{"id":243206,"date":"2025-09-09T16:44:49","date_gmt":"2025-09-09T08:44:49","guid":{"rendered":"https:\/\/www.hkubs.hku.hk\/media\/a-new-in-depth-report-of-ai-large-language-models-hallucination-control\/"},"modified":"2025-09-09T16:58:28","modified_gmt":"2025-09-09T08:58:28","slug":"a-new-in-depth-report-of-ai-large-language-models-hallucination-control","status":"publish","type":"hkubs-media","link":"https:\/\/www.hkubs.hku.hk\/tc\/media\/press-release\/a-new-in-depth-report-of-ai-large-language-models-hallucination-control\/","title":{"rendered":"\u6e2f\u5927\u7d93\u7ba1\u5b78\u9662\u6700\u65b0\u767c\u4f48: AI\u5927\u8a9e\u8a00\u6a21\u578b\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u6df1\u5ea6\u8a55\u6e2c\u5831\u544a"},"content":{"rendered":"

AI<\/strong>\u5728\u5c08\u696d\u5834\u666f\u4e2d\u7684\u53ef\u4fe1\u5ea6\u8abf\u67e5<\/strong><\/h2>\n

\u6e2f\u5927\u7d93\u7ba1\u5b78\u9662\u4eca\u65e5\u767c\u8868\u300a\u5927\u8a9e\u8a00\u6a21\u578b\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u6e2c\u8a55\u5831\u544a\u300b<\/strong>\uff0c\u91dd\u5c0d\u9078\u5b9a\u7684AI\u5927\u8a9e\u8a00\u6a21\u578b\uff08LLM\uff09\u8a55\u4f30\u5176\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u3002\u300c\u5e7b\u89ba\u300d\u662f\u6307AI\u6703\u8f38\u51fa\u770b\u4f3c\u5408\u7406\u537b\u8207\u4e8b\u5be6\u76f8\u6096\u6216\u504f\u96e2\u8a9e\u5883\u7684\u5167\u5bb9\u3002\u73fe\u6642\uff0c\u5927\u8a9e\u8a00\u6a21\u578b\uff08LLM\uff09\u6b63\u7a4d\u6975\u65bc\u77e5\u8b58\u670d\u52d9\u3001\u667a\u80fd\u5c0e\u822a\u3001\u5ba2\u6236\u670d\u52d9\u7b49\u5c08\u696d\u61c9\u7528\u5834\u666f\u4e2d\u63a8\u9032\u3002 \u7136\u800c\uff0c\u300c\u5e7b\u89ba\u300d\u73fe\u8c61\u4ecd\u662f\u5236\u7d04\u5176\u53ef\u4fe1\u5ea6\u7684\u95dc\u9375\u74f6\u9838\u3002<\/p>\n

\u8a72\u7814\u7a76\u7531\u6e2f\u5927\u7d93\u7ba1\u5b78\u9662\u5275\u65b0\u53ca\u8cc7\u8a0a\u7ba1\u7406\u5b78\u6559\u6388\u517c\u590f\u5229\u840a\u4f09\u5137\u57fa\u91d1\u6559\u6388<\/strong> \uff08\u6230\u7565\u4fe1\u606f\u7ba1\u7406\u5b78\uff09\u8523\u93ae\u8f1d<\/strong>\u6559\u6388<\/strong>\u7387\u9818\u4eba\u5de5\u667a\u80fd\u8a55\u6e2c\u5be6\u9a57\u5ba4\uff08AIEL\uff09\uff08https:\/\/hkubs.hku.hk\/aimodelrankings<\/a><\/span>\uff09\u7814\u7a76\u4eba\u54e1\u767c\u8d77\uff0c\u91dd\u5c0d 37\u500b\u4e2d\u7f8e\u5927\u8a9e\u8a00\u6a21\u578b\uff08\u5305\u62ec20\u500b\u901a\u7528\u6a21\u578b\u300115\u500b\u63a8\u7406\u6a21\u578b\u53ca2\u500b\u4e00\u9ad4\u5316\u7cfb\u7d71\uff09\u5728\u4e2d\u6587\u8a9e\u5883\u4e0b\u7684\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u958b\u5c55\u5c08\u696d\u8a55\u6e2c\uff0c\u63ed\u793a\u4e0d\u540c\u6a21\u578b\u5728\u898f\u907f\u4e8b\u5be6\u932f\u8aa4\u8207\u4fdd\u6301\u8a9e\u5883\u4e00\u81f4\u6027\u65b9\u9762\u7684\u771f\u5be6\u8868\u73fe\u3002<\/p>\n

\u8a55\u6e2c\u7d50\u679c<\/u><\/strong>\u986f\u793a\uff0cGPT-5\uff08\u601d\u8003\u6a21\u5f0f\uff09\u548cGPT-5\uff08\u81ea\u52d5\u6a21\u5f0f\uff09\u5206\u5225\u4f4d\u5217\u7b2c\u4e00\u7b2c\u4e8c\uff0cClaude 4 Opus\u7cfb\u5217\u7dca\u96a8\u5176\u5f8c\uff1b\u5b57\u7bc0\u8df3\u52d5\u516c\u53f8\u7684\u8c46\u53051.5 Pro\u7cfb\u5217\u8868\u73fe\u7a81\u51fa\uff0c\u5728\u4e2d\u570b\u5927\u8a9e\u8a00\u6a21\u578b\u9663\u71df\u4e2d\u8655\u65bc\u9818\u5148\u5730\u4f4d\uff0c\u4f46\u8207\u570b\u969b\u9802\u5c16\u6a21\u578b\u4e4b\u9593\u4ecd\u5b58\u5728\u986f\u8457\u5dee\u8ddd\u3002<\/p>\n

\u8523\u93ae\u8f1d\u6559\u6388\u8868\u793a<\/strong>\uff1a<\/strong>\u300c\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u4f5c\u70ba\u8861\u91cf\u6a21\u578b\u8f38\u51fa\u771f\u5be6\u6027\u8207\u53ef\u9760\u6027\u7684\u6838\u5fc3\u6307\u6a19\uff0c\u76f4\u63a5\u5f71\u97ff\u5927\u8a9e\u8a00\u6a21\u578b\u5728\u5c08\u696d\u5834\u666f\u4e2d\u7684\u53ef\u4fe1\u5ea6\u3002 \u8a72\u7814\u7a76\u70ba\u672a\u4f86\u6a21\u578b\u7684\u512a\u5316\u63d0\u4f9b\u4e86\u660e\u78ba\u65b9\u5411\uff0c\u63a8\u52d5AI\u5f9e\u300e\u80fd\u751f\u6210\u300f\u5411\u66f4\u95dc\u9375\u7684\u300e\u53ef\u4fe1\u8cf4\u300f\u9081\u9032\u3002 \u300d<\/p>\n

\u8a55\u6e2c\u65b9\u6cd5<\/u><\/strong><\/h5>\n

\u6839\u64da\u6a21\u578b\u751f\u6210\u5167\u5bb9\u5728\u4e8b\u5be6\u4f9d\u64da\u6216\u8a9e\u5883\u5951\u5408\u5ea6\u4e0a\u5b58\u5728\u7684\u554f\u984c\uff0c\u8a72\u7814\u7a76\u5c07\u300c\u5e7b\u89ba\u300d\u5206\u70ba\u5169\u985e\uff1a\u4e8b\u5be6\u6027\u5e7b\u89ba\u548c\u5fe0\u5be6\u6027\u5e7b\u89ba\u3002<\/p>\n

