Research map / Continual learning / Continual fine-tuning of LLMs
Spectral Imbalance Causes Forgetting in Low-Rank Continual Adaptation
Hao Gu, Mao-Lin Luo, Zi-Hao Zhou, Han-Chen Zhang, Min-Ling Zhang, Tong Wei
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 LoRA with EBLoRA (balanced energy + orthogonality) | LoRA-FT, LLaVA-1.5-7B · UCIT | MFN: 61.4 -> 72.8 | Improved |
| replace LoRA with EBLoRA | LoRA-FT, LLaVA-1.5-7B · MLLM-DCL | MFN: 54.5 -> 66.7 | Improved |
| add energy-balanced optimization (s*U*V factorization) to LoRA (ablation) | LoRA-FT · UCIT | BWT: -15.4 -> -5.1 | Improved |
| add energy-balanced optimization to LoRA (ablation) | LoRA-FT · UCIT | FWT: 26.8 -> 34.3 | Improved |
| add gradient orthogonality to EBO (ablation) | EBO · UCIT | BWT: -5.1 -> -2.6 | Improved |
| add gradient orthogonality to EBO (ablation) | EBO · UCIT | FWT: 34.3 -> 34.0 | No clear effect |
| add depth-aware initialization of s (ablation) | EBO+GO · UCIT | MFN: 72.1 -> 72.8 | Improved |
| replace scalar s with learnable diagonal S (free spectrum) (ablation) | EBO · UCIT | MFN: 70.2 -> 68.1 | Hurt |
| smooth singular values of LoRA adapters before merging (ablation) | LoRA merging · UCIT | NAI | Improved |
| use Newton-Schulz orthogonalization instead of whitening retraction (ablation) | EBLoRA · UCIT | MFN: 72.8 -> 71.6 | Hurt |
| use EBLoRA instead of LoRA | training cost · UCIT | training time (s/it): 3.28 -> 4.20 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- KeepLoRA — KeepLoRA: Continual Learning with Residual Gradient Adaptation (ICLR 2026) · better on 8 of 8 shared results
- SEFE — SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning (ICML 2025) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.91)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.89)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.89)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.89)
- Learning Continually by Spectral Regularization (ICLR 2025 · similarity 0.89)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · 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.
