Pith. sign in

REVIEW 3 major objections 5 minor 11 cited by

Assessing Adaptive World Models in Machines with Novel Games

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 This paper argues that judging AI by its play on 'novel games'—games with hidden, changing, or freshly invented rules—is the way to measure whether machines can quickly build and revise world models.

desk verdict A well-written perspective that names and organizes a real gap in AI evaluation, but its central construct—novel games as a measure of world model induction—remains unvalidated, and the paper knows it. read the letter →

arxiv 2507.12821 v2 pith:WGLT2HZ5 submitted 2025-07-17 cs.AI cs.LG

Lance Ying , Katherine M. Collins , Prafull Sharma , Cedric Colas , Kaiya Ivy Zhao , Adrian Weller , Zenna Tavares , Phillip Isola
show 6 more authors
Samuel J. Gershman Jacob D. Andreas Thomas L. Griffiths Francois Chollet Kelsey R. Allen Joshua B. Tenenbaum
This is my paper
classification cs.AIcs.LG
keywords worldmodelinductionnovelgamesevaluationframeworksampleefficiencygeneralizationhierarchicalBayesianmodelsreinforcementlearningadaptiveintelligence
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 argues that the kind of intelligence people rely on in unfamiliar situations is the ability to quickly build and refine internal models of how an environment works, and that this ability—termed world model induction—should be the target of AI evaluation. Much current evaluation, the authors contend, mostly tests static representations trained on large datasets, not the process of constructing and updating a model through limited interaction. To close the gap, they propose evaluating AI on 'novel games': games whose underlying rules, objects, or objectives are initially unknown, hidden, or dynamically changing, so that success requires inducing a new world model on the fly. The paper sets out desiderata for building such game suites, metrics for measuring sample efficiency, exploration quality, and internal model structure, and a generative design process to keep refreshing novelty.

What carries the argument

The load-bearing object is the 'novel game', defined as a game with genuine, deep, and continually refreshing novelty in its underlying structure, where rules, mechanics, object properties, or objectives are initially unknown, hidden, or dynamically changing. Supporting it is a hierarchical world-model framework in which an instance world model $\Omega_1$ is generated by more abstract world models $\Omega_2,\dots,\Omega_n$, updated through the Bayesian relation $P(\Omega|e)\propto P(e|\Omega_1)P(\Omega_1|\Omega_2)\cdots P(\Omega_n|\Omega_{n+1})$; this hierarchy explains how abstract prior knowledge makes new-model induction sample-efficient. The proposed evaluation uses this machinery to design games that force induction at several levels and to select metrics—sample efficiency, qualitative exploration patterns, and probes of internal representations—that reveal whether an agent is building and revising a model rather than exploiting familiar structure.

What would settle it

A controlled experiment in which a model with no structured or inspectable world model—for example, an end-to-end reactive policy with large capacity—matches human-level sample efficiency and robust within-domain generalization across a suite of novel games, while post-hoc probes find no representation of the latent rules, would settle against the paper's central claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that rapid adaptation in humans is enabled by hierarchical world model induction, and that this capacity can and should be measured in machines by performance in novel games rather than by performance on familiar or overtrained domains. A novel game is defined not by unseen instances or parametric variation within a familiar structure, but by genuine, deep, and continually refreshing novelty in the underlying game structure, where rules, object properties, and objectives may be latent or dynamic; a player succeeds only by actively experimenting, forming hypotheses, and revising its internal model. The authors argue that adaptation within such games, combined with direct inspection of the agent's inferred rules or internal representations, provides evidence of world model induction. They propose that benchmark suites should be generative—continually producing new variants by modifying mechanics, objectives, player counts, or by combining games—so that novelty does not decay, and that human learnability must be validated to make AI performance interpretable.

Load-bearing premise

The framework rests on the assumption, which the paper explicitly acknowledges as unresolved in Section 6, that success in novel games with hidden or changing rules requires inducing structured internal world models rather than achieving the same behavior through model-free pattern matching or massive-capacity generalization.

