Appearance
People use fast and flat simulation to reason about new games
Summary
Problem: Prior research on game reasoning has focused on expert-level or superhuman play, typically involving deep search and extensive experience. However, everyday cognition often requires people to reason about entirely novel problems—evaluating whether a game is fair or fun, choosing first moves, and predicting others' actions—before any experience. The cognitive mechanisms underlying this "zero-shot" reasoning are poorly understood. Approach: The authors conducted large-scale behavioral studies with over 1,000 participants across 121 novel two-player grid-based board games. They assessed reasoning in three settings: game evaluation (fairness and funness before play), action selection (first-time play), and action prediction (judging others' moves). They developed the 'Intuitive Gamer' computational model, which uses fast (few simulations), flat (depth-limited, single-step lookahead), goal-directed (advancing own goals, blocking opponent), and probabilistic (softmax action sampling) simulation to explain human judgments. Finding: The Intuitive Gamer model captures human payoff judgments with R² = 0.81, matching the explainable variance from human data (split-half R² = 0.82). It significantly outperforms Random Gamer, Expert Gamer, MCTS, and non-probabilistic variants. A small number of simulations (k ≈ 5–7) best fits human variance. For funness, a regression combining model-derived features (balance, reward for thinking, game length) achieves R² = 0.57, approaching the explainable variance (R² = 0.60). The same model also captures action selection and prediction behaviors. Significance: This work demonstrates that novice reasoning about novel problems is systematic and adaptively rational, occupying an intermediate point between random exploration and deep expert search. It provides a resource-rational computational account that could inform the design of more flexible, human-like AI systems capable of evaluating whether a task is worth engaging with, not just how to solve it.
Theoretical Framework
- Theoretical tradition: Computational cognitive science, resource-rational analysis, probabilistic simulation models of cognition.
- Prior theories built upon: Game-theoretic optimal play (minimax), Monte Carlo tree search (MCTS), cognitive models of expert gameplay (e.g., depth-limited search with value functions, ref. 6), bounded rationality and compute-constrained decision-making (refs. 19, 20), and evidence that people use small numbers of mental simulations for belief formation and decisions (refs. 21–26).
- Causal mechanism: Game properties (e.g., payoff, funness) are inferred by running a small number of self-play simulations using a flat, goal-directed, probabilistic policy. The policy evaluates each possible action via single-step lookahead, scoring states by how much they advance the player's own goal and block the opponent's goal, then samples actions probabilistically. Simulated traces are aggregated to answer queries about game outcomes or subjective qualities.
- Key assumptions: (1) People understand game rules and pursue their goals while preventing opponents from doing the same. (2) Cognitive resources are limited, so reasoning uses few simulations and shallow search. (3) Simulated games are independent. (4) A policy can be constructed quickly from rule understanding without game-specific tuning.
- Theoretical move: This paper extends and applies existing computational models of gameplay (expert models, MCTS) by scaling them down to a compute-bounded, novice level. It challenges the assumption that sophisticated reasoning requires deep search, proposing instead that fast, flat, probabilistic simulation is sufficient and more human-like for novel problems. It also synthesizes resource-rational analysis with probabilistic simulation to explain both objective and subjective evaluations.
- Scope conditions: The model applies to two-player, grid-based, M–N–K-style strategy games with familiar core mechanics (placing pieces on a grid). Generalizability to other game types or real-world novel problems is suggested but not directly tested. The model assumes players understand rules and are motivated to pursue goals; unmotivated or random players are modeled separately.
Research Design
This is a behavioral experiment study with multiple sub-studies. The unit of analysis is the game (n = 121) for evaluation tasks, and individual moves or predictions for action tasks. Key independent variables are game characteristics (board size, win conditions, move dynamics). Dependent variables include: (1) expected payoff (likelihood of first-player win given no draw) and funness ratings (0–100 scale) in zero-shot evaluation tasks; (2) chosen moves in first-time play; (3) likelihood judgments of others' moves in action prediction. These are operationalized via participant ratings and choices, compared against model predictions from the Intuitive Gamer and alternative models (Random, Expert, MCTS, non-probabilistic variants, linguistic models).
Data & Sample
- Study 1 (payoff evaluation): 238 participants.
- Study 2 (funness evaluation): 246 participants.
- Additional studies for action selection and prediction (total over 1,000 participants across all studies).
- Population: Online participants (recruitment method not specified in the provided text; likely via platforms like Prolific or MTurk, but not stated).
- Region: Not specified; likely international online sample.
- Stimuli: 121 two-player grid-based board games, almost all novel to participants, derived from M–N–K games (e.g., tic-tac-toe, gomoku) with variations in board size, win conditions, and move dynamics.
- Time period: Data collected prior to publication (2026); exact dates not provided.
Analytical Strategy
- Model comparison: R² (coefficient of determination) between human mean judgments and model predictions across games, with 95% confidence intervals via bootstrapping.
- Split-half correlations to estimate total explainable variance in human data.
