Research map / Label-efficient and robust learning / Recognition and long-tail
Partial and multi-label learning
Learns from partially annotated labels using class-aware, margin-based and representation blending approaches.
345 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Recognition and long-tail, keyword cluster “label · classification · classifier”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Leveling off Several papers from 2025–26 still report on 3 tracked benchmarks, but the best result has barely moved since 2024.
76 papers in 2023–24 and 81 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 6.1% to 6.2%.
Open questions the data points to
- Papers report PL-SVM on FG-NET (accuracy) anywhere from 0.043 to 0.063: comparisons against it depend on whose number you take.
- Papers report LIFT on Corel5K (average precision) anywhere from 0.291 to 0.423: comparisons against it depend on whose number you take.
What looks settled
- corel5k (average precision): best up to 2024 0.575, best in 2025–26 0.304 from 2 papers, at most 0.2% better, if at all.
- scene (Average precision): best up to 2024 0.905, best in 2025–26 0.871 from 3 papers, at most 0.2% better, if at all.
- image (Average precision): best up to 2024 0.856, best in 2025–26 0.821 from 2 papers, at most 0.2% better, if at all.
No longer reported by papers here: Amazon-670K (P@1, last in 2022).
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 · Multi-Label Manifold Learning (AAAI 2016 · 140 citations)
- 2017 · Learning Spatial Regularization With Image-Level Supervisions for Multi-Label Image Classification (CVPR 2017 · 383 citations)
- 2018 · Attention-based Deep Multiple Instance Learning (ICML 2018 · 675 citations)
- 2019 · Learning a Deep ConvNet for Multi-Label Classification With Partial Labels (CVPR 2019 · 256 citations)
- 2020 · Multi-Label Classification with Label Graph Superimposing (AAAI 2020 · 190 citations)
- 2021 · TransMIL: Transformer based Correlated Multiple Instance Learning for Whole Slide Image Classification (NeurIPS 2021 · 505 citations)
- 2022 · DTFD-MIL: Double-Tier Feature Distillation Multiple Instance Learning for Histopathology Whole Slide Image Classification (CVPR 2022 · 474 citations)
- 2023 · Interventional Bag Multi-Instance Learning on Whole-Slide Pathological Images (CVPR 2023 · 138 citations)
- 2024 · Multi-Label Supervised Contrastive Learning (AAAI 2024 · 35 citations)
- 2025 · Rebalancing Multi-Label Class-Incremental Learning (AAAI 2025 · 10 citations)
- 2026 · Multi-Label Classification with Incremental and Decremental Features (AAAI 2026 · 1 citation)
Earliest papers here
- Multi-Label Manifold Learning (AAAI 2016 · 140 citations)
- Learning Structured Inference Neural Networks With Label Relations (CVPR 2016 · 126 citations)
- Conditional Graphical Lasso for Multi-Label Image Classification (CVPR 2016 · 119 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.
