Research map / Label-efficient and robust learning / Test-time adaptation
Specialization after Generalization: Towards Understanding Test-Time Training in Foundation Models
Jonas Hübotter, Patrik Wolf, Aleksandr Shevchenko, Dennis Jüni, Andreas Krause, Gil Kur
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 test-time training (local last-layer fine-tuning on 50 neighbors) | global linear classifier · ImageNet-1k (CLIP ViT-B/32) | accuracy: 78.33 -> 79.39 | Improved |
| add test-time training (local last-layer fine-tuning) | LeNet-5 · MNIST | accuracy: 98.57 -> 99.01 | Improved |
| add TTT on SAE concept features | global linear on SAE features · ImageNet-1k | accuracy: 71.45 -> 72.64 | Improved |
| increase model size (ablation) | TTT vs global gap · MNIST, ImageNet, Pile | error / bits per byte | Mixed |
| increase training set size (ablation) | TTT vs global gap · MNIST and ImageNet | error rate | Improved |
| apply a single local TTT head globally (ablation) | TTT heads · MNIST | accuracy: 98.57 -> 36.38 | Hurt |
| apply a single local TTT head globally (ablation) | TTT heads · ImageNet-1k | accuracy: 78.33 -> 77.04 | Hurt |
| use non-adaptive mask instead of learned sparse mask (ablation) | sparse TTT · ImageNet-1k (SAE) | accuracy: 72.56 -> 71.51 | Hurt |
| add learned sparse concept mask (~40 concepts) (ablation) | TTT on SAE reconstructions · ImageNet-1k | accuracy | No clear effect |
| use TTT end-to-end instead of last layer only (ablation) | MNIST classifier · MNIST | accuracy | No clear effect |
| compare TTT to majority vote over neighborhood (ablation) | TTT · image classification with many classes | error | Improved |
Closest papers by meaning
- The Surprising Effectiveness of Test-Time Training for Few-Shot Learning (ICML 2025 · similarity 0.89)
- Test-Time Training Provably Improves Transformers as In-context Learners (ICML 2025 · similarity 0.89)
- Training on the Test Task Confounds Evaluation and Emergence (ICLR 2025 · similarity 0.88)
- In-Place Test-Time Training (ICLR 2026 · similarity 0.88)
- Understanding the Role of Training Data in Test-Time Scaling (ICLR 2026 · similarity 0.88)
- Specialized Foundation Models Struggle to Beat Supervised Baselines (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.
