EXODUS: An Explicit Outcome-Driven Uncertainty Screening Controller for Diffusion Language Model Sampling

Abstract

Masked diffusion language models can revise committed tokens to correct errors at the cost of additional computation. Because recovery benefits vary across tokens and denoising steps, fixed policies can spend this computation on unhelpful revisions. We present EXplicit Outcome-Driven Uncertainty Screening (EXODUS), a method that selects recovery requests using commitment confidence and denoising progress. EXODUS rejects high-confidence or late requests using existing sampler information while preserving scheduled denoising. We derive sufficient conditions for both decisions by relating the expected benefit of recovery to its computational cost. Experiments with LLaDA and Dream on GSM8K, HumanEval, and MBPP compare EXODUS against four baselines. The results show that EXODUS provides competitive quality--latency tradeoffs while reducing the latency and forward-pass cost of optional recovery.