Research map / Vision-language and multimodal / CLIP and zero-shot
Attribute-based zero-shot learning
Uses attribute-guided transformers, prototypes and semantic grounding to classify unseen classes.
115 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in CLIP and zero-shot, keyword cluster “zero shot · open · vocabulary”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 5 of 6 tracked benchmarks.
38 papers in 2023–24 and 56 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 10.2% to 11.2%.
Open questions the data points to
- C-GQA open-world (AUC): best went from 2.7 (up to 2024) to 4.2 in 2025 — TOMCAT: Test-time Comprehensive Knowledge Accumulation for Compositional Zero-Shot Learning (NeurIPS 2025 · 0 citations)
- MIT-States open-world (AUC): best went from 7.2 (up to 2024) to 8.7 in 2025 — LOGICZSL: Exploring Logic-induced Representation for Compositional Zero-shot Learning (CVPR 2025 · 8 citations)
- C-GQA closed-world (AUC): best went from 12.4 (up to 2024) to 16 in 2025 — TOMCAT: Test-time Comprehensive Knowledge Accumulation for Compositional Zero-Shot Learning (NeurIPS 2025 · 0 citations)
- CUB (H): best went from 73.6 (up to 2024) to 75.7 in 2025 — Visual and Semantic Prompt Collaboration for Generalized Zero-Shot Learning (CVPR 2025 · 8 citations)
- ImageNet (accuracy): best went from 68.8 (up to 2024) to 83.1 in 2025 — UNEM: UNrolled Generalized EM for Transductive Few-Shot Learning (CVPR 2025 · 2 citations)
- Papers report MeaCap on MSCOCO test (CIDEr) anywhere from 40.3 to 95.4: comparisons against it depend on whose number you take.
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
- 2016 · Learning Deep Representations of Fine-Grained Visual Descriptions (CVPR 2016 · 799 citations)
- 2017 · Learning a Deep Embedding Model for Zero-Shot Learning (CVPR 2017 · 792 citations)
- 2018 · Stacked Semantics-Guided Attention Model for Fine-Grained Zero-Shot Learning (NeurIPS 2018 · 97 citations)
- 2019 · Semantically Tied Paired Cycle Consistency for Zero-Shot Sketch-Based Image Retrieval (CVPR 2019 · 169 citations)
- 2020 · Learning Cross-Aligned Latent Embeddings for Zero-Shot Cross-Modal Retrieval (AAAI 2020 · 40 citations)
- 2021 · Open World Compositional Zero-Shot Learning (CVPR 2021 · 150 citations)
- 2022 · LiT: Zero-Shot Transfer With Locked-Image Text Tuning (CVPR 2022 · 380 citations)
- 2023 · WinCLIP: Zero-/Few-Shot Anomaly Classification and Segmentation (CVPR 2023 · 526 citations)
- 2024 · Language-only Training of Zero-shot Composed Image Retrieval (CVPR 2024 · 58 citations)
- 2025 · AA-CLIP: Enhancing Zero-Shot Anomaly Detection via Anomaly-Aware CLIP (CVPR 2025 · 69 citations)
- 2026 · AdaptCLIP: Adapting CLIP for Universal Visual Anomaly Detection (AAAI 2026 · 16 citations)
Earliest papers here
- Learning Deep Representations of Fine-Grained Visual Descriptions (CVPR 2016 · 799 citations)
- Less Is More: Zero-Shot Learning From Online Textual Documents With Noise Suppression (CVPR 2016 · 187 citations)
- Zero-Shot Event Detection by Multimodal Distributional Semantic Embedding of Videos (AAAI 2016 · 17 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.
