Research map / Label-efficient and robust learning / Test-time adaptation
Test-time adaptation of vision-language models
Adapts CLIP-style models at inference using caches, diffusion and weight averaging.
79 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Test-time adaptation, keyword cluster “time · adaptation · test”
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.
22 papers in 2023–24 and 43 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 20.0% to 24.4%.
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
- 2017 · Test of Time Award (NeurIPS 2017)
- 2019 · Real-Time Self-Adaptive Deep Stereo (CVPR 2019 · 294 citations)
- 2021 · Test-Time Fast Adaptation for Dynamic Scene Deblurring via Meta-Auxiliary Learning (CVPR 2021 · 72 citations)
- 2022 · Parameter-Free Online Test-Time Adaptation (CVPR 2022 · 135 citations)
- 2023 · TTA-COPE: Test-Time Adaptation for Category-Level Object Pose Estimation (CVPR 2023 · 38 citations)
- 2024 · Efficient Test-Time Adaptation of Vision-Language Models (CVPR 2024 · 70 citations)
- 2025 · Test-Time Domain Generalization via Universe Learning: A Multi-Graph Matching Approach for Medical Image Segmentation (CVPR 2025 · 7 citations)
- 2026 · Shedding the Facades, Connecting the Domains: Detecting Shifting Multimodal Hate Video with Test-Time Adaptation (AAAI 2026 · 2 citations)
Earliest papers here
- Test of Time Award (NeurIPS 2017)
- Real-Time Self-Adaptive Deep Stereo (CVPR 2019 · 294 citations)
- Fast and Flexible Multi-Task Classification using Conditional Neural Adaptive Processes (NeurIPS 2019 · 69 citations)
What has been tried
50 papers here report 567 measured changes (a component added, removed or swapped, compared with a baseline): helped 75%, no clear effect 4%, hurt 18%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| add the positive cache | ImageNet · accuracy | 59.81 -> 60.82 | Helped | Efficient Test-Time Adaptation of Vision-Language Models (CVPR 2024 · 70 citations) |
| Add teacher-student structure | KITTI-to-Driving Stereo CTTA · Abs Rel | 0.312 -> 0.249 | Helped | Exploring Sparse Visual Prompt for Domain Adaptive Dense Prediction (AAAI 2024 · 23 citations) |
| Add MTT to HTT | RGBNT201 · mAP | 69.0 -> 71.1 | Helped | Heterogeneous Test-Time Training for Multi-Modal Person Re-identification (AAAI 2024 · 17 citations) |
| add marginal entropy loss | ImageNet-C severity 3 (ResNet50) · accuracy | 40.9 -> 41.70 | Helped | GDA: Generalized Diffusion for Robust Test-time Adaptation (CVPR 2024 · 16 citations) |
| Add prompt ensembling to the best single prompt, with updates | CIFAR10-C, corruption level 3 (average) · accuracy | 77.72 ± 0.24 -> 79.15 ± 0.23 | Helped | Robust Test-Time Adaptation for Zero-Shot Prompt Tuning (AAAI 2024 · 9 citations) |
| use prompt tuning rather than updating all network parameters | four cross-dataset validations · gaze error (degree) | 6.47 -> 6.30; 6.10 -> 5.89; 6.71 -> 6.62; 5.58 -> 5.04 | Helped | Test-Time Personalization with Meta Prompt for Gaze Estimation (AAAI 2024 · 9 citations) |
| add meta-optimization to MeTTA | Penn Action · PCK | 86.67 -> 87.60 | Helped | Fast Adaptation for Human Pose Estimation via Meta-Optimization (CVPR 2024 · 8 citations) |
| Add learned universe embeddings rather than using randomly initialized universe matching | Retinal fundus segmentation, multi-source, average across Sites A–E · DSC (%) | 80.47 -> 88.46 | Helped | Test-Time Domain Generalization via Universe Learning: A Multi-Graph Matching Approach for Medical Image Segmentation (CVPR 2025 · 7 citations) |
| add regional-bootstrapped boosting samples to historical samples | OOD benchmark (ViT-B/16), Average · top-1 accuracy | 64.85 -> 65.57 | Helped | BoostAdapter: Improving Vision-Language Test-Time Adaptation via Regional Bootstrapping (NeurIPS 2024 · 6 citations) |
| Remove normalization of fusion weights | LLVIP · SD | 37.75 -> 29.84 | Hurt | Test-Time Dynamic Image Fusion (NeurIPS 2024 · 6 citations) |
| Add the local knowledge bank | Cityscapes semantic segmentation · mIoU | 39.2 -> 41.1 | Helped | Historical Test-time Prompt Tuning for Vision Foundation Models (NeurIPS 2024 · 6 citations) |
| Use prior-based filtering instead of manual filtering | KITTI DC · RMSE | 1.198 -> 1.413 | Hurt | Zero-shot Depth Completion via Test-time Alignment with Affine-invariant Depth Prior (AAAI 2025 · 5 citations) |
What reviewers will expect
Datasets and baselines named in the 20 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- ImageNet-C · 4 of 20 papers
- CIFAR-100-C · 4 of 20 papers
- ImageNet · 3 of 20 papers
- CIFAR-10-C · 3 of 20 papers
- REFUGE · 2 of 20 papers
- ORIGA · 2 of 20 papers
- Drishti-GS · 2 of 20 papers
- Caltech101 · 2 of 20 papers
Baselines
- TENT · 6 of 20 papers
- SAR · 5 of 20 papers
- TPT · 4 of 20 papers
- TDA · 4 of 20 papers
- CoTTA · 3 of 20 papers
- CoOp · 3 of 20 papers
- CoCoOp · 3 of 20 papers
- ETA · 3 of 20 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 65 papers; from OpenAlex and the papers' affiliations).
- University of Science and Technology of China · 3 papers
- École de technologie supérieure · 3 papers
- Tsinghua University · 3 papers
- Concordia University · 2 papers
- Huawei · 2 papers
- Nanyang Technological University · 2 papers
- Carnegie Mellon University · 2 papers
- Mohamed bin Zayed University of Artificial Intelligence · 2 papers
- Amazon · 2 papers
- Columbia University · 2 papers
Use it
Download BibTeX of 35 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
- Realistic and stable test-time adaptation (71 papers)
- Extrapolation and single-sample adaptation (57 papers)
- Pseudo-label test-time adaptation (90 papers)
Elsewhere in the topic
- Robust optimization for distribution shift (34 papers)
- Characterizing out-of-distribution failures (28 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).