Byteification retrofits LLMs to count individual letters reliably

Nature reported a technique called byteification that retrofits existing large language models to evaluate text at the individual-letter level, a task most models currently fail, such as counting the i's in "artificial intelligence." The method requires no retraining and targets a known tokenization blind spot.

• The method retrofits existing LLMs rather than building a new model
• Models fail letter-level tasks because of tokenization
• It is unclear whether gains hold on broader benchmarks