Research map / Video / Video-language understanding
Sink-Token-Aware Pruning for Fine-Grained Video Understanding in Efficient Video LLMs
Kibum Kim, Jiwan Kim, Kyle Min, Yueqi Wang, Jinyoung Moon, Julian McAuley, Chanyoung Park
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 sink-token-aware spatial pruning (STSP) (ablation) | VisionZip · fine-grained suite, 10% retention | performance drop rate: 20.46% -> 4.64% | Improved |
| add STSP (ablation) | Holitom · fine-grained suite, 10% retention | performance drop rate: 3.87% -> 1.94% | Improved |
| add sink-token-aware temporal pruning (STTP) on top of STSP (ablation) | Holitom · fine-grained suite, 10% retention | performance drop rate: 1.94% -> 1.17% | Improved |
| add SToP | FastVid · fine-grained suite, 10% retention | performance drop rate: 15.69% -> 6.32% | Improved |
| add SToP | VisionZip · EventHallusion Binary, 10% retention | accuracy: 50.37 -> 60.39 | Improved |
| add SToP | Holitom · fine-grained suite, 20% retention | performance drop rate: 2.41% -> 1.84% | Improved |
| add SToP | VisionZip on LLaVA-Video-7B · fine-grained suite, 10% retention | performance drop rate: 26.46% -> 7.04% | Improved |
| add SToP | VisionZip, FastVid, Holitom · MCQA suite, 10% retention | performance drop rate: 7.36%->3.34%, 5.12%->2.90%, 3.21%->2.02% | Improved |
| remove temporal pruning (ablation) | Holitom · EventHallusion | hallucination performance | Hurt |
| remove sink tokens at increasing ratio from selected set (ablation) | VisionZip · EventHallusion and MCQA | performance degradation | Improved |
| use attention-based pruning + SToP instead of feature-based DPC-KNN pruning (ablation) | spatial pruning · fine-grained suite, 10% retention | performance drop rate: 6.50% -> 4.41% | Improved |
| use feature-based pruning instead of attention-based pruning (ablation) | spatial pruning · fine-grained suite, 10% retention | performance drop rate: 15.85% -> 6.50% | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- FastVid — FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025) · better on 4 of 4 shared results
- Holitom — HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025) · better on 4 of 4 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
- HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.91)
- Recurrent Attention-based Token Selection for Efficient Streaming Video-LLMs (NeurIPS 2025 · similarity 0.90)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.90)
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.90)
- LearnPruner: Rethinking Attention-based Token Pruning in Vision Language Models (ICLR 2026 · similarity 0.90)
- SlowFocus: Enhancing Fine-grained Temporal Understanding in Video LLM (NeurIPS 2024 · similarity 0.90)
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.
