{"id":16745,"date":"2026-08-11T23:11:10","date_gmt":"2026-08-11T13:11:10","guid":{"rendered":"https:\/\/geoffreychen.com\/?p=16745"},"modified":"2026-08-11T23:18:12","modified_gmt":"2026-08-11T13:18:12","slug":"finding-sources-is-not-finishing-research","status":"publish","type":"post","link":"https:\/\/geoffreychen.com\/zh\/finding-sources-is-not-finishing-research\/","title":{"rendered":"AI\u627e\u5230\u4e86\u8d44\u6599\uff0c\u4e3a\u4ec0\u4e48\u8fd8\u4e0d\u80fd\u7b97\u5b8c\u6210\u7814\u7a76\uff1f"},"content":{"rendered":"<\/p>\n<p><em>\u300aAI\u8fdb\u5165\u5de5\u4f5c\u4e4b\u540e\u300b\u00b7 \u7b2c\u4e00\u5b63\u201c\u4ece\u56de\u7b54\u95ee\u9898\u5230\u53c2\u4e0e\u5de5\u4f5c\u201d\u00b7 \u7b2c\u4e94\u7bc7<\/em><\/p>\n<p>\u8bbe\u60f3\u4e00\u540d\u4f5c\u8005\u51c6\u5907\u7814\u7a76\u201c\u751f\u6210\u5f0fAI\u662f\u5426\u63d0\u9ad8\u7a0b\u5e8f\u5458\u7684\u751f\u4ea7\u7387\u201d\u3002\u4ed6\u8ba9AI\u641c\u7d22\u76f8\u5173\u7814\u7a76\u3002\u51e0\u5206\u949f\u540e\uff0c\u7cfb\u7edf\u5217\u51fa\u5341\u51e0\u7bc7\u8bba\u6587\uff0c\u5206\u522b\u652f\u6301\u201c\u663e\u8457\u63d0\u9ad8\u201d\u201c\u5f71\u54cd\u6709\u9650\u201d\u548c\u201c\u53ef\u80fd\u964d\u4f4e\u6548\u7387\u201d\u4e09\u79cd\u7ed3\u8bba\uff0c\u8fd8\u751f\u6210\u4e86\u4e00\u6bb5\u770b\u4f3c\u5e73\u8861\u7684\u7efc\u8ff0\u3002<\/p>\n<p>\u8d44\u6599\u6709\u4e86\uff0c\u89c2\u70b9\u4e5f\u88ab\u5206\u6210\u4e0d\u540c\u9635\u8425\uff0c\u6587\u7ae0\u4f3c\u4e4e\u5df2\u7ecf\u5b8c\u6210\u4e86\u4e00\u5927\u534a\u3002<\/p>\n<p>\u4f46\u4ed6\u6253\u5f00\u8bba\u6587\u540e\u53d1\u73b0\uff0c\u6709\u7684\u7814\u7a76\u6d4b\u8bd5\u7684\u662f\u65b0\u624b\u5728\u4e00\u4e2a\u5c0f\u65f6\u5185\u5199\u5c0f\u7a0b\u5e8f\uff0c\u6709\u7684\u7814\u7a76\u89c2\u5bdf\u7684\u662f\u719f\u6089\u5927\u578b\u4ee3\u7801\u5e93\u7684\u8d44\u6df1\u5f00\u53d1\u8005\uff0c\u6709\u7684\u6765\u81ea\u5458\u5de5\u81ea\u6211\u62a5\u544a\uff0c\u6709\u7684\u6d4b\u91cf\u5b8c\u6210\u901f\u5ea6\uff0c\u6709\u7684\u6d4b\u91cf\u4ee3\u7801\u884c\u6570\uff0c\u8fd8\u6709\u7684\u53ea\u662f\u5c1a\u672a\u540c\u884c\u8bc4\u5ba1\u7684\u9884\u5370\u672c\u3002\u51e0\u7bc7\u8bba\u6587\u4f7f\u7528\u7684AI\u5de5\u5177\u76f8\u9694\u6570\u5e74\uff0c\u5b9e\u9a8c\u73af\u5883\u4e5f\u5b8c\u5168\u4e0d\u540c\u3002<\/p>\n<p>\u8fd9\u4e9b\u8d44\u6599\u90fd\u4e0e\u95ee\u9898\u6709\u5173\uff0c\u5374\u4e0d\u80fd\u76f4\u63a5\u653e\u5728\u540c\u4e00\u5f20\u201c\u8d5e\u6210\u2014\u53cd\u5bf9\u201d\u8868\u91cc\u3002\u771f\u6b63\u7684\u7814\u7a76\u5de5\u4f5c\u4ece\u8fd9\u91cc\u624d\u521a\u521a\u5f00\u59cb\uff1a\u6bcf\u4e2a\u6765\u6e90\u7a76\u7adf\u6d4b\u91cf\u4e86\u4ec0\u4e48\uff0c\u5b83\u80fd\u652f\u6301\u591a\u5927\u7684\u7ed3\u8bba\uff0c\u5f7c\u6b64\u4e3a\u4ec0\u4e48\u4e0d\u540c\uff0c\u4ee5\u53ca\u54ea\u4e9b\u5dee\u5f02\u6765\u81ea\u4efb\u52a1\u3001\u4eba\u5458\u3001\u65f6\u95f4\u6216\u6d4b\u91cf\u65b9\u6cd5\u3002<\/p>\n<p>AI\u53ef\u4ee5\u5927\u5e45\u964d\u4f4e\u53d1\u73b0\u8d44\u6599\u7684\u6210\u672c\u3002\u4f46\u53d1\u73b0\u8d44\u6599\u662f\u68c0\u7d22\u7ed3\u679c\uff0c\u4e0d\u662f\u7814\u7a76\u7ed3\u8bba\u3002<\/p>\n<h2>\u4e00\u7bc7\u8d44\u6599\u5b58\u5728\uff0c\u4e0d\u7b49\u4e8e\u5b83\u652f\u6301\u773c\u524d\u7684\u4e3b\u5f20<\/h2>\n<p>\u7814\u7a76\u4e2d\u7684\u5f15\u7528\u81f3\u5c11\u8981