Research map / Architectures and efficiency / Attention and transformers
Vision transformer pruning
Prunes ViT depth, tokens, and channels using saliency and structure-aware optimization.
26 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 Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
12 papers in 2023–24 and 8 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 2.1% to 0.7%.
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 · Chasing Sparsity in Vision Transformers: An End-to-End Exploration (NeurIPS 2021 · 86 citations)
- 2022 · Width & Depth Pruning for Vision Transformers (AAAI 2022 · 94 citations)
- 2023 · Joint Token Pruning and Squeezing Towards More Aggressive Compression of Vision Transformers (CVPR 2023 · 73 citations)
- 2024 · IRPruneDet: Efficient Infrared Small Target Detection via Wavelet Structure-Regularized Soft Channel Pruning (AAAI 2024 · 155 citations)
- 2025 · Elastic ViTs from Pretrained Models without Retraining (NeurIPS 2025 · 0 citations)
- 2026 · Rethinking Depth Pruning for Vision Transformers: A Heterogeneity-Aware Perspective (ICML 2026 · 0 citations)
Earliest papers here
- Chasing Sparsity in Vision Transformers: An End-to-End Exploration (NeurIPS 2021 · 86 citations)
- Width & Depth Pruning for Vision Transformers (AAAI 2022 · 94 citations)
- Unified Visual Transformer Compression (ICLR 2022 · 29 citations)
What has been tried
19 papers here report 199 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 5%, hurt 25%, 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 |
|---|---|---|---|---|
| Use WCP instead of l1-norm channel importance pruning | NUAA-SIRST · IoU | 67.18 -> 74.25 | Helped | IRPruneDet: Efficient Infrared Small Target Detection via Wavelet Structure-Regularized Soft Channel Pruning (AAAI 2024 · 155 citations) |
| use bi-mask weight sharing instead of ordinal sharing | ImageNet-1K, 1B FLOPs search ablation · Top-1 accuracy | 75.8 -> 76.1 | Helped | Once for Both: Single Stage of Importance and Sparsity Search for Vision Transformer Compression (CVPR 2024 · 8 citations) |
| Add node masks alongside edge masks | Circuit discovery at high sparsity · sparsity | Helped | Finding Transformer Circuits With Edge Pruning (NeurIPS 2024 · 4 citations) | |
| Add pruning-aware hidden state alignment | ImageNet-1K val · top-1 accuracy | 75.4 -> 78.8 | Helped | Exploring Token Pruning in Vision State Space Models (NeurIPS 2024 · 4 citations) |
| Add cross-network interactions between FFN blocks and attention heads rather than modeling none | 50% sparsity · Avg. k-NN accuracy | 56.6 -> 63.5; +6.9% | Helped | Elastic ViTs from Pretrained Models without Retraining (NeurIPS 2025 · 0 citations) |
| Use SPP instead of DessiLBI training from scratch | ImageNet-1k · Top-1 accuracy | 78.9 -> 80.2 | Helped | Adaptive Pruning of Pretrained Transformer via Differential Inclusions (ICLR 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 9 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 · 4 of 9 papers
- CIFAR-10 · 3 of 9 papers
- CIFAR-100 · 2 of 9 papers
- ADE20K · 2 of 9 papers
- C4 · 2 of 9 papers
- Wikitext2 · 2 of 9 papers
Baselines
- ViTAS · 2 of 9 papers
- ViT-Slim · 2 of 9 papers
- S2ViTE · 2 of 9 papers
- WDPruning · 2 of 9 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 20 papers; from OpenAlex and the papers' affiliations).
- NVIDIA · 2 papers
- Northeastern University · 2 papers
- Meta · 2 papers
Use it
Download BibTeX of 26 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)
- Hybrid and interpretable ViT designs (94 papers)
- Token merging and masked pretraining (82 papers)
- Theory of self-attention dynamics (43 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).