Research map / Architectures and efficiency / Attention and transformers
Statistically Calibrated Scaling for Token Merging in Transformers
Qing Zhou, Hongyuan Zhang, Tao Yang, Junyu Gao, Qi Wang
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 sqrt r scaling | ToMe ViT (vs PA) · ImageNet-1K, r=0.75 | top-1 accuracy: +26.3% (vs PA) | Improved |
| add sqrt r scaling | ToMe ViT (vs Vanilla) · ImageNet-1K, r=0.75 | top-1 accuracy: +11.6% (vs Vanilla) | Improved |
| add sqrt r scaling | DCT pruning · ImageNet-1K, r=0.75 | top-1 accuracy: +6.4% (vs Vanilla) | Improved |
| add sqrt r scaling | ToFu · ImageNet-1K, r=0.75 | top-1 accuracy: +8.4% (vs PA) | Improved |
| add sqrt r scaling | DiffRate · ImageNet-1K, r=0.75 | top-1 accuracy: +10.1% (vs PA) | Improved |
| add proportional attention (PA) | ToMe ViT · ImageNet-1K, r=0.75 | top-1 accuracy | Hurt |
| add sqrt r scaling | PiToMe ALBEF · Flickr30k retrieval, r=0.90 | Rsum: 540.9 -> 549.1 | Improved |
| add sqrt r scaling | PiToMe BLIP · Flickr30k retrieval, r=0.90 | Rsum: 564.7 -> 567.8 | Improved |
| add sqrt r scaling | PiToMe BLIP · Flickr30k retrieval, r=0.75 | Rsum: +166.4 (vs PA) | Improved |
| add proportional attention (PA) | PiToMe BLIP2 · MS COCO retrieval, r=0.95 | Rsum: 501.8 -> 495.4 | Hurt |
| add sqrt r scaling | ToMe LLaVA-1.5 · VQA POPE, r=0.9 | accuracy: 82.88 -> 85.60 | Improved |
| add proportional attention (PA) | ToMe LLaVA-1.5 · VQA GQA, r=0.9 | accuracy: 55.38 -> 52.61 | Hurt |
Closest papers by meaning
- Efficient Time Series Processing for Transformers and State-Space Models through Token Merging (ICML 2025 · similarity 0.89)
- TokenFormer: Rethinking Transformer Scaling with Tokenized Model Parameters (ICLR 2025 · similarity 0.88)
- Accelerating Transformers with Spectrum-Preserving Token Merging (NeurIPS 2024 · similarity 0.88)
- Frequency-Aware Token Reduction for Efficient Vision Transformer (NeurIPS 2025 · similarity 0.87)
- Don't Look Twice: Faster Video Transformers with Run-Length Tokenization (NeurIPS 2024 · similarity 0.87)
- A General and Efficient Training for Transformer via Token Expansion (CVPR 2024 · similarity 0.87)
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.
