Masked diffusion models (MDMs) generate text by unmasking several tokens per step, but they are trained and sampled under different conditions. The model is …
机构:Apple
来源:arXiv 2609.37974 | AI4Papers 论文推荐平台