{"id":415297,"date":"2024-06-10T20:22:28","date_gmt":"2024-06-10T12:22:28","guid":{"rendered":"https:\/\/www.idc.net\/help\/?p=415297"},"modified":"2024-06-10T20:22:28","modified_gmt":"2024-06-10T12:22:28","slug":"%e6%95%b0%e9%87%8f%e5%8d%b3%e5%8a%9b%e9%87%8f%ef%bc%81%e8%85%be%e8%ae%af%e6%8f%ad%e7%a7%98%ef%bc%9aagent%e6%95%b0%e9%87%8f%e8%b6%8a%e5%a4%9a%ef%bc%8c%e5%a4%a7%e8%af%ad%e8%a8%80%e6%a8%a1%e5%9e%8b","status":"publish","type":"post","link":"https:\/\/idc.net\/help\/415297\/","title":{"rendered":"\u6570\u91cf\u5373\u529b\u91cf\uff01\u817e\u8baf\u63ed\u79d8\uff1aAgent\u6570\u91cf\u8d8a\u591a\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u6548\u679c\u8d8a\u597d"},"content":{"rendered":"<p>\u6765\u81ea\u817e\u8baf\u7684\u7814\u7a76\u8005\u4eec\u505a\u4e86\u4e00\u4e2a\u5173\u4e8e agent \u7684scaling property\uff08\u53ef\u62d3\u5c55\u6027\uff09\u7684\u5de5\u4f5c\u3002\u53d1\u73b0\uff1a\u901a\u8fc7\u7b80\u5355\u7684\u91c7\u6837\u6295\u7968\uff0c\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u6027\u80fd\uff0c\u4f1a\u968f\u7740\u5b9e\u4f8b\u5316agent\u6570\u91cf\u7684\u589e\u52a0\u800c\u589e\u5f3a\u3002\u5176\u7b2c\u4e00\u6b21\u5728\u5e7f\u6cdb\u7684\u573a\u666f\u4e0b\u9a8c\u8bc1\u4e86\u8be5\u73b0\u8c61\u7684\u666e\u904d\u6027\uff0c\u4e0e\u5176\u4ed6\u590d\u6742\u65b9\u6cd5\u7684\u6b63\u4ea4\u6027\uff0c\u4ee5\u53ca\u7814\u7a76\u4e86\u5176\u80cc\u540e\u7684\u539f\u56e0\uff0c\u5e76\u63d0\u51fa\u8fdb\u4e00\u6b65\u4fc3\u6210scaling\u53d1\u6325\u5a01\u529b\u7684\u529e\u6cd5\u3002<\/p>\n<p>\u672c\u6587\u4e2d\uff0c\u6765\u81ea\u817e\u8baf\u7684\u7814\u7a76\u8005\u53d1\u73b0\uff1a\u53ea\u9700\u901a\u8fc7\u4e00\u79cd\u7b80\u5355\u7684\u91c7\u6837\u6295\u7968\u6cd5\uff0c\u5927\u8bed\u8a00\u6a21\u578b\u7684\u6027\u80fd\u5c31\u4f1a\u968f\u7740\u5b9e\u4f8b\u5316 agent \u7684\u6570\u91cf\u7684\u589e\u5927\u800c\u589e\u5f3a\uff0c\u5448\u73b0scaling property\uff08\u53ef\u62d3\u5c55\u6027\uff09\uff0c\u65e0\u9700\u590d\u6742\u7684\u591a LLM agents \u534f\u4f5c\u6846\u67b6\u4ee5\u53caprompt\u5de5\u7a0b\u65b9\u6cd5\u7684\u52a0\u6301\u3002\u6b64\u5916\uff0c\u8be5\u65b9\u6cd5\u4e0e\u73b0\u6709\u7684\u590d\u6742\u65b9\u6cd5\u6b63\u4ea4\uff0c\u7ed3\u5408\u4e4b\u540e\uff0c\u53ef\u8fdb\u4e00\u6b65\u589e\u5f3a