Research map / Video / Video-language understanding
Map the Flow: Revealing Hidden Pathways of Information in VideoLLMs
Minji Kim, Taekyung Kim, Bohyung Han
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 cross-frame attention in first 16 layers (ablation) | LLaVA-NeXT-7B-Video-FT · TVBench Object Count | accuracy drop when removed: -60.8% | Improved |
| add cross-frame attention in first 16 layers (ablation) | LLaVA-NeXT-7B-Video-FT · TVBench Action Antonym | accuracy drop when removed: -24.1% | Improved |
| add cross-frame attention in first 16 layers (ablation) | LLaVA-NeXT-7B-Video-FT · TVBench Scene Transition | accuracy drop when removed: -18.0% | Improved |
| restrict attention to effective pathways instead of full causal attention | LLaVA-NeXT-7B-Video-FT · TVBench | accuracy: 51.5 -> 51.2 | No clear effect |
| restrict attention to effective pathways instead of full causal attention | LLaVA-NeXT-7B-Video-FT · TOMATO | accuracy: 30.2 -> 29.2 | No clear effect |
| restrict attention to effective pathways instead of full causal attention | LLaVA-NeXT-13B-Video-FT · TVBench | accuracy: 55.1 -> 54.6 | No clear effect |
| restrict attention to effective pathways instead of full causal attention | VideoLLaMA3-7B · TVBench | accuracy: 55.2 -> 57.2 | Improved |
| restrict attention to effective pathways instead of full causal attention | Mini-InternVL-4B-Video-FT · TOMATO | accuracy: 32.2 -> 31.2 | No clear effect |
| use identified effective pathways instead of random blocking of same edge proportion (ablation) | LLaVA-NeXT-7B-Video-FT · TVBench | accuracy: 40.1 -> 51.2 | Improved |
| use identified effective pathways instead of random blocking of same edge proportion (ablation) | VideoLLaMA3-7B · TVBench | accuracy: 22.2 -> 57.2 | Improved |
| use identified effective pathways instead of random blocking of same edge proportion (ablation) | Mini-InternVL-4B-Video-FT · TOMATO | accuracy: 25.9 -> 31.2 | Improved |
| use identified effective pathways instead of random blocking of same edge proportion (ablation) | LLaVA-NeXT-13B-Video-FT · TOMATO | accuracy: 23.8 -> 27.4 | Improved |
Closest papers by meaning
- Streaming Long Video Understanding with Large Language Models (NeurIPS 2024 · similarity 0.92)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.91)
- Language-aware Visual Semantic Distillation for Video Question Answering (CVPR 2024 · similarity 0.91)
- VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · similarity 0.91)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.90)
- Temporal Reasoning Transfer from Text to Video (ICLR 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.
