Firehose

This paper investigates how Large Language Models (LLMs) handle dynamic user intent in conversations, where users' goals change over time, and finds that current LLMs struggle with this capability, leading to significant performance drops when evaluated in a more realistic setting.