MDM-VGB enables efficient test-time scaling for constrained generation by allowing tokens to be remasked during sampling, achieving quadratic complexity while competing methods like best-of-N suffer exponential complexity—making it practical for real-world constraint satisfaction problems.
This paper introduces MDM-VGB, a sampling method for masked diffusion models that improves generation quality at test time by allowing tokens to be strategically unmasked and remasked based on reward signals.