Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

People use fast and flat simulation to reason about new games

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read People can judge, play, and predict novel two-player games with zero experience, and this paper argues the underlying mechanism is fast, flat (one-step), goal-directed probabilistic simulation — implemented as the Intuitive Gamer model, whi

desk verdict A large, well-executed package that mostly earns its central claim about fast, flat mental simulation; the zero-shot evaluation results carry a scratchpad confound that needs a quantitative control. read the letter →

arxiv 2510.11503 v2 pith:YAJH2DKT submitted 2025-10-13 q-bio.NC cs.AIcs.GT

Katherine M. Collins , Cedegao E. Zhang , Lionel Wong , Mauricio Barba da Costa , Graham Todd , Adrian Weller , Samuel J. Cheyette , Thomas L. Griffiths
show 1 more author
Joshua B. Tenenbaum
This is my paper
classification q-bio.NCcs.AIcs.GT
keywords novelgameszero-shotreasoningmentalsimulationboundedrationalitygameevaluationprobabilisticmodelsnovicegameplayintuitivetheoryof
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to establish that novice reasoning about brand-new problems is neither random nor deep: people evaluate, play, and predict novel two-player grid games by running a few stochastic mental simulations of self-play, each using only a single step of look-ahead guided by simple goal-directed heuristics. In support, it reports that an 'Intuitive Gamer' model quantitatively matches human fairness judgments (R²=0.81, essentially at the split-half noise ceiling of 0.82), captures most explainable variance in funness ratings (R²=0.57 vs 0.60), best predicts the 9,892 moves made by first-time players across 1,808 matches (per-move log-likelihood advantage of 0.51 over a deep expert model), and tracks real match outcome distributions at R²=0.72. A sympathetic reader should care because this reframes everyday intelligence: a resource-limited, compute-cheap simulation process can explain how people decide whether a new situation is fair, fun, or worth engaging, and it suggests that flexible AI systems do not need enormous search to reason about novel settings.

What carries the argument

The central object is the two-module Intuitive Gamer model. The player module scores every legal action by immediate goal progress (Uself), immediate blocking of the opponent's goal (Uopp), and a center-bias auxiliary term (Uaux), then samples an action via a Boltzmann softmax—making it 'flat' (single-step look-ahead), goal-directed, and probabilistic. The reasoning module nests this player in k self-play simulations (empirically k=6), aggregates the simulated outcomes to answer queries like payoff, funness (via balance, reward-for-thinking, and game-length features), or next-move distributions. The model's work is to show that a compute-limited, depth-limited sampler—not deep tree search an

What would settle it

Rerun the zero-shot payoff and funness experiments without any scratchpad and with a forced response delay that forbids board interaction. If human judgments then depart substantially from Intuitive Gamer predictions (R² falling well below the split-half noise ceiling), the thesis that people evaluate never-played games through internal fast-flat simulation fails. A cheaper check: pre-register that the flat model assigns higher per-move likelihood than the deep expert model to first-time moves on a new set of unseen games; a reversal would falsify the player module.

Watch

Extended reading notes

Core claim

The paper's central claim is that people reason about novel games—before playing, during first-time play, and when predicting other novices—through what the authors call the Intuitive Gamer: a model that samples a small number (about k=6) of self-play game trajectories, with each action chosen probabilistically from a softmax over a shallow, one-step look-ahead value function combining three general-purpose heuristics (advancing one's own goal, blocking the opponent's progress, and proximity to the board center). The same model accounts for judgments of objective payoff, subjective funness, actual first moves, distributions of likely moves, and even whether to accept a draw. It does so signi

Load-bearing premise

For the zero-shot 'just think' payoff and funness studies, the claim that people evaluate a game before playing it via internal mental simulation presupposes that the optional interactive scratchpad did not carry the reasoning; the paper's own supplemental analysis reports participants used it only 1.60±1.46 rollouts per game on average, but if the scratchpad supplied the simulations, the model fits partly an external aid, weakening the 'before they have played it even once'

