The llm has absolutely no way of knowing which characters are in which token.