Research map / Video / Video-language understanding
EventSTU: Event-Guided Efficient Spatio-Temporal Understanding for Video-LLMs
Wenhao Xu, Xin Dong, Yue Li, Haoyuan Shi, Yueyi Zhang, Zhiwei Xiong
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 retention ratio to ZAP (ablation) | ZAP · EventBench / LongVideoBench | accuracy: 54.09 -> 55.09 (EventBench); 59.54 -> 61.11 (LongVideoBench) | Improved |
| combine physics-aware and semantic-aware selection (ablation) | ZAP · EventBench | accuracy: 52.10 -> 54.09 | Improved |
| add physics-aware selection to semantic-only (ablation) | ZAP · LongVideoBench | accuracy: 58.12 -> 59.54 | Improved |
| accumulation (ACC) coarse sampling instead of uniform (ablation) | C2FS · EventBench | accuracy: 53.09 -> 55.09 | Improved |
| accumulation (ACC) coarse sampling instead of top density (ablation) | C2FS · EventBench | accuracy: 54.09 -> 55.09 | Improved |
| bin sampling (BIN) instead of TOP at fine stage (ablation) | C2FS · EventBench | accuracy: 53.89 -> 55.09 | Improved |
| add C2FS keyframe sampling instead of AKS | LLaVA-OV-7B · Avg (4 benchmarks) | average score: 61.06 -> 61.59 | Improved |
| add ZAP token pruning instead of DyCoke at 50% | LLaVA-OV-7B · Avg (4 benchmarks) | average score: 57.97 -> 59.07 | Improved |
| add EventSTU at 30% retention | LLaVA-OV-7B · Avg (4 benchmarks) | average score: 58.05 -> 61.86 | Improved |
| add EventSTU at 30% retention | Qwen3-VL-8B · Avg (4 benchmarks) | average score: 66.60 -> 69.29 | Improved |
| raise coarse sampling ratio S to 0.50 (ablation) | C2FS · number-of-frames sweep | accuracy | Hurt |
| add ZAP at high retention ratio R=0.7 (ablation) | LLaVA-OV-7B · Fig 10 | accuracy: +2.19 points | 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.
- FastVID (50%) — FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025) · better on 2 of 2 shared results
- HoliTom (50%) — HoliTom: Holistic Token Merging for Fast Video Large Language Models (NeurIPS 2025) · 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
- EventGPT: Event Stream Understanding with Multimodal Large Language Models (CVPR 2025 · similarity 0.90)
- E.T. Bench: Towards Open-Ended Event-Level Video-Language Understanding (NeurIPS 2024 · similarity 0.89)
- EventLens: Event-Structure Reinforcement Learning for Video Understanding (NeurIPS 2026 · similarity 0.88)
- EventPrune: Cascaded Event-Assisted Token Pruning for Efficient First-Person Dynamic Spatial Reasoning (NeurIPS 2026 · similarity 0.88)
- Unleashing Hour-Scale Video Training for Long Video-Language Understanding (NeurIPS 2025 · similarity 0.88)
- OpenESS: Event-based Semantic Scene Understanding with Open Vocabularies (CVPR 2024 · 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.
