Research map / Vision-language and multimodal / CLIP and zero-shot
Test-time adaptation of VLMs
Adapts vision-language models on unlabeled test data via prompt tuning and lightweight adapters.
93 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · an approach in CLIP and zero-shot, keyword cluster “prompt · tuning · pre trained”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 4 of 4 tracked benchmarks.
20 papers in 2023–24 and 71 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 5.4% to 14.2%.
Open questions the data points to
- ImageNet-Sketch (Accuracy): best went from 49.13 (up to 2024) to 51.65 in 2025 — SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting (CVPR 2025 · 1 citation)
- ImageNet-V2 (Accuracy): best went from 64.33 (up to 2024) to 66.37 in 2025 — Adaptive Parameter Selection for Tuning Vision-Language Models (CVPR 2025 · 4 citations)
- ImageNet-A (Accuracy): best went from 51.12 (up to 2024) to 64.52 in 2025 — SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting (CVPR 2025 · 1 citation)
- ImageNet-R (Accuracy): best went from 77.32 (up to 2024) to 81.68 in 2025 — SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting (CVPR 2025 · 1 citation)
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
- 2022 · VL-Adapter: Parameter-Efficient Transfer Learning for Vision-and-Language Tasks (CVPR 2022 · 310 citations)
- 2023 · Task Residual for Tuning Vision-Language Models (CVPR 2023 · 104 citations)
- 2024 · MMA: Multi-Modal Adapter for Vision-Language Models (CVPR 2024 · 76 citations)
- 2025 · BiomedCoOp: Learning to Prompt for Biomedical Vision-Language Models (CVPR 2025 · 26 citations)
- 2026 · Cross-Domain Few-Shot Learning via Multi-View Collaborative Optimization with Vision-Language Models (AAAI 2026 · 1 citation)
Earliest papers here
- VL-Adapter: Parameter-Efficient Transfer Learning for Vision-and-Language Tasks (CVPR 2022 · 310 citations)
- VLUE: A Multi-Task Multi-Dimension Benchmark for Evaluating Vision-Language Pre-training (ICML 2022)
- Task Residual for Tuning Vision-Language Models (CVPR 2023 · 104 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
ImageNet-Sketch · Accuracy (higher is better)
5 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 49.13 | MMA: Multi-Modal Adapter for Vision-Language Models (CVPR 2024 · 76 citations) |
| 2025 | 51.65 | SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting (CVPR 2025 · 1 citation) |
ImageNet-V2 · Accuracy (higher is better)
5 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 64.33 | MMA: Multi-Modal Adapter for Vision-Language Models (CVPR 2024 · 76 citations) |
| 2025 | 66.37 | Adaptive Parameter Selection for Tuning Vision-Language Models (CVPR 2025 · 4 citations) |
ImageNet-A · Accuracy (higher is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 51.12 | MMA: Multi-Modal Adapter for Vision-Language Models (CVPR 2024 · 76 citations) |
| 2025 | 64.52 | SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting (CVPR 2025 · 1 citation) |
ImageNet-R · Accuracy (higher is better)
4 papers, 2024–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2024 | 77.32 | MMA: Multi-Modal Adapter for Vision-Language Models (CVPR 2024 · 76 citations) |
| 2025 | 81.68 | SCAP: Transductive Test-Time Adaptation via Supportive Clique-based Attribute Prompting (CVPR 2025 · 1 citation) |
What has been tried
49 papers here report 618 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 5%, hurt 25%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add adapters to both branches and a shared projection, rather than use a single-modality adapter or replace MMA with cross-attention | Base-to-Novel average (11 datasets) · HM | 78.02; 77.16; 77.34; 79.87 | Helped | MMA: Multi-Modal Adapter for Vision-Language Models (CVPR 2024 · 76 citations) |
| Add SCCM (LLM prompt ensemble semantic consistency) on BiomedCLIP | 11 biomedical benchmarks (few-shot and base-to-novel) · HM / K=16 accuracy | HM 67.23 -> 70.19; K=16 69.62 -> 69.99 (per Table 3 rows) | Helped | BiomedCoOp: Learning to Prompt for Biomedical Vision-Language Models (CVPR 2025 · 26 citations) |