\u4e8b\u5be6\u6027\u5e7b\u89ba<\/strong>\u662f\u6307\u6a21\u578b\u8f38\u51fa\u5167\u5bb9\u8207\u771f\u5be6\u4e16\u754c\u8cc7\u8a0a\u4e0d\u7b26\uff0c\u65e2\u5305\u62ec\u5c0d\u5df2\u77e5\u77e5\u8b58\u7684\u932f\u8aa4\u8abf\u7528\uff08\u5982\u5f35\u51a0\u674e\u6234\u3001\u6578\u64da\u932f\u8a18\uff09\uff0c\u4e5f\u5305\u62ec\u5c0d\u672a\u77e5\u8cc7\u8a0a\u7684\u865b\u69cb\uff08\u5982\u7de8\u9020\u672a\u9a57\u8b49\u7684\u4e8b\u4ef6\u3001\u6578\u64da\uff09\u3002 \u7814\u7a76\u900f\u904e\u8a0a\u606f\u6aa2\u7d22\u985e\u554f\u984c\u3001\u865b\u5047\u4e8b\u5be6\u8b58\u5225\u985e\u554f\u984c\uff0c\u4ee5\u53ca\u77db\u76fe\u524d\u63d0\u8b58\u5225\u985e\u554f\u984c\uff0c\u6aa2\u6e2c\u4e0d\u540c\u6a21\u578b\u91dd\u5c0d\u4e8b\u5be6\u6027\u5e7b\u89ba\u7684\u63a7\u5236\u80fd\u529b\u3002<\/p>\n

\u5fe0\u5be6\u6027\u5e7b\u89ba<\/strong>\u662f\u6307\u6a21\u578b\u672a\u80fd\u56b4\u683c\u9075\u5faa\u4f7f\u7528\u8005\u6307\u4ee4\uff0c\u6216\u8f38\u51fa\u5167\u5bb9\u8207\u8f38\u5165\u4e0a\u4e0b\u6587\u77db\u76fe\uff0c\u5305\u62ec\u907a\u6f0f\u95dc\u9375\u8981\u6c42\u3001\u904e\u5ea6\u5f15\u7533\u3001\u683c\u5f0f\u932f\u8aa4\u7b49\u3002 \u7814\u7a76\u900f\u904e\u6307\u4ee4\u4e00\u81f4\u6027\u8a55\u4f30\u554f\u984c\u53ca\u4e0a\u4e0b\u6587\u4e00\u81f4\u6027\u8a55\u4f30\u554f\u984c\uff0c\u6aa2\u6e2c\u4e0d\u540c\u6a21\u578b\u91dd\u5c0d\u5fe0\u5be6\u6027\u5e7b\u89ba\u7684\u63a7\u5236\u80fd\u529b\u3002<\/p>\n

\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u68af\u968a<\/u><\/strong><\/h5>\n

