Research map / Continual learning / Class-incremental
Exemplar-efficient class-incremental learning
Retains old classes with compact exemplars, embedding distillation and pretrained models.
111 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Class-incremental, keyword cluster “shot · incremental cil · feature”
State History Benchmarks Tried Baselines Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 2 of 2 tracked benchmarks.
40 papers in 2023–24 and 57 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 43.5% to 42.2%.
Open questions the data points to
- CIFAR-100, 10 tasks (average accuracy): best went from 67.42 (up to 2024) to 94.43 in 2025 — Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · 0 citations)
- CIFAR-100, 20 tasks (average accuracy): best went from 62.76 (up to 2024) to 93.48 in 2025 — Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · 0 citations)
- Papers report LUCIR on ImageNet, B=100, S=50 (average incremental accuracy) anywhere from 47.55 to 52.75: 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
- 2017 · iCaRL: Incremental Classifier and Representation Learning (CVPR 2017 · 3,848 citations)
- 2019 · Learning Without Memorizing (CVPR 2019 · 534 citations)
- 2020 · Mnemonics Training: Multi-Class Incremental Learning Without Forgetting (CVPR 2020 · 289 citations)
- 2021 · Adaptive Aggregation Networks for Class-Incremental Learning (CVPR 2021 · 227 citations)
- 2022 · Federated Class-Incremental Learning (CVPR 2022 · 230 citations)
- 2023 · Dense Network Expansion for Class Incremental Learning (CVPR 2023 · 68 citations)
- 2024 · Expandable Subspace Ensemble for Pre-Trained Model-Based Class-Incremental Learning (CVPR 2024 · 117 citations)
- 2025 · Task-Agnostic Guided Feature Expansion for Class-Incremental Learning (CVPR 2025 · 17 citations)
- 2026 · BOFA: Bridge-Layer Orthogonal Low-Rank Fusion for CLIP-Based Class-Incremental Learning (AAAI 2026 · 2 citations)
Earliest papers here
- iCaRL: Incremental Classifier and Representation Learning (CVPR 2017 · 3,848 citations)
- Streaming Classification with Emerging New Class by Class Matrix Sketching (AAAI 2017 · 66 citations)
- Learning Without Memorizing (CVPR 2019 · 534 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-100, 10 tasks · average accuracy (higher is better)
4 papers, 2022–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2022 | 66.9 | Federated Class-Incremental Learning (CVPR 2022 · 230 citations) |
| 2024 | 67.42 | Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning (CVPR 2024 · 7 citations) |
| 2025 | 94.43 | Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · 0 citations) |
CIFAR-100, 20 tasks · average accuracy (higher is better)
4 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 39.32 | Rebalancing Batch Normalization for Exemplar-Based Class-Incremental Learning (CVPR 2023 · 20 citations) |
| 2024 | 62.76 | Task-Adaptive Saliency Guidance for Exemplar-free Class Incremental Learning (CVPR 2024 · 7 citations) |
| 2025 | 93.48 | Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · 0 citations) |
What has been tried
75 papers here report 794 measured changes (a component added, removed or swapped, compared with a baseline): helped 74%, no clear effect 5%, 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 task-specific adapters to the pre-trained model | ImageNet-R B0 Inc20 · accuracy | Helped | Expandable Subspace Ensemble for Pre-Trained Model-Based Class-Incremental Learning (CVPR 2024 · 117 citations) | |
| Use gradient reweighting and DAKD instead of the baseline GR + KD configuration | Food101-LT, LFS N=10 · average accuracy (%) | 27.77 -> 29.05 | Helped | Gradient Reweighting: Towards Imbalanced Class-Incremental Learning (CVPR 2024 · 57 citations) |
| add Dual-Activation Compensation (DAC) to C-RLS | ImageNet-Full, K=5 · last-phase accuracy | 56.11 -> 57.43 | Helped | DS-AL: A Dual-Stream Analytic Learning for Exemplar-Free Class-Incremental Learning (AAAI 2024 · 38 citations) |
