{"id":1385,"date":"2026-05-18T10:32:13","date_gmt":"2026-05-18T02:32:13","guid":{"rendered":"https:\/\/www.ndnlab.com\/?p=1385"},"modified":"2026-05-18T10:32:50","modified_gmt":"2026-05-18T02:32:50","slug":"rethinking-evaluation-for-llm-hallucination-detection-a-desiderata-a-new-rag-based-benchmark-new-insights","status":"publish","type":"post","link":"https:\/\/www.ndnlab.com\/?p=1385","title":{"rendered":"Rethinking Evaluation for LLM Hallucination Detection: A Desiderata, A New RAG-based Benchmark, New Insights"},"content":{"rendered":"\n

1.\u6458\u8981\uff08Abstract\uff09<\/h2>\n\n\n\n

\u672c\u6587\u7814\u7a76\u7684\u662f LLM hallucination detection benchmark\uff08\u5927\u6a21\u578b\u5e7b\u89c9\u68c0\u6d4b\u57fa\u51c6\uff09 \u7684\u8bc4\u6d4b\u95ee\u9898\u3002\u968f\u7740\u5927\u6a21\u578b\u88ab\u7528\u4e8e\u7535\u5546\u3001\u533b\u7597\u3001\u6cd5\u5f8b\u7b49\u771f\u5b9e\u573a\u666f\uff0c\u5e7b\u89c9\u95ee\u9898\u5df2\u7ecf\u4e0d\u53ea\u662f\u6a21\u578b\u6548\u679c\u95ee\u9898\uff0c\u800c\u662f\u76f4\u63a5\u5173\u7cfb\u5230\u751f\u6210\u5f0f AI \u7684\u5b89\u5168\u4f7f\u7528\u3002\u867d\u7136\u8fd1\u51e0\u5e74\u51fa\u73b0\u4e86\u5927\u91cf\u5e7b\u89c9\u68c0\u6d4b\u65b9\u6cd5\u548c\u68c0\u6d4b\u57fa\u51c6\uff0c\u4f46\u4f5c\u8005\u6307\u51fa\uff0c\u73b0\u6709 benchmark \u672c\u8eab\u5b58\u5728\u660e\u663e\u7f3a\u9677\uff1a\u5f88\u591a\u6570\u636e\u96c6\u5e76\u4e0d\u80fd\u771f\u5b9e\u53cd\u6620 RAG \u573a\u666f\u4e0b\u7684\u5927\u6a21\u578b\u5e7b\u89c9\uff0c\u4e5f\u7f3a\u5c11\u5bf9\u6807\u6ce8\u566a\u58f0\u7684\u7cfb\u7edf\u8bc4\u4f30\u3002<\/p>\n\n\n\n

\u8bba\u6587\u9996\u5148\u63d0\u51fa\u4e00\u7ec4\u8861\u91cf hallucination detection benchmark \u7684 desiderata\uff08\u7406\u60f3\u5c5e\u6027\uff09\uff0c\u5305\u62ec\u81ea\u7136\u751f\u6210\u7684\u5e7b\u89c9\u3001\u4eba\u7c7b\u9a8c\u8bc1\u6807\u7b7e\u3001\u957f\u4e0a\u4e0b\u6587 RAG \u4efb\u52a1\u3001\u771f\u5b9e\u566a\u58f0\u8bad\u7ec3\u6807\u7b7e\u3001\u591a\u79cd\u5e7b\u89c9\u7c7b\u578b\u3001\u591a\u6a21\u578b\u6765\u6e90\u548c\u591a\u9886\u57df\u8986\u76d6\u3002\u4f5c\u8005\u7528\u8fd9\u5957\u6807\u51c6\u68c0\u67e5\u5df2\u6709 23 \u4e2a benchmark\uff0c\u53d1\u73b0\u6ca1\u6709\u4e00\u4e2a\u5df2\u6709\u6570\u636e\u96c6\u540c\u65f6\u6ee1\u8db3\u8fd9\u4e9b\u6761\u4ef6\uff0c\u5c24\u5176\u7f3a\u5c11\u4e24\u7c7b\u8d44\u6e90\uff1a\u4e00\u662f\u9762\u5411 RAG \u7684\u957f\u4e0a\u4e0b\u6587 grounded benchmark\uff1b\u4e8c\u662f\u80fd\u591f\u6a21\u62df\u73b0\u5b9e\u6807\u6ce8\u566a\u58f0\u7684\u8bad\u7ec3\u6807\u7b7e\u3002<\/p>\n\n\n\n

\u4e3a\u586b\u8865\u8fd9\u4e2a\u7f3a\u53e3\uff0c\u4f5c\u8005\u6784\u5efa\u5e76\u5f00\u6e90\u4e86\u65b0\u7684 RAG-based hallucination detection benchmark\uff1aTRIVIA+\u3002\u8fd9\u4e2a\u6570\u636e\u96c6\u5305\u542b\u6765\u81ea\u591a\u4e2a\u95ee\u7b54\u6570\u636e\u6e90\u7684\u957f\u4e0a\u4e0b\u6587\u6837\u672c\uff0c\u7531\u4e09\u4e2a\u4e0d\u540c LLM \u751f\u6210\u56de\u7b54\uff0c\u5e76\u7ecf\u8fc7\u4e25\u683c\u7684\u4eba\u7c7b\u591a\u8f6e\u6807\u6ce8\u3002\u9664\u5e72\u51c0\u7684\u4eba\u5de5\u9a8c\u8bc1\u6807\u7b7e\u5916\uff0cTRIVIA+ \u8fd8\u63d0\u4f9b\u56db\u7ec4\u5e26\u566a\u58f0\u7684\u8bad\u7ec3\u6807\u7b7e\uff0c\u7528\u4e8e\u7814\u7a76\u68c0\u6d4b\u5668\u5728\u5f31\u76d1\u7763\u6216\u4eba\u5de5\u6807\u6ce8\u9519\u8bef\u4e0b\u7684\u9c81\u68d2\u6027\u3002<\/p>\n\n\n\n

