Yannic Kilcher: TiDAR Paper Analysis — Diffusion Drafting + AR Sampling Hybrid Hits 5.91x Speedup
Yannic Kilcher·medium signal
Yannic Kilcher reviews the TiDAR architecture (Think in Diffusion, Talk in Autoregression), a hybrid approach that drafts tokens using diffusion models in parallel, then samples final outputs autoregressively — all in a single forward pass via structured attention masks. TiDAR achieves 4.71x-5.91x wall-clock throughput boost over standard AR with PyTorch Flex Attention, and matches or closes the gap to AR model quality. Kilcher calls it a promising direction for breaking the sequential decoding bottleneck.