Shouldn't Jev-like models be built on top of diffusion models ? like GSAI-ML/iLLaDA-8B-Instruct ?

That showed us the best results at least