This paper digs into what actually makes continuous diffusion language models work. Key insight: self-conditioning mainly fixes a train/inference mismatch. Their new SemiCat approach reaches discrete diffusion-level performance on language tasks.
arxiv.org/abs/2610.13834v1
