The Lattice of Transition Laws
The paper treats diffusion, autoregression, and their hybrids as schedules on the same corruption lattice.
It defines a schedule’s cost as the dependence lost when steps are run in parallel. For graph-Markov data, the minimum zero-cost step count is tied to treedepth: logarithmic for sequences and linear in a grid’s side length. When schedules use fewer steps, the authors say their ranking can be predicted before decoding from pairwise dependence estimated in pretrained weights. They report checks across text, image, and video generation. HF Daily Papers' note
It defines a schedule’s cost as the dependence lost when steps are run in parallel. For graph-Markov data, the minimum zero-cost step count is tied to treedepth: logarithmic for sequences and linear in a grid’s side length. When schedules use fewer steps, the authors say their ranking can be predicted before decoding from pairwise dependence estimated in pretrained weights. They report checks across text, image, and video generation. HF Daily Papers' note
score 5