SuperBPE

A "superword" tokenizer that lets BPE merge across whitespace, producing tokens that span multiple words (COLM 2025). SuperBPE improves average performance by 4.0% across 30 tasks (+8.2% MMLU) while using 27% less inference compute at fixed model size. Clean UW attribution — Alisa Liu, Jonathan Hayase, Sewoong Oh, Noah A. Smith, and Yejin Choi.