MMFlash: Faster Speculative Decoding for Vision-Language Models
Three arXiv papers posted in November 2026 address efficiency in speculative decoding. MMFlash (arXiv:2611.18482v1) presents a method for faster speculative decoding in vision-language models, where a lightweight drafter proposes tokens for parallel verification by a target model. VisHead-Spec (arXiv:2611.21839v1) proposes head-routed visual attention for multimodal speculative decoding, noting that speedup diminishes under certain conditions. SpecSearch (arXiv:2611.22296v1) targets hardware-aware efficient speculative decoding, arguing that optimal configuration depends jointly on workload and hardware. All three are new-announcement preprint abstracts and report no deployment or benchmark results in the supplied text.
How this story unfolded
Key points7 items Model summary · ordered by when the evidence was published
- Point 01 11/16, 13:00
MMFlash is posted as arXiv:2611.18482v1 under the new-announcement category.
Based on 1 - Point 02 11/16, 13:00
MMFlash speeds generation by having a lightweight drafter propose tokens for parallel verification by a target model.
Based on 1 - Point 03 11/18, 13:00
VisHead-Spec is posted as arXiv:2611.21839v1 and targets multimodal speculative decoding for VLMs.
Based on 1 - Point 04 11/18, 13:00
VisHead-Spec proposes head-routed visual attention and states that speculative decoding preserves the target model's outputs.
Based on 1 - Point 05 11/18, 13:00
VisHead-Spec notes that the speedup from speculative decoding diminishes.
Based on 1 - Point 06 11/18, 13:00
SpecSearch is posted as arXiv:2611.22296v1 and focuses on hardware-aware efficient speculative decoding.
Based on 1 - Point 07 11/18, 13:00
SpecSearch argues that the optimal speculative decoding configuration depends jointly on workload and hardware.
Based on 1
ReportsBy publish time · newer further down
- November 16, 2026 · Mon
- Start
MMFlash: Faster Speculative Decoding for Vision-Language Models
Original summary arXiv:2611.18482v1 Announce Type: new Abstract: Speculative decoding accelerates generation by letting a lightweight drafter propose tokens for parallel verification by a target mo
arXiv cs.CV Original ↗ - November 18, 2026 · Wed
- +2d 0h
VisHead-Spec: Head-Routed Visual Attention for Multimodal Speculative Decoding
Original summary arXiv:2611.21839v1 Announce Type: new Abstract: Speculative decoding accelerates vision-language models(VLMs) while preserving the target model's outputs, but its speedup diminishe
arXiv cs.LG Original ↗ - +2d 0h
SpecSearch: Towards Hardware-Aware Efficient Speculative Decoding
Original summary arXiv:2611.22296v1 Announce Type: new Abstract: Speculative decoding (SD) accelerates large language model (LLM) inference, but its optimal configuration depends jointly on workloa
arXiv cs.AI Original ↗ - Still developing · later reports will keep extending this axis