{"id":1072593,"date":"2023-07-20T22:36:00","date_gmt":"2023-07-21T05:36:00","guid":{"rendered":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/?post_type=msr-blog-post&#038;p=1072593"},"modified":"2024-09-25T03:53:40","modified_gmt":"2024-09-25T10:53:40","slug":"evaluation-of-large-language-models","status":"publish","type":"msr-blog-post","link":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/articles\/evaluation-of-large-language-models\/","title":{"rendered":"\u5982\u4f55\u8bc4\u6d4b\u4e00\u4e2a\u5927\u8bed\u8a00\u6a21\u578b\uff1f"},"content":{"rendered":"\n<p>\u7f16\u8005\u6309\uff1a\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08Large language models, LLMs\uff09\u56e0\u5176\u5728\u5b66\u672f\u754c\u548c\u5de5\u4e1a\u754c\u5c55\u73b0\u51fa\u524d\u6240\u672a\u6709\u7684\u6027\u80fd\u800c\u5907\u53d7\u9752\u7750\u3002\u968f\u7740 LLMs \u5728\u7814\u7a76\u548c\u5b9e\u9645\u5e94\u7528\u4e2d\u88ab\u5e7f\u6cdb\u4f7f\u7528\uff0c\u5bf9\u5176\u8fdb\u884c\u6709\u6548\u8bc4\u6d4b\u53d8\u5f97\u6108\u53d1\u91cd\u8981\u3002\u8fd1\u671f\u5df2\u6709\u591a\u7bc7\u8bba\u6587\u56f4\u7ed5\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\u8fdb\u884c\u7814\u7a76\uff0c\u4f46\u5c1a\u672a\u6709\u6587\u7ae0\u5bf9\u8bc4\u6d4b\u7684\u65b9\u6cd5\u3001\u6570\u636e\u3001\u6311\u6218\u7b49\u8fdb\u884c\u5b8c\u6574\u7684\u68b3\u7406\u3002\u65e5\u524d\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u53c2\u4e0e\u5b8c\u6210\u4e86\u4ecb\u7ecd\u5927\u6a21\u578b\u8bc4\u6d4b\u9886\u57df\u7684\u7b2c\u4e00\u7bc7\u7efc\u8ff0\u6587\u7ae0\u300aA Survey on Evaluation of Large Language Models\u300b\u3002\u8be5\u8bba\u6587\u4e00\u5171\u8c03\u7814\u4e86219\u7bc7\u6587\u732e\uff0c\u4ee5\u8bc4\u6d4b\u5bf9\u8c61 (what to evaluate)\u3001\u8bc4\u6d4b\u9886\u57df (where to evaluate)\u3001\u8bc4\u6d4b\u65b9\u6cd5 (How to evaluate)\u548c\u76ee\u524d\u7684\u8bc4\u6d4b\u6311\u6218\u7b49\u51e0\u5927\u65b9\u9762\u5bf9\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\u8fdb\u884c\u4e86\u8be6\u7ec6\u7684\u68b3\u7406\u548c\u603b\u7ed3\u3002\u7814\u7a76\u5458\u4eec\u4e5f\u5c06\u6301\u7eed\u7ef4\u62a4\u5927\u6a21\u578b\u8bc4\u6d4b\u7684\u5f00\u6e90\u9879\u76ee\u4ee5\u4fc3\u8fdb\u6b64\u9886\u57df\u7684\u53d1\u5c55\u3002<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u4e3a\u4ec0\u4e48\u8981\u7814\u7a76\u5927\u6a21\u578b\u8bc4\u6d4b\">\u4e3a\u4ec0\u4e48\u8981\u7814\u7a76\u5927\u6a21\u578b\u8bc4\u6d4b\uff1f<\/h2>\n\n\n\n<p>\u901a\u4fd7\u6765\u8bb2\uff0c\u5927\u6a21\u578b\u662f\u4e00\u4e2a\u80fd\u529b\u5f88\u5f3a\u7684\u51fd\u6570 f\uff0c\u4e0e\u4e4b\u524d\u7684\u673a\u5668\u5b66\u4e60\u6a21\u578b\u5e76\u65e0\u672c\u8d28\u4e0d\u540c\u3002\u90a3\u4e48\uff0c\u4e3a\u4ec0\u4e48\u8981\u7814\u7a76\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\uff1f\u5927\u6a21\u578b\u8bc4\u6d4b\u8ddf\u4ee5\u524d\u7684\u673a\u5668\u5b66\u4e60\u6a21\u578b\u8bc4\u6d4b\u6709\u4f55\u4e0d\u540c\uff1f<\/p>\n\n\n\n<p>\u9996\u5148\uff0c\u7814\u7a76\u8bc4\u6d4b\u53ef\u4ee5\u5e2e\u52a9\u6211\u4eec\u66f4\u597d\u5730\u7406\u89e3\u5927\u6a21\u578b\u7684\u957f\u5904\u548c\u77ed\u5904\u3002\u5c3d\u7ba1\u591a\u6570\u7814\u7a76\u8868\u660e\u5927\u6a21\u578b\u5728\u8bf8\u591a\u901a\u7528\u4efb\u52a1\u4e0a\u5df2\u8fbe\u5230\u7c7b\u4eba\u6216\u8d85\u8fc7\u4eba\u7684\u6c34\u5e73\uff0c\u4f46\u4ecd\u7136\u6709\u5f88\u591a\u7814\u7a76\u5728\u8d28\u7591\u5176\u80fd\u529b\u6765\u6e90\u662f\u5426\u4e3a\u5bf9\u8bad\u7ec3\u6570\u636e\u96c6\u7684\u8bb0\u5fc6\u3002\u5982\uff0c\u4eba\u4eec\u53d1\u73b0\uff0c\u5f53\u53ea\u7ed9\u5927\u6a21\u578b\u8f93\u5165 LeetCode \u9898\u76ee\u7f16\u53f7\u800c\u4e0d\u7ed9\u4efb\u4f55\u4fe1\u606f\u7684\u65f6\u5019\uff0c\u5927\u6a21\u578b\u5c45\u7136\u4e5f\u80fd\u591f\u6b63\u786e\u8f93\u51fa\u7b54\u6848\uff0c\u8fd9\u663e\u7136\u662f\u8bad\u7ec3\u6570\u636e\u88ab\u6c61\u67d3\u4e86\u3002<\/p>\n\n\n\n<p>\u5176\u6b21\uff0c\u7814\u7a76\u8bc4\u6d4b\u53ef\u4ee5\u66f4\u597d\u5730\u4e3a\u4eba\u4e0e\u5927\u6a21\u578b\u7684\u534f\u540c\u4ea4\u4e92\u63d0\u4f9b\u6307\u5bfc\u548c\u5e2e\u52a9\u3002\u5927\u6a21\u578b\u7684\u670d\u52a1\u5bf9\u8c61\u7ec8\u7a76\u662f\u4eba\uff0c\u90a3\u4e48\u4e3a\u4e86\u66f4\u597d\u5730\u8fdb\u884c\u4eba\u673a\u4ea4\u4e92\u65b0\u8303\u5f0f\u7684\u8bbe\u8ba1\uff0c\u6211\u4eec\u4fbf\u6709\u5fc5\u8981\u5bf9\u5176\u5404\u65b9\u9762\u80fd\u529b\u8fdb\u884c\u5168\u9762\u4e86\u89e3\u548c\u8bc4\u6d4b\u3002\u5982\uff0c\u6211\u4eec\u6700\u8fd1\u7684\u7814\u7a76\u5de5\u4f5c