Sequential Task Tuning: Overcoming Catastrophic Forgetting with Memory-Augmented Language Models

Abstract

Sequential task tuning in language models often faces the challenge of catastrophic forgetting, where performance on previously learned tasks deteriorates as new tasks are introduced. This paper explores the use of memory-augmented language models to mitigate this issue. By integrating external memory components that store and retrieve knowledge from previous tasks, we propose a framework that enhances the model’s ability to retain and recall information across a sequence of tasks. Experimental results demonstrate significant improvements in task retention and overall model performance, highlighting the potential of memory-augmented architectures in overcoming catastrophic forgetting.