Research map / Continual learning / Continual fine-tuning of LLMs
Compensating Distribution Drifts in Continual Learning with Pre-trained Vision Transformers
Xuan Rao, Simian Xu, Zheng Li, Bo Zhao, Derong Liu, Mingming Ha, Cesare Alippi
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 SLDC (weak-nonlinear α2) drift compensation to sequential fine-tuning (ablation) | SeqFT · CUB-200 (MoCo-V3) | Last-Acc: 64.40 -> 78.98 | Improved |
| use weak-nonlinear operator instead of linear operator (ablation) | SLDC · Cars-196 (MoCo-V3) | Last-Acc: 61.01 -> 77.53 | Improved |
| use weak-nonlinear SLDC instead of fully nonlinear MLP compensation (ablation) | SLDC · Cars-196 (Sup-21K) | Last-Acc: +6.52% | Improved |
| add feature distillation to sequential fine-tuning (ablation) | SeqFT · CUB-200 (Sup-21K) | Last-Acc: 76.57 -> 86.75 | Improved |
| combine distillation with linear SLDC (ablation) | α1-SLDC · CUB-200 (MoCo-V3) | Last-Acc: 70.42 -> 80.55 | Improved |
| linear α1-SLDC without ADE or distillation on Sup-21K (ablation) | SeqFT · CUB-200 and Cars-196 (Sup-21K) | Last-Acc: 76.57 -> 71.49; 54.53 -> 46.78 | Hurt |
| add auxiliary data enrichment (ADE) to linear SLDC (ablation) | α1-SLDC · Cars-196 (Sup-21K) | Last-Acc: 46.78 -> 73.01 | Improved |
| add MLP-based compensation (MLPDC) to distilled sequential fine-tuning (ablation) | SeqKD · CUB-200 (Sup-21K) | Last-Acc: 86.75 -> 75.76 | Hurt |
| vary regularization coefficient γα2 over 0.1 to 2.0 (ablation) | α2-SLDC · MoCo-V3 with distillation | accuracy | No clear effect |
| add ADE samples (larger sample size) (ablation) | SLDC · CUB-200 vs ImageNet-R | accuracy | Mixed |
| lower temperature αtemp below 1.0 when ADE is used (ablation) | α1-SLDC · Cars-196 and CUB-200 (MoCo-V3, distillation) | accuracy | Mixed |
| extend to 20-task sequences (ablation) | SLDC variants · MoCo-V3 20 tasks | accuracy | Improved |
Closest papers by meaning
- SLCA++: Unleash the Power of Sequential Fine-Tuning for Continual Learning with Pre-training (IJCV 2026 · similarity 0.89)
- Exemplar-Free Continual Learning of Vision Transformers via Gated Class-Attention and Cascaded Feature Drift Compensation (IJCV 2025 · similarity 0.88)
- Navigating Semantic Drift in Task-Agnostic Class-Incremental Learning (ICML 2025 · similarity 0.86)
- Mixture of Noise for Pre-Trained Model-Based Class-Incremental Learning (NeurIPS 2025 · similarity 0.86)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.85)
- ExPLoRA: Parameter-Efficient Extended Pre-Training to Adapt Vision Transformers under Domain Shifts (ICML 2025 · similarity 0.85)
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.