Editorial extensions

If this is right

  • Zero-shot evaluation: people can estimate a game's payoff and funness from a handful of imagined self-play traces, without ever having moved a piece; the model explains essentially all explainable variance in payoff judgments (R²=0.81 vs 0.82 split-half).
  • First-time move selection follows the same flat policy: the Intuitive Gamer predicts the 9,892 moves made by novices better than a deep expert model or random play, with a per-move log-likelihood advantage of 0.51.
  • The model also captures graded expectations: in watch-and-predict experiments, its move distributions approach the human split-half noise ceiling, whereas the deeper expert model tends to be overconfident or diffuse.
  • The same model's self-play outcome distributions track actual novice match outcomes (R²=0.72), and draw accept/reject decisions are explained by expected value of continuing plus expected funness.
  • Compute efficiency is a corollary: the flat model is roughly 700× faster than the deep expert model and about 40,000× faster than Monte Carlo Tree Search, yet provides better fits to human judgments.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If this generalizes beyond games, fast few-sample depth-limited simulation could explain how people reason about novel economic mechanisms, interfaces, or social rule systems; the paper itself raises this agenda but leaves the test open.
  • The result implies a benchmark shift for AI human-likeness: systems should be evaluated not only on expert play, but on zero-shot evaluation—'is this task worth solving?'—using the same flat simulation budget.
  • Because k≈6 samples are apparently sufficient, individual judgment noise is intrinsic and predictable; time-pressure manipulations or think-aloud protocols could directly test whether people vary the number of simulations or their stopping rules.
  • The endgame finding—where a depth-3 variant matches the flat model—suggests a testable extension: as the legal-move space shrinks, people may adaptively deepen search; manipulating board size and remaining moves should shift the optimal depth.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes the 'Intuitive Gamer' model, according to which people reason about novel two-player board games by running a small number of fast, flat, goal-directed probabilistic self-play simulations. The model has two components: a one-step-lookahead heuristic player with fixed hand-set weights and softmax temperature, and a reasoning module that aggregates k simulated games to answer queries such as expected payoff, funness, and move distributions. The authors test the model on more than 1,000 participants across 121 novel M-N-K-style games in four settings: zero-shot payoff and funness judgments, first-time human-human play, watching-and-predicting others' moves, and draw-accept/reject decisions. Headline results are a zero-shot payoff R² of 0.81 [0.77, 0.85] against a split-half noise ceiling of 0.82; a funness R² of 0.57 [0.51, 0.63] against 0.60 [0.51, 0.68]; a per-move log-likelihood advantage of 0.51 over the Expert Gamer in predicting 9,892 first-time moves; and watch-and-predict TVDs near the human split-half ceiling.

Significance. If the conclusions hold, this is a substantial contribution to the study of 'pre-expertise': it provides a rare computational account of how people form judgments and make decisions about entirely new rule systems without task-specific experience. The paper's strengths are considerable: a large and carefully constructed game stimulus set; a player model that is nearly parameter-free (fixed weights and temperature) and is tested on action-prediction data it was not tuned on; systematic comparison with random, deeper-search, MCTS, and language-model baselines; and unusually candid reporting of failure modes (misère games, over-sharp P(draw), late-game depth, post-watch noise). The action-selection and watch-and-predict results are largely independent of the scratchpad and provide strong convergent evidence for the player module. However, the headline zero-shot evaluation results are entangled with an interactive scratchpad that allows external self-play simulation, and the number of simulations k is selected from the same human payoff judgments used to report the payoff fit. These issues affect the strongest quantitative claims and require substantial revision.

major comments (3)
  1. [Methods, 'Zero-shot game outcome evaluation experiment'; SI §5.1] The zero-shot payoff and funness experiments provided an interactive scratchpad board that participants could use while 'just thinking.' SI §5.1 reports that participants used a mean of 1.60 ± 1.46 rollouts per game in the payoff task. The model's reasoning module itself works by running k self-play simulations and aggregating outcomes. The scratchpad is therefore not a neutral memory aid in this context: it is an external medium for exactly the kind of self-play simulation the model posits. If those external rollouts carried part of the simulation load, the model's R²=0.81 payoff fit and R²=0.57 funness fit partly reflect the structure of the task rather than purely internal mental simulation. The SI's statement that the authors 'do not draw a direct parallel between scratchpad rollouts and mental simulation' does not resolve this concern, because the threat is not that the count equals
  2. [Methods, 'Sensitivity analyses into the number of samples k'; Fig. 3c] The reasoning module's sample size is set to k=6 because 'a small number of simulations (k more than one, but less than 10) best captures the variance in human prediction' in the payoff experiment. The same human payoff judgments are then used to report the model's payoff R²=0.81. Thus k is not a fully a priori or held-out quantity for the zero-shot payoff evaluation; the 'fastness' component is in part identified from the very data used to validate the model. This does not invalidate the result, but it weakens the claim that the model is nearly parameter-free and that the fit is out-of-sample. Please report payoff R² as a function of k (beyond the variance-matching plots), or select k on a training subset of games/participants and test on the remainder, or otherwise show that the payoff R² is insensitive to the k-selection procedure.
  3. [SI §5.2, 'Decomposed game outcome prediction tasks'] The payoff composite R²=0.81 masks a systematic miss: the Intuitive Gamer's P(draw) predictions have R²=0.60 [0.56, 0.64] against a human split-half of R²=0.78 [0.72, 0.83], and the authors state that the model is 'substantially sharper in P(draw) predictions than people.' Because payoff is the difference P(P1 win) − P(P1 lose), a good payoff fit can be produced by compensating errors in the components. The paper should either report the decomposed fits prominently in the main text or temper the statement that the model 'matches the total explainable variance' from human payoff judgments. This is especially relevant given the scratchpad and k-selection issues above.
minor comments (5)
  1. [SI §5.1] Typo: 'we include the scratchpad was to the demands on spatial working memory' should read 'we included the scratchpad to reduce the demands on spatial working memory.'
  2. [Main text and SI figure captions] 'Total Variance Distance' should be 'Total Variation Distance' (e.g., Fig. 5c caption and SI Figures 47–48).
  3. [Methods, 'Game construction'] Typo: 'games wiht more atypical rules' should be 'games with more atypical rules.'
  4. [Main text, 'Is the game likely to be fair?'] 'k more than one, but less than 10' should be written as 1 < k < 10 for clarity and to avoid reading as an interval of integers.
  5. [SI §6.8.2] The post-watch funness split-half R² is 0.07 [0.0001, 0.27], which is effectively null. This is an important negative result and is only reported in the Supplementary Information. Please at least mention it in the main text's discussion of task differences or limitations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the Intuitive Gamer's payoff, play, and prediction outputs are computed from game rules against fresh human data, with no reduction of predictions to fitted quantities.

