Does it help to add an attention module?
What 467 papers at ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV, ECCV, TPAMI, IJCV and JMLR report when they make this one change.
90% of the papers report a gain; 5% report a loss.
- Improved 90% (421)
- No clear effect 4% (20)
- Mixed 1% (3)
- Hurt 5% (23)
467 papers, 725 reported results. Each paper counts once, with the outcome it reports most often.
Compared with similar changes
This change is a part added to a method (a component). Across all 20 components we track, 82% of papers report a gain and 10% a loss. Papers report gains from added components far more often than from turning a knob: 82% against 65%.
By year of publication
| Papers from | Papers | Improved | No clear effect | Mixed | Hurt |
|---|---|---|---|---|---|
| 2016–2023 | 274 | 91% | 4% | 1% | 5% |
| 2024–2025 | 133 | 90% | 5% | 0% | 5% |
| 2026 | 60 | 88% | 5% | 2% | 5% |
Where it was tried most
- CIFAR100 (4 papers)
- ImageNet-1K (3 papers)
- Market-1501 (3 papers)
See the papers behind these numbers
Describe your own setup on the atlas and "Already tried" lists the papers that made this change closest to yours, with the dataset, the metric and the numbers they report.
Check this change for your setup Free; the paper list needs a sign-in.
How to read this
The results were read by AI from each paper's full text and matched to this change by its wording, so some are missed or misread. These are what papers chose to report, mostly from their own ablations, and published results lean toward gains; a share here is not the chance that the change helps you. All 37 changes
