Research map / Video / Video-language understanding
Bridging VideoQA and Video-Guided Agentic Tasks via Generalized Keyframe Extraction
Sunqi Fan, Qingle Liu, Runqi Yin, Meng-Hao Guo, Shuojin Yang
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 |
|---|---|---|---|
| replace BFS expansion with A* cost-function search (ablation) | TASKER (GPT-4) · EgoSchema subset | accuracy: 64.7 -> 68.0 | Improved |
| combine task-driven and scene-aware costs (A*) versus either alone (ablation) | TASKER · EgoSchema subset | accuracy: 67.0 / 66.8 -> 68.0 | Improved |
| use A* instead of other variants on VG-GUI-Bench | TASKER · VG-GUI-Bench | accuracy: 40.01 -> 40.96 | Improved |
| A* variant on VG-GUI-Bench completion and type accuracy (ablation) | TASKER-A* · VG-GUI-Bench | completion: 71.38 vs 74.39 (Dijkstra) | Mixed |
| combine self-evaluation and summarization confidence by voting (ablation) | TASKER-A* · EgoSchema subset | accuracy: 67.4 / 67.3 -> 68.0 | Improved |
| use GPT-4o as base LLM instead of GPT-4 (ablation) | TASKER · EgoSchema subset | accuracy: 68.0 -> 68.6 | Improved |
| use reasoning models (o3-mini, Deepseek-R1) as base LLM (ablation) | TASKER · EgoSchema subset | accuracy: 68.0 -> 67.3 / 67.6 | Hurt |
| use TASKER keyframe search instead of VideoTree | GPT-4 backbone · NExT-QA average | accuracy: 75.6 -> 77.4 | Improved |
| use TASKER instead of VideoTree on Qwen3-VL | Qwen3-VL · EgoSchema fullset | accuracy: 76.7 -> 77.3 | Improved |
| add 10 uniformly sampled video frames versus no video | frontier MLLMs · VG-GUI-Bench | accuracy: 35.93 -> 39.78 (Seed-2.0-Pro) | Improved |
| feed oracle keyframes instead of selected frames | Qwen3-VL-235B-A22B · VG-GUI-Bench | accuracy: 44.32 vs 40.96 | Mixed |
| add TASKER-selected frames versus uniform frames on OSWorld subset | Gemini-3-Flash · OSWorld subset | task success rate: 49.13 -> 50.63 | Mixed |
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.
- VideoTree — VideoTree: Adaptive Tree-based Video Representation for LLM Reasoning on Long Videos (CVPR 2025) · better on 7 of 7 shared results
- VideoAgent — VideoAgent: A Memory-augmented Multimodal Agent for Video Understanding (ECCV 2024) · better on 4 of 4 shared results
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
- MVBench: A Comprehensive Multi-modal Video Understanding Benchmark (CVPR 2024 · similarity 0.90)
- MAGNET: A Multi-agent Framework for Finding Audio-Visual Needles by Reasoning over Multi-Video Haystacks (NeurIPS 2025 · similarity 0.89)
- Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs (ICLR 2025 · similarity 0.89)
- VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · similarity 0.89)
- Adaptive Keyframe Sampling for Long Video Understanding (CVPR 2025 · similarity 0.89)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · 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.
