Research map / Continual learning / Domain and prompt-based
RDumb: A simple approach that questions our progress in continual test-time adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| periodically reset the adapted model to pretrained weights (T=1,000) | RDumb · CIN-C | accuracy: 41.8% points to 46.5% points | Helped |
| use periodic resetting rather than EATA's anti-collapse regularization | continual TTA on ResNet-50 · CCC-Medium | accuracy: 35.4 ± 1.0 -> 38.9 ± 1.4 | Helped |
| use periodic resetting rather than EATA's anti-collapse regularization | continual TTA on ResNet-50 · CIN-3DCC | accuracy: 43.6 ± 0.98 -> 45.2 ± 0.15 | Helped |
| evaluate long, smoothly changing corruption streams using CCC instead of CIN-C | continual TTA methods · CCC-Medium and CCC-Hard | accuracy relative to pretrained baseline | Mixed |
| add anti-collapse regularization to ETA's adaptation loss, producing EATA | continual TTA on CIN-C · CIN-C | accuracy: 2% points | Hurt |
| reset every 1,000 steps rather than every 125 steps (ablation) | RDumb · CIN-C-Holdout | accuracy: 42.1 -> 46.7 | Helped |
| reset every 1,000 steps rather than every 500 steps (ablation) | RDumb · CIN-C-Holdout | accuracy: 46.0 -> 46.7 | Helped |
| reset every 2,000 steps rather than every 1,000 steps (ablation) | RDumb · CIN-C-Holdout | accuracy: 46.7 -> 46.4 | Hurt |
| set H0 to 0.3 × ln 10^3 instead of 0.4 × ln 10^3 (ablation) | RDumb · CCC average | accuracy: 32.6 -> 33.1 | Helped |
| raise H0 from 0.4 to 0.7 × ln 10^3 (ablation) | EATA · CCC average | accuracy: 30.8 -> 0.33 | Hurt |
| increase the weight regularizer β from 2,000 to 3,000 (ablation) | EATA with ViT-B16 · CCC-Medium | accuracy: 38.5 -> 27.8 | Hurt |
| use periodic resets to repeatedly restart adaptation | ETA · ImageNet-C holdout noises and severities | accuracy over adaptation: about 12,500 samples; about 200 adaptation steps | Helped |
Closest papers by meaning
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.89)
- Towards Real-World Test-Time Adaptation: Tri-net Self-Training with Balanced Normalization (AAAI 2024 · similarity 0.89)
- SNAP: Low-Latency Test-Time Adaptation with Sparse Updates (NeurIPS 2025 · similarity 0.88)
- ReservoirTTA: Prolonged Test-time Adaptation for Evolving and Recurring Domains (NeurIPS 2025 · similarity 0.88)
- Persistent Test-time Adaptation in Recurring Testing Scenarios (NeurIPS 2024 · similarity 0.88)
- Active Test-Time Adaptation: Theoretical Analyses and An Algorithm (ICLR 2024 · similarity 0.88)
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.