CRAFT: Forgetting-Aware Intervention-Based Adaptation for Continual Learning

Researchers have introduced CRAFT, a novel continual learning framework designed to address catastrophic forgetting in Large Language Models (LLMs). Unlike traditional fine-tuning methods that update model weights, CRAFT learns low-rank interventions on hidden representations. The framework operates in three stages: routing tasks to similar groups based on output-distribution divergence, fine-tuning using Kullback-Leibler (KL) divergence against the group's prior state to control forgetting, and merging interventions into shared representations via the same KL signal. This approach unifies routing, regularization, and merging under a single objective. Experimental results indicate that CRAFT outperforms strong LoRA-based approaches across various benchmarks and model scales, significantly reducing forgetting while maintaining robustness to task ordering. The study suggests that controlling adaptation in representation space, guided by output-space divergence, offers a scalable and principled solution for continual learning in LLMs. This academic submission highlights advancements in machine learning techniques aimed at improving the efficiency and stability of AI model updates without compromising previously acquired capabilities.