GIDD

Researchers propose UNIFUSION, a continual pre-training approach that adapts pretrained autoregressive language models (e.g., GPT-2) to uniform-noise discrete diffusion under a unified reverse-rate ob…