\u901a\u8fc7\u4e09\u9053\u68c0\u67e5\u3002<\/p>\n<p>\u7b2c\u4e00\u9053\u662f<strong>\u5b58\u5728\u68c0\u67e5<\/strong>\uff1a\u8fd9\u7bc7\u8bba\u6587\u3001\u62a5\u544a\u3001\u6cd5\u5f8b\u6216\u6570\u636e\u96c6\u662f\u5426\u771f\u7684\u5b58\u5728\uff0c\u4f5c\u8005\u3001\u6807\u9898\u3001\u65e5\u671f\u548c\u94fe\u63a5\u662f\u5426\u6b63\u786e\u3002<\/p>\n<p>\u7b2c\u4e8c\u9053\u662f<strong>\u5185\u5bb9\u68c0\u67e5<\/strong>\uff1a\u6765\u6e90\u662f\u5426\u771f\u7684\u8bf4\u4e86\u6587\u7ae0\u5f52\u7ed9\u5b83\u7684\u5185\u5bb9\u3002\u4e00\u4e2a\u771f\u5b9e\u94fe\u63a5\u5b8c\u5168\u53ef\u80fd\u88ab\u7528\u6765\u652f\u6301\u5b83\u6ca1\u6709\u8ba8\u8bba\u7684\u4e3b\u5f20\u3002<\/p>\n<p>\u7b2c\u4e09\u9053\u662f<strong>\u8303\u56f4\u68c0\u67e5<\/strong>\uff1a\u5373\u4f7f\u6765\u6e90\u652f\u6301\u8fd9\u53e5\u8bdd\uff0c\u7ed3\u8bba\u662f\u5426\u88ab\u6269\u5927\u5230\u4e86\u4e0d\u9002\u7528\u7684\u4eba\u7fa4\u3001\u65f6\u671f\u3001\u5730\u533a\u6216\u4efb\u52a1\u3002<\/p>\n<p>AI\u751f\u6210\u5f15\u7528\u65f6\uff0c\u6700\u660e\u663e\u7684\u5931\u8d25\u662f\u865a\u6784\u8bba\u6587\u30022026\u5e74\u4e00\u9879\u5173\u4e8e\u53c2\u8003\u6587\u732e\u5e7b\u89c9\u7684\u7814\u7a76\u663e\u793a\uff0c\u63d0\u793a\u7ea6\u675f\u80fd\u591f\u6539\u53d8\u4e0d\u540c\u6a21\u578b\u751f\u6210\u865a\u5047\u5f15\u7528\u7684\u65b9\u5f0f\uff0c\u4f46\u5e76\u4e0d\u80fd\u628a\u95ee\u9898\u7b80\u5355\u6d88\u9664\uff1b\u7814\u7a76\u8005\u5f3a\u8c03\uff0c\u5f15\u7528\u5931\u8d25\u4e0e\u4efb\u52a1\u7ed3\u6784\u6709\u5173\uff0c\u4f7f\u7528\u8005\u4ecd\u9700\u9a8c\u8bc1\u3002<a href=\"https:\/\/proceedings.mlr.press\/v318\/davis26a.html\" target=\"_blank\" rel=\"noopener\">PMLR\uff1a\u300aCitation Constraints and Reference Hallucinations in Large Language Models\u300b<\/a><\/p>\n<p>\u4e0d\u8fc7\uff0c\u4e0d\u5b58\u5728\u7684\u6765\u6e90\u53cd\u800c\u76f8\u5bf9\u5bb9\u6613\u53d1\u73b0\u3002\u66f4\u9690\u853d\u7684\u662f\u201c\u771f\u5b9e\u6765\u6e90\u3001\u9519\u8bef\u7528\u9014\u201d\uff1a\u8bba\u6587\u5b58\u5728\uff0c\u6807\u9898\u4e5f\u76f8\u5173\uff0c\u4f46\u6b63\u6587\u6ca1\u6709\u652f\u6301\u90a3\u9879\u65ad\u8a00\uff1b\u6216\u8005\u7814\u7a76\u786e\u5b9e\u53d1\u73b0\u67d0\u79cd\u6548\u679c\uff0c\u6587\u7ae0\u5374\u5220\u6389\u4e86\u6837\u672c\u3001\u6761\u4ef6\u548c\u9650\u5236\u3002<\/p>\n<p>\u4f8b\u5982\uff0c\u4e00\u9879\u5199\u4f5c\u5b9e\u9a8c\u53d1\u73b0AI\u5e2e\u52a9\u53c2\u4e0e\u8005\u66f4\u5feb\u5b8c\u6210\u7279\u5b9a\u7684\u4e2d\u7b49\u96be\u5ea6\u4e13\u4e1a\u5199\u4f5c\u4efb\u52a1\uff0c\u5e76\u63d0\u9ad8\u5916\u90e8\u8bc4\u5206\u3002\u628a\u5b83\u51c6\u786e\u5199\u6210\u201c\u5728\u8fd9\u7c7b\u53d7\u63a7\u4efb\u52a1\u4e2d\u89c2\u5bdf\u5230\u901f\u5ea6\u548c\u8d28\u91cf\u6536\u76ca\u201d\u662f\u5408\u7406\u7684\u3002\u628a\u5b83\u6269\u5c55\u6210\u201cAI\u5df2\u7ecf\u8bc1\u660e\u53ef\u4ee5\u63d0\u9ad8\u6240\u6709\u77e5\u8bc6\u5de5\u4f5c\u7684\u8d28\u91cf\u201d\uff0c\u5c31\u8d85\u51fa\u4e86\u8bc1\u636e\u8303\u56f4\u3002<a href=\"https:\/\/doi.org\/10.1126\/science.adh2586\" target=\"_blank\" rel=\"noopener\">Science\uff1a\u300aExperimental evidence on the productivity effects of generative artificial