Research map / Continual learning / Continual fine-tuning of LLMs
Retaining Knowledge for Learning with Dynamic Definition
Zichang Liu, Benjamin Coleman, Tianyi Zhang, Anshumali Shrivastava
Read the paperCheck your idea against 137,217 papers
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 |
|---|---|---|---|
| replace final layer or model with RIDDLE instead of naive Update | MLP · MNIST Binary, original test after update | accuracy: 69.1 -> 98.3 | Improved |
| replace EWC with RIDDLE | MLP · MNIST Binary, original test after update | accuracy: 90.9 -> 98.3 | Improved |
| replace final layer with RIDDLE instead of Update-ViT | ViT · CIFAR10, original test after update | accuracy: 74.4 -> 89.0 | Improved |
| replace final layer with RIDDLE instead of EWC-ViT | ViT · CIFAR10, original test after update | accuracy: 84.5 -> 89.0 | Improved |
| replace baseline with RIDDLE instead of Update-ViT | ViT · ImageNet, original test after update | accuracy: 55.6 -> 67.1 | Improved |
| replace baseline with RIDDLE instead of Update-RoBERTa | RoBERTa · News, original test after update | accuracy: 74.0 -> 89.7 | Improved |
| RIDDLE instead of Update-ViT on update test set | ViT · ImageNet, update test set after update | accuracy: 71.5 -> 70.5 | Hurt |
| RIDDLE instead of EWC-RoBERTa on update test set | RoBERTa · News, update test set after update | accuracy: 87.5 -> 86.4 | Hurt |
| add parameter-count-dependent learning rate to RIDDLE (ablation) | RIDDLE-MLP · MNIST Binary, original test after update | accuracy: 94.9 -> 98.3 | Improved |
| use RIDDLE (hash-partitioned parameters) with global learning rate versus Update-MLP (ablation) | RIDDLE-MLP · MNIST Binary, original test after update | accuracy: 69.1 -> 94.9 | Improved |
| use pretrained ViT instead of Resnet18 under naive Update | Update · CIFAR10, original test after update | accuracy: 63.22 -> 74.4 | Improved |
| use RIDDLE instead of NN2 on tabular data | RIDDLE · HAR | accuracy: 0.9123 -> 0.9001 | Hurt |
Closest papers by meaning
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization (NeurIPS 2025 · similarity 0.85)
- System-Aware Unlearning Algorithms: Use Lesser, Forget Faster (ICML 2025 · similarity 0.83)
- Overcoming Generic Knowledge Loss with Selective Parameter Update (CVPR 2024 · similarity 0.83)
- Forgetting, Ignorance or Myopia: Revisiting Key Challenges in Online Continual Learning (NeurIPS 2024 · similarity 0.83)
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed Samples (NeurIPS 2025 · similarity 0.83)
- Language Guided Concept Bottleneck Models for Interpretable Continual Learning (CVPR 2025 · similarity 0.83)
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.
