Research map / Architectures and efficiency / Attention and transformers
Random Wins All: Rethinking Grouping Strategies for Vision Tokens
Qihang Fan, Yuang Ai, Huaibo Huang, Ran He
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 |
|---|---|---|---|
| replace Swin window grouping with random grouping (and CPE) | Swin-T · ImageNet-1K | top-1 accuracy: 81.3 -> 82.7 | Improved |
| replace Quadtree grouping with random grouping | Quadtree-b2 · ImageNet-1K | throughput: 467 -> 1926 imgs/s | Improved |
| replace Swin grouping with random grouping | Swin-T Mask R-CNN · COCO val2017 | box AP: 43.7 -> 46.0 | Improved |
| random grouping in PTv3 | PTv3 · ScanNet val | mIoU / latency: mIoU 77.6 -> 77.8; latency 88 -> 68 ms | Improved |
| apply random token-order pattern in LLM | LLaVA-1.5 · POPE | accuracy: 86.1 -> 87.4 | Improved |
| random pattern on LLaVA-1.5 | LLaVA-1.5 · VQAv2 | accuracy: 78.5 -> 78.5 | No clear effect |
| keep positional encoding (ablation) | Random-Swin-T · ImageNet-1K | top-1 accuracy: 79.3 -> 82.7 | Improved |
| use a different random tensor per head (ablation) | Random-Swin-T · ImageNet-1K | top-1 accuracy: 80.5 -> 82.7 | Improved |
| keep global receptive field in random grouping (ablation) | Random-DeiT-S · ImageNet-1K | top-1 accuracy: 79.3 -> 80.9 | Improved |
| keep global receptive field in random grouping (ablation) | Random-PVTv2-B1 · ImageNet-1K | top-1 accuracy: 79.3 -> 79.6 | No clear effect |
| fix the random grouping pattern across images (ablation) | Random-Swin-T · ImageNet-1K | top-1 accuracy: 76.4 -> 82.7 | Improved |
| add fixed pattern, then multi P, then CPE to fully random grouping (ablation) | Swin-T · ImageNet-1K | top-1 accuracy: 71.2 -> 77.6 -> 80.1 -> 82.7 | Improved |
Closest papers by meaning
- Perceptual Group Tokenizer: Building Perception with Iterative Grouping (ICLR 2024 · similarity 0.90)
- GPViT: A High Resolution Non-Hierarchical Vision Transformer with Group Propagation (ICLR 2023 · similarity 0.86)
- An Empirical Study of Attention and Diversity for Adaptive Visual Token Pruning in Large Vision-Language Models (ICLR 2026 · similarity 0.85)
- GroupViT: Semantic Segmentation Emerges From Text Supervision (CVPR 2022 · similarity 0.85)
- Exploring Token Pruning in Vision State Space Models (NeurIPS 2024 · similarity 0.85)
- GroupMamba: Efficient Group-Based Visual State Space Model (CVPR 2025 · similarity 0.85)
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.
