Research map / Video / Video-language understanding
Video question answering
Answers questions about videos using memory prompting, sampling, tool use and compositional benchmarks.
64 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 Baselines Checklist Preprints Who Export
Where it stands
Quiet None of the tracked benchmarks has a report from 2025–26.
13 papers in 2023–24 and 35 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 5.8% to 4.8%.
Open questions the data points to
- Papers report HGA on NExT-QA test (accuracy) anywhere from 25.18 to 50.01: comparisons against it depend on whose number you take.
No longer reported by papers here: TGIF-QA FrameQA (accuracy, last in 2021); TGIF-QA Action (accuracy, last in 2021); TGIF-QA Count (MSE, last in 2021); TGIF-QA Transition (accuracy, last in 2021).
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 · MovieQA: Understanding Stories in Movies Through Question-Answering (CVPR 2016 · 681 citations)
- 2018 · Motion-Appearance Co-Memory Networks for Video Question Answering (CVPR 2018 · 261 citations)
- 2019 · ActivityNet-QA: A Dataset for Understanding Complex Web Videos via Question Answering (AAAI 2019 · 267 citations)
- 2020 · Hierarchical Conditional Relation Networks for Video Question Answering (CVPR 2020 · 262 citations)
- 2021 · NExT-QA: Next Phase of Question-Answering to Explaining Temporal Actions (CVPR 2021 · 284 citations)
- 2022 · From Representation to Reasoning: Towards Both Evidence and Commonsense Reasoning for Video Question-Answering (CVPR 2022 · 57 citations)
- 2023 · MIST: Multi-Modal Iterative Spatial-Temporal Transformer for Long-Form Video Question Answering (CVPR 2023 · 88 citations)
- 2024 · Koala: Key Frame-Conditioned Long Video-LLM (CVPR 2024 · 18 citations)
- 2025 · VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · 14 citations)
- 2026 · Vista: Scene-Aware Optimization for Streaming Video Question Answering Under Post-Hoc Queries (AAAI 2026 · 2 citations)
Earliest papers here
- MovieQA: Understanding Stories in Movies Through Question-Answering (CVPR 2016 · 681 citations)
- Motion-Appearance Co-Memory Networks for Video Question Answering (CVPR 2018 · 261 citations)
- ActivityNet-QA: A Dataset for Understanding Complex Web Videos via Question Answering (AAAI 2019 · 267 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.
TGIF-QA FrameQA · accuracy (higher is better)
6 papers, 2018–2021 · not reported since 2021
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 51.5 | Motion-Appearance Co-Memory Networks for Video Question Answering (CVPR 2018 · 261 citations) |
| 2019 | 56.64 | Structured Two-Stream Attention Network for Video Question Answering (AAAI 2019) |
| 2020 | 59.7 | Divide and Conquer: Question-Guided Spatio-Temporal Contextual Attention for Video Question Answering (AAAI 2020 · 106 citations) |
| 2021 | 60.2 | Learning from Inside: Self-driven Siamese Sampling and Reasoning for Video Question Answering (NeurIPS 2021 · 18 citations) |
TGIF-QA Action · accuracy (higher is better)
5 papers, 2019–2021 · not reported since 2021
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 72.34 | Structured Two-Stream Attention Network for Video Question Answering (AAAI 2019) |
| 2020 | 75.9 | Divide and Conquer: Question-Guided Spatio-Temporal Contextual Attention for Video Question Answering (AAAI 2020 · 106 citations) |
| 2021 | 79.7 | Learning from Inside: Self-driven Siamese Sampling and Reasoning for Video Question Answering (NeurIPS 2021 · 18 citations) |
TGIF-QA Count · MSE (lower is better)
5 papers, 2019–2021 · not reported since 2021
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 4.25 | Structured Two-Stream Attention Network for Video Question Answering (AAAI 2019) |
| 2020 | 3.82 | Hierarchical Conditional Relation Networks for Video Question Answering (CVPR 2020 · 262 citations) |
| 2021 | 3.61 | Learning from Inside: Self-driven Siamese Sampling and Reasoning for Video Question Answering (NeurIPS 2021 · 18 citations) |
TGIF-QA Transition · accuracy (higher is better)
4 papers, 2019–2021 · not reported since 2021
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 79.03 | Structured Two-Stream Attention Network for Video Question Answering (AAAI 2019) |
| 2020 | 81.4 | Hierarchical Conditional Relation Networks for Video Question Answering (CVPR 2020 · 262 citations) |