LLM\uff0c\u5176\u589e\u5f3a\u7a0b\u5ea6\u4e0e\u4efb\u52a1\u96be\u5ea6\u76f8\u5173\u3002\u8be5\u8bba\u6587\u505a\u4e86\u7b2c\u4e00\u4e2a\u5173\u4e8e raw agent\uff08\u6307\u4e0d\u4f9d\u8d56\u590d\u6742\u7684prompt\u5de5\u7a0b\u548c\u534f\u4f5c\u6846\u67b6\u7684LLM agent\uff09\u7684 scaling property \u7684\u7814\u7a76\uff0c\u5176\u5bf9\u5404\u79cd LLM \u57fa\u51c6\u8fdb\u884c\u4e86\u5168\u9762\u7684\u5b9e\u9a8c\uff0c\u4ee5\u9a8c\u8bc1\u6b64\u53d1\u73b0\u7684\u666e\u904d\u6027\uff0c\u5e76\u7814\u7a76\u4e86\u53ef\u4ee5\u4fc3\u8fdb\u5176\u53d1\u751f\u7684\u7b56\u7565\u3002\u76ee\u524d\u4ee3\u7801\u5df2\u5f00\u6e90\u3002<\/p>\n<p>\u591a\u4e2a\u5c0f\u6a21\u578b\u8d85\u8fc7\u5927\u6a21\u578b<\/p>\n<p>\u8bba\u6587\u8ba8\u8bba\u4e86\u8bf8\u591a\u96c6\u6210 LLM \u7684\u76f8\u5173\u5de5\u4f5c\uff0c\u5305\u62ec LLM \u81ea\u96c6\u6210\u3001\u5f02\u6784 LLM \u96c6\u6210\u3001\u8fd8\u6709\u5173\u4e8e\u591a\u4e2a LLM Agents \u534f\u4f5c\u6846\u67b6\u7684\u5de5\u4f5c\uff0c\u5e76\u4e0e\u63d0\u51fa\u7684\u65b9\u6cd5\u8fdb\u884c\u4e86\u5bf9\u6bd4\uff0c\u53ef\u4ee5\u770b\u51fa\u8bba\u6587\u8fdb\u884c\u4e86\u66f4\u5168\u9762\u7684\u7814\u7a76\u548c\u5206\u6790\uff1a<\/p>\n<p>\u4e3a\u4e86\u7814\u7a76\u5927\u578b\u8bed\u8a00\u6a21\u578b\u7684\u6027\u80fd\u5982\u4f55\u968f\u7740\u5b9e\u4f8b\u5316 agents \u6570\u91cf\u7684\u589e\u52a0\u800c\u63d0\u5347\u3002\u8bba\u6587\u4f7f\u7528\u4e86\u4e00\u79cd\u7b80\u5355\u7684\u91c7\u6837\u548c\u6295\u7968\u65b9\u6cd5\uff08\u4f5c\u8005\u7528\u4e86 simple (st) \u7684\u8bf4\u6cd5\uff0c\u53ef\u89c1\u4ed6\u4eec\u8ba4\u4e3a\u8fd9\u4e2a\u65b9\u6cd5\u4e5f\u8bb8\u662f\u6700\u7b80\u5355\u7684\u65b9\u6cd5\u4e4b\u4e00\uff09\u3002\u503c\u5f97\u6ce8\u610f\u7684\u662f\uff0c\u6b64\u65b9\u6cd5\u53ef\u4e0e\u73b0\u6709\u7684\u590d\u6742\u65b9\u6cd5\u6b63\u4ea4\u7ed3\u5408\u3002\u5b83\u53ef\u4ee5\u88ab\u5206\u4e3a\u4e24\u4e2a\u9636\u6bb5\uff1a<\/p>\n<ul>\n<li>\u5c06\u4efb\u52a1 query \u8f93\u5165\u5230\u5355\u4e2a LLM \u6216\u591a\u4e2a LLM Agents \u534f\u4f5c\u6846\u67b6\u4e2d\uff0c\u751f\u6210\u591a\u4e2a\u8f93\u51fa\uff1b<\/li>\n<li>\u901a\u8fc7\u591a\u6570\u6295\u7968\u786e\u5b9a\u6700\u7ec8\u7ed3\u679c<\/li>\n<\/ul>\n<p>\u8bba\u6587\u4ece