{"id":864,"date":"2026-09-13T19:03:35","date_gmt":"2026-09-13T11:03:35","guid":{"rendered":"https:\/\/www.timedigital.cn\/new\/864\/"},"modified":"2026-09-13T19:03:35","modified_gmt":"2026-09-13T11:03:35","slug":"%e7%94%a8%e9%93%be%e5%bc%8f%e6%b3%95%e5%88%99%e9%80%8f%e8%a7%86%e5%a4%a7%e6%a8%a1%e5%9e%8b%e8%be%93%e5%87%ba%e8%bd%a8%e8%bf%b9%e7%9a%84%e6%a6%82%e7%8e%87%e6%9c%ac%e8%b4%a8","status":"publish","type":"post","link":"https:\/\/www.timedigital.cn\/new\/864\/","title":{"rendered":"\u7528\u94fe\u5f0f\u6cd5\u5219\u900f\u89c6\u5927\u6a21\u578b\u8f93\u51fa\u8f68\u8ff9\u7684\u6982\u7387\u672c\u8d28"},"content":{"rendered":"

\u7406\u89e3\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u751f\u6210\u8fc7\u7a0b\uff0c\u901a\u5e38\u6211\u4eec\u4f1a\u5173\u6ce8\u4e0b\u4e00\u4e2a\u8bcd\u7684\u9884\u6d4b\u6982\u7387\u3002\u4f46\u9c9c\u6709\u4eba\u601d\u8003\uff1a\u6574\u4e2a\u751f\u6210\u5e8f\u5217\u2014\u2014\u5373\u4e00\u6761\u5b8c\u6574\u7684\u201c\u8f68\u8ff9\u201d\u2014\u2014\u5176\u8054\u5408\u6982\u7387\u662f\u5982\u4f55\u88ab\u8ba1\u7b97\u548c\u5206\u89e3\u7684\uff1f\u4e00\u7bc7\u6765\u81ea\u5f00\u53d1\u8005 Sithan Kanna \u7684\u6280\u672f\u6587\u7ae0\uff0c\u4ece\u6982\u7387\u8bba\u4e2d\u6700\u57fa\u7840\u7684\u94fe\u5f0f\u6cd5\u5219\u51fa\u53d1\uff0c\u91cd\u65b0\u63a8\u5bfc\u4e86 LLM \u8f93\u51fa\u8f68\u8ff9\u7684\u6982\u7387\u8868\u8fbe\uff0c\u4e3a\u8bc4\u4f30\u6a21\u578b\u884c\u4e3a\u7684\u53ef\u89e3\u91ca\u6027\u63d0\u4f9b\u4e86\u65b0\u7684\u89c6\u89d2\u3002<\/p>\n

\u4ece\u94fe\u5f0f\u6cd5\u5219\u5230\u5e8f\u5217\u6982\u7387<\/p><\/h2>

\u94fe\u5f0f\u6cd5\u5219\u662f\u6982\u7387\u8bba\u4e2d\u7684\u57fa\u672c\u5de5\u5177\uff0c\u5b83\u5141\u8bb8\u5c06\u591a\u4e2a\u968f\u673a\u53d8\u91cf\u7684\u8054\u5408\u6982\u7387\u5206\u89e3\u4e3a\u4e00\u7cfb\u5217\u6761\u4ef6\u6982\u7387\u7684\u4e58\u79ef\u3002\u5bf9\u4e8e LLM \u751f\u6210\u7684 token \u5e8f\u5217 (x_1, x_2, dots, x_T)\uff0c\u5176\u8054\u5408\u6982\u7387\u53ef\u4ee5\u5199\u4f5c\uff1a<\/p>\n

[
\nP(x_1, x_2, dots, x_T) = prod_{t=1}^{T} P(x_t mid x_1, dots, x_{t-1})
\n]<\/p>\n