\u6839\u64da\u8a55\u6e2c\u7d50\u679c\uff0cGPT-5\uff08\u601d\u8003\u6a21\u5f0f\uff09\u548cGPT-5\uff08\u81ea\u52d5\u6a21\u5f0f\uff09\u5206\u5225\u4f4d\u5c45\u7b2c\u4e00\u540d\u548c\u7b2c\u4e8c\u540d\uff0cClaude 4 Opus\u7cfb\u5217\u7dca\u96a8\u5176\u5f8c\uff1b\u5b57\u7bc0\u8df3\u52d5\u516c\u53f8\u7684\u8c46\u53051.5 Pro\u7cfb\u5217\u5728\u4e2d\u570b\u5927\u8a9e\u8a00\u6a21\u578b\u9663\u71df\u4e2d\u8868\u73fe\u7a81\u51fa\uff0c\u4f46\u8207\u570b\u969b\u9802\u5c16\u6a21\u578b\u4e4b\u9593\u4ecd\u5b58\u5728\u986f\u8457\u5dee\u8ddd\u3002<\/p>\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n
\u6392\u540d<\/strong><\/td>\n\u6a21\u578b\u540d\u7a31<\/strong><\/td>\n\u4e8b\u5be6\u6027\u5e7b\u89ba<\/strong><\/td>\n\u5fe0\u5be6\u6027\u5e7b\u89ba<\/strong><\/td>\n\u6700\u7d42\u5f97\u5206<\/strong><\/td>\n<\/tr>\n
1<\/td>\nGPT 5\uff08\u601d\u8003\u6a21\u5f0f\uff09<\/td>\n72<\/td>\n100<\/td>\n86<\/td>\n<\/tr>\n
2<\/td>\nGPT 5\uff08\u81ea\u52d5\u6a21\u5f0f\uff09<\/td>\n68<\/td>\n100<\/td>\n84<\/td>\n<\/tr>\n
3<\/td>\nClaude 4 Opus\uff08\u601d\u8003\u6a21\u5f0f\uff09<\/td>\n73<\/td>\n92<\/td>\n83<\/td>\n<\/tr>\n
4<\/td>\nClaude 4 Opus<\/td>\n64<\/td>\n96<\/td>\n80<\/td>\n<\/tr>\n
5<\/td>\nGrok 4<\/td>\n71<\/td>\n80<\/td>\n76<\/td>\n<\/tr>\n
6<\/td>\nGPT-o3<\/td>\n49<\/td>\n100<\/td>\n75<\/td>\n<\/tr>\n
7<\/td>\n\u8c46\u53051.5 Pro<\/td>\n57<\/td>\n88<\/td>\n73<\/td>\n<\/tr>\n
8<\/td>\n\u8c46\u53051.5 Pro\uff08\u601d\u8003\u6a21\u5f0f\uff09<\/td>\n60<\/td>\n84<\/td>\n72<\/td>\n<\/tr>\n
9<\/td>\nGemini 2.5 Pro<\/td>\n57<\/td>\n84<\/td>\n71<\/td>\n<\/tr>\n
10<\/td>\nGPT-o4 mini<\/td>\n44<\/td>\n96<\/td>\n70<\/td>\n<\/tr>\n
11<\/td>\nGPT-4.1<\/td>\n59<\/td>\n80<\/td>\n69<\/td>\n<\/tr>\n
12<\/td>\nGPT-4o<\/td>\n53<\/td>\n80<\/td>\n67<\/td>\n<\/tr>\n
12<\/td>\nGemini 2.5 Flash<\/td>\n49<\/td>\n84<\/td>\n67<\/td>\n<\/tr>\n
14<\/td>\n\u6587\u5fc3\u4e00\u8a00 X1-Turbo<\/td>\n47<\/td>\n84<\/td>\n65<\/td>\n<\/tr>\n
14<\/td>\n\u901a\u7fa9\u5343\u554f3\uff08\u601d\u8003\u6a21\u5f0f\uff09<\/td>\n55<\/td>\n76<\/td>\n65<\/td>\n<\/tr>\n
14<\/td>\nDeepSeek-V3<\/td>\n49<\/td>\n80<\/td>\n65<\/td>\n<\/tr>\n
14<\/td>\n\u6df7\u5143-T1<\/td>\n49<\/td>\n80<\/td>\n65<\/td>\n<\/tr>\n
18<\/td>\nKimi<\/td>\n47<\/td>\n80<\/td>\n63<\/td>\n<\/tr>\n
18<\/td>\n\u901a\u7fa9\u5343\u554f3<\/td>\n51<\/td>\n76<\/td>\n63<\/td>\n<\/tr>\n
20<\/td>\nDeepSeek-R1<\/td>\n52<\/td>\n68<\/td>\n60<\/td>\n<\/tr>\n
20<\/td>\nGrok 3<\/td>\n36<\/td>\n84<\/td>\n60<\/td>\n<\/tr>\n
20<\/td>\n\u6df7\u5143-TurboS<\/td>\n44<\/td>\n76<\/td>\n60<\/td>\n<\/tr>\n
23<\/td>\n\u65e5\u65e5\u65b0 V6 Pro<\/td>\n41<\/td>\n76<\/td>\n59<\/td>\n<\/tr>\n
24<\/td>\nGLM-4-plus<\/td>\n35<\/td>\n80<\/td>\n57<\/td>\n<\/tr>\n
25<\/td>\nMiniMax-01<\/td>\n31<\/td>\n80<\/td>\n55<\/td>\n<\/tr>\n
25<\/td>\n360\u667a\u81662-o1<\/td>\n49<\/td>\n60<\/td>\n55<\/td>\n<\/tr>\n
27<\/td>\nYi- Lightning<\/td>\n28<\/td>\n80<\/td>\n54<\/td>\n<\/tr>\n
28<\/td>\nGrok 3\uff08\u601d\u8003\u6a21\u5f0f\uff09<\/td>\n29<\/td>\n76<\/td>\n53<\/td>\n<\/tr>\n
29<\/td>\nKimi-k1.5<\/td>\n36<\/td>\n68<\/td>\n52<\/td>\n<\/tr>\n
30<\/td>\n\u6587\u5fc3\u4e00\u8a004.5-Turbo<\/td>\n31<\/td>\n72<\/td>\n51<\/td>\n<\/tr>\n
30<\/td>\n\u65e5\u65e5\u65b0 V6\u63a8\u7406<\/td>\n37<\/td>\n64<\/td>\n51<\/td>\n<\/tr>\n
32<\/td>\nStep 2<\/td>\n32<\/td>\n68<\/td>\n50<\/td>\n<\/tr>\n
33<\/td>\nStep R1-V-Mini<\/td>\n36<\/td>\n60<\/td>\n48<\/td>\n<\/tr>\n
34<\/td>\nBaichuan4-Turbo<\/td>\n33<\/td>\n60<\/td>\n47<\/td>\n<\/tr>\n
35<\/td>\nGLM-Z1-Air<\/td>\n32<\/td>\n60<\/td>\n46<\/td>\n<\/tr>\n
36<\/td>\nLlama 3.3 70B<\/td>\n33<\/td>\n56<\/td>\n45<\/td>\n<\/tr>\n
37<\/td>\nSpark 4.0 Ultra<\/td>\n19<\/td>\n64<\/td>\n41<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

