Tag
DyStruct is a training-free Bayesian decoding framework for discrete Diffusion Language Models that enables flexible-length generation by dynamically determining expansion size and decoding order, improving accuracy on math and code tasks.