Research map / Architectures and efficiency / Attention and transformers
Vision-TTT: Efficient and Expressive Visual Representation Learning with Test-Time Training
Quan Kong, Yanru Xiao, Yuhao Shen, Cong 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 |
|---|---|---|---|
| share Q / K projections in TTT layer (ablation) | Vittt-T · ImageNet-1K | Top-1 accuracy: 74.2 -> 74.0 | No clear effect |
| add gating (ablation) | Vittt-T · ImageNet-1K | Top-1 accuracy: 74.0 -> 75.2 | Improved |
| add Conv1d (vanilla TTT) (ablation) | Vittt-T · ImageNet-1K | Top-1 accuracy: 75.2 -> 75.6 | Improved |
| add gating and Conv1d together (ablation) | Vittt-T · ImageNet, COCO, ADE20K | Acc / AP^b / mIoU / throughput: +1.6 acc, +1.5 AP^b, +1.4 mIoU, -14.1% throughput | Mixed |
| add dual dataset strategy (forward and backward) (ablation) | Vittt-T · ImageNet-1K | Top-1 accuracy: 75.6 -> 77.5 | Improved |
| add dual dataset strategy (ablation) | Vittt-T · COCO2017 / ADE20K | AP^b / mIoU: 42.7 vs 41.6 AP^b; 44.1 vs 42.6 mIoU | Improved |
| add dual dataset strategy (ablation) | Vittt-T · Throughput | FPS: 3346 -> 2068 | Hurt |
| add Conv2d dataset preprocessing (ablation) | Vittt-T · ImageNet-1K | Top-1 accuracy: 77.5 -> 77.7 | Improved |
| use MeanPool instead of head class token (ablation) | Vittt-T · ImageNet-1K | Top-1 accuracy: 76.6 -> 77.7 | Improved |
| use MeanPool instead of DoubleClassTok (ablation) | Vittt-T · ImageNet-1K | Top-1 accuracy: 75.4 -> 77.7 | Improved |
| use MeanPool instead of MaxPool (ablation) | Vittt-T · ImageNet-1K | Top-1 accuracy: 77.2 -> 77.7 | Improved |
| vary mini-batch size b from 4 to 196 (ablation) | Vittt-T · ImageNet-1K | Top-1 accuracy | Mixed |
Closest papers by meaning
- Rejuvenating image-GPT as Strong Visual Representation Learners (ICML 2024 · similarity 0.89)
- Vision Transformers with Self-Distilled Registers (NeurIPS 2025 · similarity 0.87)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.87)
- End-to-End Vision Tokenizer Tuning (NeurIPS 2025 · similarity 0.87)
- PerceptionGPT: Effectively Fusing Visual Perception into LLM (CVPR 2024 · similarity 0.87)
- TTRV: Test-Time Reinforcement Learning for Vision Language Models (CVPR 2026 · 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.
