u65e9\u671f\u9636\u6bb5[6]\uff0c\u4f46\u4e00\u4e9b\u53d7\u5230\u5173\u6ce8\u7684\u95ee\u9898\u5305\u62ec\u63a8\u8350\u7cfb\u7edf\u4e0e\u7528\u6237\u4e4b\u95f4\u5f62\u6210\u7684\u53cd\u9988\u56de\u8def \uff0c\u548c\u5927\u8bed\u8a00\u6a21\u578b\u4e0e\u4e92\u8054\u7f51\u6587\u672c\u4e4b\u95f4\u5f62\u6210\u7684\u53cd\u9988\u56de\u8def[7]\u3002<\/span><\/p>\n


<\/p>\n

\u9053\u5fb7\u4e0d\u786e\u5b9a\u6027\uff1a<\/span><\/strong><\/span><\/p>\n

\u6709\u5173\u81ea\u52a8\u9a7e\u9a76\u573a\u666f\u4e0b\u7684\u9053\u5fb7\u95ee\u9898\u8ba8\u8bba\u3001AI \u7cfb\u7edf\u4e2d\u504f\u89c1\u4e0e\u6b67\u89c6\u7684\u8ba8\u8bba\uff0c\u90fd\u662f\u8be5\u95ee\u9898\u7684\u5df2\u6709\u6848\u4f8b\u3002\u968f\u7740 AI \u7cfb\u7edf\u5728\u793e\u4f1a\u4e2d\u65e5\u6e10\u5e7f\u6cdb\u7684\u5e94\u7528\uff0cAI \u7684\u9053\u5fb7\u5224\u65ad\u7684\u5f71\u54cd\u8303\u56f4\u548c\u5f71\u54cd\u7a0b\u5ea6\u53ef\u80fd\u8fd8\u5c06\u5927\u5e45\u6269\u5927\u3002\u4ece\u66f4\u9ad8\u7684\u89d2\u5ea6\u770b\uff0c\u4ef7\u503c\u5206\u6b67\u7684\u535a\u5f08\u523b\u753b\u7b49\u95ee\u9898\uff0c\u90fd\u662f\u8fd9\u4e00\u95ee\u9898\u4e2d\u503c\u5f97\u8003\u8651\u7684\u5bf9\u8c61\u3002
<\/span><\/p>\n<\/section>\n


<\/p>\n

\n
\n
\n


<\/p>\n<\/section>\n<\/section>\n

\n
\n

\n
\n
\n
\n
\n

<\/section>\n

<\/section>\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n
\n

AI\u5bf9\u9f50\u7684\u201c\u63a7\u5236\u8bba\u8fdb\u8def\u201d<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


<\/p>\n<\/section>\n<\/section>\n<\/section>\n

\n

\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u7684\u5f3a\u5316\u5b66\u4e60<\/span>\uff08Reinforcement Learning from Human Feedback, RLHF\uff09<\/span>\u662f\u81f3\u4eca\u4e3a\u6b62\uff0cAI \u5bf9\u9f50\u4e2d\u6beb\u65e0\u4e89\u8bae\u7684\u6700\u4e3b\u6d41\u3001\u6700\u6210\u719f\u7684\u7b97\u6cd5\u4e4b\u4e00\u3002\u5b83\u7684\u601d\u8def\u662f\u5148\u4ece\u4eba\u7c7b\u6570\u636e\u4e60\u5f97\u4e00\u4e2a\u4eba\u7c7b\u504f\u597d\u6a21\u578b\uff0c\u518d\u4ee5\u8be5\u504f\u597d\u6a21\u578b\u4e3a\u4f18\u5316\u76ee\u6807\uff0c\u5bf9\u5927\u8bed\u8a00\u6a21\u578b\u7528\u5f3a\u5316\u5b66\u4e60\u4f5c\u5fae\u8c03\u3002<\/span><\/p>\n


