{"id":6136,"date":"2025-09-08T11:58:24","date_gmt":"2025-09-08T03:58:24","guid":{"rendered":"https:\/\/nullthought.net\/?p=6136"},"modified":"2025-09-08T13:18:56","modified_gmt":"2025-09-08T05:18:56","slug":"%e4%b8%ba%e4%bb%80%e4%b9%88%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b%e4%bc%9a%e4%ba%a7%e7%94%9f%e5%b9%bb%e8%a7%89","status":"publish","type":"post","link":"https:\/\/nullthought.net\/?p=6136","title":{"rendered":"\u4e3a\u4ec0\u4e48\u8bed\u8a00\u6a21\u578b\u4f1a\u4ea7\u751f\u5e7b\u89c9"},"content":{"rendered":"\n<p>\u8bba\u6587<strong><a href=\"https:\/\/cdn.openai.com\/pdf\/d04913be-3f6f-4d2b-b283-ff432ef4aaa5\/why-language-models-hallucinate.pdf\" target=\"_blank\" rel=\"noreferrer noopener\">Why Language Models Hallucinate<\/a><\/strong>\u8bd5\u56fe\u7ed9\u51fa\u4e00\u4e2a\u4ece\u7edf\u8ba1\u5b66\u4e60\u89d2\u5ea6\u201c\u53ef\u8bc1\u660e\u201d\u7684\u89e3\u91ca\uff1a\u4e3a\u4ec0\u4e48\u8bed\u8a00\u6a21\u578b\u4f1a\u4ea7\u751f\u201c\u5e7b\u89c9\u201d\uff08hallucination\uff09\u2014\u2014\u5373\u201c\u770b\u4f3c\u5408\u7406\u4f46\u4e8b\u5b9e\u4e0a\u4e3a\u5047\u7684\u5185\u5bb9\u201d\u3002\u4f5c\u8005\u9996\u5148\u628a\u6ce8\u610f\u529b\u805a\u7126\u4e8e\u201c\u5408\u7406\u6027\u201d\u7684\u6587\u672c\u7a7a\u95f4\uff0c\u627f\u8ba4\u73b0\u5b9e\u4e2d\u8fd8\u5b58\u5728\u660e\u663e\u65e0\u610f\u4e49\u7684\u8f93\u51fa\uff0c\u4f46\u6307\u51fa\u5176\u4e3b\u5b9a\u7406\u5728\u628a\u6837\u672c\u7a7a\u95f4\u5212\u5206\u4e3a\u201c\u65e0\u610f\u4e49\/\u9519\u8bef\/\u6b63\u786e\u201d\u4e09\u7c7b\u540e\u4ecd\u7136\u6210\u7acb\uff0c\u4ece\u800c\u4fdd\u8bc1\u7406\u8bba\u6846\u67b6\u7684\u7a33\u5065\u6027\u3002\u8bba\u6587\u4e5f\u8bf4\u660e\uff0c\u5f00\u653e\u5f0f\u751f\u6210\uff08\u5982\u5199\u4f20\u8bb0\uff09\u540c\u6837\u53ef\u4ee5\u7eb3\u5165\u6846\u67b6\uff1a\u53ea\u8981\u628a\u201c\u5305\u542b\u4e00\u4e2a\u6216\u591a\u4e2a\u9519\u8bef\u7684\u56de\u590d\u201d\u89c6\u4e3a\u9519\u8bef\uff0c\u5e76\u5141\u8bb8\u6309\u9519\u8bef\u6570\u91cf\u8bbe\u5b9a\u201c\u5e7b\u89c9\u5f3a\u5ea6\u201d\u3002<\/p>\n\n\n\n<p>\u8bba\u6587\u4f5c\u8005\u4e3aAdam Tauman Kalai, Ofir Nachum, Santosh S. Vempala, Edwin Zhang\uff0c\u6765\u81eaOpenAI\u548cGeorgia Tech\u3002<\/p>\n\n\n\n<p>\u4e00\u3001\u8bba\u6587\u6838\u5fc3\u601d\u8def\u4e0e\u4e3b\u8981\u8d21\u732e<br>\u6838\u5fc3\u6280\u672f\u8def\u7ebf\u662f\u628a\u201c\u751f\u6210\u662f\u5426\u51fa\u9519\u201d\u7684\u96be\u9898\u5316\u7ea6\u4e3a\u4e00\u4e2a\u4e8c\u5206\u7c7b\u5b66\u4e60\u95ee\u9898\uff1aIs-It-Valid\uff08IIV\uff0c\u201c\u8fd9\u6761\u8f93\u51fa\u662f\u5426\u6709\u6548\uff1f\u201d\uff09\u3002\u76f4\u89c9\u4e0a\uff0c\u80fd\u53ef\u9760\u751f\u6210\u6709\u6548\u8f93\u51fa\u6bd4\u56de\u7b54\u201c\u6709\u6548\/\u65e0\u6548\u201d\u66f4\u96be\u2014\u2014\u56e0\u4e3a\u751f\u6210\u5176\u5b9e\u9690\u542b\u5728\u5bf9\u6bcf\u4e2a\u5019\u9009\u56de\u590d\u8fdb\u884c\u4e00\u6b21\u201c\u662f\u5426\u6709\u6548\u201d\u7684\u5224\u65ad\u3002\u7531\u6b64\uff0c\u4f5c\u8005\u5efa\u7acb\u4e86\u751f\u6210\u9519\u8bef\u7387\u4e0eIIV\u8bef\u5206\u7c7b\u7387\u4e4b\u95f4\u7684\u5b9a\u91cf\u5173\u7cfb\uff1a<br>\uff08\u751f\u6210\u9519\u8bef\u7387\uff09\u2273 2 \u00d7\uff08IIV\u8bef\u5206\u7c7b\u7387\uff09\u3002why-language-models-hallucinate\u2026<br>\u5728\u542b\u6709\u63d0\u793a\u8bcd\uff08prompt\uff09\u7684\u66f4\u4e00\u822c\u60c5\u5f62\u4e0b\uff0c\u4f5c\u8005\u7ed9\u51fa\u4e86\u66f4\u7cbe\u786e\u7684\u4e0d\u7b49\u5f0f\u4e0b\u754c\uff0c\u628a\u201c\u6700\u4f18\u9608\u503c\u5316\u5206\u7c7b\u5668\u7684\u8bef\u5dee\u201d\u3001\u6bcf\u4e2a\u4e0a\u4e0b\u6587\u7684\u6b63\u786e\u7b54\u6848\u6570\u4e0e\u53ef\u9009\u7b54\u6848\u6570\u4e4b\u6bd4\uff0c\u4ee5\u53ca\u4e00\u4e2a\u201c\u6821\u51c6\u8bef\u5dee\u201d\u03b4\u4e00\u5e76\u7eb3\u5165\uff0c\u4ece\u800c\u5f97\u5230\uff1a<br>err \u2265 2\u00b7opt(G) \u2212 (max_c|V_c| \/ min_c|E_c|) \u2212 \u03b4\u3002<br>\u8fd9\u6761\u201c\u5316\u7ea6\u201d\u4f7f\u5f97\u5927\u91cf\u5173\u4e8e\u5206\u7c7b\u8bef\u5dee\u6765\u6e90\u7684\u6210\u719f\u8ba4\u8bc6\uff08\u53ef\u5206\/\u4e0d\u53ef\u5206\u3001\u6b20\u62df\u5408\u3001\u4e0d\u53ef\u5b66\u4e60\u6210\u5206\u7b49\uff09\u53ef\u4ee5\u76f4\u63a5\u8fc1\u79fb\u5230\u751f\u6210\u5f0f\u6a21\u578b\u7684\u9519\u8bef\u673a\u7406\u5206\u6790\u4e0a\u3002<\/p>\n\n\n\n<p>\u4e8c\u3001\u9884\u8bad\u7ec3\u9636\u6bb5\uff1a\u4e3a\u4ec0\u4e48\u9519\u8bef\u201c\u4e0d\u53ef\u907f\u514d\u201d<br>1\uff09\u8bed\u8a00\u6a21\u578b\u4f5c\u4e3a\u5bc6\u5ea6\u4f30\u8ba1\u5668\uff1a\u9884\u8bad\u7ec3\u5f97\u5230\u7684\u57fa\u7840\u6a21\u578bp^\u200b\u8fd1\u4f3c\u8bad\u7ec3\u5206\u5e03p\u3002\u5982\u679c\u53ea\u6c42\u201c\u96f6\u9519\u8bef\u201d\uff0c\u53ef\u4ee5\u8bbe\u8ba1\u9000\u5316\u7b56\u7565\uff08\u6c38\u8fdc\u56de\u7b54\u201c\u6211\u4e0d\u77e5\u9053\u201d\uff09\u6216\u673a\u68b0\u590d\u8ff0\u8bad\u7ec3\u6837\u672c\uff0c\u4f46\u8fd9\u4e9b\u90fd\u8fdd\u80cc\u4e86\u7edf\u8ba1\u5efa\u6a21\u7684\u57fa\u672c\u76ee\u6807\u3002\u7406\u60f3\u7684p^\u200b=p\u53c8\u9700\u8981\u4e0d\u53ef\u627f\u53d7\u7684\u6570\u636e\u91cf\uff0c\u56e0\u6b64\u201c\u826f\u597d\u8bad\u7ec3\u7684\u57fa\u7840\u6a21\u578b\u4ecd\u4f1a\u72af\u67d0\u4e9b\u9519\u8bef\u201d\u5728\u7edf\u8ba1\u4e0a\u662f\u53ef\u9884\u671f\u7684\u3002<br>2\uff09\u201c\u4e0d\u4ec5\u4ec5\u662f\u81ea\u52a8\u8865\u5168\u201d\uff1a\u5206\u6790\u5e76\u4e0d\u4f9d\u8d56\u201c\u4e0b\u4e00\u8bcd\u9884\u6d4b\u201d\u