Unverified50% confidenceFactExact time
大多数主流分词器（BPE、WordPiece）将数字按字符级别分割，导致数字的位值结构在输入阶段就已被破坏
1
Sources
50%
Confidence
Long-term
Relevance
7/6/2026
First Seen
Sources
补齐大模型三大短板：知识、计算与工具调用
bilibili晴天AI实战7/4/2026
Related Claims
UnverifiedLLM的分词器（如BPE）为自然语言设计，会将数值拆解为多个不相关的token，破坏数字的连续性和大小语义75% similarVerifiedBPE的核心思路是从字符级别出发，反复合并出现频率最高的相邻字符对，直到词汇表达到预设大小72% similarUnverifiedTokenization algorithms like BPE (Byte Pair Encoding) split words into sub-word segments; for example, 'understanding' might be split into 'under', 'stand', and 'ing'68% similarVerified现代大模型普遍采用的字节对编码（BPE）或WordPiece算法本质上是一种数据驱动的形态学切分方法66% similarUnverified主流大模型普遍采用字节对编码（BPE）或其变体（如SentencePiece、tiktoken）进行分词66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/112943API
curl https://kongchang.com/api/v1/knowledge/claims/112943MCP
get_claim(id=112943)