\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c\u73b0\u6709\u68c0\u6d4b\u5668\u5728\u975e\u81ea\u7136\u6784\u9020\u7684 HaluEval \u4e0a\u8868\u73b0\u5f88\u9ad8\uff0c\u4f46\u5728\u81ea\u7136\u751f\u6210\u7684 RAG hallucination benchmark \u4e0a\u660e\u663e\u4e0b\u964d\u3002\u5c24\u5176\u5728 TRIVIA+ \u8fd9\u7c7b\u957f\u4e0a\u4e0b\u6587\u573a\u666f\u4e2d\uff0c\u6240\u6709\u68c0\u6d4b\u65b9\u6cd5\u4ecd\u6709\u8f83\u5927\u63d0\u5347\u7a7a\u95f4\u3002\u4e00\u4e2a\u6709\u610f\u601d\u7684\u7ed3\u679c\u662f\uff0c\u7b80\u5355\u7684 LLM-as-a-Judge \u5728 RAG-based HDB \u4e0a\u53cd\u800c\u5177\u6709\u8f83\u5f3a\u7ade\u4e89\u529b\uff1b\u540c\u65f6\uff0c\u6837\u672c\u76f8\u5173\u7684\u6807\u7b7e\u566a\u58f0\u4f1a\u660e\u663e\u5f71\u54cd\u68c0\u6d4b\u5668\u8bad\u7ec3\u548c\u8bc4\u4f30\u3002\u6574\u4f53\u6765\u770b\uff0c\u8fd9\u7bc7\u8bba\u6587\u7684\u4ef7\u503c\u4e0d\u5728\u4e8e\u201c\u53c8\u505a\u4e86\u4e00\u4e2a\u66f4\u5f3a detector\u201d\uff0c\u800c\u5728\u4e8e\u63d0\u9192\u5927\u5bb6\uff1a\u5982\u679c\u8bc4\u6d4b\u96c6\u672c\u8eab\u4e0d\u771f\u5b9e\uff0c\u68c0\u6d4b\u5668\u5206\u6570\u518d\u9ad8\u4e5f\u53ef\u80fd\u6ca1\u6709\u610f\u4e49\u3002<\/p>\n\n\n\n

\"Rethinking<\/figure>\n\n\n\n

2.\u7814\u7a76\u80cc\u666f\u4e0e\u95ee\u9898\u52a8\u673a\uff08Introduction\uff09<\/h2>\n\n\n\n

\u5927\u6a21\u578b\u5e7b\u89c9\u901a\u5e38\u6307\u6a21\u578b\u751f\u6210\u4e86\u4e0d\u5fe0\u5b9e\u3001\u7f16\u9020\u6216\u4e0e\u4e0a\u4e0b\u6587\u4e0d\u4e00\u81f4\u7684\u5185\u5bb9\u3002\u5373\u4f7f\u5f15\u5165 RAG\uff08Retrieval-Augmented Generation\uff09\uff0c\u8ba9\u6a21\u578b\u53c2\u8003\u5916\u90e8\u6587\u6863\u751f\u6210\u7b54\u6848\uff0c\u5e7b\u89c9\u4e5f\u4e0d\u4f1a\u5b8c\u5168\u6d88\u5931\u3002\u73b0\u5b9e\u4e2d\uff0cRAG \u7cfb\u7edf\u7ecf\u5e38\u88ab\u7528\u4e8e\u9700\u8981\u4f9d\u636e\u8bc1\u636e\u56de\u7b54\u7684\u95ee\u9898\uff0c\u6bd4\u5982\u533b\u7597\u95ee\u7b54\u3001\u6cd5\u5f8b\u68c0\u7d22\u3001\u4f01\u4e1a\u77e5\u8bc6\u5e93\u548c\u79d1\u7814\u52a9\u624b\u3002\u56e0\u6b64\uff0c\u5e7b\u89c9\u68c0\u6d4b\u4e0d\u4ec5\u8981\u5224\u65ad\u7b54\u6848\u662f\u5426\u201c\u50cf\u771f\u7684\u201d\uff0c\u66f4\u8981\u5224\u65ad\u7b54\u6848\u662f\u5426\u88ab\u7ed9\u5b9a\u4e0a\u4e0b\u6587\u771f\u6b63\u652f\u6491\u3002<\/p>\n\n\n\n

\u95ee\u9898\u5728\u4e8e\uff0c\u73b0\u5728\u8bb8\u591a hallucination detection benchmark \u5e76\u4e0d\u80fd\u5f88\u597d\u5730\u6a21\u62df\u771f\u5b9e RAG \u573a\u666f\u3002\u6709\u4e9b\u6570\u636e\u96c6\u4e2d\u7684\u5e7b\u89c9\u662f\u4eba\u5de5\u6ce8\u5165\u7684\uff0c\u6216\u8005\u662f\u901a\u8fc7\u63d0\u793a\u6a21\u578b\u201c\u6545\u610f\u751f\u6210\u9519\u8bef\u7b54\u6848\u201d\u5f97\u5230\u7684\u3002\u8fd9\u7c7b\u975e\u81ea\u7136\u5e7b\u89c9\u5f80\u5f80\u6bd4\u771f\u5b9e\u5e7b\u89c9\u66f4\u5bb9\u6613\u88ab\u68c0\u6d4b\u51fa\u6765\uff0c\u53ef\u80fd\u5bfc\u81f4\u68c0\u6d4b\u5668\u5206\u6570\u865a\u9ad8\u3002\u8bba\u6587\u7528 Figure 1 \u505a\u4e86\u4e00\u4e2a\u5f88\u76f4\u89c2\u7684\u8bf4\u660e\uff1aHaluEval \u4e2d\u88ab\u63d0\u793a\u751f\u6210\u7684\u5e7b\u89c9\u6837\u672c\u548c\u975e\u5e7b\u89c9\u6837\u672c\u5728\u8868\u793a\u7a7a\u95f4\u91cc\u5206\u5f97\u5f88\u5f00\uff0c\u800c RAGTruth \u548c TRIVIA+ \u4e2d\u81ea\u7136\u751f\u6210\u7684\u5e7b\u89c9\u66f4\u5bb9\u6613\u548c\u6b63\u5e38\u56de\u7b54\u6df7\u5728\u4e00\u8d77\uff0c\u56e0\u6b64\u68c0\u6d4b\u96be\u5ea6\u66f4\u9ad8\u3002<\/p>\n\n\n\n

\u6b64\u5916\uff0cRAG-based benchmark \u7684\u6807\u6ce8\u6210\u672c\u5f88\u9ad8\u3002\u56e0\u4e3a\u6807\u6ce8\u8005\u4e0d\u4ec5\u8981\u770b\u6a21\u578b\u56de\u7b54\uff0c\u8fd8\u8981\u9605\u8bfb\u8f83\u957f\u7684\u53c2\u8003\u4e0a\u4e0b\u6587\uff0c\u5224\u65ad\u6bcf\u4e00\u53e5\u662f\u5426\u88ab\u4e0a\u4e0b\u6587\u652f\u6301\u3002\u4e0a\u4e0b\u6587\u8d8a\u957f\uff0c\u8d8a\u5bb9\u6613\u51fa\u73b0\u201cneedle in the haystack\u201d\u5f0f\u95ee\u9898\uff1a\u5e7b\u89c9\u53ef\u80fd\u53ea\u85cf\u5728\u4e00\u4e24\u53e5\u8bdd\u91cc\uff0c\u4f46\u6807\u6ce8\u8005\u9700\u8981\u4ece\u5927\u91cf\u6750\u6599\u4e2d\u6838\u5bf9\u3002\u4e5f\u6b63\u56e0\u4e3a\u5982\u6b64\uff0c\u5df2\u6709 RAG-based HDB \u5f88\u5c11\uff0c\u957f\u4e0a\u4e0b\u6587\u6837\u672c\u66f4\u5c11\u3002<\/p>\n\n\n\n

