Pith. sign in

REVIEW 3 major objections 5 minor 42 references

CogniPlay: a work-in-progress Human-like model for General Game Playing

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read CogniPlay is a general game-playing agent that mimics human decision-making through pattern-based intuition and memory-limited search, trained without human data.

desk verdict A well-written work-in-progress proposal whose central assumption—that self-play alone can yield human-like play—is untested and needs argument or evidence. read the letter →

arxiv 2507.05868 v1 pith:X4FE5V5M submitted 2025-07-08 cs.AI

Aloïs Rautureau , Éric Piette This is my paper · ORCID
classification cs.AI
keywords Human-likeAIGeneralGamePlayingcognitivepsychologyMonte-CarloTreeSearchdual-processtheoryexpertiterationaction-setpartitioningSpatialState-ActionFeatures
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 is a proposal, not a report of results. It argues that a general game-playing agent can be made to decide the way humans do by combining two components: a pattern-based intuitive module that quickly labels actions as good or bad, and a deliberative module that searches only the actions the intuitive module approves. The distinctive claim is that this dual-system design can be trained entirely by self-play, without any human game data, and that the memory bounds of the search act as a dial for strength and style. If the claim holds, the same architecture would produce believable, adjustable opponents for any formally described game and generate plausible human-like play data for game reconstruction and playtesting.

What carries the argument

The load-bearing mechanism is action-set partitioning driven by Spatial State-Action Features: a general way to derive spatial patterns from a game's rules, used as System 1's pattern vocabulary. The partition decides when the analytical System 2 runs and what it may consider. The counterpart is selective memory-bounded Monte-Carlo Tree Search, which caps how much of the game tree is retained, reproducing limited human memory and forcing the search to revisit and deepen progressively. The expert-iteration loop ties them together: System 2's policy on doubtful states becomes the training signal that reshapes System 1's intuition.

What would settle it

Train CogniPlay from game rules alone on a game with a large human move database, then compare its move choices with held-out human decisions. If its move-matching accuracy is no better than a strong, memory-unbounded MCTS baseline, or if human raters in a blind evaluation do not rate its play as more human than that baseline's, the claim that self-play alone yields human-like decisions is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper's claim is that a human-like general game player can be assembled from three existing pieces: Spatial State-Action Features serve as a general, interpretable vocabulary of perceptual chunks; a policy derived from those features partitions each state's legal actions into 'intuitively good' and 'intuitively bad'; and a memory-bounded Monte-Carlo Tree Search explores only the intuitively good actions, bypassing search entirely when the policy is certain. System 1 (the pattern module) is trained by expert iteration, using System 2's search policy as the training target only on states where System 1 is doubtful. The authors explicitly call this training scheme suboptimal, but they expect it to let biases emerge and to avoid any need for human-generated data.

Load-bearing premise

The load-bearing premise is that a policy learned from self-play, using the search's own decisions as training targets, will approximate human intuition rather than merely strong play.

Editorial extensions

If this is right

  • If the model works, human-like play no longer requires human game data, so the approach scales to the many games for which no large move databases exist.
  • Memory bounds become an explicit control knob for playing strength and style, replacing artificial difficulty tuning by depth limits or reduced search time.
  • The same agent could generate believable gameplay for partially reconstructed traditional games, improving playtesting and reconstruction pipelines.
  • Because the model is game-general, validating human-likeness on popular games with abundant expert data would support the claim across all playable games, under the paper's transfer argument.

Reading between the lines

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

  • The authors leave open whether self-play expert iteration produces human-like biases or merely a fast, strong, non-human policy; a natural extension would be to test the trained System 1 directly against human move-frequency distributions.
  • A memory-bounded, pattern-driven agent could serve as a generative model of human errors, predicting which states humans mis-play rather than only which moves they choose.
  • The paper identifies opponent modeling as crucial for adversarial human behavior but leaves it out of CogniPlay; a concrete extension is a second, learned model of the opponent's policy that modulates the action partition.
  • The certainty/doubt threshold of System 1 could itself be treated as a personality parameter, trading off intuitive reflex against analytical deliberation, which the paper does not develop.
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 CogniPlay, a work-in-progress model for human-like General Game Playing (GGP). It reviews cognitive science results on chunking, dual-process theory, and bounded selective search, and relates prior systems such as MORPH, focused Minimax, Maia, and memory-bounded MCTS. The proposed architecture combines a System 1 that partitions legal actions into intuitively good and bad sets using Spatial State-Action Features with a System 2 that performs selective memory-bounded MCTS. System 1 is trained by expert iteration on states where its policy is doubtful, without human-generated data. The paper also proposes qualitative and quantitative evaluation protocols based on move matching and Turing-style tests. No implementation, experiments, or measurements are reported.