Llama2 \u548c GPT \u7cfb\u5217\u9009\u62e9\u4e0d\u540c\u89c4\u6a21\u7684\u8bed\u8a00\u6a21\u578b\u8fdb\u884c\u8bc4\u4f30\uff0c\u4efb\u52a1\u6570\u636e\u96c6\u6db5\u76d6\u63a8\u7406\u548c\u751f\u6210\u7b49\u591a\u4e2a\u9886\u57df\u3002\u5b9e\u9a8c\u7ed3\u679c\u8868\u660e\uff0c<strong>\u5728\u6240\u6709\u4efb\u52a1\u548c\u4e0d\u540c\u79cd\u7c7b\u3001\u89c4\u6a21\u7684 LLM \u4e0a\uff0c\u53d1\u73b0 LLM \u7684\u6027\u80fd\u968f\u7740\u5b9e\u4f8b\u5316 agent \u7684\u6570\u91cf\u800c\u589e\u52a0\u3002<\/strong><\/p>\n<p>\u4f8b\u5982\uff0c\u5728 GSM8K \u4efb\u52a1\u4e0a\u63d0\u5347\u4e86 12% \u81f3 24%\uff0c\u5728 MATH \u4e0a\u63d0\u5347\u4e86 6% \u81f3 10%\u3002\u6709\u8da3\u7684\u662f\uff0c<strong>\u591a\u4e2a\u5c0f LLM \u96c6\u6210\u53ef\u4ee5\u8fbe\u5230\u751a\u81f3\u8d85\u8d8a\u8f83\u5927 LLM \u7684\u6027\u80fd\u3002<\/strong>\u4f8b\u5982\uff0c\u591a\u4e2a Llama2-13B \u7684\u96c6\u6210\u5728 GSM8K \u4e0a\u8fbe\u5230\u4e86 59% \u51c6\u786e\u7387\uff0c\u8d85\u8fc7\u4e86\u5355\u4e00 Llama2-70B \u7684 54% \u7684\u51c6\u786e\u7387\u3002<\/p>\n<p>\u8fdb\u4e00\u6b65\u5730\uff0c\u4f5c\u8005\u8fd8\u63a2\u7d22\u4e86<strong>\u4e0e\u5176\u4ed6\u65b9\u6cd5\u7684\u517c\u5bb9\u6027\u3002<\/strong>\u5c3d\u7ba1\u8fd9\u4e9b\u65b9\u6cd5\u5b9e\u73b0\u5404\u4e0d\u76f8\u540c\uff0c\u4f46\u662f\u5728\u4e0e\u4e4b\u7ed3\u5408\u4f7f\u7528\u65f6\uff0c\u6027\u80fd\u53ef\u4ee5\u8fdb\u4e00\u6b65\u63d0\u5347\uff0c<strong>\u5e76\u540c\u6837\u7b26\u5408\u5b9e\u4f8b\u5316 agent \u8d8a\u591a\uff0c\u6027\u80fd\u589e\u76ca\u8d8a\u5f3a\u7684\u73b0\u8c61\u3002<\/strong>\u5b9e\u9a8c\u7ed3\u679c\u663e\u793a\u589e\u76ca\u8303\u56f4\u4ece 1% \u5230 27% \u4e0d\u7b49\uff0c\u8bf4\u660e\u8fd9\u4e2a\u7b80\u5355\u7684\u65b9\u6cd5\u901a\u8fc7\u548c\u5176\u4ed6\u65b9\u6cd5\u6b63\u4ea4\u4f7f\u7528\u53ef\u4ee5\u8fdb\u4e00\u6b65\u589e\u5f3a LLM \u7684\u6027\u80fd\u3002<\/p>\n<p>\u57fa\u4e8e LLama13B<\/p>\n<p>\u57fa\u4e8e LLama70B<\/p>\n<p>\u57fa\u4e8e