Research map / Video / Video-language understanding
Vamba: Understanding Hour-Long Videos with Hybrid Mamba-Transformers
Weiming Ren, Wentao Ma, Huan Yang, Cong Wei, Ge Zhang, Wenhu Chen
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 |
|---|---|---|---|
| initialize cross-attention weights from pretrained self-attention (ablation) | Vamba ablation Model A · LVBench | accuracy: 23.7 -> 34.2 | Improved |
| initialize cross-attention weights from pretrained self-attention (ablation) | Vamba ablation Model A · Video-MME w/o subtitles | accuracy: 47.6 -> 51.7 | Improved |
| initialize cross-attention weights from pretrained self-attention (ablation) | Vamba ablation Model A · MVBench | accuracy: 40.9 -> 51.8 | Improved |
| add Mamba block to update video tokens (ablation) | Vamba Model B · Video-MME w/o subtitles | accuracy: 51.7 -> 53.4 | Improved |
| add Mamba block to update video tokens (ablation) | Vamba Model B · LVBench | accuracy: 34.2 -> 34.2 | No clear effect |
| add Mamba block to update video tokens (ablation) | Vamba Model B · MVBench | accuracy: 51.8 -> 53.5 | Improved |
| use Mamba-2 instead of Mamba blocks (ablation) | Vamba Model C · LVBench | accuracy: 34.2 -> 35.3 | Improved |
| use Mamba-2 instead of Mamba blocks (ablation) | Vamba Model C · G-VEval stage 1 | G-VEval score: 81.8 -> 82.2 | Improved |
| add distillation loss (lambda 0.001) in pretraining (ablation) | Vamba Model D · G-VEval stage 1 | G-VEval score: 82.19 -> 81.05 | Hurt |
| add distillation loss (lambda 2) in pretraining (ablation) | Vamba Model D · G-VEval stage 1 | G-VEval score: 82.19 -> 47.61 | Hurt |
| add token reduction to Vamba (ablation) | Vamba · LVBench | accuracy: 42.4 -> 41.6 | Hurt |
| add token reduction to Vamba (ablation) | Vamba · MLVU | accuracy: 65.9 -> 66.5 | Mixed |
Later papers that compared with this method
1 later accepted paper reports VAMBA as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- State-Space Hierarchical Compression with Gated Attention and Learnable Sampling for Hour-Long Video Understanding in Large Multimodal Models (AAAI 2026 · better on 2 of 2 shared results)
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.
- LongVU (7B) — LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025) · better on 5 of 5 shared results
- Video-XL (7B) — Video-XL: Extra-Long Vision Language Model for Hour-Scale Video Understanding (CVPR 2025) · better on 3 of 3 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
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.88)
- LLaVA-Mini: Efficient Image and Video Large Multimodal Models with One Vision Token (ICLR 2025 · similarity 0.88)
- LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.87)
- iLLaVA: An Image is Worth Fewer Than 1/3 Input Tokens in Large Multimodal Models (ICLR 2026 · similarity 0.87)
- MA-LMM: Memory-Augmented Large Multimodal Model for Long-Term Video Understanding (CVPR 2024 · similarity 0.86)
- Unleashing Hour-Scale Video Training for Long Video-Language Understanding (NeurIPS 2025 · similarity 0.86)
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.