8fc7\u7a0b\u672c\u8eab\uff0c\u800c\u662f\u9002\u7528\u4e8e\u4e00\u822c\u5bc6\u5ea6\u4f30\u8ba1\u3002\u628a\u5e7b\u89c9\u5f52\u548e\u4e8e\u201c\u524d\u7f00\u9009\u5f97\u4e0d\u597d\u201d\u7684\u76f4\u89c9\u5e76\u4e0d\u6210\u7acb\uff1b\u4ece\u7eaf\u7edf\u8ba1\u89c6\u89d2\u770b\uff0c\u9010\u8bcd\u8f93\u51fa\u53ea\u662f\u5b9e\u73b0\u7ec6\u8282\uff0c\u771f\u6b63\u9a71\u52a8\u9519\u8bef\u7684\u662f\u201c\u5c06\u6a21\u578b\u62df\u5408\u5230\u771f\u5b9e\u8bed\u8a00\u5206\u5e03\u201d\u8fd9\u4e00\u672c\u8d28\u7ea6\u675f\u3002<br>3\uff09\u4e3b\u5b9a\u7406\u4e0e\u201c\u6821\u51c6\u201d\u03b4\uff1a\u5728\u542b\u63d0\u793a\u8bcd\u7684\u8bbe\u5b9a\u4e2d\uff0c\u9519\u8bef\u4e0b\u754c\u91cc\u51fa\u73b0\u4e86\u03b4\uff0c\u5b83\u8861\u91cf\u6a21\u578b\u5728\u9608\u503ct=1\/|E|\u5904\u7684\uff08\u5931\uff09\u6821\u51c6\u7a0b\u5ea6\u3002\u4f5c\u8005\u8bc1\u660e\uff0c\u5bf9\u4ea4\u53c9\u71b5\u76ee\u6807E_x\u223cp[\u2212logp^\u200b(x)]\u800c\u8a00\uff0c\u03b4\u7b49\u4e8e\u5bf9\u201c\u628a\u6b63\u4f8b\u6574\u4f53\u7f29\u653es\u500d\u201d\u7684\u635f\u5931\u5173\u4e8es\u5728s=1\u5904\u7684\u5bfc\u6570\u5927\u5c0f\uff1b\u82e5\u03b4\u22600\uff0c\u5219\u8fd8\u53ef\u7ee7\u7eed\u4e0b\u964d\uff0c\u6545\u5728\u53ef\u8fbe\u7684\u5c40\u90e8\u6700\u4f18\u9644\u8fd1\u03b4\u5e94\u5f53\u5f88\u5c0f\u2014\u2014\u8fd9\u89e3\u91ca\u4e86\u8bb8\u591a\u57fa\u7840\u6a21\u578b\u5728\u9884\u8bad\u7ec3\u540e\u201c\u5f80\u5f80\u662f\u6821\u51c6\u7684\u201d\u3002<br>\u5b9e\u8bc1\u4e0a\uff0c\u8bba\u6587\u91cd\u5f15\u4e86GPT-4\u5728\u9009\u62e9\u9898\u4e0a\u7684\u6821\u51c6\u76f4\u65b9\u56fe\uff1a\u57fa\u7840\u6a21\u578b\u6821\u51c6\u826f\u597d\uff0c\u800c\u540e\u8bad\u7ec3\uff08\u5982RL\uff09\u53ef\u80fd\u4f7f\u4e4b\u504f\u79bb\u3002<\/p>\n\n\n\n<p>\u4e09\u3001\u4e0d\u53ef\u5b66\u4e60\u6210\u5206\uff1a\u201c\u4efb\u610f\u4e8b\u5b9e\u201d\u4e0e\u5355\u4f8b\u7387\uff08singleton rate\uff09<br>\u5f53\u6570\u636e\u4e2d\u67d0\u4e9b\u4e8b\u5b9e\u6ca1\u6709\u53ef\u5b66\u4e60\u6a21\u5f0f\uff08\u4f8b\u5982\u751f\u65e5\uff09\uff0c\u5373\u4fbf\u6a21\u578b\u5728\u5206\u5e03\u4e0a\u662f\u6821\u51c6\u7684\uff0c\u4ecd\u4e0d\u53ef\u907f\u514d\u5730\u51fa\u9519\u3002\u4f5c\u8005\u628a\u8fd9\u7c7b\u201c\u4efb\u610f\u4e8b\u5b9e\u201d\u7eb3\u5165IIV\u5316\u7ea6\uff0c\u5e76\u4e25\u683c\u5316\u4e86\u65e9\u671f\u5de5\u4f5c\uff1a\u57fa\u7840\u6a21\u578b\u7684\u5e7b\u89c9\u4e0b\u754c\u81f3\u5c11\u7b49\u4e8e\u8bad\u7ec3\u6570\u636e\u4e2d\u201c\u53ea\u51fa\u73b0\u4e00\u6b21\u7684\u4e8b\u5b9e\u201d\u6240\u5360\u6bd4\u4f8b\uff08\u5355\u4f8b\u7387\uff09\u3002\u4f8b\u5982\u82e520%\u7684\u751f\u65e5\u4e8b\u5b9e\u5728\u9884\u8bad\u7ec3\u8bed\u6599\u91cc\u53ea\u51fa\u73b0\u4e00\u6b21\uff0c\u90a3\u4