<\/p>\n

\u8fd9\u4e00\u7b97\u6cd5\u5176\u5b9e\u4ee3\u8868\u4e86 AI \u5bf9\u9f50\u4e2d\u4e24\u79cd\u4e3b\u8981\u7684\u601d\u8def\u4e4b\u4e00\uff0c\u4e0d\u59a8\u79f0\u4e4b\u4e3a\u201c\u63a7\u5236\u8bba\u8fdb\u8def\u201d\u3002\u8fd9\u79cd\u601d\u8def\u5047\u5b9a\uff0cAI \u7cfb\u7edf\u6240\u771f\u6b63\u5e94\u5bf9\u9f50\u7684\u76ee\u6807\uff0c\u5176\u5bf9\u4eba\u7c7b\u800c\u8a00\u662f\u6e05\u6670\u660e\u4e86\u7684\uff0c\u800c\u95ee\u9898\u4ec5\u5728\u4e8e\u6709\u6548\u5730\u786e\u4fdd\u8fd9\u4e00\u76ee\u6807\u88ab AI \u6240\u6267\u884c\uff0c\u786e\u4fdd\u9519\u8bef\u89c4\u8303\u548c\u9519\u8bef\u6cdb\u5316\u90fd\u4e0d\u4f1a\u53d1\u751f\u3002<\/span><\/p>\n


<\/p>\n

\u8fd9\u4e00\u8fdb\u8def\u7684\u4f18\u52bf\u5728\u4e8e\u5b83\u7684\u7b80\u6d01\u6027\uff0c\u901a\u8fc7\u628a\u95ee\u9898\u7684\u8303\u56f4\u7f29\u5c0f\u800c\u83b7\u5f97\u4e86\u66f4\u9ad8\u7684\u5b9e\u9645\u53ef\u884c\u6027\u2014\u2014RLHF \u8fd9\u4e00\u6700\u6210\u719f\u65b9\u6cd5\u5f52\u5c5e\u4e8e\u8fd9\u4e00\u7c7b\u8fdb\u8def\uff0c\u8fd9\u7edd\u4e0d\u662f\u5de7\u5408\u3002\u4f46\u540c\u65f6\uff0c\u5b83\u4e5f\u5ffd\u7565\u4e86\u4eba\u7c7b\u81ea\u5df1\u5bf9\u4e8e\u76ee\u6807\u548c\u4ef7\u503c\u89c2\u7684\u5206\u6b67\u3001\u4e0d\u786e\u5b9a\u6027\u3001\u968f\u65f6\u95f4\u6f14\u5316\u7b49\u7279\u6027\uff0c\u5e76\u4e14\u628a\u88ab\u63a7\u5236\u8005<\/span>\uff08AI \u7cfb\u7edf\uff09<\/span>\u4e0e\u63a7\u5236\u8005<\/span>\uff08\u4eba\u7c7b\uff09<\/span>\u7f6e\u4e8e\u5bf9\u6297\u7684\u5173\u7cfb\u4e0b\uff0c\u8fd9\u5bf9\u4e8e\u63a7\u5236\u80fd\u529b\u5f3a\u4e8e\u4eba\u7c7b\u7684 AI \u7cfb\u7edf\u662f\u4e0d\u5229\u7684\u3002<\/span><\/p>\n<\/section>\n


<\/p>\n

\n
\n
\n


<\/span><\/p>\n

\n
\n
\n
\n
\n

<\/section>\n

<\/section>\n
\n
<\/svg><\/section>\n<\/section>\n<\/section>\n
\n

AI\u5bf9\u9f50\u7684\u201c\u535a\u5f08\u8bba\u8fdb\u8def\u201d<\/strong><\/span><\/strong><\/strong><\/p>\n<\/section>\n