intelligence\u300b<\/a><\/p>\n<p>\u7814\u7a76\u4e0d\u662f\u628a\u5e26\u94fe\u63a5\u7684\u53e5\u5b50\u8fde\u63a5\u8d77\u6765\uff0c\u800c\u662f\u63a7\u5236\u6bcf\u4e2a\u7ed3\u8bba\u53ef\u4ee5\u8d70\u591a\u8fdc\u3002<\/p>\n<h2>\u641c\u7d22\u3001\u9605\u8bfb\u3001\u5224\u65ad\u548c\u7efc\u5408\u662f\u56db\u9879\u4e0d\u540c\u5de5\u4f5c<\/h2>\n<p>AI\u641c\u7d22\u5de5\u5177\u7ecf\u5e38\u628a\u56db\u4e2a\u6b65\u9aa4\u538b\u7f29\u6210\u4e00\u4e2a\u987a\u6ed1\u7b54\u6848\u3002<\/p>\n<p><strong>\u641c\u7d22<\/strong>\u5bfb\u627e\u53ef\u80fd\u76f8\u5173\u7684\u6750\u6599\u3002\u5b83\u4f18\u5316\u7684\u662f\u53ec\u56de\u548c\u6392\u5e8f\uff1a\u5728\u5e9e\u5927\u4fe1\u606f\u4e2d\u5c3d\u5feb\u627e\u5230\u5019\u9009\u6765\u6e90\u3002<\/p>\n<p><strong>\u9605\u8bfb<\/strong>\u786e\u5b9a\u6750\u6599\u5b9e\u9645\u8bf4\u4e86\u4ec0\u4e48\uff0c\u5305\u62ec\u65b9\u6cd5\u3001\u6570\u636e\u3001\u9650\u5236\u3001\u5b9a\u4e49\u548c\u7ed3\u8bba\uff0c\u800c\u4e0d\u53ea\u662f\u6807\u9898\u4e0e\u6458\u8981\u3002<\/p>\n<p><strong>\u5224\u65ad<\/strong>\u8bc4\u4f30\u6765\u6e90\u7684\u8d28\u91cf\u548c\u9002\u7528\u6027\u3002\u5b98\u65b9\u6587\u4ef6\u53ef\u80fd\u6700\u9002\u5408\u8bf4\u660e\u73b0\u884c\u89c4\u5219\uff0c\u4f46\u4e0d\u4e00\u5b9a\u80fd\u8bc1\u660e\u653f\u7b56\u6709\u6548\uff1b\u968f\u673a\u5b9e\u9a8c\u53ef\u4ee5\u652f\u6301\u7279\u5b9a\u56e0\u679c\u7ed3\u8bba\uff0c\u5374\u672a\u5fc5\u9002\u7528\u4e8e\u53e6\u4e00\u4e2a\u804c\u4e1a\u73af\u5883\uff1b\u65b0\u95fb\u62a5\u9053\u53ef\u4ee5\u53d1\u73b0\u6848\u4f8b\uff0c\u4f46\u901a\u5e38\u4e0d\u80fd\u66ff\u4ee3\u539f\u59cb\u8bb0\u5f55\u3002<\/p>\n<p><strong>\u7efc\u5408<\/strong>\u89e3\u91ca\u591a\u4e2a\u6765\u6e90\u4e4b\u95f4\u7684\u5173\u7cfb\u3002\u7ed3\u679c\u4e0d\u540c\u53ef\u80fd\u5e76\u975e\u4e92\u76f8\u5426\u5b9a\uff0c\u800c\u662f\u7814\u7a76\u4e86\u4e0d\u540c\u4efb\u52a1\u3002\u771f\u6b63\u7684\u7efc\u5408\u4e0d\u662f\u201c\u6709\u7814\u7a76\u652f\u6301\uff0c\u4e5f\u6709\u7814\u7a76\u53cd\u5bf9\u201d\uff0c\u800c\u662f\u8bf4\u660e\u5dee\u5f02\u4e3a\u4ec0\u4e48\u51fa\u73b0\uff0c\u4ee5\u53ca\u5728\u4ec0\u4e48\u6761\u4ef6\u4e0b\u5e94\u76f8\u4fe1\u54ea\u4e00\u79cd\u7ed3\u8bba\u3002<\/p>\n<p>\u5927\u8bed\u8a00\u6a21\u578b\u5f88\u64c5\u957f\u628a\u4e0d\u540c\u6750\u6599\u5199\u6210\u8fde\u8d2f\u53d9\u8ff0\u3002\u8fd9\u9879\u80fd\u529b\u5728\u641c\u7d22\u521d\u671f\u5f88\u6709\u4ef7\u503c\uff0c\u53ef\u4ee5\u5e2e\u52a9\u5efa\u7acb\u5173\u952e\u8bcd\u3001\u53d1\u73b0\u673a\u6784\u3001\u6574\u7406\u7814\u7a76\u5730\u56fe\u3002\u4f46\u8bed\u8a00\u8fde\u8d2f\u4f1a\u8ba9\u5c1a\u672a\u5b8c\u6210\u7684\u9605\u8bfb\u548c\u5224\u65ad\u770b\u8d77\u6765\u5df2\u7ecf\u5b8c\u6210\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u7814\u7a76\u6d41\u7a0b\u9700\u8981\u6545\u610f\u628a\u56db\u9879\u5de5\u4f5c\u91cd\u65b0\u5206\u5f00\u3002\u5019\u9009\u6765\u6e90\u6e05\u5355\u4e0d\u80fd\u81ea\u52a8\u8fdb\u5165\u6b63\u6587\uff1b\u6458\u8981\u4e0d\u80fd\u4ee3\u66ff\u65b9\u6cd5\u9605\u8bfb\uff1b\u540c\u4e00\u4e3b\u9898\u4e0d\u80fd\u4ee3\u66ff\u540c\u4e00\u8bc1\u636e\u95ee\u9898\u3002<\/p>\n<h2>\u68c0\u7d22\u589e\u5f3a\u89e3\u51b3\u4e86\u201c\u4ece\u54ea\u91cc\u53d6\u8d44\u6599\u201d\uff0c\u6ca1\u6709\u89e3\u51b3\u201c\u600e\u6837\u76f8\u4fe1\u201d<\/h2>\n<p>\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u8ba9\u6a21\u578b\u5148\u4ece\u5916\u90e8\u6587\u6863\u4e2d\u5bfb\u627e\u5185\u5bb9\uff0c\u518d\u6839\u636e\u68c0\u7d22\u7ed3\u679c\u751f\u6210\u7b54\u6848\u3002\u8fd9\u6bd4\u5b8c\u5168\u4f9d\u8d56\u6a21\u578b\u8bad\u7ec3\u4e2d\u5f62\u6210\u7684\u4e00\u822c\u77e5\u8bc6\u66f4\u5bb9\u6613\u66f4\u65b0\uff0c\u4e5f\u66f4\u6709\u673a\u4f1a\u663e\u793a\u6765\u6e90\u3002<a