PromptBench\uff1a\u9996\u4e2a\u5927\u8bed\u8a00\u6a21\u578b\u63d0\u793a\u9c81\u68d2\u6027\u7684\u8bc4\u6d4b\u57fa\u51c6\uff0c\u4fbf\u8be6\u7ec6\u5730\u8bc4\u6d4b\u4e86\u5927\u6a21\u578b\u5728\u201c\u6307\u4ee4\u7406\u89e3\u201d\u65b9\u9762\u7684\u9c81\u68d2\u6027\uff0c\u7ed3\u8bba\u662f\u5176\u666e\u904d\u5bb9\u6613\u53d7\u5230\u5e72\u6270\u3001\u4e0d\u591f\u7a33\u5b9a\uff0c\u8fd9\u4fbf\u542f\u53d1\u4e86\u6211\u4eec\u4ece prompt \u5c42\u9762\u6765\u52a0\u5f3a\u7cfb\u7edf\u7684\u5bb9\u9519\u80fd\u529b\u3002<\/p>\n\n\n\n<p>\u6700\u540e\uff0c\u7814\u7a76\u8bc4\u6d4b\u53ef\u4ee5\u66f4\u597d\u5730\u7edf\u7b79\u548c\u89c4\u5212\u5927\u6a21\u578b\u672a\u6765\u7684\u53d1\u5c55\u7684\u6f14\u53d8\u3001\u9632\u8303\u672a\u77e5\u548c\u53ef\u80fd\u7684\u98ce\u9669\u3002\u5927\u6a21\u578b\u4e00\u76f4\u5728\u4e0d\u65ad\u8fdb\u5316\uff0c\u5176\u80fd\u529b\u4e5f\u8d8a\u6765\u8d8a\u5f3a\u3002\u90a3\u4e48\uff0c\u901a\u8fc7\u5408\u7406\u3001\u79d1\u5b66\u7684\u8bc4\u6d4b\u673a\u5236\u7684\u8bbe\u8ba1\uff0c\u6211\u4eec\u80fd\u5426\u7528\u6f14\u5316\u7684\u89d2\u5ea6\u6765\u8bc4\u6d4b\u5176\u80fd\u529b\uff1f\u5982\u4f55\u63d0\u524d\u9884\u77e5\u5176\u53ef\u80fd\u7684\u98ce\u9669\uff1f\u8fd9\u90fd\u662f\u91cd\u8981\u7684\u7814\u7a76\u5185\u5bb9\u3002<\/p>\n\n\n\n<p>\u56e0\u6b64\uff0c\u7814\u7a76\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\u5177\u6709\u5341\u5206\u91cd\u8981\u7684\u610f\u4e49\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u7efc\u8ff0\u4e3b\u8981\u5185\u5bb9\">\u7efc\u8ff0\u4e3b\u8981\u5185\u5bb9<\/h2>\n\n\n\n<p>\u81ea ChatGPT \u53bb2022\u5e7410\u6708\u95ee\u4e16\u4ee5\u6765\uff0c\u5173\u4e8e\u5927\u6a21\u578b\u7684\u7814\u7a76\u53d8\u5f97\u7099\u624b\u53ef\u70ed\u8d77\u6765\u3002\u6211\u4eec\u5e0c\u671b\u63a2\u8ba8\u5927\u6a21\u578b\u7814\u7a76\u7684\u4e00\u4e2a\u91cd\u8981\u65b9\u5411\uff1a\u6a21\u578b\u8bc4\u6d4b\u3002\u6839\u636e\u4e0d\u5b8c\u5168\u7edf\u8ba1\uff08\u89c1\u4e0b\u56fe\uff09\uff0c\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\u65b9\u9762\u53d1\u8868\u7684\u6587\u7ae0\u5448\u4e0a\u5347\u8d8b\u52bf\uff0c\u8d8a\u6765\u8d8a\u591a\u7684\u7814\u7a76\u7740\u773c\u4e8e\u8bbe\u8ba1\u66f4\u79d1\u5b66\u3001\u66f4\u597d\u5ea6\u91cf\u3001\u66f4\u51c6\u786e\u7684\u8bc4\u6d4b\u65b9\u5f0f\u6765\u5bf9\u5927\u6a21\u578b\u7684\u80fd\u529b\u8fdb\u884c\u66f4\u6df1\u5165\u7684\u4e86\u89e3\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large is-resized\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"971\" src=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-1-1024x971.png\" alt=\"chart, line chart\" class=\"wp-image-1072824\" style=\"width:590px;height:auto\" srcset=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-1-1024x971.png 1024w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-1-300x284.png 300w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-1-768x728.png 768w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-1-190x180.png 190w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-1.png 1037w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u4e3a\u6b64\uff0c\u6211\u4eec\u4e8e\u8fd1\u671f\u5b8c\u6210\u4e86\u4ecb\u7ecd\u5927\u6a21\u578b\u8bc4\u6d4b\u9886\u57df\u7684\u7b2c\u4e00\u7bc7\u7efc\u8ff0\u6587\u7ae0\u300aA Survey on Evaluation of Large Language Models\u300b\u3002\u8be5\u8bba\u6587\u4e00\u5171\u8c03\u7814\u4e86219\u7bc7\u6587\u732e\uff0c\u4ee5\u8bc4\u6d4b\u5bf9\u8c61 (what to evaluate)\u3001\u8bc4\u6d4b\u9886\u57df (where to evaluate)\u3001\u8bc4\u6d4b\u65b9\u6cd5 (How to evaluate)\u548c\u76ee\u524d\u7684\u8bc4\u6d4b\u6311\u6218\u7b49\u51e0\u5927\u65b9\u9762\u5bf9\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\u8fdb\u884c\u4e86\u8be6\u7ec6\u7684\u68b3\u7406\u548c\u603b\u7ed3\u3002\u5176\u7814\u7a76\u76ee\u6807\u662f\u589e\u5f3a\u5bf9\u5927\u6a21\u578b\u5f53\u524d\u72b6\u6001\u7684\u7406\u89e3\uff0c\u9610\u660e\u5b83\u4eec\u7684\u4f18\u52bf\u548c\u5c40\u9650\u6027\uff0c\u5e76\u4e3a\u5176\u672a\u6765\u53d1\u5c55\u63d0\u4f9b\u89c1\u89e3\u3002\u540c\u65f6\uff0c\u6211\u4eec\u4e5f\u5c06\u8be5\u9879\u5de5\u4f5c\u8fdb\u884c\u4e86\u5f00\u6e90\uff0c\u5e0c\u671b\u6709\u66f4\u591a\u540c\u884c\u53c2\u4e0e\uff0c\u5171\u540c\u4fc3\u8fdb\u8be5\u9886\u57df\u7684\u53d1\u5c55\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"410\" src=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-2-1024x410.png\" alt=\"text\" class=\"wp-image-1072827\" srcset=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-2-1024x410.png 1024w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-2-300x120.png 