Research map / Video / Video-language understanding
TimeViper: A Hybrid Mamba-Transformer Vision-Language Model for Efficient Long Video Understanding
Boshen Xu, Zihan Xiao, Jiaze Li, Jianzhong Ju, Zhenbo Luo, Jian Luan, Qin Jin
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 TransV token transfer instead of plain uniform token dropping at layer 7 (ablation) | TimeViper · Charades-STA | mIoU: 26.1 -> 38.1 | Improved |
| add TransV in the LLM (50% at layer 7, 90% attention-guided at layer 39) (ablation) | TimeViper · VideoMME, Charades-STA, LVBench | accuracy / mIoU: VideoMME 58.8 -> 56.9; Charades 40.5 -> 37.9; LVBench 35.5 -> 35.6 | Mixed |
| add TransV to a Transformer Qwen2.5-7B model (ablation) | Qwen2.5-7B · LVBench and VDC | accuracy: LVBench 36.6 -> 37.0; VDC 42.0 -> 40.7 | Mixed |
| raise TransV compression at layer 7 from 50% to 90% (ablation) | TimeViper · VideoMME | accuracy: 56.7 -> 53.4 | Hurt |
| use attention-guided instead of uniform TransV at layer 39 (ablation) | TimeViper · VideoMME | accuracy: 56.2 -> 56.6 | Improved |
| place TransV at layer 7 instead of layer 2 (ablation) | TimeViper · VideoMME and VDC | accuracy: VideoMME 56.1 -> 56.7; VDC 39.7 -> 38.9 | Mixed |
| add deep-layer uniform TransV (90% at layer 39) on top of shallow TransV (ablation) | TimeViper · VideoMME | accuracy: 56.7 -> 56.2 | Mixed |
| add TransV on top of token merging (ToMe) in the projector (ablation) | TimeViper · GPU memory at 4,096 frames | memory: 54.8% less memory | Improved |
| add TransV on top of ToMe (ablation) | TimeViper · prefilling time at 4,096 frames | latency: 15.7% less prefilling time | Improved |
| add token merging (ToMe) in the projector (ablation) | TimeViper · frame capacity | max frames: out of memory at 128 frames -> about 5K frames | Improved |
| use hybrid Nanov2-9B backbone instead of Transformer Qwen2.5-7B under identical recipe | TimeViper · VideoMME | accuracy: 56.6 -> 58.8 | No clear effect |
| increase test-time input frames from 256 to 512 | TimeViper · MLVU and LVBench | accuracy: MLVU 65.64 -> 69.00; LVBench 35.53 -> 37.0 | 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.
- Video-XL — Video-XL: Extra-Long Vision Language Model for Hour-Scale Video Understanding (CVPR 2025) · better on 1 of 1 shared result
- VTimeLLM — VTimeLLM: Empower LLM to Grasp Video Moments (CVPR 2024) · better on 1 of 1 shared result
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
- TimeChat: A Time-sensitive Multimodal Large Language Model for Long Video Understanding (CVPR 2024 · similarity 0.89)
- Enhancing Temporal Understanding in Video-LLMs through Stacked Temporal Attention in Vision Encoders (NeurIPS 2025 · similarity 0.88)
- TC-LLaVA: Rethinking the Transfer of LLava from Image to Video Understanding with Temporal Considerations (AAAI 2025 · similarity 0.88)
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.87)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.87)
- TimeSuite: Improving MLLMs for Long Video Understanding via Grounded Tuning (ICLR 2025 · similarity 0.87)
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.
