Diffusion LMs do seem to be able to get more out of the same data. In a world where we are already training transformer based LLMs on all text available, diffusion LMs ability to continue learning on a fixed set of data may be able to outperform transformers
