{"id":128,"date":"2026-06-09T22:29:14","date_gmt":"2026-06-09T22:29:14","guid":{"rendered":"https:\/\/blog.joezhou.top\/ai-benchmark\/"},"modified":"2026-06-09T22:29:14","modified_gmt":"2026-06-09T22:29:14","slug":"ai-benchmark","status":"publish","type":"post","link":"https:\/\/www.joezhou.top\/?p=128","title":{"rendered":"AI \u6a21\u578b\u7684\u8bc4\u4f30\u4e0e\u8bc4\u6d4b\u2014\u2014\u4ece Benchmark \u5230\u5b9e\u9645\u5e94\u7528\u7684\u6548\u679c\u8861\u91cf"},"content":{"rendered":"<h2>\u5f15\u8a00<\/h2>\n<p>2026 \u5e74\u7684\u4eca\u5929\uff0c\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5df2\u6df1\u5ea6\u878d\u5165\u5404\u884c\u5404\u4e1a\u2014\u2014\u4ece\u4ee3\u7801\u751f\u6210\u3001\u6587\u6863\u64b0\u5199\u5230\u5ba2\u670d\u5bf9\u8bdd\u3001\u77e5\u8bc6\u68c0\u7d22\uff0cAI \u7684\u89d2\u8272\u4ece&quot;\u65b0\u5947\u73a9\u5177&quot;\u53d8\u6210\u4e86&quot;\u751f\u4ea7\u529b\u5de5\u5177&quot;\u3002\u7136\u800c\uff0c\u4e00\u4e2a\u6838\u5fc3\u95ee\u9898\u59cb\u7ec8\u60ac\u800c\u672a\u51b3\uff1a<strong>\u6211\u4eec\u600e\u4e48\u77e5\u9053\u4e00\u4e2a\u6a21\u578b\u597d\u4e0d\u597d\uff1f<\/strong><\/p>\n<p>\u8fd9\u5e76\u4e0d\u662f\u4e00\u4e2a\u7b80\u5355\u7684\u95ee\u9898\u3002\u6a21\u578b\u5728\u516c\u5f00\u6392\u884c\u699c\u4e0a\u62ff\u4e86\u9ad8\u5206\uff0c\u4e0d\u4ee3\u8868\u5b83\u5728\u4f60\u7684\u4e1a\u52a1\u573a\u666f\u4e2d\u597d\u7528\uff1b\u5b83\u5728\u6570\u5b66\u63a8\u7406\u4e0a\u8868\u73b0\u51fa\u8272\uff0c\u4e0d\u4ee3\u8868\u5b83\u80fd\u51c6\u786e\u7406\u89e3\u4f60\u7684\u884c\u4e1a\u672f\u8bed\u3002\u672c\u6587\u5c06\u4ece\u8bc4\u6d4b\u4f53\u7cfb\u7684\u6f14\u8fdb\u3001Benchmark \u7684\u5c40\u9650\u6027\u3001\u5b9e\u9645\u573a\u666f\u7684\u6548\u679c\u8861\u91cf\u4e09\u4e2a\u65b9\u9762\uff0c\u6df1\u5165\u5256\u6790 AI \u6a21\u578b\u8bc4\u4f30\u7684\u5b8c\u6574\u56fe\u666f\u3002<\/p>\n<h2>\u4e00\u3001\u4ece&quot;\u8003\u5377&quot;\u5230&quot;\u9762\u8bd5&quot;\uff1a\u8bc4\u6d4b\u8303\u5f0f\u7684\u6f14\u8fdb<\/h2>\n<h3>1.1 \u7b2c\u4e00\u4ee3\u8bc4\u6d4b\uff1a\u6807\u51c6\u5316\u8003\u8bd5<\/h3>\n<p>AI \u8bc4\u6d4b\u7684\u7b2c\u4e00\u9636\u6bb5\uff0c\u53ef\u4ee5\u79f0\u4f5c&quot;\u6807\u51c6\u5316\u8003\u8bd5\u9636\u6bb5&quot;\u3002\u5176\u5178\u578b\u4ee3\u8868\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><strong>GLUE \/ SuperGLUE<\/strong>\uff1a\u65e9\u671f\u7684\u81ea\u7136\u8bed\u8a00\u7406\u89e3\u8bc4\u6d4b\u5957\u4ef6\uff0c\u6db5\u76d6\u60c5\u611f\u5206\u6790\u3001\u6587\u672c\u8574\u542b\u3001\u95ee\u7b54\u5339\u914d\u7b49\u4efb\u52a1\u3002<\/li>\n<li><strong>SQuAD<\/strong>\uff1a\u57fa\u4e8e\u7ef4\u57fa\u767e\u79d1\u6587\u7ae0\u7684\u9605\u8bfb\u7406\u89e3\u6570\u636e\u96c6\uff0c\u6d4b\u8bd5\u6a21\u578b\u80fd\u5426\u51c6\u786e\u5730\u4ece\u6587\u672c\u4e2d\u68c0\u7d22\u5e76\u56de\u7b54\u95ee\u9898\u3002<\/li>\n<li><strong>RACE<\/strong>\uff1a\u57fa\u4e8e\u4e2d\u56fd\u4e2d\u5b66\u751f\u82f1\u8bed\u9605\u8bfb\u7406\u89e3\u9898\u7684\u6570\u636e\u96c6\uff0c\u6d4b\u8bd5\u63a8\u7406\u80fd\u529b\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e00\u9636\u6bb5\u7684\u8bc4\u6d4b\u7279\u70b9\u662f\uff1a\u56fa\u5b9a\u6570\u636e\u96c6\u3001\u56fa\u5b9a\u8bc4\u4ef7\u6307\u6807\uff08\u5982\u51c6\u786e\u7387\u3001F1 \u5206\u6570\uff09\u3001\u53ef\u590d\u73b0\u6027\u5f3a\u3002\u6a21\u578b\u50cf\u5b66\u751f\u53c2\u52a0\u8003\u8bd5\u4e00\u6837\uff0c\u5728\u540c\u4e00\u5f20\u8bd5\u5377\u4e0a\u6bd4\u8f83\u6210\u7ee9\u3002<\/p>\n<p>\u7136\u800c\uff0c\u8fd9\u79cd&quot;\u6807\u51c6\u5316\u8003\u8bd5&quot;\u7684\u5c40\u9650\u6027\u5f88\u5feb\u5c31\u663e\u73b0\u51fa\u6765\u2014\u2014\u6a21\u578b\u5f00\u59cb\u5728\u516c\u5f00\u6570\u636e\u4e0a\u51fa\u73b0\u8fc7\u62df\u5408\uff0c\u751a\u81f3\u5728\u699c\u5355\u4e0a\u51fa\u73b0\u4e13\u95e8\u5237\u5206\u7684\u6280\u672f\u624b\u6bb5\u3002\u66f4\u4e25\u91cd\u7684\u662f\uff0c\u8003\u8bd5\u8868\u73b0\u597d\u4e0d\u7b49\u4e8e\u771f\u5b9e\u80fd\u529b\u597d\u3002\u4e00\u4e2a\u5728 SQuAD \u4e0a\u51c6\u786e\u7387 95% \u7684\u6a21\u578b\uff0c\u53ef\u80fd\u5728\u7528\u6237\u95ee\u4e00\u4e2a\u7b80\u5355\u5e38\u8bc6\u6027\u95ee\u9898\u65f6\u7b54\u5f97\u83ab\u540d\u5176\u5999\u3002<\/p>\n<h3>1.2 \u7b2c\u4e8c\u4ee3\u8bc4\u6d4b\uff1a\u591a\u7ef4\u5ea6\u80fd\u529b\u77e9\u9635<\/h3>\n<p>\u968f\u7740 GPT-3\uff082020\uff09\u3001ChatGPT\uff082022\uff09\u7b49\u5bf9\u8bdd\u5f0f\u6a21\u578b\u7684\u5174\u8d77\uff0c\u4e1a\u754c\u610f\u8bc6\u5230\u4f20\u7edf\u5355\u4e00\u6307\u6807\u5df2\u7ecf\u65e0\u6cd5\u523b\u753b\u6a21\u578b\u80fd\u529b\u3002\u4e8e\u662f\u51fa\u73b0\u4e86\u7b2c\u4e8c\u4ee3\u8bc4\u6d4b\u8303\u5f0f\u2014\u2014<strong>\u591a\u7ef4\u5ea6\u80fd\u529b\u77e9\u9635<\/strong>\u3002<\/p>\n<p>\u5178\u578b\u7684\u8bc4\u6d4b\u4f53\u7cfb\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><strong>MMLU<\/strong>\uff08\u5927\u89c4\u6a21\u591a\u4efb\u52a1\u8bed\u8a00\u7406\u89e3\uff09\uff1a\u6db5\u76d6 57 \u4e2a\u5b66\u79d1\u7684\u77e5\u8bc6\u95ee\u7b54\uff0c\u8986\u76d6\u4eba\u6587\u3001\u793e\u79d1\u3001\u81ea\u7136\u79d1\u5b66\u3001\u533b\u5b66\u3001\u6cd5\u5f8b\u7b49\u9886\u57df\u3002<\/li>\n<li><strong>HellaSwag<\/strong>\uff1a\u5e38\u8bc6\u63a8\u7406\uff0c\u6d4b\u8bd5\u6a21\u578b\u662f\u5426\u80fd\u6839\u636e\u4e0a\u4e0b\u6587\u9009\u51fa\u5408\u7406\u7684\u540e\u7eed\u5185\u5bb9\u3002<\/li>\n<li><strong>GSM8K<\/strong>\uff1a\u5c0f\u5b66\u6570\u5b66\u5e94\u7528\u9898\uff0c\u6d4b\u8bd5\u6570\u5b66\u63a8\u7406\u80fd\u529b\u3002<\/li>\n<li><strong>HumanEval \/ MBPP<\/strong>\uff1a\u4ee3\u7801\u751f\u6210\uff0c\u6d4b\u8bd5\u7f16\u7a0b\u80fd\u529b\u3002<\/li>\n<li><strong>BIG-Bench<\/strong>\uff1a\u8d85\u8fc7 200 \u9879\u4efb\u52a1\u7684\u8d85\u5927\u89c4\u6a21\u8bc4\u6d4b\u5957\u4ef6\u3002<\/li>\n<\/ul>\n<p>\u8fd9\u4e00\u9636\u6bb5\u7684\u8fdb\u6b65\u5728\u4e8e\uff1a\u4e0d\u518d\u628a\u6a21\u578b\u770b\u6210\u4e00\u4e2a&quot;\u9ed1\u76d2\u5206\u6570&quot;\uff0c\u800c\u662f\u7406\u89e3\u5b83\u7684\u80fd\u529b\u5206\u5e03\u2014\u2014\u8fd9\u4e2a\u6a21\u578b\u66f4\u64c5\u957f\u6570\u5b66\uff0c\u90a3\u4e2a\u6a21\u578b\u7684\u4e2d\u6587\u77e5\u8bc6\u66f4\u4e30\u5bcc\u3002<\/p>\n<h3>1.3 \u7b2c\u4e09\u4ee3\u8bc4\u6d4b\uff1a\u573a\u666f\u5316\u4e0e\u7528\u6237\u5bf9\u9f50<\/h3>\n<p>\u8fdb\u5165 