Researchers Convert AI Models to Read Bytes
Converted models keep general abilities while gaining letter-level accuracy using less than 1% of typical training data.
Updated on Nov. 1, 2026 in Language Learning
Researchers converted three openly available AI language models to read text in bytes rather than word chunks. The converted models retained nearly all general abilities and improved letter-level task performance.
Why it matters
Models that split text into word chunks struggle with tasks that depend on individual characters. The conversion approach offers a way to address that limitation without building new models from scratch.
Bolmo averaged 62.5 on general tests against 62.3 for Olmo 3, and scored 78.1 on letter-level tasks against 54.3 for the original model. The conversion used about 49 billion tokens, less than 1% of the training data typically needed to build a comparable model from scratch.
The players
LMU Munich
Ludwig-Maximilians-Universität München is a major German research university where the study's senior author is based.
University of Cambridge
The University of Cambridge is a leading British research university where the study's lead researcher works.
The details
The team kept the core of each finished model and added smaller components that read bytes and group them into chunks. Training ran in two stages: first only the new components were trained to match the original model's behavior, then the whole model was trained.
Timeline
Article describing the byte-reading conversion study published November 1, 2026.
Culture Shift
The conversion extends the byte-level modeling line that began with ByT5, which trained models directly on raw bytes instead of word-piece tokens. Rather than training new byte models from scratch, this approach adapts models already in wide use.
The findings point to converted models that handle single-character tasks better, which may matter for spelling, word games and text-based study tools. Whether these improvements reach consumer products is not stated in the source.
The takeaway
Models that handle individual letters more reliably could make text-based learning tools more accurate for spelling and character-level practice. Improvements of this kind are most useful when general language abilities remain intact, as they did here.
Further reading
Explore more on Language Learning and how writing systems shape how words are processed.
Source note: This article includes information reported by Earth.
