Research map / Video / Video-language understanding
B-VLLM: A Vision Large Language Model with Balanced Spatio-Temporal Tokens
Zhuqiang Lu, Zhenfei Yin, Mengwei He, Zhihui Wang, Zicheng Liu, Zhiyong Wang, Kun Hu
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 adaptive frame selection (ablation) | LLaMA-VID · VideoMME | accuracy: 33.6 -> 36.5 | Improved |
| add adaptive frame selection (ablation) | LLaMA-VID · MMBench | accuracy: 49.8 -> 58.0 | Improved |
| add temporal frame token merging on top of frame selection (ablation) | LLaMA-VID · MVBench | accuracy: 39.4 -> 42.1 | Improved |
| add temporal frame token merging on top of frame selection (ablation) | LLaMA-VID · POPE | accuracy: 81.1 -> 67.3 | Hurt |
| add spatial visual token sampling on top of F.S. and T.M. (ablation) | LLaMA-VID · MVBench | accuracy: 42.1 -> 43.5 | Improved |
| add spatial visual token sampling on top of F.S. and T.M. (ablation) | LLaMA-VID · POPE | accuracy: 67.3 -> 83.8 | Improved |
| integrate the method into LLaMA-VID | LLaMA-VID · VideoMME long | accuracy: 27.1 -> 35.2 | Improved |
| integrate the method into Video-LLaMA2 | Video-LLaMA2 · VideoMME long | accuracy: 37.7 -> 41.5 | Improved |
| integrate the method into Video-LLaMA2 | Video-LLaMA2 · VideoMME short | accuracy: 48.9 -> 47.2 | Hurt |
| integrate the method into LLaMA-VID | LLaMA-VID · VizWiz | accuracy: 37 -> 36.9 | No clear effect |
| use Resampler instead of Q-Former for frame selection (ablation) | B-VLLM frame selection · MVBench | accuracy: 46.5 -> 44.3 | Hurt |
| use [CLS] token features instead of mean pooling for frame selection (ablation) | VideoLLaMA2 preliminary · MVBench | accuracy: 41.2 -> 34.5 | Hurt |
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.
- LLaMA-VID — LLaMA-VID: An Image is Worth 2 Tokens in Large Language Models (ECCV 2024) · better on 7 of 7 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
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.91)
- Visual Perception by Large Language Model’s Weights (NeurIPS 2024 · similarity 0.90)
- Jack of All Tasks Master of Many: Designing General-Purpose Coarse-to-Fine Vision-Language Model (CVPR 2024 · similarity 0.90)
- VideoLLM-MoD: Efficient Video-Language Streaming with Mixture-of-Depths Vision Computation (NeurIPS 2024 · similarity 0.90)
- Video-XL: Extra-Long Vision Language Model for Hour-Scale Video Understanding (CVPR 2025 · similarity 0.90)
- Rethinking Visual Token Reduction in LVLMs Under Cross-Modal Misalignment (AAAI 2026 · 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.
