Zhijie Deng explains methods to speed up diffusion language models
Overview
Shanghai Jiao Tong University associate professor Zhijie Deng presented methods for accelerating diffusion language models without losing output quality at CNCC 2026, according to Leiphone (雷峰网).
He covered approaches from the base LLaDA and Dream models through cross-block parallel generation, post-training, and inference-system optimization.
Deng measures efficiency by tokens committed per forward pass and speed in tokens per second, alongside answer and code-pass accuracy. The report does not give measured speedups or accuracy figures for any specific method.
Written by AI from the articles below · updated Nov 5, 2:50 AM ET
Check the sources:
Article timeline
The articles in this story. Times are ET.
- LeiphoneNewsZhijie Deng explains how to speed up diffusion language models without losing quality
AIShanghai Jiao Tong University associate professor Zhijie Deng explained at CNCC 2026 how diffusion language models can generate faster while keeping output quality. He covered methods from LLaDA and Dream through cross-block parallel generation, post-training, and inference-system optimization. He measures efficiency by tokens committed per forward pass and speed in tokens per second, alongside answer and code-pass accuracy.
Heat Trend
Not enough continuous observations to show a trend yet.