![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/0812A88C28F61597144426FB4B2C24A7C8B47382_size50_w1080_h437.jpg\")
<\/p>\n
\u4f5c\u8005\uff5c\u6c64\u4e00\u6d9b<\/strong><\/p>\n \u7f16\u8f91\uff5c\u9756\u5b87<\/strong><\/p>\n Opus 4.7 \u521a\u53d1\u5e03\u90a3\u51e0\u5929\uff0cX \u4e0a\u6028\u58f0\u8f7d\u9053\u3002\u6709\u4eba\u8bf4\u4e00\u6b21\u5bf9\u8bdd\u5c31\u628a\u5979\u7684 session \u989d\u5ea6\u7528\u5149\u4e86\uff0c\u6709\u4eba\u8bf4\u540c\u4e00\u6bb5\u4ee3\u7801\u8dd1\u5b8c\u7684\u6210\u672c\u6bd4\u4e0a\u5468\u7ffb\u4e86\u4e00\u500d\u591a\uff1b\u8fd8\u6709\u4eba\u6652\u51fa\u81ea\u5df1 200 \u7f8e\u5143 Max \u8ba2\u9605\u4e0d\u5230\u4e24\u5c0f\u65f6\u5c31\u89e6\u9876\u7684\u622a\u56fe\u3002<\/p>\n \u72ec\u7acb\u5f00\u53d1\u8005 BridgeMind \u627f\u8ba4 Claude \u662f\u4e16\u754c\u4e0a\u6700\u597d\u7684\u6a21\u578b\uff0c\u4f46\u540c\u65f6\u4e5f\u662f\u6700\u8d35\u7684\u6a21\u578b\u3002\u4ed6\u7684 Max \u8ba2\u9605\u7528\u4e0d\u5230\u4e24\u5c0f\u65f6\u5c31\u9650\u989d\u4e86\uff0c\u4f46\u5e78\u597d\u2014\u2014\u4ed6\u4e70\u4e86\u4e24\u4efd\u3002\uff5c\u56fe\u7247\u6765\u6e90\uff1aX@bridgemindai<\/p>\n Anthropic \u5b98\u65b9\u4ef7\u683c\u6ca1\u53d8\uff0c\u6bcf\u767e\u4e07\u8f93\u5165 token \u4ecd\u662f 5 \u7f8e\u5143\uff0c\u8f93\u51fa 25 \u7f8e\u5143\u3002\u4f46\u8fd9\u4e2a\u7248\u672c\u5f15\u5165\u4e86\u65b0 tokenizer\uff0c\u540c\u65f6 Claude Code \u628a\u9ed8\u8ba4 effort \u4ece high \u63d0\u5230\u4e86 xhigh\u3002\u4e24\u4ef6\u4e8b\u53e0\u52a0\uff0c\u540c\u4e00\u4efd\u5de5\u4f5c\u6d88\u8017\u7684 token \u53d8\u6210\u4e86\u4ee5\u524d\u7684 2 \u5230 2.7 \u500d\u3002<\/p>\n \u6211\u5728\u8fd9\u4e9b\u8ba8\u8bba\u91cc\u770b\u5230\u4e24\u4e2a\u548c\u4e2d\u6587\u6709\u5173\u7684\u8bf4\u6cd5\u3002\u4e00\u4e2a\u662f\uff1a\u4e2d\u6587\u5728\u65b0 tokenizer \u4e0b\u51e0\u4e4e\u6ca1\u6da8\uff0c\u4e2d\u6587\u7528\u6237\u8eb2\u8fc7\u4e86\u8fd9\u6b21\u6da8\u4ef7\u3002\u53e6\u4e00\u4e2a\u66f4\u6709\u610f\u601d\uff1a\u53e4\u6587\u6bd4\u73b0\u4ee3\u6c49\u8bed\u8fd8\u7701 token\uff0c\u7528\u6587\u8a00\u6587\u8ddf AI \u5bf9\u8bdd\u53ef\u4ee5\u8282\u7701\u6210\u672c<\/strong>\u3002<\/p>\n \u7b2c\u4e00\u4e2a\u8bf4\u6cd5\u6697\u793a Claude \u5bf9\u4e2d\u6587\u505a\u4e86\u67d0\u79cd\u4f18\u5316\uff0c\u4f46 Anthropic \u7684\u53d1\u5e03\u6587\u6863\u91cc\uff0c\u6ca1\u63d0\u8fc7\u4efb\u4f55\u548c\u4e2d\u6587\u76f8\u5173\u7684\u8c03\u6574\u3002<\/p>\n \u7b2c\u4e8c\u4e2a\u8bf4\u6cd5\u5219\u66f4\u96be\u89e3\u91ca\u3002\u53e4\u6587\u5bf9\u4eba\u7c7b\u8bfb\u8005\u6765\u8bf4\u663e\u7136\u6bd4\u73b0\u4ee3\u6c49\u8bed\u96be\u61c2\uff0c\u4e00\u4e2a\u5bf9\u4eba\u7c7b\u66f4\u590d\u6742\u7684\u6587\u672c\uff0c\u600e\u4e48\u4f1a\u5bf9 AI \u66f4\u5bb9\u6613\uff1f<\/p>\n \u4e8e\u662f\u6211\u505a\u4e86\u4e00\u6b21\u6d4b\u8bd5\uff0c\u7528 22 \u6bb5\u5e73\u884c\u6587\u672c\uff08\u5305\u542b\u5546\u4e1a\u65b0\u95fb\u3001\u6280\u672f\u6587\u6863\u3001\u53e4\u6587\u3001\u65e5\u5e38\u5bf9\u8bdd\u7b49\u7c7b\u578b\uff09\uff0c\u540c\u65f6\u9001\u8fdb 5 \u4e2a tokenizer\uff08Claude 4.6 \u548c 4.7\u3001GPT-4o\u3001Qwen 3.6\u3001DeepSeek-V3\uff09\uff0c\u8bfb\u53d6\u6bcf\u6bb5\u6587\u672c\u5728\u6bcf\u4e2a\u6a21\u578b\u4e0b\u7684 token \u6570\uff0c\u505a\u6a2a\u5411\u5bf9\u6bd4\u3002<\/p>\n \u6d4b\u8bd5\u6587\u672c\uff1a<\/p>\n 1\u3001\u65e5\u5e38\u5bf9\u8bdd\u4e2d\u82f1\u6587\uff08\u65c5\u884c\u3001\u8bba\u575b\u6c42\u52a9\u3001\u5199\u4f5c\u8bf7\u6c42\uff09<\/p>\n 2\u3001\u6280\u672f\u6587\u6863\u4e2d\u82f1\u6587\uff08python \u6587\u6863\u3001Anthropic \u6587\u6863\uff09<\/p>\n 3\u3001\u65b0\u95fb\u4e2d\u82f1\u6587\uff08NYT \u65f6\u653f\u65b0\u95fb\u3001NYT \u5546\u4e1a\u65b0\u95fb\u3001\u82f9\u679c<\/a>\u516c\u53f8\u5b98\u65b9\u58f0\u660e\uff09<\/p>\n 4\u3001\u6587\u5b66\u9009\u6bb5\u4e2d\u82f1\u53e4\u6c49\u8bed\uff08\u300a\u51fa\u5e08\u8868\u300b\u300a\u9053\u5fb7\u7ecf\u300b\uff09<\/p>\n \u6d4b\u5b8c\u4e4b\u540e\uff0c\u4e24\u4e2a\u8bf4\u6cd5\u90fd\u5f97\u5230\u4e86\u90e8\u5206\u9a8c\u8bc1\uff0c\u4f46\u4e8b\u5b9e\u4f1a\u6bd4\u4f20\u8a00\u66f4\u590d\u6742\u4e00\u4e9b\u3002<\/p>\n 01<\/strong><\/p>\n \u4e2d\u6587\u7a0e<\/strong><\/p>\n \u5148\u8bf4\u7ed3\u8bba\uff1a<\/p>\n 1\u3001\u5728 Claude \u548c GPT \u4e0a\uff0c\u4e2d\u6587\u4e00\u76f4\u6bd4\u82f1\u6587\u8d35<\/strong><\/p>\n 