LLM Task Interference: an Initial Study on the Impact of Task-Switch in Conversational History

With the recent emergence of powerful instruction-tuned large language models(LLMs), various helpful conversational Artificial Intelligence (AI) systemshave been deployed across many applications. When prompted by users, these AIsystems successfully perform a wide range of tasks as part of a conversation.To provide some sort of memory and context, such approaches typically conditiontheir output on the entire conversational history. Although this sensitivity tothe conversational history can often lead to improved performance on subsequenttasks, we find that performance can in fact also be negatively impacted, ifthere is a task-switch. To the best of our knowledge, our work makes the firstattempt to formalize the study of such vulnerabilities and interference oftasks in conversational LLMs caused by task-switches in the conversationalhistory. Our experiments across 5 datasets with 15 task switches using popularLLMs reveal that many of the task-switches can lead to significant performancedegradation.