Research map / Label-efficient and robust learning / Recognition and long-tail
Classifier rebalancing and margin losses
Separates representation and classifier learning and applies balancing margins or weight rebalancing.
58 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Recognition and long-tail, keyword cluster “long tailed · classes · tailed recognition”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Few recent reports The tracked benchmarks have at most one report from 2025–26 that could beat the old record, too few to say whether they are still moving.
24 papers in 2023–24 and 21 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 1.9% to 1.6%.
No longer reported by papers here: ImageNet-LT (Top-1 accuracy, last in 2024); iNaturalist 2018 (Top-1 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 · Accelerated Training for Massive Classification via Dynamic Class Selection (AAAI 2018 · 42 citations)
- 2019 · Class-Balanced Loss Based on Effective Number of Samples (CVPR 2019 · 2,852 citations)
- 2020 · Decoupling Representation and Classifier for Long-Tailed Recognition (ICLR 2020 · 218 citations)
- 2021 · Seesaw Loss for Long-Tailed Instance Segmentation (CVPR 2021 · 256 citations)
- 2022 · The Majority Can Help the Minority: Context-Rich Minority Oversampling for Long-Tailed Classification (CVPR 2022 · 213 citations)
- 2023 · Long-Tailed Visual Recognition via Self-Heterogeneous Integration With Knowledge Excavation (CVPR 2023 · 61 citations)
- 2024 · Long-Tailed Partial Label Learning by Head Classifier and Tail Classifier Cooperation (AAAI 2024 · 13 citations)
- 2025 · Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · 5 citations)
- 2026 · Space Alignment Matters: The Missing Piece for Inducing Neural Collapse in Long-Tailed Learning (AAAI 2026 · 1 citation)
Earliest papers here
- Accelerated Training for Massive Classification via Dynamic Class Selection (AAAI 2018 · 42 citations)
- Class-Balanced Loss Based on Effective Number of Samples (CVPR 2019 · 2,852 citations)
- Decoupling Representation and Classifier for Long-Tailed Recognition (ICLR 2020 · 218 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-LT · Top-1 accuracy (higher is better)
5 papers, 2021–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 51.8 | RSG: A Simple but Effective Module for Learning Imbalanced Datasets (CVPR 2021 · 106 citations) |
| 2023 | 60.3 | Long-Tailed Visual Recognition via Self-Heterogeneous Integration With Knowledge Excavation (CVPR 2023 · 61 citations) |
| 2024 | 60.8 | Breaking Long-Tailed Learning Bottlenecks: A Controllable Paradigm with Hypernetwork-Generated Diverse Experts (NeurIPS 2024 · 1 citation) |
CIFAR100-LT (IF 100) · top-1 accuracy (higher is better)
4 papers, 2023–2025 · one report since 2024, not better
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 49.8 | Curvature-Balanced Feature Manifold Learning for Long-Tailed Classification (CVPR 2023 · 48 citations) |
| 2024 | 54.8 | Inverse Weight-Balancing for Deep Long-Tailed Learning (AAAI 2024 · 5 citations) |
iNaturalist 2018 · Top-1 accuracy (higher is better)
4 papers, 2023–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 75.4 | Long-Tailed Visual Recognition via Self-Heterogeneous Integration With Knowledge Excavation (CVPR 2023 · 61 citations) |
What has been tried
44 papers here report 523 measured changes (a component added, removed or swapped, compared with a baseline): helped 73%, no clear effect 6%, hurt 19%, mixed 2%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| use cooperating head and tail classifiers with CWE fusion instead of only the tail classifier | CIFAR10 (η = 0.3, γ = 200) · overall accuracy | 69.31 -> 81.97 | Helped | Long-Tailed Partial Label Learning by Head Classifier and Tail Classifier Cooperation (AAAI 2024 · 13 citations) |
