We investigate the reasons masking diffusion dominates discrete diffusion.

preference fading discrete diffusion tailored for recommendation via modeling preference ratios

Rectified Flow for Discrete Flow-based Models

Absorbing state discrete diffusion models that enable remasking benefit from inference-time scaling, improving sample quality, controlled generation, and performance on downstream tasks.

We propose an informed corrector for masked discrete diffusion that reduces approximation errors, enabling faster sampling and better sample quality in both synthetic and large-scale settings.

Substantially faster diffusion LLMs using a small auxiliary autoregressive model

We show that diffusion language models are a lot more sample-efficient than standard autoregressive language models, due to their ability to learn from different token orderings.

We propose the Anchored Diffusion Language Model (ADLM), a novel two-stage framework that generates an important token mixture which guides the prediction of missing likelihoods, resulting in better likelihood modeling and generated text quality.