full rationale

The central derivation chain is self-contained. Payoffs are ψ(G) computed from k self-play simulations of the player module; human payoff/funness judgments are fresh data. Player-module parameters are fixed (w=1, τ=1) with sensitivity analyses reported; the only tuned quantity, k, is selected by matching model payoff variance to human per-game variance in the same 121-game data. This is a mild model-calibration step, but k affects sampling noise of model payoff estimates rather than their expected values, and the main per-move play, watch-and-predict, and held-out funness comparisons do not depend on it; hence it does not make the payoff or fastness claims reduce by construction. The SI's scratchpad caveat (SI §5.1: 'Participants made on average 1.60±1.46 SD rollouts for the payoff task... We do not draw a direct parallel between scratchpad rollouts and mental simulation') is a real confound for the zero-shot evaluation pillar because participants could offload self-play to the external board, but the model is not fitted to scratchpad usage, and the play and watch-and-predict experiments involve no scratchpad and independently support the player module. There is no load-bearing self-citation: the Expert Gamer's search-iteration parameter comes from van Opheusden et al. (external), and other self-citations are not used to justify the model's core mechanism. No uniqueness theorem from the authors is invoked to rule out alternatives. Accordingly, no circular step meets the evidentiary bar; the paper should be assessed on empirical/correctness grounds rather than circularity.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

The central claim rests on a small number of fixed-by-hand modeling choices (τ=1, equal weights, base-2 exponent) that are checked but not derived from human data (SI §3.3); on one post-hoc fitted quantity, the simulation budget k=6 (Figure 3c); and on regression weights for the funness model that are fit to the human funness judgments. The deepest assumptions are domain-level: that M-N-K grid games are the right probe, that contiguous-line connectivity is the natural goal-progress metric, and that softmax/Boltzmann choice with independent, run-to-termination simulations describes the mental process. The Expert Gamer comparison model carries an ad hoc 'inherited value function' and a search budget transferred from a single prior game. No new physical/causal entities are postulated; the 'Intuitive Gamer' is a model instance, and its only falsifiable handles are the human behavioral fits reported in this paper.

free parameters (6)
  • k — number of mental game simulations = 6 (range 5–7)
    Chosen post hoc by matching the variance of model payoff predictions to per-game variance in human judgments across all 121 games (Figure 3c, Methods 'Sensitivity analyses'); fixed at 6 for all headline results. Underwrites the 'fast' component of the claim.
  • softmax temperature τ = 1
    Fixed by hand; SI §3.3 sensitivity sweeps confirm robustness but are evaluated against the same human payoff data.
  • heuristic weights wconnect, wblock, wcenter = 1, 1, 1
    Chosen for simplicity; SI §3.3 and Figure 13 show fit is sensitive mainly to wconnect/wblock, so the equal-weight choice is tested but not derived.
  • exponentiation base for value features = 2
    Chosen after 'light initial exploration' prior to collecting human gameplay data (Methods 'Intuitive Gamer player module'); not derived.
  • funness regression coefficients (balance, reward-for-thinking, length, length²) = 8.7, 4.5, 9.9, −4.4 (SI Table 9)
    Fit to the 246-participant funness judgments; the funness 'prediction' is an in-sample regression, partially validated by 50/50 held-out-game splits.
  • Expert Gamer exploration rate ϵ = 0.0001
    Selected by maximizing the Expert's win rate against Intuitive Gamer variants (SI §3.2, Table 8) — a fitted parameter of the comparison model, not the central model.
assumptions (7)
  • domain assumption Two-player zero-sum M-N-K grid games are the right stimulus class, and contiguous-connectivity progress is the natural goal metric
    The model's value function only rewards extending/blocking contiguous lines plus near-center play (Methods, Eq. 3). The 121 games were constructed by the authors within this family, so the assumption is partly built into the stimulus set.
  • domain assumption Action selection follows a Boltzmann/softmax rule with fixed temperature
    Standard in the literature (Luce 1959, cited), adopted without model selection against other choice rules.
  • domain assumption People can weigh multiple goals (offense + defense) simultaneously
    Explicit in the text: 'We assume that players have therefore already developed the capacity to account for multiple goals simultaneously, unlike potentially even more naive child-like game reasoners.'
  • domain assumption Mental simulations are independent and (in the primary model) run to termination
    Explicit in the text: 'We assume simulated games are independent, though this assumption could be relaxed in the future'; early stopping is only an exploratory variant (SI §5.2).
  • domain assumption The ideal fun game is balanced, decisive, and rewards thinking
    The balance/reward-for-thinking/length features encode this normative assumption, citing game-generation literature (refs 2, 6, 13, 58); it is a modeling prior, not derived from the human data.
  • ad hoc to paper The Expert Gamer with ~636 transferred search iterations and an inherited value function fairly represents deep search
    The 636-iteration budget is transferred from van Opheusden's single-game (4-in-a-row) fit, and the inherited-value modification ('we introduce an inherited version of our heuristic', Methods) is introduced ad hoc to make deeper search work with the local heuristic.
  • domain assumption Game-theoretic optimality for 78 of 121 games is estimable via Uiterwijk results plus MCTS convergence to {−1,0,1}
    The 'optimal' comparison set is partially defined by MCTS self-convergence, which is not a guaranteed oracle; the paper discloses this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of People use fast and flat simulation to reason about new games." pith.science (2026). https://pith.science/paper/YAJH2DKT

