Research map / Video / Video-language understanding
Graph-to-Frame RAG: Visual-Space Knowledge Fusion for Training-Free and Auditable Video Reasoning
Songyuan Yang, Weijiang Yu, Ziyu Liu, Guijian Tang, Wenjing Yang, Huibin Tan, Nong Xiao
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 |
|---|---|---|---|
| deliver retrieved subgraph as a rendered reasoning frame instead of textual JSON (ablation) | G2F-RAG on Qwen2.5-VL-7B · VideoMME (w/o sub) | accuracy (%): 63.0 -> 70.6 | Improved |
| deliver retrieved subgraph as a rendered reasoning frame instead of textual JSON (ablation) | G2F-RAG on Qwen2.5-VL-7B · WildVideo | accuracy (%): 48.8 -> 55.4 | Improved |
| append the reasoning frame at the end rather than mid-sequence (ablation) | G2F-RAG on Qwen2.5-VL-7B · VideoMME (w/o sub) | accuracy (%): 64.0 -> 70.6 | Improved |
| append the reasoning frame at the end rather than the start (ablation) | G2F-RAG on Qwen2.5-VL-7B · VideoMME (w/o sub) | accuracy (%): 69.4 -> 70.6 | Improved |
| use a single end frame instead of four end frames (ablation) | G2F-RAG on Qwen2.5-VL-7B · MLVU | accuracy (%): 69.0 -> 73.4 | Improved |
| use a single end frame instead of two end frames (ablation) | G2F-RAG on Qwen2.5-VL-7B · VideoMME (w/o sub) | accuracy (%): 69.8 -> 70.6 | Improved |
| use minimal icon-and-short-label frame style instead of dense text (ablation) | G2F-RAG on Qwen2.5-VL-7B · WildVideo | accuracy (%): 50.5 -> 55.4 | Improved |
| keep both spatial topology and causality instead of topology only (ablation) | G2F-RAG on Qwen2.5-VL-7B · MLVU | accuracy (%): 71.0 -> 73.4 | Improved |
| add difficulty routing instead of applying the frame to all queries (ablation) | G2F-RAG on Qwen2.5-VL-7B · VideoMME (w/o sub) | accuracy (%): 66.8 -> 70.6 | Improved |
| add conservative fallback to routing (ablation) | G2F-RAG on Qwen2.5-VL-7B · VideoMME (w/o sub) | accuracy (%): 69.0 -> 70.6 | Improved |
| include intent and affordances in the offline graph (ablation) | G2F-RAG on Qwen2.5-VL-7B · MLVU | accuracy (%): 70.2 -> 73.4 | Improved |
| enable external tools during graph construction (ablation) | G2F-RAG on Qwen2.5-VL-7B · WildVideo | accuracy (%): 52.4 -> 55.4 | Improved |
Closest papers by meaning
- Video-of-Thought: Step-by-Step Video Reasoning from Perception to Cognition (ICML 2024 · similarity 0.89)
- Navigating Massive Visual Context in Retrieval-Augmented Generation via Multimodal Memory Graph (ICML 2026 · similarity 0.89)
- Vgent: Graph-based Retrieval-Reasoning-Augmented Generation For Long Video Understanding (NeurIPS 2025 · similarity 0.89)
- SciVideoBench: Benchmarking Scientific Video Reasoning in Large Multimodal Models (ICML 2026 · similarity 0.88)
- When Thinking Drifts: Evidential Grounding for Robust Video Reasoning (NeurIPS 2025 · similarity 0.88)
- SOK-Bench: A Situated Video Reasoning Benchmark with Aligned Open-World Knowledge (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.
