MMFlash paper: lossless speculative decoding speedups up to 3.58× for vision-language models
rsasaki0109 · x · 2026-11-16
A new Hugging Face paper page presents MMFlash, a speculative decoding framework for vision-language models.
- Problem: Diffusion drafters like DFlash propose whole token blocks in one forward pass but focus on text-only tasks; multimodal drafting is underexplored.
- Key finding: The authors introduce a token-level visual-dependence metric based on KL divergence and show acceptance rates drop from 46% in the lowest visual-dependence quartile to 22% in the highest, persisting after multimodal training.
- Method: MMFlash adaptively upweights visually dependent tokens that the drafter predicts poorly, and uses sparse attention for long video inputs to cut redundant attention computation.
- Results: Up to 3.58× lossless speedups across multimodal benchmarks, outperforming EAGLE-3 by up to 195.4% and DFlash by 17.1%.
More from Research
- ERNIE-5.5 caught benchleaking: 96% of HumanEval test set memorized, ranking collapses after decontamination — teortaxesTex · 2026-11-16
- Who Judges the LLM Judges? New Paper Proposes Multilingual Probe for Evaluation Metrics — kalyan_kpl · 2026-11-16
- VibePrevisBench: 242-task benchmark tests coding agents on 4D Blender scene generation — Kling-Keye · 2026-11-16
- V-JEPA-LM: counterfactual supervision teaches LMs to read video dynamics with 20x less data — Keuntae Kim · 2026-11-16
- No underwater IQA metric matches human perception: 73 metrics tested, best kappa just 0.23 — DelhiTechnologicalUniversity · 2026-11-16
- New paper gives finite-sample and consistency guarantees for signature kernel ridge regression — chaumian · 2026-11-16