e48\u4ec5\u5c31\u751f\u65e5\u8fd9\u7c7b\u67e5\u8be2\uff0c\u57fa\u7840\u6a21\u578b\u81f3\u5c11\u4f1a\u6709~20%\u7684\u9519\u8bef\u3002<br>\u4e3a\u8bc4\u4f30\u8fd9\u7c7b\u4e0b\u754c\uff0c\u8bba\u6587\u7ed9\u51fa\u4e86\u57fa\u4e8e\u6539\u9020\u7248Good-Turing\u4f30\u8ba1\u7684\u7ed3\u8bba\uff1a\u7528\u53d8\u4f53\u7edf\u8ba1\u91cfsr\u53ef\u4ee5\u9ad8\u6982\u7387\u903c\u8fd1\u201c\u53ef\u8bb0\u5fc6\u4f46\u975eIDK\u7684\u552f\u4e00\u4e8b\u5b9e\u8d28\u91cf\u201dMM\uff0c\u5e76\u8fdb\u4e00\u6b65\u7ed9\u51fa2\u2009err_iiv\u4e0esr\u4e4b\u95f4\u7684\u4e0b\u754c\u8054\u7cfb\uff0c\u7528\u4ee5\u652f\u6491\u201c\u5355\u4f8b\u7387\u201d\u2192\u201c\u9519\u8bef\u7387\u201d\u7684\u63a8\u65ad\u3002<\/p>\n\n\n\n<p>\u56db\u3001\u5dee\u6a21\u578b\u56e0\u7d20\u4e0e\u201c\u7eaf\u9009\u62e9\u9898\u201d\u8fb9\u754c<br>\u5f53\u6bcf\u4e2a\u4e0a\u4e0b\u6587\u53ea\u6709\u4e00\u4e2a\u6b63\u786e\u7b54\u6848\uff08\u6807\u51c6\u591a\u9009\u9898\u3001\u65e0\u201cIDK\u201d\u9009\u9879\uff09\uff0c\u4f5c\u8005\u8bc1\u660e\u4e86\u66f4\u7d27\u7684\u754c\uff1a<br>err \u2265 2(1\u22121\/C)\u00b7opt(G)\u3002\u4ee5\u4e09\u5143\u8bed\u8a00\u6a21\u578b\u4e3a\u4f8b\uff08\u53ea\u770b\u524d\u4e24\u8bcd\uff09\uff0c\u6784\u9020\u201cher mind \/ his mind\u201d\u7684\u5bf9\u5076\u63d0\u793a-\u7b54\u6848\u5bf9\uff0c\u53ef\u8bc1\u660e\u4efb\u4f55\u4e09\u5143\u6a21\u578b\u7684\u751f\u6210\u9519\u8bef\u7387\u81f3\u5c11\u4e3a1\/2\u3002\u8fd9\u8bf4\u660e\u201c\u8868\u8fbe\u80fd\u529b\u4e0d\u591f\u201d\u7684\u5dee\u6a21\u578b\u4f1a\u7cfb\u7edf\u6027\u5730\u4ea7\u751f\u9519\u8bef\u3002<br>\u8bba\u6587\u8fd8\u4ee5\u201c\u6570\u5b57\u6bcd\u201d\u4f8b\u5b50\u5bf9\u6bd4\u4e86\u63a8\u7406\u578b\u6a21\u578b\uff08DeepSeek-R1\uff09\u4e0e\u975e\u63a8\u7406\u57fa\u5ea7\u5728\u5b57\u7b26\u7c92\u5ea6\u4e0a\u7684\u8868\u5f81\u5dee\u5f02\uff0c\u63d0\u793a\u201c\u63a8\u7406\/\u8868\u5f81\u6539\u8fdb\u201d\u80fd\u7f13\u89e3\u201c\u5dee\u6a21\u578b\u578b\u201d\u9519\u8bef\u3002<\/p>\n\n\n\n<p>\u4e94\u3001\u8ba1\u7b97\u56f0\u96be\u4e0e\u201c\u5185\u5728\u5e7b\u89c9\u201d<br>\u5e76\u975e\u6240\u6709\u4efb\u52a1\u90fd\u53ef\u88ab\u9ad8\u6548\u5b66\u4e60\u3002\u4f5c\u8005\u901a\u8fc7\u4e00\u4e2a\u201c\u52a0\u5bc6\/\u89e3\u5bc6\u201d\u7684\u98ce\u683c\u5316\u4f8b\u5b50\u8bf4\u660e\uff1a\u82e5\u5206\u7c7b\u5668\u4e0d\u80fd\u6253\u7834\u52a0\u5bc6\u5b89\u5168\u6027\uff0c\u5219\u8bed\u8a00\u6a21\u578b\u5728\u8fd9\u7c7b\u4efb\u52a1\u4e0a\u4ee5\u81f3\u5c111\u2212\u03b2\u22122\/(\u2223M\u2223\u22121)\u2212\u03b4\u7684\u6982\u7387\u8f93\u51fa\u9519\u8bef\u89e3\u5bc6\u2014\u2014\u4ece\u800c\u628a\u8ba1\u7b97\u590d\u6742\u6027\u7684\u4e0b\u754c\u8f6c\u5316\u4e3a\u5e7b\u89c9\u7684\u4e0b\u754c\u3002<br>\u