The team manually designed 121 two-player competitive strategy games played on grids of varying sizes and rule structures. The games span classic “K-in-a-row” challenges across different board dimensions (including multiple variants of tic-tac-toe-like formats), plus more atypical twists such as misère objectives, asymmetric first-move dynamics, and rule sets that ignore certain line directions. To evaluate rules consistently across this diverse catalog, the authors built a flexible automated win-checking system.

They then ran multiple experiments with hundreds of participants. In one “zero-shot outcome” task, 238 people predicted win and draw likelihoods for randomly sampled novel games using continuous probability sliders. Participants were given only the linguistic game descriptions, with an optional interactive scratchpad to test ideas. Outlier responses were filtered using a distance-based criterion.

In a second experiment, 257 new participants instead rated how fun the same unfamiliar games were, again using a 0–100 confidence scale. A separate “human–human play” study recruited 302 participants to play a carefully selected subset of 40 games, after which they judged expected outcomes or funness. The authors also ran a “watch-and-predict” experiment with 314 participants who inferred where a player should move next after watching real games—linking indirect observation to next-move beliefs.

To formalize the cognitive mechanism, the researchers introduce an “Intuitive Gamer” model. It evaluates each legal move using local heuristics—rewarding immediate progress toward forming contiguous winning lines, penalizing comparable threats from the opponent, and applying a center-bias preference. Moves are chosen stochastically via a softmax (Boltzmann rationality), reflecting bounded rationality rather than exhaustive planning. For reasoning at the game level, the model samples multiple simulated play traces and converts simulated win/loss/draw outcomes into expected payoff.

Crucially, the authors show that a limited number of simulations best matches human variability: increasing the number of imagined games changes predicted dispersion, but humans’ behavior aligns with a mid-range simulation budget (reported around 5–7, using k=6 for main results). They compare this approach against alternative agents, including a deeper “Expert Gamer,” a “Random Gamer,” and an MCTS-based planning oracle, arguing that the intuitive model captures human-like reasoning efficiency.

Finally, the study extends beyond outcomes and funness to examine decision-making during draw offers, modeling whether players accept based on expected value adjusted for bonus incentives and the opportunity cost of continuing, estimated from simulation-derived expectations.

Subject of Research: Human reasoning about novel, rule-specified board games using limited simulation
Article Title: People use fast and flat simulation to reason about new games.
Article References: Collins, K.M., Zhang, C.E., Wong, L. et al. People use fast and flat simulation to reason about new games. Nature 655, 598–607 (2026). https://doi.org/10.1038/s41586-026-10722-1
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41586-026-10722-1
Keywords: board games, human cognition, bounded rationality, simulation, reinforcement learning, game theory, Monte Carlo tree search, funness, draw decisions