Research map / Video / Video-language understanding
AccKV: Towards Efficient Audio-Video LLMs Inference via Adaptive-Focusing and Cross-Calibration KV Cache Optimization
Zhonghua Jiang, Kui Chen, Kunxi Li, Keting Yin, Yiyun Zhou, Zhaode Wang, Chengfei Lv, Shengyu Zhang
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 Cross-Calibration (intra-modal merge plus cross-modal alignment) (ablation) | AccKV · AVSD (VideoLLaMA2, 20% budget) | ROUGE: 0.364630 -> 0.425179 | Improved |
| add Adaptive-Focusing (ablation) | AccKV · AVSD (VideoLLaMA2, 20% budget) | ROUGE: 0.342471 -> 0.364630 | Improved |
| add Attention Redistribution (ablation) | AccKV · AVSD (VideoLLaMA2, 20% budget) | ROUGE: 0.334883 -> 0.342471 | Improved |
| evict video KV in high layers instead of audio KV (ablation) | VideoLLaMA2 · Audio tasks | accuracy: 0.691272 -> 0.523281 | Hurt |
| merge unimportant audio and video KV across modalities instead of evicting (ablation) | VideoLLaMA2 · Audio tasks | accuracy: 0.651984 -> 0.166571 | Hurt |
| evict all audio KV cache (ablation) | VideoLLaMA2 · Audio tasks | accuracy: 0.700490 -> 0.018601 | Hurt |
| evict all video KV cache (ablation) | VideoLLaMA2 · Video tasks | accuracy: 0.541667 -> 0.09015 | Hurt |
| lower cache budget from 20% to 10% under AccKV (tau=0.9) | AccKV · AVSD (VideoLLaMA2) | ROUGE: 0.425179 -> 0.390516 | Hurt |
| use AccKV instead of H2O at 10% budget | VideoLLaMA2 · AVSD (10% budget) | ROUGE: 0.332905 -> 0.390516 | Improved |
| use AccKV instead of SnapKV at 20% budget | VideoLLaMA2 · AVSD (20% budget) | ROUGE: 0.424881 -> 0.425179 | No clear effect |
| lower threshold tau from 0.9 to 0.6 (ablation) | AccKV · AVSD (VideoLLaMA2, 20% budget) | ROUGE: 0.425179 -> 0.423846 | No clear effect |
| use AccKV instead of FastV at 120 token budget | AVicuna · AVSD (120 tokens) | ROUGE: 0.213519 -> 0.315302 | Improved |
Closest papers by meaning
- SmallKV: Small Model Assisted Compensation of KV Cache Compression for Efficient LLM Inference (NeurIPS 2025 · similarity 0.89)
- CateKV: On Sequential Consistency for Long-Context LLM Inference Acceleration (ICML 2025 · similarity 0.89)
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.89)
- Aligned Better, Listen Better for Audio-Visual Large Language Models (ICLR 2025 · similarity 0.88)
- SnapKV: LLM Knows What You are Looking for Before Generation (NeurIPS 2024 · similarity 0.88)
- SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget (ICLR 2025 · similarity 0.88)
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.
