Research map / Continual learning / Continual fine-tuning of LLMs
Task-Driven Subspace Decomposition for Knowledge Sharing and Isolation in LoRA-based Continual Learning
Lingfeng He, De Cheng, Huaijie Wang, Xi Yang, Nannan Wang, Xinbo Gao
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 |
|---|---|---|---|
| add general-subspace LoRA branch to single-branch LoRA (ablation) | sequential LoRA · ImageNetR/ImageNetA, 10 and 20 sessions | ALast: 72.09 -> 79.19 (10S-ImageNetR) | Improved |
| add isolated-subspace LoRA branch to single-branch LoRA (ablation) | sequential LoRA · ImageNetR/ImageNetA, 10 and 20 sessions | ALast: 72.09 -> 79.22 (10S-ImageNetR) | Improved |
| combine general and isolated branches (ablation) | LoDA · ImageNetR/ImageNetA, 10 and 20 sessions | ALast: 79.22 -> 81.05 (10S-ImageNetR) | Improved |
| add Gradient-Aligned Optimization (GAO) (ablation) | LoDA dual-branch · ImageNetR/ImageNetA, 10 and 20 sessions | ALast: 81.05 -> 81.93 (10S-ImageNetR) | Improved |
| build isolated subspace by maximizing new-to-old relative energy instead of null space (InfLoRA) (ablation) | LoRAI · 10S-ImageNetR | ALast: 75.15 -> 79.22 | Improved |
| build isolated subspace by relative energy maximization instead of null space (InfLoRA) (ablation) | LoRAI · 10S-ImageNetA | ALast: 57.10 -> 59.89 | Improved |
| use relative energy maximization instead of random orthogonal bases (ablation) | LoRAI · 10S-ImageNetR | ALast: 68.23 -> 79.22 | Improved |
| use closed-form rescaling instead of CoMA merging for general branch (ablation) | LoRAG · 10S-ImageNetR | ALast: 78.39 -> 79.19 | Improved |
| use per-unit closed-form rescaling instead of simple averaging (ablation) | LoRAG · 10S-ImageNetR | ALast: 78.43 -> 79.19 | Improved |
| raise rescaling lambda from 3.0 to 10.0 (ablation) | closed-form rescaling · 10S-ImageNetR / 10S-ImageNetA | ALast: 79.19 -> 78.96 (ImageNetR); 59.62 -> 59.75 (ImageNetA) | No clear effect |
| add classifier alignment (feature replay) to LoDA | LoDA · 10S-ImageNetA | ALast: 62.59 -> 66.71 | Improved |
| add classifier alignment (feature replay) to LoDA | LoDA · 10S-CIFAR100 | ALast: 90.47 -> 92.15 | Improved |
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.
- SD-LoRA — SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning (ICLR 2025) · better on 2 of 2 shared results
- InfLoRA — InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning (CVPR 2024) · better on 1 of 1 shared result
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
- CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning (CVPR 2025 · similarity 0.92)
- SD-LoRA: Scalable Decoupled Low-Rank Adaptation for Class Incremental Learning (ICLR 2025 · similarity 0.92)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.91)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.91)
- LoRA Subtraction for Drift-Resistant Space in Exemplar-Free Continual Learning (CVPR 2025 · similarity 0.89)
- When the Merge Coefficient Stops Mattering: Proximity Regularized Merging for Continual LoRA Adaptation (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.