GPT-3.5-Turbo<\/p>\n<p>\u6b64\u5916\uff0c\u8bba\u6587\u8fd8\u5206\u6790\u4e86<strong>\u6027\u80fd\u63d0\u5347\u4e0e\u95ee\u9898\u96be\u5ea6\u4e4b\u95f4\u7684\u5173\u7cfb\u3002<\/strong><\/p>\n<ul>\n<li>\u56fa\u6709\u96be\u5ea6\uff1a\u968f\u7740\u4efb\u52a1\u56fa\u6709\u96be\u5ea6\u7684\u589e\u52a0\uff0c\u6027\u80fd\u63d0\u5347\uff08\u5373\u76f8\u5bf9\u6027\u80fd\u589e\u76ca\uff09\u4e5f\u4f1a\u589e\u52a0\uff0c\u4f46\u5f53\u96be\u5ea6\u8fbe\u5230\u4e00\u5b9a\u7a0b\u5ea6\u540e\uff0c\u589e\u76ca\u4f1a\u9010\u6e10\u51cf\u5c11\u3002\u8fd9\u8868\u660e\u5728\u4efb\u52a1\u8fc7\u4e8e\u590d\u6742\u65f6\uff0c\u6a21\u578b\u7684\u63a8\u7406\u80fd\u529b\u53ef\u80fd\u65e0\u6cd5\u8ddf\u4e0a\uff0c\u5bfc\u81f4\u6027\u80fd\u63d0\u5347\u7684\u8fb9\u9645\u6548\u5e94\u9012\u51cf\u3002<\/li>\n<li>\u6b65\u9aa4\u6570\u91cf\uff1a\u968f\u7740\u89e3\u51b3\u4efb\u52a1\u6240\u9700\u7684\u6b65\u9aa4\u6570\u91cf\u589e\u52a0\uff0c\u6027\u80fd\u63d0\u5347\u4e5f\u4f1a\u589e\u52a0\u3002\u8fd9\u8868\u660e\u5728\u591a\u6b65\u9aa4\u4efb\u52a1\u4e2d\uff0c\u901a\u8fc7\u589e\u52a0 agent \u6570\u91cf\u53ef\u4ee5\u5e2e\u52a9\u6a21\u578b\u66f4\u597d\u5730\u5904\u7406\u6bcf\u4e00\u6b65\uff0c\u4ece\u800c\u6574\u4f53\u63d0\u9ad8\u4efb\u52a1\u7684\u89e3\u51b3\u6027\u80fd\u3002<\/li>\n<li>\u5148\u9a8c\u6982\u7387\uff1a\u6b63\u786e\u7b54\u6848\u7684\u5148\u9a8c\u6982\u7387\u8d8a\u9ad8\uff0c\u6027\u80fd\u63d0\u5347\u8d8a\u5927\u3002\u8fd9\u610f\u5473\u7740\u5728\u6b63\u786e\u7b54\u6848\u66f4\u6709\u53ef\u80fd\u7684\u60c5\u51b5\u4e0b\uff0c\u589e\u52a0 agent \u6570\u91cf\u66f4\u6709\u53ef\u80fd\u5e26\u6765\u663e\u8457\u7684\u6027\u80fd\u63d0\u5347\u3002<\/li>\n<\/ul>\n<p>\u8282\u70b9\uff1a\u6b65\u9aa4\uff0c\u865a\u7ebf\uff1a\u53ef\u80fd\u7684\u66ff\u4ee3\u6b65\u9aa4\u3002\u8282\u70b9\u7684\u6df1\u5ea6\uff1a\u6b65\u9aa4\u7684\u6570\u91cf\uff0c\u989c\u8272\u7684\u5f3a\u5ea6\uff1a\u56fa\u6709\u96be\u5ea6\u7684\u6c34\u5e73\u3002\u56fe\u793a\u5e2e\u52a9\u8bfb\u8005\u7406\u89e3\u4efb\u52a1\u7684\u590d\u6742\u6027\u662f\u5982\u4f55\u901a\u8fc7\u8fd9\u4e9b\u7ef4\u5ea6\u6765\u8861\u91cf\u7684\u3002<\/p>\n<p>\u57fa\u4e8e\u6b64\uff0c\u8bba\u6587\u63d0\u51fa\u4e86\u4e24\u79cd\u4f18\u5316\u7b56\u7565\u6765\u8fdb\u4e00\u6b65\u63d0\u5347\u65b9\u6cd5\u7684\u6709\u6548\u6027\uff1a<\/p>\n<ul>\n<li>\u9010\u6b65\u91c7\u6837\u548c\u6295\u7968\uff08Step-wise