8bba\u6587\u4e5f\u660e\u786e\u6307\u51fa\uff0c\u73b0\u5b9e\u4e2d\u5df2\u89c2\u5bdf\u5230AI\u5728\u8ba1\u7b97\u56f0\u96be\u4efb\u52a1\u4e0a\u7684\u7cfb\u7edf\u6027\u9519\u8bef\uff0c\u6240\u4ee5\u201c\u8ba1\u7b97\u786c\u5ea6\u201d\u662f\u5e7b\u89c9\u7684\u6839\u6e90\u4e4b\u4e00\u3002<\/p>\n\n\n\n<p>\u516d\u3001\u540e\u8bad\u7ec3\u9636\u6bb5\uff1a\u4e3a\u4ec0\u4e48\u201c\u8fc7\u5ea6\u81ea\u4fe1\u7684\u5e7b\u89c9\u201d\u4f1a\u5b58\u6d3b<br>\u9884\u8bad\u7ec3\u89e3\u91ca\u4e86\u201c\u4e3a\u4ec0\u4e48\u4f1a\u51fa\u9519\u201d\uff0c\u800c\u540e\u8bad\u7ec3\uff08\u5982\u6307\u4ee4\u5fae\u8c03\u3001RL\uff09\u672c\u5e94\u9f13\u52b1\u201c\u4e0d\u786e\u5b9a\u5c31\u8bf4\u4e0d\u77e5\u9053\u201d\u3002\u4f46\u4f5c\u8005\u7ed9\u51fa\u4e00\u4e2a\u793e\u4f1a-\u6280\u672f\u5408\u6210\u7684\u89e3\u91ca\uff1a\u5f53\u524d\u4e3b\u6d41\u8bc4\u6d4b\u5927\u591a\u662f\u201c\u4e8c\u5143\u8bc4\u5206\u201d\uff08\u5bf9\/\u9519\uff09\uff0c\u5b83\u4eec\u7cfb\u7edf\u6027\u5730\u201c\u5956\u52b1\u731c\u6d4b\u4e0e\u865a\u5f20\u58f0\u52bf\u201d\uff0c\u4ece\u800c\u538b\u5236\u4e86\u201c\u5982\u5b9e\u8868\u8fbe\u4e0d\u786e\u5b9a\u201d\u7684\u5b66\u4e60\u4fe1\u53f7\u3002\u4e8e\u662f\uff0c\u54ea\u6015\u4f60\u5728\u201c\u4e13\u95e8\u7684\u5e7b\u89c9\u8bc4\u6d4b\u201d\u4e0a\u505a\u5f97\u518d\u597d\uff0c\u4e5f\u53ef\u80fd\u56e0\u4e3a\u5728\u201c\u66f4\u4e3b\u6d41\u7684\u4e8c\u5143\u8bc4\u6d4b\u201d\u4e0a\u5403\u4e8f\u800c\u88ab\u653e\u5f03\u3002<\/p>\n\n\n\n<p>\u4e03\u3001\u8bc4\u6d4b\u673a\u5236\u5982\u4f55\u201c\u5956\u52b1\u80e1\u7f16\u201d\uff1aObservation 1 \u4e0e\u4e8c\u5143\u8bc4\u5206<br>Observation 1\u4e25\u8c28\u5730\u523b\u753b\u4e86\u8fd9\u4e2a\u6096\u8bba\uff1a\u5728\u4efb\u4f55\u201c\u4e8c\u5143\u6253\u5206\u201d\u7684\u8bc4\u6d4b\u5206\u5e03\u4e0a\uff0c\u6700\u4f18\u7b56\u7565\u90fd\u7edd\u4e0d\u4f1a\u9009\u62e9\u201c\u5f03\u7b54\/IDK\u201d\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u53ea\u8981\u4e3b\u6307\u6807\u662f\u201c\u5bf9\/\u9519\u201d\uff0c\u8bda\u5b9e\u5730\u62a5\u544a\u4e0d\u786e\u5b9a\u6027\u5c31\u4e0d\u662f\u6700\u4f18\u7b56\u7565\u3002<br>\u4f5c\u8005\u5bf9\u5f53\u524d\u6d41\u884c\u57fa\u51c6\u7684\u201c\u662f\u5426\u4e8c\u5143+\u662f\u5426\u7ed9IDK\u52a0\u5206\u201d\u505a\u4e86\u5c0f\u578b\u7efc\u8ff0\uff0c\u53d1\u73b0\u7edd\u5927\u591a\u6570\u70ed\u95e8\u8bc4\u6d4b\u90fd\u662f\u4e25\u683c\u4e8c\u5143\uff0c\u51e0\u4e4e\u4e0d\u7ed9IDK\u4efb\u4f55\u5206\u6570\uff1b\u66f4\u7cdf\u7cd5\u7684\u662f\uff0c\u4e00\u4e9b\u7528\u8bed\u8a00\u6a21\u578b\u5f53\u88c1\u5224\u7684\u8bc4\u6d4b\u8fd8\u4f1a\u628a\u201c\u770b\u8d77\u6765\u50cf\u5bf9\u7684\u957f\u7b54\u6848\u201d\u8bef\u5224\u4e3a\u6b63\u786e\uff0c\u4ece\u673a\u5236\u5c42\u9