Research map / Vision-language and multimodal / Multimodal LLMs
AirCache: Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference
Kai Huang, Hao Zou, Bochen Wang, Ye Xi, Zhen Xie, Hao Wang
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 elite observation window instead of continuous 32-token window (ablation) | AirCache · ChatQA (LLaVA-OV-7B, 1% retained) | score: 72.9 -> 76.4 | Improved |
| add elite observation window instead of all text tokens (ablation) | AirCache · ChatQA (LLaVA-OV-7B, 1% retained) | score: 72.2 -> 76.4 | Improved |
| add elite observation window instead of continuous 16-token window (ablation) | AirCache · ChatQA (LLaVA-OV-7B, 1% retained) | score: 70.4 -> 76.4 | Improved |
| use text tokens instead of visual-token window (ablation) | AirCache · ChatQA (LLaVA-OV-7B, 1% retained) | score: 68.8 -> 76.4 | Improved |
| add strength and skewness budget instead of average allocation (ablation) | AirCache · ChatQA (LLaVA-OV-7B, 1% retained) | score: 72.2 -> 76.4 | Improved |
| add strength and skewness budget instead of PyramidKV allocation (ablation) | AirCache · ChatQA (LLaVA-OV-7B, 1% retained) | score: 69.6 -> 76.4 | Improved |
| add skewness to strength-only budget (ablation) | AirCache · ChatQA (LLaVA-OV-7B, 1% retained) | score: 74.2 -> 76.4 | Improved |
| add strength to skewness-only budget (ablation) | AirCache · ChatQA (LLaVA-OV-7B, 1% retained) | score: 74.7 -> 76.4 | Improved |
| compress only visual tokens instead of unified compression (ablation) | H2O · ChatQA (LLaVA-OV-7B, 1% retained) | score: 27.8 -> 71.0 | Improved |
| compress only visual tokens instead of unified compression (ablation) | AirCache · ChatQA (LLaVA-OV-7B, 1% retained) | score: 30.1 -> 76.4 | Improved |
| AirCache vs SnapKV at 1% retained | SnapKV · DocVQA (LLaVA-OV-7B, 1% retained) | score: 64.1 -> 73.2 | Improved |
| AirCache KV compression vs FastV token pruning | FastV · ChatQA (LLaVA-OV-7B, 1% retained) | score: 16.9 -> 76.4 | Improved |
Closest papers by meaning
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.95)
- Mixing Importance with Diversity: Joint Optimization for KV Cache Compression in Large Vision-Language Models (ICLR 2026 · similarity 0.93)
- Q Cache: Visual Attention Is Valuable in Less than Half of Decode Layers for Multimodal Large Language Model (AAAI 2026 · similarity 0.93)
- PrefixKV: Adaptive Prefix KV Cache is What Vision Instruction-Following Models Need for Efficient Generation (NeurIPS 2025 · similarity 0.91)
- TGV-KV: Text-Grounded KV Eviction for Vision-Language Models (ICML 2026 · similarity 0.91)
- MHA2MLA-VLM: Enabling DeepSeek’s Economical Multi-Head Latent Attention Across Vision-Language Models (AAAI 2026 · similarity 0.91)
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.