Editorial extensions

If this is right

  • Benchmarks built on novel games would measure the adaptation process itself, not just final task performance, so progress could be tracked in how quickly agents reach proficiency.
  • The benchmark cannot be a fixed suit of games, because novelty decays as models experience a game; it must continually generate fresh variants, otherwise it silently becomes a test of familiar skill.
  • Games that remain intuitive to human players let AI results be benchmarked directly against human learning curves, grounding claims about human-like adaptation.
  • Probing internal representations would distinguish genuine model building from reactive strategy: if no internal model is found despite task success, the game is not measuring world model induction.

Reading between the lines

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

  • A direct consequence the authors leave implicit is that the benchmark's value decays against any agent that has already encountered a game variant, so a credible evaluation needs strict rules about which variants count as novel for each model, and those rules are themselves part of the measurement.
  • The framework suggests a training-time shift: agents should be rewarded for information-gathering and hypothesis-testing behavior, not only for task reward, since the metric of sample efficiency rewards epistemic actions; the paper does not develop this.
  • If the central claim holds, the same latent-mechanic evaluation could be ported from games to domains like tool use, social interaction, or scientific discovery, where the environment itself withholds the rules; the paper flags external validity as open, so extending the paradigm is a natural next step.
  • One testable extension is to pit model-free agents with large capacity against a novel-game suite with probes, testing the paper's own acknowledged open question in Section 6: whether such agents could match human sample efficiency without building structured internal models.
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. This Perspective argues that human-like rapid adaptation depends on the ability to construct and refine internal world models, a process the authors call world model induction. The paper critiques existing AI world-model evaluations for focusing on static representations learned from large datasets, and proposes a new benchmarking paradigm based on 'novel games': games with genuine, deep, and continually refreshing structural novelty in rules, mechanics, objects, or objectives. It specifies desiderata for such games (structural novelty, human intuitiveness/learnability, diversity), proposes evaluation metrics (sample-efficiency curves, qualitative exploration analysis, and probing of internal representations), and gives practical guidance for generating novel games and validating them with human players. The paper is a position/proposal piece: it contains no new empirical results, and it explicitly acknowledges open questions about the necessity of world models, measurement of internal representations, and external validity.

Significance. If the proposed framework is adopted, it could help redirect AI evaluation toward adaptive, sample-efficient behavior and provide a shared vocabulary for studying world models in machines. The paper's strengths are its grounding in cognitive science, its clear articulation of desiderata, and its honest acknowledgment of alternatives (e.g., Brooks 1991) and open questions. It also usefully connects to concrete existing benchmarks (ARC-AGI-3, AutumnBench, VGDL games, Virtual Tools) and recent generative-game efforts. However, the central measurement claim—that performance on novel games provides a valid measure of world model induction—is currently not secured, because the proposed metrics cannot rule out model-free or architecture-general adaptation. The paper would be a more convincing call to action if it specified differential tests or controls that could falsify model-free interpretations, and if it made internal-model probing a required rather than optional component of the evaluation.