300w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-2-768x307.png 768w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-2-1536x615.png 1536w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-2-240x96.png 240w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-2.png 1597w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p><em>\u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.2307.03109<\/em><\/p>\n\n\n\n<p><em>\u5f00\u6e90\u94fe\u63a5\uff1ahttps:\/\/github.com\/MLGroupJLU\/LLM-eval-survey<\/em><\/p>\n\n\n\n<p><em>\u5927\u6a21\u578b\u8bc4\u6d4b\u76f8\u5173\u7814\u7a76\uff1ahttps:\/\/llm-eval.github.io\/<\/em><\/p>\n\n\n\n<p>\u4f5c\u4e3a\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08Large language models, LLMs\uff09\u8bc4\u6d4b\u7684\u9996\u6b21\u5168\u9762\u7efc\u8ff0\uff0c\u672c\u6587\u4e3b\u8981\u4ece\u4e09\u4e2a\u65b9\u9762\u5bf9\u73b0\u6709\u5de5\u4f5c\u8fdb\u884c\u4e86\u63a2\u7d22\uff1a<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"104\" src=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-3-1024x104.png\" alt=\"diagram\" class=\"wp-image-1072830\" srcset=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-3-1024x104.png 1024w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-3-300x31.png 300w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-3-768x78.png 768w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-3-240x24.png 240w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-3.png 1226w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<p>\u2022 \u8bc4\u6d4b\u5185\u5bb9 (What to evaluate)\uff0c\u5bf9\u6d77\u91cf\u7684 LLMs \u8bc4\u6d4b\u4efb\u52a1\u8fdb\u884c\u5206\u7c7b\u5e76\u603b\u7ed3\u8bc4\u6d4b\u7ed3\u679c\uff1b<br>\u2022 \u8bc4\u6d4b\u9886\u57df (Where to evaluate)\uff0c\u5bf9 LLMs \u8bc4\u6d4b\u5e38\u7528\u7684\u6570\u636e\u96c6\u548c\u57fa\u51c6\u8fdb\u884c\u4e86\u603b\u7ed3\uff1b<br>\u2022 \u8bc4\u6d4b\u65b9\u6cd5 (How to evaluate)\uff0c\u603b\u7ed3\u4e86\u76ee\u524d\u6d41\u884c\u7684\u4e24\u79cd LLMs \u8bc4\u6d4b\u65b9\u6cd5\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"619\" src=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-4-1024x619.png\" alt=\"\u7814\u7a76\u6846\u67b6\" class=\"wp-image-1072833\" srcset=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-4-1024x619.png 1024w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-4-300x181.png 300w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-4-768x464.png 768w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-4-1536x928.png 1536w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-4-240x145.png 240w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-4.png 1951w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u7814\u7a76\u6846\u67b6<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45573\"><\/p>\n\n\n\n<p>\u6b64\u5916\uff0c\u7814\u7a76\u8fd8\u5bf9\u5927\u6a21\u578b\u8bc4\u6d4b\u4e0d\u53ef\u6216\u7f3a\u7684\u4e09\u4e2a\u7ef4\u5ea6\u5185\u5bb9\u8fdb\u884c\u4e86\u7efc\u5408\u603b\u7ed3\u3002\u6700\u540e\uff0c\u7814\u7a76\u8ba8\u8bba\u4e86\u5927\u6a21\u578b\u8bc4\u6d4b\u65f6\u53ef\u80fd\u9762\u4e34\u7684\u91cd\u5927\u6311\u6218\uff0c\u4e3a\u4eca\u540e\u7684\u7814\u7a76\u63d0\u4f9b\u4e86\u5efa\u8bae\u3002<\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u8bc4\u6d4b\u4ec0\u4e48\">\u8bc4\u6d4b\u4ec0\u4e48<\/h3>\n\n\n\n<p>\u672c\u6587\u7684\u4e3b\u8981\u76ee\u7684\u662f\u603b\u7ed3\u548c\u8ba8\u8bba\u76ee\u524d\u5728\u5927\u578b\u8bed\u8a00\u6a21\u578b\u4e0a\u7684\u8bc4\u6d4b\u5de5\u4f5c\u3002\u5728\u8bc4\u6d4b LLMs \u7684\u6027\u80fd\u65f6\uff0c\u9009\u62e9\u5408\u9002\u7684\u4efb\u52a1\u548c\u9886\u57df\u5bf9\u4e8e\u5c55\u793a\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u8868\u73b0\u3001\u4f18\u52bf\u548c\u52a3\u52bf\u81f3\u5173\u91cd\u8981\u3002\u4e3a\u4e86\u66f4\u6e05\u6670\u5730\u5c55\u793a LLMs \u7684\u80fd\u529b\u6c34\u5e73\uff0c\u6587\u7ae0\u5c06\u73b0\u6709\u7684\u4efb\u52a1\u5212\u5206\u4e3a\u4ee5\u4e0b7\u4e2a\u4e0d\u540c\u7684\u7c7b\u522b\uff1a<\/p>\n\n\n\n<p>1. \u81ea\u7136\u8bed\u8a00\u5904\u7406\uff1a\u5305\u62ec\u81ea\u7136\u8bed\u8a00\u7406\u89e3\u3001\u63a8\u7406\u3001\u81ea\u7136\u8bed\u8a00\u751f\u6210\u548c\u591a\u8bed\u8a00\u4efb\u52a1<br>2. \u9c81\u68d2\u6027\u3001\u4f26\u7406\u3001\u504f\u89c1\u548c\u771f\u5b9e\u6027<br>3. \u533b\u5b66\u5e94\u7528\uff1a\u5305\u62ec\u533b\u5b66\u95ee\u7b54\u3001\u533b\u5b66\u8003\u8bd5\u3001\u533b\u5b66\u6559\u80b2\u548c\u533b\u5b66\u52a9\u624b<br>4. \u793e\u4f1a\u79d1\u5b66<br>5. \u81ea\u7136\u79d1\u5b66\u4e0e\u5de5\u7a0b\uff1a\u5305\u62ec\u6570\u5b66\u3001\u901a\u7528\u79d1\u5b66\u548c\u5de5\u7a0b<br>6. \u4ee3\u7406\u5e94\u7528\uff1a\u5c06 