\u4f5c\u8005\u7684\u7814\u7a76\u52a8\u673a\u53ef\u4ee5\u6982\u62ec\u4e3a\u4e00\u53e5\u8bdd\uff1a\u5728\u7814\u7a76 hallucination detector \u4e4b\u524d\uff0c\u5fc5\u987b\u5148\u628a benchmark \u505a\u5bf9\u3002 \u5982\u679c benchmark \u7f3a\u5c11\u81ea\u7136\u5e7b\u89c9\u3001\u7f3a\u5c11\u4eba\u5de5\u9a8c\u8bc1\u6807\u7b7e\u3001\u7f3a\u5c11\u957f\u4e0a\u4e0b\u6587\u548c\u771f\u5b9e\u566a\u58f0\uff0c\u90a3\u4e48\u68c0\u6d4b\u5668\u5728\u8fd9\u4e9b\u6570\u636e\u96c6\u4e0a\u7684\u9ad8\u5206\uff0c\u5f88\u53ef\u80fd\u4e0d\u80fd\u4ee3\u8868\u771f\u5b9e\u573a\u666f\u80fd\u529b\u3002<\/p>\n\n\n\n

\"Rethinking<\/figure>\n\n\n\n

3.Benchmark \u7406\u60f3\u5c5e\u6027\uff1aDesiderata for HDBs<\/h2>\n\n\n\n

\u8bba\u6587\u7684\u4e00\u4e2a\u6838\u5fc3\u8d21\u732e\u662f\u63d0\u51fa\u4e86 hallucination detection benchmark \u5e94\u5177\u5907\u7684\u4e03\u4e2a\u5c5e\u6027\u3002\u8fd9\u91cc\u7684\u91cd\u70b9\u4e0d\u662f\u7b80\u5355\u5217\u6807\u51c6\uff0c\u800c\u662f\u7ed9\u540e\u7eed\u6570\u636e\u96c6\u5ba1\u67e5\u63d0\u4f9b\u4e00\u4e2a\u7edf\u4e00\u89c6\u89d2\u3002<\/p>\n\n\n\n

\u7b2c\u4e00\uff0cbenchmark \u5e94\u5305\u542b organic hallucinations\uff0c\u4e5f\u5c31\u662f\u6a21\u578b\u81ea\u7136\u751f\u6210\u8fc7\u7a0b\u4e2d\u7684\u771f\u5b9e\u5e7b\u89c9\uff0c\u800c\u4e0d\u662f\u4eba\u5de5\u6ce8\u5165\u6216\u63d0\u793a\u6a21\u578b\u6545\u610f\u5236\u9020\u7684\u5e7b\u89c9\u3002\u4f5c\u8005\u8ba4\u4e3a\uff0c\u975e\u81ea\u7136\u5e7b\u89c9\u867d\u7136\u5bb9\u6613\u63a7\u5236\u6807\u7b7e\uff0c\u4f46\u4e5f\u53ef\u80fd\u8ba9\u68c0\u6d4b\u4efb\u52a1\u53d8\u5f97\u8fc7\u4e8e\u7b80\u5355\uff0c\u9020\u6210\u201c\u770b\u8d77\u6765\u6548\u679c\u5f88\u597d\u201d\u7684\u5047\u8c61\u3002\u7b2c\u4e8c\uff0c\u6d4b\u8bd5\u6807\u7b7e\u5e94\u7ecf\u8fc7 human verification\uff0c\u56e0\u4e3a\u5e7b\u89c9\u68c0\u6d4b\u672c\u8d28\u4e0a\u662f\u8bc4\u4f30\u4efb\u52a1\uff0c\u5982\u679c\u6d4b\u8bd5\u6807\u7b7e\u4e0d\u53ef\u9760\uff0c\u5c31\u5f88\u96be\u516c\u5e73\u6bd4\u8f83\u4e0d\u540c detector \u7684\u6c34\u5e73\u3002<\/p>\n\n\n\n

\u7b2c\u4e09\uff0cbenchmark \u5e94\u8986\u76d6 long-context RAG tasks\u3002RAG \u662f\u5f53\u524d\u5927\u6a21\u578b\u843d\u5730\u7684\u91cd\u8981\u4f7f\u7528\u65b9\u5f0f\uff0c\u800c\u957f\u4e0a\u4e0b\u6587\u4e5f\u662f\u73b0\u5b9e\u4efb\u52a1\u4e2d\u5e38\u89c1\u60c5\u51b5\u3002\u7b2c\u56db\uff0cbenchmark \u5e94\u63d0\u4f9b realistic noisy training labels\u3002\u73b0\u5b9e\u4e2d\u5f88\u591a detector \u4e0d\u53ef\u80fd\u62ff\u5230\u5b8c\u7f8e\u4eba\u5de5\u6807\u7b7e\uff0c\u5e38\u5e38\u4f9d\u8d56 LLM-as-a-Judge\u3001\u5f31\u76d1\u7763\u6216\u8d28\u91cf\u4e0d\u7a33\u5b9a\u7684\u4eba\u5de5\u6807\u6ce8\u3002\u56e0\u6b64\uff0c\u4e00\u4e2a\u597d\u7684 benchmark \u4e0d\u5e94\u53ea\u7ed9\u5e72\u51c0\u6d4b\u8bd5\u96c6\uff0c\u4e5f\u5e94\u8be5\u63d0\u4f9b\u5e26\u566a\u58f0\u8bad\u7ec3\u6807\u7b7e\uff0c\u7528\u6765\u6d4b\u8bd5\u6a21\u578b\u5bf9\u6807\u7b7e\u566a\u58f0\u7684\u9c81\u68d2\u6027\u3002<\/p>\n\n\n\n