@misc{pith2026251011503,
  author       = {Pith},
  title        = {Pith review of: People use fast and flat simulation to reason about new games},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YAJH2DKT}},
  note         = {Machine review of arXiv:2510.11503}
}
read the original abstract

Games have long been a microcosm for studying planning and reasoning in both natural and artificial intelligence (AI), often focusing on expert-level or even super-human play. But real life also pushes human intelligence along a different frontier, requiring people to flexibly navigate decision-making problems that they have never thought about before. Here, we use novice gameplay to study how people reason about new problem settings. Through a series of large-scale behavioral studies with over 1000 participants and 121 two-player strategic board games (almost all novel to our participants), we show that people are systematic and adaptively rational in how they play a game for the first time, or evaluate a game (e.g., how fair or how fun it is likely to be) before they have played it even once. 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. Our work offers new 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 AI systems that can determine not just how to solve new tasks, but whether a task is worth thinking about at all.

Figures

Figures reproduced from arXiv: 2510.11503 by the authors.

Figure 1
Figure 1. Our novel game dataset and suite of game tasks [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The Intuitive Gamer model, compared to prior models of game reasoning. a, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Evaluating games without ever playing. 238 participants judged the expected payoff of games drawn from our 121 game suite. a, The expected payoff computed under the Intuitive Gamer model captures human predictions well when compared to alternate models that scale up or down compute. Each point represents the payoff for one of the n = 121 game stimuli. Error bars depict 95% CIs around the mean estimated payoff from p… view at source ↗
Figures from the paper (39 more)
Figure 4
Figure 4. Figure 4: Evaluating whether games are likely to be fun, before ever playing them. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Modeling people’s actions and distribution over predicted actions in the first encounter [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Example human- and model-predicted distributions over the next action in real games. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 13
Figure 13. Figure 13: Parameter sensitivity analyses. Analyzing the impact of the goal progress (offensive “connect” component), goal blocking (defensive “block” component), and center weight on fit to people’s judgments in the “just think” experiment, with subset of simulations for differ…
Figure 14
Figure 14. Figure 14: “Just think” evaluating games before any experience example interfaces. [PITH_FULL_IMAGE:figures/full_fig_p045_14.png]
Figure 15
Figure 15. Figure 15: Example interfaces from the human-human gameplay study. a, [PITH_FULL_IMAGE:figures/full_fig_p046_15.png]
Figure 16
Figure 16. Figure 16: Additional example interfaces from the human-human gameplay study. a, [PITH_FULL_IMAGE:figures/full_fig_p046_16.png]
Figure 17
Figure 17. Figure 17: Example interfaces from the people watching others play and predicting moves study. [PITH_FULL_IMAGE:figures/full_fig_p047_17.png]
Figure 18
Figure 18. Figure 18: Scratchpad usage. Explicit scratchpad “rollouts” per game, per participant. A “rollout” is counted as any interaction with the scratchpad, involving at least one click. 2 rollouts means the participant pressed “RESET” once. Most participants engage at least once with …
Figure 19
Figure 19. Figure 19: Empirical model comparisons for P(P1 win) against people. Bootstrapped 95% CIs over participants and samplings of k = 6 simulations for N = 20 simulated participants for each model. Our early explorations into the impact over the game reasoning module running only par…
Figure 20
Figure 20. Figure 20: Empirical model comparisons for P(draw) against people. Bootstrapped 95% CIs over participants and samplings of k = 6 simulations for N = 20 simulated participants for each model. 0.0 0.2 0.4 0.6 0.8 1.0 Model 0.0 0.2 0.4 0.6 0.8 1.0 Human Intuitive Gamer R 2=0.72 (0.…
Figure 21
Figure 21. Figure 21: Empirical model comparisons for P(P1 win|no draw) against people. Games under the model that only have draws are imputed with 0.5. Bootstrapped 95% CIs over participants and samplings of k = 6 simulations for N = 20 simulated participants for each model. 5.4 Language …
Figure 22
Figure 22. Figure 22: Full versus partial simulations to estimate [PITH_FULL_IMAGE:figures/full_fig_p050_22.png]
Figure 23
