DBLAST Improves Speculative Decoding via Dependent Block Drafting
August 7, 2026
DBLAST addresses performance degradation in block diffusion drafters during non-greedy decoding by modeling token position dependencies. It uses a low-rank latent mixture and an acceptance-oriented training objective to maximize the expected verified draft length.
HOW THIS AFFECTS YOU
●
builderImplementing this could improve inference throughput for stochastic sampling in production environments.
●
researcherThis provides a method to maintain draft efficiency even as target distribution entropy increases.