\u8fd9\u6b63\u662f LLM \u81ea\u56de\u5f52\u751f\u6210\u7684\u6838\u5fc3\u5f62\u5f0f\uff1a\u6bcf\u4e00\u6b65\u53ea\u4f9d\u8d56\u524d\u9762\u5df2\u751f\u6210\u7684\u6240\u6709 token\u3002\u6587\u7ae0\u6307\u51fa\uff0c\u8fd9\u4e00\u516c\u5f0f\u770b\u4f3c\u7b80\u5355\uff0c\u5374\u6784\u6210\u4e86\u7406\u89e3\u6a21\u578b\u201c\u4fe1\u5fc3\u201d\u3001\u201c\u8fde\u8d2f\u6027\u201d\u548c\u201c\u9000\u5316\u98ce\u9669\u201d\u7684\u5173\u952e\u3002<\/p>\n

\u8f68\u8ff9\u6982\u7387\u7684\u4e24\u4e2a\u5173\u952e\u6027\u8d28<\/p><\/h2>

– \u4e58\u6cd5\u7d2f\u79ef\u6027<\/strong>\uff1a\u5e8f\u5217\u8d8a\u957f\uff0c\u8054\u5408\u6982\u7387\u901a\u5e38\u4f1a\u6307\u6570\u7ea7\u4e0b\u964d\u3002\u8fd9\u610f\u5473\u7740\u4e00\u6761 50 \u4e2a token \u7684\u8f68\u8ff9\u5373\u4f7f\u6bcf\u4e00\u6b65\u90fd\u9ad8\u8fbe 0.9\uff0c\u6574\u4f53\u6982\u7387\u4e5f\u4f1a\u5c0f\u5230\u4e0d\u53ef\u60f3\u8c61\u3002\u6587\u7ae0\u501f\u6b64\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48\u957f\u6587\u672c\u751f\u6210\u4e2d\uff0c\u6a21\u578b\u5bb9\u6613\u504f\u79bb\u7528\u6237\u610f\u56fe\u2014\u2014\u4efb\u4f55\u4e00\u6b65\u7684\u5fae\u5c0f\u5931\u8bef\u90fd\u4f1a\u88ab\u8fde\u4e58\u653e\u5927\u3002<\/p>\n

– \u6761\u4ef6\u4f9d\u8d56\u7684\u5c40\u90e8\u6027<\/strong>\uff1a\u5728\u8ba1\u7b97\u7b2c (t) \u6b65\u7684\u6982\u7387\u65f6\uff0c\u6a21\u578b\u53ea\u663e\u5f0f\u4f9d\u8d56\u524d (t-1) \u4e2a token\uff0c\u4f46\u5b9e\u9645\u5185\u90e8\u72b6\u6001\u901a\u8fc7 attention \u673a\u5236\u9690\u5f0f\u643a\u5e26\u4e86\u5168\u90e8\u5386\u53f2\u4fe1\u606f\u3002\u8f68\u8ff9\u6982\u7387\u56e0\u6b64\u4e0d\u53ea\u662f\u7b80\u5355\u4e58\u79ef\uff0c\u800c\u662f\u5bf9\u6a21\u578b\u5185\u90e8\u72b6\u6001\u6f14\u5316\u7684\u4e00\u79cd\u201c\u8fd8\u539f\u201d\u3002<\/p>\n