Figure 23. Figure 23: Predicting payoff from non-simulation game traits. [PITH_FULL_IMAGE:figures/full_fig_p051_23.png]
Figure 24
Figure 24. Figure 24: Comparing payoff predictions against language models. [PITH_FULL_IMAGE:figures/full_fig_p051_24.png]
Figure 25
Figure 25. Figure 25: Language- and reasoning-model predicted funness per game, compared to people. [PITH_FULL_IMAGE:figures/full_fig_p053_25.png]
Figure 26
Figure 26. Figure 26: Predicting fun from non-simulation game traits. a, [PITH_FULL_IMAGE:figures/full_fig_p053_26.png]
Figure 27
Figure 27. Figure 27: Aggregate move likelihood, per game, in the human-human play experiment. [PITH_FULL_IMAGE:figures/full_fig_p054_27.png]
Figure 28
Figure 28. Figure 28: Accuracy and rank of human played move in human gameplay experiments. [PITH_FULL_IMAGE:figures/full_fig_p054_28.png]
Figure 29
Figure 29. Figure 29: “Watch-and-predict” results are robust to the choice of distributional metric. [PITH_FULL_IMAGE:figures/full_fig_p055_29.png]
Figure 30
Figure 30. Figure 30: Average admxiture weights for each game, fit to the average human watch-and-predict [PITH_FULL_IMAGE:figures/full_fig_p055_30.png]
Figure 31
Figure 31. Figure 31: Mixture weight and temperature sensitivity in models’ fit to people’s action selection and [PITH_FULL_IMAGE:figures/full_fig_p057_31.png]
Figure 32
Figure 32. Figure 32: People’s distribution of predicted moves capture real human player’s actions comparably [PITH_FULL_IMAGE:figures/full_fig_p058_32.png]
Figure 33
Figure 33. Figure 33: Full set of “watch-and-predict” distributions per match. [PITH_FULL_IMAGE:figures/full_fig_p059_33.png]
Figure 34
Figure 34. Figure 34: Full set of “watch-and-predict” distributions per match (continued). 60 [PITH_FULL_IMAGE:figures/full_fig_p060_34.png]
Figure 35
Figure 35. Figure 35: Full set of “watch-and-predict” distributions per match (continued). 61 [PITH_FULL_IMAGE:figures/full_fig_p061_35.png]
Figure 36
Figure 36. Figure 36: Full set of “watch-and-predict” distributions per match (continued). 62 [PITH_FULL_IMAGE:figures/full_fig_p062_36.png]
Figure 37
Figure 37. Figure 37: Full set of “watch-and-predict” distributions per match (continued). 63 [PITH_FULL_IMAGE:figures/full_fig_p063_37.png]
Figure 38
Figure 38. Figure 38: Full set of “watch-and-predict” distributions per match (continued). 64 [PITH_FULL_IMAGE:figures/full_fig_p064_38.png]
Figure 39
Figure 39. Figure 39: Full set of “watch-and-predict” distributions per match (continued). 0 20 40 60 Predicted Length 0 10 20 30 40 50 60 Empirical Length R 2 = 0.71 Intuitive Gamer 0 20 40 60 Predicted Length R 2 = 0.64 Expert Gamer 0 20 40 60 Predicted Length R 2 = 0.57 Random [PITH_FU…
Figure 40
Figure 40. Figure 40: Empirical human- vs. model-predicted expected game length [PITH_FULL_IMAGE:figures/full_fig_p065_40.png]
Figure 41
Figure 41. Figure 41: Draw and surrender rates and people’s “just think” game evaluation. [PITH_FULL_IMAGE:figures/full_fig_p066_41.png]
Figure 42
Figure 42. Figure 42: Impact of a single instance of play experience on judgments [PITH_FULL_IMAGE:figures/full_fig_p066_42.png]
Figure 43
Figure 43. Figure 43: Post-play funness modeling. a-c, Features derived from the Intuitive Gamer model compared against post-play participant funness judgments. d, Regression model fit to the post-play data. 95% CIs are bootstrapped over participant ratings per games. e, Comparison of boot…
Figure 44
Figure 44. Figure 44: Relationship between attained outcome and funness judgments, after a single round of [PITH_FULL_IMAGE:figures/full_fig_p067_44.png]
Figure 45
Figure 45. Figure 45: Judgments after watching a match. Comparing post-watch judgments of predicted payoff against those made in the a, “just think” and b, “play” experiments, on only the subset of 21 games considered in the watch experiment, as well as the predicted funness for c, “just-t…
Figure 46
Figure 46. Figure 46: Intermediate depth payoff predictions. Non-flat (depth-3) variant of the Intuitive Gamer model’s predicted payoffs for the 121 games in the “just think” experiment, compared against people’s predicted payoffs. Error bars depict 95% CIs over bootstrapped human means an…
Figure 47
Figure 47. Figure 47: “Human-human play” analyses broken down by game stage. [PITH_FULL_IMAGE:figures/full_fig_p070_47.png]
Figure 48
Figure 48. Figure 48: “Watch-and-predict” analyses broken down by game stage. [PITH_FULL_IMAGE:figures/full_fig_p070_48.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. A Matter of Interest: Understanding Interestingness of Math Problems in Humans and Language Models

    cs.AI 2025-11 conditional novelty 6.0 of 10

    LLMs track average human interestingness ratings for contest math problems but mostly fail to reproduce the diversity of human ratings or the reasons behind them.

