Research map / Label-efficient and robust learning / Test-time adaptation
TEA: Test-time Energy Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Test-time Energy Adaptation (TEA), which reinterprets a trained classifier's logits as an energy-based model and adapts normalization layers on test data via contrastive divergence with Langevin-sampled negatives.
Problem
Test-time adaptation under covariate shift without access to training data or training process; existing TTA methods (normalization, entropy, pseudo-label based) can harm calibration and suffer confirmation bias.
Approach
Treat the classifier as an energy-based model using negative log-sum-exp of logits as energy, and adapt normalization-layer parameters with contrastive divergence (test samples as positives, SGLD samples as negatives) so the model's distribution aligns with the test distribution.
Key results
- TEA outperforms the best-performing baseline by an average of 4.7% accuracy at the most severe corruption level on WRN-28-10 BatchNorm across CIFAR-10(C), CIFAR-100(C), and TinyImageNet(C).
- On PACS with ResNet-18, TEA improves over the best baseline by 7.1% (photo source) and 8.5% (art source) per the paper's text.
- On CIFAR-10 calibration, TEA improves ECE by 2.43% and MCE by 18.31% relative to the source model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Adapt the classifier with TEA energy-based test-time adaptation | WRN-28-10 on CIFAR-10-C · CIFAR-10-C, severity 5 | accuracy: 81.41 -> 83.34 | Helped |
| Adapt the classifier with TEA energy-based test-time adaptation | WRN-28-10 on CIFAR-100-C · CIFAR-100-C, severity 5 | accuracy: 63.09 -> 65.10 | Helped |
| Adapt the classifier with TEA energy-based test-time adaptation | WRN-28-10 on TinyImageNet-C · TinyImageNet-C, severity 5 | accuracy: 26.31 -> 31.67 | Helped |
| Adapt the classifier with TEA energy-based test-time adaptation | ResNet-50 on CIFAR-10-C · CIFAR-10-C, severity 1-5 | accuracy: 81.98 -> 83.05 | Helped |
| Adapt the classifier with TEA energy-based test-time adaptation | ResNet-50 on CIFAR-100-C · CIFAR-100-C, severity 1-5 | accuracy: 58.31 -> 59.67 | Helped |
| Adapt the classifier with TEA energy-based test-time adaptation | ResNet-50 on TinyImageNet-C · TinyImageNet-C, severity 1-5 | accuracy: 29.25 -> 30.41 | Helped |
| Adapt the classifier with TEA energy-based test-time adaptation | ResNet-18 trained on PACS Photo · PACS targets Art, Cartoon, Sketch | accuracy: 25.80 -> 27.64 | Helped |
| Adapt the classifier with TEA energy-based test-time adaptation | ResNet-18 trained on PACS Art · PACS targets Photo, Cartoon, Sketch | accuracy: 37.97 -> 41.19 | Helped |
| Adapt the classifier with TEA energy-based test-time adaptation | ResNet-18 trained on PACS Cartoon · PACS targets Photo, Art, Sketch | accuracy: 27.84 -> 30.12 | Helped |
| Adapt the classifier with TEA energy-based test-time adaptation | ResNet-18 trained on PACS Sketch · PACS targets Photo, Art, Cartoon | accuracy: 22.99 -> 24.69 | Helped |
| Use TEA instead of the unadapted source model | CIFAR-10 calibration evaluation · CIFAR-10 | ECE: 4.11% -> 4.02% | Helped |
| Use TEA instead of the unadapted source model | CIFAR-10 calibration evaluation · CIFAR-10 | MCE: 57.99% -> 47.37% | Helped |
Surprising findings
- TEA ranks second on TinyImageNet, trailing the best baseline by only 0.1%.
- TENT and SHOT reduce calibration quality below the unadapted source model on CIFAR-10.
Limitations
- Gains on TinyImageNet are not best across all levels (TEA second by 0.1%).
- Relies on Langevin sampling for negatives, adding sampling cost per adaptation step.
Experimental setup
Datasets: CIFAR-10 CIFAR-10-C CIFAR-100 CIFAR-100-C TinyImageNet-200 TinyImageNet-200-C PACS MNIST
Benchmarks: CIFAR-10-C CIFAR-100-C TinyImageNet-200-C PACS
Metrics: accuracy mean corruption error (mCE) expected calibration error (ECE) maximum calibration error (MCE)
Baselines: BN DUA PL SHOT TENT ETA EATA SAR
Builds on: JEM TENT contrastive divergence SGLD
Closest papers by meaning
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.89)
- Unified Entropy Optimization for Open-Set Test-Time Adaptation (CVPR 2024 · similarity 0.88)
- L-TTA: Lightweight Test-Time Adaptation Using a Versatile Stem Layer (NeurIPS 2024 · similarity 0.88)
- Evaluation of Test-Time Adaptation Under Computational Time Constraints (ICML 2024 · similarity 0.88)
- Bayesian Weight Enhancement with Steady-State Adaptation for Test-time Adaptation in Dynamic Environments (ICML 2025 · similarity 0.88)
- Test-Time Adaptation by Causal Trimming (NeurIPS 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.