Research map / Label-efficient and robust learning / Test-time adaptation
ClusT3: Information Invariant Test-Time Training
Gustavo A. Vargas Hakim, David Osowiechi, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Ismail Ben Ayed, Christian Desrosiers
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 |
|---|---|---|---|
| add multi-head clustering (15 projectors per layer vs 1) (ablation) | ClusT3 on layers 1-2 · CIFAR-10-C Level 5 (average) | accuracy (%): 77.97 -> 82.08 | Improved |
| increase heads from 15 to 20 per layer (ablation) | ClusT3 on layers 1-2 · CIFAR-10-C Level 5 (average) | accuracy (%): 82.08 -> 81.73 | Hurt |
| place projectors on early layers (1-2) rather than later layers (3-4) (ablation) | ClusT3 · CIFAR-10-C Level 5 (Gaussian Noise) | accuracy (%): 65.21 -> 71.36 | Improved |
| add projectors on all four layers instead of only layers 1-2 (ablation) | ClusT3 · CIFAR-10-C Level 5 (Gaussian Noise) | accuracy (%): 71.36 -> 68.71 | Hurt |
| use K=10 clusters rather than K=100 (ablation) | ClusT3 · CIFAR-10-C Level 5 (average) | accuracy (%): 79.77 -> 80.94 | Improved |
| vary number of clusters K from 2 to 50 (ablation) | ClusT3 · CIFAR-10-C Level 5 (average) | accuracy (%): 80.39 (K=2) vs 80.94 (K=10) vs 80.70 (K=50) | No clear effect |
| apply ClusT3 test-time adaptation vs no adaptation | ResNet50 · CIFAR-10-C Level 5 (average) | accuracy (%): 53.82 -> 82.08 | Improved |
| adapt with ClusT3 under small natural shift | ResNet50 · CIFAR-10.1 | accuracy (%): 88.45 -> 85.57 | Hurt |
| adapt with ClusT3 (5 heads) under small natural shift | ResNet50 · CIFAR-10.1 | accuracy (%): 88.45 -> 87.43 | Hurt |
| apply ClusT3 test-time adaptation on sim-to-real shift | ResNet50 · VisDA-C | accuracy (%): 46.31 -> 61.91 | Improved |
| use ClusT3 instead of TTT auxiliary task for joint training | TTT-style joint training · CIFAR-10 joint training time per epoch | minutes per epoch: 12.4941 -> 2.7947 | Improved |
| ClusT3-H15 vs TTT per-corruption | CIFAR-10-C Level 5 · CIFAR-10-C Level 5 (Contrast) | accuracy (%): 88.58 (TTT) vs 84.40 (ClusT3-H15) | Mixed |
Later papers that compared with this method
2 later accepted papers report ClusT3 as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.
- Synchronizing Task Behavior: Aligning Multiple Tasks during Test-Time Training (ICCV 2025 · better on 3 of 3 shared results)
- NC-TTT: A Noise Constrastive Approach for Test-Time Training (CVPR 2024 · better on 3 of 3 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TENT — Tent: Fully Test-Time Adaptation by Entropy Minimization (ICLR 2021) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- NC-TTT: A Noise Constrastive Approach for Test-Time Training (CVPR 2024 · similarity 0.91)
- BoostAdapter: Improving Vision-Language Test-Time Adaptation via Regional Bootstrapping (NeurIPS 2024 · similarity 0.86)
- Efficient Test-time Adaptation through Candidate Verification and Divergence Shifts (NeurIPS 2026 · similarity 0.86)
- Test-Time Adaptive Object Detection with Foundation Model (NeurIPS 2025 · similarity 0.85)
- Self-Bootstrapping for Versatile Test-Time Adaptation (ICML 2025 · similarity 0.85)
- Feature-Based Instance Neighbor Discovery: Advanced Stable Test-Time Adaptation in Dynamic World (NeurIPS 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.
