Research map / Label-efficient and robust learning / Semi-supervised learning
Confidence-based pseudo-label selection
Selects reliable pseudo-labels using uncertainty, local variance and bias-adaptive classifiers.
134 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Semi-supervised learning, keyword cluster “pseudo labels · segmentation · pseudo labeling”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 6 of 6 tracked benchmarks.
37 papers in 2023–24 and 50 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 23.3% to 31.6%.
Open questions the data points to
- CIFAR-10, 4000 labels (error rate (%)): best went from 3.8 (up to 2024) to 3.38 in 2025 — RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations)
- CIFAR-100 400 labels (error rate (%)): best went from 35.87 (up to 2024) to 35.27 in 2025 — RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations)
- CIFAR-10 40 labels (error rate (%)): best went from 4.87 (up to 2024) to 4.24 in 2025 — RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations)
- CIFAR-10, 250 labels (error rate (%)): best went from 4.78 (up to 2024) to 4.21 in 2025 — RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations)
- CIFAR-100, 10000 labels (error rate (%)): best went from 21.68 (up to 2024) to 19.41 in 2025 — RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations)
- CIFAR-100, 2500 labels (error rate (%)): best went from 26.47 (up to 2024) to 23.78 in 2025 — RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations)
- Papers report UDA on CIFAR-10 40 labels (error rate (%)) anywhere from 10.62 to 29.05: 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 · Robust Semi-Supervised Learning through Label Aggregation (AAAI 2016 · 54 citations)
- 2017 · Semi-Supervised Adaptive Label Distribution Learning for Facial Age Estimation (AAAI 2017 · 41 citations)
- 2018 · Learning From Semi-Supervised Weak-Label Data (AAAI 2018 · 53 citations)
- 2019 · Learning to Self-Train for Semi-Supervised Few-Shot Classification (NeurIPS 2019 · 232 citations)
- 2020 · Rethinking the Value of Labels for Improving Class-Imbalanced Learning (NeurIPS 2020 · 211 citations)
- 2021 · Meta Pseudo Labels (CVPR 2021 · 520 citations)
- 2022 · ACPL: Anti-Curriculum Pseudo-Labelling for Semi-Supervised Medical Image Classification (CVPR 2022 · 131 citations)
- 2023 · FreeMatch: Self-adaptive Thresholding for Semi-supervised Learning (ICLR 2023 · 140 citations)
- 2024 · Pseudo-Label Calibration Semi-supervised Multi-Modal Entity Alignment (AAAI 2024 · 22 citations)
- 2025 · CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · 10 citations)
- 2026 · Neural Collapse Priors Driven Trust Semi-Supervised Multi-View Classification (AAAI 2026 · 1 citation)
Earliest papers here
- Robust Semi-Supervised Learning through Label Aggregation (AAAI 2016 · 54 citations)
- Semi-Supervised Adaptive Label Distribution Learning for Facial Age Estimation (AAAI 2017 · 41 citations)
- Discriminative Semi-Supervised Dictionary Learning with Entropy Regularization for Pattern Classification (AAAI 2017 · 18 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.