C-GQA open-world · AUC (higher is better)
5 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 2.4 | Decomposed Soft Prompt Guided Fusion Enhancing for Compositional Zero-Shot Learning (CVPR 2023 · 57 citations) |
| 2024 | 2.7 | Troika: Multi-Path Cross-Modal Traction for Compositional Zero-Shot Learning (CVPR 2024 · 33 citations) |
| 2025 | 4.2 | TOMCAT: Test-time Comprehensive Knowledge Accumulation for Compositional Zero-Shot Learning (NeurIPS 2025 · 0 citations) |
MIT-States open-world · AUC (higher is better)
5 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 6.8 | Decomposed Soft Prompt Guided Fusion Enhancing for Compositional Zero-Shot Learning (CVPR 2023 · 57 citations) |
| 2024 | 7.2 | Troika: Multi-Path Cross-Modal Traction for Compositional Zero-Shot Learning (CVPR 2024 · 33 citations) |
| 2025 | 8.7 | LOGICZSL: Exploring Logic-induced Representation for Compositional Zero-shot Learning (CVPR 2025 · 8 citations) |
AwA2 · H (higher is better)
4 papers, 2022–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 68.7 | Learning Aligned Cross-Modal Representation for Generalized Zero-Shot Classification (AAAI 2022 · 31 citations) |
| 2024 | 86.39 | TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning (NeurIPS 2024 · 3 citations) |
C-GQA closed-world · AUC (higher is better)
4 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 10.5 | Decomposed Soft Prompt Guided Fusion Enhancing for Compositional Zero-Shot Learning (CVPR 2023 · 57 citations) |
| 2024 | 12.4 | Troika: Multi-Path Cross-Modal Traction for Compositional Zero-Shot Learning (CVPR 2024 · 33 citations) |
| 2025 | 16 | TOMCAT: Test-time Comprehensive Knowledge Accumulation for Compositional Zero-Shot Learning (NeurIPS 2025 · 0 citations) |
CUB · H (higher is better)
4 papers, 2022–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 55.2 | Learning Aligned Cross-Modal Representation for Generalized Zero-Shot Classification (AAAI 2022 · 31 citations) |
| 2024 | 73.6 | Progressive Semantic-Guided Vision Transformer for Zero-Shot Learning (CVPR 2024 · 52 citations) |
| 2025 | 75.7 | Visual and Semantic Prompt Collaboration for Generalized Zero-Shot Learning (CVPR 2025 · 8 citations) |
ImageNet · accuracy (higher is better)
4 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 62.7 | Text-to-Image Diffusion Models are Zero Shot Classifiers (NeurIPS 2023 · 17 citations) |
| 2024 | 68.8 | PerceptionCLIP: Visual Classification by Inferring and Conditioning on Contexts (ICLR 2024 · 0 citations) |
| 2025 | 83.1 | UNEM: UNrolled Generalized EM for Transductive Few-Shot Learning (CVPR 2025 · 2 citations) |
What has been tried
84 papers here report 952 measured changes (a component added, removed or swapped, compared with a baseline): helped 75%, no clear effect 5%, hurt 16%, 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 |
|---|---|---|---|---|
| Replace the linear adapter with a residual adapter | VisA-trained 64-shot setup · pixel-level AUROC | 48.9(-40.0) -> 91.3(+2.4) | Helped | AA-CLIP: Enhancing Zero-Shot Anomaly Detection via Anomaly-Aware CLIP (CVPR 2025 · 69 citations) |
| Replace keyword masking with the prompt “a photo of [$]” | CIRR dev · R@1 | 25.66 -> 21.65 | Hurt | Language-only Training of Zero-shot Composed Image Retrieval (CVPR 2024 · 58 citations) |
| Add visual reconstruction loss LV R | CUB · acc | 76.4 -> 78.9 | Helped | Progressive Semantic-Guided Vision Transformer for Zero-Shot Learning (CVPR 2024 · 52 citations) |
| Add memory-augmented key-concept retrieval and filtering | MSCOCO · CIDEr | 12.7 -> 31.1 | Helped | MeaCap: Memory-Augmented Zero-shot Image Captioning (CVPR 2024 · 50 citations) |