\n

<\/section>\n

<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n<\/h3>\n


<\/p>\n<\/section>\n<\/section>\n<\/section>\n

\n

\u5408\u4f5c\u9006\u5f3a\u5316\u5b66\u4e60<\/span>\uff08Cooperative Inverse Reinforcement Learning, CIRL\uff09<\/span>\u662f\u53e6\u4e00\u7c7b\u65b9\u6cd5\u4e2d\u7684\u4ee3\u8868[1][8]\u3002\u5b83\u7684\u6838\u5fc3\u601d\u60f3\u662f\uff0c\u628a\u4eba\u7c7b\u4e0e AI \u7cfb\u7edf\u89c6\u4e3a\u540c\u4e00\u73af\u5883\u4e2d\u7684\u4e24\u4e2a\u5e73\u7b49\u884c\u52a8\u8005\uff0c\u4e8c\u8005\u5171\u4eab\u4e00\u4e2a\u76ee\u6807<\/span>\uff08\u5373\u5956\u52b1\u51fd\u6570\uff09<\/span>\uff0c\u4f46\u53ea\u6709\u4eba\u7c7b\u80fd\u83b7\u5f97\u5956\u52b1\u4fe1\u53f7\uff0c\u800c AI \u7cfb\u7edf\u5219\u53ea\u80fd\u4ece\u4eba\u7c7b\u884c\u4e3a\u4e2d\u63a8\u65ad\u5956\u52b1\u51fd\u6570\u7684\u5185\u5bb9 \u2014\u2014 \u5373 \u201c\u4eba\u7c7b\u5230\u5e95\u60f3\u8981\u4ec0\u4e48\u201d \u3002\u5e76\u4e14\uff0c\u56e0\u4e3a AI \u59cb\u7ec8\u6301\u6709\u5bf9\u5956\u52b1\u51fd\u6570\u7684\u4e0d\u786e\u5b9a\u6027\uff0c\u4eba\u7c7b\u4f5c\u4e3a\u4fe1\u606f\u6765\u6e90\u7684\u91cd\u8981\u6027\u610f\u5473\u7740 AI \u8bef\u5bfc\u548c\u64cd\u7eb5\u4eba\u7c7b\u7684\u52a8\u673a\u5c06\u4f1a\u964d\u4f4e<\/span>\uff08\u4f46\u4e0d\u4e00\u5b9a\u6d88\u5931\uff09<\/span>\u3002<\/span><\/p>\n


<\/p>\n

\u8fd9\u4e00\u65b9\u6cd5\uff0c\u672c\u8d28\u4e0a\u662f\u901a\u8fc7\u5c06\u4eba\u7c7b\u4e0e AI \u7cfb\u7edf\u7f6e\u4e8e\u5408\u4f5c\u7684\u5173\u7cfb\u4e2d\uff0c\u4ee5\u51cf\u5c11\u4e8c\u8005\u5bf9\u6297\u7684\u52a8\u673a\u3002<\/span><\/p>\n


<\/p>\n

\u9664\u4e86\u8be5\u65b9\u6cd5\u5916\uff0c\u4e0e\u793e\u4f1a\u9009\u62e9\u7406\u8bba<\/span>\uff08Social Choice Theory\uff09<\/span>\u3001\u535a\u5f08\u8bba\u7b49\u7ed3\u5408\u7684\u4e00\u4e9b\u5176\u4ed6 AI \u5bf9\u9f50\u65b9\u6cd5\uff0c\u5219\u6709\u7740\u4e0d\u540c\u7684\u4f18\u70b9[9]\u3002\u5b83\u4eec\u901a\u8fc7\u663e\u5f0f\u5730\u523b\u753b\u4e0d\u540c\u884c\u52a8\u8005\u4e4b\u95f4\u76ee\u6807\u548c\u4ef7\u503c\u89c2\u7684\u51b2\u7a81\uff0c\u4f7f\u5f97\u6211\u4eec\u53ef\u4ee5\u76f4\u9762\u9053\u5fb7\u4e0d\u786e\u5b9a\u6027\u3001\u590d\u6742\u793e\u4f1a\u4e92\u52a8\u7b49\u56f0\u96be\u95ee\u9898\u3002<\/span><\/p>\n


<\/p>\n

\u53e6\u4e00\u65b9\u9762\uff0c\u8fd9\u7c7b\u65b9\u6cd5\u8f83\u9ad8\u7684\u590d\u6742\u7a0b\u5ea6\uff0c\u4e5f\u610f\u5473\u7740\u5b83\u4eec\u7684\u5de5\u7a0b\u53ef\u5b9e\u73b0\u6027\u4e5f\u5f80\u5f80\u8f83\u4f4e\u3002\u5982\u4f55\u80fd\u5c06\u8fd9\u4e9b\u65b9\u6cd5\u4f7f\u7528\u5728\u5b9e\u9645\u89c4\u6a21\u7684 AI \u5e94\u7528\u4e0a\uff0c\u662f\u4e00\u4e2a\u4e9f\u5f85\u89e3\u51b3\u7684\u95ee\u9898\u3002<\/span><\/p>\n<\/section>\n