Significance. If the proposed architecture were shown to produce human-like play across arbitrary games without human data, it would fill a genuine gap in GGP and could enable more believable NPCs, better playtesting data, and improved reconstruction of traditional games. The literature review is relevant and the architecture is coherent, and the paper is honest about its early stage. However, because no experiments or formal arguments are provided, the central claim that the self-play loop yields human-like behavior is currently a hypothesis rather than a demonstrated result. The paper's present value is as a research proposal or position statement, not as a validated contribution.

major comments (3)
  1. [Section III-A] The central claim that expert iteration without human data yields human-like biases is unsupported. The training objective minimizes the discrepancy between System 1's policy and System 2's MCTS-derived policy; nothing in this objective rewards human-likeness. The paper states, 'While suboptimal, we expect this method to allow the emergence of biases in the model,' but no mechanism or evidence is provided. Since this is the load-bearing argument for the model's central claim, the authors need either a theoretical justification for why this self-referential loop would produce human-like rather than merely strong or arbitrary play, or a proof-of-concept experiment comparing the learned policy against human data.
  2. [Section III-B] Section III-B explicitly admits, 'no measurements have yet been made on its impact to their playing style.' Because the model's entire purpose is to replicate human-like playing style, the proposed memory-bounded MCTS mechanism is central to the contribution. Without any measurement, or even a precise operational definition of 'style' in this context, the statement that this scheme is 'better aligned with human-like planning' is not established. The paper needs at least a preliminary characterization of how memory bounds affect move distributions, decision times, or other style-relevant metrics.
  3. [Section III-C] The evaluation section describes plans rather than results. There are no move-matching scores, no Turing test outcomes, and no comparisons with human data from the Ludii database or any other source. The paper's own future-work paragraph confirms that implementation is pending. As a result, the manuscript does not yet support the claimed human-likeness even for a single game. A proof-of-concept on at least Renju, which the authors mention as the first implementation target, is necessary before the central claim can be assessed.
minor comments (5)
  1. [Sections II and III-A] The terminology is inconsistent: the paper refers to 'Spatial State-Action Features' in Section II and to 'Spatial Action-State Features' in Section III-A, while reference [36] is titled 'Spatial state-action features.' Please unify the terminology.
  2. [Figure 1] Figure 1 is dense and not walked through in the text. The authors should add a paragraph explaining the arrows and the roles of Systems 1 and 2 in both training and evaluation.
  3. [Section III-A] The notion of training only on 'doubtful' states needs clarification. If certainty is determined by whether System 1 selects a single action, then training on doubtful states alone may create a distribution shift or non-stationary training signal as the policy becomes more confident. The authors should discuss this dynamic.
  4. [Section III-C] The claim that validating on popular games 'supports the broader claim that it mimics human behavior consistently across all playable games' is a generalization that needs an argument. Human-likeness may be game-dependent, and the paper should address how results on Chess or Go would transfer to arbitrary Ludii games.
  5. [Section II] The paper would benefit from a more explicit comparison with Maia [27], since Maia is the most direct prior attempt at human-like play through supervised learning on human moves. The proposed expert-iteration approach without human data is clearly different, but the novelty would be easier to see if the distinction were made explicit.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: CogniPlay is a work-in-progress proposal with no fitted predictions; its reliance on the authors' prior tools is architectural reuse, not a self-referential derivation.

full rationale

CogniPlay makes no quantitative prediction and performs no derivation that would reduce to its inputs. The architecture combines Spatial State-Action Features [36], memory-bounded MCTS [32], and expert iteration [1]; System 1 learns from System 2's policy in doubtful states, which is standard expert iteration rather than a fitted parameter being renamed as a prediction. The human-likeness claim is a motivating aspiration, not a theorem or measured output, and the paper explicitly admits in Section III-B that no measurements of playing style have been made. Self-citations to Ludii [31], the Ludii database [10], and Spatial State-Action Features [36] supply reusable infrastructure from the same group, but they do not force or pre-encode the claimed conclusion that the model will be human-like; the conclusion is untested rather than assumed-by-construction. No equation or definition is circular, so the correct finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The central proposal leans on several unproven domain assumptions: that dual-process theory applies to GGP, that chunking is captured by Spatial State-Action Features, that self-play learning from a memory-bounded search produces human-like intuitions, and that the planned evaluation metrics measure human-likeness. The only explicit free parameter is the memory bound, with no value specified.

