LLMs Get Lost in Evolving User Intent

Academics including Alan Ritter and Mohit Iyyer retweeted an announcement from Jihoon Tack at Microsoft Research about a new paper. It examines how large language models now operate as collaborative agents on user-delegated tasks through iterative interaction. The work shows these models lose track of evolving user intent during genuine dynamic exchanges. The paper highlights challenges in the agentic era of LLMs and their limits in sustained collaboration.