How Games Are Teaching AI to Think Like Humans

Researchers found a surprising way to train AI: by making it play Battleship. While today’s AI excels at answering questions, it struggles with asking the right ones—a critical skill for solving complex problems. Scientists at MIT and Harvard tested this by creating a version of Battleship where AI agents had to guess hidden ship positions by questioning each other in real time. The experiment showed that big AI models like GPT-5 could perform decently but often lacked logical strategy. Meanwhile, smaller models like Llama 4 Scout made irrational moves unless given extra help. To fix this, researchers added a step where the AI calculates the likelihood of correct answers before responding. This simple tweak boosted Llama 4 Scout’s success rate from 8% to 82%, proving that even small AI can improve with the right tools.

The team also noticed that smaller AI models frequently gave wrong answers about ship locations. By forcing the AI to verify data through code, they reduced errors by 15%. This method didn’t just help Battleship—it could be used for real science, like finding rare drug molecules. When tested on another game, "Guess Who? ", the same techniques raised success rates dramatically. What’s interesting is that pure size doesn’t guarantee smarter questions. The researchers realized AI needs a "world model"—a way to predict outcomes—to ask better questions. This shift from answering to questioning could change how AI tackles problems in coding, math, and beyond. Instead of just providing answers, AI might soon explore possibilities like a human would.

Actions