Study Profiles and Fixes Sampler Overhead in MoE Diffusion Language Model Decoding
A new arXiv paper examines how masked diffusion language models incur per-token sampling and sorting costs during iterative denoising, and proposes a fix for the sampler bottleneck.
Read the full story at arXiv cs.LG ↗
Timeline · 1 report
- 2026-11-18 05:00 · arXiv cs.LG
The Hidden Cost of MoE Diffusion Decoding: Profiling and Fixing the Sampler