- Lesion analysis: Removing model components (flatness, probabilistic simulation, goal-directedness) to assess their contribution to fit.
- Sample complexity analysis: Varying number of simulations k to find the best fit to human variance (RMSE and Wasserstein distance).
- Regression models for funness: Combining model-derived features (balance, reward for thinking, quadratic game length) with R² and bootstrap CIs.
- Generalization tests: Cross-validation to ensure fits are not game-specific; comparison against alternative models (Random, Expert, linguistic features).
- Game-theoretic optimal analysis: Computed for 78 of 121 games to compare human judgments against objective optimal play.
- Control variables: Not explicitly mentioned; model parameters (search depth, value function sophistication) are varied systematically across model variants.
Results & Findings
- Payoff evaluation (fairness): The Intuitive Gamer model correlates strongly with human payoff judgments (R² = 0.81), matching the explainable variance from human data (R² = 0.82). It outperforms Random Gamer (R² = 0.47), Expert Gamer (R² = 0.65), MCTS (R² = 0.60), and non-probabilistic Intuitive Gamer (R² = 0.53). Lesioning any component (flatness, probabilistic, goal-directed) reduces fit, confirming each is necessary. A small number of simulations (k ≈ 5–7) best captures human variance, supporting the "fast" component. Human judgments track game-theoretic optimal payoffs (R² = 0.62) but are better explained by the Intuitive Gamer, suggesting people are reasonable but not optimal.
- Funness evaluation: Funness ratings varied widely across games. Games judged most fun tended to have larger boards (10×10) and moderate win conditions (4 or 5 in a row), but the relationship was nonlinear (e.g., the most fun game was a 5×5 board with a misère rule). Three model-derived features—balance, reward for thinking, and game length (inverted U-shape)—each captured significant variance individually, and combined in a regression achieved R² = 0.57, approaching the explainable variance (R² = 0.60). Generalization tests confirmed these fits were robust and superior to alternative models.
- Action selection and prediction: The same Intuitive Gamer player module (flat, goal-directed, probabilistic) generally captured how novices actually played and how they predicted others' moves, though specific quantitative details for these studies are not fully reported in the provided text.
- Null/unexpected findings: The most fun game was a misère game (3 in a row loses), which is counterintuitive and highlights the nonlinearity of funness. Funness judgments varied more across individuals than payoff judgments, reflecting its subjective nature. Human judgments were not perfectly aligned with game-theoretic optimality, indicating bounded rationality.
Limitations
- The paper focuses on a specific subclass of games (grid-based M–N–K variants), limiting generalizability to other game types or real-world novel problems.
- The model assumes independent simulations; the authors note this could be relaxed in future work.
- The number of simulations (k) is estimated empirically rather than derived from first principles.
- Funness is a subjective construct; the model captures much but not all explainable variance (R² = 0.57 vs. 0.60), leaving room for individual differences or unmodeled features.
- Action selection and prediction results are mentioned but not detailed in the provided text, so their full quantitative support is not assessable here.
- The recruitment method and participant demographics are not specified in the provided text.
Key Contributions
- Introduces the 'Intuitive Gamer' model, a novel computational account of novice reasoning based on fast, flat, goal-directed, probabilistic simulation.
- Demonstrates that people are systematic and adaptively rational in evaluating novel games before any play, with judgments closely matching model predictions.
- Shows that a small number of mental simulations (k ≈ 5–7) is sufficient for human-like reasoning, supporting resource-rational accounts of cognition.
- Provides a unified framework that explains both objective evaluations (payoff/fairness) and subjective evaluations (funness) from the same simulation mechanism.
- Identifies specific game features (balance, reward for thinking, game length) that drive subjective enjoyment, offering insights into game design.
- Offers a path to scale the model up to expert-level performance, suggesting a continuum from novice to expert reasoning.
- Informs AI design by showing how systems could evaluate whether a task is worth engaging with, not just how to solve it.
Key Claims
"We explain these capacities via a computational cognitive model that we call the 'Intuitive Gamer': a model based on mechanisms of fast and flat (depth-limited) goal-directed probabilistic simulation." (Abstract) "Our work offers insights into how people rapidly evaluate, act and make suggestions when encountering novel problems, and could inform the design of more flexible and human-like artificial intelligence systems that can determine not just how to solve new tasks but also whether a task is worth thinking about at all." (Abstract) "We hypothesize that novice thinking occupies a valuable intermediate point between these extremes: people are non-random in assessing new problems and run computations analogous to those that underlie state-of-the-art artificial intelligence (AI) game systems and cognitive models of expert human gameplay—but scaled down to a level that is more realistic for everyday thought." (Introduction) "The Intuitive Gamer model correlates well with human estimates (R² = 0.81 (95% confidence interval (CI), 0.77, 0.85)), matching the total explainable variance from the human data as estimated by split-half correlations (R² = 0.82 (95% CI, 0.77, 0.86))." (Results, "Is the game likely to be fair?") "A small number of simulations (k more than one, but less than ten) best captures the variance in human prediction." (Results, "Is the game likely to be fair?")