ReTok: Replacing Tokenizer to Enhance Representation Efficiency in Large Language Model