When and Where to Reset Matters for Long-Term Test-Time Adaptation
On the map: Synthetic data and self-improvement for LLMs → Model collapse from synthetic data
Abstract (arXiv)
When continual test-time adaptation (TTA) persists over the long term, errors accumulate in the model and further cause it to predict only a few classes for all inputs, a phenomenon known as model collapse. Recent studies have explored reset strategies that completely erase these accumulated errors. However, their periodic resets lead to suboptimal adaptation, as they occur independently of the actual risk of collapse. Moreover, their full resets cause catastrophic loss of knowledge acquired over time, even though such knowledge could be beneficial in the future. To this end, we propose (1) an Adaptive and Selective Reset (ASR) scheme that dynamically determines when and where to reset, (2) an importance-aware regularizer to recover essential knowledge lost due to reset, and (3) an on-the-fly adaptation adjustment scheme to enhance adaptability under challenging domain shifts. Extensive experiments across long-term TTA benchmarks demonstrate the effectiveness of our approach, particularly under challenging conditions. Our code is available at https://github.com/YonseiML/asr.
Key facts. Cited by 2 papers in the map (data through 7 Oct 2026). Placed in Model collapse from synthetic data, one of 13 directions in Synthetic data and self-improvement for LLMs. Builds on 42 of its 50 references that are in the map. First posted on arXiv 4 Mar 2026.
Builds on
42 of 50 references.
References · 42
- 2015Deep Residual Learning for Image Recognitioncited by 51,809
- 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 29,915
- 2019PyTorch: An Imperative Style, High-Performance Deep Learning Librarycited by 20,160
- 2016Overcoming catastrophic forgetting in neural networkscited by 4,718
- 2019Benchmarking Neural Network Robustness to Common Corruptions and Perturbationscited by 2,857
- 2016On Large-Batch Training for Deep Learning: Generalization Gap and Sharp Minimacited by 1,337
- 2018Moment Matching for Multi-Source Domain Adaptationcited by 1,327
- 2020Do We Really Need to Access the Source Data? Source Hypothesis Transfer for Unsupervised Domain Adaptationcited by 844
- 2020RobustBench: a standardized adversarial robustness benchmarkcited by 659
- 2022Continual Test-Time Domain Adaptationcited by 520
- 2018Progress & Compress: A scalable framework for continual learningcited by 516
- 2022Efficient Test-Time Model Adaptation without Forgettingcited by 467
30 more
- 2020Improving robustness against common corruptions by covariate shift adaptationcited by 427
- 2023Towards Stable Test-Time Adaptation in Dynamic Wild Worldcited by 409
- 2021MEMO: Test Time Robustness via Adaptation and Augmentationcited by 376
- 2022Mitigating Neural Network Overconfidence with Logit Normalizationcited by 251
- 2020Universal Source-Free Domain Adaptationcited by 219
- 2023Robust Test-Time Adaptation in Dynamic Scenarioscited by 185
- 2022NOTE: Robust Continual Test-time Adaptation Against Temporal Correlationcited by 150
- 2021The Norm Must Go On: Dynamic Unsupervised Domain Adaptation by Normalizationcited by 131
- 2018Don't forget, there is more than forgetting: new metrics for Continual Learningcited by 113
- 2021Understanding and Improving Early Stopping for Learning with Noisy Labelscited by 112
- 2022Robust Mean Teacher for Continual and Gradual Test-Time Adaptationcited by 102
- 2022Test-Time Adaptation via Conjugate Pseudo-labelscited by 98
- 2022Decorate the Newcomers: Visual Domain Prompt for Continual Test Time Adaptationcited by 98
- 2024Test-Time Model Adaptation with Only Forward Passescited by 74
- 2022Improving Test-Time Adaptation via Shift-agnostic Weight Regularization and Nearest Source Prototypescited by 67
- 2023Universal Test-time Adaptation through Weight Ensembling, Diversity Weighting, and Prior Correctioncited by 65
- 2023SoTTA: Robust Test-Time Adaptation on Noisy Data Streamscited by 63
- 2023ViDA: Homeostatic Visual Domain Adapter for Continual Test Time Adaptationcited by 58
- 2023RDumb: A simple approach that questions our progress in continual test-time adaptationcited by 50
- 2022A Probabilistic Framework for Lifelong Test-Time Adaptationcited by 44
- 2025Model Adaptation: Unsupervised Domain Adaptation without Source Datacited by 34
- 2023A Path to Simpler Models Starts With Noisecited by 33
- 2024DPCore: Dynamic Prompt Coreset for Continual Test-Time Adaptationcited by 33
- 2024COME: Test-time adaption by Conservatively Minimizing Entropycited by 31
- 2023Continual-MAE: Adaptive Distribution Masked Autoencoders for Continual Test-Time Adaptationcited by 30
- 2023Persistent Test-time Adaptation in Recurring Testing Scenarioscited by 18
- 2025Ranked Entropy Minimization for Continual Test-Time Adaptationcited by 17
- 2023Layer-wise Auto-Weighting for Non-Stationary Test-Time Adaptationcited by 14
- 2024Distribution Alignment for Fully Test-Time Adaptation with Dynamic Online Data Streamscited by 13
- 2024PALM: Pushing Adaptive Learning Rate Mechanisms for Continual Test-Time Adaptationcited by 8
Built on by
Citing papers per year: 2026: 2.
2 papers.
Similar, not cited
Close in wording, no citation either way
- 2026What Should World Models Forget? Stratified Retention for Continual Adaptation
- 2025ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains
- 2025Adapt in the Wild: Test-Time Entropy Minimization with Sharpness and Feature Regularization
- 2026RAG or Learning? Understanding the Limits of LLM Adaptation under Continuous Knowledge Drift in the Real World
- 2026More Edits, More Stable: Understanding the Lifelong Normalization in Sequential Model Editing
- 2026Boosting ASR Robustness via Test-Time Reinforcement Learning with Audio-Text Semantic Rewards
- 2025Test-Time Distillation for Continual Model Adaptation
- 2024Dual-Phase Continual Learning: Supervised Adaptation Meets Unsupervised Retention