762\u8fdb\u4e00\u6b65\u9f13\u52b1\u4e86\u5e7b\u89c9\u3002<br>\u6b64\u5916\uff0c\u201c\u68c0\u7d22\u589e\u5f3a\uff08RAG\uff09\/\u641c\u7d22\u201d\u5e76\u975e\u4e07\u7075\u836f\uff1a\u5f53\u68c0\u7d22\u4e0d\u5230\u8db3\u591f\u8bc1\u636e\u65f6\uff0c\u4e8c\u5143\u6253\u5206\u4ecd\u7136\u5956\u52b1\u731c\u6d4b\uff1b\u800c\u50cf\u201c\u6570\u82f1\u6587\u5b57\u6bcd\u201d\u8fd9\u6837\u7684\u5185\u5728\u8ba1\u7b97\u9519\u8bef\uff0c\u68c0\u7d22\u4e5f\u5e2e\u4e0d\u4e0a\u5fd9\u3002<\/p>\n\n\n\n<p>\u516b\u3001\u6539\u8fdb\u65b9\u5411\uff1a\u663e\u5f0f\u7f6e\u4fe1\u9608\u503c\u4e0e\u201c\u884c\u4e3a\u6821\u51c6\u201d<br>\u8bba\u6587\u5efa\u8bae\u5728\u4e3b\u6d41\u8bc4\u6d4b\u8bf4\u660e\u4e2d\u52a0\u5165\u201c\u663e\u5f0f\u7f6e\u4fe1\u9608\u503c\u201dt\uff1a\u53ea\u6709\u5f53\u4f60\u81ea\u4fe1\u8d85\u8fc7t\u65f6\u624d\u5e94\u4f5c\u7b54\uff1b\u82e5\u7b54\u9519\uff0c\u5c06\u6309t\/(1\u2212t)\u7684\u6bd4\u4f8b\u6263\u5206\uff1b\u7b54\u5bf9\u5f971\u5206\uff1b\u56de\u7b54\u201c\u6211\u4e0d\u77e5\u9053\u201d\u5f970\u5206\u3002t\u53ef\u53d60.5\/0.75\/0.9\u7b49\u3002\u8fd9\u4e00\u6539\u52a8\u628a\u201c\u5982\u5b9e\u8868\u8fbe\u4e0d\u786e\u5b9a\u6027\u201d\u4ece\u201c\u7b56\u7565\u4e0a\u4e0d\u4f18\u201d\u6539\u9020\u6210\u201c\u7b56\u7565\u4e0a\u6700\u4f18\u201d\uff0c\u4ece\u6839\u4e0a\u6291\u5236\u8fc7\u5ea6\u81ea\u4fe1\u7684\u5e7b\u89c9\u3002<br>\u8fdb\u4e00\u6b65\u5730\uff0c\u4f5c\u8005\u63d0\u51fa\u201c\u884c\u4e3a\u6821\u51c6\u201d\uff08behavioral calibration\uff09\uff1a\u5728\u4e0d\u540ct\u9608\u503c\u4e0b\u6bd4\u8f83\u51c6\u786e\u7387\u4e0e\u9519\u8bef\u7387\uff0c\u4e0d\u8981\u6c42\u6a21\u578b\u62a5\u6570\u503c\u7f6e\u4fe1\u5ea6\uff0c\u800c\u662f\u8981\u6c42\u201c\u5728\u81f3\u5c11t\u81ea\u4fe1\u65f6\u624d\u8bf4\u51fa\u90a3\u6761\u6700\u6709\u7528\u7684\u56de\u590d\u201d\u3002\u8fd9\u65e2\u4fbf\u4e8e\u5ba1\u8ba1\uff0c\u53c8\u80fd\u7ed5\u5f00\u201c\u6b63\u786e\u7b54\u6848\u8868\u8fbe\u65b9\u5f0f\u5448\u6307\u6570\u7ea7\u591a\u6837\u201d\u7684\u8bc4\u6d4b\u96be\u9898\u3002<br>\u4f5c\u8005\u8fd8\u6307\u51fa\uff1a\u5982\u679c\u8bc4\u6d4b\u4e0d\u58f0\u660et\uff0c\u5c31\u5b58\u5728\u201c\u51c6\u786e\u7387-\u9519\u8bef\u7387\u201d\u7684\u56fa\u6709\u6743\u8861\u2014\u2014\u6ca1\u6709\u4e00\u4e2a\u5355\u4e00\u6a21\u578b\u80fd\u5728\u6240\u6709\u9690\u542b\u9608\u503c\u4e0a\u6700\u4f18\uff08\u9664\u975e\u5b83\u6c38\u8fdc\u6b63\u786e\uff09\u3002<\/p>\n\n\n\n<p>\u4e5d\u3001\u8ba8\u8bba\u4e0e\u5c40\u9650<br>\u8bba\u6587\u627f\u8ba4\u5176\u7edf\u8ba1\u6846\u67b6\u4e3a\u7b80\u5316\u8d77\u89c1\u5ffd\u7565\u4e86\u82e5\u5e72\u7ef4\u5ea6\uff1a\u6bd4\u5982\u5bf9\u201c\u65e0\u610f\u4e49\u8f93\u51fa\u201d\u7684\u5904\u7406\u3001\u5f00\u653e\u5f0f\u751