MIT researchers found that a game resembling Battleship can teach small AI models to ask better questions - and may reduce the case for defaulting to large, expensive ones.
The research used a grid-based guessing game as a benchmark for how AI agents handle incomplete information. The setup rewards models that ask targeted, high-value questions rather than guessing at random. Smaller models evaluated on this task showed measurable improvement in their information-seeking behavior. The finding suggests that part of what makes large models useful isn't raw scale - it's that they tend to reason better about what they don't yet know.
For anyone paying inference costs at scale, that distinction matters. The AI industry's default move has been to throw more parameters at hard problems, but if smaller models can be coached into sharper questioning strategies, cheaper agents become genuinely more viable. That's a meaningful shift for use cases where a frontier model is overkill but an underpowered small one fails too often.
A Battleship benchmark is a clean, bounded environment - a long way from the open-ended messiness of real deployment. Whether this questioning behavior transfers outside the grid is the question the research doesn't yet answer.