LLMs \u4f5c\u4e3a\u4ee3\u7406\u4f7f\u7528<br>7. \u5176\u4ed6\u5e94\u7528<\/p>\n\n\n\n<p>\u8fd9\u6837\u7684\u5206\u7c7b\u65b9\u5f0f\u80fd\u591f\u66f4\u597d\u5730\u5c55\u793a LLMs \u5728\u5404\u9886\u57df\u7684\u8868\u73b0\u3002\u9700\u8981\u6ce8\u610f\u7684\u662f\uff0c\u51e0\u4e2a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u9886\u57df\u6709\u4ea4\u53c9\u70b9\uff0c\u56e0\u6b64\u8fd9\u79cd\u9886\u57df\u7684\u5206\u7c7b\u53ea\u662f\u4e00\u79cd\u53ef\u80fd\u7684\u5206\u7c7b\u65b9\u5f0f\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"820\" src=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-5-1024x820.png\" alt=\"\u8bc4\u6d4b\u5185\u5bb9\" class=\"wp-image-1072836\" srcset=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-5-1024x820.png 1024w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-5-300x240.png 300w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-5-768x615.png 768w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-5-225x180.png 225w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-5.png 1389w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u8bc4\u6d4b\u5185\u5bb9<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45574\"><\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u5728\u54ea\u8bc4\u6d4b\">\u5728\u54ea\u8bc4\u6d4b<\/h3>\n\n\n\n<p>\u6211\u4eec\u901a\u8fc7\u6df1\u5165\u63a2\u8ba8\u8bc4\u6d4b\u57fa\u51c6\u6765\u56de\u7b54\u5728\u54ea\u91cc\u8bc4\u6d4b\u7684\u95ee\u9898\uff0c\u5982\u4e0b\u56fe\u6240\u793a\uff0c\u8bc4\u6d4b\u57fa\u51c6\u4e3b\u8981\u5206\u4e3a\u901a\u7528\u57fa\u51c6\uff08General benchmarks\uff09\u548c\u5177\u4f53\u57fa\u51c6\uff08Specific benchmarks\uff09\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"79\" src=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-6-1024x79.png\" alt=\"\u8bc4\u6d4b\u9886\u57df\" class=\"wp-image-1072839\" srcset=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-6-1024x79.png 1024w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-6-300x23.png 300w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-6-768x59.png 768w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-6-240x18.png 240w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-6.png 1382w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u8bc4\u6d4b\u9886\u57df<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45575\"><\/p>\n\n\n\n<p>\u968f\u7740 LLMs \u57fa\u51c6\u6d4b\u8bd5\u7684\u4e0d\u65ad\u53d1\u5c55\uff0c\u76ee\u524d\u5df2\u6709\u8bb8\u591a\u53d7\u6b22\u8fce\u7684\u8bc4\u6d4b\u57fa\u51c6\u3002\u4e0b\u8868\u7efc\u8ff0\u603b\u7ed3\u4e8619\u4e2a\u6d41\u884c\u7684\u57fa\u51c6\u6d4b\u8bd5\uff0c\u6bcf\u4e2a\u57fa\u51c6\u5173\u6ce8\u4e0d\u540c\u7684\u65b9\u9762\u548c\u8bc4\u6d4b\u6807\u51c6\uff0c\u4e3a\u5404\u81ea\u7684\u9886\u57df\u505a\u51fa\u4e86\u8d21\u732e\u3002<\/p>\n\n\n\n<p><\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"346\" src=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-7-1024x346.png\" alt=\"\u8bc4\u6d4b\u57fa\u51c6\" class=\"wp-image-1072842\" srcset=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-7-1024x346.png 1024w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-7-300x101.png 300w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-7-768x260.png 768w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-7-240x81.png 240w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-7.png 1293w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><figcaption class=\"wp-element-caption\">\u8bc4\u6d4b\u57fa\u51c6<\/figcaption><\/figure>\n\n\n\n<p id=\"caption-attachment-45576\"><\/p>\n\n\n\n<h3 class=\"wp-block-heading\" id=\"\u5982\u4f55\u8bc4\u6d4b\">\u5982\u4f55\u8bc4\u6d4b<\/h3>\n\n\n\n<p>\u5728\u672c\u8282\u4e2d\uff0c\u6587\u7ae0\u4ecb\u7ecd\u4e86\u4e24\u79cd\u5e38\u7528\u7684\u8bc4\u6d4b\u65b9\u6cd5\uff1a\u81ea\u52a8\u8bc4\u6d4b\u548c\u4eba\u5de5\u8bc4\u6d4b\u3002\u8fd9\u4e24\u79cd\u65b9\u6cd5\u5728\u8bc4\u6d4b\u8bed\u8a00\u6a21\u578b\u548c\u673a\u5668\u7ffb\u8bd1\u7b49\u4efb\u52a1\u65f6\u8d77\u7740\u91cd\u8981\u7684\u4f5c\u7528\u3002\u81ea\u52a8\u8bc4\u6d4b\u65b9\u6cd5\u57fa\u4e8e\u8ba1\u7b97\u673a\u7b97\u6cd5\u548c\u81ea\u52a8\u751f\u6210\u7684\u6307\u6807\uff0c\u80fd\u591f\u5feb\u901f\u4e14\u9ad8\u6548\u5730\u8bc4\u6d4b\u6a21\u578b\u7684\u6027\u80fd\u3002\u800c\u4eba\u5de5\u8bc4\u6d4b\u5219\u4fa7\u91cd\u4e8e\u4eba\u7c7b\u4e13\u5bb6\u7684\u4e3b\u89c2\u5224\u65ad\u548c\u8d28\u91cf\u8bc4\u6d4b\uff0c\u80fd\u591f\u63d0\u4f9b\u66f4\u6df1\u5165\u3001\u7ec6\u81f4\u7684\u5206\u6790\u548c\u610f\u89c1\u3002\u4e86\u89e3\u548c\u638c\u63e1\u8fd9\u4e24\u79cd\u8bc4\u6d4b\u65b9\u6cd5\u5bf9\u51c6\u786e\u8bc4\u6d4b\u548c\u6539\u8fdb\u8bed\u8a00\u6a21\u578b\u7684\u80fd\u529b\u5341\u5206\u91cd\u8981\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u7efc\u8ff0\u603b\u7ed3\">\u7efc\u8ff0\u603b\u7ed3<\/h2>\n\n\n\n<p>\u5728\u8fd9\u4e00\u90e8\u5206\uff0c\u6587\u7ae0\u603b\u7ed3\u4e86 