\u540e\u4e09\u4e2a\u5c5e\u6027\u5219\u662f\u8986\u76d6\u9762\u7684\u8981\u6c42\uff1abenchmark \u5e94\u5305\u542b\u4e0d\u540c\u7c7b\u578b\u7684 hallucination\uff0c\u6bd4\u5982\u4e0e\u4e0a\u4e0b\u6587\u77db\u76fe\u7684 intrinsic hallucination\uff0c\u4ee5\u53ca\u4e0a\u4e0b\u6587\u65e0\u6cd5\u9a8c\u8bc1\u7684 extrinsic hallucination\uff1b\u8fd8\u5e94\u6765\u81ea\u591a\u4e2a LLM\uff0c\u907f\u514d\u53ea\u6d4b\u8bd5\u67d0\u4e00\u4e2a\u6a21\u578b\u7684\u751f\u6210\u98ce\u683c\uff1b\u540c\u65f6\u5e94\u8986\u76d6\u591a\u4e2a\u9886\u57df\uff0c\u6d4b\u8bd5\u68c0\u6d4b\u5668\u8de8\u9886\u57df\u6cdb\u5316\u80fd\u529b\u3002<\/p>\n\n\n\n

\u4f5c\u8005\u7528\u8fd9\u4e03\u4e2a\u5c5e\u6027\u91cd\u65b0\u5ba1\u67e5\u5df2\u6709 benchmark\uff0c\u53d1\u73b0\u73b0\u6709\u6570\u636e\u96c6\u5404\u6709\u4fa7\u91cd\uff0c\u4f46\u90fd\u4e0d\u5b8c\u6574\u3002\u5c24\u5176\u7f3a\u5931\u6700\u4e25\u91cd\u7684\u662f\u957f\u4e0a\u4e0b\u6587 RAG benchmark \u548c realistic noisy labels\u3002\u8fd9\u4e5f\u662f TRIVIA+ \u88ab\u63d0\u51fa\u7684\u76f4\u63a5\u539f\u56e0\u3002<\/p>\n\n\n\n

4.TRIVIA+ \u6570\u636e\u96c6\u6784\u5efa\u4e0e\u4eba\u5de5\u6807\u6ce8\uff08Proposed Benchmark\uff09<\/h2>\n\n\n\n

TRIVIA+ \u662f\u4f5c\u8005\u63d0\u51fa\u7684\u65b0 RAG-based hallucination detection benchmark\u3002\u5b83\u7684\u6570\u636e\u6765\u6e90\u6bd4\u8f83\u5e7f\uff0c\u5305\u542b TRIVIAQA\u3001NaturalQuestions\u3001MS-MARCO\u3001CovidQA \u548c DROP \u7b49\u591a\u4e2a\u95ee\u7b54\u6570\u636e\u96c6\uff0c\u8986\u76d6 Wikipedia\u3001web search\u3001medical documents \u548c paragraph reasoning \u7b49\u9886\u57df\u3002\u6bcf\u4e2a\u6837\u672c\u90fd\u5305\u542b reference context\u3001question \u548c\u591a\u4e2a LLM \u7684\u56de\u7b54\u3002\u751f\u6210\u6a21\u578b\u5305\u62ec\u4e00\u4e2a\u5546\u7528 SOTA LLM\u3001Gemma-7B \u548c Mixtral 8x7B\u3002<\/p>\n\n\n\n

\u4e3a\u4e86\u63d0\u9ad8\u5e7b\u89c9\u6837\u672c\u6bd4\u4f8b\uff0c\u540c\u65f6\u4e0d\u7834\u574f\u201c\u81ea\u7136\u751f\u6210\u201d\u7684\u6027\u8d28\uff0c\u4f5c\u8005\u6ca1\u6709\u76f4\u63a5\u8ba9\u6a21\u578b\u6545\u610f\u72af\u9519\uff0c\u800c\u662f\u91c7\u7528\u4e86\u4e00\u4e2a\u8fc7\u6ee4\u7b56\u7565\uff1a\u5148\u7528\u5546\u7528 SOTA LLM \u56de\u7b54\u95ee\u9898\uff0c\u518d\u7528 ROUGE \u5206\u6570\u7b5b\u9009\u751f\u6210\u7b54\u6848\u548c\u6807\u51c6\u7b54\u6848\u91cd\u5408\u5ea6\u8f83\u4f4e\u7684\u6837\u672c\u3002\u4f5c\u8005\u7684\u5047\u8bbe\u662f\uff0c\u4f4e\u91cd\u5408\u5ea6\u56de\u7b54\u66f4\u53ef\u80fd\u5305\u542b\u5e7b\u89c9\u3002\u8fd9\u6837\u505a\u53ef\u4ee5\u63d0\u9ad8\u4eba\u5de5\u6807\u6ce8\u8d44\u6e90\u7684\u5229\u7528\u7387\uff0c\u4f46\u6a21\u578b\u56de\u7b54\u672c\u8eab\u4ecd\u7136\u662f\u81ea\u7136\u751f\u6210\u7684\uff0c\u6ca1\u6709\u88ab\u4eba\u4e3a\u6ce8\u5165\u9519\u8bef\u3002<\/p>\n\n\n\n

\u6807\u6ce8\u90e8\u5206\u662f\u8fd9\u7bc7\u8bba\u6587\u6bd4\u8f83\u624e\u5b9e\u7684\u5730\u65b9\u3002TRIVIA+ \u91c7\u7528 sentence-level \u6807\u6ce8\uff0c\u6bcf\u4e2a\u53e5\u5b50\u88ab\u6807\u6210 Supported\u3001Contradicted\u3001Not Mentioned \u6216 Supplementary\u3002\u968f\u540e\u518d\u805a\u5408\u5230 response-level \u4e8c\u5206\u7c7b\u6807\u7b7e\uff0c\u5176\u4e2d Contradicted \u548c Not Mentioned \u88ab\u89c6\u4e3a unfaithful\uff0cSupported \u548c Supplementary \u88ab\u89c6\u4e3a faithful\u3002\u8fd9\u4e2a\u8bbe\u7f6e\u548c RAG \u573a\u666f\u5f88\u8d34\uff0c\u56e0\u4e3a\u5b83\u5173\u5fc3\u7684\u662f\u56de\u7b54\u662f\u5426\u5fe0\u5b9e\u4e8e\u7ed9\u5b9a\u4e0a\u4e0b\u6587\u3002<\/p>\n\n\n\n