CIFAR-10, 4000 labels · error rate (%) (lower is better)
8 papers, 2020–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 4.26 | FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020) |
| 2021 | 3.8 | All Labels Are Not Created Equal: Enhancing Semi-Supervision via Label Grouping and Co-Training (CVPR 2021 · 57 citations) |
| 2025 | 3.38 | RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations) |
CIFAR-100 400 labels · error rate (%) (lower is better)
7 papers, 2020–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 48.85 | FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020) |
| 2021 | 39.94 | FlexMatch: Boosting Semi-Supervised Learning with Curriculum Pseudo Labeling (NeurIPS 2021 · 476 citations) |
| 2023 | 37.98 | FreeMatch: Self-adaptive Thresholding for Semi-supervised Learning (ICLR 2023 · 140 citations) |
| 2024 | 35.87 | Leveraging Local Variance for Pseudo-Label Selection in Semi-supervised Learning (AAAI 2024 · 5 citations) |
| 2025 | 35.27 | RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations) |
CIFAR-10 40 labels · error rate (%) (lower is better)
6 papers, 2020–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 11.39 | FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020) |
| 2021 | 4.97 | FlexMatch: Boosting Semi-Supervised Learning with Curriculum Pseudo Labeling (NeurIPS 2021 · 476 citations) |
| 2023 | 4.9 | FreeMatch: Self-adaptive Thresholding for Semi-supervised Learning (ICLR 2023 · 140 citations) |
| 2024 | 4.87 | Leveraging Local Variance for Pseudo-Label Selection in Semi-supervised Learning (AAAI 2024 · 5 citations) |
| 2025 | 4.24 | RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations) |
CIFAR-10, 250 labels · error rate (%) (lower is better)
6 papers, 2020–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 5.07 | FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020) |
| 2021 | 4.78 | DP-SSL: Towards Robust Semi-supervised Learning with A Few Labeled Samples (NeurIPS 2021 · 6 citations) |
| 2025 | 4.21 | RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations) |
CIFAR-100, 10000 labels · error rate (%) (lower is better)
6 papers, 2020–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 22.6 | FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020) |
| 2021 | 22.24 | DP-SSL: Towards Robust Semi-supervised Learning with A Few Labeled Samples (NeurIPS 2021 · 6 citations) |
| 2023 | 21.68 | FreeMatch: Self-adaptive Thresholding for Semi-supervised Learning (ICLR 2023 · 140 citations) |
| 2025 | 19.41 | RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations) |
CIFAR-100, 2500 labels · error rate (%) (lower is better)
6 papers, 2020–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 28.29 | FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence (NeurIPS 2020) |
| 2021 | 28 | DP-SSL: Towards Robust Semi-supervised Learning with A Few Labeled Samples (NeurIPS 2021 · 6 citations) |
| 2023 | 26.47 | FreeMatch: Self-adaptive Thresholding for Semi-supervised Learning (ICLR 2023 · 140 citations) |
| 2025 | 23.78 | RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations) |
What has been tried
104 papers here report 1,172 measured changes (a component added, removed or swapped, compared with a baseline): helped 72%, no clear effect 7%, hurt 17%, 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 image modality | FB15K-DB15K (20% seeds) · Hits@1 | 0.6510 -> 0.6763 | Helped | Pseudo-Label Calibration Semi-supervised Multi-Modal Entity Alignment (AAAI 2024 · 22 citations) |
| remove CDMAD refinement of training pseudo-labels | CIFAR-10-LT, γl = 100, γu = 1 · bACC | 87.5 -> 78.2 | Hurt | CDMAD: Class-Distribution-Mismatch-Aware Debiasing for Class-Imbalanced Semi-Supervised Learning (CVPR 2024 · 20 citations) |
| Enable augmentation in the second branch while leaving the first branch unaugmented | CIFAR-10 · NMI | 0.752 ± 0.003 -> 0.754 ± 0.003 | Helped | RPSC: Robust Pseudo-Labeling for Semantic Clustering (AAAI 2024 · 19 citations) |
| Replace confidence-aware k selection with any tested fixed k | Semi-Fungi, from scratch · Top-1 accuracy | 39.4 -> 36.5; 39.4 -> 35.8; 39.4 -> 36.0; 39.4 -> 36.6 | Hurt | Roll with the Punches: Expansion and Shrinkage of Soft Label Selection for Semi-supervised Fine-Grained Learning (AAAI 2024 · 12 citations) |
