Research map / Label-efficient and robust learning / Test-time adaptation
Persistent Test-time Adaptation in Recurring Testing Scenarios
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.
Introduces recurring test-time adaptation (revisiting the same domain sequences multiple times) to expose error accumulation, analyzes collapse on an epsilon-perturbed GMM classifier, and proposes PeTTA, which adaptively sets regularization and update rate from a feature-space divergence signal.
Problem
Test-time adaptation methods may degrade over long, recurring data streams where environments reappear, leading to model collapse with error accumulation.
Approach
Define a recurring TTA scenario that repeats practical TTA domain sequences, analyze collapse theoretically on an epsilon-GMM classifier, and adaptively tune the mean-teacher regularization weight and EMA rate using a Mahalanobis-based feature divergence from the source model, plus an anchor loss.
Key results
- In the recurring CIFAR-10-C setting over 20 cycles, RoTTA's running error rises from about 0.25 to about 0.70 while PeTTA stays around 0.25.
- Experiments span four tasks, and the paper reports PeTTA's stability across recurring TTA scenarios and benchmarks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Persistent TTA instead of RoTTA | continual TTA on CIFAR-10-C · CIFAR-10-C, 20th recurrence (step 285) | testing error: 0.68899934569638 -> 0.265042789319986 | Helped |
| randomize corruption order across recurrences instead of using the same order | Persistent TTA and other TTA methods · CIFAR-10-C | testing error | No clear effect |
Surprising findings
- Existing continual TTA methods such as RoTTA degrade below the no-adaptation source model when the same environments recur, even though they perform well in the first passes.
Limitations
- Theoretical analysis is limited to a binary epsilon-perturbed GMM classifier with static label distribution
- Recurring setting is a diagnostic construction and may not reflect all real-world drift
Experimental setup
Datasets: CIFAR-10-C CIFAR-100-C ImageNet-C
Benchmarks: CIFAR-10-C CIFAR-100-C ImageNet-C
Metrics: classification error rate
Baselines: RoTTA CoTTA No TTA LAME CCC
Builds on: RoTTA CoTTA mean teacher practical test-time adaptation
Closest papers by meaning
- Monitoring Risks in Test-Time Adaptation (NeurIPS 2025 · similarity 0.90)
- Evaluation of Test-Time Adaptation Under Computational Time Constraints (ICML 2024 · similarity 0.88)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.88)
- AETTA: Label-Free Accuracy Estimation for Test-Time Adaptation (CVPR 2024 · similarity 0.87)
- Bayesian Weight Enhancement with Steady-State Adaptation for Test-time Adaptation in Dynamic Environments (ICML 2025 · similarity 0.87)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (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.