RelayGen is a training-free way to switch between a big model and a small model while one answer is being generated.
The paper introduces a new way to sample text from masked diffusion language models that is smarter and less greedy.