“Text diffusion models operate similarly to BERT models by iteratively filling in gaps or "masked" portions of a text, rather than generating one token at a time.”