Research map / Architectures and efficiency / Attention and transformers
Hybrid and interpretable ViT designs
Enhances vision transformers with convolutional features, local attention, and interpretable decoders.
94 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 3 tracked benchmarks.
35 papers in 2023–24 and 32 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 6.0% to 2.7%.
Open questions the data points to
- ImageNet-1K (Top-1 accuracy): best went from 86.1 (up to 2024) to 94.9 in 2025 — SATA: Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers (CVPR 2025 · 3 citations)
No longer reported by papers here: ImageNet-1K val (Top-1 accuracy, last in 2024); ADE20K val (mIoU, 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 · ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases (ICML 2021 · 788 citations)
- 2022 · Scaling Vision Transformers (CVPR 2022 · 834 citations)
- 2023 · Scaling Vision Transformers to 22 Billion Parameters (ICML 2023 · 115 citations)
- 2024 · RMT: Retentive Networks Meet Vision Transformers (CVPR 2024 · 229 citations)
- 2025 · Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking (CVPR 2025 · 56 citations)
- 2026 · Learning Spatial Decay for Vision Transformers (AAAI 2026 · 1 citation)
Earliest papers here
- ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases (ICML 2021 · 788 citations)
- Not All Images are Worth 16x16 Words: Dynamic Transformers for Efficient Image Recognition (NeurIPS 2021 · 19 citations)
- Scaling Vision Transformers (CVPR 2022 · 834 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)
18 papers, 2022–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 84.3 | MiniViT: Compressing Vision Transformers With Weight Multiplexing (CVPR 2022 · 123 citations) |
| 2024 | 86.1 | RMT: Retentive Networks Meet Vision Transformers (CVPR 2024 · 229 citations) |
| 2025 | 94.9 | SATA: Spatial Autocorrelation Token Analysis for Enhancing the Robustness of Vision Transformers (CVPR 2025 · 3 citations) |
ImageNet-1K val · Top-1 accuracy (higher is better)
6 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 83.6 | Fast Vision Transformers with HiLo Attention (NeurIPS 2022 · 114 citations) |
| 2024 | 84.2 | Mobile Attention: Mobile-Friendly Linear-Attention for Vision Transformers (ICML 2024) |
ADE20K val · mIoU (higher is better)
5 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 48.6 | Sequencer: Deep LSTM for Image Classification (NeurIPS 2022 · 53 citations) |
What has been tried
65 papers here report 771 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 8%, hurt 22%, 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 MaSA with vanilla attention | ImageNet-1K · Top-1 accuracy | 82.4 -> 81.6(-0.8) | Hurt | RMT: Retentive Networks Meet Vision Transformers (CVPR 2024 · 229 citations) |
| Add MRFP to plain ViT-S | COCO val2017, Mask R-CNN 1× · box AP | 40.2 -> 41.5 | Helped | ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions (CVPR 2024 · 129 citations) |
| add layer adaptation | UAV123 · AUC | 65.3 -> 65.1 | Hurt | Similarity-Guided Layer-Adaptive Vision Transformer for UAV Tracking (CVPR 2025 · 56 citations) |
| remove ViT-Adapter | COCO val2017 panoptic · PQ | 57.1 -> 56.7 | Hurt | Your ViT is Secretly an Image Segmentation Model (CVPR 2025 · 37 citations) |
| replace SS2D with MS2D at roughly equivalent FLOPs | ImageNet-1K · Top-1 accuracy | 69.6% to 71.9% | Helped | Multi-Scale VMamba: Hierarchy in Hierarchy Visual State Space Model (NeurIPS 2024 · 28 citations) |
| Add OOD distillation | CIFAR-10 LT and CIFAR-100 LT · accuracy | 70.2 -> 84.5; 31.3 -> 48.9 | Helped | DeiT-LT: Distillation Strikes Back for Vision Transformer Training on Long-Tailed Datasets (CVPR 2024 · 20 citations) |
| Add a learnable scaling factor | ImageNet · Top-1 accuracy | 6.6 -> 24.4; +17.8 | Helped | Bi-ViT: Pushing the Limit of Vision Transformer Quantization (AAAI 2024 · 16 citations) |
| Pretrain ViTEraser with SegMIM | SCUT-EnsText · PSNR | 36.32 -> 36.80 | Helped | ViTEraser: Harnessing the Power of Vision Transformers for Scene Text Removal with SegMIM Pretraining (AAAI 2024 · 16 citations) |
| Replace NGCA with negative NGCA region selection | ImageNet val, focus stage · top-1 accuracy | 80.8 -> 79.7 | Hurt | LF-ViT: Reducing Spatial Redundancy in Vision Transformer for Efficient Image Recognition (AAAI 2024 · 15 citations) |
| Set r0/r1=1/3 instead of the default | ImageNet-1K · Top-1 accuracy | 72.6 -> 72.3 | Hurt | A General and Efficient Training for Transformer via Token Expansion (CVPR 2024 · 7 citations) |
| Use progressive compression instead of top-k compression | ImageNet-1k validation, 15% and 30% MACs reduction · Top-1 accuracy | 82.84 -> 83.25; 67.95 -> 79.28; 83.05 -> 83.25; 78.50 -> 79.28 | Helped | Dense Vision Transformer Compression with Few Samples (CVPR 2024 · 7 citations) |
| add HCS during training | ImageNet val, R-only · accuracy | 29.39 to 68.86 | Helped | Channel Vision Transformers: An Image Is Worth 1 x 16 x 16 Words (ICLR 2024 · 6 citations) |
What reviewers will expect
Datasets and baselines named in the 17 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 · 5 of 17 papers
- COCO · 3 of 17 papers
- ADE20K · 3 of 17 papers
- CIFAR-10 · 2 of 17 papers
- CIFAR-100 · 2 of 17 papers
Baselines
- VIT_Base · 2 of 17 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 67 papers; from OpenAlex and the papers' affiliations).
- Tencent · 4 papers
- Tsinghua University · 3 papers
- Chinese Academy of Sciences · 3 papers
- University of Chinese Academy of Sciences · 3 papers
- Xiamen University · 2 papers
- ETH Zurich · 2 papers
- Northeastern University · 2 papers
- Huawei · 2 papers
- Beihang University · 2 papers
- South China University of Technology · 2 papers
Use it
Download BibTeX of 37 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
- Efficient and robust vision transformers (124 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).