cs.LGNov 14, 2026

DynaDFlash: Dynamic Feature Projection for Block Diffusion Speculators

Authors: Aadi Kulshrestha, Reese Chong

Abstract

DFlash trains a parallel draft model on hidden features from its target model. Inspired by the previous work of DFlare that introduced per layer combinations of target hidden states, we test whether context dependent weighting of these states helps improve acceptance length of block diffusion speculators. We trained both a DFlash baseline for the Qwen 3 of block size 16, and a dynamic model that computes gates for each layer to affect the magnitude of hidden features before projection. While mean acceptance length between the dynamic model and the DFlash baseline didn't significantly differ, we notice that the dynamic model excels for drafting in the context of math. In an evaluation of 700 prompts from the MetaMathQA and Orca-Math datasets, the mean accepted length improves by 5%. We also found this lead to a 4.1% higher throughput for a subset of the same dataset in the context of speculative decoding.

Figures & tables

Explore similar work

CardsList
  1. DFlare: Scaling Up Draft Capacity for Block Diffusion Speculative Decoding

    Jun 1, 2026Jiebin Zhang, Zhenghan Yu, Song Liu +9Speculative DecodingLLM Inference Acceleration

  2. Accelerating Speculative Decoding with Block Diffusion Draft Trees

    Apr 14, 2026Liran Ringel, Yaniv RomanoSpeculative DecodingParallel Speculative Decoding