major comments (3)
  1. [§4.2 and §6] The benchmark's construct validity is not secured. The proposed metrics—sample-efficiency curves, qualitative exploration analyses, and optional probing—do not provide a differential test capable of ruling out world-model-free adaptation. A large-capacity transformer or deep RL agent trained on many generated games could plausibly produce steep learning curves and directed exploration through compositional pattern matching, without constructing a structured causal model. The paper itself acknowledges this possibility in Section 6, citing Brooks (1991), but leaves it as an open question rather than designing the benchmark to answer it. Probing of internal representations could in principle disambiguate, but it is presented as an architecture-dependent complement and not as a scoring requirement. As written, a high benchmark score could be achieved by an agent that fails the intended construct. I recommend that the paper require internal-model probes (or another scoreable model-identity diagnostic) for any claim of world model induction, and that it propose control conditions with matched surface novelty but known simple latent rules, or games where the latent rule is explicitly perturbed, to create falsifiable predictions that distinguish model-based from model-free adaptation.
  2. [§4.1 and §4 intro] The definition of 'novel games' is partly circular with respect to the measured construct. The paper defines novel games as games that 'require players to construct new world models or modify their models,' and then treats adaptation performance on such games as evidence of world model induction. To break this circularity, the paper should provide an operational, agent-independent criterion for 'genuine, deep, and continually refreshing novelty.' For instance, novelty could be quantified by the minimal description length of the rule change relative to a generative grammar, or by ensuring that the rule set is not expressible as a small parametric perturbation of prior training games. Without such an operational criterion, the dichotomy between 'parametric variation' and 'structural novelty' in Section 3 remains empirically underspecified, and the benchmark cannot distinguish between adaptation that reuses existing compositional patterns and adaptation that genuinely constructs a new model.
  3. [§2, Eq. (1) and §4.2] The paper's formal hierarchical Bayesian framework (Eq. 1) is not sufficiently connected to the proposed metrics. The framework distinguishes rapid learning (efficiency of posterior inference) from robust generalization (quality of the prior over world models), but the metrics in Section 4.2—sample efficiency curves and exploration traces—conflate these components. A system with a powerful prior over game structures could show fast adaptation without a more efficient induction process, while a system with strong within-game inference could show steep learning curves even with a poor prior. The paper should specify how the proposed metrics decompose these contributions, or propose additional metrics (e.g., transfer efficiency across a held-out game family, posterior calibration, or counterfactual prediction accuracy) that target the hierarchical structure claimed in Eq. (1).
minor comments (5)
  1. [§1] The phrase 'especially inhuman worlds' appears to be a typo; it should likely read 'especially in human worlds' or 'in human-like worlds, particularly...'.
  2. [§6] The sentence 'This view has met with some skepticism by recent work in AI' is slightly awkward; consider rephrasing to 'recent work in AI has expressed skepticism toward this view'.
  3. [References] In the reference to Pourcel et al., the title contains 'with with autocatalytic generative models'; the duplicated 'with' should be removed.
  4. [§5.1] The examples in Figure 4 are helpful, but the text would benefit from a short discussion of how these variant-creation methods relate to the operational criterion for structural novelty requested in the major comments, since 'changed objective' could range from a trivial re-labeling to a genuinely new causal rule.
  5. [§4.2] When describing probing of internal world models, the paper mentions question answering for natural-language agents but does not discuss how to handle agents with no linguistic interface; a brief note on probing methods for non-linguistic RL or VLM agents would improve the completeness of this subsection.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is a self-aware perspective whose construct-validity gap is explicitly acknowledged, not a derivation that reduces to its inputs.

full rationale

This is a Perspective paper that proposes an evaluation paradigm rather than deriving predictions from fitted parameters. It contains no equations that equate a fitted quantity with a reported result, no 'uniqueness theorem' imported from the authors' prior work to forbid alternatives, and no empirical benchmark results that could be tautological. The closest candidate for circularity is the construct-level claim that success in 'novel games' requires world model induction (Section 4). However, the paper explicitly disclaims this inference in Section 6: it states that 'one may question if a reactive or purely end-to-end system with massive capacity might exhibit surprising levels of adaptation through pattern generalization across diverse training data, even if it lacks a structured, predictive model of environmental dynamics,' and calls the question 'an open empirical question.' It also separately acknowledges that 'directly measuring the quality and efficiency of internal world model induction at different levels of hierarchy presents its own challenges' and that metrics 'beyond just task success' are still needed. These passages show that the authors do not treat benchmark performance as being definitionally equivalent to world model induction; they present the link as a motivated hypothesis with an unresolved model-free alternative. That is a construct-validity risk, not a circularity of the kind where an input is renamed as an output. The self-citations to the authors' prior cognitive-science work (e.g., Tenenbaum et al. 2011; Tsividis et al. 2021) are used as domain evidence rather than as an unverified forced premise. Accordingly, no circular step is exhibited.

