Minixhofer et al. introduce byteification, retrofitting token-based LLMs to run at the byte level so they can count individual letters in words

The work introduces byteification, which retrofits LLMs that process text as tokens (sequences of letters) so that they operate at the byte level, where bytes are the binary sequences encoding individual characters. Token-based LLMs can perform excellently but cannot access the individual characters within each word; byteification restores character-level access in the retrofitted model. A Nature News & Views article describes the Nature paper by Minixhofer et al., stating that the authors show byteified models achieve competitive performance while retaining the ability to read individual characters; this summary-level page provides no specific benchmarks, datasets, or numbers.

The retrofitted models can handle tasks that require character-level information, such as counting how many times a letter appears in a word. This directly addresses the characteristic failure of token-based models on character-counting questions, illustrated by many LLMs answering that 'strawberry' contains the letter 'r' twice when it appears three times. The claim comes from the News & Views article's account of the original paper's results, reported as the authors' finding without experimental scale or control details.

The significance of byteification lies in bringing byte-level operation to existing token-based base models rather than training byte-level models from scratch. It offers a retrofit path to character-level capability on top of existing LLMs while preserving their performance level. The article describes the method as retrofitting token-based LLMs to operate at the byte level and states that byteified models can achieve competitive performance; the exact retrofit procedure and evaluation conditions require the original paper.

The result targets language tasks that need character-level information, such as counting letter occurrences in a word; it applies to existing token-based base LLMs that are retrofitted via byteification to operate at the byte level. For developers and researchers who want to add character-level capability without giving up the performance of an existing model, this offers a reference direction.

This is a summary-level page and does not include the original paper's benchmark names, datasets, model scales, or specific numbers, so it is not possible to judge under which tasks and controls the 'competitive performance' holds. The cost of byteification, its effect on training and inference efficiency, and how character-level capability behaves across languages and character sets remain open questions for the original paper. The page also does not state whether the finding has been validated across multiple model families.