Research map / Continual learning / Continual fine-tuning of LLMs
Continual Learners are Incremental Model Generalizers
Jaehong Yoon, Sung Ju Hwang, Yue Cao
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use masked modeling instead of supervised continual learning | Base continual learner · CIFAR-100 Split, fine-tuning | Top-1 accuracy: 79.3 -> 88.9 | Improved |
| Add GLAD attention adaptors and regularization | Supervised continual pre-training · ImageNet-1K Split, T9 fine-tuning after T5-T8 | accuracy: 64.56 -> 65.78; 69.06 -> 69.84; 75.06 -> 76.12; 77.80 -> 79.04 | Improved |
| Add GLAD attention adaptors and regularization | SimMIM continual pre-training · ImageNet-1K Split, T9 fine-tuning after T5-T8 | accuracy: 68.34 -> 68.24; 72.22 -> 73.24; 77.38 -> 78.94; 80.12 -> 81.62 | Mixed |
| Add GLAD attention adaptors and regularization | MAE continual pre-training · ImageNet-1K Split, T9 fine-tuning after T5-T8 | accuracy: 42.19 -> 41.75; 52.07 -> 54.74; 63.07 -> 68.01; 71.19 -> 74.30 | Mixed |
| Freeze the two lowest or two deepest layers during subsequent continual learning (ablation) | Supervised continual learning · ImageNet-1K Split, unseen-task fine-tuning | accuracy | Hurt |
| Freeze the two lowest or two deepest layers during subsequent continual learning (ablation) | MIM-based unsupervised continual learning · ImageNet-1K Split, unseen-task fine-tuning | accuracy | No clear effect |
| Train on additional sequential tasks | Continual learners · ImageNet-1K Split, unseen-task T9 | fine-tuning accuracy | Improved |
Closest papers by meaning
- Theory on Forgetting and Generalization of Continual Learning (ICML 2023 · similarity 0.90)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.89)
- Continual Learning with Global Alignment (NeurIPS 2024 · similarity 0.89)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.89)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.89)
- Advancing Prompt-Based Methods for Replay-Independent General Continual Learning (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