f\u6210\u4e2d\u201c\u591a\u9519\u8bef\u8ba1\u6570\u201d\u3001\u63d0\u793a\u4e4b\u5916\u7684\u201c\u6f5c\u5728\u4e0a\u4e0b\u6587\u6b67\u4e49\u201d\u3001\u4ee5\u53ca\u201c\u6b63\u786e\/\u9519\u8bef\/IDK\u201d\u4e09\u5206\u6cd5\u7684\u7c97\u7cd9\u6027\u3002\u4f5c\u8005\u540c\u65f6\u63d0\u9192\uff1a\u9664\u4e86\u663e\u5f0fIDK\uff0c\u73b0\u5b9e\u4e2d\u8fd8\u6709\u201c\u6a21\u7cca\u8868\u8fbe\u3001\u5c11\u62a5\u7ec6\u8282\u3001\u53cd\u95ee\u6c42\u8bc1\u201d\u7b49\u66f4\u7ec6\u7684\u201c\u4e0d\u786e\u5b9a\u6027\u8868\u8fbe\u201d\u624b\u6bb5\uff0c\u8bed\u8a00\u5b66\uff08\u8bed\u7528\u5b66\uff09\u5173\u4e8e\u201c\u5982\u4f55\u7528\u8bed\u8a00\u4f20\u8fbe\u786e\u5b9a\u6027\/\u610f\u56fe\u201d\u7684\u7814\u7a76\u5bf9\u6539\u8fdb\u6a21\u578b\u8868\u8fbe\u81f3\u5173\u91cd\u8981\u3002<\/p>\n\n\n\n<p>\u5341\u3001\u7ed9\u6a21\u578b\u8bad\u7ec3\u4e0e\u8bc4\u6d4b\/\u4ea7\u54c1\u56e2\u961f\u7684\u5b9e\u8df5\u8981\u70b9<br>1\uff09\u628a\u201cIDK\/\u4e0d\u4f5c\u7b54\u201d\u53d8\u6210\u201c\u8bc4\u6d4b\u4e0a\u7684\u7406\u6027\u9009\u62e9\u201d\u800c\u4e0d\u662f\u201c\u88ab\u60e9\u7f5a\u7684\u9009\u62e9\u201d\uff0c\u5728\u4e3b\u57fa\u51c6\u4e2d\u5f15\u5165\u663e\u5f0f\u7f6e\u4fe1\u9608\u503c\u4e0e\u76f8\u5e94\u6263\u5206\u89c4\u5219\u3002<br>2\uff09\u5728\u6c47\u62a5\u6a21\u578b\u80fd\u529b\u65f6\uff0c\u914d\u5408\u5c55\u793a\u201c\u884c\u4e3a\u6821\u51c6\u66f2\u7ebf\u201d\uff08\u4e0d\u540ct\u4e0b\u7684\u51c6\u786e-\u9519\u8bef\u6743\u8861\uff09\uff0c\u4ee5\u907f\u514d\u5355\u4e00\u4e8c\u5143\u6307\u6807\u63a9\u76d6\u8bda\u5b9e\u4e0d\u786e\u5b9a\u6027\u7684\u4ef7\u503c\u3002<br>3\uff09\u533a\u5206\u4e09\u79cd\u8bef\u5dee\u6765\u6e90\u5e76\u5bf9\u75c7\u4e0b\u836f\uff1a<br>\u2014\u2014\u201c\u4efb\u610f\u4e8b\u5b9e\u201d\u2192\u7528\u6570\u636e\u7b56\u5c55\/\u8986\u76d6\u7387\u5ea6\u91cf\uff08\u5982\u5355\u4f8b\u7387\uff09\u4e0e\u8bb0\u5fc6\u6a21\u5757\uff1b<br>\u2014\u2014\u201c\u5dee\u6a21\u578b\u201d\u2192\u6539\u8fdb\u8868\u5f81\u4e0e\u63a8\u7406\uff08\u5b57\u7b26\/\u7ed3\u6784\u5bf9\u9f50\u3001\u8fc7\u7a0b\u76d1\u7763\uff09\uff1b<br>\u2014\u2014\u201c\u8ba1\u7b97\u56f0\u96be\u201d\u2192\u8bbe\u5b9a\u5408\u7406\u4efb\u52a1\u8fb9\u754c\u4e0eIDK\u7b56\u7565\uff0c\u52ff\u671f\u671b\u5728\u4e0d\u53ef tractable \u95ee\u9898\u4e0a\u96f6\u5e7b\u89c9\u3002<br>4\uff09\u5bf9RAG\u7684\u6b63\u786e\u9884\u671f\uff1a\u5b83\u80fd\u964d\u5e7b\u89c9\uff0c\u4f46\u5728\u68c0\u7d22\u5931\u8d25\u6216\u5185\u5728\u8ba1\u7b97\u9519\u8bef\u4e0a\u65e0\u80fd\u4e3a\u529b\uff1b\u8bc4\u6d4b\u4ecd\u9700\u9f13\u52b1\u201c\u68c0\u7d22\u5931\u8d25\u2192IDK\u201d\u7684\u6700\u4f18\u7b56\u7565\u3002<\/p>\n\n\n\n<p>\u5341\u4e00\u3001\u7ed3\u8bba<br>\u8bba\u6587\u628