Research map / Continual learning / Continual fine-tuning of LLMs
Beyond Prompt Learning: Continual Adapter for Efficient Rehearsal-Free Continual Learning
Xinyuan Gao, Songlin Dong, Yuhang He, Qiang Wang, Yihong Gong
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 the scale-and-shift (S&S) module before incremental training (ablation) | C-ADA · ImageNet-R (P=10) | A_N: 0.73% drop when removed | Improved |
| add the orthogonal loss between new and old CAL weights (ablation) | C-ADA · ImageNet-R (P=10) | A_N: 3.76% drop when removed | Improved |
| add the continual adapter layer (CAL) instead of tuning only the classifier (ablation) | C-ADA · ImageNet-R (P=10) | A_N: 7.72% drop when removed | Improved |
| replace key-query-matched prompts with an adapter without matching (single forward pass) | C-ADA vs CODA-P-S · ImageNet-R (P=10) | A_N: 73.93 -> 76.66 | Improved |
| attach CAL to shallow layers 0-5 instead of deep layers 6-11 (ablation) | C-ADA · ImageNet-R (P=10) | A_N: 72.51 -> 76.21 | Improved |
| attach CAL to layers 0-4 instead of all 12 layers (ablation) | C-ADA · ImageNet-R (P=10) | A_N: 75.23 -> 76.66 | Improved |
| attach CAL to projection layer instead of QKV attention layer (ablation) | C-ADA · ImageNet-R (P=10) | A_N: 76.04 -> 76.25 | Improved |
| attach CAL to both Proj and MLP instead of a single position (ablation) | C-ADA · ImageNet-R (P=10) | A_N: 76.25 -> 76.66 | Improved |
| vary the split of middle dimensions between Proj and MLP CALs (ablation) | C-ADA · ImageNet-R (P=10) | A_N: 76.51 to 76.66 | No clear effect |
| increase trainable parameters (middle dimension of CAL) (ablation) | C-ADA · ImageNet-R (P=20) | A_N: more parameters give higher accuracy; 0.35% params gives 71.92 | Improved |
| use C-ADA instead of S-Prompts on domain-incremental benchmarks | C-ADA · CORe50 | A_N: 83.13 -> 89.45 | Improved |
| use C-ADA instead of CODA for forgetting on ImageNet-R P=20 | C-ADA · ImageNet-R (P=20) | F_N: 1.00 (CODA) vs 1.02 (C-ADA) | No clear effect |
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.
- DualPrompt — DualPrompt: Complementary Prompting for Rehearsal-Free Continual Learning (ECCV 2022) · 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
- Convolutional Prompting meets Language Models for Continual Learning (CVPR 2024 · similarity 0.90)
- CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning (CVPR 2025 · similarity 0.89)
- Consistent Prompting for Rehearsal-Free Continual Learning (CVPR 2024 · similarity 0.89)
- Self-Expansion of Pre-trained Models with Mixture of Adapters for Continual Learning (CVPR 2025 · similarity 0.89)
- REP: Resource-Efficient Prompting for Rehearsal-Free Continual Learning (NeurIPS 2025 · similarity 0.89)
- Adaptive Prompting for Continual Relation Extraction: A Within-Task Variance Perspective (AAAI 2025 · 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.
