Research map / Continual learning / Continual fine-tuning of LLMs
Sparse Orthogonal Parameters Tuning for Continual Learning
Kun-Peng Ning, Hai-Jian Ke, Yuyang Liu, Jiayu Yao, Yonghong Tian, Yuan Li
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 |
|---|---|---|---|
| use SoTU delta masking and merging instead of RanPAC | ViT-B/16-IN21K · Cars196 (10 tasks) | final accuracy: +3.7% | Improved |
| use SoTU instead of RanPAC | ViT-B/16-IN21K · ImageNet-A (10 tasks) | final accuracy: +2.3% | Improved |
| use SoTU instead of RanPAC | ViT-B/16-IN21K · ImageNet-R (10 tasks) | final accuracy: +1.6% | Improved |
| use SoTU features without nonlinear projection instead of RanPAC features | NCM classifier, buffer 10/cls · ImageNet-R (10 tasks, NCM feature space) | final accuracy: 52.2% -> 67.9% (+12.7%) | Improved |
| use SoTU features without nonlinear projection instead of RanPAC features | NCM classifier, buffer 10/cls · CIFAR100 (10 tasks, NCM feature space) | final accuracy: 75.4% -> 83.8% (+8.4%) | Improved |
| use SoTU features without nonlinear projection instead of RanPAC features | NCM classifier, buffer 10/cls · CUB200 (10 tasks, NCM feature space) | final accuracy: 74.6% -> 78.8% (+4.2%) | Improved |
| merge deltas with random masking (high sparsity) instead of low sparsity (ablation) | SoTU delta merging · CIFAR10, CIFAR100, Tiny-ImageNet | accuracy | Improved |
| randomly mask 80%-90% of delta parameters (ablation) | single fine-tuned ViT · CIFAR10, CIFAR100, Tiny-ImageNet | accuracy: 80%~90% masking keeps competitive performance | No clear effect |
| use larger ViT (ViT-L vs ViT-B vs ViT-S) for delta masking (ablation) | delta masking · CIFAR10, CIFAR100, Tiny-ImageNet | accuracy: ViT-L tolerates most sparsity, then ViT-B, then ViT-S | Improved |
| apply delta masking to ResNet50 instead of ViT (ablation) | delta masking · CIFAR10, CIFAR100, Tiny-ImageNet | accuracy | Hurt |
| mask excessive delta parameters when merging (ablation) | SoTU delta merging · CIFAR10, CIFAR100, Tiny-ImageNet | accuracy | Hurt |
| use masking rate p=0.7 (30% delta kept) instead of p=0.9 over 10 tasks (ablation) | SoTU · CIFAR100 (10 tasks) | accuracy: better in first 3 tasks, fails in following 7 | Mixed |
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.
- RanPAC — RanPAC: Random Projections and Pre-trained Models for Continual Learning (NeurIPS 2023) · better on 12 of 12 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
- BiLoRA: Almost-Orthogonal Parameter Spaces for Continual Learning (CVPR 2025 · similarity 0.90)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.90)
- SAFE: Slow and Fast Parameter-Efficient Tuning for Continual Learning with Pre-Trained Models (NeurIPS 2024 · similarity 0.89)
- Task-aware Orthogonal Sparse Network for Exploring Shared Knowledge in Continual Learning (ICML 2024 · similarity 0.89)
- Improving Continual Learning Performance and Efficiency with Auxiliary Classifiers (ICML 2025 · similarity 0.88)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · similarity 0.88)
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.
