MMFlash: Faster Speculative Decoding for Vision-Language Models
MMFlash improves speculative decoding for vision-language models by focusing drafter training on visually dependent tokens and using sparse attention for long videos.
ls ./papers | grep "#speculative decoding"
Trending AI research papers from Hugging Face, summarized in plain language: what each paper does, why it matters and where to read more.
6 papers · #speculative decoding
MMFlash improves speculative decoding for vision-language models by focusing drafter training on visually dependent tokens and using sparse attention for long videos.
VFlash uses known visual primitive structure to decode trajectories, boxes, and motion plans in parallel, cutting latency while often improving spatial accuracy.
SpecOPD trains speculative decoding drafters on their own sampled histories while parallelizing target supervision, improving acceptance length and training throughput.
SpecPilot uses cost models, coding agents, and token-level checks to optimize speculative decoding across models, GPUs, and serving regimes without quality loss.
Soft-MTP removes a noisy component of standard multi-token prediction, improving model training, speculative decoding, and downstream math and code results without added parameters.
This study shows that lexical alignment between pre-training data and benchmarks, measured by unigram cross-entropy, predicts zero-shot accuracy and guides data selection and prompt routing.