Diffusion Language Models — systematic benchmark of eight architectures
arXiv paper compares eight state-of-the-art diffusion language models (DLMs) side-by-side under controlled conditions.
DLMs generate text by iterative denoising rather than next-token prediction, enabling parallel refinement of full sequences — but prior work used incompatible evaluation protocols, datasets, and inference budgets, making direct comparison impossible.
This systematic analysis isolates trade-offs between speed, quality, and compute across the landscape.