{"id":"3aa19dff-c0f0-4a22-99ff-e2bd46464d01","arxiv_id":"2507.05868","paper_version":1,"verdict":"UNVERDICTED","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A position paper proposing CogniPlay, a dual-process architecture for human-like general game playing, with no implementation or evaluation yet.","lead":"This paper sketches CogniPlay, a proposed general game-playing agent that mimics human thinking by combining pattern recognition with selective, memory-limited search. It is a work-in-progress blueprint, not a tested system, so no code, experiments, or results are reported.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Expert iteration with memory-bounded MCTS has no training signal for human-likeness; the model may converge to weak non-human play, so the central claim is unsupported.","rationale":"I agree with the reader's identification of the weakest assumption: self-play expert iteration will approximate human intuition. The paper is an honest work-in-progress, but the core design has a specific, identifiable gap. The training objective in Section III-A is purely self-referential: System 1 learns to imitate System 2 on states where System 1 is uncertain, and System 2 is guided by System 1's heuristic. This is a value/policy distillation loop, not a model of human cognition. Human-like behavior would have to emerge from the interaction of Spatial State-Action Features and memory-bounded search; the paper provides no inductive bias or constraint that would make that happen. The evaluation section describes how human-likeness could be measured but does not provide evidence. Because there is no implementation, the verdict remains UNVERDICTED; however, the theoretical concern is real and should be addressed in future work. I therefore leave the reader's UNVERDICTED unchanged.","tokens_in":6642,"tokens_out":4263,"duration_ms":49568,"concrete_test":"Implement CogniPlay in Ludii for Renju, train from scratch as specified, then evaluate per-state move matching against the Ludii human-game database. Compare against random moves, unrestricted MCTS, and memory-bounded MCTS with System 1 disabled. If CogniPlay's human move-match rate is not significantly higher than the memory-bounded MCTS baseline—or if match rate tracks playing strength rather than human agreement—the claimed emergence of human-likeness from expert iteration is falsified. A stronger variant: replace Spatial State-Action Features with a raw MCTS-visit-count policy for System 1; if move-matching is unchanged, the chunk-based mechanism contributes no human-likeness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CogniPlay produces human-like decisions in GGP without human-generated data. The load-bearing step is in Section III-A: System 1 is trained by expert iteration 'only on states where it exhibits a doubtful policy,' using the policy of System 2 (memory-bounded MCTS) as the target. Nothing in this self-referential loop biases the learned policy toward human play. Expert iteration minimizes the discrepancy between System 1 and a search policy that is itself an MCTS value estimate; MCTS optimizes expected outcome, not human style. Bounding memory or pruning actions with System 1 may reduce strength or inject arbitrary biases, but reduced strength is not human-likeness: a random or shallow search is also weak yet not human-like. The paper offers no argument, and no measurement (Section III-B admits 'no measurements have yet been made on its impact to their playing style'), that intermediate memory bounds trace a path through human-like behavior rather than simply degrading strong play. Since the claimed human-likeness is never an objective in training, the central claim rests on an unstated emergence assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":6846,"tokens_out":6579,"duration_ms":72455,"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":[{"comment":"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.","section":"Section III-A"},{"comment":"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.","section":"Section III-B"},{"comment":"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.","section":"Section III-C"}],"minor_comments":[{"comment":"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.","section":"Sections II and III-A"},{"comment":"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.","section":"Figure 1"},{"comment":"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.","section":"Section III-A"},{"comment":"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.","section":"Section III-C"},{"comment":"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.","section":"Section II"}],"recommendation":"major_revision","confidential_remarks":"This is a work-in-progress position paper with no experimental results. If the journal publishes vision papers or short proposals, it may be within scope after revisions. Otherwise, the lack of any implementation and evaluation is a serious scope problem. The authors also lean heavily on their own group's tools (Ludii, Spatial State-Action Features); this is not inappropriate, but independent validation of the human-likeness claims would be needed before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a clearly written position paper, not a research result. The authors propose CogniPlay, a dual-process GGP architecture: System 1 uses Spatial State-Action Features to partition legal actions into 'good' and 'bad' sets, and System 2 runs a memory-bounded MCTS only on the 'good' set. Training uses expert iteration without human data. The paper also sketches a sensible evaluation plan using move-matching and Turing-style tests, noting the need for a performance ceiling.