Why one paper uses "tokenizer encoder": the term encoder is overloaded
BaadeAlan · x · 2026-10-21
In the ongoing terminology thread, researcher Alan Baade responds that he has no allegiance to "tokenizer," but "encoder" alone is too overloaded: encoders in enc-dec Transformers and in MAE/MAR are different things, and the latter need not have modelable or predictable outputs. "Tokenizer encoder" is used to specify the space the model generates in.
Related event: Researchers debate whether encoders should be called tokenizers(3 posts)→