LLMs \u5728\u4e0d\u540c\u4efb\u52a1\u4e2d\u7684\u6210\u529f\u548c\u5931\u8d25\u6848\u4f8b\u3002<\/p>\n\n\n\n<p>LLMs \u80fd\u591f\u5728\u54ea\u4e9b\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff1f1. LLMs \u5728\u751f\u6210\u6587\u672c\u65b9\u9762\u5c55\u73b0\u51fa\u719f\u7ec3\u5ea6\uff0c\u80fd\u591f\u4ea7\u751f\u6d41\u7545\u4e14\u51c6\u786e\u7684\u8bed\u8a00\u8868\u8fbe\u30022. LLMs \u5728\u8bed\u8a00\u7406\u89e3\u65b9\u9762\u8868\u73b0\u51fa\u8272\uff0c\u80fd\u591f\u8fdb\u884c\u60c5\u611f\u5206\u6790\u548c\u6587\u672c\u5206\u7c7b\u7b49\u4efb\u52a1\u30023. LLMs \u5177\u5907\u5f3a\u5927\u7684\u8bed\u5883\u7406\u89e3\u80fd\u529b\uff0c\u80fd\u591f\u751f\u6210\u4e0e\u8f93\u5165\u4e00\u81f4\u7684\u8fde\u8d2f\u56de\u7b54\u30024. LLMs \u5728\u591a\u4e2a\u81ea\u7136\u8bed\u8a00\u5904\u7406\u4efb\u52a1\u4e2d\u8868\u73b0\u51fa\u4ee4\u4eba\u79f0\u8d5e\u7684\u6027\u80fd\uff0c\u5305\u62ec\u673a\u5668\u7ffb\u8bd1\u3001\u6587\u672c\u751f\u6210\u548c\u95ee\u7b54\u4efb\u52a1\u3002<\/p>\n\n\n\n<p>LLMs \u5728\u4ec0\u4e48\u60c5\u51b5\u4e0b\u53ef\u80fd\u4f1a\u5931\u8d25\uff1f1. LLMs \u5728\u751f\u6210\u8fc7\u7a0b\u4e2d\u53ef\u80fd\u4f1a\u8868\u73b0\u51fa\u504f\u5dee\u548c\u4e0d\u51c6\u786e\u6027\uff0c\u5bfc\u81f4\u4ea7\u751f\u6709\u504f\u5dee\u7684\u8f93\u51fa\u30022. LLMs \u5728\u7406\u89e3\u590d\u6742\u7684\u903b\u8f91\u548c\u63a8\u7406\u4efb\u52a1\u65b9\u9762\u80fd\u529b\u6709\u9650\uff0c\u5728\u590d\u6742\u7684\u73af\u5883\u4e2d\u7ecf\u5e38\u51fa\u73b0\u6df7\u4e71\u6216\u9519\u8bef\u30023. LLMs \u5728\u5904\u7406\u5927\u91cf\u6570\u636e\u96c6\u548c\u957f\u671f\u8bb0\u5fc6\u65b9\u9762\u9762\u4e34\u9650\u5236\uff0c\u8fd9\u53ef\u80fd\u4f1a\u5728\u5904\u7406\u5197\u957f\u7684\u6587\u672c\u548c\u6d89\u53ca\u957f\u671f\u4f9d\u8d56\u7684\u4efb\u52a1\u65b9\u9762\u5e26\u6765\u6311\u6218\u30024. LLMs \u5728\u6574\u5408\u5b9e\u65f6\u6216\u52a8\u6001\u4fe1\u606f\u65b9\u9762\u5b58\u5728\u5c40\u9650\u6027\uff0c\u4f7f\u5f97\u5b83\u4eec\u4e0d\u592a\u9002\u5408\u9700\u8981\u6700\u65b0\u77e5\u8bc6\u6216\u5feb\u901f\u9002\u5e94\u53d8\u5316\u73af\u5883\u7684\u4efb\u52a1\u30025. LLMs \u5bf9\u63d0\u793a\u975e\u5e38\u654f\u611f\uff0c\u5c24\u5176\u662f\u654c\u5bf9\u63d0\u793a \uff0c\u8fd9\u4e9b\u63d0\u793a\u4f1a\u89e6\u53d1\u65b0\u7684\u8bc4\u6d4b\u548c\u7b97\u6cd5\uff0c\u63d0\u9ad8\u5176\u9c81\u68d2\u6027\u30026. \u5728\u6587\u672c\u6458\u8981\u9886\u57df\uff0c\u53ef\u4ee5\u89c2\u5bdf\u5230 LLMs \u53ef\u80fd\u5728\u7279\u5b9a\u7684\u8bc4\u6d4b\u6307\u6807\u4e0a\u8868\u73b0\u51fa\u4f4e\u4e8e\u6807\u51c6\u7684\u6027\u80fd\uff0c\u8fd9\u53ef\u80fd\u5f52\u56e0\u4e8e\u90a3\u4e9b\u7279\u5b9a\u6307\u6807\u7684\u5185\u5728\u9650\u5236\u6216\u4e0d\u8db3\u30027. LLMs \u5728\u53cd\u4e8b\u5b9e\u4efb\u52a1\u4e2d \u7684\u8868\u73b0\u4e0d\u4ee4\u4eba\u6ee1\u610f\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-large\"><img loading=\"lazy\" decoding=\"async\" width=\"1024\" height=\"143\" src=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-8-1024x143.png\" alt=\"text\" class=\"wp-image-1072845\" srcset=\"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-8-1024x143.png 1024w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-8-300x42.png 300w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-8-768x107.png 768w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-8-240x34.png 240w, https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-content\/uploads\/2024\/08\/evaluation-of-large-language-models-8.png 1181w\" sizes=\"auto, (max-width: 1024px) 100vw, 1024px\" \/><\/figure>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u91cd\u5927\u6311\u6218\">\u91cd\u5927\u6311\u6218<\/h2>\n\n\n\n<p>\u8bc4\u6d4b\u4f5c\u4e3a\u4e00\u95e8\u65b0\u5b66\u79d1\uff1a\u6211\u4eec\u5bf9\u5927\u6a21\u578b\u8bc4\u6d4b\u7684\u603b\u7ed3\u542f\u53d1\u6211\u4eec\u91cd\u65b0\u8bbe\u8ba1\u4e86\u8bb8\u591a\u65b9\u9762\u3002\u5728\u672c\u8282\u4e2d\uff0c\u6211\u4eec\u4ecb\u7ecd\u4e86\u4ee5\u4e0b7\u4e2a\u91cd\u5927\u6311\u6218\u3002<\/p>\n\n\n\n<p>1. \u8bbe\u8ba1 