2\u3001\u5728 Qwen \u548c DeepSeek \u4e0a\uff0c\u4e2d\u6587\u53cd\u800c\u6bd4\u82f1\u6587\u4fbf\u5b9c<\/strong><\/p>\n 3\u3001Opus 4.7 \u8fd9\u6b21\u5f15\u53d1\u9707\u8361\u7684 tokenizer \u5347\u7ea7\uff0c\u901a\u80c0<\/a>\u51e0\u4e4e\u53ea\u53d1\u751f\u5728\u82f1\u6587\u4e0a\uff0c\u4e2d\u6587\u7eb9\u4e1d\u4e0d\u52a8<\/strong><\/p>\n \u770b\u5177\u4f53\u6570\u5b57\u3002Claude Opus 4.7 \u4e4b\u524d\u7684\u5168\u7cfb\u5217\u6a21\u578b\uff08\u5305\u62ec Opus 4.6\u3001Sonnet\u3001Haiku\uff09\uff0c\u4f7f\u7528\u7684\u662f\u540c\u4e00\u4e2a tokenizer\u3002\u5728\u8fd9\u4e2a tokenizer \u4e0b\uff0c\u4e2d\u6587\u7684 token \u6d88\u8017\u5168\u7ebf\u9ad8\u4e8e\u7b49\u91cf\u82f1\u6587\u5185\u5bb9\uff0ccn\/en \u6bd4\u503c\u8303\u56f4\u5728 1.11× \u5230 1.64× \u4e4b\u95f4\u3002<\/p>\n \u6700\u6781\u7aef\u7684\u573a\u666f\u51fa\u73b0\u5728 NYT \u98ce\u683c\u7684\u5546\u4e1a\u65b0\u95fb\uff1a\u540c\u4e00\u6bb5\u5185\u5bb9\uff0c\u4e2d\u6587\u7248\u8981\u591a\u6d88\u8017 64% \u7684 token\uff0c\u7b49\u4e8e\u591a\u4ed8 64% \u7684\u94b1\u3002<\/p>\n Opus 4.6 \u53ca\u5176\u4e4b\u524d\u7684 Claude \u6a21\u578b\uff0c\u4e2d\u6587 token \u7684\u6d88\u8017\u91cf\u663e\u8457\u9ad8\u4e8e\u5176\u5b83\u6a21\u578b\uff08\u7ea2\u6846\uff09<\/p>\n \u6700\u6781\u7aef\u7684\u573a\u666f\u51fa\u73b0\u5728 NYT \u98ce\u683c\u7684\u5546\u4e1a\u65b0\u95fb\uff1a\u540c\u4e00\u6bb5\u5185\u5bb9\uff0c\u4e2d\u6587\u7248\u8981\u591a\u6d88\u8017 64% \u7684 token\uff08\u7eff\u6846\uff09<\/p>\n GPT-4o \u7684 o200k tokenizer \u597d\u4e00\u4e9b\uff0ccn\/en \u6bd4\u503c\u591a\u6570\u843d\u5728 1.0 \u5230 1.35× \u4e4b\u95f4\uff0c\u90e8\u5206\u573a\u666f\u4f4e\u4e8e 1\u3002\u4e2d\u6587\u4ecd\u7136\u6574\u4f53\u504f\u8d35\uff0c\u4f46\u5dee\u8ddd\u6bd4 Claude \u5c0f\u5f97\u591a\u3002<\/p>\n \u56fd\u4ea7\u6a21\u578b Qwen 3.6 \u548c DeepSeek-V3 \u7684\u6570\u636e\u5219\u5b8c\u5168\u53cd\u4e86\u8fc7\u6765\u3002\u4e24\u8005\u7684 cn\/en \u6bd4\u503c\u5927\u9762\u79ef\u4f4e\u4e8e 1\uff0c\u8fd9\u610f\u5473\u7740\u540c\u6837\u7684\u5185\u5bb9\uff0c\u4e2d\u6587\u7248\u53cd\u800c\u6bd4\u82f1\u6587\u7248\u7701 token\u3002DeepSeek \u6700\u4f4e\u505a\u5230\u4e86 0.65×\uff0c\u540c\u4e00\u6bb5\u8bdd\u4e2d\u6587\u7248\u6bd4\u82f1\u6587\u7248\u4fbf\u5b9c\u4e09\u5206\u4e4b\u4e00<\/strong>\u3002<\/p>\n Opus 4.7 \u7684\u65b0 tokenizer \u901a\u80c0\u51e0\u4e4e\u53ea\u53d1\u751f\u5728\u82f1\u6587\u4e0a\u3002\u82f1\u6587 token \u6570\u81a8\u80c0\u4e86 1.24× \u5230 1.63×\uff0c\u4e2d\u6587\u5927\u91cf\u7ef4\u6301\u5728 1.000×\uff0c\u51e0\u4e4e\u6ca1\u6709\u53d8\u5316\u3002\u5f00\u5934\u90a3\u4e9b\u82f1\u6587\u5f00\u53d1\u8005\u7684\u8d26\u5355\u9707\u8361\uff0c\u4e2d\u6587\u7528\u6237\u786e\u5b9e\u6ca1\u611f\u53d7\u5230\u3002\u539f\u56e0\u53ef\u80fd\u662f\u4e2d\u6587\u5728\u65e7\u7248\u4e0a\u5df2\u7ecf\u88ab\u5207\u5230\u4e86\u5355\u5b57\u9897\u7c92\u5ea6\uff0c\u53ef\u62c6\u5206\u7684\u7a7a\u95f4\u6781\u5c0f\u3002<\/p>\n Opus 4.7 \u5bf9\u6bd4 4.6\uff0c\u82f1\u6587\u6d88\u8017\u7684 token \u66f4\u591a\u4e86\uff0c\u4e2d\u6587\u53cd\u800c\u6ca1\u53d8<\/p>\n \u6d4b\u8bd5\u8fc7\u7a0b\u4e2d\u6211\u8fd8\u6ce8\u610f\u5230\u4e00\u4ef6\u4e8b\u3002token \u6d88\u8017\u7684\u5dee\u5f02\u4e0d\u53ea\u662f\u8d26\u5355\u95ee\u9898\uff0c\u5b83\u76f4\u63a5\u5f71\u54cd\u5de5\u4f5c\u7a7a\u95f4\u7684\u5927\u5c0f\u3002\u540c\u6837 200k \u4e0a\u4e0b\u6587\u7a97\u53e3\uff0c\u7528\u65e7\u7248 Claude tokenizer \u88c5\u4e2d\u6587\u8d44\u6599\uff0c\u80fd\u585e\u8fdb\u53bb\u7684\u5185\u5bb9\u91cf\u6bd4\u82f1\u6587\u5c11 40% \u5230 70%\u3002<\/p>\n \u540c\u4e00\u7c7b\u5de5\u4f5c\uff0c\u6bd4\u5982\u8ba9 AI \u5206\u6790\u4e00\u4efd\u957f\u6587\u6863\u6216\u8005\u662f\u603b\u7ed3\u4e00\u7ec4\u4f1a\u8bae\u8bb0\u5f55\uff0c\u4e2d\u6587\u7528\u6237\u80fd\u5582\u7ed9\u6a21\u578b\u7684\u6750\u6599\u66f4\u5c11\uff0c\u6a21\u578b\u80fd\u53c2\u8003\u7684\u4e0a\u4e0b\u6587\u66f4\u77ed\u3002\u7ed3\u679c\u5c31\u662f\u4ed8\u4e86\u66f4\u591a\u7684\u94b1\uff0c\u4f46\u5f97\u5230\u7684\u662f\u66f4\u5c0f\u7684\u5de5\u4f5c\u7a7a\u95f4\u3002<\/p>\n \u56db\u7ec4\u6570\u636e\u653e\u5728\u4e00\u8d77\u770b\uff0c\u4e00\u4e2a\u95ee\u9898\u81ea\u7136\u6d6e\u51fa\u6765\uff1a<\/p>\n \u4e3a\u4ec0\u4e48\u540c\u4e00\u6bb5\u5185\u5bb9\u6362\u4e2a\u8bed\u8a00\uff0ctoken \u6570\u5c31\u4e0d\u4e00\u6837\uff1f\u4e3a\u4ec0\u4e48 Claude \u548c GPT \u7684\u4e2d\u6587\u8d35\uff0cQwen \u548c DeepSeek \u7684\u4e2d\u6587\u53cd\u800c\u4fbf\u5b9c<\/strong>\uff1f<\/p>\n \u7b54\u6848\u85cf\u5728\u4e0a\u6587\u591a\u6b21\u63d0\u5230\u7684\u6982\u5ff5 tokenizer\uff08\u5206\u8bcd\u5668\uff09\u4e0a\u3002<\/p>\n 02<\/strong><\/p>\n \u4e00\u4e2a\u6c49\u5b57\uff0c\u53ef\u4ee5\u5207\u6210\u51e0\u5757\uff1f<\/strong><\/p>\n \u6a21\u578b\u5728\u8bfb\u5230\u4efb\u4f55\u6587\u5b57\u4e4b\u524d\uff0c\u4f1a\u901a\u8fc7 tokenizer \u628a\u8f93\u5165\u5207\u6210\u4e00\u4e2a\u4e2a token\u3002\u4f60\u53ef\u4ee5\u628a tokenizer \u60f3\u8c61\u6210 AI \u7684\u300c\u79ef\u6728\u5207\u5272\u673a\u300d\u3002\u4f60\u8f93\u5165\u4e00\u53e5\u8bdd\uff0c\u5b83\u8d1f\u8d23\u628a\u8fd9\u53e5\u8bdd\u62c6\u6210\u4e00\u5757\u5757\u6807\u51c6\u5316\u7684\u79ef\u6728\uff08\u4e5f\u5c31\u662f token\uff09\u3002AI \u6a21\u578b\u4e0d\u770b\u6587\u5b57\uff0c\u53ea\u8ba4\u79ef\u6728\u7684\u7f16\u53f7\u3002\u4f60\u7528\u591a\u5c11\u5757\u79ef\u6728\uff0c\u5c31\u4ed8\u591a\u5c11\u94b1\u3002<\/p>\n \u82f1\u6587\u7684\u5207\u6cd5\u6bd4\u8f83\u7b26\u5408\u76f4\u89c9\uff0c\u6bd4\u5982\u300cintelligence\u300d\u5927\u6982\u7387\u662f\u4e00\u4e2a token\uff0c\u300cinformation\u300d\u4e5f\u662f\u4e00\u4e2a token\uff0c\u4e00\u4e2a\u5355\u8bcd\u5bf9\u5e94\u4e00\u4e2a\u8ba1\u8d39\u5355\u4f4d\u3002<\/p>\n \u4f46\u4e2d\u6587\u5230\u4e86\u8fd9\u4e00\u6b65\u5c31\u51fa\u95ee\u9898\u4e86\u3002\u628a\u540c\u4e00\u53e5\u8bdd\u300c\u4eba\u5de5\u667a\u80fd<\/a>\u6b63\u5728\u91cd\u5851\u5168\u7403\u7684\u4fe1\u606f\u57fa\u7840\u8bbe\u65bd\u300d\u5206\u522b\u9001\u8fdb GPT-4 \u7684 cl100k tokenizer \u548c Qwen 2.5 \u7684 tokenizer\uff0c\u5207\u51fa\u6765\u7684\u7ed3\u679c\u5b8c\u5168\u4e0d\u540c\u3002<\/p>\n GPT-4 \u57fa\u672c\u628a\u6bcf\u4e00\u4e2a\u6c49\u5b57\u90fd\u62c6\u6210\u4e86\u4e00\u4e2a token\uff1bQwen \u5219\u4f1a\u628a\u8bcd\u8bed\u8bc6\u522b\u6210\u4e00\u4e2a token\uff0c\u4f8b\u5982\u300c\u4eba\u5de5\u667a\u80fd\u300d\u8fd9 4 \u4e2a\u5b57\u5728\u5343\u95ee\u53ea\u7b97\u4e00\u4e2a token\u3002<\/p>\n \u540c\u4e00\u53e5 16 \u4e2a\u6c49\u5b57\u7684\u8bdd\uff0cGPT-4 \u5207\u51fa\u6765 19 \u4e2a token\uff0cQwen \u5207\u51fa\u6765\u53ea\u6709 6 \u4e2a\u3002<\/p>\n \u4e3a\u4ec0\u4e48\u4f1a\u5207\u6210\u8fd9\u6837\uff1f\u539f\u56e0\u5728\u4e00\u4e2a\u53eb BPE\uff08Byte Pair Encoding\uff09\u7684\u7b97\u6cd5\u3002<\/p>\n BPE \u7684\u5de5\u4f5c\u65b9\u5f0f\uff0c\u662f\u7edf\u8ba1\u8bad\u7ec3\u8bed\u6599\u91cc\u54ea\u4e9b\u5b57\u7b26\u7ec4\u5408\u51fa\u73b0\u9891\u7387\u6700\u9ad8\uff0c\u7136\u540e\u628a\u9ad8\u9891\u7ec4\u5408\u5408\u5e76\u6210\u4e00\u4e2a token\uff0c\u7eb3\u5165\u8bcd\u8868\u3002<\/p>\n GPT-2 \u65f6\u4ee3\uff0c\u8bad\u7ec3\u8bed\u6599\u7684\u7edd\u5927\u591a\u6570\u662f\u82f1\u6587\u3002\u82f1\u6587\u5b57\u6bcd\u7ec4\u5408\uff08th\u3001ing\u3001tion\uff09\u53cd\u590d\u51fa\u73b0\uff0c\u5f88\u5feb\u5c31\u88ab\u5408\u5e76\u6210 token\u3002\u4e2d\u6587\u5b57\u7b26\u5728\u90a3\u4e2a\u8bed\u6599\u6c60\u91cc\u51fa\u73b0\u7684\u9891\u7387\u592a\u4f4e\uff0c\u6392\u4e0d\u8fdb\u8bcd\u8868\uff0c\u53ea\u80fd\u88ab\u5f53\u4f5c\u539f\u59cb\u5b57\u8282\u6765\u5904\u7406\uff0c\u4e00\u4e2a\u6c49\u5b57\u5360 3 \u4e2a\u5b57\u8282\uff0c\u5c31\u53d8\u6210\u4e86 3 \u4e2a token\u3002<\/p>\n BPE \u6309\u8bad\u7ec3\u8bed\u6599\u4e2d\u7684\u5b57\u7b26\u9891\u7387\u51b3\u5b9a\u5408\u5e76\u3002\u82f1\u6587\u8bed\u6599\u4e3b\u5bfc\u4e0b\uff0c\u4e2d\u6587 UTF-8 \u5b57\u8282\u65e0\u6cd5\u5408\u5e76\u4e3a\u6574\u5b57<\/p>\n \u540e\u6765 GPT-4 \u7684 cl100k \u8bcd\u8868\u6269\u5927\u4e86\uff0c\u5e38\u7528\u6c49\u5b57\u5f00\u59cb\u88ab\u7eb3\u5165\uff0c\u4e00\u4e2a\u5b57\u901a\u5e38\u7f29\u5230 1 \u5230 2 \u4e2a token\uff0c\u4f46\u6574\u4f53\u6548\u7387\u4ecd\u7136\u4e0d\u5982\u82f1\u6587\u3002<\/p>\n \u5230\u4e86 GPT-4o \u7684 o200k \u8bcd\u8868\uff0c\u4e2d\u6587\u6548\u7387\u518d\u8fdb\u4e86\u4e00\u6b65\u3002\u8fd9\u4e5f\u89e3\u91ca\u4e86\u4e3a\u4ec0\u4e48\u7b2c\u4e00\u6bb5\u7684\u6570\u636e\u91cc GPT-4o \u7684 cn\/en \u6bd4\u503c\u6bd4 Claude \u4f4e\u3002<\/p>\n Qwen \u548c DeepSeek \u4f5c\u4e3a\u56fd\u4ea7\u6a21\u578b\uff0c\u4ece\u4e00\u5f00\u59cb\u5c31\u628a\u5927\u91cf\u5e38\u7528\u6c49\u5b57\u548c\u9ad8\u9891\u8bcd\u7ec4\u4f5c\u4e3a\u6574\u5b57\u3001\u6574\u8bcd\u7eb3\u5165\u8bcd\u8868\u3002\u4e00\u4e2a\u5b57\u4e00\u4e2a