CORA-Diff Accelerates Diffusion Language Model Inference by 3x

Diffusion language models, while powerful, often suffer from inefficient inference due to fixed denoising horizons that repeatedly process already stable predictions. Existing acceleration techniques typically involve learned filters or modified scores. This new research proposes CORA-Diff, a novel approach that leverages native confidence and persistence signals within the model's trajectory to identify tokens that are likely to match the final deterministic output early. CORA-Diff operates without requiring any changes to the model's backbone, learned acceptance models, or logit modifications. It selectively applies confidence and persistence gating only to positions that the standard rule leaves unresolved, accepting stable tokens and using them as context. This method has demonstrated significant speedups, achieving 2.70x to 3.32x faster inference on benchmarks like GSM8K and HumanEval, and up to 13.14x under specific isolation protocols, all while preserving or even enhancing task performance.