Research map / Vision-language and multimodal / CLIP and zero-shot
Recaptioning and text-supervised pretraining
Uses synthetic captions and interleaved image-text data to learn better visual representations.
73 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in CLIP and zero-shot, keyword cluster “contrastive · language image · image text”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 1 of 2 tracked benchmarks.
35 papers in 2023–24 and 30 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 9.4% to 6.0%.
Open questions the data points to
- ImageNet-1K (Top-1 accuracy): best went from 84 (up to 2024) to 89.5 in 2025 — Multimodal Autoregressive Pre-training of Large Vision Encoders (CVPR 2025 · 19 citations)
No longer reported by papers here: ImageNet zero-shot (accuracy, last in 2024).
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
- 2018 · Stack-Captioning: Coarse-to-Fine Learning for Image Captioning (AAAI 2018 · 187 citations)
- 2021 · Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision (ICML 2021 · 1,195 citations)
- 2022 · LAION-5B: An open large-scale dataset for training next generation image-text models (NeurIPS 2022 · 1,032 citations)
- 2023 · MaskCLIP: Masked Self-Distillation Advances Contrastive Language-Image Pretraining (CVPR 2023 · 169 citations)
- 2024 · MLIP: Enhancing Medical Visual Representation with Divergence Encoder and Knowledge-guided Contrastive Learning (CVPR 2024 · 34 citations)
- 2025 · Multimodal Autoregressive Pre-training of Large Vision Encoders (CVPR 2025 · 19 citations)
- 2026 · MaskInversion: Localized Embeddings via Optimization of Explainability Maps (ICLR 2026 · 1 citation)
Earliest papers here
- Stack-Captioning: Coarse-to-Fine Learning for Image Captioning (AAAI 2018 · 187 citations)
- Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision (ICML 2021 · 1,195 citations)
- LAION-5B: An open large-scale dataset for training next generation image-text models (NeurIPS 2022 · 1,032 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 zero-shot · accuracy (higher is better)
4 papers, 2022–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 84.1 | Multimodal Contrastive Learning with LIMoE: the Language-Image Mixture of Experts (NeurIPS 2022 · 71 citations) |
ImageNet-1K · Top-1 accuracy (higher is better)
4 papers, 2022–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 81.8 | Unified Contrastive Learning in Image-Text-Label Space (CVPR 2022 · 193 citations) |
| 2024 | 84 | Scalable Pre-training of Large Autoregressive Image Models (ICML 2024 · 6 citations) |
| 2025 | 89.5 | Multimodal Autoregressive Pre-training of Large Vision Encoders (CVPR 2025 · 19 citations) |
What has been tried
53 papers here report 642 measured changes (a component added, removed or swapped, compared with a baseline): helped 71%, no clear effect 7%, hurt 19%, 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 category-level ITA | RSNA Pneumonia · Dice | 71.7 -> 73.5 | Helped | MLIP: Enhancing Medical Visual Representation with Divergence Encoder and Knowledge-guided Contrastive Learning (CVPR 2024 · 34 citations) |
| Use the full synthetic caption configuration instead of cc12m captions | ImageNet-1K linear evaluation · top-1 accuracy | 77.1 -> 78.8 | Helped | Learning Vision from Models Rivals Learning Vision from Data (CVPR 2024 · 25 citations) |
| Add minimal fine-tuning (Freeze) | OOD mean across target datasets · mIoU (%) | 14.69 -> 32.44 | Helped | A Simple Recipe for Language-guided Domain Generalized Segmentation (CVPR 2024 · 23 citations) |
| Use the multimodal autoregressive objective instead of captioning-only training | ImageNet-1k · top-1 accuracy | 85.1 -> 85.6 | Helped | Multimodal Autoregressive Pre-training of Large Vision Encoders (CVPR 2025 · 19 citations) |
| Add image-text co-decomposition | Six-dataset average · mIoU | 29.8 -> 31.0 | Helped | Image-Text Co-Decomposition for Text-Supervised Semantic Segmentation (CVPR 2024 · 14 citations) |
| use CLIP ViT-L/14 instead of ViT-B/16 as image encoder | GOD test · Top-1 accuracy | 25.6% -> 25.8% | Helped | Bridging the Semantic Latent Space between Brain and Machine: Similarity Is All You Need (AAAI 2024 · 12 citations) |
| add human-refined training data to PLM | AuroraCap · Score | 2.2 -> 3.4 | Helped | Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · 11 citations) |
| Add CLIP Image Adapter (CIA) before pre-training | Objaverse-LVIS zero-shot · Top-1 accuracy | 12.2 -> 13.7 (single CIA-only ablation row 13.7 vs 12.2 one-stage) | Helped | TAMM: TriAdapter Multi-Modal Learning for 3D Shape Understanding (CVPR 2024 · 11 citations) |
| Pre-train on image-text pairs before COCO fine-tuning | OpenImages zero-shot tagging · mAP | 57.5 -> 83.4 | Helped | Tag2Text: Guiding Vision-Language Model via Image Tagging (ICLR 2024 · 11 citations) |
| add CKA rather than use normalized correlation matrices | COCO · matching accuracy | 10.1 -> 48.8 | Helped | Do Vision and Language Encoders Represent the World Similarly? (CVPR 2024 · 6 citations) |
| Add Screenshot Titling to Screen Parsing pre-training | Screen2words · Cider | 101.3 -> 101.4 | No clear effect | Enhancing Vision-Language Pre-training with Rich Supervisions (CVPR 2024 · 6 citations) |
| Add InterAug without data pruning | ChestXpert zero-shot · AUC | 86.24 -> 84.91 | Hurt | PairAug: What Can Augmented Image-Text Pairs Do for Radiology? (CVPR 2024 · 6 citations) |
What reviewers will expect
Datasets and baselines named in the 15 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- COCO · 6 of 15 papers
- ImageNet-1k · 6 of 15 papers
- Flickr30K · 4 of 15 papers
- CC3M · 3 of 15 papers
- CC12M · 3 of 15 papers
- ImageNet · 3 of 15 papers
- MSCOCO · 3 of 15 papers
- DataComp-1B · 2 of 15 papers
Baselines
- SigLIP · 3 of 15 papers
- CLIP · 3 of 15 papers
- NegCLIP · 2 of 15 papers
- VeCLIP · 2 of 15 papers
- Long-CLIP · 2 of 15 papers
- GPT-4V · 2 of 15 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).
- Meta · 7 papers
- Apple · 6 papers
- Google · 5 papers
- University of Michigan · 3 papers
- Tsinghua University · 3 papers
- University of California, Berkeley · 3 papers
- Amazon · 3 papers
- ByteDance · 3 papers
- Zhejiang University · 2 papers
- University of Science and Technology of China · 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
- CLIP training recipes and data curation (72 papers)
- Fine-grained improvements to CLIP (63 papers)
- Self-distillation for vision-language pretraining (48 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)
- Prompt learning for CLIP (102 papers)
- Test-time adaptation of VLMs (93 papers)
- Language models for zero-shot reasoning (64 papers)
- Feature calibration for unseen-class detection (50 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).