token\uff0c\u6548\u7387\u76f4\u63a5\u7ffb\u500d\u751a\u81f3\u66f4\u591a\u3002<\/p>\n \u540c\u4e00\u53e5\u8bdd\u5728\u4e0d\u540c tokenizer \u4e0b\u7684\u62c6\u5206\u7ed3\u679c\u793a\u610f\u56fe<\/p>\n \u8fd9\u5c31\u662f\u4e3a\u4ec0\u4e48\u5b83\u4eec\u7684 cn\/en \u6bd4\u503c\u80fd\u4f4e\u4e8e 1\uff0c\u4e2d\u6587\u5b57\u5747\u4fe1\u606f\u5bc6\u5ea6\u672c\u6765\u5c31\u9ad8\u4e8e\u82f1\u6587\u5355\u8bcd\uff0c\u5f53 tokenizer \u4e0d\u518d\u4eba\u4e3a\u62c6\u788e\u6c49\u5b57\uff0c\u8fd9\u4e2a\u5929\u7136\u4f18\u52bf\u5c31\u663e\u73b0\u51fa\u6765\u4e86<\/strong>\u3002<\/p>\n \u6240\u4ee5\u4e0a\u4e00\u8282\u90a3\u56db\u7ec4\u6570\u636e\u7684\u5dee\u5f02\uff0c\u6839\u6e90\u4e0d\u5728\u6a21\u578b\u7684\u80fd\u529b\uff0c\u800c\u5728 tokenizer \u7684\u8bcd\u8868\u91cc\uff0c\u7ed9\u4e2d\u6587\u7559\u4e86\u591a\u5c11\u4f4d\u7f6e\u3002<\/p>\n Claude \u548c\u65e9\u671f GPT \u7684\u8bcd\u8868\u662f\u4ee5\u82f1\u6587\u4e3a\u9ed8\u8ba4\u503c\u6784\u5efa\u7684\uff0c\u4e2d\u6587\u662f\u540e\u6765\u88ab\u300c\u585e\u8fdb\u53bb\u300d\u7684\uff1bQwen \u548c DeepSeek \u7684\u8bcd\u8868\u4ece\u8bbe\u8ba1\u4e4b\u521d\u5c31\u628a\u4e2d\u6587\u5f53\u4f5c\u9ed8\u8ba4\u8bed\u8a00\u5bf9\u5f85\u3002\u8fd9\u4e2a\u8d77\u70b9\u7684\u5dee\u5f02\uff0c\u4e00\u8def\u4f20\u5bfc\u5230 token \u6570\u3001\u8d26\u5355\u3001\u4e0a\u4e0b\u6587\u7a97\u53e3\u5927\u5c0f\u3002<\/p>\n 03<\/strong><\/p>\n \u53e4\u6587\u771f\u7684\u66f4\u4fbf\u5b9c\u5417\uff1f<\/strong><\/p>\n \u518d\u770b\u5f00\u5934\u7684\u7b2c\u4e8c\u4e2a\u4f20\u8a00\uff1a\u53e4\u6587\u6bd4\u73b0\u4ee3\u6c49\u8bed\u66f4\u7701 token<\/strong>\u3002<\/p>\n \u6570\u636e\u786e\u8ba4\u4e86\u8fd9\u4e2a\u8bf4\u6cd5\u3002\u5728\u6d4b\u8bd5\u91cc\uff0c\u53e4\u6587\u6837\u672c\u7684 cn\/en \u6bd4\u503c\u5168\u7ebf\u4f4e\u4e8e 1\uff0c\u5728\u6240\u6709\u4e94\u4e2a tokenizer \u4e0a\u90fd\u4e00\u81f4\u3002\u540c\u4e00\u6bb5\u5185\u5bb9\u7684\u53e4\u6587\u7248\u672c\uff0ctoken \u6570\u6bd4\u5bf9\u5e94\u82f1\u6587\u7ffb\u8bd1\u8fd8\u5c11\u3002<\/p>\n \u5728\u6240\u6709\u6a21\u578b\u4e2d\uff0c\u53e4\u6587\u6d88\u8017\u7684 token \u6570\u4e0d\u4f46\u6bd4\u73b0\u4ee3\u4e2d\u6587\u5c11\uff0c\u751a\u81f3\u6bd4\u82f1\u6587\u8fd8\u5c11<\/p>\n \u539f\u56e0\u4e5f\u4e0d\u590d\u6742\uff0c\u53e4\u6587\u7528\u5b57\u6781\u5ea6\u7cbe\u70bc\u3002\u300c\u5b66\u800c\u4e0d\u601d\u5219\u7f54\uff0c\u601d\u800c\u4e0d\u5b66\u5219\u6b86\u300d\u662f 12 \u4e2a\u5b57\u3002\u7ffb\u8bd1\u6210\u73b0\u4ee3\u6c49\u8bed\u5c31\u662f\u300c\u53ea\u662f\u5b66\u4e60\u800c\u4e0d\u601d\u8003\u5c31\u4f1a\u8ff7\u60d1\uff0c\u53ea\u662f\u601d\u8003\u800c\u4e0d\u5b66\u4e60\u5c31\u4f1a\u9677\u5165\u56f0\u5883\u300d\uff0c\u5b57\u6570\u76f4\u63a5\u7ffb\u500d\uff0ctoken \u6570\u81ea\u7136\u4e5f\u8ddf\u7740\u7ffb\u500d\u3002<\/p>\n \u800c\u4e14\u53e4\u6587\u7684\u5e38\u7528\u5b57\uff08\u4e4b\u3001\u4e5f\u3001\u8005\u3001\u800c\u3001\u4e0d\uff09\u90fd\u662f\u9ad8\u9891\u5b57\u7b26\uff0c\u5728\u4efb\u4f55 tokenizer \u7684\u8bcd\u8868\u91cc\u90fd\u6709\u72ec\u7acb\u4f4d\u7f6e\uff0c\u4e0d\u4f1a\u88ab\u62c6\u6210\u5b57\u8282\u3002\u6240\u4ee5\u53e4\u6587\u5728\u7f16\u7801\u5c42\u9762\u786e\u5b9e\u662f\u9ad8\u6548\u7684\u3002<\/p>\n \u4f46\u8fd9\u91cc\u85cf\u7740\u4e00\u4e2a\u9677\u9631\u3002<\/p>\n \u53e4\u6587\u7684 token \u7701\u5728\u7f16\u7801\u7aef\uff0c\u4f46\u6a21\u578b\u7684\u63a8\u7406\u8d1f\u62c5\u6ca1\u6709\u51cf\u8f7b<\/strong>\u3002\u300c\u7f54\u300d\u4e00\u4e2a\u5b57\uff0c\u6a21\u578b\u9700\u8981\u5224\u65ad\u5b83\u5728\u8fd9\u4e2a\u8bed\u5883\u91cc\u662f\u300c\u8ff7\u60d1\u300d\u300c\u88ab\u8499\u853d\u300d\u8fd8\u662f\u300c\u6ca1\u6709\u300d\u3002\u73b0\u4ee3\u6c49\u8bed\u53ef\u4ee5\u7528 26 \u4e2a\u5b57\u628a\u8fd9\u5c42\u610f\u601d\u8bf4\u6e05\u695a\uff0c\u7528\u53e4\u6587\u7b49\u4e8e\u628a\u94fa\u5f00\u7684\u90e8\u5206\u538b\u4e86\u56de\u53bb\uff0c\u628a\u63a8\u7406\u7684\u6d3b\u7559\u7ed9\u4e86\u6a21\u578b\u3002\u6253\u4e2a\u6bd4\u65b9\uff0c\u4e00\u4efd\u538b\u7f29\u6210 zip \u7684\u6587\u4ef6\u4f53\u79ef\u66f4\u5c0f\uff0c\u4f46\u89e3\u538b\u5b83\u9700\u8981\u66f4\u591a\u8ba1\u7b97\u3002<\/p>\n token \u7701\u4e86\uff0c\u63a8\u7406\u7684\u6d88\u8017\u53cd\u800c\u4e0a\u5347\u4e86\uff0c\u7406\u89e3\u51c6\u786e\u5ea6\u8fd8\u4e0b\u964d\u4e86<\/strong>\u3002\u8fd9\u7b14\u8d26\u7b97\u4e0d\u8fc7\u6765\u3002<\/p>\n \u53e4\u6587\u8fd9\u4e2a\u4f8b\u5b50\u8ba9\u6211\u610f\u8bc6\u5230\uff0ctoken \u6570\u91cf\u672c\u8eab\u4e0d\u80fd\u8bf4\u660e\u592a\u591a\u95ee\u9898\u3002\u4f46\u987a\u7740\u8fd9\u4e2a\u65b9\u5411\u60f3\u4e0b\u53bb\uff0c\u8fd8\u6709\u4e00\u5c42\u6211\u4e4b\u524d\u5ffd\u7565\u4e86\u7684\u4e1c\u897f\u3002<\/p>\n \u4e0a\u9762\u8bf4\u8fc7\uff0cGPT-2 \u65f6\u4ee3\u7684 tokenizer \u4f1a\u628a\u300c\u4eba\u300d\u8fd9\u4e2a\u5b57\u62c6\u6210\u4e09\u4e2a UTF-8 \u5b57\u8282 token\uff0c\u540e\u6765 GPT-4 \u7684\u8bcd\u8868\u6269\u5927\uff0c\u5e38\u7528\u6c49\u5b57\u53d8\u6210\u4e86\u4e00\u4e2a\u5b57\u4e00\u4e2a token\uff0cQwen \u66f4\u8fdb\u4e00\u6b65\uff0c\u628a\u300c\u4eba\u5de5\u667a\u80fd\u300d\u56db\u4e2a\u5b57\u5408\u6210\u4e00\u4e2a token\u3002<\/p>\n \u76f4\u89c9\u4e0a\u8fd9\u662f\u4e00\u4e2a\u4e0d\u65ad\u6539\u8fdb\u7684\u8fc7\u7a0b\uff1a\u5408\u5e76\u5f97\u8d8a\u591a\uff0c\u6548\u7387\u8d8a\u9ad8\uff0c\u6a21\u578b\u5e94\u8be5\u4e5f\u7406\u89e3\u5f97\u8d8a\u597d\u3002<\/p>\n \u4f46\u771f\u7684\u662f\u8fd9\u6837\u5417\uff1f\u6211\u4eec\u4e0d\u59a8\u56de\u5fc6\u4e00\u4e0b\uff0c\u6211\u4eec\u662f\u5982\u4f55\u8ba4\u8bc6\u6c49\u5b57\u7684\u3002<\/p>\n \u6c49\u5b57\u662f\u8868\u610f\u6587\u5b57\uff0c\u73b0\u4ee3\u6c49\u5b57\u91cc\u8d85\u8fc7 80% \u662f\u5f62\u58f0\u5b57\uff0c\u7531\u4e00\u4e2a\u8868\u4e49\u7684\u504f\u65c1\u548c\u4e00\u4e2a\u8868\u97f3\u7684\u90e8\u4ef6\u7ec4\u5408\u800c\u6210\u3002\u300c\u6c35\u300d\u65c1\u7684\u5b57\u591a\u548c\u6db2\u4f53\u6709\u5173\uff0c\u300c\u6728\u300d\u65c1\u7684\u5b57\u591a\u548c\u690d\u7269\u6709\u5173\uff0c\u300c\u706b\u300d\u65c1\u7684\u5b57\u591a\u548c\u70ed\u91cf\u6709\u5173\u3002\u504f\u65c1\u90e8\u9996\u5c31\u662f\u4eba\u7c7b\u8bc6\u5b57\u65f6\u6700\u57fa\u7840\u7684\u8bed\u4e49\u7ebf\u7d22\uff0c\u4e00\u4e2a\u4e0d\u8ba4\u8bc6\u300c\u7131\u300d\u5b57\u7684\u4eba\uff0c\u770b\u5230 3 \u4e2a\u300c\u706b\u300d\u4e5f\u80fd\u731c\u5230\u5b83\u548c\u706b\u6709\u5173\u3002<\/strong><\/p>\n \u56e0\u4e3a\u504f\u65c1\u90e8\u9996\u662f\u4eba\u7c7b\u8bc6\u5b57\u65f6\u6700\u57fa\u7840\u7684\u8bed\u4e49\u7ebf\u7d22\uff0c\u4eba\u4f1a\u5148\u4ece\u7ed3\u6784\u63a8\u65ad\u610f\u4e49\u8303\u7574\uff0c\u518d\u7ed3\u5408\u8bed\u5883\u7406\u89e3\u5177\u4f53\u542b\u4e49\u3002<\/p>\n \u706b\u82b1\u3001\u706b\u7130\u3001\u5149\u7130\uff0c\u4e66\u9762\u8bed\u4e0e\u4eba\u540d\u4e2d\u591a\u89c1\uff0c\u5bd3\u610f\u5149\u660e\u3001\u70bd\u70ed\u3002<\/p>\n \u4f46\u662f\u5728 tokenizer \u7684\u8bcd\u8868\u91cc\uff0c\u300c\u7131\u300d\u8fd9\u4e2a\u5b57\u5bf9\u5e94\u7684\u662f\u4e00\u4e2a\u7f16\u53f7\u3002\u6211\u4eec\u5047\u8bbe\u5b83\u662f 38721 \u53f7\uff0c\u5b83\u4ee3\u8868\u7684\u662f\u8bcd\u8868\u91cc\u7684\u4e00\u4e2a\u7d22\u5f15\u4f4d\u7f6e\uff0c\u6a21\u578b\u901a\u8fc7\u5b83\u67e5\u627e\u5230\u4e00\u7ec4\u6570\u5b57\u5411\u91cf\uff0c\u7528\u8fd9\u7ec4\u5411\u91cf\u6765\u8868\u5f81\u300c\u7131\u300d\u8fd9\u4e2a\u5b57\u3002<\/p>\n \u7f16\u53f7\u672c\u8eab\u4e0d\u643a\u5e26\u4efb\u4f55\u5173\u4e8e\u8fd9\u4e2a\u5b57\u5185\u90e8\u7ed3\u6784\u7684\u4fe1\u606f\u300238721 \u548c 38722 \u7684\u5173\u7cfb\uff0c\u5bf9\u6a21\u578b\u6765\u8bf4\u548c 1 \u548c 10000 \u7684\u5173\u7cfb\u6ca1\u6709\u533a\u522b\u3002\u4e8e\u662f\uff0c\u300c\u6c49\u5b57\u7684\u7ed3\u6784\u300d\u8fd9\u4e00\u5c42\u4fe1\u606f\uff0c\u5c31\u88ab\u5c01\u88c5\u8d77\u6765\u4e86\u3002\u4e09\u4e2a\u300c\u706b\u300d\u53e0\u5728\u4e00\u8d77\u8fd9\u4ef6\u4e8b\uff0c\u5728\u7f16\u53f7\u91cc\u4e0d\u5b58\u5728\u3002<\/p>\n \u6a21\u578b\u5f53\u7136\u53ef\u4ee5\u901a\u8fc7\u5927\u91cf\u8bad\u7ec3\u6570\u636e\u95f4\u63a5\u5b66\u5230\u300c\u7131\u300d\u300c\u708e\u300d\u300c\u707c\u300d\u7ecf\u5e38\u51fa\u73b0\u5728\u76f8\u4f3c\u7684\u8bed\u5883\u91cc\uff0c\u4f46\u8fd9\u6761\u8def\u6bd4\u76f4\u63a5\u5229\u7528\u504f\u65c1\u4fe1\u606f\u8981\u66f4\u95f4\u63a5\u4e00\u4e9b\u3002<\/p>\n \u6240\u4ee5\u6a21\u578b\u80fd\u4e0d\u80fd\u4ece\u62c6\u5f00\u7684\u5b57\u8282\u91cc\uff0c\u300c\u770b\u5230\u300d\u67d0\u4e9b\u7c7b\u4f3c\u504f\u65c1\u7684\u7ed3\u6784\u7ebf\u7d22\uff0c\u7136\u540e\u5728\u540e\u7eed\u7684\u8ba1\u7b97\u5c42\u91cc\u91cd\u65b0\u7ec4\u5408\u5462\uff1f\u8fd9\u6761\u8def\u867d\u7136 