People reason through new games with surprisingly shallow search

A test built around first impressions

When people face a new game, they do not usually search every possible future. They try a few things in their head, notice what seems important and make a move.

That sounds obvious. Turning it into a model that works across many different games is less obvious.

A study published in Nature this week tested more than 1,000 people on 121 two-player grid games. Almost all of the games were new to the participants. The researchers wanted to know how people could make sensible first decisions without long practice or exhaustive search.

Their answer is a model they call the Intuitive Gamer. It uses a small number of fast, goal-directed simulations. The simulations are shallow — usually looking only one move ahead at each step — but they are not random. The model focuses on moves that appear to matter for winning.

Across the study, that relatively small amount of imagined play came surprisingly close to human choices.

The researchers did not ask people to master one game. They gave them many unfamiliar ones.

Participants completed three kinds of task. Some looked at a game before playing and judged things such as whether it seemed fair or enjoyable. Others selected moves during their first play. A third group predicted what another person would do.

This matters because most game-playing systems are designed for a different target. They become strong by searching deeply, playing many times or training on large numbers of examples. The study was interested in the first encounter instead.

Humans are often decent at that. Not perfect, just useful. We bring a general sense of goals, obstacles, threats and promising moves from one situation to another.

Fast and flat

The Intuitive Gamer model tries to capture that behaviour with short probabilistic rollouts. In plain English, it imagines a few plausible sequences, rather than mapping the whole tree.

The best fit to the participants used roughly five to seven simulations. The implementation reported in the paper used six. At each point, the look-ahead was no more than one move deep.

In one zero-shot task involving 238 participants, the model’s predictions of how people would judge a game came close to the estimated ceiling set by the consistency of human responses. It reached an R-squared value of 0.81, while the paper estimated explainable variance at 0.82.

That number should not be read as a general measure of intelligence. It describes a fit on a specific task and dataset. But it is a strong result inside that setting.

Matching people is not the same as playing perfectly

Deeper systems were better at moving towards game-theoretic optimal play. That was expected. The Intuitive Gamer was built to explain beginners, not beat them.

Its attraction is efficiency. In the researchers’ self-play comparisons, it was about 700 times faster than their Expert Gamer model and used roughly 500 times fewer board evaluations. Against Monte Carlo tree search, the difference was larger still: almost 40,000 times faster, with around 10,000 times fewer node evaluations.

Those figures depend on the particular implementations and experimental setup. They are not a universal benchmark. Still, they show the trade-off clearly. A small, well-directed search can produce behaviour that looks flexible even when it is not optimal.

Why AI researchers may care

Much of modern AI improves with more computation. That can work very well, but it is expensive. It can also be awkward in a new environment where there is little time to search or learn.

The paper suggests another direction: systems that use simple, reusable intuitions to decide which possibilities are worth considering. The interesting part is not merely doing less work. It is choosing a small amount of work that carries useful information.

That could matter in multi-agent settings, where another person or system changes what happens next. A model that can form a quick, imperfect view of an unfamiliar situation may sometimes be more practical than one that needs a long search.

This is still an interpretation, not a demonstrated general AI method. The paper models human behaviour in games. It does not show a production system outperforming large AI models on open-ended tasks.

The limits are important

All 121 games were two-player, competitive and based on perfect information. Players could see the board. There were no hidden cards, incomplete observations or shifting real-world goals.

The games were novel, but they belonged to a familiar family. People already understand boards, turns and winning. Reasoning in a negotiation, a laboratory or a crowded street is a much larger jump.

The model also focuses on first-time reasoning. It does not explain in detail how people learn better heuristics after repeated play.

Those limits do not make the result unimportant. They make it specific. The study offers a careful account of how a little structured imagination can go a long way. Whether that lesson travels beyond the board is now the more interesting question.