free parameters (1)
  • memory bound for MCTS = not specified
    Section III-B introduces 'an additional parameter, memory bounds, that can be adjusted to influence both playing strength and style.' No value or rule is given, and the paper presents no experiments that would fit it.
assumptions (5)
  • domain assumption Kahneman's Dual Process Theory of Cognition is a valid model for human game decision-making
    Section II cites [23] and uses System 1 and System 2 to justify the architecture.
  • domain assumption Chunking is the right granularity for representing game states in a GGP agent
    Section II reviews chunking literature [8, 11, 19, 20, 34, 41] and equates it with Spatial State-Action Features in Section III-A.
  • ad hoc to paper Self-play expert iteration without human data will produce human-like biases rather than merely strong play
    Section III-A states that training on doubtful states will 'allow the emergence of biases', but no evidence or mechanism is given.
  • ad hoc to paper Spatial State-Action Features provide a generalized chunked representation applicable to arbitrary games
    Section III-A relies on [36] for the feature policy without discussing failure modes or feature coverage limits.
  • domain assumption Move-matching accuracy plus qualitative ratings is a valid and sufficient measure of human-likeness
    Section III-C adopts this measurement framework and acknowledges limitations but still builds the evaluation plan on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CogniPlay: a work-in-progress Human-like model for General Game Playing." pith.science (2026). https://pith.science/paper/X4FE5V5M

@misc{pith2026250705868,
  author       = {Pith},
  title        = {Pith review of: CogniPlay: a work-in-progress Human-like model for General Game Playing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X4FE5V5M}},
  note         = {Machine review of arXiv:2507.05868}
}
read the original abstract

While AI systems have equaled or surpassed human performance in a wide variety of games such as Chess, Go, or Dota 2, describing these systems as truly "human-like" remains far-fetched. Despite their success, they fail to replicate the pattern-based, intuitive decision-making processes observed in human cognition. This paper presents an overview of findings from cognitive psychology and previous efforts to model human-like behavior in artificial agents, discusses their applicability to General Game Playing (GGP) and introduces our work-in-progress model based on these observations: CogniPlay.

Figures

Figures reproduced from arXiv: 2507.05868 by the authors.

Figure 1
Figure 1. Our model and experimental setup for a human-like agent in GGP. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 40 canonical work pages

  1. [1]

    Thinking fast and slow with deep learning and tree search

    Thomas Anthony, Zheng Tian, and David Barber. Thinking fast and slow with deep learning and tree search. CoRR, abs/1705.08439, 2017

    arXiv 2017

  2. [2]

    Dota 2 with large scale deep reinforcement learning

    Christopher Berner, Greg Brockman, Brooke Chan, Vicki Cheung, Przemysław D˛ ebiak, Christy Dennison, David Farhi, Quirin Fischer, Shariq Hashme, Chris Hesse, et al. Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680 , 2019

    arXiv 1912

  3. [3]

    Modern techniques for ancient games

    Cameron Browne. Modern techniques for ancient games. In 2018 IEEE Conference on Computational Intelligence and Games (CIG), pages 1–8. IEEE, 2018

    work page 2018

  4. [4]

    Foundations of Digital Arch{\ae}oludology

    Cameron Browne, Dennis JNJ Soemers, Éric Piette, Matthew Stephen- son, Michael Conrad, Walter Crist, Thierry Depaulis, Eddie Duggan, Fred Horn, Steven Kelk, et al. Foundations of digital arch {\ae} oludology. arXiv preprint arXiv:1905.13516 , 2019

    work page Pith review arXiv 1905

  5. [5]

    Deep blue

    Murray Campbell, A.Joseph Hoane, and Feng hsiung Hsu. Deep blue. Artificial Intelligence, 134(1):57–83, 2002

    work page 2002

  6. [6]

    On the utility of learning about humans for human-ai coordination

    Micah Carroll, Rohin Shah, Mark K Ho, Tom Griffiths, Sanjit Seshia, Pieter Abbeel, and Anca Dragan. On the utility of learning about humans for human-ai coordination. In Advances in Neural Information Processing Systems, volume 32. Curran Associates, Inc., 2019

    work page 2019

  7. [7]

