Researchers from Hugging Face and collaborators examine how large language models (LLMs) handle evolving user intent in multi-turn interactions, finding that current evaluation and training methods—largely based on single-turn, fully-specified tasks—may not adequately capture real-world usage where user intent is gradually disclosed and revised. The study highlights a gap between static training paradigms and the dynamic nature of genuine conversational collaboration, raising questions about LLM performance in tracking and responding to shifting user goals over time.