TiDAR: Think in Diffusion, Talk in Autoregression — Hybrid Architecture Uses Parallel Diffusion for Reasoning, AR for Output

Yannic Kilcher (YouTube)·medium signal

Yannic Kilcher's analysis of TiDAR (arXiv 2511.08923) covers a hybrid language model architecture where internal reasoning uses parallel diffusion-based generation while output tokens are produced autoregressively. The approach combines the quality advantage of autoregression with the speed potential of diffusion models, enabling faster parallel generation in latent reasoning space without sacrificing output quality. This is a direct architectural response to the CoT token efficiency problem and complements recent TERMINATOR-style exit-signal approaches.