token \u6570\u591a\u3001\u6210\u672c\u9ad8\uff0c\u4f46\u6709\u6ca1\u6709\u53ef\u80fd\u5728\u8bed\u4e49\u7406\u89e3\u4e0a\uff0c\u53cd\u800c\u6bd4\u76f4\u63a5\u541e\u4e0b\u4e00\u4e2a\u4e0d\u900f\u660e\u7684\u7f16\u53f7\u66f4\u6709\u6548\uff1f<\/p>\n 2025 \u5e74\u53d1\u8868\u5728 MIT Press\u300aComputational Linguistics\u300b\u4e0a\u7684\u4e00\u7bc7\u8bba\u6587\uff08\u300aTokenization Changes Meaning in Large Language Models: Evidence from Chinese\u300b\uff09\uff0c\u56de\u7b54\u4e86\u8fd9\u4e2a\u95ee\u9898\u3002<\/p>\n 04<\/strong><\/p>\n \u788e\u7247\u91cc\u957f\u51fa\u504f\u65c1<\/strong><\/p>\n \u8bba\u6587\u4f5c\u8005 David Haslett \u6ce8\u610f\u5230\u4e00\u4e2a\u5386\u53f2\u5de7\u5408\u3002<\/p>\n 1990 \u5e74\u4ee3\uff0cUnicode \u8054\u76df\u5728\u7ed9\u6c49\u5b57\u5206\u914d UTF-8 \u7f16\u7801\u65f6\uff0c\u6392\u5217\u987a\u5e8f\u662f\u6309\u90e8\u9996\u5f52\u7c7b\u6392\u7684\u3002\u540c\u4e00\u4e2a\u90e8\u9996\u4e0b\u7684\u6c49\u5b57\uff0cUTF-8 \u7f16\u7801\u662f\u76f8\u90bb\u7684\u3002\u300c\u8336\u300d\u548c\u300c\u830e\u300d\u90fd\u542b\u6709\u300c\u8279\u300d\u90e8\uff08\u8349\u5b57\u5934\uff09\uff0c\u5b83\u4eec\u7684 UTF-8 \u5b57\u8282\u5e8f\u5217\u4ee5\u76f8\u540c\u7684\u5b57\u8282\u5f00\u5934\u3002\u300c\u6cb3\u300d\u548c\u300c\u6d77\u300d\u90fd\u542b\u6709\u300c\u6c35\u300d\u90e8\uff0c\u5b57\u8282\u5e8f\u5217\u540c\u6837\u5171\u4eab\u5f00\u5934\u3002<\/p>\n UTF-8 \u6309\u7167\u90e8\u5206\u90e8\u9996\u987a\u5e8f\u7ed9\u4e2d\u6587\u6392\u5e8f\uff0c\u90e8\u9996\u76f8\u540c\u7684\u5b57\uff0c\u7f16\u7801\u76f8\u8fd1\uff5c\u56fe\u7247\u6765\u6e90\uff1aGithub<\/p>\n \u8fd9\u610f\u5473\u7740\uff0c\u5f53 tokenizer \u628a\u6c49\u5b57\u62c6\u6210\u4e09\u4e2a UTF-8 \u5b57\u8282 token \u7684\u65f6\u5019\uff0c\u5171\u4eab\u90e8\u9996\u7684\u6c49\u5b57\u4f1a\u5171\u4eab\u7b2c\u4e00\u4e2a token\u3002\u6a21\u578b\u5728\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u53cd\u590d\u770b\u5230\u8fd9\u4e9b\u5171\u4eab\u7684\u5b57\u8282\u6a21\u5f0f\uff0c\u6709\u53ef\u80fd\u4ece\u4e2d\u5b66\u5230\u300c\u7b2c\u4e00\u4e2a token \u76f8\u540c\u7684\u5b57\uff0c\u5f80\u5f80\u5c5e\u4e8e\u540c\u4e00\u4e2a\u610f\u4e49\u8303\u7574\u300d\u3002\u8fd9\u5728\u529f\u80fd\u4e0a\u5c31\u63a5\u8fd1\u4e8e\u4eba\u7c7b\u901a\u8fc7\u504f\u65c1\u5224\u65ad\u8bed\u4e49\u7684\u8fc7\u7a0b\u3002<\/p>\n Haslett \u8bbe\u8ba1\u4e86\u4e09\u4e2a\u5b9e\u9a8c\u6765\u9a8c\u8bc1\u8fd9\u4ef6\u4e8b\u3002<\/p>\n \u7b2c\u4e00\u4e2a\u5b9e\u9a8c\u8be2\u95ee GPT-4\u3001GPT-4o \u548c Llama 3\uff1a\u300c\u8336\u300d\u548c\u300c\u830e\u300d\u662f\u5426\u542b\u6709\u76f8\u540c\u7684\u8bed\u4e49\u90e8\u9996<\/strong>\uff1f<\/p>\n \u7b2c\u4e8c\u4e2a\u5b9e\u9a8c\u8ba9\u6a21\u578b\u7ed9\u4e24\u4e2a\u6c49\u5b57\u7684\u8bed\u4e49\u76f8\u4f3c\u5ea6\u8bc4\u5206<\/strong>\u3002<\/p>\n \u7b2c\u4e09\u4e2a\u5b9e\u9a8c\u8ba9\u6a21\u578b\u505a\u300c\u627e\u51fa\u4e0d\u540c\u7c7b\u300d\u7684\u6392\u9664\u4efb\u52a1<\/strong>\u3002<\/p>\n \u6bcf\u4e2a\u5b9e\u9a8c\u90fd\u63a7\u5236\u4e86\u4e24\u4e2a\u53d8\u91cf\uff1a\u4e24\u4e2a\u6c49\u5b57\u662f\u5426\u771f\u7684\u5171\u4eab\u90e8\u9996\u3001\u4e24\u4e2a\u6c49\u5b57\u5728 tokenizer \u4e0b\u662f\u5426\u5171\u4eab\u7b2c\u4e00\u4e2a token\u3002\u8fd9\u4e2a 2×2 \u7684\u8bbe\u8ba1\uff0c\u8ba9\u5979\u80fd\u5206\u79bb\u51fa\u90e8\u9996\u6548\u5e94\u548c token \u6548\u5e94\u5404\u81ea\u7684\u5f71\u54cd\u3002<\/p>\n \u4e09\u4e2a\u5b9e\u9a8c\u7684\u7ed3\u8bba\u4e00\u81f4\uff1a\u5f53\u6c49\u5b57\u88ab\u5207\u6210\u591a\u4e2a token \u65f6<\/strong>\uff08\u6bd4\u5982 GPT-4 \u7684\u65e7 tokenizer \u4e0b\uff0c89% \u7684\u6c49\u5b57\u88ab\u5207\u6210\u4e86\u591a token\uff09\uff0c\u6a21\u578b\u8bc6\u522b\u5171\u4eab\u90e8\u9996\u7684\u51c6\u786e\u7387\u66f4\u9ad8<\/strong>\uff1b\u5f53\u6c49\u5b57\u88ab\u7f16\u7801\u4e3a\u5355\u4e2a token \u65f6<\/strong>\uff08GPT-4o \u7684\u65b0 tokenizer \u4e0b\uff0c\u53ea\u6709 57% \u7684\u6c49\u5b57\u8fd8\u662f\u591a token\uff09\uff0c\u51c6\u786e\u7387\u4e0b\u964d\u4e86<\/strong>\u3002<\/p>\n \u6362\u53e5\u8bdd\u8bf4\uff0c\u4e0a\u4e00\u6bb5\u7684\u90a3\u4e2a\u731c\u60f3\u6210\u7acb\u4e86\u3002\u628a\u6c49\u5b57\u5207\u788e\uff0c\u6210\u672c\u786e\u5b9e\u66f4\u9ad8\uff0c\u4f46\u5207\u788e\u540e\u7684\u5b57\u8282\u5e8f\u5217\u91cc\u4fdd\u7559\u4e86\u90e8\u9996\u7684\u75d5\u8ff9\uff0c\u6a21\u578b\u771f\u7684\u4ece\u4e2d\u5b66\u5230\u4e86\u4e00\u4e9b\u4e1c\u897f<\/strong>\u3002\u800c\u628a\u6c49\u5b57\u7f16\u7801\u4e3a\u6574\u5b57 token\uff0c\u6210\u672c\u964d\u4e0b\u6765\u4e86\uff0c\u4f46\u90e8\u9996\u4fe1\u606f\u88ab\u5c01\u88c5\u5728\u4e00\u4e2a\u4e0d\u900f\u660e\u7684\u7f16\u53f7\u91cc\uff0c\u6a21\u578b\u65e0\u6cd5\u518d\u901a\u8fc7\u5b57\u8282\u5e8f\u5217\u83b7\u53d6\u8fd9\u4e00\u7ebf\u7d22\u3002<\/p>\n \u9700\u8981\u7279\u522b\u8bf4\u660e\u7684\u662f\uff0c\u8fd9\u4e00\u7ed3\u8bba\u4ec5\u5c40\u9650\u4e8e\u5b57\u5f62\u76f8\u5173\u7684\u7ec6\u5206\u8bed\u4e49\u4efb\u52a1\uff0c\u4e0d\u80fd\u7b49\u540c\u4e8e\u6a21\u578b\u6574\u4f53\u7684\u4e2d\u6587\u7406\u89e3\u3001\u903b\u8f91\u63a8\u7406\u3001\u957f\u6587\u672c\u751f\u6210\u80fd\u529b\u4e0b\u964d<\/strong>\u3002\u540c\u65f6\uff0c\u5b9e\u9a8c\u5bf9\u6bd4\u7684 GPT-4 \u4e0e GPT-4o\uff0c\u9664\u4e86\u5206\u8bcd\u5668\u5dee\u5f02\u5916\uff0c\u6a21\u578b\u67b6\u6784\u3001\u8bad\u7ec3\u8bed\u6599\u3001\u53c2\u6570\u91cf\u5747\u6709\u663e\u8457\u53d8\u5316\uff0c\u65e0\u6cd5\u5c06\u51c6\u786e\u7387\u53d8\u5316 100% \u5f52\u56e0\u4e8e\u5206\u8bcd\u7c92\u5ea6\u7684\u8c03\u6574\u3002<\/p>\n \u8fd9\u4e2a\u53d1\u73b0\u8fd8\u5f97\u5230\u4e86\u5de5\u7a0b\u4fa7\u7684\u9a8c\u8bc1\u30022024 \u5e74\u4e00\u9879\u9488\u5bf9 GPT-4o \u7684\u7814\u7a76\u53d1\u73b0\uff0cGPT-4o \u7684\u65b0 tokenizer \u628a\u67d0\u4e9b\u4e2d\u6587\u5b57\u7b26\u7ec4\u5408\u5408\u6210\u4e86\u4e00\u4e2a\u957f token \u4e4b\u540e\uff0c\u6a21\u578b\u53cd\u800c\u51fa\u73b0\u4e86\u7406\u89e3\u9519\u8bef\u3002\u5f53\u7814\u7a76\u8005\u7528\u4e13\u4e1a\u7684\u4e2d\u6587\u5206\u8bcd\u5668\uff0c\u628a\u8fd9\u4e9b\u957f token \u91cd\u65b0\u62c6\u5f00\u518d\u5582\u7ed9\u6a21\u578b\uff0c\u7406\u89e3\u51c6\u786e\u5ea6\u6062\u590d\u4e86\u3002<\/p>\n \u76ee\u524d\u5168\u7403\u5927\u6a21\u578b\u884c\u4e1a\u7684\u4e3b\u6d41\u5171\u8bc6\uff0c\u4f9d\u7136\u662f\u9488\u5bf9\u76ee\u6807\u8bed\u8a00\u4f18\u5316\u7684\u6574\u8bcd \/ \u6574\u5b57\u5206\u8bcd\u5668\uff0c\u80fd\u663e\u8457\u63d0\u5347\u6a21\u578b\u7684\u6574\u4f53\u6027\u80fd<\/strong>\u3002\u6574\u5b57 \/ \u6574\u8bcd\u7f16\u7801\u4e0d\u4ec5\u80fd\u5927\u5e45\u964d\u4f4e token \u6210\u672c\u3001\u63d0\u5347\u4e0a\u4e0b\u6587\u7a97\u53e3\u7684\u6709\u6548\u4fe1\u606f\u91cf\uff0c\u8fd8\u80fd\u7f29\u77ed\u5e8f\u5217\u957f\u5ea6\u3001\u964d\u4f4e\u63a8\u7406\u5ef6\u8fdf\u3001\u63d0\u5347\u957f\u6587\u672c\u5904\u7406\u7684\u7a33\u5b9a\u6027\u3002\u8bba\u6587\u4e2d\u53d1\u73b0\u7684\u7ec6\u5206\u4efb\u52a1\u4f18\u52bf\uff0c\u65e0\u6cd5\u8986\u76d6\u7edd\u5927\u591a\u6570\u4e2d\u6587 NLP \u573a\u666f\u7684\u6027\u80fd\u6536\u76ca\u3002<\/p>\n \u4f46\u8fd9\u4ef6\u4e8b\u4f9d\u7136\u6233\u4e2d\u4e86\u5927\u578b\u7cfb\u7edf\u91cc\u6700\u96be\u5904\u7406\u7684\u4e00\u7c7b\u95ee\u9898\uff1a\u4f60\u80fd\u4f18\u5316\u4f60\u8bbe\u8ba1\u8fc7\u7684\u90e8\u5206\uff0c\u4f46\u4f60\u6ca1\u6cd5\u4f18\u5316\u4f60\u4e0d\u77e5\u9053\u81ea\u5df1\u62e5\u6709\u7684\u90e8\u5206\u3002<\/strong>Unicode \u8054\u76df\u6309\u90e8\u9996\u6392\u5217\u7f16\u7801\uff0c\u662f\u4e3a\u4e86\u4eba\u7c7b\u68c0\u7d22\u7684\u65b9\u4fbf\u3002BPE \u628a\u6c49\u5b57\u62c6\u6210\u5b57\u8282\uff0c\u662f\u56e0\u4e3a\u4e2d\u6587\u5728\u8bed\u6599\u91cc\u7684\u9891\u7387\u592a\u4f4e\u3002\u4e24\u4e2a\u4e0d\u76f8\u5173\u7684\u5de5\u7a0b\u51b3\u7b56\u78b0\u5de7\u53e0\u5728\u4e00\u8d77\uff0c\u4ea7\u751f\u4e86\u4e00\u6761\u8c01\u90fd\u6ca1\u89c4\u5212\u8fc7\u7684\u8bed\u4e49\u901a\u9053\u3002<\/p>\n \u7136\u540e\uff0c\u5f53\u65b0\u4e00\u4ee3\u5de5\u7a0b\u5e08\u300c\u6539\u8fdb\u300dtokenizer\u3001\u628a\u6c49\u5b57\u5408\u5e76\u4e3a\u6574\u5b57 token \u7684\u65f6\u5019\uff0c\u4ed6\u4eec\u540c\u65f6\u62b9\u6389\u4e86\u4e00\u6761\u81ea\u5df1\u4e0d\u77e5\u9053\u5b58\u5728\u7684\u8def\u3002\u6548\u7387\u63d0\u5347\u4e86\uff0c\u6210\u672c\u964d\u4f4e\u4e86\uff0c\u67d0\u4e9b\u4e1c\u897f\u4e5f\u5b89\u9759\u5730\u6d88\u5931\u4e86\uff0c\u800c\u4f60\u751a\u81f3\u4e0d\u4f1a\u6536\u5230\u4e00\u6761\u62a5\u9519\u4fe1\u606f\u3002<\/p>\n \u6240\u4ee5\u4e8b\u60c5\u6bd4\u300c\u4e2d\u6587\u5728 AI \u91cc\u591a\u4ed8\u94b1\u300d\u8fd9\u4e2a\u5224\u65ad\u66f4\u590d\u6742\u3002\u6bcf\u4e00\u79cd tokenizer \u90fd\u5728\u4e3a\u67d0\u4e2a\u9ed8\u8ba4\u503c\u4f18\u5316\uff0c\u4ee3\u4ef7\u85cf\u5728\u4e86\u522b\u5904<\/strong>\u3002<\/p>\n 05<\/strong><\/p>\n \u6797\u8bed\u5802<\/strong><\/p>\n \u4e2d\u6587\u9002\u914d\u897f\u65b9\u6280\u672f\u57fa\u7840\u8bbe\u65bd\u7684\u4ee3\u4ef7\uff0c\u4e0d\u662f AI \u65f6\u4ee3\u624d\u5f00\u59cb\u4ed8\u7684\u3002<\/p>\n 2025 \u5e74 1 \u6708\uff0c\u7ebd\u7ea6<\/a>\u5c45\u6c11 Nelson Felix \u5728 Facebook<\/a> \u4e00\u4e2a\u6253\u5b57\u673a\u7231\u597d\u8005\u5c0f\u7ec4\u91cc\u53d1\u4e86\u51e0\u5f20\u7167\u7247\u3002\u4ed6\u5728\u59bb\u5b50\u7956\u7236\u7684\u9057\u7269\u91cc\u53d1\u73b0\u4e86\u4e00\u53f0\u523b\u6ee1\u4e2d\u6587\u7684\u6253\u5b57\u673a\uff0c\u4e0d\u77e5\u9053\u662f\u4ec0\u4e48\u6765\u5386\u3002\u5f88\u5feb\u6570\u767e\u6761\u8bc4\u8bba\u6d8c\u5165\u3002<\/p>\n Nelson Felix \u7684\u95ee\u9898\uff1a\u660e\u5feb\u6253\u5b57\u673a\u503c\u94b1\u5417\uff1f\uff5c\u56fe\u7247\u6765\u6e90\uff1aFacebook<\/p>\n \u65af\u5766\u798f\u5927\u5b66\u6c49\u5b66\u5bb6\u58a8\u78ca\u5b81\uff08Thomas S. Mullaney\uff09\u770b\u5230\u7167\u7247\u540e\u7acb\u523b\u8ba4\u51fa\u6765\u4e86\uff0c\u8fd9\u662f\u6797\u8bed\u5802 1947 \u5e74\u53d1\u660e\u7684\u300c\u660e\u5feb\u6253\u5b57\u673a\u300d\u7684\u552f\u4e00\u539f\u578b\u673a\uff0c\u5931\u8e2a\u4e86\u5c06\u8fd1 80 \u5e74\u3002\u540c\u5e74 4 \u6708\uff0cFelix \u592b\u5987\u5c06\u6253\u5b57\u673a\u5356\u7ed9\u65af\u5766\u798f\u5927\u5b66\u56fe\u4e66\u9986\u3002<\/p>\n \u660e\u5feb\u6253\u5b57\u673a\u8981\u89e3\u51b3\u7684\u95ee\u9898\uff0c\u548c\u4eca\u5929 tokenizer \u9762\u5bf9\u7684\u95ee\u9898\u5728\u7ed3\u6784\u4e0a\u662f\u540c\u4e00\u4e2a\uff1a\u600e\u4e48\u628a\u4e2d\u6587\u9ad8\u6548\u5730\u5d4c\u5165\u4e00\u5957\u4e3a\u897f\u65b9\u8bed\u8a00\u8bbe\u8ba1\u7684\u6280\u672f\u57fa\u7840\u8bbe\u65bd\u3002<\/strong><\/p>\n 1940 \u5e74\u4ee3\u7684\u82f1\u6587\u6253\u5b57\u673a\u6709 26 \u4e2a\u5b57\u6bcd\u952e\uff0c\u4e00\u952e\u4e00\u5b57\uff0c\u7b80\u5355\u76f4\u63a5\u3002\u4e2d\u6587\u6709\u51e0\u5343\u4e2a\u5e38\u7528\u5b57\uff0c\u4e0d\u53ef\u80fd\u4e00\u952e\u4e00\u5b57\u3002\u5f53\u65f6\u7684\u4e2d\u6587\u6253\u5b57\u673a\u662f\u4e00\u4e2a\u5de8\u5927\u7684\u5b57\u76d8\uff0c\u6392\u7740\u51e0\u5343\u4e2a\u94c5\u5b57\uff0c\u6253\u5b57\u5458\u7528\u624b\u9010\u4e2a\u6361\u5b57\uff0c\u6bcf\u5206\u949f\u53ea\u80fd\u6253\u5341\u51e0\u4e2a\u5b57\u3002<\/p>\n![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/316C27FFA8F7485DF9DB54976533EED472466A06_size47_w1080_h484.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/0D1DE200EFAAFF6471CA7359391646B6F8CD478F_size97_w1080_h1295.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/39E3302D3C907D8DC9B7CCBA266A4BC2E63207EE_size107_w1080_h1241.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/2A841F1E5079A3BB9D7386A4A93589CDB2B48555_size125_w1080_h1658.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/E958333F449237AC297EF48402868D14460CAB66_size26_w1080_h504.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/1CE74ABA7413978A6CC47B132FB46FE835CA93DC_size36_w1080_h944.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/43BF2C2CFAF564942F20EA096DE71D1A3413AF21_size22_w1080_h450.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/5085C62DFD2AEAED8CF18E690DD845745210974A_size27_w1080_h479.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/0B56720F2EDB49BF96E8654825CBB427EB7B1635_size71_w1080_h670.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/8976910A87DB75230C0ED42C3E32C316352642B9_size11_w400_h300.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/5C719E70FEB07BD32C822AF1BA77679C4876BFE2_size46_w496_h968.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/914E6F16C72C969566AFE08C408FC6547D154A09_size167_w1080_h1032.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/AB141A2375516417837A7CEBE0A89372D3A7934C_size120_w1080_h577.jpg\")
![\"AI\u5927\u6a21\u578b\u7684\u201c\u4e2d\u6587\u7a0e\u201d\uff1a\u4e2d\u6587\u6bd4\u82f1\u6587\u66f4\u8d39Token\uff0c\u4e3a\u4ec0\u4e48\uff1f\"](\"https:\/\/x0.ifengimg.com\/res\/2026\/FCFED529648965A1410E87D3AB1230B97C2C6CDF_size165_w1080_h720.jpg\")
<\/p>\n