Research map / Video / Video-language understanding
Towards Sparse Video Understanding and Reasoning
Chenwei Xu, Zhen Ye, Shang Wu, Weijian Li, Zihan Wang, Zhuofan Xia, Lie Lu, Pranav Maneriker, Fan Du, Manling Li, Han Liu
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 state carryover (summary-as-state) (ablation) | ReViSe · VideoEspresso | accuracy (%): 23.14 -> 41.48 | Improved |
| add structured P/O/H/U/R summary fields (ablation) | ReViSe · VideoEspresso | accuracy (%): 24.27 -> 41.48 | Improved |
| add both state carryover and structured fields (ablation) | ReViSe · VideoEspresso | accuracy (%): 20.24 -> 41.48 | Improved |
| allow 4 turns instead of 1 (ablation) | ReViSe · VideoEspresso | accuracy (%): 38.3 -> 42.1 | Improved |
| allow 4 turns instead of 1 (ablation) | ReViSe · VideoEspresso | frames used: 4.60 -> 2.89 | Improved |
| balanced rounds-frames allocation (5x4) instead of skewed 10x2 (ablation) | ReViSe · VideoEspresso | accuracy (%): 63.3 -> 66.9 | Improved |
| increase max rounds from 5 to 7 (4 frames each) (ablation) | ReViSe · VideoEspresso | accuracy (%): 66.9 -> 67.4 | Mixed |
| reinforcement fine-tuning (EAGER GRPO) instead of plug-and-play | ReViSe Qwen2.5-VL-3B · NExT-QA | accuracy (%): 31.7 -> 51.3 | Improved |
| reinforcement fine-tuning instead of supervised format fine-tuning | ReViSe Qwen2.5-VL-3B · VideoEspresso | accuracy (%): 21.3 -> 27.8 | Improved |
| drop confidence-gain term (lambda1 = 0) (ablation) | EAGER reward · NExT-QA | accuracy (%): 44.4 -> 46.4 | No clear effect |
| add correct-and-early-stop term (lambda3 = 0.5) (ablation) | EAGER reward · NExT-QA | accuracy (%): 39.9 -> 44.4 | Improved |
| add ReViSe to GPT-4o | GPT-4o · VideoEspresso | accuracy (%): 26.4 -> 48.9 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 3 on most shared results.
- VideoTree — VideoTree: Adaptive Tree-based Video Representation for LLM Reasoning on Long Videos (CVPR 2025) · better on 0 of 2 shared results
- VideoAgent — VideoAgent: A Memory-augmented Multimodal Agent for Video Understanding (ECCV 2024) · better on 1 of 2 shared results
- MDP3 — MDP3: A Training-free Approach for List-wise Frame Selection in Video-LLMs (ICCV 2025) · 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
- VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · similarity 0.91)
- Flexible Frame Selection for Efficient Video Reasoning (CVPR 2025 · similarity 0.91)
- Not Another Text Benchmark: Putting the “Visual" Back in Visual Question Answering for Large Video Models (NeurIPS 2026 · similarity 0.90)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.90)
- LongVideoBench: A Benchmark for Long-context Interleaved Video-Language Understanding (NeurIPS 2024 · similarity 0.89)
- MoReVQA: Exploring Modular Reasoning Models for Video Question Answering (CVPR 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.
