AIM introduces a new scheme to improve merging performance in LLMs by combining continual learning principles and activation space-based model compression.