Research map / Architectures and efficiency / Attention and transformers
Vision transformer design
Improves vision transformer robustness, efficiency, and training through architectural changes.
30 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Attention and transformers, keyword cluster “vision · vit · transformers”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Few recent reports The tracked benchmarks have at most one report from 2025–26 that could beat the old record, too few to say whether they are still moving.
17 papers in 2023–24 and 6 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 2.9% to 0.5%.
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
- 2021 · Intriguing Properties of Vision Transformers (NeurIPS 2021 · 302 citations)
- 2022 · When Vision Transformers Outperform ResNets without Pre-training or Strong Data Augmentations (ICLR 2022 · 102 citations)
- 2023 · Conditional Positional Encodings for Vision Transformers (ICLR 2023 · 406 citations)
- 2024 · You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · 21 citations)
- 2025 · Mamba-Reg: Vision Mamba Also Needs Registers (CVPR 2025 · 19 citations)
- 2026 · From Per-Image Low-Rank to Encoding Mismatch: Rethinking Feature Distillation in Vision Transformers (ICML 2026 · 0 citations)
Earliest papers here
- Intriguing Properties of Vision Transformers (NeurIPS 2021 · 302 citations)
- IA-RED$^2$: Interpretability-Aware Redundancy Reduction for Vision Transformers (NeurIPS 2021 · 68 citations)
- Learning to Iteratively Solve Routing Problems with Dual-Aspect Collaborative Transformer (NeurIPS 2021 · 24 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.
ImageNet-1K · Top-1 accuracy (%) (higher is better)
5 papers, 2021–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 80.9 | IA-RED$^2$: Interpretability-Aware Redundancy Reduction for Vision Transformers (NeurIPS 2021 · 68 citations) |
| 2023 | 81.3 | Masked Jigsaw Puzzle: A Versatile Position Embedding for Vision Transformers (CVPR 2023 · 19 citations) |
| 2024 | 83.4 | Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · 17 citations) |
What has been tried
27 papers here report 330 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 8%, hurt 23%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Replace Less-Attention layers with MHSA | ImageNet-1K · Top-1 accuracy | 79.2 -> 78.7; 82.6 -> 82.0 | Hurt | You Only Need Less Attention at Each Stage in Vision Transformers (CVPR 2024 · 21 citations) |
| Use reduce-and-concatenate register prediction instead of one register or the mean, with registers at the beginning, middle, or evenly spaced | ImageNet-1k · accuracy (%) | 81.3 -> 82.1; 81.8 -> 82.6; 81.7 -> 83.0; 81.4 -> 82.1; 82.0 -> 82.6; 82.2 -> 83.0 | Helped | Mamba-Reg: Vision Mamba Also Needs Registers (CVPR 2025 · 19 citations) |
| Add similarity and informativeness criteria to the size criterion | ImageNet-1K · Top-1 accuracy (%) | 79.8% -> 80.1% | Helped | Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · 17 citations) |
| Train MViTv2-S with bipartite matching instead of RoI align on precomputed boxes | AVA 2.2 · mAP | 26.8 -> 30.0 | Helped | Multiscale Vision Transformers Meet Bipartite Matching for Efficient Single-stage Action Localization (CVPR 2024 · 12 citations) |
| Add Mixup to Cutmix 1.0 | ImageNet-1K · top-1 accuracy (%) | ViT-S/16: 79.540 -> 78.069 (Mixup 0.2), 78.128 (Mixup 0.4); ViT-B/16: 80.240 -> 79.448 (Mixup 0.2), 79.248 (Mixup 0.4) | Hurt | Configuring Data Augmentations to Reduce Variance Shift in Positional Embedding of Vision Transformers (AAAI 2025 · 1 citation) |
| Use VCA in Stage I only | ImageNet-1K · Top-1 accuracy | 75.1 -> 75.5 | Helped | Linear Differential Vision Transformer: Learning Visual Contrasts via Pairwise Differentials (NeurIPS 2025 · 1 citation) |
| add token calibration alone | CIFAR-100 · accuracy | +0.05 | Helped | ViT-Calibrator: Decision Stream Calibration for Vision Transformer (AAAI 2024 · 1 citation) |
| Aggregate model outputs instead of intermediate features | NYU-V2 · RMSE | 0.323 -> 0.314 | Hurt | Sub-token ViT Embedding via Stochastic Resonance Transformers (ICML 2024 · 1 citation) |
| use the reported neighbor-repelling penalty settings across λ values | ImageNet-100 · accuracy (%) | 70.24; 76.20; 78.26; 77.89; 72.17 | Mixed | Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · 0 citations) |
| Keep the key sequence uncompressed while applying token dropout to the query | ImageNet-100, 100M budget · NN accuracy | 65.4 -> 69.6 | Helped | Accelerating Augmentation Invariance Pretraining (NeurIPS 2024 · 0 citations) |
| Use two-step sampling rather than uniform configuration sampling | ImageNet-1K, around 1.6G FLOPs · Top-1 accuracy | nearly 1% | Helped | ELSA: Exploiting Layer-wise N:M Sparsity for Vision Transformer Acceleration (NeurIPS 2024) |
What reviewers will expect
Datasets and baselines named in the 13 papers of this approach from 2024–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ImageNet-1K · 9 of 13 papers
- ADE20K · 5 of 13 papers
- CIFAR-100 · 2 of 13 papers
- ImageNet-100 · 2 of 13 papers
Baselines
- DeiT · 3 of 13 papers
- EViT · 2 of 13 papers
- DynamicViT · 2 of 13 papers
- PVT · 2 of 13 papers
- MoCo-V3 · 2 of 13 papers
- DINO · 2 of 13 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 23 papers; from OpenAlex and the papers' affiliations).
- Microsoft · 2 papers
Use it
Download BibTeX of 29 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
- Transformer expressivity and learnability (217 papers)
- Attention head mechanisms (75 papers)
- Training dynamics of in-context learning (62 papers)
Elsewhere in the topic
- State space and recurrent sequence models (172 papers)
- Sparse and structured attention (163 papers)
- Sparse attention for long-context LLMs (147 papers)
- Attention head and block redesign (127 papers)
- Efficient and robust vision transformers (124 papers)
- Efficient language transformers (105 papers)
- Language model pretraining and induction (99 papers)
- Hybrid and interpretable ViT designs (94 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).