Teaching AI agents to ask better questions by playing "Battleship"

MIT and Harvard researchers improved AI agents' questioning in a modified Battleship game by training them on human interactions. Their Collaborative Battleship game uses natural language questions and real-time answers to evaluate how models weigh options. Smaller models like Llama 4 Scout outperformed leading LMs such as GPT-5 in fewer turns. The BattleshipQA dataset was built from over 40 human sessions playing the game together.