Research map / Label-efficient and robust learning / Test-time adaptation
Characterizing out-of-distribution failures
Benchmarks and diagnoses generalization failures under distribution shift, including test-time training.
28 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2020–2026) · an approach in Test-time adaptation, keyword cluster “shift · distribution · machine”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
11 papers in 2023–24 and 4 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 10.0% to 2.3%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2020 · AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (ICLR 2020 · 567 citations)
- 2021 · WILDS: A Benchmark of in-the-Wild Distribution Shifts (ICML 2021 · 278 citations)
- 2022 · DDG-DA: Data Distribution Generation for Predictable Concept Drift Adaptation (AAAI 2022 · 52 citations)
- 2023 · Neural Priming for Sample-Efficient Adaptation (NeurIPS 2023 · 3 citations)
- 2024 · Any-Shift Prompting for Generalization over Distributions (CVPR 2024 · 8 citations)
- 2025 · WATCH: Adaptive Monitoring for AI Deployments via Weighted-Conformal Martingales (ICML 2025 · 1 citation)
- 2026 · Mitigating Label Shift in Tabular In-Context Learning via Test-Time Posterior Adjustment (ICML 2026 · 0 citations)
Earliest papers here
- AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (ICLR 2020 · 567 citations)
- Test-Time Training with Self-Supervision for Generalization under Distribution Shifts (ICML 2020 · 70 citations)
- The Effect of Natural Distribution Shift on Question Answering Models (ICML 2020 · 22 citations)
What has been tried
24 papers here report 237 measured changes (a component added, removed or swapped, compared with a baseline): helped 61%, no clear effect 11%, hurt 22%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| use only the training prompt | Open-Office-Home · accuracy | 82.62 vs 84.50 | Hurt | Any-Shift Prompting for Generalization over Distributions (CVPR 2024 · 8 citations) |
| Remove data augmentation | CIFAR-10 · variance between runs | 3.5\times more variance | Hurt | On the Variance of Neural Network Training with respect to Test Sets and Distributions (ICLR 2024 · 7 citations) |
| increase fine-tuning data from 5% to 20% of the left-out subject's data | Myo Dataset · accuracy | 93.2 -> 95.0 | Helped | EMGBench: Benchmarking Out-of-Distribution Generalization and Adaptation for Electromyography (NeurIPS 2024 · 4 citations) |
| Use AbdomenAtlas 1.0 with enhanced label quality to train ResEncL | JHH · average DSC | 88.8±8.3 -> 89.5±7.8 | Helped | Touchstone Benchmark: Are We on the Right Way for Evaluating AI Algorithms for Medical Segmentation? (NeurIPS 2024 · 2 citations) |
| train the predictor using only 10% of the available training data | Fashion MNIST, 5 nodes · average accuracy | 0.7520 ± 0.0209 -> 0.7376 ± 0.0099 | Hurt | Addressing Label Shift in Distributed Learning via Entropy Regularization (ICLR 2025 · 0 citations) |
| Update the feature extractor before OLS instead of after OLS | CIFAR-10 online label shift, sinusoidal · average error | 11.3% -> 12.33%; 11.2% -> 12.12%; 13.9% -> 14.35%; 11.6% -> 12.10%; 11.4% -> 11.91%; 11.2% -> 12.08% | Hurt | Online Feature Updates Improve Online (Generalized) Label Shift Adaptation (NeurIPS 2024 · 0 citations) |
| Use I-Div instead of HDR alone for assessing hypothesis applicability on noisy test data | CIFAR10 with Gaussian, Salt & Pepper, Uniform, and Speckle noise · AUROC | Helped | Revealing Distribution Discrepancy by Sampling Transfer in Unlabeled Data (NeurIPS 2024 · 0 citations) | |
| Use more acquisition rounds | ACS PUMS, Income and PublicCoverage tasks · accuracy | Mixed | Domain-wise Data Acquisition to Improve Performance under Distribution Shift (ICML 2024) |
What reviewers will expect
Datasets and baselines named in the 10 papers of this approach from 2024–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-10 · 5 of 10 papers
- ImageNet · 3 of 10 papers
- PACS · 2 of 10 papers
- Office-Home · 2 of 10 papers
- ImageNet-V2 · 2 of 10 papers
- ImageNet-Sketch · 2 of 10 papers
- ImageNet-R · 2 of 10 papers
- MNIST · 2 of 10 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 15 papers; from OpenAlex and the papers' affiliations).
- Johns Hopkins University · 2 papers
- The Chinese University of Hong Kong · 2 papers
Use it
Download BibTeX of 26 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Robust optimization for distribution shift (34 papers)
Elsewhere in the topic
- Pseudo-label test-time adaptation (90 papers)
- Test-time adaptation of vision-language models (79 papers)
- Realistic and stable test-time adaptation (71 papers)
- Extrapolation and single-sample adaptation (57 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).