Research map / Video / Video-language understanding
Event and anomaly reasoning in video
Reasons about events, anomalies and object state changes in video with datasets and grounded reasoning.
184 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Video-language understanding, keyword cluster “video understanding · question · long”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 1 tracked benchmark.
30 papers in 2023–24 and 130 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 13.5% to 17.9%.
Open questions the data points to
- UCF-Crime (AUC (%)): best went from 85.3 (up to 2024) to 88.96 in 2025 — Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity (CVPR 2025 · 31 citations)
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2016 · Qualitative Spatio-Temporal Stream Reasoning with Unobservable Intertemporal Spatial Relations Using Landmarks (AAAI 2016 · 13 citations)
- 2017 · Automatic Understanding of Image and Video Advertisements (CVPR 2017 · 163 citations)
- 2018 · Attend and Interact: Higher-Order Object Interactions for Video Understanding (CVPR 2018 · 144 citations)
- 2019 · Video Relationship Reasoning Using Gated Spatio-Temporal Energy Graph (CVPR 2019 · 99 citations)
- 2020 · CLEVRER: Collision Events for Video Representation and Reasoning (ICLR 2020 · 68 citations)
- 2021 · AGQA: A Benchmark for Compositional Spatio-Temporal Reasoning (CVPR 2021 · 86 citations)
- 2022 · End-to-End Referring Video Object Segmentation With Multimodal Transformers (CVPR 2022 · 167 citations)
- 2023 · FakeSV: A Multimodal Benchmark with Rich Social Context for Fake News Detection on Short Video Platforms (AAAI 2023 · 103 citations)
- 2024 · Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · 89 citations)
- 2025 · Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity (CVPR 2025 · 31 citations)
- 2026 · The Human Brain as a Dynamic Mixture of Expert Models in Video Understanding (ICLR 2026 · 2 citations)
Earliest papers here
- Qualitative Spatio-Temporal Stream Reasoning with Unobservable Intertemporal Spatial Relations Using Landmarks (AAAI 2016 · 13 citations)
- Automatic Understanding of Image and Video Advertisements (CVPR 2017 · 163 citations)
- Checking the Consistency of Combined Qualitative Constraint Networks (AAAI 2017 · 6 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
UCF-Crime · AUC (%) (higher is better)
5 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 85.3 | Towards Surveillance Video-and-Language Understanding: New Dataset Baselines and Challenges (CVPR 2024 · 39 citations) |
| 2025 | 88.96 | Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity (CVPR 2025 · 31 citations) |
What has been tried
104 papers here report 1,071 measured changes (a component added, removed or swapped, compared with a baseline): helped 74%, no clear effect 5%, hurt 19%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add image-text caption cleaning | UCF-Crime test · AUC | −3.8% | Helped | Harnessing Large Language Models for Training-free Video Anomaly Detection (CVPR 2024 · 89 citations) |
| Add event-level training data to clip-level training | Video-level anomaly reasoning · BLEU | 0.351 -> 0.349 | No clear effect | Holmes-VAU: Towards Long-term Video Anomaly Understanding at Any Granularity (CVPR 2025 · 31 citations) |
| Add soft prompt | CUVA Description · MMEval | 78.92 -> 79.65 | Helped | Uncovering What Why and How: A Comprehensive Benchmark for Causation Understanding of Video Anomaly (CVPR 2024 · 26 citations) |
| Add text queries | VideoQA · accuracy | 40.4 -> 42.3 | Helped | OmniViD: A Generative Framework for Universal Video Understanding (CVPR 2024 · 21 citations) |
| replace uniform 16-frame clip segmentation with scene-based segmentation | MLVU · accuracy | 59.8 -> 61.7 | Helped | Dispider: Enabling Video LLMs with Active Real-Time Interaction via Disentangled Perception, Decision, and Reaction (CVPR 2025 · 20 citations) |
| Remove video input | Expert commentary generation · BLEU-4 | 44.9 -> 44.6 | Hurt | ExpertAF: Expert Actionable Feedback from Video (CVPR 2025 · 17 citations) |
| Add SoccerReplay-1988 to supervised classification pretraining | SoccerReplay-test event classification · Acc.@1 | 62.67 -> 68.03 | Helped | Towards Universal Soccer Video Understanding (CVPR 2025 · 17 citations) |
| replace global text context with local text context | UCF-Crime, average across data splits · AUC (%) | 84.10 vs 83.62 | Hurt | Federated Weakly Supervised Video Anomaly Detection with Multimodal Prompt (AAAI 2025 · 16 citations) |
| Pre-train on Touchdown only | Touchdown validation · TC | 1.0% | Helped | VLN-Video: Utilizing Driving Videos for Outdoor Vision-and-Language Navigation (AAAI 2024 · 14 citations) |
| add shared state vocabulary | HowToChange · Prec.@1 (%) | 58.5 -> 60.7 | Helped | Learning Object State Changes in Videos: An Open-World Perspective (CVPR 2024 · 14 citations) |
| halve the number of video frames | ASPIRe, position · R@20 | 13.02 / 0.09 -> 10.13 / 0.08 | Hurt | HIG: Hierarchical Interlacement Graph Approach to Scene Graph Generation in Video Understanding (CVPR 2024 · 12 citations) |
| Add subject/object motion features to a context-only representation | SportsHHI validation HHICls · mAP | 2.07 -> 5.19 | Helped | SportsHHI: A Dataset for Human-Human Interaction Detection in Sports Videos (CVPR 2024 · 12 citations) |
What reviewers will expect
Datasets and baselines named in the 72 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- UCF-Crime · 7 of 72 papers
- XD-Violence · 7 of 72 papers
- Ego4D · 4 of 72 papers
- Ego-Exo4D · 4 of 72 papers
- MSR-VTT · 3 of 72 papers
- CH-SIMS · 2 of 72 papers
- MOSI · 2 of 72 papers
- MOSEI · 2 of 72 papers
Baselines
- GPT-4o · 14 of 72 papers
- Video-LLaVA · 8 of 72 papers
- Gemini-1.5-Pro · 8 of 72 papers
- CLIP · 7 of 72 papers
- Video-ChatGPT · 5 of 72 papers
- VideoLLaMA2 · 5 of 72 papers
- LLaVA-OneVision · 5 of 72 papers
- LAVAD · 5 of 72 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 160 papers; from OpenAlex and the papers' affiliations).
- Meta · 8 papers
- National University of Singapore · 7 papers
- Beijing University of Posts and Telecommunications · 6 papers
- Shanghai Jiao Tong University · 5 papers
- Fudan University · 5 papers
- The Chinese University of Hong Kong · 5 papers
- Google DeepMind · 5 papers
- Carnegie Mellon University · 4 papers
- Shanghai Innovation Institute · 4 papers
- Peking University · 4 papers
Use it
Download BibTeX of 35 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Video LLMs with fine-grained understanding (229 papers)
- Long video understanding (236 papers)
- Video question answering (64 papers)
- Egocentric video understanding (46 papers)
Elsewhere in the topic
- Temporal video grounding (88 papers)
- Video-language pretraining (82 papers)
- Dense video captioning (65 papers)
- Text-video retrieval (59 papers)
- Encoder-decoder video captioning (52 papers)
- Video moment retrieval (41 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).