Yuyu Guo

LoRA's catastrophic forgetting problem can be solved by explicitly decoupling its factor updates and pushing new feature representations away from old ones, leading to state-of-the-art continual learning performance.