Reproduce “MMFlash: Faster Speculative Decoding for Vision-Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Sicheng Feng, Zigeng Chen, Song Wang, Gongfan Fang, Xinchao Wang

MMFlash adapts DFlash’s parallel block-diffusion drafter for vision-language models by weighting training toward image-dependent tokens the drafter predicts poorly. It uses sparse attention to reduce the drafter’s context for long videos. On Qwen3.5-9B ChartQA, it achieves 3.58× decoding speedup versus DFlash’s 3.20×, with greedy decoding in its test setup.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.