\u4f5c\u8005\u8fd8\u8bbe\u8ba1\u4e86\u591a\u8f6e\u591a\u6295\u7968\u6807\u6ce8\u6d41\u7a0b\u3002\u7b2c\u4e00\u8f6e\u4e2d\uff0c\u6bcf\u4e2a\u6837\u672c\u5148\u7531\u4e24\u540d\u6807\u6ce8\u8005\u6807\u6ce8\uff1b\u5982\u679c\u6709\u5206\u6b67\uff0c\u5c31\u8ffd\u52a0\u4e24\u540d\uff1b\u5982\u679c\u4ecd\u6ca1\u6709\u660e\u786e\u591a\u6570\uff0c\u518d\u8ffd\u52a0\u4e24\u540d\uff0c\u56e0\u6b64\u6bcf\u4e2a\u6837\u672c\u6700\u591a\u53ef\u6709\u516d\u4e2a\u6807\u6ce8\u3002\u7b2c\u4e8c\u8f6e\u5219\u5148\u7528 Dawid\u2013Skene \u6a21\u578b\u8fc7\u6ee4\u6389\u4f4e\u8d28\u91cf\u6807\u6ce8\u8005\uff0c\u518d\u8ba9\u5269\u4f59\u6570\u636e\u7531\u4e09\u540d\u6807\u6ce8\u8005\u5b8c\u6210\u3002\u8fd9\u6837\u505a\u7684\u76ee\u7684\u4e0d\u662f\u8ffd\u6c42\u6d41\u7a0b\u590d\u6742\uff0c\u800c\u662f\u627f\u8ba4\u957f\u4e0a\u4e0b\u6587 RAG \u6807\u6ce8\u672c\u8eab\u5f88\u96be\uff0c\u9700\u8981\u901a\u8fc7\u591a\u6295\u7968\u673a\u5236\u63d0\u9ad8\u6807\u7b7e\u53ef\u9760\u6027\u3002<\/p>\n\n\n\n

\u4ece\u6570\u636e\u7edf\u8ba1\u770b\uff0cTRIVIA+ \u7684\u5e73\u5747\u4e0a\u4e0b\u6587\u957f\u5ea6\u8fbe\u5230 9.3K \u5b57\u7b26\uff0c\u6700\u5927\u8fbe\u5230 94K \u5b57\u7b26\uff0c\u660e\u663e\u957f\u4e8e\u5df2\u6709 RAG-based HDB\u3002\u5b83\u5171\u6709 3224 \u4e2a\u6837\u672c\uff0c\u5e7b\u89c9\u6bd4\u4f8b\u7ea6 35%\uff0c\u6765\u81ea 3 \u4e2a LLM \u548c\u591a\u4e2a\u9886\u57df\u3002\u76f8\u6bd4 HaluEval\u3001RAGTruth \u548c Dolly(NC)\uff0cTRIVIA+ \u66f4\u80fd\u4ee3\u8868\u957f\u4e0a\u4e0b\u6587 RAG \u573a\u666f\u3002<\/p>\n\n\n\n

\"Rethinking<\/figure>\n\n\n\n
\"Rethinking<\/figure>\n\n\n\n

5.\u566a\u58f0\u6807\u7b7e\u8bbe\u8ba1\u4e0e\u5b9e\u9a8c\u8bbe\u7f6e\uff08Noisy Labels and Experimental Setup\uff09<\/h2>\n\n\n\n

\u9664\u4e86\u5e72\u51c0\u4eba\u5de5\u6807\u7b7e\uff0cTRIVIA+ \u8fd8\u63d0\u4f9b\u56db\u7ec4 noisy labels\uff0c\u8fd9\u662f\u672c\u6587\u533a\u522b\u4e8e\u5f88\u591a benchmark \u7684\u5173\u952e\u8bbe\u8ba1\u3002\u4f5c\u8005\u8ba4\u4e3a\uff0c\u73b0\u5b9e\u4e2d\u8bad\u7ec3 hallucination detector \u5f88\u5c11\u80fd\u5b8c\u5168\u4f9d\u8d56\u9ad8\u8d28\u91cf\u4eba\u5de5\u6807\u7b7e\uff0c\u66f4\u591a\u65f6\u5019\u4f1a\u7528 LLM-as-a-Judge\u3001\u5f31\u76d1\u7763\u6216\u8d28\u91cf\u53c2\u5dee\u4e0d\u9f50\u7684\u4eba\u5de5\u6807\u6ce8\u3002\u56e0\u6b64\uff0cbenchmark \u5982\u679c\u53ea\u63d0\u4f9b\u5e72\u51c0\u6807\u7b7e\uff0c\u53cd\u800c\u4e0d\u80fd\u5f88\u597d\u53cd\u6620\u771f\u5b9e\u8bad\u7ec3\u73af\u5883\u3002<\/p>\n\n\n\n

\u56db\u7c7b\u566a\u58f0\u5305\u62ec\uff1a\u7b2c\u4e00\uff0cWeak Supervision\uff08WS\uff09\uff0c\u5373\u4f7f\u7528\u4e00\u4e2a\u5546\u7528 SOTA LLM \u4f5c\u4e3a judge \u6765\u5224\u65ad\u56de\u7b54\u662f\u5426\u5fe0\u5b9e\u4e8e\u4e0a\u4e0b\u6587\uff1b\u7b2c\u4e8c\uff0cDissenting Worker\uff08DW\uff09\uff0c\u6a21\u62df\u67d0\u4e9b\u6807\u6ce8\u8005\u7cfb\u7edf\u6027\u504f\u5dee\uff1b\u7b2c\u4e09\uff0cDissenting Label\uff08DL\uff09\uff0c\u6a21\u62df\u4eba\u5de5\u6807\u6ce8\u4e2d\u7684\u968f\u673a\u5206\u6b67\uff1b\u7b2c\u56db\uff0cRandom Flip\uff08RF\uff09\uff0c\u76f4\u63a5\u968f\u673a\u7ffb\u8f6c\u4e00\u90e8\u5206\u6807\u7b7e\u3002\u4f5c\u8005\u7279\u522b\u533a\u5206\u4e86 sample-dependent noise \u548c sample-independent noise\uff0c\u5176\u4e2d WS\u3001DW\u3001DL \u66f4\u63a5\u8fd1\u73b0\u5b9e\u4e2d\u7684\u6837\u672c\u76f8\u5173\u566a\u58f0\uff0c\u800c RF \u53ea\u662f\u4f5c\u4e3a\u5bf9\u7167\u3002\u5b9e\u9a8c\u4e2d\u566a\u58f0\u6bd4\u4f8b\u7edf\u4e00\u8bbe\u7f6e\u4e3a 15%\u3002<\/p>\n\n\n\n