href=\"https:\/\/papers.neurips.cc\/paper\/2020\/hash\/6b493230205f780e1bc26945df7481e5-Abstract.html\" target=\"_blank\" rel=\"noopener\">NeurIPS\uff1a\u300aRetrieval-Augmented Generation for Knowledge-Intensive NLP Tasks\u300b<\/a><\/p>\n<p>\u4f46RAG\u662f\u4e00\u79cd\u83b7\u53d6\u4fe1\u606f\u7684\u67b6\u6784\uff0c\u4e0d\u662f\u81ea\u52a8\u7684\u8bc1\u636e\u8bc4\u5ba1\u5236\u5ea6\u3002<\/p>\n<p>\u68c0\u7d22\u7cfb\u7edf\u53ef\u80fd\u8fd4\u56de\u9519\u8bef\u7248\u672c\u3001\u76f8\u4f3c\u4f46\u65e0\u5173\u7684\u6bb5\u843d\uff0c\u6216\u8005\u53ea\u6293\u5230\u652f\u6301\u95ee\u9898\u9884\u8bbe\u7684\u4e00\u9762\u3002\u6587\u6863\u5207\u5206\u8fd8\u53ef\u80fd\u628a\u7ed3\u8bba\u4e0e\u9650\u5236\u62c6\u5f00\uff1a\u6a21\u578b\u770b\u89c1\u201c\u7ed3\u679c\u663e\u8457\u201d\uff0c\u5374\u6ca1\u6709\u540c\u65f6\u770b\u89c1\u201c\u6837\u672c\u5f88\u5c0f\u201d\u6216\u201c\u4ec5\u9002\u7528\u4e8e\u67d0\u79cd\u8bbe\u7f6e\u201d\u3002<\/p>\n<p>\u957f\u4e0a\u4e0b\u6587\u4e5f\u4e0d\u80fd\u4fdd\u8bc1\u6240\u6709\u6750\u6599\u88ab\u540c\u7b49\u4f7f\u7528\u3002\u5173\u4e8e\u201c\u8ff7\u5931\u5728\u4e2d\u95f4\u201d\u7684\u7814\u7a76\u663e\u793a\uff0c\u76f8\u5173\u4fe1\u606f\u5728\u957f\u8f93\u5165\u4e2d\u7684\u4f4d\u7f6e\u53ef\u80fd\u5f71\u54cd\u6a21\u578b\u80fd\u5426\u6709\u6548\u5229\u7528\u5b83\u3002<a href=\"https:\/\/direct.mit.edu\/tacl\/article\/doi\/10.1162\/tacl_a_00638\/119630\/Lost-in-the-Middle-How-Language-Models-Use-Long\" target=\"_blank\" rel=\"noopener\">TACL\uff1a\u300aLost in the Middle\u300b<\/a><\/p>\n<p>\u6240\u4ee5\uff0c\u4e00\u5957\u7814\u7a76\u7cfb\u7edf\u81f3\u5c11\u9700\u8981\u68c0\u9a8c\u4e24\u4e2a\u73af\u8282\uff1a\u68c0\u7d22\u5230\u7684\u6750\u6599\u662f\u5426\u6b63\u786e\uff0c\u751f\u6210\u7684\u4e3b\u5f20\u662f\u5426\u5fe0\u5b9e\u4e8e\u6750\u6599\u3002\u53ea\u68c0\u67e5\u6700\u7ec8\u6587\u5b57\u901a\u4e0d\u901a\u987a\uff0c\u4f1a\u540c\u65f6\u653e\u8fc7\u68c0\u7d22\u9519\u8bef\u4e0e\u89e3\u91ca\u9519\u8bef\u3002<\/p>\n<h2>AI\u81ea\u6211\u6838\u67e5\u6709\u5e2e\u52a9\uff0c\u4f46\u4e0d\u7b49\u4e8e\u72ec\u7acb\u9a8c\u8bc1<\/h2>\n<p>\u4e00\u79cd\u5e38\u89c1\u505a\u6cd5\u662f\u8981\u6c42\u6a21\u578b\u68c0\u67e5\u81ea\u5df1\u7684\u7b54\u6848\u3001\u5217\u51fa\u53ef\u80fd\u9519\u8bef\uff0c\u6216\u8005\u8ba9\u53e6\u4e00\u4e2aAI\u62c5\u4efb\u8bc4\u5ba1\u3002\u7814\u7a76\u8868\u660e\uff0c\u628a\u751f\u6210\u4e0e\u6838\u67e5\u5206\u6210\u6b65\u9aa4\uff0c\u53ef\u4ee5\u964d\u4f4e\u90e8\u5206\u4e8b\u5b9e\u5e7b\u89c9\u3002\u4f8b\u5982\u201c\u9a8c\u8bc1\u94fe\u201d\u65b9\u6cd5\u5148\u4ea7\u751f\u8349\u7a3f\uff0c\u518d\u63d0\u51fa\u6838\u67e5\u95ee\u9898\u3001\u72ec\u7acb\u56de\u7b54\u8fd9\u4e9b\u95ee\u9898\uff0c\u6700\u540e\u4fee\u8ba2\u7b54\u6848\u3002<a href=\"https:\/\/aclanthology.org\/2024.findings-acl.212\/\" target=\"_blank\" rel=\"noopener\">ACL Anthology\uff1a\u300aChain-of-Verification Reduces Hallucination in Large Language