DUET introduces dual-sparse inference to cut costs of block diffusion language models
A new paper, DUET, proposes dual-sparse inference for block diffusion language models, reducing the cost of repeated attention computations during parallel token generation.
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-11-09 05:00 · arXiv cs.CL
DUET: Dual-Sparse Inference for Block Diffusion Language Models