Lead author highlights finding on tokenizer merging strategy
2Sep 17 10:06 AM · 9d ago · 1 post · 1 source · development 2 of 2
Tiago Pimentel posted on X that the paper, submitted to EMNLP 2026, shows that BPE's step-by-step merging strategy is what matters, not its compression objective, challenging how tokenizers are typically compared.
x.com
“Standard tokeniser comparisons (eg, BPE ⚔️ UnigramLM) conflate the two. Our #emnlp2026 paper separates tokenisation objective from search procedure and shows BPE's merging strategy is what matters”
Why is BPE so good: the compression objective or the step-by-step merging? 🤔 Standard tokeniser comparisons (eg, BPE ⚔️ UnigramLM) conflate the two. Our #emnlp2026 paper separates tokenisation objective from search procedure and shows BPE's merging strategy is what matters 🧵