corel5k · average precision (higher is better)
10 papers, 2016–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2016 | 0.123 | Teaching-to-Learn and Learning-to-Teach for Multi-label Propagation (AAAI 2016 · 43 citations) |
| 2017 | 0.575 | Efficient Multiple Instance Metric Learning Using Weakly Supervised Data (CVPR 2017 · 15 citations) |
Amazon-670K · P@1 (higher is better)
9 papers, 2018–2022 · not reported since 2022
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 39.9 | A no-regret generalization of hierarchical softmax to extreme multi-label classification (NeurIPS 2018 · 37 citations) |
| 2019 | 47.58 | AttentionXML: Label Tree-based Attention-Aware Deep Model for High-Performance Extreme Multi-Label Text Classification (NeurIPS 2019 · 92 citations) |
| 2021 | 50.11 | Fast Multi-Resolution Transformer Fine-tuning for Extreme Multi-label Text Classification (NeurIPS 2021 · 16 citations) |
| 2022 | 53.02 | ELIAS: End-to-End Learning to Index and Search in Large Output Spaces (NeurIPS 2022 · 2 citations) |
scene · Average precision (higher is better)
9 papers, 2016–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2016 | 0.859 | Teaching-to-Learn and Learning-to-Teach for Multi-label Propagation (AAAI 2016 · 43 citations) |
| 2019 | 0.897 | Collaboration Based Multi-Label Learning (AAAI 2019 · 59 citations) |
| 2022 | 0.905 | End-to-End Probabilistic Label-Specific Feature Learning for Multi-Label Classification (AAAI 2022 · 12 citations) |
image · Average precision (higher is better)
8 papers, 2018–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 0.833 | Feature-Induced Labeling Information Enrichment for Multi-Label Learning (AAAI 2018 · 86 citations) |
| 2019 | 0.843 | Collaboration Based Multi-Label Learning (AAAI 2019 · 59 citations) |
| 2022 | 0.856 | End-to-End Probabilistic Label-Specific Feature Learning for Multi-Label Classification (AAAI 2022 · 12 citations) |
mirflickr · average precision (higher is better)
8 papers, 2019–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2019 | 0.827 | Partial Multi-Label Learning via Credible Label Elicitation (AAAI 2019 · 133 citations) |
| 2020 | 0.835 | Partial Multi-Label Learning with Label Distribution (AAAI 2020 · 80 citations) |
yeast · Average precision (higher is better)
7 papers, 2018–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2018 | 0.738 | Partial Multi-Label Learning (AAAI 2018 · 183 citations) |
| 2019 | 0.775 | Collaboration Based Multi-Label Learning (AAAI 2019 · 59 citations) |
What has been tried
287 papers here report 2,786 measured changes (a component added, removed or swapped, compared with a baseline): helped 68%, no clear effect 8%, hurt 18%, mixed 6%. 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 global projection head | TripAdvisor, 30–360 samples · accuracy | 72.50 -> 72.9; 74.81 -> 75.1; 77.30 -> 77.9; 78.44 -> 79.0 | Helped | Multi-Label Supervised Contrastive Learning (AAAI 2024 · 35 citations) |
| Omit the last self-attention operation | PASCAL VOC 2012 val · mIoU | 16.2 -> 41.6 | Helped | TagCLIP: A Local-to-Global Framework to Enhance Open-Vocabulary Multi-Label Classification of CLIP without Training (AAAI 2024 · 28 citations) |
| Use pseudo-bag size 512 rather than the baseline size 512 random-sampling setup alternatives | CAMELYON16 · accuracy | 0.923 (128); 0.923 (256); 0.954 (512); 0.930 (1024); 0.907 (2048) | Mixed | Dynamic Policy-Driven Adaptive Multi-Instance Learning for Whole Slide Image Classification (CVPR 2024 · 22 citations) |
| Vary the number of clusters to 2, 4, 8, or 16 | TCGA-NSCLC · classification performance | Helped | CaMIL: Causal Multiple Instance Learning for Whole Slide Image Classification (AAAI 2024 · 21 citations) | |
| Add aggregated features to most-relevant-modality features | CMU-MOSEI aligned · Micro-F1 | 0.563 -> 0.569 | Helped | CARAT: Contrastive Feature Reconstruction and Aggregation for Multi-Modal Multi-Label Emotion Recognition (AAAI 2024 · 19 citations) |