| use SemEvo paraphrased descriptions instead of class names | MiniImageNet 5-way 1-shot · accuracy | 73.98 -> 77.63 | Helped | Simple Semantic-Aided Few-Shot Learning (CVPR 2024 · 48 citations) |
| Add SPT, LAT, and SaSC to VLBase | FSC147 test set · MAE | 32.20 -> 17.05 | Helped | VLCounter: Text-Aware Visual Representation for Zero-Shot Object Counting (AAAI 2024 · 43 citations) |
| add the state and object branches alongside the composition branch during training and inference | UT-Zappos closed-world · HM | 49.9 -> 54.6 | Helped | Troika: Multi-Path Cross-Modal Traction for Compositional Zero-Shot Learning (CVPR 2024 · 33 citations) |
| Replace Shift-GCN with another tested skeleton encoder | NTU-RGB+D 60 55/5 · accuracy (%) | 79.23; 76.75; 79.97; 80.41; 81.14 | Mixed | Part-aware Unified Representation of Language and Skeleton for Zero-shot Action Recognition (CVPR 2024 · 30 citations) |
| add prompt alignment using P1 | ChestXray14 · AUC | 0.764 -> 0.795 | Helped | CARZero: Cross-Attention Alignment for Radiology Zero-Shot Classification (CVPR 2024 · 29 citations) |
| Add top-K retrieved image features | CIRR validation · R10 | 67.0 -> 69.2 | Helped | Knowledge-Enhanced Dual-stream Zero-shot Composed Image Retrieval (CVPR 2024 · 28 citations) |
| Add a learned visual prompt | Sketchy · mAP@200 | 0.513 -> 0.746 | Helped | Text-to-Image Diffusion Models are Great Sketch-Photo Matchmakers (CVPR 2024 · 24 citations) |
| add triplet-to-instance matching instead of scoring-and-ranking | RefCOCOg · accuracy | 55.35 -> 57.60; 54.33 -> 56.64 | Helped | Zero-shot Referring Expression Comprehension via Structural Similarity Between Images and Captions (CVPR 2024 · 17 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| MeaCap | MSCOCO test | CIDEr | 40.3 – 95.4 | 2 papers, e.g. CVPR 2025 and AAAI 2025 |
What reviewers will expect
Datasets and baselines named in the 30 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CUB · 4 of 30 papers
- SUN · 3 of 30 papers
- AWA2 · 3 of 30 papers
- MSCOCO · 3 of 30 papers
- MVTec AD · 2 of 30 papers
- VisA · 2 of 30 papers
- MPDD · 2 of 30 papers
- BTAD · 2 of 30 papers
Baselines
- CLIP · 9 of 30 papers
- CoOp · 7 of 30 papers
- TransZero++ · 3 of 30 papers
- WinCLIP · 2 of 30 papers
- AnomalyCLIP · 2 of 30 papers
- ZSLViT · 2 of 30 papers
- DUET · 2 of 30 papers
- DPPN · 2 of 30 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 94 papers; from OpenAlex and the papers' affiliations).
- Zhejiang University · 4 papers
- Alibaba Group · 4 papers
- Hong Kong University of Science and Technology · 4 papers
- University of Science and Technology of China · 4 papers
- Sungkyunkwan University · 3 papers
- Tsinghua University · 3 papers
- Huazhong University of Science and Technology · 3 papers
- National University of Singapore · 3 papers
- Harbin Institute of Technology · 3 papers
- Nanyang Technological University · 3 papers
Use it
Download BibTeX of 37 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
- Open-vocabulary alignment from VLMs (130 papers)
- Adapting VLMs for zero-shot classification (108 papers)
- Language models for zero-shot reasoning (64 papers)
Elsewhere in the topic
- Prompt learning for CLIP (102 papers)
- Test-time adaptation of VLMs (93 papers)
- Recaptioning and text-supervised pretraining (73 papers)
- CLIP training recipes and data curation (72 papers)
- Fine-grained improvements to CLIP (63 papers)
- Feature calibration for unseen-class detection (50 papers)
- Self-distillation for vision-language pretraining (48 papers)
- Parameter-efficient visual tuning (41 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).