\u4e3a\u4ec0\u4e48\u73b0\u5728\u8ba8\u8bba\u8f68\u8ff9\u6982\u7387<\/p><\/h2>

\u5f53\u524d\u4e1a\u754c\u5bf9\u5927\u6a21\u578b\u7684\u8bc4\u4f30\u591a\u96c6\u4e2d\u5728\u5355\u6b65\u51c6\u786e\u7387\uff08next-token prediction accuracy\uff09\u6216\u6574\u4f53\u4efb\u52a1\u7684\u6700\u7ec8\u7ed3\u679c\u4e0a\u3002\u4f46\u8fd9\u4e24\u8005\u4e4b\u95f4\u5b58\u5728\u9e3f\u6c9f\uff1a\u4e00\u4e2a\u5728\u5c40\u90e8\u8868\u73b0\u4f18\u5f02\u7684\u6a21\u578b\uff0c\u53ef\u80fd\u5728\u957f\u5e8f\u5217\u4e0a\u4ea7\u751f\u4f4e\u6982\u7387\u7684\u201c\u6f02\u79fb\u201d\u3002\u6587\u7ae0\u63d0\u51fa\uff0c\u901a\u8fc7\u8ba1\u7b97\u6574\u6761\u8f68\u8ff9\u7684\u5bf9\u6570\u6982\u7387\uff0c\u53ef\u4ee5\u66f4\u65e9\u5730\u53d1\u73b0\u751f\u6210\u4e2d\u7684\u4e0d\u7a33\u5b9a\u70b9\uff0c\u751a\u81f3\u7528\u4e8e\u68c0\u6d4b\u6a21\u578b\u662f\u5426\u5728\u201c\u80e1\u7f16\u4e71\u9020\u201d\uff08\u5373\u4ea7\u751f\u9ad8\u7f6e\u4fe1\u5ea6\u4f46\u8bed\u4e49\u4e0a\u4e0d\u53ef\u80fd\u7684 token \u7ec4\u5408\uff09\u3002<\/p>\n

\u53e6\u4e00\u4e2a\u503c\u5f97\u5173\u6ce8\u7684\u5e94\u7528\u662f\u5bf9\u6bd4\u4e0d\u540c\u91c7\u6837\u7b56\u7565\uff08\u5982 temperature\u3001top-p\uff09\u5bf9\u8f68\u8ff9\u6982\u7387\u5206\u5e03\u7684\u5f71\u54cd\u3002\u4f20\u7edf\u4e0a\u6211\u4eec\u901a\u8fc7\u4eba\u5de5\u89c2\u5bdf\u6765\u51b3\u5b9a\u91c7\u6837\u53c2\u6570\uff0c\u4f46\u8f68\u8ff9\u6982\u7387\u63d0\u4f9b\u4e86\u4e00\u79cd\u65e0\u76d1\u7763\u7684\u91cf\u5316\u6307\u6807\uff1a\u4f4e\u6982\u7387\u8f68\u8ff9\u901a\u5e38\u610f\u5473\u7740\u751f\u6210\u8def\u5f84\u504f\u79bb\u4e86\u6a21\u578b\u5b66\u5230\u7684\u5178\u578b\u5206\u5e03\u3002<\/p>\n

\u5bf9\u6a21\u578b\u89e3\u91ca\u4e0e\u8c03\u8bd5\u7684\u610f\u4e49<\/p><\/h2>

\u6587\u7ae0\u5e76\u4e0d\u53ea\u662f\u91cd\u8ff0\u6559\u79d1\u4e66\u77e5\u8bc6\uff0c\u800c\u662f\u628a\u94fe\u5f0f\u6cd5\u5219\u5f53\u4f5c\u4e00\u4e2a\u8bca\u65ad\u5de5\u5177\u3002\u5f53\u5f00\u53d1\u8005\u9700\u8981\u5b9a\u4f4d\u6a21\u578b\u8f93\u51fa\u4e2d\u201c\u54ea\u4e00\u6b65\u5bfc\u81f4\u6574\u4f53\u8d28\u91cf\u6076\u5316\u201d\u65f6\uff0c\u53ef\u4ee5\u9010 token \u67e5\u770b\u6761\u4ef6\u6982\u7387\u7684\u9aa4\u964d\u70b9\u3002\u8fd9\u79cd\u7ec6\u7c92\u5ea6\u7684\u5206\u6790\uff0c\u6bd4\u53ea\u770b\u6700\u7ec8 BLEU \u6216 ROUGE \u5206\u6570\u66f4\u80fd\u63ed\u793a\u95ee\u9898\u672c\u6e90\u3002<\/p>\n

