Research map / Label-efficient and robust learning / Semi-supervised learning
Safe and realistic semi-supervised learning
Evaluates and debiases use of unlabeled samples, including moderately confident ones.
70 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Semi-supervised learning, keyword cluster “semi · examples · classification”
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.
22 papers in 2023–24 and 10 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 13.8% to 6.3%.
Open questions the data points to
- Papers report FixMatch on CIFAR-10, 40 labels (error rate (%)) anywhere from 7.47 to 13.81: 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 · Towards Safe Semi-Supervised Learning for Multivariate Performance Measures (AAAI 2016 · 30 citations)
- 2017 · Variational Autoencoder for Semi-Supervised Text Classification (AAAI 2017 · 192 citations)
- 2018 · Realistic Evaluation of Deep Semi-Supervised Learning Algorithms (NeurIPS 2018 · 205 citations)
- 2019 · Joint Semi-Supervised Feature Selection and Classification through Bayesian Approach (AAAI 2019 · 41 citations)
- 2020 · Semi-Supervised Learning under Class Distribution Mismatch (AAAI 2020 · 135 citations)
- 2021 · Dash: Semi-Supervised Learning with Dynamic Thresholding (ICML 2021 · 53 citations)
- 2022 · CoSSL: Co-Learning of Representation and Classifier for Imbalanced Semi-Supervised Learning (CVPR 2022 · 54 citations)
- 2023 · FlatMatch: Bridging Labeled Data and Unlabeled Data with Cross-Sharpness for Semi-Supervised Learning (NeurIPS 2023 · 11 citations)
- 2024 · ANEDL: Adaptive Negative Evidential Deep Learning for Open-Set Semi-supervised Learning (AAAI 2024 · 13 citations)
- 2025 · Semi-Supervised Online Cross-Modal Hashing (AAAI 2025 · 6 citations)
- 2026 · Let the Void Be Void: Robust Open-Set Semi-Supervised Learning via Selective Non-Alignment (AAAI 2026 · 0 citations)
Earliest papers here
- Towards Safe Semi-Supervised Learning for Multivariate Performance Measures (AAAI 2016 · 30 citations)
- Variational Autoencoder for Semi-Supervised Text Classification (AAAI 2017 · 192 citations)
- Learning Safe Prediction for Semi-Supervised Regression (AAAI 2017 · 63 citations)
What has been tried
65 papers here report 711 measured changes (a component added, removed or swapped, compared with a baseline): helped 72%, no clear effect 7%, hurt 16%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| remove adaptive loss weights while retaining negative optimization | CIFAR-100 · AUROC | 90.3±0.1 -> 89.0±0.3 | Hurt | ANEDL: Adaptive Negative Evidential Deep Learning for Open-Set Semi-supervised Learning (AAAI 2024 · 13 citations) |
| Add CamMix | CIFAR10-LT · test accuracy | 74.0 -> 75.7 | Helped | BEM: Balanced and Entropy-based Mix for Long-Tailed Semi-Supervised Learning (CVPR 2024 · 11 citations) |
| replace one-hot labels with discriminative Hadamard embedding | IAPR TC-12, Image2Text and Text2Image, 64 bits · mAP@all | 0.4463 -> 0.4992; 0.3380 -> 0.6306 | Helped | Semi-Supervised Online Cross-Modal Hashing (AAAI 2025 · 6 citations) |
| Add NDNS target selection | Office-Home A→C · accuracy | 64.0 -> 64.0 (1-shot); 65.8 -> 69.9 (3-shot) | Mixed | Enhancing Semi-supervised Domain Adaptation via Effective Target Labeling (AAAI 2024 · 6 citations) |
| Increase data heterogeneity from γ=0.1 to γ=0.05 | CIFAR-10 and SVHN · accuracy | 70.06 -> 43.82; 62.94 -> 43.22 | Hurt | Robust Training of Federated Models with Extremely Label Deficiency (ICLR 2024 · 5 citations) |
| add conditional self-labeling | CIFAR-100 · novel accuracy | Helped | OwMatch: Conditional Self-Labeling with Consistency for Open-World Semi-Supervised Learning (NeurIPS 2024 · 3 citations) | |
| Add the forward-and-backward strategy to the labeled-only baseline | Flowers · accuracy | Helped | Learning Textual Prompts for Open-World Semi-Supervised Learning (CVPR 2025 · 2 citations) | |
| Add feature decoupling loss Ldec | MVSA-S and MVSA-M · accuracy | Helped | Seek Common Ground While Reserving Differences: Semi-Supervised Image-Text Sentiment Recognition (CVPR 2025 · 2 citations) | |
| Add similarity loss Ls | CIFAR100 #100, CoreSet selection · accuracy (%) | 56.43 -> 58.72 | Helped | Towards Cost-Effective Learning: A Synergy of Semi-Supervised and Active Learning (CVPR 2025 · 2 citations) |
| Add embedding fusion | CIFAR-10, 40 labels, CNN · error rate (%) | 4.82±0.12 -> 4.51±0.01 | Helped | InterLUDE: Interactions between Labeled and Unlabeled Data to Enhance Semi-Supervised Learning (ICML 2024 · 2 citations) |
| replace PCA with sparse PCA (IPU) after labeled screening | Sparse Gaussian classification simulations · support recovery accuracy | Helped | Semi-Supervised Sparse Gaussian Classification: Provable Benefits of Unlabeled Data (NeurIPS 2024 · 2 citations) | |
| add discriminator filtering of unlabeled data | ImageNet-30 · accuracy | 72.65 -> 83.60 | Helped | Data Augmentation with Diffusion for Open-Set Semi-Supervised Learning (NeurIPS 2024 · 1 citation) |
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 |
|---|---|---|---|---|
| FixMatch | CIFAR-10, 40 labels | error rate (%) | 7.47 – 13.81 | 2 papers, e.g. ICML 2022 and CVPR 2022 |
What reviewers will expect
Datasets and baselines named in the 8 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-100 · 4 of 8 papers
- CUB · 2 of 8 papers
- CIFAR-10 · 2 of 8 papers
Baselines
- FixMatch · 3 of 8 papers
- FreeMatch · 3 of 8 papers
- SoftMatch · 3 of 8 papers
- Supervised · 2 of 8 papers
- FlexMatch · 2 of 8 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 32 papers; from OpenAlex and the papers' affiliations).
- University of Science and Technology of China · 3 papers
- The University of Hong Kong · 2 papers
- Shanghai Jiao Tong University · 2 papers
- Nanjing University · 2 papers
- MBZUAI · 2 papers
- Mohamed bin Zayed University of Artificial Intelligence · 2 papers
- Southeast University · 2 papers
- The University of Sydney · 2 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
- Holistic semi-supervised pipelines (125 papers)
- Self-training theory (86 papers)
- GAN-based semi-supervised learning (61 papers)
Elsewhere in the topic
- Confidence-based pseudo-label selection (134 papers)
- Semi-supervised medical segmentation (50 papers)
- Consistency training for semi-supervised segmentation (49 papers)
- Pseudo-labeling for specialized tasks (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).