<\/p>\n

\u53c2\u8003\u6587\u732e<\/span><\/strong><\/span><\/p>\n

\n

[1] Ji, J. et al. (2023). AI Alignment: A Comprehensive survey. arXiv. https:\/\/doi.org\/10.48550\/arxiv.2310.19852<\/span><\/p>\n

[2] https:\/\/docs.google.com\/spreadsheets\/d\/e\/2PACX-1vRPiprOaC3HsCf5Tuum8bRfzYUiKLRqJmbOoC-32JorNdfyTiRRsR7Ea5eWtvsWzuxo8bjOxCG84dAg\/pubhtml<\/span><\/p>\n

[3] Gao, L., John, S., & Hilton, J. Scaling laws for reward model overoptimization. ICML 2023. https:\/\/dl.acm.org\/doi\/10.5555\/3618408.3618845<\/span><\/p>\n

[4] https:\/\/docs.google.com\/spreadsheets\/d\/e\/2PACX-1vTo3RkXUAigb25nP7gjpcHriR6XdzA_L5loOcVFj_u7cRAZghWrYKH2L2nU4TA_Vr9KzBX5Bjpz9G_l\/pubhtml<\/span><\/p>\n

[5] Perez, E. et al. Discovering Language Model Behaviors with Model-Written Evaluations. Findings of the Association for Computational Linguistics: ACL 2023. https:\/\/doi.org\/10.18653\/v1\/2023.findings-acl.847<\/span><\/p>\n

[6] Krueger, D. A., Maharaj, T., & Leike, J. (2020). Hidden Incentives for Auto-Induced Distributional Shift. arXiv. https:\/\/doi.org\/10.48550\/arxiv.2009.09153<\/span><\/p>\n

[7] Taori, R., & Hashimoto, T. (2022). Data feedback loops: model-driven amplification of dataset biases. arXiv. https:\/\/doi.org\/10.48550\/arxiv.2209.03942<\/span><\/p>\n

[8] Hadfield-Menell, D., Russell, S., Abbeel, P., & Dragan, A. D. Cooperative inverse reinforcement learning. NeurIPS 2016. https:\/\/dl.acm.org\/doi\/10.5555\/3157382.3157535<\/span><\/p>\n

[9] Critch, A., & Krueger, D. W. (2020). AI Research Considerations for Human Existential Safety (ARCHES). arXiv. http:\/\/export.arxiv.org\/pdf\/2006.04948<\/span><\/p>\n<\/section>\n

\n

<\/span><\/p>\n


<\/span><\/strong><\/span><\/section>\n
\n
\n
\n

\u4f5c\u8005\u7b80\u4ecb<\/strong><\/p>\n<\/section>\n<\/section>\n<\/section>\n

\n


<\/p>\n<\/section>\n

<\/p>\n

\u90b1\u5929\u5f02<\/span><\/strong>\u662f\u5317\u4eac\u5927\u5b66\u8ba1\u7b97\u673a\u56fe\u7075\u73ed\u7684\u4e09\u5e74\u7ea7\u672c\u79d1\u751f\uff0c\u5317\u5927\u5bf9\u9f50\u5c0f\u7ec4 (PKU-Alignment) \u6210\u5458\uff0c\u5728\u5317\u5927PAIR-Lab\u8fdb\u884cAI\u5bf9\u9f50\u7814\u7a76\uff0c\u5173\u6ce8\u5bf9\u9f50\u4e0e\u9053\u5fb7\u4ef7\u503c\u7684\u4ea4\u53c9\uff0c\u5e76\u64b0\u5199AI Alignment: A Comprehensive Survey\uff0c\u7efc\u8ff0\u4e3b\u9875\uff1awww.alignmentsurvey.com\u3002\u4ed6\u66fe\u83b7John Hopcroft\u5956\u5b66\u91d1\u3001NOI 2020\u91d1\u724c\u3002<\/span><\/section>\n

