C4: Coordinating Adaptive Sampling and Early Exit in Diffusion Language Models

A recent study published on arXiv (2607.28166v2) presents C4, a technique that synchronizes two acceleration strategies in diffusion language models: early sequence termination and early block exit. These models produce predictions at each denoising stage, often stabilizing the candidate answer before the entire process concludes. Current approaches usually focus on a single aspect, with exit gates based on fixed-region confidence or rules dependent on the schedule. C4 introduces distinct gates for each decision point: the Confidence-Verified Early Exit (CVEE) assesses when to halt the sequence, requiring both confidence and consistent argmax stability over a re-extracted candidate range at each step. Additionally, the Commit-Core-Then-Confirm component addresses block-level early exits. This paper was marked as a revision on arXiv, aiming to enhance efficiency while maintaining answer quality.