Models\u300b<\/a><\/p>\n<p>\u8fd9\u7c7b\u65b9\u6cd5\u503c\u5f97\u4f7f\u7528\uff0c\u4f46\u5fc5\u987b\u7406\u89e3\u5176\u8fb9\u754c\u3002<\/p>\n<p>\u540c\u4e00\u4e2a\u6a21\u578b\u53ef\u80fd\u5728\u751f\u6210\u548c\u590d\u6838\u65f6\u5171\u4eab\u76f8\u4f3c\u76f2\u70b9\u3002\u4e24\u4e2a\u6a21\u578b\u4e5f\u53ef\u80fd\u4f9d\u636e\u540c\u4e00\u4efd\u9519\u8bef\u68c0\u7d22\u7ed3\u679c\uff0c\u6216\u8005\u90fd\u88ab\u4e00\u4e2a\u770b\u4f3c\u6743\u5a01\u7684\u9519\u8bef\u524d\u63d0\u5f15\u5bfc\u3002\u5982\u679c\u6838\u67e5\u53ea\u662f\u518d\u751f\u6210\u4e00\u6bb5\u201c\u6211\u68c0\u67e5\u8fc7\uff0c\u6ca1\u6709\u95ee\u9898\u201d\u7684\u6587\u5b57\uff0c\u9a8c\u8bc1\u5e76\u6ca1\u6709\u63a5\u89e6\u72ec\u7acb\u8bc1\u636e\u3002<\/p>\n<p>\u771f\u6b63\u72ec\u7acb\u7684\u9a8c\u8bc1\u5e94\u5c3d\u91cf\u56de\u5230\u6a21\u578b\u4e4b\u5916\uff1a\u6253\u5f00\u539f\u59cb\u9875\u9762\u3001\u6838\u5bf9\u6570\u636e\u5e93\u8bb0\u5f55\u3001\u91cd\u65b0\u8ba1\u7b97\u6570\u5b57\u3001\u67e5\u770b\u6cd5\u5f8b\u6b63\u6587\u3001\u8fd0\u884c\u4ee3\u7801\u6d4b\u8bd5\uff0c\u6216\u8005\u7531\u5177\u5907\u9886\u57df\u77e5\u8bc6\u7684\u4eba\u5224\u65ad\u7814\u7a76\u8bbe\u8ba1\u3002<\/p>\n<p>AI\u53ef\u4ee5\u5e2e\u52a9\u63d0\u51fa\u68c0\u67e5\u95ee\u9898\uff0c\u5374\u4e0d\u80fd\u4ec5\u51ed\u518d\u6b21\u8868\u8fbe\u4fe1\u5fc3\u8bc1\u660e\u7b2c\u4e00\u6b21\u56de\u7b54\u6b63\u786e\u3002<\/p>\n<h2>\u5efa\u7acb\u4e00\u4efd\u201c\u4e3b\u5f20\u2014\u8bc1\u636e\u8d26\u672c\u201d<\/h2>\n<p>\u5bf9\u4e8e\u957f\u6587\u3001\u8c03\u67e5\u6216\u4e13\u4e1a\u62a5\u544a\uff0c\u4e00\u4e2a\u7b80\u5355\u800c\u6709\u6548\u7684\u65b9\u6cd5\uff0c\u662f\u5728\u5199\u4f5c\u524d\u5efa\u7acb\u4e3b\u5f20\u4e0e\u8bc1\u636e\u7684\u5bf9\u5e94\u8868\u3002<\/p>\n<p>\u6bcf\u9879\u91cd\u8981\u4e3b\u5f20\u81f3\u5c11\u8bb0\u5f55\uff1a<\/p>\n<ul>\n<li>\u51c6\u5907\u5728\u6b63\u6587\u4e2d\u8868\u8fbe\u4ec0\u4e48\uff1b<\/li>\n<li>\u54ea\u4e2a\u6765\u6e90\u76f4\u63a5\u652f\u6301\u5b83\uff1b<\/li>\n<li>\u6765\u6e90\u4e2d\u7684\u5177\u4f53\u9875\u7801\u3001\u6bb5\u843d\u3001\u8868\u683c\u6216\u6570\u636e\u5b57\u6bb5\uff1b<\/li>\n<li>\u6765\u6e90\u5c5e\u4e8e\u6cd5\u5f8b\u3001\u5b98\u65b9\u65b9\u6cd5\u3001\u539f\u59cb\u7814\u7a76\u3001\u8c03\u67e5\u62a5\u544a\u8fd8\u662f\u65b0\u95fb\uff1b<\/li>\n<li>\u9002\u7528\u7684\u65e5\u671f\u3001\u5730\u533a\u3001\u4eba\u7fa4\u4e0e\u4efb\u52a1\uff1b<\/li>\n<li>\u8fd9\u662f\u4e8b\u5b9e\u3001\u6765\u6e90\u4f5c\u8005\u7684\u89e3\u91ca\uff0c\u8fd8\u662f\u672c\u6587\u4f5c\u8005\u7684\u63a8\u65ad\uff1b<\/li>\n<li>\u662f\u5426\u5b58\u5728\u53cd\u4f8b\u6216\u76f8\u53cd\u8bc1\u636e\uff1b<\/li>\n<li>\u5f53\u524d\u72b6\u6001\u662f\u5f85\u6838\u67e5\u3001\u5df2\u6838\u67e5\u8fd8\u662f\u4e0d\u80fd\u652f\u6301\u3002<\/li>\n<\/ul>\n<p>AI\u53ef\u4ee5\u534f\u52a9\u4ece\u6587\u4ef6\u4e2d\u63d0\u53d6\u5019\u9009\u6bb5\u843d\uff0c\u751f\u6210\u6bd4\u8f83\u8868\u5e76\u53d1\u73b0\u8bba\u8bc1\u7a7a\u7f3a\u3002\u4f46\u201c\u5df2\u6838\u67e5\u201d\u4e0d\u5e94\u7531AI\u770b\u5230\u94fe\u63a5\u540e\u81ea\u52a8\u586b\u5199\u3002\u81f3\u5c11\u5bf9\u6587\u7ae0\u7684\u6838\u5fc3\u4e8b\u5b9e\uff0c\u4f5c\u8005\u9700\u8981\u6253\u5f00\u6765\u6e90\uff0c\u786e\u8ba4\u9875\u9762\u771f\u5b9e\u5b58\u5728\u5e76\u652f\u6301\u9644\u8fd1\u7684\u8868\u8ff0\u3002<\/p>\n<p>\u8fd9\u79