Research map / Continual learning / Continual fine-tuning of LLMs
Achieving Forgetting Prevention and Knowledge Transfer in Continual Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add CTR's KSM and TSM instead of using Adapter-BERT alone (ablation) | CTR with BERT · ASC | MF1: 0.4481 -> 0.8362 | Helped |
| Add TSM for task-specific protection (ablation) | CTR with BERT · ASC | MF1: 0.7107 -> 0.8362 | Helped |
| Add KSM for cross-task knowledge sharing (ablation) | CTR with BERT · ASC | MF1: 0.7852 -> 0.8362 | Helped |
| Add the task router to KSM (ablation) | CTR with BERT · ASC | MF1: 0.8155 -> 0.8362 | Helped |
| Use transfer routing rather than dynamic routing (ablation) | CTR with BERT · ASC | MF1: 0.8140 -> 0.8362 | Helped |
| Place the CL-plugin within BERT rather than on top (ablation) | CTR with BERT · ASC | MF1: 0.6390 -> 0.8362 | Helped |
| Insert paired CL-plugins rather than only after the first FF layer (ablation) | CTR with BERT · ASC | MF1: 0.8014 -> 0.8362 | Helped |
| Insert paired CL-plugins rather than only after the other two FF layers (ablation) | CTR with BERT · ASC | MF1: 0.7863 -> 0.8362 | Helped |
| Add TSM for task-specific protection (ablation) | CTR with BERT · 20News | MF1: 0.8064 -> 0.9523 | Helped |
| Add KSM for knowledge transfer (ablation) | CTR with BERT · 20News | MF1: 0.9521 -> 0.9523 | Helped |
| Add the task router to KSM (ablation) | CTR with BERT · 20News | MF1: 0.9469 -> 0.9523 | Helped |
| Use transfer routing rather than dynamic routing (ablation) | CTR with BERT · 20News | MF1: 0.9504 -> 0.9523 | Helped |
Closest papers by meaning
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.89)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.89)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.89)
- Spurious Forgetting in Continual Learning of Language Models (ICLR 2025 · similarity 0.89)
- Improving Plasticity in Online Continual Learning via Collaborative Learning (CVPR 2024 · similarity 0.89)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.89)
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.