Research map / Continual learning / Online and replay-based
Unified and unsupervised continual frameworks
Proposes general frameworks, benchmarks and architecture-growing methods for task-agnostic and unsupervised continual learning.
120 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Online and replay-based, keyword cluster “weight · transfer · growing”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Few recent reports None of the tracked benchmarks has a report from 2025–26.
38 papers in 2023–24 and 42 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 23.3% to 21.3%.
No longer reported by papers here: Split CIFAR-100 (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 · Progress & Compress: A scalable framework for continual learning (ICML 2018 · 281 citations)
- 2019 · Compacting, Picking and Growing for Unforgetting Continual Learning (NeurIPS 2019 · 158 citations)
- 2020 · A Neural Dirichlet Process Mixture Model for Task-Free Continual Learning (ICLR 2020 · 109 citations)
- 2021 · Curriculum-Meta Learning for Order-Robust Continual Relation Extraction (AAAI 2021 · 69 citations)
- 2022 · Self-Supervised Models Are Continual Learners (CVPR 2022 · 144 citations)
- 2023 · Computationally Budgeted Continual Learning: What Does Matter? (CVPR 2023 · 38 citations)
- 2024 · Continual Relation Extraction via Sequential Multi-Task Learning (AAAI 2024 · 16 citations)
- 2025 · MalCL: Leveraging GAN-Based Generative Replay to Combat Catastrophic Forgetting in Malware Classification (AAAI 2025 · 9 citations)
- 2026 · Unlocking the Potential of Continual Model Merging: An ODE Perspective (ICML 2026 · 0 citations)
Earliest papers here
- Progress & Compress: A scalable framework for continual learning (ICML 2018 · 281 citations)
- Compacting, Picking and Growing for Unforgetting Continual Learning (NeurIPS 2019 · 158 citations)
- Uncertainty-based Continual Learning with Adaptive Regularization (NeurIPS 2019 · 113 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.
Split CIFAR-100 · accuracy (higher is better)
4 papers, 2020–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2020 | 20.1 | A Neural Dirichlet Process Mixture Model for Task-Free Continual Learning (ICLR 2020 · 109 citations) |
| 2022 | 82.3 | Representational Continuity for Unsupervised Continual Learning (ICLR 2022 · 28 citations) |
What has been tried
90 papers here report 909 measured changes (a component added, removed or swapped, compared with a baseline): helped 70%, no clear effect 7%, hurt 18%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| use AUGD instead of no MOO | FewRel, T10 · mean overall accuracy | 83.3 -> 84.8 | Helped | Continual Relation Extraction via Sequential Multi-Task Learning (AAAI 2024 · 16 citations) |
| Add importance sampling to the prototype-based contrastive loss without PRD | Seq-Cifar-10 Class-IL, buffer 200 · accuracy | 56.43±1.54 -> 58.37± 1.17 | Helped | Contrastive Continual Learning with Importance Sampling and Prototype-Instance Relation Distillation (AAAI 2024 · 14 citations) |
| add DSGD | CIFAR100-5, CIFAR100-20, CIFAR100-80 · average accuracy | 31.79 -> 35.91; 45.75 -> 48.44; 53.46 -> 54.40 | Helped | Dynamic Sub-graph Distillation for Robust Semi-supervised Continual Learning (AAAI 2024 · 12 citations) |
| Add FML instead of BCE | EMBER, 11 tasks · mean accuracy | 52.1 -> 54.5 | Helped | MalCL: Leveraging GAN-Based Generative Replay to Combat Catastrophic Forgetting in Malware Classification (AAAI 2025 · 9 citations) |
| Remove replay memory | TinyImagenet · capacity | 93.6% -> 32.15% | Helped | TinySubNets: An Efficient and Low Capacity Continual Learning Strategy (AAAI 2025 · 6 citations) |
| add Tail Anchor | CIFAR-100, Task 5 · accuracy (%) | 72.3 -> 89.4 | Helped | Handling Spatial-Temporal Data Heterogeneity for Federated Continual Learning via Tail Anchor (CVPR 2025 · 6 citations) |
| Add reliable weight consolidation during tuning | CIFAR-100 · OOD detection AUROC | 71.97 -> 82.24 | Helped | RCL: Reliable Continual Learning for Unified Failure Detection (CVPR 2024 · 6 citations) |
| Select random, highest-attention, and lowest-attention patches rather than only highest-attention patches | Camelyon-TCGA · AACC | 0.595±0.121 -> 0.729±0.041 | Helped | Advancing Multiple Instance Learning with Continual Learning for Whole Slide Imaging (CVPR 2025 · 5 citations) |
| replace KAC spline functions with a trainable MLP | ImageNet-R, 20 steps · Avg | 83.59 -> 80.56 | Hurt | KAC: Kolmogorov-Arnold Classifier for Continual Learning (CVPR 2025 · 5 citations) |
| Decrease the expansion threshold γ | Split CIFAR10 · accuracy | No clear effect | Task-Free Dynamic Sparse Vision Transformer for Continual Learning (AAAI 2024 · 5 citations) | |
| Set unlearning rate gamma to 0.02, 0.03, or 0.04 | CIFAR-100, CIFAR-10, and Tiny-ImageNet Task-IL · accuracy | CIFAR100: 77.23 ± 0.97; 77.71 ± 0.85; 77.08 ± 0.90. CIFAR10: 94.27 ± 0.42; 94.64 ± 0.38; 94.82 ± 0.51. Tiny-ImageNet: 53 | Mixed | A Unified and General Framework for Continual Learning (ICLR 2024 · 3 citations) |
| Set λ=0.2 instead of λ=0.8 | CIFAR10, α=0.1 · test accuracy | 39.93 -> 42.58 | Helped | FedSSI: Rehearsal-Free Continual Federated Learning with Synergistic Synaptic Intelligence (ICML 2025 · 2 citations) |
What reviewers will expect
Datasets and baselines named in the 20 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 · 7 of 20 papers
- Tiny-ImageNet · 6 of 20 papers
- ImageNet-R · 4 of 20 papers
- CIFAR10 · 3 of 20 papers
- Permuted MNIST · 2 of 20 papers
- split CIFAR100 · 2 of 20 papers
- Imagenet100 · 2 of 20 papers
- CUB200 · 2 of 20 papers
Baselines
- EWC · 7 of 20 papers
- ER · 5 of 20 papers
- LwF · 5 of 20 papers
- DER++ · 4 of 20 papers
- Fine-tuning · 4 of 20 papers
- SI · 3 of 20 papers
- PackNet · 3 of 20 papers
- FedAvg · 3 of 20 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 80 papers; from OpenAlex and the papers' affiliations).
- University of York · 4 papers
- Mohamed bin Zayed University of Artificial Intelligence · 4 papers
- Sun Yat-sen University · 3 papers
- JD Explore Academy · 3 papers
- Seoul National University · 3 papers
- Tsinghua University · 3 papers
- Peng Cheng Laboratory · 2 papers
- Tencent · 2 papers
- Hong Kong University of Science and Technology · 2 papers
- Tianjin University · 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
- Theoretical analysis of forgetting (141 papers)
- Gradient projection and plasticity preservation (80 papers)
- Dynamic expansion and isolation (70 papers)
- Plasticity-preserving continual reinforcement learning (43 papers)
Elsewhere in the topic
- Memory-based and Bayesian online learning (56 papers)
- Plug-and-play online training enhancements (39 papers)
- Replay buffer construction and update (33 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).