\u66f4\u91cd\u8981\u7684\u662f\uff0c\u8f68\u8ff9\u6982\u7387\u5bf9\u8f93\u5165\u6270\u52a8\u9ad8\u5ea6\u654f\u611f\u3002\u901a\u8fc7\u6bd4\u8f83\u539f\u59cb\u8f93\u5165\u548c\u5bf9\u6297\u6270\u52a8\u8f93\u5165\u4e0b\u7684\u8f68\u8ff9\u6982\u7387\u5dee\u5f02\uff0c\u7814\u7a76\u4eba\u5458\u53ef\u4ee5\u91cf\u5316\u6a21\u578b\u9c81\u68d2\u6027\uff0c\u8fdb\u800c\u8bbe\u8ba1\u66f4\u6709\u9488\u5bf9\u6027\u7684\u9632\u5fa1\u673a\u5236\u3002\u6587\u7ae0\u6697\u793a\uff0c\u8fd9\u4e00\u601d\u8def\u4e5f\u53ef\u7528\u4e8e\u77e5\u8bc6\u7f16\u8f91\u2014\u2014\u5f53\u4fee\u6539\u67d0\u4e2a\u4e8b\u5b9e\u6027 token \u540e\uff0c\u65b0\u65e7\u8f68\u8ff9\u6982\u7387\u7684\u843d\u5dee\u53ef\u4ee5\u7528\u6765\u8861\u91cf\u7f16\u8f91\u7684\u5f71\u54cd\u8303\u56f4\u3002<\/p>\n

\u4ece\u7406\u8bba\u5230\u5de5\u7a0b\u5b9e\u8df5\u7684\u6865\u6881<\/p><\/h2>

\u5728\u5de5\u7a0b\u5c42\u9762\uff0c\u8f68\u9053\u6982\u7387\u7684\u8ba1\u7b97\u6210\u672c\u5e76\u4e0d\u9ad8\u2014\u2014\u5b83\u672c\u6765\u5c31\u5728\u6bcf\u6b21\u524d\u5411\u4f20\u64ad\u4e2d\u9690\u542b\u4ea7\u751f\u3002\u771f\u6b63\u7f3a\u4e4f\u7684\u662f\u5c06\u5176\u4f5c\u4e3a\u4e00\u79cd\u8bc4\u4f30\u7ef4\u5ea6\u7684\u610f\u8bc6\u3002\u6587\u7ae0\u547c\u5401\uff0c\u5f00\u53d1\u8005\u5728\u8bb0\u5f55\u65e5\u5fd7\u65f6\uff0c\u9664\u4e86 token \u548c loss\uff0c\u4e5f\u5e94\u8bb0\u5f55\u8f68\u8ff9\u5bf9\u6570\u6982\u7387\uff0c\u8fd9\u5c06\u4e3a\u7ebf\u4e0a\u6a21\u578b\u7684\u5f02\u5e38\u68c0\u6d4b\u63d0\u4f9b\u5b9d\u8d35\u7684\u4fe1\u53f7\u3002\u4f8b\u5982\uff0c\u5f53\u7528\u6237\u89e6\u53d1\u67d0\u6761\u67e5\u8be2\u65f6\uff0c\u5982\u679c\u6a21\u578b\u751f\u6210\u8f68\u8ff9\u7684\u7d2f\u79ef\u6982\u7387\u7a81\u7136\u4f4e\u4e8e\u8bad\u7ec3\u5206\u5e03\u4e2d\u7684\u5e38\u89c1\u8303\u56f4\uff0c\u5f88\u53ef\u80fd\u610f\u5473\u7740\u8f93\u5165\u5c5e\u4e8e\u5206\u5e03\u5916\uff08OOD\uff09\u6216\u6a21\u578b\u5185\u90e8\u72b6\u6001\u5f02\u5e38\u3002<\/p>\n

