Research map / Vision-language and multimodal / CLIP and zero-shot
Fine-grained improvements to CLIP
Strengthens CLIP's local detail understanding and studies its scaling behaviour.
63 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2022–2026) · an approach in CLIP and zero-shot, keyword cluster “contrastive · language image · image text”
State History Tried Baselines 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.
25 papers in 2023–24 and 35 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 6.7% to 7.0%.
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 · CRIS: CLIP-Driven Referring Image Segmentation (CVPR 2022 · 401 citations)
- 2023 · Exploring CLIP for Assessing the Look and Feel of Images (AAAI 2023 · 693 citations)
- 2024 · VadCLIP: Adapting Vision-Language Models for Weakly Supervised Video Anomaly Detection (AAAI 2024 · 206 citations)
- 2025 · SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining (CVPR 2025 · 36 citations)
- 2026 · LLM2CLIP: Powerful Language Model Unlocks Richer Cross-Modality Representation (AAAI 2026 · 3 citations)
Earliest papers here
- CRIS: CLIP-Driven Referring Image Segmentation (CVPR 2022 · 401 citations)
- CLOOB: Modern Hopfield Networks with InfoLOOB Outperform CLIP (NeurIPS 2022 · 40 citations)
- Quality Not Quantity: On the Interaction between Dataset Design and Robustness of CLIP (NeurIPS 2022 · 18 citations)
What has been tried
43 papers here report 510 measured changes (a component added, removed or swapped, compared with a baseline): helped 72%, no clear effect 6%, 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 temporal modeling with LGT-Adapter | XD-Violence · AP | 72.22 -> 84.51 | Helped | VadCLIP: Adapting Vision-Language Models for Weakly Supervised Video Anomaly Detection (AAAI 2024 · 206 citations) |
| Use a whole-image alpha map instead of a region mask | ImageNet-S · Top-1 accuracy | 77.41 -> 73.37 | Hurt | Alpha-CLIP: A CLIP Model Focusing on Wherever You Want (CVPR 2024 · 91 citations) |
| Add synthetic captions alongside real captions during CLIP training | ImageNet-val · zero-shot accuracy | 44.5 -> 51.9 | Helped | MobileCLIP: Fast Image-Text Models through Multi-Modal Reinforced Training (CVPR 2024 · 46 citations) |
| Add VTP to the HDMNet baseline | PASCAL-5i · mIoU | 4.3% | Helped | Rethinking Prior Information Generation with CLIP for Few-Shot Segmentation (CVPR 2024 · 46 citations) |
| Add ROI/tag soft loss | ImageNet ZS; CC3M; ResNet50 and ViT-B/32 · top-1 accuracy | +4.0; +1.0 | Helped | SoftCLIP: Softer Cross-Modal Alignment Makes CLIP Stronger (AAAI 2024 · 42 citations) |
| add SR-CLIP | NUAA-SIRST · IoU | 76.82 -> 80.02 | Helped | SAIST: Segment Any Infrared Small Target Model Guided by Contrastive Language-Image Pretraining (CVPR 2025 · 36 citations) |
| Add auxiliary features | ImageNet-1K, 16-shot · accuracy | 65.34 -> 69.35 | Helped | AMU-Tuning: Effective Logit Bias for CLIP-based Few-shot Learning (CVPR 2024 · 31 citations) |
| Use MoPE rather than top-layer selection | MSCOCO image-to-text retrieval, prune 6 layers at 0.375 width · TR@1 | 57.6 -> 69.7 | Helped | MoPE-CLIP: Structured Pruning for Efficient Vision-Language Models with Module-wise Pruning Error Metric (CVPR 2024 · 21 citations) |
| add F2E contrastive distillation | DDD17-Seg test, voxel grid, linear probing · mIoU | 12.30 -> 52.60 | Helped | OpenESS: Event-based Semantic Scene Understanding with Open Vocabularies (CVPR 2024 · 21 citations) |
| Replace global loss only with text-conditioned loss | COCO retrieval · T2I R@1 | 28.3 -> 32.0 | Helped | FLAIR: VLM with Fine-grained Language-informed Image Representations (CVPR 2025 · 15 citations) |
| add edge detection and optimal transport normalization | ImageNet-1K zero-shot classification · top-1 accuracy (%) | 35.2 -> 38.0 | Helped | Seeing What Matters: Empowering CLIP with Patch Generation-to-Selection (CVPR 2025 · 8 citations) |
| Add IVFusion to LoRA | MSRVTT · R@sum | 193.0 -> 201.7 | Helped | DiscoVLA: Discrepancy Reduction in Vision, Language, and Alignment for Parameter-Efficient Video-Text Retrieval (CVPR 2025 · 7 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 |
|---|---|---|---|---|
| OpenCLIP | ImageNet-1K zero-shot | accuracy | 72.8 – 78 | 2 papers, e.g. CVPR 2023 and CVPR 2023 |
What reviewers will expect
Datasets and baselines named in the 22 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-1K · 5 of 22 papers
- MS-COCO · 4 of 22 papers
- Flickr30k · 4 of 22 papers
- ImageNet · 4 of 22 papers
- DCI · 3 of 22 papers
- ShareGPT4V · 3 of 22 papers
- CIFAR10 · 3 of 22 papers
- CC12M · 3 of 22 papers
Baselines
- CLIP · 7 of 22 papers
- Long-CLIP · 3 of 22 papers
- CoOp · 2 of 22 papers
- DIVA · 2 of 22 papers
- TripletCLIP · 2 of 22 papers
- NegCLIP · 2 of 22 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 60 papers; from OpenAlex and the papers' affiliations).
- Microsoft · 3 papers
- Eberhard-Karls-Universität Tübingen · 3 papers
- University of Tübingen · 3 papers
- East China Normal University · 2 papers
- Nanyang Technological University · 2 papers
- City University of Hong Kong · 2 papers
- Meta · 2 papers
- Huawei · 2 papers
- Google DeepMind · 2 papers
- EPFL · 2 papers
Use it
Download BibTeX of 30 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)
- Recaptioning and text-supervised pretraining (73 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).