\u5b9e\u9a8c\u90e8\u5206\u4e3b\u8981\u6bd4\u8f83\u591a\u79cd hallucination detector \u5728 RAG-based HDB \u4e0a\u7684\u8868\u73b0\u3002\u65e0\u76d1\u7763\u65b9\u6cd5\u5305\u62ec SelfCheckGPT \u548c LLM-as-a-Judge\uff1b\u5229\u7528\u6807\u7b7e\u7684\u65b9\u6cd5\u5305\u62ec few-shot prompt\u3001prompt-optimized \u65b9\u6cd5\uff0c\u4ee5\u53ca\u57fa\u4e8e Mistral-7B-Instruct-v0.2 \u7684 SFT\u3002\u4f5c\u8005\u4f7f\u7528 HaluEval\u3001RAGTruth\u3001Dolly(NC) \u548c TRIVIA+ \u8fdb\u884c\u6bd4\u8f83\uff0c\u5e76\u62a5\u544a Precision\u3001Recall\u3001F1 \u548c Accuracy\u3002<\/p>\n\n\n\n

\u8fd9\u4e2a\u5b9e\u9a8c\u8bbe\u8ba1\u6709\u4e24\u4e2a\u91cd\u70b9\uff1a\u4e00\u662f\u6bd4\u8f83\u4e0d\u540c detector \u5728\u81ea\u7136\/\u975e\u81ea\u7136 hallucination benchmark \u4e0a\u7684\u5dee\u8ddd\uff1b\u4e8c\u662f\u7814\u7a76 noisy labels \u5bf9\u8bad\u7ec3\u548c\u8bc4\u4f30\u7684\u5f71\u54cd\u3002\u4e5f\u5c31\u662f\u8bf4\uff0c\u8bba\u6587\u4e0d\u662f\u53ea\u95ee\u201c\u8c01\u5206\u6570\u6700\u9ad8\u201d\uff0c\u800c\u662f\u5728\u95ee\u201c\u54ea\u4e9b benchmark \u4f1a\u8ba9\u5206\u6570\u5931\u771f\uff0c\u54ea\u4e9b\u6807\u7b7e\u566a\u58f0\u4f1a\u5f71\u54cd detector \u5224\u65ad\u201d\u3002<\/p>\n\n\n\n

\"Rethinking<\/figure>\n\n\n\n

6.\u5b9e\u9a8c\u7ed3\u679c\u4e0e\u5206\u6790\uff08Experiments\uff09<\/h2>\n\n\n\n

\u9996\u5148\u770b\u68c0\u6d4b\u7ed3\u679c\u3002\u8bba\u6587\u5728 Table 4 \u4e2d\u6bd4\u8f83\u4e86\u591a\u4e2a detector \u5728\u56db\u4e2a RAG-QA HDB \u4e0a\u7684\u8868\u73b0\u3002\u6700\u660e\u663e\u7684\u73b0\u8c61\u662f\uff0c\u68c0\u6d4b\u5668\u5728 HaluEval \u4e0a\u8868\u73b0\u975e\u5e38\u9ad8\uff0cSFT \u7684 F1 \u751a\u81f3\u8fbe\u5230 0.996\uff1b\u4f46\u5728 RAGTruth\u3001Dolly(NC) \u548c TRIVIA+ \u8fd9\u4e9b\u81ea\u7136\u5e7b\u89c9\u6570\u636e\u96c6\u4e0a\uff0c\u6027\u80fd\u660e\u663e\u4e0b\u964d\uff0cF1 \u57fa\u672c\u4f4e\u4e8e 0.7\u3002\u8fd9\u4e2a\u7ed3\u679c\u548c Figure 1 \u7684\u89c2\u5bdf\u4e00\u81f4\uff1a\u975e\u81ea\u7136\u6784\u9020\u7684\u5e7b\u89c9\u66f4\u5bb9\u6613\u88ab\u6a21\u578b\u5206\u5f00\uff0c\u800c\u81ea\u7136\u751f\u6210\u7684 RAG \u5e7b\u89c9\u66f4\u96be\u68c0\u6d4b\u3002<\/p>\n\n\n\n

\u7b2c\u4e8c\u4e2a\u91cd\u8981\u53d1\u73b0\u662f\uff0cLLM-as-a-Judge \u7684\u8868\u73b0\u6bd4\u5f88\u591a\u4eba\u9884\u671f\u66f4\u5f3a\u3002\u5728 TRIVIA+ \u4e0a\uff0cLLM-as-a-Judge \u7684 F1 \u8fbe\u5230 0.694\uff0c\u548c few-shot \u65b9\u6cd5 0.692 \u63a5\u8fd1\uff0c\u5e76\u4e14\u9ad8\u4e8e SFT \u7684 0.663\u3002\u4f5c\u8005\u8ba4\u4e3a\uff0c\u8fd9\u53ef\u80fd\u548c\u8fd1\u671f LLM \u672c\u8eab\u80fd\u529b\u63d0\u5347\u4ee5\u53ca carefully engineered prompt \u6709\u5173\u3002\u8fd9\u4e2a\u7ed3\u679c\u633a\u6709\u610f\u601d\uff0c\u56e0\u4e3a\u5b83\u8bf4\u660e\u5728 RAG hallucination detection \u4e0a\uff0c\u7b80\u5355 judge baseline \u4e0d\u80fd\u8f7b\u6613\u88ab\u5ffd\u7565\u3002<\/p>\n\n\n\n

\u7b2c\u4e09\uff0c\u957f\u4e0a\u4e0b\u6587\u4f1a\u660e\u663e\u589e\u52a0\u68c0\u6d4b\u96be\u5ea6\u3002Table 3 \u6309\u4e0a\u4e0b\u6587\u957f\u5ea6\u5212\u5206 TRIVIA+ \u6837\u672c\u540e\u53d1\u73b0\uff0c\u6240\u6709 detector \u5728\u957f\u4e0a\u4e0b\u6587\u6837\u672c\uff08>5K characters\uff09\u4e0a\u90fd\u4f1a\u660e\u663e\u4e0b\u964d\u3002\u4f8b\u5982 SFT \u4ece\u77ed\u4e0a\u4e0b\u6587\u7684 0.725 \u964d\u5230\u957f\u4e0a\u4e0b\u6587\u7684 0.504\uff1bSelfCheckGPT \u4e5f\u51fa\u73b0\u7c7b\u4f3c\u4e0b\u964d\u3002\u8fd9\u8bf4\u660e TRIVIA+ \u4e0d\u53ea\u662f\u591a\u4e86\u4e00\u4e2a\u6570\u636e\u96c6\uff0c\u800c\u662f\u63d0\u4f9b\u4e86\u5df2\u6709\u77ed\u4e0a\u4e0b\u6587 benchmark \u5f88\u96be\u6d4b\u8bd5\u5230\u7684\u538b\u529b\u573a\u666f\u3002<\/p>\n\n\n\n