\n\nWhat's genuinely new is the assembly: applying the DPTC framework to GGP with these specific components, and the training-on-doubtful-states scheme. The literature review is accurate and highlights a real gap—opponent modeling is underdeveloped in GGP. The authors are transparent that this is work-in-progress; they admit no measurements have been made on playing style.\n\nThe soft spot is the load-bearing assumption. The paper claims, in Section III-A, that training only on states where System 1 is doubtful, with System 2's policy as the target, 'allows the emergence of biases' approximating human behavior. The stress-test note is correct: expert iteration toward a search policy optimizes expected outcome, not human-likeness. Reducing memory bounds may degrade strength, but weakness is not human-like behavior. That is not necessarily a flaw in a proposal, but it is an unstated emergence assumption, and the paper currently offers neither a mechanism nor evidence. This affects the significance of the whole project, not just the final evaluation.\n\nMinor concerns: the design leans heavily on Ludii and Spatial State-Action Features, which are the authors' own prior work. That is legitimate given their relevance, but the generalizability claim would be stronger with independent components or baseline comparisons. The evaluation plan is thoughtful but depends on the Ludii database, which may limit breadth.\n\nFor a proposal, the internal logic is coherent and the writing is honest. It is not a full research paper and should not be judged as one. I'd send this to a workshop or short-paper track where the design can get community feedback; with implementation and preliminary results on Renju or Ludii games, it could become a solid conference paper. As it stands, I would not accept it for a full journal without results.","headline":"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.","tokens_in":7346,"tokens_out":3449,"would_cite":false,"duration_ms":37251,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CogniPlay is a general game-playing agent that mimics human decision-making through pattern-based intuition and memory-limited search, trained without human data.","keywords":["Human-like AI","General Game Playing","cognitive psychology","Monte-Carlo Tree Search","dual-process theory","expert iteration","action-set partitioning","Spatial State-Action Features"],"falsifier":"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.","tokens_in":6427,"feed_emoji":"🧠","tokens_out":6102,"duration_ms":62898,"temperature":0.7,"pith_summary":"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.","feed_headline":"Human-like game AI trained on no human moves","feed_subtitle":"CogniPlay pairs pattern-based intuition with memory-limited search to mimic human decision-making in any formal game.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the dual-process theory of cognition (System 1 and System 2) that the architecture instantiates.","marker":"[23]"},{"why":"Provides the precedent of a pattern-oriented chess engine whose piece relations form chunks and guide search.","marker":"[25]"},{"why":"Defines Spatial State-Action Features, the general, interpretable pattern vocabulary used for action partitioning.","marker":"[36]"},{"why":"Supplies the expert-iteration training loop through which System 1 learns from System 2's search policy.","marker":"[1]"},{"why":"Provides the memory-bounded MCTS algorithm used to model human memory limits in System 2.","marker":"[32]"},{"why":"Supplies the quantitative move-matching evaluation method the paper plans to use for human-likeness.","marker":"[27]"}],"fun_headline_variants":["Game AI trained without human moves","Pattern-based game AI mimics human intuition","CogniPlay: intuition-driven search without human data","No human moves needed for human-like game AI","Expert iteration replaces human game data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Game AI trained without human moves","Pattern-based game AI mimics human intuition","CogniPlay: intuition-driven search without human data","No human moves needed for human-like game AI","Expert iteration replaces human game data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000457,"raw_usage":{"total_tokens":2214,"prompt_tokens":784,"completion_tokens":1430,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":400,"completion_tokens_details":{"reasoning_tokens":1365}},"tokens_in":400,"tokens_out":1430,"duration_ms":11451,"temperature":1.0,"reasoning_tokens":1365,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:15:52.592632+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"A perspective on judgment and choice: mapping bounded rationality","cited_arxiv_id":null,"evidence_quote":"Supplies the dual-process theory of cognition (System 1 and System 2) that the architecture instantiates."},{"cited_title":"Adaptive pattern-oriented chess","cited_arxiv_id":null,"evidence_quote":"Provides the precedent of a pattern-oriented chess engine whose piece relations form chunks and guide search."},{"cited_title":"Spatial State-Action Features for General Games","cited_arxiv_id":"2201.06401","evidence_quote":"Defines Spatial State-Action Features, the general, interpretable pattern vocabulary used for action partitioning."},{"cited_title":"Memory Bounded Monte Carlo Tree Search","cited_arxiv_id":null,"evidence_quote":"Provides the memory-bounded MCTS algorithm used to model human memory limits in System 2."},{"cited_title":"Aligning superhuman ai with human behavior: Chess as a model system","cited_arxiv_id":null,"evidence_quote":"Supplies the quantitative move-matching evaluation method the paper plans to use for human-likeness."}],"review_version":1}