TiDAR: Hybrid Diffusion+Autoregressive LM Achieves 4.71–5.91x Speed vs Pure AR at Same Quality

arXiv·medium signalXBlueskyLinkedInCopy link

arXiv:2511.08923 proposes TiDAR — Think in Diffusion, Talk in Autoregression — a hybrid language model that drafts tokens in fully parallelized diffusion (the 'thinking' phase) then outputs them autoregressively (the 'talking' phase), all within a single forward pass using structured attention masks. The draft model is the base model itself, preserving capacity without a separate smaller model. Yannic Kilcher's 18K-view analysis highlights this as the first architecture to close the quality gap with pure AR while delivering 4.71x–5.91x more tokens per second.