MIT Makes Small AI 82% Smarter at 1% of the Cost

Scientists just figured out how to make small AI models punch way above their weight class, and the secret weapon was a childhood board game.

Researchers at MIT and Harvard turned to Battleship, the classic guessing game, to solve one of artificial intelligence's biggest problems. While today's large AI models excel at answering questions, they struggle to ask good ones, especially in high-stakes fields like medical diagnosis and scientific research.

The team created "Collaborative Battleship," where players ask and answer questions in natural language to find hidden ships. They first watched over 40 humans play together, then tested both cutting-edge models like GPT-5 and smaller systems like Llama 4 Scout.

The results revealed a costly problem. Large models could beat humans at the game, but smaller, more affordable systems performed terribly, winning only 8 percent of the time.

Then came the breakthrough. The researchers taught the AI models to think more carefully about each question using Monte Carlo inference, a strategy that weighs different possibilities before choosing what to ask next. Think of it like inflating or deflating game balls based on how likely each guess is to be correct.

The transformation was stunning. That same small Llama model jumped from an 8 percent win rate to 82 percent against humans. Even better, it outperformed GPT-5 while operating at roughly 1 percent of the cost.

The team also solved another puzzle: getting AI to answer questions more accurately. They converted each question into Python code that gave models explicit instructions on how to verify their answers. One lightweight system saw its accuracy jump nearly 30 percent.

The Ripple Effect

This isn't just about winning board games. The ability to ask smart questions could revolutionize fields where AI agents need to explore uncertain environments and make discoveries.

Medical diagnosis requires doctors to narrow down thousands of possible conditions through strategic questioning. Scientific research demands careful hypothesis testing. Software debugging needs systematic problem-solving. All of these could become dramatically more accessible with smaller, cheaper AI that thinks strategically.

The cost difference matters enormously for real-world applications. Running AI at 1 percent of the typical expense means small clinics, independent researchers, and developing countries could access tools previously reserved for tech giants with massive budgets.

The researchers are already planning to scale this approach beyond games to actual coding and mathematical problem-solving. By teaching AI to explore and gather information more efficiently, they're opening doors that were previously locked behind expensive computational power.

A childhood game just showed us how to make breakthrough AI technology available to everyone who needs it.