Research map / Label-efficient and robust learning / Test-time adaptation
Test-Time Training with Masked Autoencoders
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 test-time training with MAE reconstruction | ViT probe on MAE ViT-L · ImageNet-C L5 gaussian noise | accuracy: 17.4 -> 30.5 | Helped |
| add test-time training with MAE reconstruction | ViT probe on MAE ViT-L · ImageNet-C L5 elastic | accuracy: 31.6 -> 50.7 | Helped |
| add test-time training with MAE reconstruction | ViT probe on MAE ViT-L · ImageNet-C L5 brightness | accuracy: 68.3 -> 69.1 | Helped |
| add test-time training with MAE reconstruction | ViT probe · ImageNet-A | accuracy: 15.3 -> 21.3 | Helped |
| add test-time training with MAE reconstruction | ViT probe · ImageNet-R | accuracy: 31.3 -> 38.9 | Helped |
| add test-time training with MAE reconstruction | ViT probe · Portraits train split 3 | accuracy: 78.2 -> 79.4 | Helped |
| use ViT probing instead of joint training (ablation) | MAE encoder · ImageNet-C L5 gaussian noise | accuracy: 5.8 -> 17.4 | Helped |
| use ViT probing instead of fine-tuning (before TTT) (ablation) | MAE encoder · ImageNet-C L5 | accuracy: 22.0 -> 17.4 (gaussian) | Mixed |
| use TTT-MAE instead of rotation-prediction TTT on rotation-invariant classes | TTT · ImageNet-C rotation-invariant classes | correct images change out of 50: -3 -> 0 (jellyfish, frost) | Helped |
| use normalized pixels as reconstruction target (ablation) | TTT-MAE · ImageNet-C L5 fog | accuracy: 33.3 -> 42.8 | Helped |
| also train decoder during TTT (ablation) | TTT-MAE · ImageNet-C L5 | accuracy: 42.8 -> 43.3 (fog) | No clear effect |
| use mask ratio 75% instead of 90% at TTT (ablation) | TTT-MAE · ImageNet-C L5 glass blur | accuracy: 28.3 -> 31.8 | Helped |
Closest papers by meaning
- Continual-MAE: Adaptive Distribution Masked Autoencoders for Continual Test-Time Adaptation (CVPR 2024 · similarity 0.87)
- Exploring Target Representations for Masked Autoencoders (ICLR 2024 · similarity 0.85)
- IT$^3$: Idempotent Test-Time Training (ICML 2025 · similarity 0.84)
- Learning to Generate Gradients for Test-Time Adaptation via Test-Time Training Layers (AAAI 2025 · similarity 0.84)
- Self-Bootstrapping for Versatile Test-Time Adaptation (ICML 2025 · similarity 0.84)
- Protected Test-Time Adaptation via Online Entropy Matching: A Betting Approach (NeurIPS 2024 · similarity 0.84)
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.