40 min
Seminar, Hosted by Simons Institute for the Theory of Computing
Optimal Inference Schedules for Masked Diffusion Models
University of Washington
Recording
Abstract
Masked diffusion language models can reveal tokens out of order and in parallel, potentially reducing sequential inference costs. This talk asks how much parallelism is possible before distributional accuracy degrades. It gives an exact expression for expected divergence under arbitrary data distributions and unmasking schedules, exposing a connection with classical approximation of one-variable functions.
Topics
- Masked diffusion language models, broader topic: diffusion language models
- sequential inference costs, search World Wide
- distributional accuracy, search World Wide
- expected divergence, broader topic: divergence
- unmasking schedules, search World Wide
- arbitrary data distributions, broader topic: data distributions
- classical approximation, search World Wide
- one-variable functions, search World Wide