A new study shows LLMs represent different data types based on their underlying meaning and reason about data in their dominant language.