 <\/p>\n

\u5716\u8868 1\uff1a\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u7d9c\u5408\u6392\u540d<\/p>\n

\"\"<\/p>\n

\u5716\u8868 2\uff1a\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u68af\u968a<\/p>\n

\u662f\u6b21\u7814\u7a76\u6db5\u84cb\u768437\u500b\u6a21\u578b\u7684\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u5f97\u5206\u53ca\u6392\u540d\u6574\u9ad4\u5448\u73fe\u51fa\u986f\u8457\u7684\u68af\u5ea6\u5dee\u7570\uff0c\u4e14\u5728\u4e8b\u5be6\u6027\u8207\u5fe0\u5be6\u6027\u5e7b\u89ba\u63a7\u5236\u4e0a\u8868\u73fe\u51fa\u4e0d\u540c\u7684\u80fd\u529b\u7279\u5fb5\u3002<\/p>\n

\u7d9c\u5408\u800c\u8a00\uff0c\u76ee\u524d\u5927\u6a21\u578b\u5728\u5fe0\u5be6\u6027\u5e7b\u89ba\u63a7\u5236\u4e0a\u5df2\u76f8\u7576\u5f37\u52c1\uff0c\u4f46\u5728\u4e8b\u5be6\u6027\u5e7b\u89ba\u9818\u57df\u4e0a\u4ecd\u6709\u4e0d\u8db3\u4e4b\u8655\u53cd\u6620\u51fa\u9019\u4e9b\u6a21\u578b\u76ee\u524d\u666e\u904d\u5b58\u5728<\/strong>\u300c<\/strong>\u56b4\u5b88\u6307\u4ee4\uff0c\u4f46\u6613\u865b\u69cb\u4e8b\u5be6<\/strong>\u300d<\/strong>\u7684\u50be\u5411\u3002<\/strong><\/p>\n

\u6b64\u5916\uff0c\u63a8\u7406\u6a21\u578b\u7684\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u76f8\u5c0d\u8868\u73fe\u8f03\u597d\uff0c\u4f8b\u5982\u901a\u7fa9\u5343\u554f3\uff08\u601d\u8003\u6a21\u5f0f\uff09\uff0c\u6587\u5fc3\u4e00\u8a00 X1-Turbo\u548cClaude 4 Opus\uff08\u601d\u8003\u6a21\u5f0f\uff09\u7b49\u63a8\u7406\u6a21\u578b\u7684\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u6bd4\u540c\u7cfb\u5217\u901a\u7528\u7248\u672c\u66f4\u4f73\u3002<\/p>\n

