Nature · 7 October 2026 · From issue 48
Retrofitting language models to operate over bytes
Converted letter-by-letter AI models now match word-chunk models in performance.
Benjamin Minixhofer, Tyler Murray et al.
AI language models usually break text into word chunks before reading it. This hides small details, which causes problems when every letter matters, like in computer code or biological sequences. A new method changes existing word-chunk models into models that read text one letter at a time. These new models catch up to the older ones and do better on tasks that need letter-by-letter attention.
Shows letter-level models can match word-chunk models while reading text more precisely.
Method paper, no reported test scores or sample sizes yet.
Integrity screen: passed (3 checks) Checked 20 October 2026. Retraction record: none. DOI resolves at doi.org. Metadata record found (Nature). Read the source
This screen checks process signals. It does not tell you the findings are true.