Discrete Diffusion Models: A Unified Framework from Tokenization to Generation
The paper frames discrete diffusion around the design of the state space itself.
It treats tokenization, vocabulary structure, and domain-specific alphabets as core modeling choices, not preprocessing details. From that view, transition-matrix, masking or absorbing-state, and score or ratio-based methods become variants inside one design space. The authors use the framework to compare trade-offs in objectives, inference, scaling, systems work, and evaluation. HF Daily Papers' note
It treats tokenization, vocabulary structure, and domain-specific alphabets as core modeling choices, not preprocessing details. From that view, transition-matrix, masking or absorbing-state, and score or ratio-based methods become variants inside one design space. The authors use the framework to compare trade-offs in objectives, inference, scaling, systems work, and evaluation. HF Daily Papers' note
score 5