xPress Restores Causality to Diffusion Drafters, Lifting Speculative-Decoding Acceptance Length 30% and Throughput 1.3x on Qwen3-8B

arXiv 2608.02438·medium signalXBlueskyLinkedInCopy link

Block-diffusion drafters like dFlash emit a whole block of draft tokens in one forward pass but sample per-position marginals, so tokens are individually likely yet jointly improbable under the target model that verifies them conditionally — causing early rejection. xPress adds a lightweight causal refiner that reconciles the entire diffusion block at once through parallel refinement, propagating dependencies without a token-by-token loop. On Qwen3-8B across seven math, code, and chat benchmarks it raises acceptance length about 30% on average (up to +56%) and end-to-end throughput about 1.3x (up to 1.7x) over the original dFlash drafter.