Research map / Vision-language and multimodal / Multimodal LLMs
Efficient Inference of Vision Instruction-Following Models with Elastic Cache
Zuyan Liu, Benlin Liu, Jiahui Wang, Yuhao Dong, Guangyi Chen, Yongming Rao, Ranjay Krishna, Jiwen Lu
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 |
|---|---|---|---|
| use fixed-point discard during generation instead of discarding the most recent cache (ablation) | Elastic Cache · LLaVA-Description, budget 0.5 | PPL: 3.93 -> 3.60 | Improved |
| use fixed-point discard instead of frequency policy during generation (ablation) | Elastic Cache · LLaVA-Description, budget 0.5 | PPL: 3.75 -> 3.60 | Improved |
| merge less-important cache into anchors instead of evicting (ablation) | Elastic Cache · LLaVA-Description, budget 0.5 | PPL: 3.68 -> 3.60 | Improved |
| use simple bucket averaging instead of clustering for merging (ablation) | Elastic Cache · LLaVA-Description, budget 0.5 | PPL: 3.61 -> 3.60 | No clear effect |
| use layer-wise importance rather than head-wise or shared (ablation) | Elastic Cache · LLaVA-Description, budget 0.5 | PPL: 3.75 -> 3.60 | Improved |
| use sum of attention scores as importance instead of mean or moving average (ablation) | Elastic Cache · LLaVA-Description, budget 0.5 | PPL: 8.70 -> 3.60 | Improved |
| replace Local cache with Elastic Cache at budget 0.5 | LLaVA-1.5-13B · LLaVA-Description | PPL: 28.72 lower PPL | Improved |
| replace H2O with Elastic Cache at budget 0.5 | LLaVA-1.5-13B · LLaVA-Description | PPL: 4.34 lower PPL | Improved |
| apply Elastic Cache at budget 0.2 versus full cache | LLaVA-1.5-7B · batch 16, 1024+512 tokens | throughput (token/s): 267.7 -> 476.3 | Improved |
| increase pruning ratio to 0.3 (budget 0.3) for Local cache | Local cache · GPT-4V win rate | win rate (%): 46.37 -> 10.10 | Hurt |
Closest papers by meaning
- PrefixKV: Adaptive Prefix KV Cache is What Vision Instruction-Following Models Need for Efficient Generation (NeurIPS 2025 · similarity 0.92)
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.90)
- Mixing Importance with Diversity: Joint Optimization for KV Cache Compression in Large Vision-Language Models (ICLR 2026 · similarity 0.90)
- CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences (ICLR 2025 · similarity 0.89)
- TGV-KV: Text-Grounded KV Eviction for Vision-Language Models (ICML 2026 · similarity 0.89)
- 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.