cd\u8d26\u672c\u8fd8\u6709\u4e00\u4e2a\u597d\u5904\uff1a\u5f53\u89c4\u5219\u66f4\u65b0\u6216\u6587\u7ae0\u4fee\u8ba2\u65f6\uff0c\u53ef\u4ee5\u8fc5\u901f\u627e\u5230\u54ea\u4e9b\u6bb5\u843d\u4f9d\u8d56\u65e7\u8d44\u6599\u3002\u7814\u7a76\u4e0d\u518d\u662f\u4e00\u7bc7\u6587\u5b57\u548c\u4e00\u4e32\u6587\u672b\u94fe\u63a5\uff0c\u800c\u6210\u4e3a\u53ef\u4ee5\u7ef4\u62a4\u7684\u8bc1\u636e\u7ed3\u6784\u3002<\/p>\n<h2>\u8d44\u6599\u6570\u91cf\u4e0d\u662f\u7814\u7a76\u8d28\u91cf<\/h2>\n<p>\u751f\u6210\u5f0fAI\u8ba9\u201c\u5217\u51fa\u4e8c\u5341\u4e2a\u6765\u6e90\u201d\u53d8\u5f97\u5bb9\u6613\uff0c\u8fd9\u53ef\u80fd\u4f7f\u4eba\u8bef\u4ee5\u4e3a\u66f4\u591a\u94fe\u63a5\u4ee3\u8868\u66f4\u5145\u5206\u7684\u7814\u7a76\u3002<\/p>\n<p>\u4f46\u4e8c\u5341\u7bc7\u4e92\u76f8\u5f15\u7528\u540c\u4e00\u4efd\u4e8c\u624b\u62a5\u9053\u7684\u6587\u7ae0\uff0c\u8bc1\u636e\u4e0a\u53ef\u80fd\u53ea\u6709\u4e00\u4e2a\u6765\u6e90\u3002\u5341\u7bc7\u89c2\u70b9\u76f8\u4f3c\u7684\u8bc4\u8bba\uff0c\u4e5f\u4e0d\u80fd\u66ff\u4ee3\u4e00\u4efd\u539f\u59cb\u6570\u636e\u3002\u5927\u91cf\u6458\u8981\u5982\u679c\u6ca1\u6709\u9605\u8bfb\u65b9\u6cd5\u548c\u9650\u5236\uff0c\u53ea\u4f1a\u589e\u52a0\u9a8c\u8bc1\u8d1f\u62c5\u3002<\/p>\n<p>\u66f4\u597d\u7684\u7814\u7a76\u4e0d\u4e00\u5b9a\u62e5\u6709\u6700\u591a\u8d44\u6599\uff0c\u800c\u662f\u6765\u6e90\u5c42\u7ea7\u6e05\u695a\uff1a\u5173\u952e\u5236\u5ea6\u4e8b\u5b9e\u56de\u5230\u6cd5\u5f8b\u4e0e\u5b98\u65b9\u6587\u4ef6\uff0c\u5173\u952e\u7ecf\u9a8c\u5224\u65ad\u56de\u5230\u539f\u59cb\u7814\u7a76\u548c\u6570\u636e\uff0c\u65b0\u95fb\u7528\u4e8e\u65f6\u95f4\u7ebf\u4e0e\u6848\u4f8b\u53d1\u73b0\uff0c\u8bc4\u8bba\u7528\u4e8e\u7406\u89e3\u4e89\u8bae\uff0c\u4f46\u4e0d\u6df7\u6dc6\u5404\u81ea\u4f5c\u7528\u3002<\/p>\n<p>\u7f8e\u56fd\u56fd\u5bb6\u6807\u51c6\u4e0e\u6280\u672f\u7814\u7a76\u9662\u7684\u751f\u6210\u5f0fAI\u98ce\u9669\u6846\u67b6\u628a\u201c\u865a\u6784\u4fe1\u606f\u201d\u53ca\u4fe1\u606f\u5b8c\u6574\u6027\u89c6\u4e3a\u9700\u8981\u6d4b\u91cf\u548c\u7ba1\u7406\u7684\u98ce\u9669\uff0c\u5e76\u5f3a\u8c03\u6d4b\u8bd5\u3001\u8bc4\u4f30\u3001\u9a8c\u8bc1\u4e0e\u786e\u8ba4\u4e0d\u80fd\u53ea\u53d1\u751f\u5728\u6a21\u578b\u5f00\u53d1\u9636\u6bb5\uff0c\u4e5f\u8981\u8fdb\u5165\u5177\u4f53\u4f7f\u7528\u73af\u5883\u3002<a href=\"https:\/\/www.nist.gov\/publications\/artificial-intelligence-risk-management-framework-generative-artificial-intelligence\" target=\"_blank\" rel=\"noopener\">NIST\uff1a\u300aGenerative Artificial Intelligence Profile\u300b<\/a><\/p>\n<p>\u5bf9\u4e8e\u7814\u7a76\u5199\u4f5c\uff0c\u8fd9\u610f\u5473\u7740\u4e0d\u80fd\u56e0\u4e3a\u5de5\u5177\u5e26\u6709\u201c\u7814\u7a76\u201d\u540d\u79f0\u6216\u63d0\u4f9b\u5f15\u7528\u754c\u9762\uff0c\u5c31\u964d\u4f4e\u6765\u6e90\u5ba1\u67e5\u6807\u51c6\u3002<\/p>\n<h2>\u7ed3\u8bba\uff1aAI\u53ef\u4ee5\u7f29\u77ed\u627e\u5230\u8d44\u6599\u7684\u8def\uff0c\u4e0d\u80fd\u66ff\u4f60\u51b3\u5b9a\u8bc1\u636e\u591f\u4e0d\u591f<\/h2>\n<p>AI\u641c\u7d22\u548c\u7814\u7a76\u5de5\u5177\u53ef\u4ee5\u8fc5\u901f\u6269\u5c55\u89c6\u91ce\uff1a\u5bfb\u627e\u5173\u952e\u8bcd\u3001\u53d1\u73b0\u539f\u59cb\u673a\u6784\u3001\u6bd4\u8f83\u591a\u7bc7\u8bba\u6587\u3001\u6574\u7406\u65f6\u95f4\u7ebf\u3001\u63d0\u53d6\u8868\u683c\u548c\u6307\u