改造后的 LLM 能数清 "artificial intelligence" 里有几个字母 "i"

• NEWS AND VIEWS
• 07 October 2026

Most LLMs cannot reliably evaluate text on the level of individual letters. A technique called byteification retrofits existing models to enable it.

By

• Zhao Zhang⁰ &
• Yingfei Xiong¹

1. Zhao Zhang

   1. Zhao Zhang is in the Key Lab of High Confidence Software Technologies, School of Computer Science, Peking University, Beijing 100871, China.

2. Yingfei Xiong

   1. Yingfei Xiong is in the Key Lab of High Confidence Software Technologies, School of Computer Science, Peking University, Beijing 100871, China.

Strawberry contains the letter ‘r’ three times, but when asked, many large language models (LLMs) answer that the letter appears twice. This happens because most LLMs encode words as ‘tokens’ that represent sequences of letters. LLMs that operate in this way can achieve excellent performance, but they cannot access the individual characters in each word, which are encoded as binary sequences called bytes. Writing in Nature, Minixhofer et al.¹ now report an approach called byteification that retrofits token-based LLMs to operate at the byte level. The authors show that byteified models can achieve competitive performance while retaining the ability to read individual characters.

References

1. Minixhofer, B. et al. Nature https://doi.org/10.1038/s41586-026-11111-4 (2026).

   Article  Google Scholar 

2. Sennrich, R., Haddow, B. & Birch, A. in Proc. 54th Ann. Meet. Assoc. Comput. Linguist. 1715–1725 (2016).

3. Cosma, A., Ruseti, S., Radoi, E. & Dascalu, M. in Proc. 2025 Conf. Empir. Meth. Nat. Lang. Proc. 28252–28263 (2025).

4. Pagnoni, A. et al. Proc. 63rd Ann. Meet. Assoc. Comput. Linguist. 9238–9258 (2025).

Download references

Competing Interests

The authors declare no competing interests.