Research map / Video / Video-language understanding
Seeing the Scene Matters: Revealing Forgetting in Video Understanding Models with a Scene-Aware Long-Video Benchmark
Seng Nam Chen, Hao Chen, Chenglam Ho, Xinyu Mao, Jinping Wang, Yu Zhang, Chao Li
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 |
|---|---|---|---|
| move from clip-level to scene-level questions | open-source MLLMs · SceneBench | accuracy: ClipQA 51.73 -> SceneQA 26.11 | Hurt |
| add Scene-RAG to base MLLMs | LongVA, Long-LLaVA, LLaVA-OneVision · SceneBench | avg accuracy: 53.3 -> 54.4 (LongVA); 52.9 -> 54.6 (Long-LLaVA); 58.1 -> 59.4 (LLaVA-OV) | Improved |
| add frame-level Video-RAG to base MLLMs | LongVA, Long-LLaVA, LLaVA-OneVision · SceneBench | avg accuracy: +0.1 / +0.0 / +0.2 | No clear effect |
| add Video-RAG to LLaVA-OneVision on I-VQA | LLaVA-OneVision-7B · SceneBench I-VQA | accuracy: 46.5 -> 43.7 | Hurt |
| add Scene-RAG to base MLLMs | LongVA, Long-LLaVA, LLaVA-OneVision · Video-MME overall | accuracy: 51.4 -> 63.4 (LongVA); 52.0 -> 64.8 (Long-LLaVA); 43.0 -> 50.6 (LLaVA-OV) | Improved |
| add visual retrieval only (ablation) | LongVA · SceneBench | avg accuracy: 53.3 -> 53.5 | No clear effect |
| add audio retrieval (ablation) | LongVA Scene-RAG · SceneBench | avg accuracy: 53.5 -> 54.0 | Improved |
| add scene tiling (ablation) | LongVA Scene-RAG · SceneBench | avg accuracy: 54.0 -> 54.4 | Improved |
| add audio retrieval (ablation) | LLaVA-Video-7B Scene-RAG · MLVU | accuracy: 71.2 -> 73.4 | Improved |
| add scene tiling (ablation) | LLaVA-Video-7B Scene-RAG · MLVU | accuracy: 73.4 -> 74.1 | Improved |
| add audio track as input (SceneQA-Audio vs SceneQA) | open-source MLLMs · SceneBench | accuracy: 26.11 -> 29.08 | Improved |
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.
- TimeChat — TimeChat: A Time-sensitive Multimodal Large Language Model for Long Video Understanding (CVPR 2024) · 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
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.90)
- MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding (NeurIPS 2024 · similarity 0.90)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.90)
- MotionBench: Benchmarking and Improving Fine-grained Video Motion Understanding for Vision Language Models (CVPR 2025 · similarity 0.90)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.89)
- MMLongBench: Benchmarking Long-Context Vision-Language Models Effectively and Thoroughly (NeurIPS 2025 · 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.