| 2021 | 85.3 | Learning from Inside: Self-driven Siamese Sampling and Reasoning for Video Question Answering (NeurIPS 2021 · 18 citations) |
What has been tried
40 papers here report 431 measured changes (a component added, removed or swapped, compared with a baseline): helped 74%, no clear effect 6%, hurt 18%, 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 the Conditioned Segment tokenizer | EgoSchema · Top 1 Acc (%) | 33.25 -> 36.93 | Helped | Koala: Key Frame-Conditioned Long Video-LLM (CVPR 2024 · 18 citations) |
| use ground-truth CoT | VideoEspresso benchmark · Accuracy | 34.13 -> 72.95 | Helped | VideoEspresso: A Large-Scale Chain-of-Thought Dataset for Fine-Grained Video Reasoning via Core Frame Selection (CVPR 2025 · 14 citations) |
| Remove [CTRL] tokens and use fixed top-k selection | NExT-QA validation subset · accuracy | Hurt | Flexible Frame Selection for Efficient Video Reasoning (CVPR 2025 · 9 citations) | |
| Initialize visual queries with average pooling rather than learnable queries | NExT-QA · accuracy | 67.67 -> 68.61 | Helped | BIMBA: Selective-Scan Compression for Long-Range Video Question Answering (CVPR 2025 · 8 citations) |
| add video context | MSVD-QA · accuracy | 35.7 -> 45.2 | Helped | Video-Context Aligned Transformer for Video Question Answering (AAAI 2024 · 8 citations) |
| Replace EIGV with the proposed hierarchical video aligner | AGQA-Decomp balanced · overall main-question accuracy | Helped | Align and Aggregate: Compositional Reasoning with Video Alignment and Answer Aggregation for Video Question-Answering (CVPR 2024 · 8 citations) | |
| add intermediate supervision | AGQA · Overall accuracy | 50.75 -> 51.14 | Helped | STAIR: Spatial-Temporal Reasoning with Auditable Intermediate Results for Video Question Answering (AAAI 2024 · 5 citations) |
| Add the evidence pool | NExT-QA · Avg. | 71.05 -> 76.14 | Helped | VITED: Video Temporal Evidence Distillation (CVPR 2025 · 3 citations) |
| use STAR's spatiotemporal-interleaved toolchain instead of temporal-then-spatial disentanglement | VideoMME · accuracy | 68.6 -> 70.0 | Helped | Tool-Augmented Spatiotemporal Reasoning for Streamlining Video Question Answering Task (NeurIPS 2025 · 2 citations) |
| add granularity weighting module | MSVD-QA · accuracy | 72.0 -> 73.4 | Helped | Granularity-Adaptive Spatial Evidence Tokenization for Video Question Answering (AAAI 2025 · 1 citation) |
| Use both audio and visual modalities rather than visual only | AVHaystacks-Full · BLEU@4 | 45.28 -> 53.64 | Helped | MAGNET: A Multi-agent Framework for Finding Audio-Visual Needles by Reasoning over Multi-Video Haystacks (NeurIPS 2025 · 1 citation) |
| Replace uniform sampling with external retrieval | QaEgo4D test-mc · accuracy | 53.0 -> 54.2 | Helped | Streaming Video Question-Answering with In-context Video KV-Cache Retrieval (ICLR 2025 · 0 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| HGA | NExT-QA test | accuracy | 25.18 – 50.01 | 2 papers, e.g. AAAI 2023 and CVPR 2021 |
What reviewers will expect
Datasets and baselines named in the 16 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- NExT-QA · 6 of 16 papers
- EgoSchema · 5 of 16 papers
- STAR · 4 of 16 papers
- IntentQA · 3 of 16 papers
- MVBench · 3 of 16 papers
- MSRVTT-QA · 2 of 16 papers
- ActivityNet-QA · 2 of 16 papers
- YouCook2 · 2 of 16 papers
Baselines
- Video-LLaVA · 5 of 16 papers
- SeViLA · 4 of 16 papers
- VideoLLaMA2 · 3 of 16 papers
- GPT-4o · 3 of 16 papers
- Chat-UniVi · 2 of 16 papers
- TimeChat · 2 of 16 papers
- Video-ChatGPT · 2 of 16 papers
- VideoChat2 · 2 of 16 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 48 papers; from OpenAlex and the papers' affiliations).
- Meta · 3 papers
- Hangzhou Dianzi University · 2 papers
- Peking University · 2 papers
- Chinese Academy of Sciences · 2 papers
- Zhejiang University · 2 papers
- Google · 2 papers
- Shanghai Jiao Tong University · 2 papers
- Adobe · 2 papers
- UNC Chapel Hill · 2 papers
- Tsinghua University · 2 papers
Use it
Download BibTeX of 38 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
- Event and anomaly reasoning in video (184 papers)
- Video LLMs with fine-grained understanding (229 papers)
- Long video understanding (236 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).