Abstract Language models can greatly benefit from efficient tokenization. However, they still
mostly utilize the classical Byte-Pair Encoding (BPE) algorithm, a simple and reliable …