Tag
This paper introduces a framework for interactive task alignment under ambiguity, formalized as a POMDP, and shows that current LLMs recover intended tasks only 22–32% of the time, lagging behind human performance.