<\/span><\/strong><\/span><\/section>\n

<\/section>\n
\n
\n
\n
\n

\u5927\u6a21\u578b\u5b89\u5168\u4e0e\u5bf9\u9f50\u8bfb\u4e66\u4f1a<\/strong>
<\/p>\n<\/section>\n<\/section>\n<\/section>\n<\/section>\n


<\/section>\n

\u5927\u6a21\u578b\u7684\u72c2\u98d9\u7a81\u8fdb\u5524\u9192\u4e86\u4eba\u4eec\u5bf9AI\u6280\u672f\u7684\u70ed\u60c5\u548c\u61a7\u61ac\uff0c\u4e5f\u5f15\u53d1\u4e86\u5bf9AI\u6280\u672f\u672c\u8eab\u5b58\u5728\u7684\u793e\u4f1a\u4f26\u7406\u98ce\u9669\u53ca\u5176\u5bf9\u4eba\u7c7b\u751f\u5b58\u6784\u6210\u7684\u6f5c\u5728\u5a01\u80c1\u7684\u666e\u904d\u62c5\u5fe7\u3002\u5728\u6b64\u80cc\u666f\u4e0b\uff0cAI\u5b89\u5168\u4e0e\u5bf9\u9f50\u5f97\u5230\u5e7f\u6cdb\u5173\u6ce8\uff0c\u8fd9\u662f\u4e00\u4e2a\u81f4\u529b\u4e8e\u8ba9AI\u9020\u798f\u4eba\u7c7b\uff0c\u907f\u514dAI\u6a21\u578b\u5931\u63a7\u6216\u88ab\u6ee5\u7528\u800c\u5bfc\u81f4\u707e\u96be\u6027\u540e\u679c\u7684\u7814\u7a76\u65b9\u5411\u3002\u96c6\u667a\u4ff1\u4e50\u90e8\u548c\u5b89\u8fdcAI\u8054\u5408\u4e3e\u529e\u300c\u5927\u6a21\u578b\u5b89\u5168\u4e0e\u5bf9\u9f50\u300d\u8bfb\u4e66\u4f1a<\/a>\uff0c\u7531\u591a\u4f4d\u6d77\u5185\u5916\u4e00\u7ebf\u7814\u7a76\u8005\u8054\u5408\u53d1\u8d77\uff0c\u65e8\u5728\u6df1\u5165\u63a2\u8ba8AI\u5b89\u5168\u4e0e\u5bf9\u9f50\u6240\u6d89\u53ca\u7684\u6838\u5fc3\u6280\u672f\u3001\u7406\u8bba\u67b6\u6784\u3001\u89e3\u51b3\u8def\u5f84\u4ee5\u53ca\u5b89\u5168\u6cbb\u7406\u7b49\u4ea4\u53c9\u8bfe\u9898\u3002<\/span><\/p>\n


<\/strong><\/span><\/section>\n
<\/span><\/a><\/section>\n

<\/section>\n
\u8be6\u60c5\u8bf7\u89c1\uff1a<\/span>
<\/section>\n
2024\u5f00\u5e74\u8bfb\u4e66\u4f1a\uff1aAI\u5b89\u5168\u4e0e\u5bf9\u9f50\u2014\u2014\u5e94\u5bf9\u524d\u6cbfAI\u5931\u63a7\u4e0e\u6ee5\u7528\u7684\u6280\u672f\u8def\u7ebf<\/a><\/span>
<\/span><\/section>\n

