Shuchen Xue

New research frames LLMs as masked diffusion models for faster inference

Researchers have proposed a new approach to scaling large language models (LLMs) by treating them as masked diffusion models (MDMs). This formulation decouples the modeling choice from architectural differences, allowin…