Masked Diffusion Language Models (MDLMs) promise fast, non-autoregressive text generation, but existing samplers reduce to slow autoregressive behaviors by ignoring interactions when unmasking multiple positions in parallel, based on the model’s confidence level. In this paper, we propose a diluted unmask scheduler (DUS). DUS partitions sequence positions into non-adjacent diluted groups in an inference-only, planner-model-free manner and unmasks them in parallel so as to minimize an upper bound on the joint entropy gain at each denoising step. By making the trade-off between the number of network calls and the quality of the generation explicit, DUS recovers most of the performance lost from existing parallel unmasking strategies. On math (GSM8K, MATH500), code (HumanEval, MBPP), and general knowledge benchmarks (BBH, MMLU-Pro), DUS outperforms confidence-based planners without modifying the underlying denoiser, demonstrating the real speed-quality frontier of MDLM.
