Research map / Vision-language and multimodal / Multimodal LLMs
ZipVL: Accelerating Vision-Language Models through Dynamic Token Sparsity
Yefei He, Feng Chen, Jing Liu, Wenqi Shao, Hong Zhou, Kaipeng Zhang, Bohan Zhuang
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 layer-wise adaptive token ratio instead of fixed ratio (ablation) | ZipVL sparse attention · Video-MME (128 frames) | accuracy (%): 51.1 -> 52.4 | Improved |
| use adaptive ratio in mixed-precision KV quantization (ablation) | ZipCache on LLaMA3-8B · GSM8k | accuracy (%): 53.75 -> 54.06 | Improved |
| use adaptive ratio in mixed-precision KV quantization (ablation) | ZipCache on LLaMA3-8B · GSM8k | compression ratio: 4.69x -> 6.18x | Improved |
| replace FastV fixed ratio with ZipVL adaptive ratio (tau=0.975) | LLaVA-Next-13B · ChartQA | accuracy: 51.6 -> 58.5 | Improved |
| replace FastV with ZipVL (tau=0.975) | LLaVA-Next-13B · VQAv2 | accuracy: 76.8 -> 80.4 | Improved |
| replace HiRED 40% with ZipVL (tau=0.975) | LLaVA-Next-7B · ChartQA | accuracy: 46.5 -> 52.6 | Improved |
| raise threshold tau from 0.96 to 0.975 (ablation) | ZipVL on LLaVA-Next-7B · ChartQA | accuracy: 51.0 -> 52.6 | Improved |
| add ZipVL sparse attention to prefill | LLaVA-Next-7B 32K input · TTFT | TTFT (s): 9.35 -> 6.40 | Improved |
| add approximate attention score computation (ablation) | LLaVA-Next-7B 32K input · TTFT | TTFT (s): 9.35 -> 10.95 | Hurt |
| replace FlashAttention with ZipVL at 128K input | LLaVA-Next-13B · prefill latency | TTFT (s): 48.01 -> 20.57 | Improved |
| replace FlashAttention with ZipVL at 16K input | LLaVA-Next-13B · decoding throughput | tokens/s: 14.10 -> 40.87 | Improved |
| replace MInference with ZipVL on 128-frame video | LongVA-7B · Video-MME | attention FLOPs reduction (%): 77.1 -> 82.3 | Improved |
Later papers that compared with this method
1 later accepted paper reports ZIPVL as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Efficient Multimodal Large Language Model via Dynamic KV Cache Quantization (AAAI 2026 · better on 2 of 2 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.93)
- Dynamic-LLaVA: Efficient Multimodal Large Language Models via Dynamic Vision-language Context Sparsification (ICLR 2025 · similarity 0.93)
- Attention-aware Inference Optimizations for Large Vision-Language Models with Memory-efficient Decoding (CVPR 2026 · similarity 0.92)
- A-VL: Adaptive Attention for Large Vision-Language Models (AAAI 2025 · similarity 0.92)
- A Stitch in Time Saves Nine: Small VLM is a Precise Guidance for Accelerating Large VLMs (CVPR 2025 · similarity 0.91)
- SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (ICML 2025 · 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.
