Research map / Label-efficient and robust learning / Test-time adaptation
TTT++: When Does Self-Supervised Test-Time Training Fail or Thrive?
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 feature alignment (TFA) on top of contrastive TTT (ablation) | TTT-C · CIFAR10-C avg | error (%): 10.7 -> 9.8 | Helped |
| add test-time feature alignment on top of contrastive TTT (ablation) | TTT-C · CIFAR100-C avg | error (%): 36.9 -> 34.1 | Helped |
| add test-time feature alignment on top of contrastive TTT (ablation) | TTT-C · VisDA-C | error (%): 25.72 -> 22.46 | Helped |
| replace rotation SSL with contrastive SSL (SimCLR) (ablation) | TTT-R · CIFAR10-C avg | error (%): 14.3 -> 10.7 | Helped |
| replace rotation SSL with contrastive SSL (SimCLR) (ablation) | TTT-R · CIFAR10.1 | error (%): 11.0 -> 9.7 | Helped |
| add feature alignment to unconstrained TTT under large shift | TTT · inter-twinning moons toy | test accuracy: 50% -> 93% | Helped |
| add dynamic queue of features at small batch size (ablation) | feature alignment · CIFAR100-C fog level 5 | error (%): 40.31 -> 36.02 | Helped |
| use dynamic queue vs single large batch of same sample size (ablation) | feature alignment · CIFAR100-C fog level 5 | error (%): 38.67 -> 39.84 | Hurt |
| add covariance matching to mean matching (ablation) | feature alignment · CIFAR10-C fog | error (%): 19.22 -> 12.82 | Helped |
| add mean matching to covariance matching (ablation) | feature alignment · CIFAR10-C fog | error (%): 12.98 -> 12.82 | Helped |
| align features at both encoder and SSL head instead of encoder only (ablation) | feature alignment · CIFAR10-C fog | error (%): 14.33 -> 12.82 | Helped |
| TTT++ vs BN/TENT on natural shift CIFAR10.1 | BN-based methods · CIFAR10.1 | error (%): 12.1 -> 14.1 | Mixed |
Closest papers by meaning
- Improved Self-Training for Test-Time Adaptation (CVPR 2024 · similarity 0.89)
- NC-TTT: A Noise Constrastive Approach for Test-Time Training (CVPR 2024 · similarity 0.87)
- Test-Time Training Provably Improves Transformers as In-context Learners (ICML 2025 · similarity 0.87)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.86)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.86)
- Domain-Specific Block Selection and Paired-View Pseudo-Labeling for Online Test-Time Adaptation (CVPR 2024 · similarity 0.86)
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.