\u5728\u4e2d\u570b\u5927\u8a9e\u8a00\u6a21\u578b\u4e2d\uff0c\u8c46\u53051.5 Pro\u7cfb\u5217\u4f4d\u65bc\u9818\u5148\u5730\u4f4d\uff0c\u5728\u4e8b\u5be6\u6027\u5e7b\u89ba\u63a7\u5236\u548c\u5fe0\u5be6\u6027\u5e7b\u89ba\u63a7\u5236\u7684\u9818\u57df\u5f97\u5206\u5747\u8861\uff0c\u5c55\u73fe\u51fa\u7a69\u5065\u7684\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\uff1b\u7136\u800c\uff0c\u5176\u7d9c\u5408\u80fd\u529b\u8207GPT-5\u548cClaude\u7cfb\u5217\u4ecd\u6709\u5dee\u8ddd\u3002 \u76f8\u6bd4\u4e4b\u4e0b\uff0cDeepSeek\u7cfb\u5217\u7684\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u7a0d\u986f\u905c\u8272\uff0c\u6709\u5f85\u52a0\u5f37\u3002<\/p>\n

\u8acb\u6309\u6b64<\/a><\/span>\u700f\u89bd\u300a\u5927\u8a9e\u8a00\u6a21\u578b\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u6df1\u5ea6\u6e2c\u8a55\u5831\u544a\u300b\u5168\u6587\u3002<\/p>\n

\u7e3d\u62ec\u800c\u8a00\uff0c\u672c\u6b21\u8a55\u4f30\u900f\u904e\u4e8b\u5be6\u6027\u540c\u5fe0\u5be6\u6027\u7684\u5206\u985e\u6846\u67b6\uff0c\u6df1\u5165\u95e1\u8ff0\u5927\u8a9e\u8a00\u6a21\u578b\u5e7b\u89ba\u63a7\u5236\u80fd\u529b\u7684\u6838\u5fc3\u7406\u8ad6\u57fa\u790e\u8207\u6a21\u578b\u8868\u73fe\u3002\u672a\u4f86\uff0c AI\u53ef\u4fe1\u5ea6\u9700\u517c\u9867\u63a7\u5236\u4e8b\u5be6\u6027\u5e7b\u89ba\u8207\u5fe0\u5be6\u6027\u5e7b\u89ba\u5169\u500b\u9818\u57df\u7684\u80fd\u529b\uff0c\u4ee5\u7522\u51fa\u66f4\u53ef\u4fe1\u8cf4\u7684\u5167\u5bb9\u3002<\/p>\n

\u5716\u7247<\/u><\/strong><\/h5>\n

\"\"<\/p>\n

    \n
  1. \u6e2f\u5927\u7d93\u7ba1\u5b78\u9662\u5275\u65b0\u53ca\u8cc7\u8a0a\u7ba1\u7406\u5b78\u6559\u6388\u517c\u590f\u5229\u840a\u4f09\u5137\u57fa\u91d1\u6559\u6388 \uff08\u6230\u7565\u4fe1\u606f\u7ba1\u7406\u5b78\uff09\u8523\u93ae\u8f1d\u6559\u6388<\/strong><\/a><\/li>\n<\/ol>\n

    \u8acb\u6309\u6b64<\/a><\/span>\u4e0b\u8f09\u9ad8\u6e05\u5716\u7247\u3002<\/p>\n","protected":false},"featured_media":0,"template":"","meta":{"_lmt_disableupdate":"no","_lmt_disable":"","footnotes":""},"class_list":["post-243206","hkubs-media","type-hkubs-media","status-publish","hentry","media-categories-press-release-tc","media-topic-categories-research-tc"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t