Research map / Architectures and efficiency / Attention and transformers
Efficient and robust vision transformers
Improves vision transformer efficiency, robustness, and registers through adaptive computation and architecture changes.
124 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2021–2026) · an approach in Attention and transformers, keyword cluster “self attention · attention mechanism · softmax”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 6 tracked benchmarks.
41 papers in 2023–24 and 45 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 7.0% to 3.9%.
Open questions the data points to
- ADE20K val (mIoU): best went from 50 (up to 2024) to 53.5 in 2025 — Vision-RWKV: Efficient and Scalable Visual Perception with RWKV-Like Architectures (ICLR 2025 · 10 citations)
What looks settled
- ImageNet-1K (Top-1 accuracy): best up to 2024 86.9, best in 2025–26 85.1 from 5 papers, at most 0.2% better, if at all.
- ADE20K (mIoU): best up to 2024 51.56, best in 2025–26 50.2 from 4 papers, at most 0.2% better, if at all.
No longer reported by papers here: CIFAR-10 (accuracy, last in 2024); CIFAR-100 (accuracy, last in 2024).
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 · Training data-efficient image transformers & distillation through attention (ICML 2021 · 1,042 citations)
- 2022 · MetaFormer Is Actually What You Need for Vision (CVPR 2022 · 1,239 citations)
- 2023 · BiFormer: Vision Transformer With Bi-Level Routing Attention (CVPR 2023 · 1,147 citations)
- 2024 · SHViT: Single-Head Vision Transformer with Memory Efficient Macro Design (CVPR 2024 · 179 citations)
- 2025 · LSNet: See Large, Focus Small (CVPR 2025 · 90 citations)
- 2026 · Vision Transformers Are Circulant Attention Learners (AAAI 2026 · 5 citations)
Earliest papers here
- Training data-efficient image transformers & distillation through attention (ICML 2021 · 1,042 citations)
- CoAtNet: Marrying Convolution and Attention for All Data Sizes (NeurIPS 2021 · 739 citations)
- DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification (NeurIPS 2021 · 313 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)
22 papers, 2021–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 81.6 | Pay Attention to MLPs (NeurIPS 2021 · 32 citations) |
| 2022 | 85.15 | Vision Transformers Are Robust Learners (AAAI 2022 · 213 citations) |
| 2023 | 86.9 | Hiera: A Hierarchical Vision Transformer without the Bells-and-Whistles (ICML 2023 · 61 citations) |
ImageNet-1K val · top-1 accuracy (higher is better)
17 papers, 2021–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 85.2 | Training data-efficient image transformers & distillation through attention (ICML 2021 · 1,042 citations) |
ADE20K · mIoU (higher is better)
10 papers, 2021–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 45.4 | ViTAE: Vision Transformer Advanced by Exploring Intrinsic Inductive Bias (NeurIPS 2021 · 153 citations) |
| 2022 | 51.56 | Expediting Large-Scale Vision Transformer for Dense Prediction without Fine-tuning (NeurIPS 2022 · 14 citations) |
CIFAR-10 · accuracy (higher is better)
6 papers, 2021–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 94.47 | Adder Attention for Vision Transformer (NeurIPS 2021 · 8 citations) |
| 2022 | 99.2 | DearKD: Data-Efficient Early Knowledge Distillation for Vision Transformers (CVPR 2022 · 71 citations) |
CIFAR-100 · accuracy (higher is better)
6 papers, 2021–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 74.49 | Adder Attention for Vision Transformer (NeurIPS 2021 · 8 citations) |
| 2022 | 91.1 | DearKD: Data-Efficient Early Knowledge Distillation for Vision Transformers (CVPR 2022 · 71 citations) |
ADE20K val · mIoU (higher is better)
5 papers, 2021–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 50 | HRFormer: High-Resolution Vision Transformer for Dense Predict (NeurIPS 2021 · 156 citations) |
| 2025 | 53.5 | Vision-RWKV: Efficient and Scalable Visual Perception with RWKV-Like Architectures (ICLR 2025 · 10 citations) |
What has been tried
92 papers here report 1,133 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 9%, hurt 23%, 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 |