| Add feature perturbation | CIFAR-100 B50-5S · average incremental accuracy | 65.07(˘0.28) -> 66.76(˘0.14) | Helped | Class Incremental Learning with Multi-Teacher Distillation (CVPR 2024 · 31 citations) |
| Add FCN to calibrate old-class prototypes | CIFAR-100, 5 stages · top-1 accuracy | 60.51 -> 61.77 | Helped | FCS: Feature Calibration and Separation for Non-Exemplar Class Incremental Learning (CVPR 2024 · 30 citations) |
| Use task-oriented feature generation instead of naive feature generation | CIFAR-100, 4 equally divided tasks · A | 57.30 -> 59.82 | Helped | eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · 21 citations) |
| add continual task-agnostic learning | CIFAR100 10-10 · Last accuracy | 64.45 -> 65.86 | Helped | Task-Agnostic Guided Feature Expansion for Class-Incremental Learning (CVPR 2025 · 17 citations) |
| adapter merging (EMA) plus self-refined retrieval mechanism added to the progressively adapted PTM | ImageNet-R B0 Inc20 · accuracy | Helped | MOS: Model Surgery for Pre-Trained Model-Based Class-Incremental Learning (AAAI 2025 · 15 citations) | |
| remove LTE constraints from data-free generation | CIFAR-100, T=5 and T=10 · accuracy | Hurt | Text-Enhanced Data-free Approach for Federated Class-Incremental Learning (CVPR 2024 · 13 citations) | |
| Add prototypical routing | CIFAR-100 P=5, P=10, and P=20 · average accuracy | 58.1 -> 67.9; 56.8 -> 66.8; 49.5 -> 64.0 | Helped | Dual-Consistency Model Inversion for Non-Exemplar Class Incremental Learning (CVPR 2024 · 12 citations) |
| Add sub-prototype construction loss Lcon | CIFAR100, Shuffled LT-CIL, ρ = 0.01, 5 tasks · Accuracy | 11.3 -> 10.4 | Hurt | Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 2024 · 12 citations) |
| add self-supervised learning loss to base-session training | CIFAR100 · accuracy | 48.6 -> 50.4 | Helped | Non-exemplar Online Class-Incremental Continual Learning via Dual-Prototype Self-Augment and Refinement (AAAI 2024 · 11 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 |
|---|---|---|---|---|
| LUCIR | ImageNet, B=100, S=50 | average incremental accuracy | 47.55 – 52.75 | 2 papers, e.g. AAAI 2024 and CVPR 2022 |
| LUCIR | ImageNet, B=100, S=100 | average incremental accuracy | 52.47 – 56.4 | 2 papers, e.g. AAAI 2024 and CVPR 2022 |
What reviewers will expect
Datasets and baselines named in the 30 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 · 21 of 30 papers
- ImageNet-R · 10 of 30 papers
- ImageNet-A · 8 of 30 papers
- CUB200 · 8 of 30 papers
- TinyImageNet · 6 of 30 papers
- OmniBenchmark · 5 of 30 papers
- ObjectNet · 4 of 30 papers
- VTAB · 4 of 30 papers
Baselines
- L2P · 13 of 30 papers
- iCaRL · 12 of 30 papers
- DualPrompt · 11 of 30 papers
- CODA-Prompt · 11 of 30 papers
- DER · 9 of 30 papers
- SimpleCIL · 9 of 30 papers
- EASE · 9 of 30 papers
- Finetune · 8 of 30 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 97 papers; from OpenAlex and the papers' affiliations).
- Nanjing University · 8 papers
- Chinese Academy of Sciences · 6 papers
- University of Chinese Academy of Sciences · 5 papers
- South China University of Technology · 4 papers
- University of Illinois at Chicago · 4 papers
- Xi’an Jiaotong University · 3 papers
- Xidian University · 3 papers
- University of Illinois Chicago · 3 papers
- Nankai University · 3 papers
- Tencent · 3 papers
Use it
Download BibTeX of 36 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
- Class-balanced knowledge retention (73 papers)
- Few-shot class-incremental prototypes (47 papers)
- Incremental object detection with replay (32 papers)
Elsewhere in the topic
- Unknown-class mining for incremental segmentation (23 papers)
- Distillation-based continual segmentation (17 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).