Research map / Continual learning / Continual fine-tuning of LLMs
Understanding Catastrophic Forgetting In LoRA via Mean-Field Attention Dynamics
Hugo Koubbi, Louis Hernandez, Matthieu Boussard
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 |
|---|---|---|---|
| increase Frobenius norm of random Gaussian low-rank perturbation added to weights (ablation) | Qwen3-0.6B · next-token loss on base dataset | perturbed loss | Hurt |
| increase rank of random Gaussian perturbation at fixed Frobenius norm (ablation) | Qwen3-0.6B · next-token loss on base dataset | perturbed loss | No clear effect |
| vary LoRA rank during actual training (ablation) | Qwen3-0.6B LoRA · base-dataset perplexity | non-trained perplexity | Mixed |
| continue LoRA fine-tuning (accumulating perturbation) | Qwen3-0.6B LoRA · hidden-state drift vs base model | empirical bifurcation depth (layer): 28 -> 16 | Hurt |
| low-rank LoRA adapters (r=8, 32) during training | Qwen3-0.6B LoRA · alignment with dominant value direction u1 | normalized subspace alignment | Hurt |
Closest papers by meaning
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.88)
- LoRA vs Full Fine-tuning: An Illusion of Equivalence (NeurIPS 2025 · similarity 0.88)
- LoRA Training Provably Converges to a Low-Rank Global Minimum Or It Fails Loudly (But it Probably Won't Fail) (ICML 2025 · similarity 0.87)
- Measuring Representational Shifts in Continual Learning: A Linear Transformation Perspective (ICML 2025 · similarity 0.86)
- Recurrent Self-Attention Dynamics: An Energy-Agnostic Perspective from Jacobians (NeurIPS 2025 · similarity 0.86)
- Beyond Zero Initialization: Investigating the Impact of Non-Zero Initialization on LoRA Fine-Tuning Dynamics (ICML 2025 · similarity 0.86)
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.
