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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add Skip-Vision (skip FFN and KV-cache) to LLaVA | LLaVA · LLaVA setting | inference latency (ms): 125 -> 90 | Improved |
| add Skip-Vision to LLaVA-HD | LLaVA-HD · LLaVA-HD setting | training hours: 19 -> 12.2 | Improved |
| add Skip-Vision to CoS | CoS · CoS setting | training hours: 15.3 -> 9.8 | Improved |
| add Skip-Vision to CoS | CoS · CoS setting | Avg score: 46.0 -> 46.0 | No clear effect |
| increase retained/skipped tokens from 100/156 to 256/320 (ablation) | SV-LLaVA · LLaVA setting | Avg score: 41.3 -> 42.6 | Improved |
| scale SFT data from LLaVA-665K to SV-1M | SV-CoS · CoS setting | Avg score: 46.0 -> 49.3 | Improved |
| add skip KV-cache at inference (ablation) | SV-LLaVA · LLaVA setting | inference latency (ms): 103 -> 90 | Improved |
| add skip KV-cache at inference (ablation) | SV-LLaVA-CoS · CoS setting | MMVet: 40.2 -> 41.7 | Improved |
| add skip KV-cache at inference (ablation) | SV-LLaVA · LLaVA setting | TextVQA: 56.7 -> 56.5 | Hurt |
| skip middle and small window tokens in KV-cache instead of small only (ablation) | SV-CoS · CoS setting | OCRBench: 37.0 -> 22.6 | Hurt |
| add local token merge to CoS baseline without skip FFN (ablation) | CoS-1296 · CoS setting | Overall score: 46 -> 45.6 | Hurt |
| add local token merge within skip-vision (ablation) | SV-CoS · CoS setting | Overall score: 45.0 -> 45.3 | Improved |
Closest papers by meaning
- Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inference (AAAI 2025 · similarity 0.91)
- Variation-aware Vision Token Dropping for Faster Large Vision-Language Models (CVPR 2026 · similarity 0.90)
- One Layer's Trash is Another Layer's Treasure: Adaptive Layer-wise Visual Token Selection in LVLMs (CVPR 2026 · similarity 0.90)
- Skip-It? Theoretical Conditions for Layer Skipping in Vision–Language Models (ICML 2026 · similarity 0.90)
- SwiftVLM: Efficient Vision-Language Model Inference via Cross-Layer Token Bypass (NeurIPS 2026 · similarity 0.90)
- Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves (CVPR 2025 · similarity 0.90)
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.