51fa\u53ef\u80fd\u53cd\u4f8b\u3002\u8fd9\u4e9b\u80fd\u529b\u80fd\u591f\u8282\u7701\u5927\u91cf\u673a\u68b0\u52b3\u52a8\uff0c\u4e5f\u53ef\u80fd\u8ba9\u4e2a\u4eba\u63a5\u89e6\u8fc7\u53bb\u96be\u4ee5\u8986\u76d6\u7684\u8d44\u6599\u8303\u56f4\u3002<\/p>\n<p>\u4f46\u7814\u7a76\u5b8c\u6210\u7684\u6807\u51c6\u4e0d\u662f\u201c\u7cfb\u7edf\u8fd4\u56de\u4e86\u7b54\u6848\u201d\uff0c\u751a\u81f3\u4e0d\u662f\u201c\u6bcf\u6bb5\u8bdd\u540e\u9762\u90fd\u6709\u94fe\u63a5\u201d\u3002\u771f\u6b63\u7684\u5b8c\u6210\u9700\u8981\u8bc1\u660e\uff1a\u6765\u6e90\u5b58\u5728\uff0c\u5185\u5bb9\u652f\u6301\u4e3b\u5f20\uff0c\u9002\u7528\u8303\u56f4\u6ca1\u6709\u88ab\u6269\u5927\uff0c\u91cd\u8981\u53cd\u8bc1\u5f97\u5230\u5904\u7406\uff0c\u4f5c\u8005\u81ea\u5df1\u7684\u63a8\u65ad\u88ab\u660e\u786e\u6807\u51fa\u3002<\/p>\n<p>\u56e0\u6b64\uff0c\u6700\u91cd\u8981\u7684\u8fb9\u754c\u662f\uff1a<\/p>\n<blockquote>\n<p>AI\u53ef\u4ee5\u8d1f\u8d23\u53d1\u73b0\u5019\u9009\u8bc1\u636e\u3001\u6574\u7406\u6750\u6599\u548c\u63d0\u51fa\u6bd4\u8f83\uff1b\u91c7\u7528\u4ec0\u4e48\u4f5c\u4e3a\u8bc1\u636e\uff0c\u4ee5\u53ca\u8bc1\u636e\u5141\u8bb8\u5f97\u51fa\u4ec0\u4e48\u7ed3\u8bba\uff0c\u4ecd\u5fc5\u987b\u7531\u7814\u7a76\u8005\u627f\u62c5\u3002<\/p>\n<\/blockquote>\n<p>\u5f53\u5bfb\u627e\u8d44\u6599\u53d8\u5f97\u5ec9\u4ef7\uff0c\u7814\u7a76\u80fd\u529b\u4e0d\u4f1a\u56e0\u6b64\u5931\u53bb\u4ef7\u503c\u3002\u76f8\u53cd\uff0c\u5224\u65ad\u6765\u6e90\u3001\u63a7\u5236\u7ed3\u8bba\u8303\u56f4\u548c\u7ef4\u62a4\u8bc1\u636e\u94fe\u4f1a\u6210\u4e3a\u66f4\u7a00\u7f3a\u7684\u5de5\u4f5c\u3002<\/p>\n<p>AI\u627e\u5230\u8d44\u6599\uff0c\u53ea\u662f\u6253\u5f00\u4e86\u7814\u7a76\u7684\u5927\u95e8\u3002\u9605\u8bfb\u3001\u5224\u65ad\u3001\u7efc\u5408\u548c\u8d1f\u8d23\uff0c\u624d\u51b3\u5b9a\u6211\u4eec\u662f\u5426\u771f\u6b63\u8d70\u4e86\u8fdb\u53bb\u3002<\/p>\n<h3>\u4e3b\u8981\u8d44\u6599<\/h3>\n<ul>\n<li><a href=\"https:\/\/proceedings.mlr.press\/v318\/davis26a.html\" target=\"_blank\" rel=\"noopener\">Kimberly Davis and Qusay Mahmoud\uff1aCitation Constraints and Reference Hallucinations in Large Language Models<\/a><\/li>\n<li><a href=\"https:\/\/papers.neurips.cc\/paper\/2020\/hash\/6b493230205f780e1bc26945df7481e5-Abstract.html\" target=\"_blank\" rel=\"noopener\">Patrick Lewis et al.\uff1aRetrieval-Augmented Generation for Knowledge-Intensive NLP Tasks<\/a><\/li>\n<li><a href=\"https:\/\/direct.mit.edu\/tacl\/article\/doi\/10.1162\/tacl_a_00638\/119630\/Lost-in-the-Middle-How-Language-Models-Use-Long\" target=\"_blank\" rel=\"noopener\">Nelson F. Liu et al.\uff1aLost in the Middle: How Language Models Use Long Contexts<\/a><\/li>\n<li><a href=\"https:\/\/aclanthology.org\/2024.findings-acl.212\/\" target=\"_blank\" rel=\"noopener\">Shehzaad Dhuliawala et al.