AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

Skip-Vision: Efficient and Scalable Acceleration of Vision-Language Models via Adaptive Token Skipping

Weili Zeng, Ziyuan Huang, Kaixiang Ji, Yichao Yan

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning · 9 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add Skip-Vision (skip FFN and KV-cache) to LLaVALLaVA · LLaVA settinginference latency (ms): 125 -> 90Improved
add Skip-Vision to LLaVA-HDLLaVA-HD · LLaVA-HD settingtraining hours: 19 -> 12.2Improved
add Skip-Vision to CoSCoS · CoS settingtraining hours: 15.3 -> 9.8Improved
add Skip-Vision to CoSCoS · CoS settingAvg score: 46.0 -> 46.0No clear effect
increase retained/skipped tokens from 100/156 to 256/320 (ablation)SV-LLaVA · LLaVA settingAvg score: 41.3 -> 42.6Improved
scale SFT data from LLaVA-665K to SV-1MSV-CoS · CoS settingAvg score: 46.0 -> 49.3Improved
add skip KV-cache at inference (ablation)SV-LLaVA · LLaVA settinginference latency (ms): 103 -> 90Improved
add skip KV-cache at inference (ablation)SV-LLaVA-CoS · CoS settingMMVet: 40.2 -> 41.7Improved
add skip KV-cache at inference (ablation)SV-LLaVA · LLaVA settingTextVQA: 56.7 -> 56.5Hurt
skip middle and small window tokens in KV-cache instead of small only (ablation)SV-CoS · CoS settingOCRBench: 37.0 -> 22.6Hurt
add local token merge to CoS baseline without skip FFN (ablation)CoS-1296 · CoS settingOverall score: 46 -> 45.6Hurt
add local token merge within skip-vision (ablation)SV-CoS · CoS settingOverall score: 45.0 -> 45.3Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy