The authors explore the tradeoff between model efficiency and recall ability in linear attention language models, proposing a novel architecture to balance these factors effectively.
