Research map / Label-efficient and robust learning
Test-time adaptation: research map
437 papers on Test-time adaptation in Label-efficient and robust learning, from ICML, NeurIPS, ICLR, CVPR, AAAI, ICCV and ECCV, and the journals TPAMI, IJCV and JMLR (2016–2026), grouped into 2 clusters and 6 approaches. The busiest year so far is 2026.
What it is
A deployed model often meets data unlike its training data (new weather, sensors or users) and its accuracy drops. Test-time adaptation updates the model during deployment using only the unlabeled test data, for example by adjusting normalisation statistics or training on its own confident predictions.
Terms you will see on this page:
- Distribution shift
- a difference between the data a model was trained on and the data it is used on.
- Test-time adaptation
- updating a model at deployment without labels.
- Pseudo-labels
- the model's own predictions used as if they were labels.
Start reading with the most cited papers here:
- AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (ICLR 2020 · 567 citations)
- Real-Time Self-Adaptive Deep Stereo (CVPR 2019 · 294 citations)
Within Label-efficient and robust learning, its share grew from 3.8% in 2023–24 to 5.8% in 2025–26 (110 → 176 papers at ICML, NeurIPS, ICLR, CVPR and AAAI, the venues with data for all four years).
Explore Test-time adaptation in the interactive map
Working on something in this topic? Describe your idea in scime atlas to see which approach it falls under, the closest papers by meaning and how crowded the spot has become.
What changed recently
Approaches whose share of the topic moved most between 2023–24 and 2025–26 (ICML, NeurIPS, ICLR, CVPR and AAAI):
- Pseudo-label test-time adaptation rising: 18% → 38% of the topic (20 → 67 papers)
- Test-time adaptation of vision-language models rising: 20% → 24% of the topic (22 → 43 papers)
- Characterizing out-of-distribution failures falling: 10% → 2% of the topic (11 → 4 papers)
- Extrapolation and single-sample adaptation falling: 17% → 10% of the topic (19 → 17 papers)
Papers from 2025–2026 that are already cited most:
- Sufficient Invariant Learning for Distribution Shift (CVPR 2025 · 3 citations)
Approaches and key papers
Adapting models at test time · 374 papers
Keywords: time, adaptation, test
Approaches in this cluster:
- Test-time adaptation of vision-language models (119 papers)
Adapts CLIP-style models at inference using caches, diffusion and weight averaging. - Realistic and stable test-time adaptation (82 papers)
Tackles layer selection, long-horizon drift and sequential streams in practical test-time adaptation. - Extrapolation and single-sample adaptation (63 papers)
Generalizes to new domains or tasks from single test samples with uncertainty estimates. - Pseudo-label test-time adaptation (110 papers)
Uses pseudo-labels, feedback and proxy data for robust adaptation on noisy streams.
Most cited and most cited since 2024:
- Real-Time Self-Adaptive Deep Stereo (CVPR 2019 · 294 citations)
- Tent: Fully Test-Time Adaptation by Entropy Minimization (ICLR 2021 · 283 citations)
- Efficient Test-Time Adaptation of Vision-Language Models (CVPR 2024 · 70 citations)
- Test-Time Domain Generalization for Face Anti-Spoofing (CVPR 2024 · 59 citations)
Understanding distribution shift · 63 papers
Keywords: shift, distribution, machine
Approaches in this cluster:
- Robust optimization for distribution shift (35 papers)
Trains for robustness via risk extrapolation, selective augmentation and adaptive risk minimization. - Characterizing out-of-distribution failures (28 papers)
Benchmarks and diagnoses generalization failures under distribution shift, including test-time training.
Most cited and most cited since 2024:
- AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (ICLR 2020 · 567 citations)
- WILDS: A Benchmark of in-the-Wild Distribution Shifts (ICML 2021 · 278 citations)
- Any-Shift Prompting for Generalization over Distributions (CVPR 2024 · 8 citations)
- On the Variance of Neural Network Training with respect to Test Sets and Distributions (ICLR 2024 · 7 citations)
Related topics in Label-efficient and robust learning
- Noisy labels (576)
- Zero-/few-shot and meta-learning (1,029)
- OOD and anomaly detection (836)
- Recognition and long-tail (6,306)
- Semi-supervised learning (897)
- Self-supervised and contrastive (1,267)
- Domain adaptation and generalization (2,012)
- Knowledge distillation (783)