    Search in chess: Age and skill differences

    Neil Charness. Search in chess: Age and skill differences. Journal of Ex- perimental Psychology: Human Perception and Performance, 7(2):467– 476, 1981

    work page 1981

  8. [8]

    Chase and Herbert A

    William G. Chase and Herbert A. Simon. Perception in chess. Cognitive Psychology, 4(1):55–81, 1973

    work page 1973

Show all 42 references
  1. [9]

    Computational approaches for recognising and reconstructing ancient games: The case of Ludus Latrunculorum

    Walter Crist, Éric Piette, Dennis JNJ Soemers, Matthew Stephenson, and Cameron Browne. Computational approaches for recognising and reconstructing ancient games: The case of Ludus Latrunculorum. The Archaeology of Play: Material Approaches to Games and Gaming in , 2023

    2023

  2. [10]

    The ludii games database: A resource for computational and cultural research on traditional board games

    Walter Crist, Matthew Stephenson, Eric Piette, and Cameron Browne. The ludii games database: A resource for computational and cultural research on traditional board games. Digital Humanities Quarterly , 2024

    2024

  3. [11]

    A. D. De Groot. Thought of the Chess-Player, as study in Experimental Psychology. Noord-Holl. Uitg., Oxford, England, 1946

    1946

  4. [12]

    Thought and choice in chess , volume 4

    Adriaan D De Groot and Adrianus Dingeman De Groot. Thought and choice in chess , volume 4. Walter de Gruyter, 1978

    1978

  5. [13]

    Donkers, J.W.H.M

    H.H.L.M. Donkers, J.W.H.M. Uiterwijk, and H.J. van den Herik. Prob- abilistic opponent-model search. Information Sciences, 135(3):123–149,

  6. [14]

    Mastering the game of go with deep neural networks and tree search

    David Silver et al. Mastering the game of go with deep neural networks and tree search. Nature, 529(7587):484–489, Jan 2016

    2016

  7. [15]

    Plug-and-play policy planner for large language model powered dialogue agents

    Yang Deng et al. Plug-and-play policy planner for large language model powered dialogue agents. In The Twelfth International Conference on Learning Representations, 2024

    2024

  8. [16]

    Robert M. French. Subcognition and the limits of the turing test. Mind, 99(393):53–65, 1990

    1990

  9. [17]

    General game playing: Overview of the aaai competition

    Michael Genesereth, Nathaniel Love, and Barney Pell. General game playing: Overview of the aaai competition. AI magazine, 26(2):62–62, 2005

    2005

  10. [18]

    A Pattern-recognition Theory of Search in Expert Problem Solving

    Fernand Gobet. A Pattern-recognition Theory of Search in Expert Problem Solving. Thinking & Reasoning , 3(4):291–313, November 1997

    1997

  11. [19]

    Fernand Gobet and Herbert A. Simon. Expert Chess Memory: Revisiting the Chunking Hypothesis. Memory, 6(3):225–255, 1998

    1998

  12. [20]

    Fernand Gobet and Herbert A. Simon. Five Seconds or Sixty? Pre- sentation Time in Expert Memory. Cognitive Science , 24(4):651–682, 2000

    2000

  13. [21]

    Planning like human: A dual-process framework for dialogue planning

    Tao He, Lizi Liao, Yixin Cao, Yuanxing Liu, Ming Liu, Zerui Chen, and Bing Qin. Planning like human: A dual-process framework for dialogue planning. arXiv preprint arXiv:2406.05374 , 2024

    2024 arXiv

  14. [22]

    Using Knowledge about the Opponent in Game-Tree Search

    Peter Jozef Jansen. Using Knowledge about the Opponent in Game-Tree Search. phdthesis, Carnegie Mellon University, 1992

    1992

  15. [23]

    A perspective on judgment and choice: mapping bounded rationality

    Daniel Kahneman. A perspective on judgment and choice: mapping bounded rationality. Am Psychol, 58(9):697–720, 2003

    2003

  16. [24]

    Bandit based monte-carlo plan- ning

    Levente Kocsis and Csaba Szepesvári. Bandit based monte-carlo plan- ning. In Johannes Fürnkranz, Tobias Scheffer, and Myra Spiliopoulou, editors, Machine Learning: ECML 2006, pages 282–293. Springer Berlin Heidelberg, 2006

    2006

  17. [25]

    Adaptive pattern-oriented chess

    Robert Levinson and Richard Snyder. Adaptive pattern-oriented chess. In Lawrence A. Birnbaum and Gregg C. Collins, editors, Machine Learning Proceedings 1991 , pages 85–89. Morgan Kaufmann, 1991

