MIT researchers used a Battleship-style test to show how smaller AI models can improve by asking sharper questions, potentially making cheaper AI agents more useful without relying on bigger systems.