Sampling-and-Voting\uff09\uff1a\u8fd9\u79cd\u65b9\u6cd5\u5c06\u4efb\u52a1\u5206\u89e3\u4e3a\u591a\u4e2a\u6b65\u9aa4\uff0c\u5e76\u5728\u6bcf\u4e2a\u6b65\u9aa4\u4e2d\u5e94\u7528\u91c7\u6837\u548c\u6295\u7968\uff0c\u4ee5\u51cf\u5c11\u7d2f\u79ef\u9519\u8bef\u5e76\u63d0\u9ad8\u6574\u4f53\u6027\u80fd\u3002<\/li>\n<li>\u5206\u5c42\u91c7\u6837\u548c\u6295\u7968\uff08Hierarchical Sampling-and-Voting\uff09\uff1a\u8fd9\u79cd\u65b9\u6cd5\u5c06\u4f4e\u6982\u7387\u4efb\u52a1\u5206\u89e3\u4e3a\u591a\u4e2a\u9ad8\u6982\u7387\u5b50\u4efb\u52a1\uff0c\u5e76\u5206\u5c42\u89e3\u51b3\uff0c\u540c\u65f6\u53ef\u4ee5\u4f7f\u7528\u4e0d\u540c\u6a21\u578b\u6765\u5904\u7406\u4e0d\u540c\u6982\u7387\u7684\u5b50\u4efb\u52a1\u4ee5\u964d\u4f4e\u6210\u672c\u3002<\/li>\n<\/ul>\n<p>\u6700\u540e\uff0c\u63d0\u51fa\u4e86\u672a\u6765\u7684\u5de5\u4f5c\u65b9\u5411\uff0c\u5305\u62ec\u4f18\u5316\u91c7\u6837\u9636\u6bb5\u4ee5\u964d\u4f4e\u6210\u672c\uff0c\u5e76\u7ee7\u7eed\u5f00\u53d1\u76f8\u5173\u673a\u5236\u6765\u51cf\u8f7b LLM \u5e7b\u89c9\uff08hallucinations\uff09\u7684\u5e26\u6765\u7684\u6f5c\u5728\u8d1f\u9762\u5f71\u54cd\uff0c\u786e\u4fdd\u8fd9\u4e9b\u5f3a\u5927\u6a21\u578b\u7684\u90e8\u7f72\u65e2\u8d1f\u8d23\u4efb\u53c8\u6709\u76ca\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>\u6765\u81ea\u817e\u8baf\u7684\u7814\u7a76\u8005\u4eec\u505a\u4e86\u4e00\u4e2a\u5173\u4e8e agent \u7684scaling property\uff08\u53ef\u62d3\u5c55\u6027\uff09\u7684\u5de5\u4f5c\u3002\u53d1\u73b0\uff1a\u901a\u8fc7\u7b80 [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[7023],"tags":[],"class_list":["post-415297","post","type-post","status-publish","format-standard","hentry","category-chatgpt"],"_links":{"self":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts\/415297","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/comments?post=415297"}],"version-history":[{"count":0,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/posts\/415297\/revisions"}],"wp:attachment":[{"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/media?parent=415297"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/categories?post=415297"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/idc.net\/help\/wp-json\/wp\/v2\/tags?post=415297"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}