\u7b2c\u56db\uff0c\u6807\u7b7e\u566a\u58f0\u4f1a\u5f71\u54cd\u8bc4\u4f30\u7ed3\u8bba\u3002Table 5 \u663e\u793a\uff0c\u5982\u679c\u7528 noisy test labels \u8bc4\u4f30 detector\uff0c\u5f97\u5230\u7684 measured performance \u53ef\u80fd\u548c clean labels \u4e0a\u7684 true performance \u4e0d\u4e00\u81f4\u3002\u5c24\u5176\u662f LLM-based weak supervision \u6807\u7b7e\u53ef\u80fd\u5e26\u6765\u504f\u4e50\u89c2\u7684\u8bc4\u4f30\u7ed3\u679c\u3002\u6362\u53e5\u8bdd\u8bf4\uff0c\u5982\u679c\u6d4b\u8bd5\u6807\u7b7e\u672c\u8eab\u6765\u81ea\u4e0d\u53ef\u9760 judge\uff0c\u90a3\u4e48 detector \u7684\u5206\u6570\u4e5f\u53ef\u80fd\u88ab judge \u7684\u504f\u5dee\u6c61\u67d3\u3002<\/p>\n\n\n\n

\u7b2c\u4e94\uff0cnoisy train labels \u4e5f\u4f1a\u5f71\u54cd\u76d1\u7763\u5f0f\u68c0\u6d4b\u5668\u3002Table 6 \u8868\u660e\uff0cfew-shot \u548c prompt-optimized \u65b9\u6cd5\u56e0\u4e3a\u53ea\u5c40\u90e8\u4f7f\u7528\u5c11\u91cf\u6837\u672c\uff0c\u53d7\u5230\u566a\u58f0\u5f71\u54cd\u76f8\u5bf9\u6709\u9650\uff1b\u4f46\u5168\u5c40\u5fae\u8c03\u7684 SFT \u66f4\u5bb9\u6613\u88ab noisy labels \u62d6\u7d2f\u3002\u8fd9\u8bf4\u660e\u5728\u5e7b\u89c9\u68c0\u6d4b\u4efb\u52a1\u4e2d\uff0c\u9c81\u68d2\u5b66\u4e60\u4e0d\u662f\u53ef\u6709\u53ef\u65e0\u7684\u7ec6\u8282\uff0c\u800c\u662f\u9700\u8981\u5355\u72ec\u7814\u7a76\u7684\u95ee\u9898\u3002<\/p>\n\n\n\n

\u603b\u4f53\u6765\u770b\uff0c\u5b9e\u9a8c\u90e8\u5206\u5f97\u51fa\u7684\u7ed3\u8bba\u6bd4\u8f83\u6e05\u695a\uff1a\u5f53\u524d detector \u5728\u771f\u5b9e RAG hallucination detection \u4e0a\u8fd8\u8fdc\u6ca1\u5230 ceiling\uff1b\u957f\u4e0a\u4e0b\u6587\u662f\u4e00\u4e2a\u660e\u663e\u96be\u70b9\uff1bLLM-as-a-Judge \u662f\u4e00\u4e2a\u7b80\u5355\u4f46\u5f3a\u7684 baseline\uff1b\u6807\u7b7e\u566a\u58f0\u4f1a\u8ba9\u8bad\u7ec3\u548c\u8bc4\u4f30\u90fd\u53d8\u5f97\u4e0d\u7a33\u5b9a\u3002<\/p>\n\n\n\n

\"Rethinking<\/figure>\n\n\n\n
\"Rethinking<\/figure>\n\n\n\n
\"Rethinking<\/figure>\n\n\n\n

7.\u8d21\u732e\u3001\u5c40\u9650\u4e0e\u7ed3\u8bba\uff08Conclusion\uff09<\/h2>\n\n\n\n

\u672c\u6587\u7684\u4e3b\u8981\u8d21\u732e\u53ef\u4ee5\u6982\u62ec\u4e3a\u56db\u70b9\u3002\u7b2c\u4e00\uff0c\u63d0\u51fa hallucination detection benchmark \u7684\u4e03\u4e2a desiderata\uff0c\u4e3a\u8bc4\u4f30\u6570\u636e\u96c6\u8d28\u91cf\u63d0\u4f9b\u4e86\u7cfb\u7edf\u89c6\u89d2\u3002\u7b2c\u4e8c\uff0c\u57fa\u4e8e\u8fd9\u5957\u6807\u51c6\u5ba1\u67e5\u73b0\u6709 HDB\uff0c\u6307\u51fa\u5f53\u524d\u6700\u5927\u7f3a\u53e3\u662f\u957f\u4e0a\u4e0b\u6587 RAG benchmark \u548c realistic noisy training labels\u3002\u7b2c\u4e09\uff0c\u6784\u5efa\u5e76\u5f00\u6e90 TRIVIA+\uff0c\u63d0\u4f9b\u81ea\u7136\u751f\u6210\u5e7b\u89c9\u3001\u4eba\u7c7b\u9a8c\u8bc1\u6807\u7b7e\u3001\u957f\u4e0a\u4e0b\u6587\u3001\u591a\u9886\u57df\u3001\u591a\u6a21\u578b\u548c\u56db\u7c7b\u566a\u58f0\u6807\u7b7e\u3002\u7b2c\u56db\uff0c\u901a\u8fc7\u5b9e\u9a8c\u53d1\u73b0\u73b0\u6709 detector \u5728 RAG-based HDB \u4e0a\u4ecd\u6709\u5f88\u5927\u63d0\u5347\u7a7a\u95f4\uff0c\u4e14\u6807\u7b7e\u566a\u58f0\u4f1a\u663e\u8457\u5f71\u54cd\u8bad\u7ec3\u548c\u8bc4\u4f30\u3002<\/p>\n\n\n\n

