Decoupling the Benefits of Subword Tokenization for Language Model Training via Byte-level Simulation Paper • 2604.27263 • Published May 14 • 11