|---|---|---|---|---|
| Replace single-head attention with MHSA | ImageNet-1K · Top-1 accuracy | 77.7 -> 77.4 | Helped | SHViT: Single-Head Vision Transformer with Memory Efficient Macro Design (CVPR 2024 · 179 citations) |
| Replace identity token mixing with LS convolution | ImageNet-1K · Top-1 accuracy | 69.3 -> 71.6 | Helped | LSNet: See Large, Focus Small (CVPR 2025 · 90 citations) |
| Vary the number of register tokens (0, 1, 2, 4, 8 or 16) | ImageNet, ADE-20k and NYUd downstream evaluation · performance | Mixed | Vision Transformers Need Registers (ICLR 2024 · 54 citations) | |
| Replace the S2WAT encoder with a Swin-based encoder | MS-COCO/WikiArt style transfer · SSIM | 0.651 -> 0.567 | Hurt | S2WAT: Image Style Transfer via Hierarchical Vision Transformer Using Strips Window Attention (AAAI 2024 · 50 citations) |
| add ORR (spatial Cox process masking, MSE invariance loss) | UAVDT · Prec. | +3.3 | Helped | Learning Occlusion-Robust Vision Transformers for Real-Time UAV Tracking (CVPR 2025 · 36 citations) |
| Add AVTrack-ViT and AVTrack-EVA as additional teachers | Five UAV tracking datasets (Avg.) · precision | 81.7 -> 82.5 -> 83.6 | Helped | Learning Adaptive and View-Invariant Vision Transformer for Real-Time UAV Tracking (ICML 2024 · 34 citations) |
| Add full-precision downsampling layers | CIFAR-10 · Top-1 accuracy | 92.1 -> 95.0 | Helped | BHViT: Binarized Hybrid Vision Transformer (CVPR 2025 · 32 citations) |
| Add gradient information to final-layer attention | ImageNet · SaCo | 0.1835 -> 0.3783 | Helped | On the Faithfulness of Vision Transformer Explanations (CVPR 2024 · 14 citations) |
| add aggregation framework (AF) | ImageNet · Pixel accuracy | 52.94 -> 76.30 | Helped | Token Transformation Matters: Towards Faithful Post-hoc Explanation for Vision Transformer (CVPR 2024 · 13 citations) |
| add Q-Shift instead of no token shift | ImageNet-1K · top-1 accuracy | 71.5 -> 75.1 | Helped | Vision-RWKV: Efficient and Scalable Visual Perception with RWKV-Like Architectures (ICLR 2025 · 10 citations) |
| Remove position embeddings | ImageNet supervised classification · Acc@1 | 82.7 -> 81.2 | Hurt | An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · 10 citations) |
| vary sparsity ratio r | ACDC · Dice | baseline 88.56; 1/2 89.08; 1/4 89.28; 1/8 89.64; 1/16 89.53; 1/32 89.19 | Mixed | DTMFormer: Dynamic Token Merging for Boosting Transformer-Based Medical Image Segmentation (AAAI 2024 · 9 citations) |
What reviewers will expect
Datasets and baselines named in the 25 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ImageNet-1K · 16 of 25 papers
- ADE20K · 9 of 25 papers
- COCO · 4 of 25 papers
- CIFAR-100 · 3 of 25 papers
- CIFAR-10 · 2 of 25 papers
- COCO-2017 · 2 of 25 papers
- MSCOCO · 2 of 25 papers
- CUB-200-2011 · 2 of 25 papers
Baselines
- DeiT · 4 of 25 papers
- Swin Transformer · 3 of 25 papers
- DTEM · 3 of 25 papers
- Evo-ViT · 2 of 25 papers
- EViT · 2 of 25 papers
- PiToMe · 2 of 25 papers
- ATC · 2 of 25 papers
- DPS · 2 of 25 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 86 papers; from OpenAlex and the papers' affiliations).
- Alibaba Group · 4 papers
- Shanghai Jiao Tong University · 3 papers
- Tsinghua University · 3 papers
- National University of Singapore · 3 papers
- Chinese Academy of Sciences · 3 papers
- Zhejiang University · 2 papers
- Baidu · 2 papers
- Sichuan University · 2 papers
- Carnegie Mellon University · 2 papers
- Illinois Institute of Technology · 2 papers
Use it
Download BibTeX of 39 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
- Hybrid and interpretable ViT designs (94 papers)
- Token merging and masked pretraining (82 papers)
- Theory of self-attention dynamics (43 papers)
- Vision transformer pruning (26 papers)
Elsewhere in the topic
- Transformer expressivity and learnability (217 papers)
- 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 language transformers (105 papers)
- Language model pretraining and induction (99 papers)
- KV cache eviction and compression (92 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).