\u8fd9\u7bc7\u8bba\u6587\u7684\u7814\u7a76\u95ee\u9898\u6293\u5f97\u6bd4\u8f83\u51c6\u3002\u5b83\u6ca1\u6709\u76f4\u63a5\u5377 detector \u5206\u6570\uff0c\u800c\u662f\u4ece\u8bc4\u6d4b\u57fa\u7840\u8bbe\u65bd\u5165\u624b\uff0c\u6307\u51fa\u201cbenchmark \u4e0d\u53ef\u9760\u201d\u4f1a\u5bfc\u81f4\u6574\u4e2a\u7814\u7a76\u65b9\u5411\u7684\u6bd4\u8f83\u5931\u771f\u3002\u5c24\u5176\u662f\u5728 RAG \u8d8a\u6765\u8d8a\u5e38\u89c1\u7684\u80cc\u666f\u4e0b\uff0c\u68c0\u6d4b\u5668\u4e0d\u80fd\u53ea\u5728\u77ed\u4e0a\u4e0b\u6587\u3001\u4eba\u5de5\u6784\u9020\u5e7b\u89c9\u6216\u5e72\u51c0\u6807\u7b7e\u4e0a\u8868\u73b0\u597d\uff0c\u800c\u5fc5\u987b\u9762\u5bf9\u957f\u4e0a\u4e0b\u6587\u3001\u81ea\u7136\u5e7b\u89c9\u548c\u566a\u58f0\u6807\u7b7e\u8fd9\u4e9b\u771f\u5b9e\u95ee\u9898\u3002<\/p>\n\n\n\n

\u8bba\u6587\u4e5f\u6709\u660e\u786e\u5c40\u9650\u3002TRIVIA+ \u4e3b\u8981\u5173\u6ce8 faithfulness\uff0c\u4e5f\u5c31\u662f\u56de\u7b54\u662f\u5426\u4e0e\u7ed9\u5b9a context \u4e00\u81f4\uff0c\u800c\u4e0d\u662f\u5b8c\u6574\u610f\u4e49\u4e0a\u7684 factuality\uff1b\u5b83\u805a\u7126\u77e5\u8bc6\u5bc6\u96c6\u578b QA\uff0c\u6ca1\u6709\u8986\u76d6 summarization\u3001translation\u3001\u591a\u8f6e\u5bf9\u8bdd\u7b49\u5176\u4ed6 reference-based \u4efb\u52a1\uff1b\u540c\u65f6\u5b83\u662f\u7eaf\u6587\u672c benchmark\uff0c\u6ca1\u6709\u6d89\u53ca\u591a\u6a21\u6001 RAG\u3002\u4f5c\u8005\u4e5f\u627f\u8ba4\uff0cROUGE-based prefiltering \u53ef\u80fd\u8ba9\u6570\u636e\u504f\u5411\u4f4e\u8bcd\u9762\u91cd\u5408\u7684\u9519\u8bef\uff0c\u5c3d\u7ba1\u4ed6\u4eec\u7684\u5206\u6790\u8ba4\u4e3a\u8fd9\u79cd\u504f\u5dee\u4e0d\u4f1a\u663e\u8457\u964d\u4f4e\u68c0\u6d4b\u96be\u5ea6\u3002<\/p>\n\n\n\n

\u6574\u4f53\u6765\u8bf4\uff0c\u8fd9\u7bc7\u6587\u7ae0\u9002\u5408\u4f5c\u4e3a\u201cRAG \u53ef\u9760\u6027 \/ hallucination evaluation \/ evidence grounding\u201d\u65b9\u5411\u7684\u9605\u8bfb\u6750\u6599\u3002\u5b83\u7684\u4ef7\u503c\u4e0d\u662f\u63d0\u51fa\u590d\u6742\u6a21\u578b\uff0c\u800c\u662f\u628a\u4e00\u4e2a\u57fa\u7840\u4f46\u5bb9\u6613\u88ab\u5ffd\u7565\u7684\u95ee\u9898\u8bb2\u6e05\u695a\uff1a\u6211\u4eec\u5230\u5e95\u5728\u7528\u4ec0\u4e48\u6570\u636e\u96c6\u5224\u65ad\u68c0\u6d4b\u5668\u662f\u5426\u53ef\u9760\uff1f \u5bf9\u540e\u7eed\u505a RAG \u7cfb\u7edf\u3001\u79d1\u7814\u8bc1\u636e\u52a9\u624b\u3001\u6587\u732e\u95ee\u7b54\u7cfb\u7edf\u90fd\u5f88\u6709\u53c2\u8003\u610f\u4e49\u3002<\/p>\n\n\n\n

\"Rethinking<\/figure>\n\n\n\n

<\/p>\n","protected":false},"excerpt":{"rendered":"

1.\u6458\u8981\uff08Abstract\uff09 \u672c\u6587\u7814\u7a76\u7684\u662f LLM hallucination detection benchmark\uff08\u5927\u6a21\u578b\u5e7b\u89c9\u68c0\u6d4b\u57fa\u51c6\uff09 \u7684\u8bc4\u6d4b\u95ee\u9898\u3002\u968f\u7740\u5927\u6a21\u578b\u88ab\u7528\u4e8e\u7535\u5546\u3001\u533b\u7597\u3001\u6cd5\u5f8b\u7b49\u771f\u5b9e\u573a\u666f\uff0c\u5e7b\u89c9\u95ee\u9898\u5df2\u7ecf\u4e0d\u53ea\u662f\u6a21\u578b\u6548\u679c\u95ee\u9898\uff0c\u800c\u662f\u76f4\u63a5\u5173\u7cfb\u5230\u751f\u6210\u5f0f AI \u7684\u5b89\u5168\u4f7f\u7528\u3002\u867d\u7136\u8fd1\u51e0\u5e74\u51fa\u73b0\u4e86\u5927\u91cf\u5e7b\u89c9\u68c0\u6d4b\u65b9\u6cd5\u548c\u68c0\u6d4b\u57fa\u51c6\uff0c\u4f46\u4f5c\u8005\u6307\u51fa\uff0c\u73b0\u6709 benchmark \u672c\u8eab\u5b58\u5728\u660e\u663e\u7f3a\u9677\uff1a\u5f88\u591a\u6570\u636e\u96c6\u5e76\u4e0d\u80fd\u771f\u5b9e\u53cd\u6620 RAG \u573a\u666f\u4e0b\u7684\u5927\u6a21\u578b\u5e7b\u89c9\uff0c\u4e5f\u7f3a\u5c11\u5bf9\u6807\u6ce8\u566a\u58f0\u7684\u7cfb\u7edf\u8bc4\u4f30\u3002 \u8bba\u6587\u9996\u5148\u63d0\u51fa\u4e00\u7ec4\u8861\u91cf ha … \u7ee7\u7eed\u9605\u8bfb →<\/span><\/a><\/p>\n","protected":false},"author":7,"featured_media":1387,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[],"class_list":["post-1385","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-rengongzhineng"],"_links":{"self":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1385","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/users\/7"}],"replies":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=1385"}],"version-history":[{"count":1,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1385\/revisions"}],"predecessor-version":[{"id":1395,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/posts\/1385\/revisions\/1395"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=\/wp\/v2\/media\/1387"}],"wp:attachment":[{"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=1385"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=1385"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.ndnlab.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=1385"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}