Apple researchers propose learned unmasking policies for diffusion LMs

Diffusion large language models (dLLMs) match autoregressive performance while promising greater inference efficiency. This paper explores learned unmasking policies for token selection during sampling, a key design aspect of dLLMs.