Research map / Video / Video-language understanding
Enhancing Visual Token Representations for Video Large Language Models via Training-free Spatial-Temporal Pooling and Gridding
Bingjun Luo, Tony Wang, Hanqi Chen, Xinpeng Ding
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 ST-GridPool (training-free) to LLaVA-Video-7B | LLaVA-Video-7B · LongVideoBench | accuracy: 58.2 -> 60.1 | Improved |
| add ST-GridPool to LLaVA-OneVision-7B | LLaVA-OneVision-7B · EgoSchema | accuracy: 60.1 -> 62.1 | Improved |
| add pyramid temporal gridding on top of norm-based pooling (ablation) | LLaVA-Video-7B · MVBench | accuracy: 58.8 -> 59.8 | Improved |
| add norm-based spatial pooling on top of PTG (ablation) | LLaVA-Video-7B · MVBench | accuracy: 59.1 -> 59.8 | Improved |
| use ST-GridPool as token reducer at 30% budget instead of FrameFusion | LLaVA-Video-7B · EgoSchema | accuracy: 53.0 -> 56.0 | Improved |
| use ST-GridPool at 50% budget vs FrameFusion on VideoMME | LLaVA-Video-7B · VideoMME (50% budget) | accuracy: 62.6 -> 62.5 | No clear effect |
| apply IG-VLM image gridding to LLaVA-Video | LLaVA-Video-7B · MVBench | accuracy: 58.6 -> 54.2 | Hurt |
| apply TS-LLaVA to LLaVA-Video | LLaVA-Video-7B · MVBench | accuracy: VideoMME 63.3 -> 63.8; MVBench 58.6 -> 57.1 | Mixed |
| token-level gridding instead of image-level gridding (ablation) | ST-GridPool on LLaVA-Video-7B · VideoMME | accuracy: 62.8 -> 64.2 | Improved |
| increase softmax temperature beta to 5 or 10 in NSP (ablation) | ST-GridPool · VideoMME / LongVideoBench | accuracy | Hurt |
| use norm order p above 2 in NSP (ablation) | ST-GridPool · VideoMME / LongVideoBench | accuracy | Hurt |
| use spatial pooling kernel larger than 2x2 (ablation) | ST-GridPool · VideoMME / LongVideoBench | accuracy | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- FrameFusion — FrameFusion: Combining Similarity and Importance for Video Token Reduction on Large Vision Language Models (ICCV 2025) · better on 1 of 2 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
- What Do Visual Tokens Really Encode? Uncovering Sparsity and Redundancy in Multimodal Large Language Models (CVPR 2026 · similarity 0.90)
- Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization (ICML 2024 · similarity 0.90)
- VQToken: Neural Discrete Token Representation Learning for Extreme Token Reduction in Video Large Language Models (NeurIPS 2025 · similarity 0.89)
- ST3: Accelerating Multimodal Large Language Model by Spatial-Temporal Visual Token Trimming (AAAI 2025 · similarity 0.89)
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.89)
- Task-Related Token Compression in Multimodal Large Language Models from an Explainability Perspective (ICLR 2026 · similarity 0.89)
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.
