Research map / Architectures and efficiency / Attention and transformers
CAIT: Triple-Win C ompression Toward High A ccuracy, Fast I nference, and Favorable T ransferability for ViTs
Ao Wang, Hui Chen, Zijia Lin, Sicheng Zhao, Jungong Han, Guiguang Ding
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add asymmetric merging (HTM and VTM alternating) instead of symmetric merging (ablation) | ATME · ImageNet DeiT-Tiny | Top-1: 71.2 -> 71.9 | Improved |
| add head-level consistency (ablation) | CDCP · ImageNet DeiT-Tiny | Top-1: 71.3 -> 72.7 | Improved |
| add attention-level consistency (ablation) | CDCP · ImageNet DeiT-Tiny | Top-1: 72.1 -> 72.7 | Improved |
| add both consistencies (ablation) | CDCP · ImageNet DeiT-Tiny | Top-1: 71.0 -> 72.7 | Improved |
| combine token merging with channel pruning (ablation) | ATME · ImageNet DeiT-Small | Top-1: 79.8 -> 80.2 | Improved |
| fine-tune from pretrained instead of training ATME from scratch (ablation) | ATME · ImageNet DeiT-Tiny | Top-1: 71.1 -> 71.9 | Improved |
| lengthen fine-tuning from 30 to 300 epochs (ablation) | CAIT · ImageNet DeiT-Small | Top-1: 76.7 -> 80.2 | Improved |
| change warmup epochs before raising target ratio from 0 to 30 (ablation) | CDCP · ImageNet DeiT-Tiny | Top-1: 72.26 -> 72.34 | No clear effect |
| change reconstruction interval iterations from 25 to 75 (ablation) | CDCP · ImageNet DeiT-Tiny | Top-1: 72.34 -> 72.46 | No clear effect |
| use HTM only instead of alternating merging (ablation) | ATME · ImageNet DeiT-Tiny | Top-1: 71.9 -> 71.5 | Hurt |
| use structured token merging instead of unstructured pruning for transfer | ATME · ADE20k Semantic FPN | mIoU: 42.7 -> 44.5 | Improved |
| use ATME instead of STViT-R reproduced | ATME · ImageNet Swin-Tiny | Top-1: 80.5 -> 81.1 | Improved |
Later papers that compared with this method
1 later accepted paper reports CaiT as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Revisiting [CLS] and Patch Token Interaction in Vision Transformers (ICLR 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VTC-LFC — VTC-LFC: Vision Transformer Compression with Low-Frequency Components (NeurIPS 2022) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- ToaSt: Token Channel Selection and Structured Pruning for Efficient ViT (ICML 2026 · similarity 0.91)
- Evo-ViT: Slow-Fast Token Evolution for Dynamic Vision Transformer (AAAI 2022 · similarity 0.91)
- Joint Token Pruning and Squeezing Towards More Aggressive Compression of Vision Transformers (CVPR 2023 · similarity 0.90)
- Peeling the Onion: Hierarchical Reduction of Data Redundancy for Efficient Vision Transformer Training (AAAI 2023 · similarity 0.90)
- DeepCompress-ViT: Rethinking Model Compression to Enhance Efficiency of Vision Transformers at the Edge (CVPR 2025 · similarity 0.89)
- Multi-criteria Token Fusion with One-step-ahead Attention for Efficient Vision Transformers (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
