Research map / Continual learning / Continual fine-tuning of LLMs
Catastrophic Forgetting Meets Negative Transfer: Batch Spectral Shrinkage for Safe Transfer 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 BSS regularizer | L2 fine-tuning · Stanford Cars 15% | top-1 accuracy: 36.77 -> 40.57 | Helped |
| add BSS regularizer | DELTA · CUB-200-2011 15% | top-1 accuracy: 46.83 -> 49.77 | Helped |
| add BSS regularizer | DELTA · Stanford Cars 100% | top-1 accuracy: 86.32 -> 86.32 | No clear effect |
| add BSS regularizer | L2 fine-tuning · Stanford Dogs 15% | top-1 accuracy: 81.05 -> 81.86 | Helped |
| use L2-SP instead of L2 (pull weights to pre-trained) | fine-tuning · MIT Indoors 67 low-data | error rate | Hurt |
| add BSS regularizer | BERT-base · GLUE (4 tasks avg) | accuracy: 88.0 -> 88.9 | Helped |
| add BSS regularizer | EWC incremental learning · Permuted MNIST task A (source) | accuracy: 96.60 -> 96.46 | Hurt |
| truncate smallest r singular components of batch features (larger r) (ablation) | L2 fine-tuning · Stanford Dogs | error rate | Mixed |
| penalize more singular values (larger k) (ablation) | L2+BSS · Stanford Dogs | error rate | Mixed |
Closest papers by meaning
- Upweighting Easy Samples in Fine-Tuning Mitigates Forgetting (ICML 2025 · similarity 0.87)
- ZeroFlow: Overcoming Catastrophic Forgetting is Easier than You Think (ICML 2025 · similarity 0.85)
- Learning Continually by Spectral Regularization (ICLR 2025 · similarity 0.85)
- Mitigating Forgetting in LLM Fine-Tuning via Low-Perplexity Token Learning (NeurIPS 2025 · similarity 0.85)
- SAFE: Slow and Fast Parameter-Efficient Tuning for Continual Learning with Pre-Trained Models (NeurIPS 2024 · similarity 0.85)
- Spectral Reversal: Counteracting Singular Value Bias for Graph Prompting (NeurIPS 2026 · similarity 0.84)
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.