<\/section>\n


<\/span><\/p>\n

\n
\n

\u63a8\u8350\u9605\u8bfb<\/span><\/strong><\/strong><\/p>\n

1. <\/strong>\u5927\u6a21\u578b\u5b89\u5168\u4e0e\u5bf9\u9f50\uff1a\u590d\u6742\u7cfb\u7edf\u89c6\u89d2\u4e0b\u7684AI\u5b89\u5168<\/a><\/strong><\/span><\/strong><\/section>\n
2. <\/strong>\u4e07\u5b57\u957f\u6587\u8be6\u89e3\uff1a\u5927\u6a21\u578b\u65f6\u4ee3AI\u4ef7\u503c\u5bf9\u9f50\u7684\u95ee\u9898\u3001\u5bf9\u7b56\u548c\u5c55\u671b<\/a><\/strong><\/span><\/strong><\/section>\n
3. <\/strong>AI\u5b89\u5168\u524d\u6cbf | \u5927\u6a21\u578b\u8c04\u5a9a\u73b0\u8c61\u3001RLHF\u540e\u95e8\u653b\u51fb\u3001AI4Science\u6a21\u578b\u7684\u6ee5\u7528\u98ce\u9669\u3001\u6001\u52bf\u611f\u77e5\u80fd\u529b\u3001\u8868\u5f81\u5de5\u7a0b<\/a><\/strong><\/span><\/strong><\/section>\n
4. <\/strong><\/span>\u5f20\u6c5f\uff1a\u7b2c\u4e09\u4ee3\u4eba\u5de5\u667a\u80fd\u6280\u672f\u57fa\u7840\u2014\u2014\u4ece\u53ef\u5fae\u5206\u7f16\u7a0b\u5230\u56e0\u679c\u63a8\u7406 | \u96c6\u667a\u5b66\u56ed\u5168\u65b0\u8bfe\u7a0b<\/a><\/strong><\/strong><\/span><\/section>\n
5. <\/strong>\u52a0\u5165\u96c6\u667a\u5b66\u56edVIP\uff0c\u4e00\u6b21\u6027\u83b7\u53d6\u96c6\u667a\u5e73\u53f0\u6240\u6709\u5185\u5bb9\u8d44\u6e90<\/a><\/strong><\/strong><\/strong><\/strong><\/span><\/section>\n
6. <\/strong>\u52a0\u5165\u96c6\u667a\uff0c\u4e00\u8d77\u590d\u6742\uff01<\/strong><\/span><\/a><\/section>\n<\/section>\n


<\/p>\n<\/section>\n


<\/p>\n

\u70b9\u51fb\u201c\u9605\u8bfb\u539f\u6587\u201d\uff0c\u62a5\u540d\u8bfb\u4e66\u4f1a<\/span><\/strong><\/section>\n<\/section>\n<\/section>\n

<\/mp-style-type><\/p>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

\u5bfc\u8bed \u672c\u6587\u662f\u5317\u4eac\u5927\u5b66\u8ba1\u7b97\u673a\u56fe\u7075\u73ed\u672c\u79d1\u751f\u3001\u5317\u5927\u5bf9\u9f50\u5c0f\u7ec4 (PKU-Alignemnt\uff09\u6210\u5458\u90b1\u5929\u5f02\u64b0\u5199\u7684\u6587\u7ae0\uff0c\u4ecb\u7ecd\u4e86AI\u5bf9\u9f50\u95ee\u9898\u7684\u4e24\u79cd\u4e3b\u8981\u601d\u8def\uff1a\u57fa\u4e8e\u4eba\u7c7b\u53cd\u9988\u5f3a\u5316\u5b66\u4e60\uff08\u63a7\u5236\u8bba\u8fdb\u8def\uff09\uff0c\u5408\u4f5c\u9006\u5f3a\u5316\u5b66\u4e60\uff08\u535a\u5f08\u8bba\u8fdb\u8def\uff09\u3002\u90b1\u5929\u5f02\u5728\u5317\u5927PAIR-Lab\u8fdb\u884cAI\u5bf9\u9f50\u7814\u7a76\uff0c\u5173\u6ce8\u5bf9\u9f50\u4e0e\u9053\u5fb7\u4ef7\u503c\u7684\u4ea4\u53c9\uff0c\u5e76\u5408\u4f5c\u64b0\u5199\u7efc\u8ff0\u6587\u7ae0 A…<\/p>\n","protected":false},"author":0,"featured_media":48153,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[1],"tags":[],"special":[],"_links":{"self":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/48159"}],"collection":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=48159"}],"version-history":[{"count":0,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/posts\/48159\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=\/wp\/v2\/media\/48153"}],"wp:attachment":[{"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=48159"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=48159"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=48159"},{"taxonomy":"special","embeddable":true,"href":"https:\/\/swarma.org\/index.php?rest_route=%2Fwp%2Fv2%2Fspecial&post=48159"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}