Research map / Continual learning / Continual fine-tuning of LLMs
KeepLoRA: Continual Learning with Residual Gradient Adaptation
Mao-Lin Luo, Zi-Hao Zhou, Yi-Lin Zhang, Yuanyu Wan, 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 |
|---|---|---|---|
| freeze the LoRA down-projection A (rank 16) instead of standard LoRA (ablation) | KeepLoRA · MTIL ablation | Last: 59.4 -> 69.5 | Improved |
| add gradient-informed initialization of A (ablation) | KeepLoRA · MTIL ablation | Last: 69.5 -> 75.4 | Improved |
| add projection orthogonal to principal weight subspace W_p (ablation) | KeepLoRA · MTIL ablation | Last: 75.4 -> 76.5 | Improved |
| add projection orthogonal to previous task directions M_{t-1} (ablation) | KeepLoRA · MTIL ablation | Average: 70.2 -> 77.2 | Improved |
| add W_p projection on top of M_{t-1} projection (ablation) | KeepLoRA · MTIL ablation | Transfer: 68.1 -> 69.0 | Improved |
| add W_p projection on top of M_{t-1} projection (ablation) | KeepLoRA · MTIL ablation | Last: 86.1 -> 86.1 | No clear effect |
| add W_p projection on top of M_{t-1} projection (ablation) | KeepLoRA · MTIL ablation | Average: 77.2 -> 77.5 | Improved |
| use KeepLoRA instead of SEFE | KeepLoRA · MLLM-DCL (LLaVA) | Last: 59.91 -> 64.41 | Improved |
| use KeepLoRA instead of SEFE | KeepLoRA · UCIT (LLaVA) | Last: 66.48 -> 67.84 | Improved |
| add prototype vector structure extension (KeepLoRA+) | KeepLoRA · MTIL order-I | Last: 86.1 -> 88.2 | Improved |
Later papers that compared with this method
1 later accepted paper reports KeepLoRA as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Spectral Imbalance Causes Forgetting in Low-Rank Continual Adaptation (ICML 2026 · better on 8 of 8 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SEFE — SEFE: Superficial and Essential Forgetting Eliminator for Multimodal Continual Instruction Tuning (ICML 2025) · better on 6 of 6 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
- TreeLoRA: Efficient Continual Learning via Layer-Wise LoRAs Guided by a Hierarchical Gradient-Similarity Tree (ICML 2025 · similarity 0.91)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.90)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.90)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.89)
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.89)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (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.