| Replace Self-Adaptive Thresholding with Global Thresholding | CIFAR10, CIFAR100, SVHN, and STL10 · top-1 error rate (%) | CIFAR10-4: 5.19±0.22 -> 4.87±0.18; SVHN-25: 5.57±1.72 -> 2.13±0.05; STL10-10: 16.38±2.19 -> 16.23±2.39; CIFAR10-25: 4.83 | Mixed | CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · 10 citations) |
| Add three experts to FixMatch | CIFAR-10-LT, γu=100 · top-1 accuracy | 76.49 -> 84.02 | Helped | Three Heads Are Better than One: Complementary Experts for Long-Tailed Semi-supervised Learning (AAAI 2024 · 10 citations) |
| add Generative Adversarial Networks for feature extraction | Yeast · AP | 71.93 -> 75.51 | Helped | Semi-Supervised Multi-View Multi-Label Learning with View-Specific Transformer and Enhanced Pseudo-Label (AAAI 2025 · 9 citations) |
| Add contrastive loss with RNS | CIFAR10-LT · accuracy | 83.95 -> 84.30 | Helped | BaCon: Boosting Imbalanced Semi-supervised Learning via Balanced Feature-Level Contrastive Learning (AAAI 2024 · 9 citations) |
| add mixed samples to clean-sample training | STL-10, 40 labels · error rate (%) | 15.56 -> 11.74 | Helped | RegMixMatch: Optimizing Mixup Utilization in Semi-Supervised Learning (AAAI 2025 · 7 citations) |
| Add model bias correction to FixMatch | γu=100 · accuracy (%) | 67.8 -> 79.2 | Helped | Twice Class Bias Correction for Imbalanced Semi-supervised Learning (AAAI 2024 · 6 citations) |
| Add CPG to FedAvg+FixMatch-LPL | CIFAR-100, α = 0.1/0.5/1 · accuracy | 49.32 -> 52.25; 49.67 -> 53.85; 49.55 -> 53.50 | Helped | Mind the Gap: Confidence Discrepancy Can Guide Federated Semi-Supervised Learning Across Pseudo-Mismatch (CVPR 2025 · 5 citations) |
| add DCC, CGPL, and PGLS to the labeled-only baseline | DVM, CAD, and Infarction, 1% labeled data · accuracy (%) / AUC (%) | 82.98 -> 91.92; 75.75 -> 83.54; 73.70 -> 82.64 | Helped | STiL: Semi-supervised Tabular-Image Learning for Comprehensive Task-Relevant Information Exploration in Multimodal Classification (CVPR 2025 · 5 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 |
|---|---|---|---|---|
| ReMixMatch | CIFAR-100, 2500 labels | error rate (%) | 26.03 – 27.43 | 3 papers, e.g. AAAI 2024 and NeurIPS 2020 |
| UDA | CIFAR-10 40 labels | error rate (%) | 10.62 – 29.05 | 2 papers, e.g. NeurIPS 2021 and NeurIPS 2020 |
| FlexMatch | STL-10, 40 labels | top-1 error rate (%) | 11.82 – 29.15 | 2 papers, e.g. ICLR 2024 and ICLR 2023 |
| ReMixMatch | CIFAR-100 (10000 labels) | error rate | 20.02 – 23.03 | 2 papers, e.g. ICLR 2023 and NeurIPS 2020 |
| FixMatch | CIFAR100 (400 labels) | Top-1 accuracy | 51.15 – 57.45 | 2 papers, e.g. CVPR 2022 and CVPR 2023 |
What reviewers will expect
Datasets and baselines named in the 33 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 · 11 of 33 papers
- CIFAR-100 · 10 of 33 papers
- STL-10 · 7 of 33 papers
- SVHN · 6 of 33 papers
- ImageNet · 4 of 33 papers
- CIFAR-100-LT · 4 of 33 papers
- CIFAR-10-LT · 3 of 33 papers
- SVHN-LT · 3 of 33 papers
Baselines
- FixMatch · 13 of 33 papers
- FreeMatch · 9 of 33 papers
- FlexMatch · 8 of 33 papers
- SoftMatch · 8 of 33 papers
- Mean Teacher · 6 of 33 papers
- ReMixMatch · 6 of 33 papers
- MixMatch · 5 of 33 papers
- UDA · 4 of 33 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 87 papers; from OpenAlex and the papers' affiliations).
- Southeast University · 9 papers
- Peking University · 6 papers
- Jilin University · 5 papers
- Nanjing University · 5 papers
- Tsinghua University · 5 papers
- Google · 4 papers
- City University of Hong Kong · 4 papers
- Mohamed bin Zayed University of Artificial Intelligence · 4 papers
- Xi’an Jiaotong University · 4 papers
- Zhejiang 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
- Consistency training for semi-supervised segmentation (49 papers)
- Semi-supervised medical segmentation (50 papers)
- Pseudo-labeling for specialized tasks (48 papers)
Elsewhere in the topic
- Holistic semi-supervised pipelines (125 papers)
- Self-training theory (86 papers)
- Safe and realistic semi-supervised learning (70 papers)
- GAN-based semi-supervised learning (61 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).