Research map / Architectures and efficiency / Attention and transformers
Linearizing Vision Transformer with Test-Time Training
Yining Li, Dongchen Han, Zeyu Liu, Hanyi Wang, Yulin Wang, Gao Huang
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 |
|---|---|---|---|
| inherit both attention and MLP weights instead of MLP only (ablation) | T5 (DiT-S/2) · ImageNet 256 generation | FID: 89.71 -> 68.52 | Improved |
| inherit both attention and MLP weights instead of attention only (ablation) | T5 (DiT-S/2) · ImageNet 256 generation | FID: 93.18 -> 68.52 | Improved |
| add InstanceNorm on keys (vs none) (ablation) | TTT linearization · DeiT-Tiny ImageNet | Top-1 (%): 0.37 -> 71.19 | Improved |
| use InstanceNorm instead of LayerNorm on keys (ablation) | TTT linearization · DeiT-Tiny ImageNet | Top-1 (%): 57.25 -> 71.19 | Improved |
| add mean subtraction to key InstanceNorm (ablation) | TTT linearization · DeiT-Tiny ImageNet | Top-1 (%): 51.43 -> 71.19 | Improved |
| add division-by-std to key InstanceNorm (ablation) | TTT linearization · DeiT-Tiny ImageNet | Top-1 (%): 71.15 -> 71.19 | No clear effect |
| add depthwise convolution on queries and keys (vs none) (ablation) | TTT · DeiT-Tiny ImageNet | Top-1 (%): 69.25 -> 71.19 | Improved |
| use DWC on QK instead of DWC on values (ablation) | TTT · DeiT-Tiny ImageNet | Top-1 (%): 70.47 -> 71.19 | Improved |
| add Neighborhood Attention (window 5) to T5 (ablation) | T5 · DeiT-Tiny ImageNet | Top-1 (%): 71.19 -> 72.06 | Improved |
| add NAT (ablation) | T5 (DiT-S/2) · ImageNet 256 generation | FID: 72.98 -> 68.52 | Improved |
| use 2x learning rate for TTT-specific params (ablation) | T5 (DiT-S/2) · ImageNet 256 generation | FID: 68.84 -> 68.52 | Improved |
| raise TTT-parameter learning rate multiplier from 1x to 20x (ablation) | T5 · DeiT-Tiny ImageNet | Top-1 (%): 68.14 -> 71.19 | Improved |
Closest papers by meaning
- Bridging the Divide: Reconsidering Softmax and Linear Attention (NeurIPS 2024 · similarity 0.88)
- MimiQ: Low-Bit Data-Free Quantization of Vision Transformers with Encouraging Inter-Head Attention Similarity (AAAI 2025 · similarity 0.88)
- MonarchAttention: Zero-Shot Conversion to Fast, Hardware-Aware Structured Attention (NeurIPS 2025 · similarity 0.88)
- Attention Transfer Is Not Universally Effective for Vision Transformers (NeurIPS 2026 · similarity 0.87)
- Breaking the Low-Rank Dilemma of Linear Attention (CVPR 2025 · similarity 0.87)
- Token Statistics Transformer: Linear-Time Attention via Variational Rate Reduction (ICLR 2025 · 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.