\uff1aChain-of-Verification Reduces Hallucination in Large Language Models<\/a><\/li>\n<li><a href=\"https:\/\/www.nist.gov\/publications\/artificial-intelligence-risk-management-framework-generative-artificial-intelligence\" target=\"_blank\" rel=\"noopener\">NIST\uff1aArtificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile<\/a><\/li>\n<\/ul>\n<p><strong>\u7ee7\u7eed\u9605\u8bfb\uff1a<\/strong><a href=\"https:\/\/geoffreychen.com\/zh\/category\/ai-frontiers\/after-ai-enters-the-workflow\/\">\u6d4f\u89c8\u300aAI\u8fdb\u5165\u5de5\u4f5c\u4e4b\u540e\u300b\u7cfb\u5217\u6587\u7ae0<\/a><\/p>\n<p>","protected":false},"excerpt":{"rendered":"","protected":false},"author":17897162,"featured_media":0,"comment_status":"closed","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"_coblocks_attr":"","_coblocks_dimensions":"","_coblocks_responsive_height":"","_coblocks_accordion_ie_support":"","_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":true,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_geoff_meta_description_en":"When AI finds sources and drafts a synthesis, research still isn\u2019t finished: claims must be checked for relevance, accuracy, scope and evidence strength.","_geoff_meta_description_zh":"AI\u867d\u7136\u80fd\u68c0\u7d22\u5230\u5e76\u751f\u6210\u7efc\u8ff0\uff0c\u7814\u7a76\u4ecd\u672a\u5b8c\u6210\uff1a\u5173\u952e\u4e3b\u5f20\u5fc5\u987b\u9010\u6761\u6838\u5bf9\u6765\u6e90\u5b58\u5728\u6027\u3001\u5185\u5bb9\u652f\u6301\u53ca\u9002\u7528\u8303\u56f4\u7b49\u8bc1\u636e\u5f3a\u5ea6\u3002","_wpcom_ai_launchpad_first_post":false,"_jetpack_feature_clip_id":0,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2},"jetpack_post_was_ever_published":false},"categories":[787296333],"tags":[787296336,787296348,787296349,787296338,787296334,787296351,787296350,787296335],"class_list":["post-16745","post","type-post","status-publish","format-standard","hentry","category-after-ai-enters-the-workflow","tag-after-ai-enters-the-workflow","tag-ai-research","tag-ai-research-zh","tag-after-ai-enters-the-workflow-zh","tag-from-answering-questions-to-participating-in-work","tag-source-verification","tag-source-verification-zh","tag-from-answering-questions-to-participating-in-work-zh",""],"translatepress_titles":{"en_AU":"If AI Has Found the Sources, Why Is the Research Not Finished?","zh_CN":"AI\u627e\u5230\u4e86\u8d44\u6599\uff0c\u4e3a\u4ec0\u4e48\u8fd8\u4e0d\u80fd\u7b97\u5b8c\u6210\u7814\u7a76\uff1f"},"jetpack_publicize_connections":[],"jetpack_likes_enabled":false,"jetpack_sharing_enabled":true,"jetpack_shortlink":"https:\/\/wp.me\/p1ar9H-4m5","jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/geoffreychen.com\/zh\/wp-json\/wp\/v2\/posts\/16745","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/geoffreychen.com\/zh\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/geoffreychen.com\/zh\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/geoffreychen.com\/zh\/wp-json\/wp\/v2\/users\/17897162"}],"replies":[{"embeddable":true,"href":"https:\/\/geoffreychen.com\/zh\/wp-json\/wp\/v2\/comments?post=16745"}],"version-history":[{"count":2,"href":"https:\/\/geoffreychen.com\/zh\/wp-json\/wp\/v2\/posts\/16745\/revisions"}],"predecessor-version":[{"id":16770,"href":"https:\/\/geoffreychen.com\/zh\/wp-json\/wp\/v2\/posts\/16745\/revisions\/16770"}],"wp:attachment":[{"href":"https:\/\/geoffreychen.com\/zh\/wp-json\/wp\/v2\/media?parent=16745"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/geoffreychen.com\/zh\/wp-json\/wp\/v2\/categories?post=16745"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/geoffreychen.com\/zh\/wp-json\/wp\/v2\/tags?post=16745"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}