CORA-Diff: Training-Free Acceleration for Diffusion Language Models

Researchers have introduced CORA-Diff, a training-free method to accelerate inference in diffusion language models (DLMs). DLMs update multiple tokens in parallel but often use a fixed denoising horizon, causing unnecessary computation when predictions stabilize early. CORA-Diff leverages native trajectory signals to identify residual positions likely to match the deterministic dense endpoint, applying confidence-and-persistence gating only to unresolved positions. This method requires no backbone changes, learned acceptance models, or logit modifications, preserving the original transfer rule. The approach is grounded in theory explaining why high-confidence, persistent predictions are more likely to be correct. The paper is available on arXiv under identifier 2608.11235.