Causal Drafting Mask Impact

Background: Block-diffusion models offer parallelism by operating on blocks, but their interaction with autoregressive verification must be carefully managed for optimal performance.

Question / Future Work: Further investigate the effect of the optional partially causal drafting/caching mask (Equation 4) on the overall hybrid decoding trajectory, specifically assessing whether a more strictly causal drafting mask provides better local AR-guidance at the cost of KV cache utilization or overall speed.