| Select synthetic examples with the filter instead of randomly | SemEval CG · accuracy | 6.93 -> 8.29 | Helped | Compositional Generalization for Multi-Label Text Classification: A Data-Augmentation Approach (AAAI 2024 · 14 citations) |
| Emphasize negative labels in classification loss | PASCAL VOC B4-C2 · last mAP | 22.6 -> 36.3 | Helped | Rebalancing Multi-Label Class-Incremental Learning (AAAI 2025 · 10 citations) |
| add knowledge-based representation refinement with λ = 0.1 | Fashion-MNIST · accuracy | 91.48 ± 0.21% -> 92.01 ± 0.24% | Helped | Distilling Reliable Knowledge for Instance-Dependent Partial Label Learning (AAAI 2024 · 9 citations) |
| Add the regression-based full branch | THUMOS14 · mAP@IoU (0.1:0.7) | 46.4 -> 48.2 | Helped | Bridge the Gap: From Weak to Full Supervision for Temporal Action Localization with PseudoFormer (CVPR 2025 · 7 citations) |
| Add CO and DC group prompts (GVPT) to vanilla VPT | COCO · mAP | 78.20 -> 79.61 | Helped | Correlative and Discriminative Label Grouping for Multi-Label Visual Prompt Tuning (CVPR 2025 · 7 citations) |
| Add removal of high-loss negative labels | MS-COCO, 25% labels left · mAP | Helped | Addressing Multi-Label Learning with Partial Labels: From Sample Selection to Label Selection (AAAI 2025 · 6 citations) | |
| Use QR rather than Q alone for feature selection | CAL, CHD_49, Chess, Corel5K, LLOG_F, Water, and Yeast · Micro-F1 | 0.149 vs 0.538; 0.477 vs 0.637; 0.000 vs 0.579; 0.000 vs 0.263; 0.022 vs 0.339; 0.222 vs 0.556; 0.121 vs 0.496 | Helped | Reconsidering Feature Structure Information and Latent Space Alignment in Partial Multi-label Feature Selection (AAAI 2025 · 6 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 |
|---|---|---|---|---|
| PL-SVM | FG-NET | accuracy | 0.043 – 0.063 | 2 papers, e.g. AAAI 2024 and AAAI 2019 |
| LIFT | Corel5K | average precision | 0.291 – 0.423 | 2 papers, e.g. AAAI 2018 and AAAI 2018 |
| LIFT | Medical | average precision | 0.609 – 0.874 | 2 papers, e.g. AAAI 2018 and AAAI 2018 |
| LIFT | yeast | Average precision | 0.714 – 0.769 | 2 papers, e.g. AAAI 2018 and AAAI 2022 |
| AttentionXML | Wiki-500k | P@1 | 72.62 – 76.74 | 2 papers, e.g. ICML 2020 and NeurIPS 2021 |
What reviewers will expect
Datasets and baselines named in the 53 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- CIFAR-10 · 9 of 53 papers
- CIFAR-100 · 7 of 53 papers
- NUS-WIDE · 6 of 53 papers
- scene · 6 of 53 papers
- MS-COCO · 4 of 53 papers
- PASCAL VOC 2007 · 4 of 53 papers
- Corel5k · 4 of 53 papers
- MS-COCO 2014 · 4 of 53 papers
Baselines
- PRODEN · 6 of 53 papers
- LWS · 5 of 53 papers
- ROLE · 4 of 53 papers
- PiCO · 4 of 53 papers
- CAVL · 4 of 53 papers
- AA-kNN · 3 of 53 papers
- WAN · 3 of 53 papers
- SCL · 3 of 53 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 157 papers; from OpenAlex and the papers' affiliations).
- Southeast University · 28 papers
- Nanjing University of Science and Technology · 8 papers
- Nanjing University of Aeronautics and Astronautics · 7 papers
- Key Laboratory of Computer Network and Information Integration (Southeast University) · 5 papers
- Shanghai Jiao Tong University · 5 papers
- Beihang University · 4 papers
- Amazon · 4 papers
- Zhejiang University · 4 papers
- Hong Kong Baptist University · 4 papers
- The Hong Kong Polytechnic University · 4 papers
Use it
Download BibTeX of 40 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
- Loss design and classifier theory (273 papers)
- Group-robust classification (251 papers)
- Learning from unlabeled and positive-unlabeled data (240 papers)
- Confidence calibration objectives (129 papers)
- Deep active learning (76 papers)
Elsewhere in the topic
- Multilingual language representation pretraining (507 papers)
- Contrastive representation theory (495 papers)
- Dense features and foundation priors (482 papers)
- Correspondence and causal representation learning (447 papers)
- Data augmentation and generalization bias (409 papers)
- Recommendation representation learning (207 papers)
- Margin-based face recognition losses (124 papers)
- Supervised person re-identification (108 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).