    1991

  18. [26]

    Towards cognitively plausible game playing systems

    Jacek Ma ´ndziuk. Towards cognitively plausible game playing systems. IEEE Computational Intelligence Magazine , 6(2):38–51, 2011

    2011

  19. [27]

    Aligning superhuman ai with human behavior: Chess as a model system

    Reid McIlroy-Young, Siddhartha Sen, Jon Kleinberg, and Ashton An- derson. Aligning superhuman ai with human behavior: Chess as a model system. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’20. ACM, 2020

    2020

  20. [28]

    Navigates like me: Understanding how people evaluate human-like ai in video games

    Stephanie Milani, Arthur Juliani, Ida Momennejad, Raluca Georgescu, Jaroslaw Rzepecki, Alison Shaw, Gavin Costello, Fei Fang, Sam Devlin, and Katja Hofmann. Navigates like me: Understanding how people evaluate human-like ai in video games. In Proceedings of the 2023 CHI Confer...

    2023

  21. [29]

    Moriarty and Risto Miikkulainen

    David E. Moriarty and Risto Miikkulainen. Evolving neural networks to focus minimax search. In Proceedings of the Twelfth National Con- ference on Artificial Intelligence (AAAI-94) , pages 1371–1377, Seattle, W A, 1994. Cambridge, MA: MIT Press

    1994

  22. [30]

    Gametable cost action: Kickoff report

    Éric Piette, Walter Crist, Dennis JNJ Soemers, Lisa Rougetet, Summer Courts, Tim Penn, and Achille Morenville. Gametable cost action: Kickoff report. ICGA Journal, 46(1):11–27, 2024

    2024

  23. [31]

    Ludii–the ludemic general game system

    Eric Piette, Dennis JNJ Soemers, Matthew Stephenson, Chiara F Sironi, Mark HM Winands, and Cameron Browne. Ludii–the ludemic general game system. In ECAI 2020, pages 411–418. IOS Press, 2020

    2020

  24. [32]

    Memory Bounded Monte Carlo Tree Search

    Edward Powley, Peter Cowling, and Daniel Whitehouse. Memory Bounded Monte Carlo Tree Search. Proceedings of the AAAI Con- ference on Artificial Intelligence and Interactive Digital Entertainment , 13(1):94–100, 2017

    2017

  25. [33]

    Does the chimpanzee have a theory of mind? Behavioral and brain sciences , 1(4):515–526, 1978

    David Premack and Guy Woodruff. Does the chimpanzee have a theory of mind? Behavioral and brain sciences , 1(4):515–526, 1978

    1978

  26. [34]

    Skilled perception in Go: Deducing memory structures from inter-response times

    Judith S Reitman. Skilled perception in Go: Deducing memory structures from inter-response times. Cognitive Psychology, 8(3):336–356, 1976

    1976

  27. [35]

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

    1950

  28. [36]

    Dennis J. N. J. Soemers, Éric Piette, Matthew Stephenson, and Cameron Browne. Spatial state-action features for general games. CoRR, abs/2201.06401, 2022

    2022 arXiv

  29. [37]

    Soemers, Spyridon Samothrakis, Éric Piette, and Matthew Stephenson

    Dennis J.N.J. Soemers, Spyridon Samothrakis, Éric Piette, and Matthew Stephenson. Extracting tactics learned from self-play in general games. Information Sciences, 624:277–298, 2023

    2023

  30. [38]

    Adversarial problem solving: Modeling an opponent using explanatory coherence

    Paul Thagard. Adversarial problem solving: Modeling an opponent using explanatory coherence. Cognitive Science, 16(1):123–149, 1992

    1992

  31. [39]

    A. M. Turing. I.—computing machinery and intelligence. Mind, LIX(236):433–460, 10 1950

    1950

  32. [40]

    Human-like motion planning based on game theoretic decision making

    Annemarie Turnwald and Dirk Wollherr. Human-like motion planning based on game theoretic decision making. International Journal of Social Robotics, 11(1):151–170, 2019

    2019

  33. [41]

    Wolff, Donald H

    Alan S. Wolff, Donald H. Mitchell, and Peter W. Frey. Perceptual Skill in the Game of Othello. The Journal of Psychology, 118(1):7–16, 1984

    1984

  34. [2001]

    Heuristic Search and Computer Game Playing

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.