Research map / Video / Video-language understanding
Divid: Disentangled Spatial-Temporal Modeling within LLMs for Temporally Grounded Video Understanding
Yepeng Tang, Weining Wang, Longteng Guo, Tongtian Yue, Wenxuan Wang, Chunjie Zhang, Jing Liu
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 Top-K temporal-attention keyframe selection instead of uniform sampling (ablation) | Divid · Charades-STA | mIoU: 46.27 -> 47.33 | Improved |
| add Top-K temporal-attention keyframe selection instead of uniform sampling (ablation) | Divid · ReXTime | mIoU: 27.61 -> 29.71 | Improved |
| add token-wise soft-router instead of plain addition (ablation) | Divid · Charades-STA | mIoU: 46.81 -> 47.33 | Improved |
| add token-wise soft-router instead of plain addition (ablation) | Divid · ReXTime | mIoU: 29.11 -> 29.71 | Improved |
| disentangle spatial and temporal modeling inside the LLM instead of Slow-Fast concatenation (ablation) | Divid · Charades-STA | mIoU: 46.71 -> 47.33 | Improved |
| disentangle spatial and temporal modeling inside the LLM instead of Slow-Fast concatenation (ablation) | Divid · ReXTime | mIoU: 29.49 -> 29.71 | Improved |
| use disentangled Divid instead of full high-resolution frames (ablation) | Divid · Charades-STA | TFLOPs: 28.2 -> 10.5 TFLOPs | Improved |
| train on TempGCap instead of Momentor-10M (ablation) | Divid · Charades-STA | mIoU: 46.42 -> 47.33 | Improved |
| train on TempGCap instead of Momentor-1M (ablation) | Divid · Charades-STA | mIoU: 44.90 -> 47.33 | Improved |
| fuse at every decoder layer instead of only at output level (ablation) | Divid · Charades-STA | mIoU: 42.51 -> 47.33 | Improved |
| use separate LoRA attention parameters per branch instead of shared (ablation) | Divid · Charades-STA/ReXTime | mIoU: 47.33 -> 47.38 (Charades-STA mIoU); 29.71 -> 29.24 (ReXTime mIoU) | No clear effect |
| increase temporal frames from 128 to 256 (ablation) | Divid · ReXTime | mIoU: 29.71 -> 30.35 | No clear effect |
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.
- VideoMind — VideoMind: A Chain-of-LoRA Agent for Temporal-Grounded Video Reasoning (ICLR 2026) · better on 6 of 6 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
- VideoRefer Suite: Advancing Spatial-Temporal Object Understanding with Video LLM (CVPR 2025 · similarity 0.90)
- Improving LLM Video Understanding with 16 Frames Per Second (ICML 2025 · similarity 0.90)
- SlowFocus: Enhancing Fine-grained Temporal Understanding in Video LLM (NeurIPS 2024 · similarity 0.89)
- Re-thinking Temporal Search for Long-Form Video Understanding (CVPR 2025 · similarity 0.89)
- Scaling Video-Language Models to 10K Frames via Hierarchical Differential Distillation (ICML 2025 · similarity 0.89)
- SF2T: Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding (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.