AGI \u57fa\u51c6\u6d4b\u8bd5\u3002\u4ec0\u4e48\u662f\u53ef\u9760\u3001\u53ef\u4fe1\u4efb\u3001\u53ef\u8ba1\u7b97\u7684\u80fd\u6b63\u786e\u8861\u91cf AGI \u4efb\u52a1\u7684\u8bc4\u6d4b\u6307\u6807\uff1f<br>2. \u8bbe\u8ba1 AGI \u57fa\u51c6\u5b8c\u6210\u884c\u4e3a\u8bc4\u6d4b\u3002\u9664\u53bb\u6807\u51c6\u4efb\u52a1\u4e4b\u5916\uff0c\u5982\u4f55\u8861\u91cf AGI \u5728\u5176\u4ed6\u4efb\u52a1\uff0c\u5982\u673a\u5668\u4eba\u4ea4\u4e92\u4e2d\u7684\u8868\u73b0\uff1f<br>3. \u7a33\u5065\u6027\u8bc4\u6d4b\u3002\u76ee\u524d\u7684\u5927\u6a21\u578b\u5bf9\u8f93\u5165\u7684 prompt \u975e\u5e38\u4e0d\u9c81\u68d2\uff0c\u5982\u4f55\u6784\u5efa\u66f4\u597d\u7684\u9c81\u68d2\u6027\u8bc4\u6d4b\u51c6\u5219\uff1f<br>4. \u52a8\u6001\u6f14\u5316\u8bc4\u6d4b\u3002\u5927\u6a21\u578b\u7684\u80fd\u529b\u5728\u4e0d\u65ad\u8fdb\u5316\u3001\u4e5f\u5b58\u5728\u8bb0\u5fc6\u8bad\u7ec3\u6570\u636e\u7684\u95ee\u9898\u3002\u5982\u4f55\u8bbe\u8ba1\u66f4\u52a8\u6001\u66f4\u8fdb\u5316\u5f0f\u7684\u8bc4\u6d4b\u65b9\u6cd5\uff1f<br>5. \u53ef\u4fe1\u8d56\u7684\u8bc4\u6d4b\u3002\u5982\u4f55\u4fdd\u8bc1\u6240\u8bbe\u8ba1\u7684\u8bc4\u6d4b\u51c6\u5219\u662f\u53ef\u4fe1\u4efb\u7684\uff1f<br>6. \u652f\u6301\u6240\u6709\u5927\u6a21\u578b\u4efb\u52a1\u7684\u7edf\u4e00\u8bc4\u6d4b\u3002\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\u5e76\u4e0d\u662f\u7ec8\u70b9\u3001\u5982\u4f55\u5c06\u8bc4\u6d4b\u65b9\u6848\u4e0e\u5927\u6a21\u578b\u6709\u5173\u7684\u4e0b\u6e38\u4efb\u52a1\u8fdb\u884c\u878d\u5408\uff1f<br>7. \u8d85\u8d8a\u5355\u7eaf\u7684\u8bc4\u6d4b:\u5927\u6a21\u578b\u7684\u589e\u5f3a\u3002\u8bc4\u6d4b\u51fa\u5927\u6a21\u578b\u7684\u4f18\u7f3a\u70b9\u4e4b\u540e\uff0c\u5982\u4f55\u5f00\u53d1\u65b0\u7684\u7b97\u6cd5\u6765\u589e\u5f3a\u5176\u5728\u67d0\u65b9\u9762\u7684\u8868\u73b0\uff1f<\/p>\n\n\n\n<p>\u7814\u7a76\u7684\u91cd\u70b9\u662f\uff0c\u8bc4\u6d4b\u5e94\u8be5\u88ab\u89c6\u4e3a\u63a8\u52a8 LLMs \u548c\u5176\u4ed6\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u6210\u529f\u7684\u57fa\u672c\u5b66\u79d1\u3002\u73b0\u6709\u7684\u7814\u7a76\u65b9\u6848\u4e0d\u8db3\u4ee5\u5bf9 LLMs \u8fdb\u884c\u5168\u9762\u7684\u8bc4\u6d4b\uff0c\u8fd9\u53ef\u80fd\u4e3a\u672a\u6765\u7684 LLMs \u8bc4\u6d4b\u7814\u7a76\u5e26\u6765\u65b0\u7684\u673a\u9047\u3002<\/p>\n\n\n\n<h2 class=\"wp-block-heading\" id=\"\u7ed3\u8bba\">\u7ed3\u8bba<\/h2>\n\n\n\n<p>\u8bc4\u6d4b\u5177\u6709\u6df1\u8fdc\u7684\u610f\u4e49\uff0c\u5728\u4eba\u5de5\u667a\u80fd\u6a21\u578b\u7684\u53d1\u5c55\u4e2d\u53d8\u5f97\u52bf\u5728\u5fc5\u884c\uff0c\u5728 LLMs \u4e0d\u65ad\u53d1\u5c55\u7684\u80cc\u666f\u4e0b\u5c24\u5176\u5982\u6b64\u3002\u672c\u6587\u9996\u6b21\u4ece\u8bc4\u6d4b\u4ec0\u4e48\u3001\u5982\u4f55\u8bc4\u6d4b\u3001\u5728\u54ea\u91cc\u8bc4\u6d4b\u4e09\u4e2a\u65b9\u9762\u5bf9 LLMs \u7684\u8bc4\u6d4b\u8fdb\u884c\u4e86\u5168\u9762\u7684\u6982\u8ff0\u3002\u901a\u8fc7\u5c01\u88c5\u8bc4\u6d4b\u4efb\u52a1\u3001\u534f\u8bae\u548c\u57fa\u51c6\uff0c\u7814\u7a76\u7684\u76ee\u6807\u662f\u589e\u5f3a\u5bf9 LLMs \u5f53\u524d\u72b6\u6001\u7684\u7406\u89e3\uff0c\u9610\u660e\u5b83\u4eec\u7684\u4f18\u52bf\u548c\u5c40\u9650\u6027\uff0c\u5e76\u4e3a\u672a\u6765 LLMs \u7684\u53d1\u5c55\u63d0\u4f9b\u89c1\u89e3\u3002<\/p>\n\n\n\n<p>\u7814\u7a76\u7684\u8c03\u67e5\u663e\u793a\uff0c\u76ee\u524d\u7684 LLMs \u5728\u8bb8\u591a\u4efb\u52a1\u4e2d\u90fd\u5b58\u5728\u4e00\u5b9a\u7684\u5c40\u9650\u6027\uff0c\u5c24\u5176\u662f\u63a8\u7406\u548c\u9c81\u68d2\u6027\u4efb\u52a1\u3002\u4e0e\u6b64\u540c\u65f6\uff0c\u5bf9\u5f53\u4ee3\u8bc4\u6d4b\u7cfb\u7edf\u8fdb\u884c\u8c03\u6574\u548c\u53d1\u5c55\u7684\u9700\u6c42\u4f9d\u7136\u660e\u663e\uff0c\u4ee5\u786e\u4fdd\u5bf9 LLMs \u7684\u5185\u5728\u80fd\u529b\u548c\u5c40\u9650\u6027\u8fdb\u884c\u51c6\u786e\u8bc4\u6d4b\u3002\u6700\u540e\uff0c\u672c\u6587\u786e\u5b9a\u4e86\u672a\u6765\u7814\u7a76\u5e94\u8be5\u89e3\u51b3\u7684\u51e0\u4e2a\u91cd\u5927\u6311\u6218\uff0c\u5e76\u5e0c\u671b LLMs \u80fd\u591f\u9010\u6b65\u63d0\u9ad8\u5927\u8bed\u8a00\u6a21\u578b\u4e3a\u4eba\u7c7b\u670d\u52a1\u7684\u6c34\u51c6\u3002<\/p>\n\n\n\n<p>\u6211\u4eec\u8fd8\u5728\u4ee5\u4e0b\u7f51\u7ad9\u4e2d\u6c47\u603b\u4e86\u672c\u56e2\u961f\u6240\u6709\u7684\u5927\u6a21\u578b\u8bc4\u6d4b\u76f8\u5173\u7814\u7a76\uff0c\u6b22\u8fce\u5173\u6ce8\uff1a<br>https:\/\/llm-eval.github.io\/<br>https:\/\/github.com\/microsoft\/promptbench<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u7f16\u8005\u6309\uff1a\u5927\u578b\u8bed\u8a00\u6a21\u578b\uff08Large