Block Diffusion

Overview June 12, 2026

diffusion block-diffusion systems

SDAR: A Synergistic Diffusion-AutoRegression Paradigm for Scalable Sequence Generation — Cheng et al.
TiDAR: Think in Diffusion, Talk in Autoregression — Liu et al.
Introspective Diffusion Language Models — Yu et al.
Fast-dLLM v2: Efficient Block-Diffusion LLM — Wu et al.
DiffusionBlocks: Block-wise Neural Network Training via Diffusion Interpretation — Shing et al.

Overview of block diffusion and hybrid diffusion–autoregressive paradigms for scalable sequence generation, drawn from the NAR LMs Reading Group collection.

Discussion points:

How do block-wise decoding strategies trade off throughput and quality?
What is shared across SDAR, TiDAR, and Fast-dLLM v2 — and where do they diverge?
Can introspective consistency close the gap with autoregressive models?
How does DiffusionBlocks relate to training vs. inference-time block structure?