Assumptions & free parameters 0 free parameters · 3 assumptions · 2 invented entities

The framework rests on a small set of domain assumptions from cognitive science and AI evaluation; no numerical free parameters are fitted, and no new physical entities are proposed. The two introduced constructs, 'world model induction' and 'novel games', are defined conceptually and await external validation.

assumptions (3)
  • domain assumption Human rapid adaptation is enabled by constructing and refining hierarchical internal world models.
    Borrowed from Tenenbaum et al. 2006, 2011; underpins the entire argument in Section 2.
  • domain assumption Games are a uniquely well-suited domain for evaluating adaptive intelligence.
    Stated in Section 3, following Allen et al. 2024; no direct evidence given here.
  • ad hoc to paper Genuinely novel games with hidden or changing rules necessitate induction of new world models.
    Core construct-validity premise in Section 4.1; acknowledged in Section 6 as an open question regarding model-free adaptation.
invented entities (2)
  • World model induction
    purpose: Named capacity for efficiently constructing and refining internal world models through interaction, posited as critical for AGI; used as the target construct for the proposed benchmarks.
    Introduced in Section 2; the paper provides no independent measure of it, and acknowledges in Section 6 that direct measurement remains challenging.
  • Novel games
    purpose: Proposed benchmark class of games with genuine, deep, continually refreshing structural novelty to trigger world model induction.
    Defined in Section 4; no concrete instance validated, and the paper notes novelty is ephemeral as games become familiar.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Assessing Adaptive World Models in Machines with Novel Games." pith.science (2026). https://pith.science/paper/WGLT2HZ5

@misc{pith2026250712821,
  author       = {Pith},
  title        = {Pith review of: Assessing Adaptive World Models in Machines with Novel Games},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WGLT2HZ5}},
  note         = {Machine review of arXiv:2507.12821}
}
read the original abstract

Human intelligence exhibits a remarkable capacity for rapid adaptation and effective problem-solving in novel and unfamiliar contexts. We argue that this profound adaptability is fundamentally linked to the efficient construction and refinement of internal representations of the environment, commonly referred to as world models, and we refer to this adaptation mechanism as world model induction. However, current understanding and evaluation of world models in artificial intelligence (AI) remains narrow, often focusing on static representations learned from training on massive corpora of data, instead of the efficiency and efficacy in learning these representations through interaction and exploration within a novel environment. In this Perspective, we provide a view of world model induction drawing on decades of research in cognitive science on how humans learn and adapt so efficiently; we then call for a new evaluation framework for assessing adaptive world models in AI. Concretely, we propose a new benchmarking paradigm based on suites of carefully designed games with genuine, deep and continually refreshing novelty in the underlying game structures -- we refer to this class of games as novel games. We detail key desiderata for constructing these games and propose appropriate metrics to explicitly challenge and evaluate the agent's ability for rapid world model induction. We hope that this new evaluation framework will inspire future evaluation efforts on world models in AI and provide a crucial step towards developing AI systems capable of human-like rapid adaptation and robust generalization -- a critical component of artificial general intelligence.

Figures

Figures reproduced from arXiv: 2507.12821 by the authors.

Figure 1
Figure 1. Framework for characterizing world models across different levels of abstractions. a). World models within a hierarchical Bayesian framework. The structured probabilistic model Ω1 (ad-hoc world model) generates expectations about possible observations e, while abstract knowl￾edge and principles (abstract world model) Ω2, Ω3, . . . generate the space of possible structures for Ω1. Each level of abstraction generates … view at source ↗
Figure 2
Figure 2. Case studies of novel environments for testing world model induction a) ARC-AGI-3 (ARC Prize, 2025) is an interactive reasoning benchmark. Players are not given instructions about the gameplay or win conditions; instead, they must infer game rules and objectives through interaction. In the featured example, players can move the tiles horizontally. The objective is to align the tiles so that the cells matching the st… view at source ↗
Figure 3
Figure 3. Evaluation methods used by Tsividis et al. (2021) for comparing learning behaviors across humans and models (EMPA and DDQN). a) Quantitative learning efficiencies plotted as the number of levels passed as a function of exploration steps. b) Qualitative analysis of different exploration patterns demonstrated by different learning agents. Similarly, the EMPA model proposed in the same paper shows more structured but s… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Creating novel games. There are many ways that researchers can create novel games. Variants could be sourced from existing games by modifying the environment and number of players, mechanics, or objective function. Novel games could also be formed by combining existing…

Discussion (0). Sign in to comment.

