Stop the Flip-Flop: Context-Preserving Verification for Fast Revocable Diffusion Decoding

COVER (Cache Override Verification for Efficient Revision) reduces unnecessary token revisions in diffusion language models. It preserves context while combining token verification and parallel drafting in a single forward pass, achieving up to 11.64× decoding speedup while preserving generation quality.

Recommended citation: Yanzheng Xiang, Lan Wei, Yizhen Yao, Qinglin Zhu, Hanqi Yan, Chen Jin, Philip Alexander Teare, Dandan Zhang, Lin Gui, Amrutha Saseendran, Yulan He. 2026. "Stop the Flip-Flop: Context-Preserving Verification for Fast Revocable Diffusion Decoding." In ICML 2026.
Download Paper

Code