LLMs Struggle with Ambiguous User Tasks, Lagging Human Alignment

Large language models (LLMs) are highly capable when tasks are clearly defined, but their performance significantly drops when faced with ambiguous or underspecified user requests. Researchers have introduced a framework that models this "task alignment" problem as a Partially Observable Markov Decision Process (POMDP), where the AI must infer the user's true intent from partial and evolving interactions. Experiments across various domains like shopping and coding show that while fine-tuning and reinforcement learning can improve alignment, current LLMs still fall short of human capabilities. Models only correctly identify user intent in 22-32% of ambiguous cases, whereas humans achieve 48%. This highlights a critical gap in LLMs' interactive abilities, suggesting they lack the nuanced understanding and adaptive interaction needed for reliable agency in real-world, open-ended scenarios.