ETD

• Catastrophic Forgetting
• Class-Incremental Learning.
• Continual Learning
• LoRA
• Parameter-Efficient Fine-Tuning
• Primal–Dual Control

Continual learning over long horizons is challenging because small per-task interference can compound into catastrophic forgetting. We propose Dual-Controlled LoRA (DuCo-LoRA), a parameter-efficient continual learning method that treats stability as a stream-level budget and enforces it online via primal–dual control of a drift surrogate. DuCo-LoRA continually updates a single shared LoRA adapter while keeping the pretrained backbone frozen, and carries a persistent dual variable across tasks to balance stability and plasticity without replay or parameter growth. With a task-agnostic whitened nearest-class-mean (W-NCM) classifier, DuCo-LoRA achieves 71.19% final average accuracy on 100-task ImageNet-R without replay (88% of joint-training accuracy), outperforming strong prompt-based and continual-adaptation baselines.