Learning Permutation Distributions via Reflected Diffusion on Ranks
Researchers have developed a new diffusion model called Soft-Rank Diffusion for learning probability distributions on permutations. This method improves upon existing techniques by using a soft-rank forward process, which relaxes discrete ranks into continuous representations for smoother trajectories. The model also incorporates contextualized generalized Plackett-Luce denoisers for enhanced expressivity. Experiments demonstrate that Soft-Rank Diffusion outperforms previous diffusion baselines, especially in sequential tasks and with longer sequences. AI
IMPACT Introduces a novel diffusion model that could improve performance on permutation-based tasks in machine learning.