\u8fd9\u7bc7\u6280\u672f\u7b14\u8bb0\u7684\u53ef\u8d35\u4e4b\u5904\uff0c\u5728\u4e8e\u5b83\u7528\u6700\u57fa\u7840\u7684\u6570\u5b66\u5de5\u5177\u89e3\u91ca\u4e86 LLM \u884c\u4e3a\u4e2d\u4e00\u4e2a\u88ab\u5ffd\u89c6\u7684\u7ef4\u5ea6\u3002\u968f\u7740\u5927\u6a21\u578b\u5728\u957f\u4e0a\u4e0b\u6587\u3001\u590d\u6742\u63a8\u7406\u7b49\u573a\u666f\u7684\u666e\u53ca\uff0c\u8f68\u8ff9\u6982\u7387\u6709\u671b\u6210\u4e3a\u7ee7 loss\u3001perplexity \u4e4b\u540e\u7684\u65b0\u4e00\u4ee3\u8c03\u8bd5\u6307\u6807\u2014\u2014\u5b83\u4e0d\u53ea\u770b\u6a21\u578b\u201c\u7b54\u6ca1\u7b54\u5bf9\u201d\uff0c\u66f4\u770b\u6a21\u578b\u201c\u8d70\u7684\u8def\u662f\u5426\u5178\u578b\u201d\u3002\u5bf9\u4e8e\u6240\u6709\u8bd5\u56fe\u8ba9\u5927\u6a21\u578b\u66f4\u53ef\u63a7\u3001\u66f4\u53ef\u89e3\u91ca\u7684\u5f00\u53d1\u8005\u800c\u8a00\uff0c\u8fd9\u4e2a\u89c6\u89d2\u503c\u5f97\u88ab\u7eb3\u5165\u65e5\u5e38\u5de5\u5177\u7bb1\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

\u7406\u89e3\u5927\u8bed\u8a00\u6a21\u578b\uff08LLM\uff09\u7684\u751f\u6210\u8fc7\u7a0b\uff0c\u901a\u5e38\u6211\u4eec\u4f1a\u5173\u6ce8\u4e0b\u4e00\u4e2a\u8bcd\u7684\u9884\u6d4b\u6982\u7387\u3002\u4f46\u9c9c\u6709\u4eba\u601d\u8003\uff1a\u6574\u4e2a\u751f\u6210\u5e8f\u5217\u2014\u2014\u5373\u4e00\u6761\u5b8c\u6574\u7684\u201c\u8f68\u8ff9\u201d\u2014\u2014\u5176\u8054\u5408\u6982\u7387\u662f\u5982\u4f55\u88ab\u8ba1\u7b97\u548c\u5206\u89e3\u7684\uff1f\u4e00\u7bc7\u6765\u81ea\u5f00\u53d1\u8005 Sithan Kanna \u7684\u6280\u672f\u6587\u7ae0\uff0c\u4ece\u6982\u7387\u8bba\u4e2d\u6700\u57fa\u7840\u7684\u94fe\u5f0f\u6cd5\u5219\u51fa\u53d1\uff0c\u91cd\u65b0\u63a8\u5bfc\u4e86 LLM \u8f93\u51fa\u8f68\u8ff9\u7684\u6982\u7387\u8868\u8fbe\uff0c\u4e3a\u8bc4\u4f30\u6a21\u578b\u884c\u4e3a\u7684\u53ef\u89e3\u91ca\u6027\u63d0\u4f9b\u4e86\u65b0\u7684\u89c6\u89d2\u3002<\/p>\n

## \u4ece\u94fe\u5f0f\u6cd5\u5219\u5230\u5e8f\u5217\u6982\u7387<\/p>\n

\u94fe\u5f0f\u6cd5\u5219\u662f\u6982\u7387\u8bba\u4e2d\u7684\u57fa\u672c\u5de5\u5177\uff0c\u5b83\u5141\u8bb8\u5c06\u591a\u4e2a\u968f<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-864","post","type-post","status-publish","format-standard","hentry","category-3"],"_links":{"self":[{"href":"https:\/\/www.timedigital.cn\/new\/wp-json\/wp\/v2\/posts\/864","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.timedigital.cn\/new\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.timedigital.cn\/new\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.timedigital.cn\/new\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.timedigital.cn\/new\/wp-json\/wp\/v2\/comments?post=864"}],"version-history":[{"count":0,"href":"https:\/\/www.timedigital.cn\/new\/wp-json\/wp\/v2\/posts\/864\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.timedigital.cn\/new\/wp-json\/wp\/v2\/media?parent=864"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.timedigital.cn\/new\/wp-json\/wp\/v2\/categories?post=864"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.timedigital.cn\/new\/wp-json\/wp\/v2\/tags?post=864"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}