Reference graph

Works this paper leans on

69 extracted references · 3 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Almaatouq, J

    A. Almaatouq, J. Becker, J. P . Houghton, N. Paton, D. J. Watts, and M. E. Whiting. Empirica: a virtual lab for high-throughput macro-level experiments.Behavior Research Methods, 53 (5):2158–2171, 2021

    2021

  2. [2]

    Althofer

    I. Althofer. Computer-aided game inventing.Friedrich Schiller University, Jena, Germany, Tech. Rep, 2003

    2003

  3. [3]

    O. Amir, L. Tyomkin, and Y. Hart. Adaptive search space pruning in complex strategic problems.PLoS Computational Biology, 18(8):e1010358, 2022. 15

    2022

  4. [4]

    C. L. Baker, J. Jara-Ettinger, R. Saxe, and J. B. Tenenbaum. Rational quantitative attribution of beliefs, desires and percepts in human mentalizing.Nature Human Behaviour, 1(4):0064, 2017

    2017

  5. [5]

    P . W. Battaglia, J. B. Hamrick, and J. B. Tenenbaum. Simulation as an engine of physical scene understanding.Proceedings of the National Academy of Sciences, 110(45):18327–18332, 2013

    2013

  6. [6]

    C. B. Browne.Automatic generation and evaluation of recombination games. PhD thesis, Queensland University of Technology, 2008

    2008

  7. [7]

    C. B. Browne, E. Powley, D. Whitehouse, S. M. Lucas, P . I. Cowling, P . Rohlfshagen, S. Tavener, D. Perez, S. Samothrakis, and S. Colton. A survey of monte carlo tree search methods.IEEE Transactions on Computational Intelligence and AI in Games, 4(1):1–43, 2012

    2012

  8. [8]

    Campbell, A

    M. Campbell, A. J. Hoane Jr, and F.-h. Hsu. Deep Blue.Artificial Intelligence, 134(1-2):57–83, 2002

    2002

Show all 69 references
  1. [9]

    Cao and Y

    X. Cao and Y. Lin. Uct-adp progressive bias algorithm for solving gomoku. In2019 IEEE Symposium Series on Computational Intelligence (SSCI), pages 50–56. IEEE, 2019

    2019

  2. [10]

    Carroll, R

    M. Carroll, R. Shah, M. K. Ho, T. Griffiths, S. Seshia, P . Abbeel, and A. Dragan. On the utility of learning about humans for human-AI coordination.Advances in neural information processing systems, 32, 2019

    2019

  3. [11]

    Charness

    N. Charness. Expertise in chess: The balance between knowledge and search. In K. A. Ericsson and J. Smith, editors,Toward a general theory of expertise: Prospects and limits, pages 39–63. Cambridge University Press, Cambridge, 1991

    1991

  4. [12]

    W. G. Chase and H. A. Simon. The mind’s eye in chess. InVisual information processing, pages 215–281. Elsevier, 1973

    1973

  5. [13]

    J. Chu, J. B. Tenenbaum, and L. E. Schulz. In praise of folly: flexible goals and human cognition.Trends in Cognitive Sciences, 2023

    2023

  6. [14]

    A. A. Cleveland. The psychology of chess and of learning to play it.The American Journal of Psychology, 18(3):269–308, 1907

    1907

  7. [15]

    K. M. Collins, I. Sucholutsky, U. Bhatt, K. Chandra, L. Wong, M. Lee, C. E. Zhang, T. Zhi- Xuan, M. Ho, V . Mansinghka, et al. Building machines that learn and think with people. Nature Human Behavior, 2024

    2024

  8. [16]

    R. Coulom. Efficient selectivity and backup operators in monte-carlo tree search. In International conference on computers and games, pages 72–83. Springer, 2006

    2006

  9. [17]

    Crowley and R

    K. Crowley and R. S. Siegler. Flexible strategy use in young children’s tic-tac-toe.Cognitive Science, 17(4):531–561, 1993

    1993

  10. [18]

    Bakhtin, N

    FAIR, A. Bakhtin, N. Brown, E. Dinan, G. Farina, C. Flaherty, D. Fried, A. Goff, J. Gray, H. Hu, et al. Human-level play in the game of diplomacy by combining language models with strategic reasoning.Science, 378(6624):1067–1074, 2022

    2022

  11. [19]

    Franke and J

    M. Franke and J. Degen. The softmax function: Properties, motivation, and interpretation. OSF Preprints, 2023

    2023

  12. [20]

    Genesereth and M

    M. Genesereth and M. Thielscher.General Game Playing. Synthesis Lectures on Artificial Intelligence and Machine Learning. Morgan & Claypool Publishers, 2014. 16

    2014

  13. [21]

    Gobet, J

    F. Gobet, J. Retschitzki, and A. de Voogt.Moves in mind: The psychology of board games. Psychology Press, 2004

    2004

  14. [22]

    T. L. Griffiths, A. N. Sanborn, K. R. Canini, and D. J. Navarro. Categorization as nonparametric Bayesian density estimation.The probabilistic mind: Prospects for Bayesian cognitive science, pages 303–328, 2008

    2008

  15. [23]

    Hintikka

    J. Hintikka. On the logic of an interrogative model of scientific inquiry.Synthese, pages 69–83, 1981

    1981

  16. [24]