2024-2026 \u5e74\uff0c\u8bc4\u6d4b\u8303\u5f0f\u7684\u7b2c\u4e09\u6b21\u6f14\u8fdb\u6b63\u5728\u53d1\u751f\u3002\u4eba\u4eec\u53d1\u73b0\uff1a<strong>\u7ef4\u5ea6\u77e9\u9635\u4ecd\u7136\u4e0d\u591f<\/strong>\u3002\u6a21\u578b\u4ec0\u4e48\u90fd\u4f1a\u4e00\u70b9\uff0c\u4f46\u5728\u5177\u4f53\u573a\u666f\u4e2d\u53ef\u80fd\u8868\u73b0\u5f97&quot;\u667a\u80fd\u4f46\u4e0d\u597d\u7528&quot;\u3002<\/p>\n<p>\u7b2c\u4e09\u4ee3\u8bc4\u6d4b\u5173\u6ce8\u7684\u662f\uff1a<\/p>\n<ul>\n<li><strong>\u6307\u4ee4\u9075\u5faa\uff08Instruction Following\uff09<\/strong>\uff1a\u6a21\u578b\u80fd\u5426\u51c6\u786e\u7406\u89e3\u7528\u6237\u610f\u56fe\uff0c\u800c\u4e0d\u662f\u5b57\u9762\u610f\u601d\uff1f\u4f8b\u5982\u7528\u6237\u8bf4&quot;\u7528\u7b80\u5355\u7684\u8bdd\u89e3\u91ca\u91cf\u5b50\u7ea0\u7f20&quot;\uff0c\u6a21\u578b\u662f\u5426\u771f\u7684\u505a\u5230\u4e86\u7b80\u5355\u6613\u61c2\uff1f<\/li>\n<li><strong>\u957f\u6587\u672c\u7406\u89e3<\/strong>\uff1a\u5f53\u4e0a\u4e0b\u6587\u8fbe\u5230\u51e0\u5341\u4e0a\u767e K Token \u65f6\uff0c\u6a21\u578b\u662f\u5426\u8fd8\u80fd\u51c6\u786e\u68c0\u7d22\u548c\u63a8\u7406\uff1f<\/li>\n<li><strong>\u591a\u8f6e\u5bf9\u8bdd\u4e00\u81f4\u6027<\/strong>\uff1a\u5728\u591a\u8f6e\u5bf9\u8bdd\u4e2d\uff0c\u6a21\u578b\u662f\u5426\u80fd\u8bb0\u4f4f\u4e0a\u4e0b\u6587\u3001\u4fdd\u6301\u7acb\u573a\u4e00\u81f4\uff1f<\/li>\n<li><strong>\u5b89\u5168\u4e0e\u5bf9\u9f50<\/strong>\uff1a\u9762\u5bf9\u8d8a\u72f1\u653b\u51fb\u3001\u8bf1\u5bfc\u6027\u95ee\u9898\u65f6\uff0c\u6a21\u578b\u662f\u5426\u5b89\u5168\u53ef\u9760\uff1f\u504f\u89c1\u548c\u6709\u5bb3\u5185\u5bb9\u662f\u5426\u5f97\u5230\u6709\u6548\u63a7\u5236\uff1f<\/li>\n<\/ul>\n<p>\u4ee3\u8868\u8bc4\u6d4b\u5982 MT-Bench\uff08\u591a\u8f6e\u5bf9\u8bdd\uff09\u3001Chatbot Arena\uff08\u4eba\u5de5\u8bc4\u5206\uff09\u3001Safety Benchmarks \u7b49\u3002Chatbot Arena \u91c7\u7528 Elo \u8bc4\u5206\u673a\u5236\u7684\u65b9\u6848\u5c24\u5176\u503c\u5f97\u5173\u6ce8\u2014\u2014\u5b83\u8ba9\u771f\u5b9e\u7528\u6237\u5728\u76f2\u6d4b\u4e2d\u7ed9\u6a21\u578b\u6253\u5206\uff0c\u67d0\u79cd\u7a0b\u5ea6\u4e0a\u56de\u5f52\u4e86&quot;AI \u7684\u80fd\u529b\uff0c\u8bf4\u5230\u5e95\u662f\u7531\u4eba\u8bf4\u4e86\u7b97&quot;\u7684\u539f\u5219\u3002<\/p>\n<h2>\u4e8c\u3001Benchmark \u7684\u8ff7\u601d\uff1a\u4e3a\u4ec0\u4e48\u9ad8\u5206\u4e0d\u4e00\u6837\u5f97\u597d\u7528\uff1f<\/h2>\n<h3>2.1 \u6570\u636e\u6cc4\u9732\u95ee\u9898<\/h3>\n<p>\u8fd9\u662f\u4e00\u4e2a\u88ab\u5e7f\u6cdb\u8ba8\u8bba\u4f46\u81f3\u4eca\u672a\u80fd\u6839\u672c\u89e3\u51b3\u7684\u95ee\u9898\u3002\u8bad\u7ec3\u6570\u636e\u96c6\u7684\u89c4\u6a21\u5df2\u7ecf\u4ece GPT-3 \u7684 1750 \u4ebf\u53c2\u6570\u6269\u5c55\u5230\u4eca\u5929\u5341\u4e07\u4ebf\u7ea7\u522b\uff0c\u800c\u8bad\u7ec3\u6570\u636e\u5f80\u5f80\u8986\u76d6\u4e86\u4e92\u8054\u7f51\u4e0a\u51e0\u4e4e\u6240\u6709\u516c\u5f00\u6587\u672c\u2014\u2014\u5305\u62ec\u540e\u6765\u88ab\u7528\u4f5c\u8bc4\u6d4b\u7684\u6570\u636e\u96c6\u3002<\/p>\n<p>\u5f53\u4e00\u4e2a\u6a21\u578b\u5728\u8bad\u7ec3\u9636\u6bb5\u5c31\u5df2\u7ecf&quot;\u89c1\u8fc7&quot;\u4e86\u8bc4\u6d4b\u7b54\u6848\u65f6\uff0c\u8bc4\u6d4b\u6210\u7ee9\u5c31\u5931\u53bb\u4e86\u610f\u4e49\u3002\u8fd9\u5c31\u50cf\u4e00\u573a\u5f00\u5377\u8003\u8bd5\uff0c\u5b66\u751f\u8003\u524d\u5df2\u7ecf\u7ffb\u8fc7\u4e86\u6807\u51c6\u7b54\u6848\u3002<\/p>\n<p>\u89e3\u51b3\u601d\u8def\uff1a\u6784\u5efa\u975e\u516c\u5f00\u7684\u3001\u5b9a\u671f\u66f4\u65b0\u7684\u52a8\u6001\u8bc4\u6d4b\u96c6\uff0c\u5982 Nvidia \u7684 Nemotron \u7cfb\u5217\u5c31\u91c7\u7528\u65f6\u95f4\u6233\u5207\u5272\u7684\u65b9\u5f0f\uff0c\u786e\u4fdd\u8bad\u7ec3\u6570\u636e\u548c\u8bc4\u6d4b\u6570\u636e\u7684\u65f6\u95f4\u7a97\u53e3\u4e0d\u91cd\u53e0\u3002<\/p>\n<h3>2.2 \u6307\u6807\u7684\u8bef\u5bfc\u6027<\/h3>\n<p>Benchmark \u6307\u6807\u4e5f\u4f1a\u4ea7\u751f\u8bef\u5bfc\u3002\u4f8b\u5982\uff1a<\/p>\n<ul>\n<li><strong>\u51c6\u786e\u7387<\/strong>\uff1a\u5728\u7c7b\u522b\u4e0d\u5e73\u8861\u7684\u6570\u636e\u96c6\u4e0a\uff0c\u4e00\u4e2a\u6a21\u578b\u53ea\u9700\u8981\u9884\u6d4b&quot;\u591a\u6570\u7c7b&quot;\u5c31\u80fd\u83b7\u5f97\u9ad8\u51c6\u786e\u7387\uff0c\u4f46\u5b8c\u5168\u6ca1\u6709\u5b9e\u9645\u4ef7\u503c\u3002<\/li>\n<li><strong>BLEU \/ ROUGE<\/strong>\uff1a\u7528\u4e8e\u6587\u672c\u751f\u6210\u8bc4\u6d4b\u7684\u6307\u6807\uff0c\u504f\u91cd n-gram \u5339\u914d\uff0c\u53ef\u80fd\u8bc4\u5206\u76f8\u5f53\u9ad8\u4f46\u751f\u6210\u7684\u662f&quot;\u6b7b\u677f\u7684\u5957\u8bdd&quot;\uff0c\u800c\u4e00\u4e2a\u6709\u521b\u610f\u7684\u3001\u63aa\u8f9e\u65b0\u9896\u7684\u56de\u7b54\u53cd\u800c\u5f97\u5206\u66f4\u4f4e\u3002<\/li>\n<li><strong>\u56f0\u60d1\u5ea6\uff08Perplexity\uff09<\/strong>\uff1a\u8d8a\u4f4e\u8d8a\u597d\uff0c\u4f46\u53ef\u80fd\u5b58\u5728\u8fc7\u62df\u5408\u73b0\u8c61\u2014\u2014\u6a21\u578b\u5bf9\u8bad\u7ec3\u6570\u636e\u6781\u4f4e\u7684\u56f0\u60d1\u5ea6\u4e0d\u4ee3\u8868\u5b83\u80fd\u6cdb\u5316\u5230\u65b0\u573a\u666f\u3002<\/li>\n<\/ul>\n<p>\u5b66\u672f\u754c\u5c1d\u8bd5\u7528\u66f4\u590d\u6742\u7684\u6307\u6807\uff08\u5982 BERTScore\u3001Unieval\uff09\u66ff\u4ee3\u4f20\u7edf\u6307\u6807\uff0c\u4f46\u8fd9\u4e9b\u65b0\u6307\u6807\u672c\u8eab\u4e5f\u4f9d\u8d56\u53e6\u4e00\u4e2a\u6a21\u578b\u6765\u8bc4\u5206\uff0c\u5f15\u5165\u4e86\u65b0\u7684\u504f\u5dee\u3002<\/p>\n<h3>2.3 \u573a\u666f\u9002\u914d\u6027\u7684\u7f3a\u5931<\/h3>\n<p>\u8fd9\u662f Benchmark \u6700\u6839\u672c\u7684\u95ee\u9898\uff1a<strong>Benchmark \u6d4b\u8bd5\u7684\u662f\u4e00\u822c\u80fd\u529b\uff0c\u800c\u5b9e\u9645\u90e8\u7f72\u9700\u8981\u7684\u662f\u7279\u5b9a\u573a\u666f\u80fd\u529b\u3002<\/strong><\/p>\n<ul>\n<li>\u4e00\u4e2a\u901a\u7528\u8bc4\u6d4b\u699c TOP 1 \u7684\u6a21\u578b\uff0c\u5728\u533b\u5b66\u95ee\u7b54\u573a\u666f\u4e0b\u53ef\u80fd\u4e0d\u5982\u4e00\u4e2a\u7ecf\u8fc7\u9886\u57df\u5fae\u8c03\u7684\u5c0f\u6a21\u578b\u3002<\/li>\n<li>\u4e00\u4e2a\u82f1\u6587\u80fd\u529b\u6781\u5f3a\u7684\u6a21\u578b\uff0c\u5728\u4e2d\u6587\u573a\u666f\u4e0b\u53ef\u80fd\u8868\u73b0\u4e0d\u4f73\u3002<\/li>\n<li>\u4e00\u4e2a\u63a8\u7406\u80fd\u529b\u5f3a\u5927\u7684\u6a21\u578b\uff0c\u53ef\u80fd\u5728\u5ef6\u8fdf\u8981\u6c42\u6781\u4f4e\u7684\u5b9e\u65f6\u573a\u666f\u4e2d\u4e0d\u9002\u7528\u3002<\/li>\n<\/ul>\n<p><strong>\u573a\u666f\u9002\u914d\u6027\u7684\u7f3a\u5931<\/strong>\u610f\u5473\u7740\uff1a\u4f01\u4e1a\u4e0d\u80fd\u53ea\u770b\u6392\u884c\u699c\u9009\u6a21\u578b\uff0c\u800c\u5fc5\u987b\u7528\u81ea\u5df1\u7684\u4e1a\u52a1\u6570\u636e\u505a\u5b9a\u5236\u5316\u8bc4\u4f30\u3002<\/p>\n<h2>\u4e09\u3001\u5b9e\u6218\uff1a\u5982\u4f55\u6784\u5efa\u4f01\u4e1a\u7ea7 