ResearchNovember 04, 2026
-37
🧊
RECAP Speeds Up Vision Models Without Training
The EchoKV paper describes a self-corrective KV cache eviction method that relies on iterative query feedback.
#Speculative Decoding#Multimodal#LLM Inference#Vision-Language#Optimization
🔥 What happened
Researchers dropped RECAP, a training-free method that recycles a vision-language model's own predictions to speed up speculative decoding. Unlike prior work that indexed by last token only, RECAP uses one-, two-, and three-token keys plus request-local and persistent memories.
💡 Why it matters
It hits 1.66x and 1.56x mean speedups on Qwen2.5-VL-7B and LLaVA-1.6-7B across seven benchmarks—beating the strongest training-free baseline, Token Recycling (1.51x/1.37x). That's a free latency win for multimodal inference, no draft model required.
âš¡ Our take
A clever hack that exposes how much performance is left on the table. If you deploy VLMs, test this immediately—the baseline just got crushed.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions — when in doubt, read the linked original source.
