Research map / Continual learning / Continual fine-tuning of LLMs
Gradient Projection for Continual Parameter-Efficient Tuning
Jingyang Qiao, Zhizhong Zhang, Xin Tan, Yanyun Qu, Wensheng Zhang, Zhi Han, Yuan Xie
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 orthogonal gradient projection (PEGP) to Adapter-10 | LAE Adapter-10 (ViT) · 10-Split-CIFAR100 (class-incremental) | average accuracy: 85.10 -> 85.43 | Improved |
| add gradient projection to Adapter-5 on ImageNet-R | LAE Adapter-5 (ViT) · 10-Split-ImageNet-R (class-incremental) | forgetting: 9.62 -> 5.62 | Improved |
| add gradient projection to the plain Adapter-10 baseline | Adapter-10 baseline · 10-Split-CIFAR100 (class-incremental) | average accuracy: 79.17 -> 85.43 | Improved |
| add gradient projection under one-epoch online training | LAE Adapter-5 · 10-Split-CIFAR100 (online class-incremental) | forgetting: 11.44 -> 6.82 | Improved |
| add gradient projection to LoRA-5 in online setting | LAE LoRA-5 · 10-Split-CIFAR100 (online class-incremental) | average accuracy: 78.30 -> 83.21 | Improved |
| add PEGP to CLIP prompt tuning (CoOP) | CoOP Prompt (CLIP) · 10-Split-CIFAR100 (class-incremental, CLIP) | average accuracy: 73.76 -> 82.36 | Improved |
| add PEGP to CLIP linear adapter | CoOP Adapter (CLIP) · 10-Split-CIFAR100 (class-incremental, CLIP) | average accuracy: 63.95 -> 70.37 | Improved |
| add PEGP on task-incremental CLIP, trading plasticity for stability | CLIP-EWC Prompt · 10-Split-CIFAR100 (task-incremental, CLIP) | average accuracy: 94.42 -> 93.99 | Mixed |
| add PEGP to Adapter-10 on DomainNet | LAE Adapter-10 · 5-Split-DomainNet (domain-incremental) | forgetting: 10.38 -> 6.84 | Improved |
| add PEGP to CLIP prompt on cross-modality BITM | CoOP Prompt · 5-Split-BITM | average accuracy: 52.73 -> 62.66 | Improved |
| add PEGP to CLIP before continual training on CIFAR100, test zero-shot ImageNet100 | CLIP baseline · zero-shot CIFAR100->ImageNet100 | zero-shot accuracy: 70.56 -> 72.84 | Improved |
| add PEGP, zero-shot on CIFAR100 after ImageNet100 training | CLIP baseline · zero-shot ImageNet100->CIFAR100 | zero-shot accuracy: 52.85 -> 59.91 | Improved |
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.
- AttriCLIP — AttriCLIP: A Non-Incremental Learner for Incremental Knowledge Learning (CVPR 2023) · 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
- SAFE: Slow and Fast Parameter-Efficient Tuning for Continual Learning with Pre-Trained Models (NeurIPS 2024 · similarity 0.89)
- Estimating and Orthogonalizing Unknown Pre-training Gradients for Continual Fine-tuning of Large Language Models (NeurIPS 2026 · similarity 0.88)
- Visual Prompt Tuning in Null Space for Continual Learning (NeurIPS 2024 · similarity 0.88)
- BiLoRA: Almost-Orthogonal Parameter Spaces for Continual Learning (CVPR 2025 · similarity 0.87)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.87)
- Faster Parameter-Efficient Tuning with Token Redundancy Reduction (CVPR 2025 · similarity 0.87)
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.