Forward citations

Cited by 11 Pith papers

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

  1. Open-Ended Video Game Glitch Detection with Agentic Reasoning and Temporal Grounding

    cs.MA 2026-04 unverdicted novelty 7.0 of 10

    Introduces the first benchmark for open-ended video game glitch detection with temporal localization and proposes GliDe, an agentic framework that achieves stronger performance than vanilla multimodal models.

  2. Tycho: Active Abstraction with Programmatic World Models for ARC-AGI-3

    cs.AI 2026-07 conditional novelty 6.5 of 10

    Selective programmatic world modeling (actor-requested builder) yields 100 RHAE on all 183 public ARC-AGI-3 levels, while automatic repair is more transition-exact but weaker at play.

  3. Training Language Agents to Learn from Experience

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    Introduces the ICT framework and an RL pipeline to train language agent reflectors that distill experience into reusable prompts, outperforming baselines on held-out tasks in ALFWorld and MiniHack.

  4. Mini Amusement Parks (MAPs): A Testbed for Modelling Business Decisions

    cs.AI 2025-11 conditional novelty 6.0 of 10

    MAPs is a new amusement-park simulator benchmark on which frontier LLM agents score 7–15% of human performance, exposing persistent gaps in long-horizon planning, active learning, spatial reasoning, and handling stoch...

  5. BuilderBench: The Building Blocks of Intelligent Agents

    cs.AI 2025-10 conditional novelty 6.0 of 10

    BuilderBench is a fast, open-source 3D block-building benchmark where current RL and LLM agents fail at all non-trivial construction tasks, exposing weak open-ended exploration.

  6. Identifying Informative Environments for Cognition Parameter Inference via Bayesian Experimental Design

    cs.AI 2026-07 conditional novelty 5.0 of 10

    Environment design for cognitive parameter inference is formulated as Bayesian experimental design, with an amortized surrogate recovering exact Monte Carlo environment rankings on Mouselab-MDP.

  7. stable-worldmodel: A Platform for Reproducible World Modeling Research and Evaluation

    cs.LG 2026-05 unverdicted novelty 5.0 of 10

    The paper presents stable-worldmodel (swm), a platform with high-performance data layer, modern world model baselines, planning solvers, and extended environments for reproducible research and generalization evaluation.

  8. Scalable Environments Drive Generalizable Agents

    cs.AI 2026-05 unverdicted novelty 5.0 of 10

    Generalizable agents require environment scaling via diverse executable rule-sets, distinguished from trajectory and task scaling in a new taxonomy.

  9. People use fast and flat simulation to reason about new games

    q-bio.NC 2025-10 conditional novelty 5.0 of 10

    Novices' first-contact judgments, moves, and predictions across 121 new games are best explained by an 'Intuitive Gamer' model that runs a handful of shallow goal-directed self-play simulations, not by deep search or ...

  10. Identifying Informative Environments for Cognition Parameter Inference via Bayesian Experimental Design

    cs.AI 2026-07 conditional novelty 4.0 of 10

    On Mouselab-MDP, amortized Bayesian experimental design ranks planning environments by expected information gain almost as well as exact Monte Carlo BED, and no single environment wins on all objectives.

  11. Hypothesis Generation and Inductive Inference in Children and Language Models

    cs.AI 2026-05 unverdicted novelty 4.0 of 10

    Children and LLM agents show parallel adaptations to evidence reliability in a Bayesian program induction task but differ in information-seeking costs and compliance.