    M. K. Ho, D. Abel, C. G. Correa, M. L. Littman, J. D. Cohen, and T. L. Griffiths. People construct simplified mental representations to plan.Nature, 606(7912):129–136, 2022

    2022

  17. [25]

    L. Hurwicz. The design of mechanisms for resource allocation.The American Economic Review, 63(2):1–30, 1973

    1973

  18. [26]

    T. Icard. Subjective probability as sampling propensity.Review of Philosophy and Psychology, 7(4):863–903, 2016

    2016

  19. [27]

    J. Kiefer. Sequential minimax search for a maximum.Proceedings of the American mathematical society, 4(3):502–506, 1953

    1953

  20. [28]

    Kocsis and C

    L. Kocsis and C. Szepesvari. Bandit based Monte-Carlo planning. InEuropean Conference on Machine Learning, pages 282–293. Springer, 2006

    2006

  21. [29]

    Koster, J

    R. Koster, J. Balaguer, A. Tacchetti, A. Weinstein, T. Zhu, O. Hauser, D. Williams, L. Campbell-Gillingham, P . Thacker, M. Botvinick, et al. Human-centred mechanism design with democratic ai.Nature Human Behaviour, 6(10):1398–1407, 2022

    2022

  22. [30]

    Kuperwajs, M

    I. Kuperwajs, M. K. Ho, and W. J. Ma. Heuristics for meta-planning from a normative model of information search.Planning, 1:a2, 2024

    2024

  23. [31]

    Laudan.Progress and its problems: Towards a theory of scientific growth, volume 282

    L. Laudan.Progress and its problems: Towards a theory of scientific growth, volume 282. Univ of California Press, 1978

    1978

  24. [32]

    H. Li. gobang: JavaScript Gobang AI based on Alpha-Beta Pruning. https://github. com/lihongxun945/gobang, 2025

    2025

  25. [33]

    R. D. Luce.Individual choice behavior, volume 4. Wiley New York, 1959

    1959

  26. [34]

    E. S. Maskin. Mechanism design: How to implement social goals.American Economic Review, 98(3):567–576, 2008

    2008

  27. [35]

    McFadden

    D. McFadden. The human side of mechanism design: a tribute to Leo Hurwicz and Jean-Jacque Laffont.Review of Economic Design, 13(1):77–100, 2009

    2009

  28. [36]

    M. L. Menéndez, J. A. Pardo, L. Pardo, and M. d. C. Pardo. The Jensen-Shannon divergence. Journal of the Franklin Institute, 334(2):307–318, 1997

    1997

  29. [37]

    P . R. Milgrom.Putting auction theory to work. Cambridge University Press, 2004

    2004

  30. [38]

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, et al. Human-level control through deep reinforcement learning.nature, 518(7540):529–533, 2015

    2015

  31. [39]

    R. B. Myerson. Mechanism design by an informed principal.Econometrica: Journal of the Econometric Society, pages 1767–1797, 1983. 17

    1983

  32. [40]

    M. R. Nassar and M. J. Frank. Taming the beast: extracting generalizable knowledge from computational models of cognition.Current opinion in behavioral sciences, 11:49–54, 2016

    2016

  33. [41]

    Newell and H

    A. Newell and H. A. Simon.Human problem solving. Prentice-Hall, Englewood Cliffs, NJ, 1972

    1972

  34. [42]

    Newell, J

    A. Newell, J. C. Shaw, and H. A. Simon. Chess-playing programs and the problem of complexity.IBM Journal of Research and Development, 2(4):320–335, 1958

    1958

  35. [43]

    C. T. Nguyen. Games and the art of agency.Philosophical Review, 128(4):423–462, 2019

    2019

  36. [44]

    Jaech, A

    OpenAI Team, A. Jaech, A. Kalai, A. Lerer, A. Richardson, A. El-Kishky, A. Low, A. Helyar, A. Madry, A. Beutel, A. Carney, A. Iftimie, A. Karpenko, A. T. Passos, A. Neitz, A. Prokofiev, A. Wei, A. Tam, A. Bennett, A. Kumar, A. Saraiva, A. Vallone, A. Duberstein, A. Kondrich, A...

    2024 arXiv

  37. [45]

    Palan and C

    S. Palan and C. Schitter. Prolific.ac—A subject pool for online experiments.Journal of Behavioral and Experimental Finance, 17:22–27, 2018

    2018

  38. [46]

    Piette, D

    E. Piette, D. J. Soemers, M. Stephenson, C. F. Sironi, M. H. Winands, and C. Browne. Ludii–the ludemic general game system. InECAI 2020, pages 411–418. IOS Press, 2020

    2020

  39. [47]

    Rubner, C

    Y. Rubner, C. Tomasi, and L. J. Guibas. A metric for distributions with applications to image databases. InSixth international conference on computer vision (IEEE Cat. No. 98CH36271), pages 59–66. IEEE, 1998. 18

    1998

  40. [48]

    R. B. Rutledge, N. Skandali, P . Dayan, and R. J. Dolan. A computational and neural model of momentary subjective well-being.Proceedings of the National Academy of Sciences, 111(33): 12252–12257, 2014

    2014

  41. [49]

