CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning
CRAFT is a continual learning framework that avoids updating model weights by learning low-rank interventions on hidden representations. It routes new tasks to existing similar task groups based on output distribution divergence, uses a Kullback-Leibler penalty to prevent forgetting, and merges interventions for updated tasks. This approach improves performance and reduces forgetting compared to standard LoRA methods.