New Diffusion Language Models have finally bridged the gap: they are now as fast as parallel generation and as smart as ChatGPT.
April 15, 2026
Original Paper
Introspective Diffusion Language Models
arXiv · 2604.11035
The Takeaway
I-DLM is a Diffusion Language Model that matches the quality of autoregressive (AR) models while maintaining the speed of parallel generation. It outperformed prior models by over 26 points on AIME-24, a massive leap. Traditionally, we had to choose between 'slow and smart' (AR) or 'fast and dumb' (Diffusion). This claims to bridge that gap, offering the 'holy grail' of LLMs: high-quality output without the token-by-token bottleneck. For developers, this means the potential for 10x or 20x faster inference without sacrificing the reasoning capabilities we've come to expect.
From the abstract
Diffusion language models promise parallel generation, yet still lag behind autoregressive (AR) models in quality. We stem this gap to a failure of introspective consistency: AR models agree with their own generations, while DLMs often do not. We define the introspective acceptance rate, which measures whether a model accepts its previously generated tokens. This reveals why AR training has a structural advantage: causal masking and logit shifting implicitly enforce introspective consistency. Mo