NAVIRA Decouples Token Detection and Regeneration in Masked Diffusion LMs

NAVIRA splits masked diffusion inference into two forward passes: one scores token quality, a second regenerates from the cleaned context, preventing erroneous tokens from conditioning their own replacements. Temperature-controlled stochastic remasking reduces repeated corrections and improves fluency-diversity tradeoff over PRISM.