    A. N. Sanborn, T. L. Griffiths, and D. J. Navarro. Rational approximations to rational models: alternative algorithms for category learning.Psychological review, 117(4):1144, 2010

    2010

  42. [50]

    C. E. Shannon. Xxii. programming a computer for playing chess.The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science, 41(314):256–275, 1950

    1950

  43. [51]

    Sheoran, G

    K. Sheoran, G. Dhand, M. Dabas, N. Dahiya, and P . Pushparaj. Solving connect 4 using optimized minimax and monte carlo tree search.Advances and Applications in Mathematical Sciences, 21(6):3303–3313, 2022

    2022

  44. [52]

    Silver, A

    D. Silver, A. Huang, C. J. Maddison, A. Guez, L. Sifre, G. van den Driessche, J. Schrittwieser, I. Antonoglou, V . Panneershelvam, M. Lanctot, S. Dieleman, D. Grewe, J. Nham, N. Kalchbrenner, I. Sutskever, T. Lillicrap, M. Leach, K. Kavukcuoglu, T. Graepel, and D. Hassabis. Ma...

    2016

  45. [53]

    Silver, T

    D. Silver, T. Hubert, J. Schrittwieser, I. Antonoglou, M. Lai, A. Guez, M. Lanctot, L. Sifre, D. Kumaran, T. Graepel, et al. Mastering chess and shogi by self-play with a general reinforcement learning algorithm.arXiv preprint arXiv:1712.01815, 2017

    2017 arXiv

  46. [54]

    Silver, J

    D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker, M. Lai, A. Bolton, Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, T. Graepel, and D. Hassabis. Mastering the game of go without human knowledge.Nature, 550(7676):35...

    2017

  47. [55]

    Simon and W

    H. Simon and W. Chase. Skill in chess. InComputer chess compendium, pages 175–188. Springer, 1988

    1988

  48. [56]

    Steyvers, H

    M. Steyvers, H. Tejeda, G. Kerrigan, and P . Smyth. Bayesian modeling of human–ai complementarity.Proceedings of the National Academy of Sciences, 119(11):e2111547119, 2022

    2022

  49. [57]

    T. Tao. There’s more to mathematics than rigour and proofs, 2007

    2007

  50. [58]

    G. Todd, A. G. Padula, M. Stephenson, É. Piette, D. Soemers, and J. Togelius. Gavel: Generating games via evolution and language models.Advances in Neural Information Processing Systems, 37:110723–110745, 2024

    2024

  51. [59]

    K. E. Train.Discrete choice methods with simulation. Cambridge University Press, 2009

    2009

  52. [60]

    J. W. Uiterwijk. Solving strong and weak 4-in-a-row. In2019 IEEE conference on games (CIG), pages 1–8. IEEE, 2019

    2019

  53. [61]

    T. D. Ullman, E. Spelke, P . Battaglia, and J. B. Tenenbaum. Mind games: Game engines as an architecture for intuitive physics.Trends in cognitive sciences, 21(9):649–665, 2017

    2017

  54. [62]

    van Opheusden, I

    B. van Opheusden, I. Kuperwajs, G. Galbiati, Z. Bnaya, Y. Li, and W. J. Ma. Expertise increases planning depth in human gameplay.Nature, pages 1–6, 2023

    2023

  55. [63]

    Vinyals, I

    O. Vinyals, I. Babuschkin, W. M. Czarnecki, M. Mathieu, A. Dudzik, J. Chung, D. H. Choi, R. Powell, T. Ewalds, P . Georgiev, et al. Grandmaster level in starcraft ii using multi-agent reinforcement learning.Nature, 575(7782):350–354, 2019

    2019

  56. [64]

    E. Vul, N. Goodman, T. L. Griffiths, and J. B. Tenenbaum. One and done? Optimal decisions from very few samples.Cognitive science, 38(4):599–637, 2014. 19

    2014

  57. [65]

    X. Wang, Z. Lu, and M. Yin. Will you accept the AI recommendation? predicting human behavior in ai-assisted decision making. InProceedings of the ACM Web Conference 2022, pages 1697–1708, 2022

    2022

  58. [66]

    G. N. Yannakakis and J. Togelius.Artificial Intelligence and Games. Springer, 2018

    2018

  59. [67]

    L. Ying, K. M. Collins, P . Sharma, C. Colas, K. I. Zhao, A. Weller, Z. Tavares, P . Isola, S. J. Gershman, J. D. Andreas, et al. Assessing adaptive world models in machines with novel games.arXiv preprint arXiv:2507.12821, 2025

    2025 arXiv

  60. [68]

    Zhi-Xuan, J

    T. Zhi-Xuan, J. Mann, T. Silver, J. Tenenbaum, and V . Mansinghka. Online Bayesian goal inference for boundedly rational planning agents.Advances in neural information processing systems, 33:19238–19250, 2020

    2020

  61. [69]

    infinitely

    J.-Q. Zhu, A. N. Sanborn, and N. Chater. The bayesian sampler: Generic bayesian inference causes incoherence in human probability judgments.Psychological review, 127(5):719, 2020. Methods Game construction We manually constructed the 121 two-player competitive strategy game va...

    2020

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.