AI Paper Atlas

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

ICCV 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Attention-based visual token pruning · 9 citations (OpenAlex)

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 changedWhereMeasuredEffect
move sparsity from prefill-only (90%) to decoupled 70% prefill / 85% decode (ablation)SparseVILA on Cosmos-Reason1-7B · RoboVQAaccuracy: 80.0 -> 89.1Improved
add decoupled decoding-stage query-aware retrieval to prefill pruning, versus VisionZipLLaVA-NeXT-7B · ChartQAaccuracy: 38.2 -> 47.8Improved
SparseVILA versus VisionZipLLaVA-NeXT-7B · DocVQAaccuracy: 48.5 -> 58.0Improved
SparseVILA versus unpruned model on image tasksLLaVA-NeXT-7B · DocVQAaccuracy: 63.6 -> 58.0Hurt
SparseVILA versus unpruned model on general VQALLaVA-NeXT-7B · POPEaccuracy: 84.5 -> 85.8Improved
SparseVILA versus FastV on InternVL3.5-8BInternVL3.5-8B · ChartQAaccuracy: 28.4 -> 61.8Improved
SparseVILA versus unpruned on long video understandingQwen2.5-VL-7B · Video-MME overallaccuracy: 62.3 -> 66.3Improved
SparseVILA versus VisionZip and PruMerge (95% prefill pruning)LongVILA-7B · Video-MME overallaccuracy: 52.2 -> 58.7Improved
SparseVILA versus unpruned on video captioningLongVILA-7B · Video-ChatGPT overallGPT score: 2.31 -> 2.35Improved
SparseVILA versus unpruned on temporal understandingLongVILA-7B · Video-ChatGPT TUGPT score: 1.70 -> 1.90Improved
SparseVILA versus unpruned on video reasoningLongVILA-R1-7B · LongVideo-Reason overallaccuracy: 74.5 -> 74.4No clear effect
SparseVILA versus unpruned on physical reasoning at 24fpsCosmos-Reason1-7B · Physical reasoning average (24fps)accuracy: 71.4 -> 75.9Improved

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.

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

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy