Research map / Vision-language and multimodal / Multimodal LLMs
SparseVILA: Decoupling Visual Sparsity for Efficient VLM Inference
Samir Khaki, Junxian Guo, Jiaming Tang, Shang Yang, Yukang Chen, Konstantinos N. Plataniotis, Yao Lu, Song Han, Zhijian Liu
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 |
|---|---|---|---|
| move sparsity from prefill-only (90%) to decoupled 70% prefill / 85% decode (ablation) | SparseVILA on Cosmos-Reason1-7B · RoboVQA | accuracy: 80.0 -> 89.1 | Improved |
| add decoupled decoding-stage query-aware retrieval to prefill pruning, versus VisionZip | LLaVA-NeXT-7B · ChartQA | accuracy: 38.2 -> 47.8 | Improved |
| SparseVILA versus VisionZip | LLaVA-NeXT-7B · DocVQA | accuracy: 48.5 -> 58.0 | Improved |
| SparseVILA versus unpruned model on image tasks | LLaVA-NeXT-7B · DocVQA | accuracy: 63.6 -> 58.0 | Hurt |
| SparseVILA versus unpruned model on general VQA | LLaVA-NeXT-7B · POPE | accuracy: 84.5 -> 85.8 | Improved |
| SparseVILA versus FastV on InternVL3.5-8B | InternVL3.5-8B · ChartQA | accuracy: 28.4 -> 61.8 | Improved |
| SparseVILA versus unpruned on long video understanding | Qwen2.5-VL-7B · Video-MME overall | accuracy: 62.3 -> 66.3 | Improved |
| SparseVILA versus VisionZip and PruMerge (95% prefill pruning) | LongVILA-7B · Video-MME overall | accuracy: 52.2 -> 58.7 | Improved |
| SparseVILA versus unpruned on video captioning | LongVILA-7B · Video-ChatGPT overall | GPT score: 2.31 -> 2.35 | Improved |
| SparseVILA versus unpruned on temporal understanding | LongVILA-7B · Video-ChatGPT TU | GPT score: 1.70 -> 1.90 | Improved |
| SparseVILA versus unpruned on video reasoning | LongVILA-R1-7B · LongVideo-Reason overall | accuracy: 74.5 -> 74.4 | No clear effect |
| SparseVILA versus unpruned on physical reasoning at 24fps | Cosmos-Reason1-7B · Physical reasoning average (24fps) | accuracy: 71.4 -> 75.9 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VisionZip — VisionZip: Longer is Better but Not Necessary in Vision Language Models (CVPR 2025) · better on 8 of 8 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
- SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (ICML 2025 · similarity 0.92)
- NVILA: Efficient Frontier Visual Language Models (CVPR 2025 · similarity 0.90)
- SwiftVLM: Efficient Vision-Language Model Inference via Cross-Layer Token Bypass (NeurIPS 2026 · similarity 0.89)
- A Stitch in Time Saves Nine: Small VLM is a Precise Guidance for Accelerating Large VLMs (CVPR 2025 · similarity 0.89)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.89)
- Inference Optimal VLMs Need Fewer Visual Tokens and More Parameters (ICLR 2025 · similarity 0.89)
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.
