AI research map: 137,217 papers in 20 areas
Every paper from ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR, 2016–2026, placed by what it is about. Papers on similar problems sit close together; colour is the research area.
A sample of 6,484 papers is drawn. Names mark the ten largest areas.
The 20 areas
The largest area is Label-efficient and robust learning with 14,143 papers. Between 2016–2018 and 2024–2026, Large language models gained the most share of all papers (+10.1 points) and Learning theory and algorithms lost the most (-11.2).
| Research area | Papers | Topics | Share 2016–18 | Share 2024–26 | Change |
|---|---|---|---|---|---|
| Label-efficient and robust learning | 14,143 | 9 | 11.8% | 7.8% | -4.0 |
| 3D vision | 11,404 | 6 | 9.5% | 8.2% | -1.3 |
| Learning theory and algorithms | 11,342 | 9 | 16.7% | 5.5% | -11.2 |
| Image generation | 11,161 | 7 | 5.7% | 9.7% | +3.9 |
| Architectures and efficiency | 10,039 | 8 | 7.5% | 7.0% | -0.5 |
| Large language models | 9,881 | 7 | 1.5% | 11.7% | +10.1 |
| Vision-language and multimodal | 9,839 | 6 | 3.7% | 10.2% | +6.5 |
| Probabilistic and scientific modeling | 9,481 | 8 | 8.8% | 6.2% | -2.6 |
| Video | 8,033 | 4 | 6.1% | 6.4% | +0.2 |
| Detection and segmentation | 7,932 | 5 | 9.1% | 3.9% | -5.2 |
| Reinforcement learning | 6,414 | 6 | 3.6% | 4.5% | +0.9 |
| Graph learning | 5,185 | 4 | 2.8% | 3.4% | +0.6 |
| Optimization | 4,371 | 4 | 5.0% | 2.2% | -2.8 |
| Multi-agent systems and games | 3,786 | 4 | 1.7% | 3.3% | +1.6 |
| Adversarial robustness and security | 3,554 | 4 | 1.2% | 2.3% | +1.1 |
| Federated learning and privacy | 2,557 | 4 | 0.6% | 2.1% | +1.4 |
| Point clouds | 2,494 | 4 | 1.1% | 1.5% | +0.5 |
| Bandits and online learning | 2,385 | 4 | 2.2% | 1.2% | -1.0 |
| Time series | 1,620 | 4 | 0.8% | 1.5% | +0.6 |
| Continual learning | 1,596 | 4 | 0.3% | 1.3% | +1.0 |
Share: the area's papers among all papers of those years at all ten venues (each window holds three editions of ICCV and ECCV together). Each area page lists its topics, approaches and most cited papers.
What is an AI research map?
A picture of a research field in which each paper is a point and papers about similar things sit close together. This one places 137,217 papers from ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR (2016–2026) by the meaning of their title and abstract, and groups them into 20 research areas and 111 topics.
How is the map made?
Each paper's title and abstract are turned into a vector by a text embedding model. Papers with similar vectors are placed near each other in two dimensions and grouped into clusters, and each cluster is named from its most distinctive words. Area and topic names are generated automatically and checked by hand.
What can I do with it?
Describe a research idea on the interactive map and see which area and approach it falls in, the closest published papers, how crowded that spot has become since 2016, and what papers reported when they tried a similar change.
More on scime atlas
- All 111 topics and 1,260 approach pages, with benchmark records and key papers
- Does it actually help? What papers report for 37 common changes, from a larger learning rate to dropout
- Accepted papers by venue and year
