Researchers improve diffusion language models by using temperature-based sampling to boost both diversity and speed without sacrificing quality.

New approach uses tempered, softened versions of confidence-based remasking heuristics to increase sample diversity in diffusion language models

Method maintains computational efficiency while improving pass@k scores, closing the exploration gap between confidence-based and autoregressive sampling methods

Introduces formal model of fork tokens to analyze how remasking affects expected entropy and sample variety

Achieves better performance than existing approaches when controlling for computational cost (pass@NFE metric)