Most score-based generative models rely on Brownian diffusion to gradually add noise to data, then reverse the process to generate new samples. A new preprint on arXiv explores an alternative driven by alpha-stable Lévy processes, which generalize Brownian motion and allow for heavier-tailed noise. The authors position this approach as a way to improve sampling and to model data that is not well described by Gaussian assumptions.
The paper's title, From Global to Tail Convergence, indicates a focus on how these models behave both in the bulk of the distribution and in the tails. The abstract reports empirical evidence of better sampling and improved heavy-tail modeling, but it stops short of detailing the methods or comparing against specific baselines. Because this is a single source, there are no conflicting findings to weigh; the claims stand only as reported in the preprint.
As an arXiv announcement, the work has not yet gone through peer review. Still, it points to a promising direction for generative models that need to handle extreme values or skewed data, where standard diffusion models often struggle. Future work will need to confirm the extent of these advantages with rigorous benchmarks.