| replace gradient-based grouping with random grouping | ImageNet-LT · top-1 accuracy | 57.49±0.29 -> 56.37±0.41 | Hurt | Long-Tailed Learning as Multi-Objective Optimization (AAAI 2024 · 8 citations) |
| Replace the linear classifier with a cosine classifier only | CIFAR-100-LT (IF 100) · top-1 accuracy | 44.60 -> 44.29; 49.35 -> 49.84 | No clear effect | Difficulty-aware Balancing Margin Loss for Long-tailed Recognition (AAAI 2025 · 5 citations) |
| Add HC to NC using vector cosine similarity | CIFAR100-LT (IF 100) · top-1 accuracy | 54.6 -> 54.8 | Helped | Inverse Weight-Balancing for Deep Long-Tailed Learning (AAAI 2024 · 5 citations) |
| Use random auxiliary categories instead of selected neighbor categories | ImageNet-LT · Overall accuracy | 79.6 -> 77.2 | Hurt | Learning from Neighbors: Category Extrapolation for Long-Tail Learning (CVPR 2025 · 3 citations) |
| Add Long-Tailed Graph Adjustment | Gowalla · Acc@1 | 0.1617 -> 0.1668 | Helped | Taming the Long Tail in Human Mobility Prediction (NeurIPS 2024 · 3 citations) |
| Add the class prior term u(y) | ImageNet-GLT GLT · accuracy | 42.51% -> 44.71% | Helped | Class and Attribute-Aware Logit Adjustment for Generalized Long-Tail Learning (AAAI 2025 · 2 citations) |
| Use Distribution-agnostic Matching | CIFAR-10-LT (ω=200, IPC=50) · accuracy | 14.3% | Helped | Distilling Long-tailed Datasets (CVPR 2025 · 2 citations) |
| Use the two-stage schedule instead of SAM (ext.) | CIFAR-100-LT (IF=100, ρtail=0.10) · Avg accuracy | 47.8 -> 49.3 | Helped | SSE-SAM: Balancing Head and Tail Classes Gradually Through Stage-Wise SAM (AAAI 2025 · 1 citation) |
| Add multiplicative logit adjustment after WD, FR, and ETF one-stage feature training | CIFAR100-LT · Average accuracy (%) | 47.6 (±0.5) -> 53.0 (±0.3) | Helped | Exploring Weight Balancing on Long-Tailed Recognition Problem (ICLR 2024 · 1 citation) |
| Add the hypernetwork | CIFAR100-LT with unknown test class distributions · top-1 accuracy | Helped | Breaking Long-Tailed Learning Bottlenecks: A Controllable Paradigm with Hypernetwork-Generated Diverse Experts (NeurIPS 2024 · 1 citation) | |
| use CLIP, OpenCLIP, and MetaCLIP instead of only OpenCLIP as semantic factors | ImageNet-LT · D-Few accuracy | 71.70 -> 74.99 | Helped | Rethinking the Bias of Foundation Model under Long-tailed Distribution (ICML 2025 · 0 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 |
|---|---|---|---|---|
| CE | ImageNet | Top-1 accuracy | 70.1 – 76 | 2 papers, e.g. ICLR 2023 and CVPR 2023 |
What reviewers will expect
Datasets and baselines named in the 14 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-LT · 8 of 14 papers
- CIFAR-100-LT · 7 of 14 papers
- CIFAR-10-LT · 5 of 14 papers
- Places-LT · 4 of 14 papers
- iNaturalist2018 · 4 of 14 papers
- Tiny-ImageNet-LT · 2 of 14 papers
Baselines
- LA · 3 of 14 papers
- DisAlign · 3 of 14 papers
- CE · 3 of 14 papers
- LDAM-DRW · 2 of 14 papers
- BCL · 2 of 14 papers
- GML · 2 of 14 papers
- RIDE · 2 of 14 papers
- CB · 2 of 14 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 45 papers; from OpenAlex and the papers' affiliations).
- Tianjin University · 3 papers
- Xidian University · 3 papers
- University of Chinese Academy of Sciences · 3 papers
- Korea Advanced Institute of Science and Technology · 3 papers
- University of Science and Technology of China · 3 papers
- Jilin University · 3 papers
- Nanjing University · 2 papers
- Zhejiang University · 2 papers
- Shenzhen University · 2 papers
- Southeast University · 2 papers
Use it
Download BibTeX of 38 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
- Decoupled and calibrated long-tail training (82 papers)
- Contrastive learning for imbalanced data (22 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)
- Partial and multi-label learning (345 papers)
- Loss design and classifier theory (273 papers)
- Group-robust classification (251 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).