language models, LLMs\uff09\u56e0\u5176\u5728\u5b66\u672f\u754c\u548c\u5de5\u4e1a\u754c\u5c55\u73b0\u51fa\u524d\u6240\u672a\u6709\u7684\u6027\u80fd\u800c\u5907\u53d7\u9752\u7750\u3002\u968f\u7740 LLMs \u5728\u7814\u7a76\u548c\u5b9e\u9645\u5e94\u7528\u4e2d\u88ab\u5e7f\u6cdb\u4f7f\u7528\uff0c\u5bf9\u5176\u8fdb\u884c\u6709\u6548\u8bc4\u6d4b\u53d8\u5f97\u6108\u53d1\u91cd\u8981\u3002\u8fd1\u671f\u5df2\u6709\u591a\u7bc7\u8bba\u6587\u56f4\u7ed5\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\u8fdb\u884c\u7814\u7a76\uff0c\u4f46\u5c1a\u672a\u6709\u6587\u7ae0\u5bf9\u8bc4\u6d4b\u7684\u65b9\u6cd5\u3001\u6570\u636e\u3001\u6311\u6218\u7b49\u8fdb\u884c\u5b8c\u6574\u7684\u68b3\u7406\u3002\u65e5\u524d\uff0c\u5fae\u8f6f\u4e9a\u6d32\u7814\u7a76\u9662\u7684\u7814\u7a76\u5458\u4eec\u53c2\u4e0e\u5b8c\u6210\u4e86\u4ecb\u7ecd\u5927\u6a21\u578b\u8bc4\u6d4b\u9886\u57df\u7684\u7b2c\u4e00\u7bc7\u7efc\u8ff0\u6587\u7ae0\u300aA Survey on Evaluation of Large Language Models\u300b\u3002\u8be5\u8bba\u6587\u4e00\u5171\u8c03\u7814\u4e86219\u7bc7\u6587\u732e\uff0c\u4ee5\u8bc4\u6d4b\u5bf9\u8c61 (what to evaluate)\u3001\u8bc4\u6d4b\u9886\u57df (where to evaluate)\u3001\u8bc4\u6d4b\u65b9\u6cd5 (How to evaluate)\u548c\u76ee\u524d\u7684\u8bc4\u6d4b\u6311\u6218\u7b49\u51e0\u5927\u65b9\u9762\u5bf9\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\u8fdb\u884c\u4e86\u8be6\u7ec6\u7684\u68b3\u7406\u548c\u603b\u7ed3\u3002\u7814\u7a76\u5458\u4eec\u4e5f\u5c06\u6301\u7eed\u7ef4\u62a4\u5927\u6a21\u578b\u8bc4\u6d4b\u7684\u5f00\u6e90\u9879\u76ee\u4ee5\u4fc3\u8fdb\u6b64\u9886\u57df\u7684\u53d1\u5c55\u3002 \u901a\u4fd7\u6765\u8bb2\uff0c\u5927\u6a21\u578b\u662f\u4e00\u4e2a\u80fd\u529b\u5f88\u5f3a\u7684\u51fd\u6570 f\uff0c\u4e0e\u4e4b\u524d\u7684\u673a\u5668\u5b66\u4e60\u6a21\u578b\u5e76\u65e0\u672c\u8d28\u4e0d\u540c\u3002\u90a3\u4e48\uff0c\u4e3a\u4ec0\u4e48\u8981\u7814\u7a76\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\uff1f\u5927\u6a21\u578b\u8bc4\u6d4b\u8ddf\u4ee5\u524d\u7684\u673a\u5668\u5b66\u4e60\u6a21\u578b\u8bc4\u6d4b\u6709\u4f55\u4e0d\u540c\uff1f \u9996\u5148\uff0c\u7814\u7a76\u8bc4\u6d4b\u53ef\u4ee5\u5e2e\u52a9\u6211\u4eec\u66f4\u597d\u5730\u7406\u89e3\u5927\u6a21\u578b\u7684\u957f\u5904\u548c\u77ed\u5904\u3002\u5c3d\u7ba1\u591a\u6570\u7814\u7a76\u8868\u660e\u5927\u6a21\u578b\u5728\u8bf8\u591a\u901a\u7528\u4efb\u52a1\u4e0a\u5df2\u8fbe\u5230\u7c7b\u4eba\u6216\u8d85\u8fc7\u4eba\u7684\u6c34\u5e73\uff0c\u4f46\u4ecd\u7136\u6709\u5f88\u591a\u7814\u7a76\u5728\u8d28\u7591\u5176\u80fd\u529b\u6765\u6e90\u662f\u5426\u4e3a\u5bf9\u8bad\u7ec3\u6570\u636e\u96c6\u7684\u8bb0\u5fc6\u3002\u5982\uff0c\u4eba\u4eec\u53d1\u73b0\uff0c\u5f53\u53ea\u7ed9\u5927\u6a21\u578b\u8f93\u5165 LeetCode \u9898\u76ee\u7f16\u53f7\u800c\u4e0d\u7ed9\u4efb\u4f55\u4fe1\u606f\u7684\u65f6\u5019\uff0c\u5927\u6a21\u578b\u5c45\u7136\u4e5f\u80fd\u591f\u6b63\u786e\u8f93\u51fa\u7b54\u6848\uff0c\u8fd9\u663e\u7136\u662f\u8bad\u7ec3\u6570\u636e\u88ab\u6c61\u67d3\u4e86\u3002 \u5176\u6b21\uff0c\u7814\u7a76\u8bc4\u6d4b\u53ef\u4ee5\u66f4\u597d\u5730\u4e3a\u4eba\u4e0e\u5927\u6a21\u578b\u7684\u534f\u540c\u4ea4\u4e92\u63d0\u4f9b\u6307\u5bfc\u548c\u5e2e\u52a9\u3002\u5927\u6a21\u578b\u7684\u670d\u52a1\u5bf9\u8c61\u7ec8\u7a76\u662f\u4eba\uff0c\u90a3\u4e48\u4e3a\u4e86\u66f4\u597d\u5730\u8fdb\u884c\u4eba\u673a\u4ea4\u4e92\u65b0\u8303\u5f0f\u7684\u8bbe\u8ba1\uff0c\u6211\u4eec\u4fbf\u6709\u5fc5\u8981\u5bf9\u5176\u5404\u65b9\u9762\u80fd\u529b\u8fdb\u884c\u5168\u9762\u4e86\u89e3\u548c\u8bc4\u6d4b\u3002\u5982\uff0c\u6211\u4eec\u6700\u8fd1\u7684\u7814\u7a76\u5de5\u4f5c PromptBench\uff1a\u9996\u4e2a\u5927\u8bed\u8a00\u6a21\u578b\u63d0\u793a\u9c81\u68d2\u6027\u7684\u8bc4\u6d4b\u57fa\u51c6\uff0c\u4fbf\u8be6\u7ec6\u5730\u8bc4\u6d4b\u4e86\u5927\u6a21\u578b\u5728\u201c\u6307\u4ee4\u7406\u89e3\u201d\u65b9\u9762\u7684\u9c81\u68d2\u6027\uff0c\u7ed3\u8bba\u662f\u5176\u666e\u904d\u5bb9\u6613\u53d7\u5230\u5e72\u6270\u3001\u4e0d\u591f\u7a33\u5b9a\uff0c\u8fd9\u4fbf\u542f\u53d1\u4e86\u6211\u4eec\u4ece prompt \u5c42\u9762\u6765\u52a0\u5f3a\u7cfb\u7edf\u7684\u5bb9\u9519\u80fd\u529b\u3002 \u6700\u540e\uff0c\u7814\u7a76\u8bc4\u6d4b\u53ef\u4ee5\u66f4\u597d\u5730\u7edf\u7b79\u548c\u89c4\u5212\u5927\u6a21\u578b\u672a\u6765\u7684\u53d1\u5c55\u7684\u6f14\u53d8\u3001\u9632\u8303\u672a\u77e5\u548c\u53ef\u80fd\u7684\u98ce\u9669\u3002\u5927\u6a21\u578b\u4e00\u76f4\u5728\u4e0d\u65ad\u8fdb\u5316\uff0c\u5176\u80fd\u529b\u4e5f\u8d8a\u6765\u8d8a\u5f3a\u3002\u90a3\u4e48\uff0c\u901a\u8fc7\u5408\u7406\u3001\u79d1\u5b66\u7684\u8bc4\u6d4b\u673a\u5236\u7684\u8bbe\u8ba1\uff0c\u6211\u4eec\u80fd\u5426\u7528\u6f14\u5316\u7684\u89d2\u5ea6\u6765\u8bc4\u6d4b\u5176\u80fd\u529b\uff1f\u5982\u4f55\u63d0\u524d\u9884\u77e5\u5176\u53ef\u80fd\u7684\u98ce\u9669\uff1f\u8fd9\u90fd\u662f\u91cd\u8981\u7684\u7814\u7a76\u5185\u5bb9\u3002 \u56e0\u6b64\uff0c\u7814\u7a76\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\u5177\u6709\u5341\u5206\u91cd\u8981\u7684\u610f\u4e49\u3002 \u81ea ChatGPT \u53bb2022\u5e7410\u6708\u95ee\u4e16\u4ee5\u6765\uff0c\u5173\u4e8e\u5927\u6a21\u578b\u7684\u7814\u7a76\u53d8\u5f97\u7099\u624b\u53ef\u70ed\u8d77\u6765\u3002\u6211\u4eec\u5e0c\u671b\u63a2\u8ba8\u5927\u6a21\u578b\u7814\u7a76\u7684\u4e00\u4e2a\u91cd\u8981\u65b9\u5411\uff1a\u6a21\u578b\u8bc4\u6d4b\u3002\u6839\u636e\u4e0d\u5b8c\u5168\u7edf\u8ba1\uff08\u89c1\u4e0b\u56fe\uff09\uff0c\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\u65b9\u9762\u53d1\u8868\u7684\u6587\u7ae0\u5448\u4e0a\u5347\u8d8b\u52bf\uff0c\u8d8a\u6765\u8d8a\u591a\u7684\u7814\u7a76\u7740\u773c\u4e8e\u8bbe\u8ba1\u66f4\u79d1\u5b66\u3001\u66f4\u597d\u5ea6\u91cf\u3001\u66f4\u51c6\u786e\u7684\u8bc4\u6d4b\u65b9\u5f0f\u6765\u5bf9\u5927\u6a21\u578b\u7684\u80fd\u529b\u8fdb\u884c\u66f4\u6df1\u5165\u7684\u4e86\u89e3\u3002 \u4e3a\u6b64\uff0c\u6211\u4eec\u4e8e\u8fd1\u671f\u5b8c\u6210\u4e86\u4ecb\u7ecd\u5927\u6a21\u578b\u8bc4\u6d4b\u9886\u57df\u7684\u7b2c\u4e00\u7bc7\u7efc\u8ff0\u6587\u7ae0\u300aA Survey on Evaluation of Large Language Models\u300b\u3002\u8be5\u8bba\u6587\u4e00\u5171\u8c03\u7814\u4e86219\u7bc7\u6587\u732e\uff0c\u4ee5\u8bc4\u6d4b\u5bf9\u8c61 (what to evaluate)\u3001\u8bc4\u6d4b\u9886\u57df (where to evaluate)\u3001\u8bc4\u6d4b\u65b9\u6cd5 (How to evaluate)\u548c\u76ee\u524d\u7684\u8bc4\u6d4b\u6311\u6218\u7b49\u51e0\u5927\u65b9\u9762\u5bf9\u5927\u6a21\u578b\u7684\u8bc4\u6d4b\u8fdb\u884c\u4e86\u8be6\u7ec6\u7684\u68b3\u7406\u548c\u603b\u7ed3\u3002\u5176\u7814\u7a76\u76ee\u6807\u662f\u589e\u5f3a\u5bf9\u5927\u6a21\u578b\u5f53\u524d\u72b6\u6001\u7684\u7406\u89e3\uff0c\u9610\u660e\u5b83\u4eec\u7684\u4f18\u52bf\u548c\u5c40\u9650\u6027\uff0c\u5e76\u4e3a\u5176\u672a\u6765\u53d1\u5c55\u63d0\u4f9b\u89c1\u89e3\u3002\u540c\u65f6\uff0c\u6211\u4eec\u4e5f\u5c06\u8be5\u9879\u5de5\u4f5c\u8fdb\u884c\u4e86\u5f00\u6e90\uff0c\u5e0c\u671b\u6709\u66f4\u591a\u540c\u884c\u53c2\u4e0e\uff0c\u5171\u540c\u4fc3\u8fdb\u8be5\u9886\u57df\u7684\u53d1\u5c55\u3002 \u8bba\u6587\u94fe\u63a5\uff1ahttps:\/\/arxiv.2307.03109 \u5f00\u6e90\u94fe\u63a5\uff1ahttps:\/\/github.com\/MLGroupJLU\/LLM-eval-survey [&hellip;]<\/p>\n","protected":false},"author":34512,"featured_media":1072821,"template":"","meta":{"msr-url-field":"","msr-podcast-episode":"","msrModifiedDate":"","msrModifiedDateEnabled":false,"ep_exclude_from_search":false,"_classifai_error":"","msr-content-parent":1012650,"msr_hide_image_in_river":0,"footnotes":""},"research-area":[],"msr-locale":[268881],"msr-post-option":[],"class_list":["post-1072593","msr-blog-post","type-msr-blog-post","status-publish","has-post-thumbnail","hentry","msr-locale-zh_cn"],"msr_assoc_parent":{"id":1012650,"type":"lab"},"_links":{"self":[{"href":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072593","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post"}],"about":[{"href":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/types\/msr-blog-post"}],"author":[{"embeddable":true,"href":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/users\/34512"}],"version-history":[{"count":3,"href":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072593\/revisions"}],"predecessor-version":[{"id":1072848,"href":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-blog-post\/1072593\/revisions\/1072848"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/media\/1072821"}],"wp:attachment":[{"href":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/media?parent=1072593"}],"wp:term":[{"taxonomy":"msr-research-area","embeddable":true,"href":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/research-area?post=1072593"},{"taxonomy":"msr-locale","embeddable":true,"href":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-locale?post=1072593"},{"taxonomy":"msr-post-option","embeddable":true,"href":"https:\/\/new-cm-edgedigital.pages.dev\/en-us\/research\/wp-json\/wp\/v2\/msr-post-option?post=1072593"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}