| remove text-side representation tokens | Base-to-novel generalization, 11-dataset average · HM | 81.20 -> 80.08 | Hurt | MMRL: Multi-Modal Representation Learning for Vision-Language Models (CVPR 2025 · 22 citations) |
| Add multi-modal in-context tuning (Stage II) | VSR · score | 61.54 -> 70.54 | Helped | MMICL: Empowering Vision-language Model with Multi-Modal In-Context Learning (ICLR 2024 · 18 citations) |
| add intra-HOI feature fusion | HICO-DET UV · Seen mAP | +7.41 mAP | Helped | EZ-HOI: VLM Adaptation via Guided Prompt Learning for Zero-Shot HOI Detection (NeurIPS 2024 · 13 citations) |
| add HouseHolder transformation | 11-dataset average · ECE | 4.45 -> 4.23 | Helped | O-TPT: Orthogonality Constraints for Calibrating Test-time Prompt Tuning in Vision-Language Models (CVPR 2025 · 10 citations) |
| Add classifier tuning after first-stage LayerNorm tuning | Stanford Cars all-to-all · Accuracy | +10.6 | Helped | Rethinking Few-Shot Adaptation of Vision-Language Models in Two Stages (CVPR 2025 · 10 citations) |
| add hand-crafted prompt ensemble | Natural Distribution Shifts (mean) · top-1 accuracy | 63.74 -> 66.10 | Helped | Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · 10 citations) |
| Add likelihood adaptation to CLIP | ImageNet · accuracy (%) | 59.81 -> 60.95 | Helped | Bayesian Test-Time Adaptation for Vision-Language Models (CVPR 2025 · 9 citations) |
| Add PCB to BADGE without description augmentation | Average across seven datasets · Avg Acc | 72.43 -> 72.95 | Helped | Active Prompt Learning in Vision Language Models (CVPR 2024 · 9 citations) |
| use cumulative average rather than no textual prototype update | ImageNet · accuracy | 62.93 -> 63.41 | Helped | Dual Prototype Evolving for Test-Time Generalization of Vision-Language Models (NeurIPS 2024 · 9 citations) |
| Remove the imposed layer-wise distribution | MSCOCO Image-to-Text retrieval, 63% sparsity · R@1 | 77.35 -> 0.01±0.01 | Hurt | MULTIFLOW: Shifting Towards Task-Agnostic Vision-Language Pruning (CVPR 2024 · 8 citations) |
What reviewers will expect
Datasets and baselines named in the 35 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Caltech101 · 16 of 35 papers
- ImageNet · 15 of 35 papers
- Food101 · 14 of 35 papers
- DTD · 14 of 35 papers
- EuroSAT · 14 of 35 papers
- OxfordPets · 12 of 35 papers
- FGVCAircraft · 12 of 35 papers
- SUN397 · 12 of 35 papers
Baselines
- CoOp · 13 of 35 papers
- CoCoOp · 10 of 35 papers
- TPT · 10 of 35 papers
- CLIP · 8 of 35 papers
- MaPLe · 7 of 35 papers
- KgCoOp · 6 of 35 papers
- C-TPT · 6 of 35 papers
- Tip-Adapter · 5 of 35 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 91 papers; from OpenAlex and the papers' affiliations).
- University of Science and Technology of China · 8 papers
- National University of Singapore · 4 papers
- Tsinghua University · 4 papers
- University of Trento · 3 papers
- Chinese Academy of Sciences · 3 papers
- Shanghai Jiao Tong University · 2 papers
- Fondazione Bruno Kessler · 2 papers
- Shenzhen Technology University · 2 papers
- Peking University · 2 papers
- Hong Kong Baptist University · 2 papers
Use it
Download BibTeX of 25 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
- Prompt learning for CLIP (102 papers)
- Feature calibration for unseen-class detection (50 papers)
- Parameter-efficient visual tuning (41 papers)
Elsewhere in the topic
- Open-vocabulary alignment from VLMs (130 papers)
- Attribute-based zero-shot learning (115 papers)
- Adapting VLMs for zero-shot classification (108 papers)
- Recaptioning and text-supervised pretraining (73 papers)
- CLIP training recipes and data curation (72 papers)
- Language models for zero-shot reasoning (64 papers)
- Fine-grained improvements to CLIP (63 papers)
- Self-distillation for vision-language pretraining (48 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).