a\u201c\u8bed\u8a00\u6a21\u578b\u4e3a\u4ec0\u4e48\u4f1a\u5e7b\u89c9\u201d\u4ece\u7ecf\u9a8c\u53d9\u4e8b\u8f6c\u5316\u4e3a\u53ef\u8bc1\u660e\u7684\u7edf\u8ba1\u7ed3\u8bba\uff1a\u9884\u8bad\u7ec3\u9636\u6bb5\uff0c\u6765\u81ea\u201c\u4e0d\u53ef\u5b66\u4e60\u6210\u5206\/\u6570\u636e\u8986\u76d6\u4e0d\u8db3\/\u8868\u5f81\u53d7\u9650\/\u8ba1\u7b97\u56f0\u96be\u201d\u7684\u5206\u7c7b\u8bef\u5dee\u5fc5\u7136\u5728\u751f\u6210\u7aef\u663e\u5316\uff1b\u540e\u8bad\u7ec3\u9636\u6bb5\uff0c\u53ea\u8981\u4e3b\u6d41\u8bc4\u6d4b\u575a\u6301\u201c\u4e8c\u5143\u3001\u4e0d\u7ed9IDK\u5206\u201d\uff0c\u673a\u5236\u4e0a\u5c31\u4f1a\u5956\u52b1\u8fc7\u5ea6\u81ea\u4fe1\u3002\u89e3\u51b3\u4e4b\u9053\u4e0d\u662f\u4ec5\u589e\u52a0\u201c\u5e7b\u89c9\u4e13\u9879\u8bc4\u6d4b\u201d\uff0c\u800c\u662f\u628a\u201c\u663e\u5f0f\u7f6e\u4fe1\u9608\u503c\u4e0e\u884c\u4e3a\u6821\u51c6\u201d\u690d\u5165\u4e3b\u8bc4\u6d4b\u4e0e\u4ea7\u54c1\u6307\u6807\uff0c\u628a\u201c\u8bda\u5b9e\u7684\u4e0d\u786e\u5b9a\u6027\u8868\u8fbe\u201d\u53d8\u6210\u6700\u4f18\u7b56\u7565\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u8bba\u6587Why Language Models Hallucinate\u8bd5\u56fe\u7ed9\u51fa\u4e00\u4e2a\u4ece\u7edf\u8ba1\u5b66\u4e60\u89d2\u5ea6\u201c\u53ef\u8bc1\u660e\u201d\u7684\u89e3\u91ca\uff1a [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"set","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[8],"tags":[39,86],"class_list":["post-6136","post","type-post","status-publish","format-standard","hentry","category-tech","tag-ai","tag-llm"],"rttpg_featured_image_url":null,"rttpg_author":{"display_name":"NullThought","author_link":"https:\/\/nullthought.net\/?author=1"},"rttpg_comment":0,"rttpg_category":"<a href=\"https:\/\/nullthought.net\/?cat=8\" rel=\"category\">Tech<\/a>","rttpg_excerpt":"\u8bba\u6587Why Language Models Hallucinate\u8bd5\u56fe\u7ed9\u51fa\u4e00\u4e2a\u4ece\u7edf\u8ba1\u5b66\u4e60\u89d2\u5ea6\u201c\u53ef\u8bc1\u660e\u201d\u7684\u89e3\u91ca\uff1a&hellip;","_links":{"self":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6136","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=6136"}],"version-history":[{"count":2,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6136\/revisions"}],"predecessor-version":[{"id":6138,"href":"https:\/\/nullthought.net\/index.php?rest_route=\/wp\/v2\/posts\/6136\/revisions\/6138"}],"wp:attachment":[{"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=6136"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=6136"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/nullthought.net\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=6136"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}