Research map / Video / Video-language understanding
Linear Scaling Video VLMs for Long Video Understanding
Cristobal Eyzaguirre, Jiajun Wu, Juan Carlos Niebles
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 importance-based StateKV state instead of ReKV sliding window (compute-matched) | StateKV · VideoMME, MLVU, OVOBench | accuracy: about 10 points on average | Improved |
| use StateKV (B=4096) instead of full self-attention | StateKV · VideoMME, MLVU, OVOBench | accuracy: within roughly a point on average | No clear effect |
| run InternVL3-8B with StateKV instead of InternVL3-1B with full SA at similar compute | StateKV · VideoMME (no subtitles, 512 frames) | accuracy: 46.2 -> 62.5 | Improved |
| increase StateKV cache budget B (ablation) | StateKV · VideoMME short videos (InternVL3) | accuracy: 75.0 | Improved |
| replace full SA with ReKV (R=16) on InternVL3-2B | ReKV · MLVU (512 frames) | accuracy: 56.6104 -> 5.4931 | Hurt |
| use StateKV (B=4096) instead of full SA on InternVL3-2B | StateKV · MLVU (512 frames) | accuracy: 56.6104 -> 57.3099 | No clear effect |
| use StateKV (B=4096) instead of full SA on InternVL3-8B | StateKV · MLVU (512 frames) | accuracy: 61.14 -> 62.85 | No clear effect |
| use StateKV (B=4096) instead of full SA on Qwen3-VL-8B | StateKV · MLVU (512 frames) | accuracy: 75.82 -> 71.38 | Hurt |
| use ReKV (R=16) instead of full SA on Qwen3-VL-8B | ReKV · VideoMME (no subtitles, 512 frames) | accuracy: 70.52 -> 55.52 | Hurt |
| select tokens by attention instead of keeping recent frames (256 tokens vs 1 frame) (ablation) | StateKV · Attention mass analysis (InternVL3-1B/2B/8B) | fraction of historical attention mass: gap about 0.59/0.57/0.62 | Improved |
| raise cache budget B from 256 to 4096 tokens (ablation) | StateKV · Attention mass analysis (InternVL3-1B) | historical attention mass captured: 0.71 -> 0.93 | Improved |
| use Triton fused score kernel instead of eager attention for cache building (ablation) | StateKV · wall time per frame (L40S) | wall time | Improved |
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.90)
- BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding (CVPR 2025 · similarity 0.89)
- VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration (ICLR 2025 · similarity 0.89)
- Streaming Long Video Understanding with Large Language Models (NeurIPS 2024 · similarity 0.88)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.88)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 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.