AI \u8bc4\u4f30\u4f53\u7cfb<\/h2>\n<h3>3.1 \u660e\u786e\u8bc4\u4f30\u76ee\u6807<\/h3>\n<p>\u5728\u5f00\u59cb\u8bc4\u4f30\u4e4b\u524d\uff0c\u9996\u5148\u8981\u56de\u7b54\u4e09\u4e2a\u95ee\u9898\uff1a<\/p>\n<ol>\n<li><strong>\u8bc4\u4f30\u8c01\uff1f<\/strong>\uff08\u54ea\u4e2a\u6a21\u578b\u3001\u54ea\u4e2a\u7248\u672c\uff1f\uff09<\/li>\n<li><strong>\u8bc4\u4f30\u4ec0\u4e48\uff1f<\/strong>\uff08\u4ec0\u4e48\u80fd\u529b\u7ef4\u5ea6\uff1f\u4ec0\u4e48\u4e1a\u52a1\u573a\u666f\uff1f\uff09<\/li>\n<li><strong>\u600e\u4e48\u7b97\u597d\uff1f<\/strong>\uff08\u5b9a\u91cf\u6307\u6807\uff1f\u5b9a\u6027\u8bc4\u4f30\uff1fA\/B \u6d4b\u8bd5\uff1f\uff09<\/li>\n<\/ol>\n<p>\u8bc4\u4f30\u76ee\u6807\u51b3\u5b9a\u4e86\u540e\u7eed\u6240\u6709\u5de5\u4f5c\u7684\u65b9\u5411\u3002\u4f8b\u5982\uff1a<\/p>\n<ul>\n<li>\u5ba2\u670d\u573a\u666f\uff1a\u5173\u6ce8\u51c6\u786e\u6027\uff08\u56de\u7b54\u662f\u5426\u6b63\u786e\uff09\u3001\u4e00\u81f4\u6027\uff08\u662f\u5426\u524d\u540e\u77db\u76fe\uff09\u3001\u5b89\u5168\u6027\uff08\u662f\u5426\u6cc4\u9732\u9690\u79c1\uff09\u3002<\/li>\n<li>\u4ee3\u7801\u751f\u6210\u573a\u666f\uff1a\u5173\u6ce8\u8bed\u6cd5\u6b63\u786e\u6027\u3001\u7f16\u7a0b\u98ce\u683c\u662f\u5426\u7b26\u5408\u89c4\u8303\u3001\u751f\u6210\u901f\u5ea6\u3002<\/li>\n<li>\u5185\u5bb9\u521b\u4f5c\u573a\u666f\uff1a\u5173\u6ce8\u521b\u9020\u529b\u3001\u53ef\u8bfb\u6027\u3001\u4e8b\u5b9e\u51c6\u786e\u6027\u3002<\/li>\n<\/ul>\n<h3>3.2 \u6784\u5efa\u573a\u666f\u5316\u8bc4\u6d4b\u6570\u636e\u96c6<\/h3>\n<p>\u516c\u5f00\u6570\u636e\u96c6\u4e0d\u591f\u7528\uff0c\u4f01\u4e1a\u9700\u8981\u6784\u5efa\u81ea\u5df1\u7684\u8bc4\u6d4b\u96c6\u3002\u6b65\u9aa4\u5305\u62ec\uff1a<\/p>\n<ol>\n<li><strong>\u91c7\u96c6\u771f\u5b9e\u7528\u6237\u8bf7\u6c42<\/strong>\uff1a\u4ece\u65e5\u5fd7\u4e2d\u63d0\u53d6\u7528\u6237\u8f93\u5165\uff0c\u8131\u654f\u540e\u6784\u5efa\u6d4b\u8bd5\u6837\u672c\u3002<\/li>\n<li><strong>\u6807\u6ce8\u6807\u51c6\u7b54\u6848<\/strong>\uff1a\u7531\u4e1a\u52a1\u4e13\u5bb6\u7ed9\u51fa&quot;\u7406\u60f3\u56de\u7b54&quot;\uff0c\u4f5c\u4e3a\u53c2\u8003\u6807\u51c6\u3002<\/li>\n<li><strong>\u6784\u5efa\u56f0\u96be\u6837\u672c<\/strong>\uff1a\u5305\u62ec\u8fb9\u7f18\u60c5\u51b5\u3001\u6a21\u7cca\u610f\u56fe\u3001\u957f\u5c3e\u77e5\u8bc6\u70b9\u7b49\u3002<\/li>\n<li><strong>\u6301\u7eed\u66f4\u65b0<\/strong>\uff1a\u968f\u7740\u4e1a\u52a1\u53d8\u5316\u548c\u6570\u636e\u79ef\u7d2f\uff0c\u5b9a\u671f\u6269\u5145\u8bc4\u6d4b\u96c6\u3002<\/li>\n<\/ol>\n<h3>3.3 \u591a\u7ef4\u5ea6\u8bc4\u5206\u4f53\u7cfb<\/h3>\n<p>\u5728\u5b9e\u9645\u8bc4\u4f30\u4e2d\uff0c\u5efa\u8bae\u91c7\u7528\u591a\u7ef4\u5ea6\u8bc4\u5206\uff0c\u6bcf\u4e2a\u7ef4\u5ea6\u72ec\u7acb\u6253\u5206\uff1a<\/p>\n<table>\n<thead>\n<tr>\n<th>\u7ef4\u5ea6<\/th>\n<th>\u6743\u91cd<\/th>\n<th>\u8bc4\u4f30\u65b9\u6cd5<\/th>\n<th>\u8bc4\u5206\u6807\u51c6\uff081-5\u5206\uff09<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>\u51c6\u786e\u6027<\/td>\n<td>30%<\/td>\n<td>\u4e13\u5bb6\u6807\u6ce8\u6bd4\u5bf9<\/td>\n<td>5=\u5b8c\u5168\u6b63\u786e\uff0c1=\u4e25\u91cd\u9519\u8bef<\/td>\n<\/tr>\n<tr>\n<td>\u5b8c\u6574\u6027<\/td>\n<td>15%<\/td>\n<td>\u8986\u76d6\u8981\u70b9\u68c0\u67e5<\/td>\n<td>5=\u65e0\u9057\u6f0f\uff0c1=\u7f3a\u5931\u5173\u952e\u5185\u5bb9<\/td>\n<\/tr>\n<tr>\n<td>\u5b89\u5168\u6027<\/td>\n<td>20%<\/td>\n<td>\u5b89\u5168\u6d4b\u8bd5\u7528\u4f8b<\/td>\n<td>5=\u96f6\u8fdd\u89c4\uff0c1=\u51fa\u73b0\u6709\u5bb3\u5185\u5bb9<\/td>\n<\/tr>\n<tr>\n<td>\u53ef\u8bfb\u6027<\/td>\n<td>10%<\/td>\n<td>\u4eba\u5de5\u4e3b\u89c2\u8bc4\u5206<\/td>\n<td>5=\u6d41\u7545\u6613\u61c2\uff0c1=\u96be\u4ee5\u7406\u89e3<\/td>\n<\/tr>\n<tr>\n<td>\u901f\u5ea6<\/td>\n<td>15%<\/td>\n<td>\u5ef6\u8fdf\u6d4b\u91cf<\/td>\n<td>\u6a21\u578b\u54cd\u5e94\u65f6\u95f4<\/td>\n<\/tr>\n<tr>\n<td>\u6210\u672c<\/td>\n<td>10%<\/td>\n<td>Token \u6d88\u8017\u8ba1\u7b97<\/td>\n<td>\u8f93\u51fa\u957f\u5ea6\u4e0e\u51c6\u786e\u6027\u6bd4\u7387<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<p>\u4e0d\u540c\u4e1a\u52a1\u573a\u666f\u53ef\u4ee5\u8c03\u6574\u6743\u91cd\u3002\u4f8b\u5982\u91d1\u878d\u5408\u89c4\u573a\u666f\u5b89\u5168\u6027\u6743\u91cd\u53ef\u4ee5\u63d0\u9ad8\u5230 40%\uff0c\u800c\u521b\u610f\u5199\u4f5c\u573a\u666f\u53ef\u8bfb\u6027\u5e94\u5360\u66f4\u9ad8\u6bd4\u4f8b\u3002<\/p>\n<h3>3.4 \u81ea\u52a8\u5316\u4e0e\u4eba\u5de5\u7ed3\u5408<\/h3>\n<p>\u6700\u7406\u60f3\u7684\u8bc4\u4f30\u65b9\u5f0f\u662f&quot;\u81ea\u52a8\u5316\u521d\u7b5b + \u4eba\u5de5\u590d\u5ba1&quot;\uff1a<\/p>\n<ul>\n<li><strong>\u81ea\u52a8\u5316\u8bc4\u4f30<\/strong>\uff1a\u7528\u9884\u8bbe\u7684\u81ea\u52a8\u5316\u6d4b\u8bd5\u7528\u4f8b\u3001\u4e8b\u5b9e\u4e00\u81f4\u6027\u68c0\u67e5\u5de5\u5177\uff08\u5982 FactScore\uff09\u3001\u4ee3\u7801\u7f16\u8bd1\u6d4b\u8bd5\u7b49\u65b9\u5f0f\uff0c\u5feb\u901f\u6dd8\u6c70\u660e\u663e\u4e0d\u5408\u683c\u7684\u6a21\u578b\u7248\u672c\u3002<\/li>\n<li><strong>\u4eba\u5de5\u8bc4\u4f30<\/strong>\uff1a\u7cbe\u9009 200-500 \u4e2a\u4ee3\u8868\u6027\u6837\u672c\uff0c\u7531\u4e1a\u52a1\u4e13\u5bb6\u6216\u7ecf\u8fc7\u57f9\u8bad\u7684\u6807\u6ce8\u4eba\u5458\u8fdb\u884c\u76f2\u6d4b\u6253\u5206\u3002<\/li>\n<li><strong>A\/B \u6d4b\u8bd5<\/strong>\uff1a\u7ebf\u4e0a\u5c0f\u6d41\u91cf\u5bf9\u6bd4\u5b9e\u9a8c\uff0c\u89c2\u5bdf\u771f\u5b9e\u7528\u6237\u7684\u884c\u4e3a\u6307\u6807\uff08\u5982\u7528\u6237\u6ee1\u610f\u5ea6\u3001\u4efb\u52a1\u5b8c\u6210\u7387\u3001\u5bf9\u8bdd\u8f6e\u6b21\u7b49\uff09\u3002<\/li>\n<\/ul>\n<h2>\u56db\u3001\u6700\u65b0\u8d8b\u52bf\u4e0e\u524d\u6cbf\u65b9\u6cd5<\/h2>\n<h3>4.1 \u81ea\u6211\u8bc4\u6d4b\u4e0e Agent \u8bc4\u6d4b<\/h3>\n<p>\u968f\u7740 AI Agent \u7684\u5174\u8d77\uff0c\u4f20\u7edf\u7684&quot;\u4e00\u95ee\u4e00\u7b54&quot;\u8bc4\u6d4b\u5df2\u4e0d\u8db3\u4ee5\u8bc4\u4f30\u590d\u6742\u7684\u591a\u6b65\u9aa4\u4efb\u52a1\u80fd\u529b\u3002\u65b0\u5174\u7684\u8bc4\u6d4b\u65b9\u6cd5\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><strong>Agent \u4efb\u52a1\u5b8c\u6210\u5ea6<\/strong>\uff1a\u7ed9\u5b9a\u4e00\u4e2a\u9700\u8981\u591a\u6b65\u64cd\u4f5c\u7684\u4efb\u52a1\uff08\u5982&quot;\u5e2e\u6211\u5b9a\u4e00\u5f20\u53bb\u4e0a\u6d77\u7684\u673a\u7968&quot;\uff09\uff0c\u8bc4\u4f30\u6a21\u578b\u662f\u5426\u5b8c\u6574\u51c6\u786e\u5730\u5b8c\u6210\u3002<\/li>\n<li><strong>\u5de5\u5177\u8c03\u7528\u8bc4\u6d4b<\/strong>\uff1a\u6a21\u578b\u662f\u5426\u80fd\u6b63\u786e\u9009\u62e9\u5de5\u5177\u3001\u4f20\u9012\u53c2\u6570\u3001\u5904\u7406\u8fd4\u56de\u7ed3\u679c\u3002<\/li>\n<li><strong>\u9519\u8bef\u6062\u590d\u80fd\u529b<\/strong>\uff1a\u5f53\u5de5\u5177\u8c03\u7528\u5931\u8d25\u6216\u7528\u6237\u4fee\u6539\u9700\u6c42\u65f6\uff0c\u6a21\u578b\u662f\u5426\u80fd\u81ea\u4e3b\u7ea0\u9519\u3002<\/li>\n<\/ul>\n<h3>4.2 RAG \u6548\u679c\u8bc4\u4f30<\/h3>\n<p>\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff08RAG\uff09\u5df2\u7ecf\u6210\u4e3a\u4f01\u4e1a\u77e5\u8bc6\u5e93\u7684\u6807\u51c6\u67b6\u6784\u3002RAG \u7cfb\u7edf\u7684\u8bc4\u4f30\u5305\u62ec\uff1a<\/p>\n<ul>\n<li><strong>\u68c0\u7d22\u8d28\u91cf<\/strong>\uff1a\u53ec\u56de\u7387\u3001\u7cbe\u786e\u7387\u3001MRR\uff08\u5e73\u5747\u5012\u6570\u6392\u540d\uff09\u7b49\u3002<\/li>\n<li><strong>\u751f\u6210\u8d28\u91cf<\/strong>\uff1a\u68c0\u7d22\u7ed3\u679c\u662f\u5426\u88ab\u6b63\u786e\u5f15\u7528\u3001\u751f\u6210\u5185\u5bb9\u662f\u5426\u5fe0\u5b9e\u4e8e\u68c0\u7d22\u7ed3\u679c\uff08\u51cf\u5c11\u5e7b\u89c9\uff09\u3002<\/li>\n<li><strong>\u7aef\u5230\u7aef\u6548\u679c<\/strong>\uff1a\u7528\u6237\u5728\u77e5\u8bc6\u5e93\u573a\u666f\u4e2d\u7684\u95ee\u9898\u6ee1\u610f\u5ea6\u3002<\/li>\n<\/ul>\n<p>\u5de5\u5177\u5982 RAGAS\u3001ARES \u7b49\u63d0\u4f9b\u4e86 RAG \u7cfb\u7edf\u7684\u6807\u51c6\u5316\u8bc4\u4f30\u6846\u67b6\uff0c\u4f46\u6700\u4f73\u5b9e\u8df5\u4ecd\u7136\u9700\u8981\u7ed3\u5408\u4e1a\u52a1\u573a\u666f\u5b9a\u5236\u3002<\/p>\n<h3>4.3 \u5b89\u5168\u7ea2\u961f\u4e0e\u6301\u7eed\u76d1\u63a7<\/h3>\n<p>\u5728 AI \u6a21\u578b\u4e0a\u7ebf\u540e\uff0c\u8bc4\u4f30\u5de5\u4f5c\u5e76\u6ca1\u6709\u7ed3\u675f\u3002\u6301\u7eed\u7684\u5b89\u5168\u76d1\u63a7\u540c\u6837\u5173\u952e\uff1a<\/p>\n<ul>\n<li><strong>\u7ea2\u961f\u6d4b\u8bd5<\/strong>\uff1a\u6a21\u62df\u653b\u51fb\u884c\u4e3a\uff08\u8d8a\u72f1\u63d0\u793a\u3001Prompt Injection\u3001\u6570\u636e\u6295\u6bd2\uff09\uff0c\u6d4b\u8bd5\u6a21\u578b\u7684\u9632\u5fa1\u80fd\u529b\u3002<\/li>\n<li><strong>\u6f02\u79fb\u68c0\u6d4b<\/strong>\uff1a\u968f\u7740\u65f6\u95f4\u63a8\u79fb\uff0c\u6a21\u578b\u7684\u8f93\u51fa\u8d28\u91cf\u662f\u5426\u51fa\u73b0\u4e0b\u964d\uff08\u6570\u636e\u6f02\u79fb\u6216\u6982\u5ff5\u6f02\u79fb\uff09\u3002<\/li>\n<li><strong>\u7528\u6237\u53cd\u9988\u95ed\u73af<\/strong>\uff1a\u6536\u96c6\u7528\u6237\u7684&quot;\u8e29&quot;\u64cd\u4f5c\u548c\u8d1f\u9762\u53cd\u9988\uff0c\u4f5c\u4e3a\u6301\u7eed\u6539\u8fdb\u7684\u8f93\u5165\u3002<\/li>\n<\/ul>\n<h2>\u4e94\u3001\u7ed3\u8bed<\/h2>\n<p>AI \u6a21\u578b\u7684\u8bc4\u6d4b\u662f\u4e00\u4e2a\u6c38\u8fdc\u5728\u8def\u4e0a\u7684\u8bfe\u9898\u3002\u6bcf\u4e00\u6b21\u6a21\u578b\u80fd\u529b\u7684\u8dc3\u5347\uff0c\u90fd\u610f\u5473\u7740\u8bc4\u6d4b\u6807\u51c6\u9700\u8981\u91cd\u65b0\u5b9a\u4e49\u30022026 \u5e74\u7684\u4eca\u5929\uff0c\u6211\u4eec\u4e0d\u518d\u6ee1\u8db3\u4e8e&quot;\u67d0\u4e2a\u6392\u884c\u699c\u7b2c\u4e00\u540d&quot;\u7684\u5934\u8854\uff0c\u800c\u662f\u8ffd\u6c42&quot;\u8fd9\u4e2a\u6a21\u578b\u5728\u6211\u7684\u4e1a\u52a1\u4e2d\u5230\u5e95\u6709\u591a\u5927\u4ef7\u503c&quot;\u7684\u52a1\u5b9e\u8bc4\u4f30\u3002<\/p>\n<p>\u5bf9\u4e8e\u4f01\u4e1a\u6765\u8bf4\uff0c\u642d\u5efa\u81ea\u5df1\u7684\u8bc4\u6d4b\u4f53\u7cfb\u4e0d\u4ec5\u4ec5\u662f\u4e3a\u4e86\u9009\u6a21\u578b\uff0c\u66f4\u662f\u4e3a\u4e86\uff1a<\/p>\n<ul>\n<li><strong>\u7406\u89e3\u6a21\u578b\u7684\u8fb9\u754c<\/strong>\uff1a\u77e5\u9053\u5b83\u80fd\u505a\u4ec0\u4e48\u3001\u4e0d\u80fd\u505a\u4ec0\u4e48\u3002<\/li>\n<li><strong>\u5efa\u7acb\u8fed\u4ee3\u57fa\u51c6<\/strong>\uff1a\u6bcf\u6b21\u6a21\u578b\u5347\u7ea7\uff0c\u90fd\u80fd\u7528\u6570\u636e\u8bf4\u8bdd\uff0c\u800c\u4e0d\u662f\u51ed\u611f\u89c9\u3002<\/li>\n<li><strong>\u63a7\u5236\u98ce\u9669<\/strong>\uff1a\u5728\u5173\u952e\u4e1a\u52a1\u4e2d\u4f7f\u7528 AI\uff0c\u786e\u4fdd\u5b89\u5168\u3001\u5408\u89c4\u3001\u53ef\u63a7\u3002<\/li>\n<\/ul>\n<p>\u8bc4\u6d4b\u4e0d\u662f\u7ec8\u70b9\uff0c\u800c\u662f\u901a\u5f80\u53ef\u4fe1 AI \u7684\u5fc5\u7ecf\u4e4b\u8def\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u5f15\u8a00 2026 \u5e74\u7684\u4eca\u5929\uff0c\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u5df2\u6df1\u5ea6\u878d\u5165\u5404\u884c\u5404\u4e1a\u2014\u2014\u4ece\u4ee3\u7801\u751f\u6210\u3001\u6587\u6863\u64b0\u5199\u5230\u5ba2\u670d\u5bf9\u8bdd\u3001\u77e5\u8bc6\u68c0\u7d22\uff0c [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5],"tags":[18,26,25,27],"class_list":["post-128","post","type-post","status-publish","format-standard","hentry","category-ai","tag-ai","tag-26","tag-25","tag-27"],"_links":{"self":[{"href":"https:\/\/www.joezhou.top\/index.php?rest_route=\/wp\/v2\/posts\/128","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.joezhou.top\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.joezhou.top\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.joezhou.top\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.joezhou.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=128"}],"version-history":[{"count":0,"href":"https:\/\/www.joezhou.top\/index.php?rest_route=\/wp\/v2\/posts\/128\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.joezhou.top\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=128"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.joezhou.top\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=128"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.joezhou.top\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=128"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}