Research map / Video / Video-language understanding
An Empirical Study on How Video-LLMs Answer Video Questions
Chenhui Gou, Ziyu Ma, Zicheng Duan, Haoyu He, Feng Chen, Akide Liu, Bohan Zhuang, Jianfei Cai, Hamid Rezatofighi
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 |
|---|---|---|---|
| block language-to-video attention in all layers (no visual info reaches text) (ablation) | Video-LLMs · Video-MME / EgoSchema | performance retained: retain at most 70% (LongVA, Video-MME), as little as 48% (LLaVA-Video, EgoSchema) | Hurt |
| block language-to-video attention beyond about 60% of layers (ablation) | Video-LLMs · Video-MME, MVBench, EgoSchema | normalized performance | No clear effect |
| block language-to-video attention in layers between 50% and 60% depth (ablation) | Video-LLMs · all benchmarks | normalized performance | Hurt |
| knock out video temporal or spatial self-attention in all layers (ablation) | Video-LLMs · Video-MME, MVBench, EgoSchema | accuracy | No clear effect |
| knock out language-to-video attention in all layers (ablation) | Video-LLMs · Video-MME, MVBench, EgoSchema | accuracy | Hurt |
| knock out a window of 4 critical intermediate layers (e.g. layers 12-16 of 28) (ablation) | Video-LLMs · VideoQA benchmarks | normalized performance | Hurt |
| exit video tokens after layer 18 | LLaVA-Video-7B · Video-MME | accuracy: 62.4 -> 62.0 (attention FLOPs 64.3%) | No clear effect |
| exit video tokens after layer 18 and spatial-only attention in layers 1-8 | LLaVA-Video-7B · Video-MME | accuracy: 62.4 -> 60.0 (attention FLOPs 37.6%) | Hurt |
| exit video tokens after layer 18 and spatial-only attention in layers 1-8 | LLaVA-OneVision-7B · EgoSchema | accuracy: 65.2 -> 65.2 (attention FLOPs 37.5%) | No clear effect |
| exit video tokens after layer 18 | LLaVA-OneVision-7B · MVBench | accuracy: 58.3 -> 57.3 | Hurt |
Closest papers by meaning
- Can Video LLMs Refuse to Answer? Alignment for Answerability in Video Large Language Models (ICLR 2025 · similarity 0.91)
- Recurrent Attention-based Token Selection for Efficient Streaming Video-LLMs (NeurIPS 2025 · similarity 0.91)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.91)
- SlowFocus: Enhancing Fine-grained Temporal Understanding in Video LLM (NeurIPS 2024 · similarity 0.90)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.90)
- SF2T: Self-supervised Fragment Finetuning of Video-LLMs for Fine-Grained Understanding (CVPR 2025 · 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.