Reference graph

Works this paper leans on

72 extracted references · 54 canonical work pages · cited by 10 Pith papers

  1. [1]

    E., Gershman, S

    Allen, K., Br \"a ndle, F., Botvinick, M., Fan, J. E., Gershman, S. J., Gopnik, A., Griffiths, T. L., Hartshorne, J. K., Hauser, T. U., Ho, M. K., et al. (2024). Using games to understand the mind. Nature Human Behaviour , pages 1--9

    work page 2024

  2. [2]

    R., Smith, K

    Allen, K. R., Smith, K. A., and Tenenbaum, J. B. (2020). Rapid trial-and-error learning with simulation supports flexible tool use and physical reasoning. Proceedings of the National Academy of Sciences , 117(47):29302--29310

    work page 2020

  3. [3]

    Andreas, J. (2024). Language models, world models, and human model-building

    work page 2024

  4. [4]

    Arc-agi-3

    ARC Prize (2025). Arc-agi-3. Retrieved from https://arcprize.org/arc-agi/3

    work page 2025

  5. [5]

    Autumnbench: World model learning in humans and ai

    Basis (2025). Autumnbench: World model learning in humans and ai

    work page 2025

  6. [6]

    Bauer, J., Baumli, K., Behbahani, F., Bhoopchand, A., Bradley-Schmieg, N., Chang, M., Clay, N., Collister, A., Dasagi, V., Gonzalez, L., et al. (2023). Human-timescale adaptation in an open-ended task space. In International Conference on Machine Learning , pages 1887--1935. PMLR

    work page 2023

  7. [7]

    G., Naddaf, Y., Veness, J., and Bowling, M

    Bellemare, M. G., Naddaf, Y., Veness, J., and Bowling, M. (2013). The arcade learning environment: An evaluation platform for general agents. Journal of artificial intelligence research , 47:253--279

    work page 2013

  8. [8]

    Bengio, Y., Cohen, M., Fornasiere, D., Ghosn, J., Greiner, P., MacDermott, M., Mindermann, S., Oberman, A., Richardson, J., Richardson, O., et al. (2025). Superintelligent agents pose catastrophic risks: Can scientist ai offer a safer path? arXiv preprint arXiv:2502.15657

    arXiv 2025

Show all 72 references
  1. [9]

    Brooks, R. A. (1991). Intelligence without representation. Artificial intelligence , 47(1-3):139--159

    1991

  2. [10]

    J., and Hsu, F.-h

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

    2002

  3. [11]

    Chase, W. G. and Simon, H. A. (1973). The mind's eye in chess. In Visual information processing , pages 215--281. Elsevier

    1973

  4. [12]

    Chollet, F. (2019). On the measure of intelligence. arXiv preprint arXiv:1911.01547

    2019 arXiv

  5. [13]

    Chollet, F., Knoop, M., Kamradt, G., Landers, B., and Pinkard, H. (2025). Arc-agi-2: A new challenge for frontier ai reasoning systems. arXiv preprint arXiv:2505.11831

    2025 arXiv

  6. [14]

    and Schulz, L

    Chu, J. and Schulz, L. (2020). Exploratory play, rational action, and efficient search. In Proceedings of the Annual Meeting of the Cognitive Science Society , volume 42

    2020

  7. [15]

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

    1907

  8. [16]

    Cobbe, K., Hesse, C., Hilton, J., and Schulman, J. (2020). Leveraging procedural generation to benchmark reinforcement learning. In International conference on machine learning , pages 2048--2056. PMLR

    2020

  9. [17]

    M., Sucholutsky, I., Bhatt, U., Chandra, K., Wong, L., Lee, M., Zhang, C

    Collins, K. M., Sucholutsky, I., Bhatt, U., Chandra, K., Wong, L., Lee, M., Zhang, C. E., Zhi-Xuan, T., Ho, M., Mansinghka, V., et al. (2024). Building machines that learn and think with people. Nature human behaviour , 8(10):1851--1863

    2024

  10. [18]

    M., Todd, G., Wong, L., Zhang, C., Togelius, J., Weller, A., Chu, J., Griffiths, T., and Tenenbaum, J

    Collins, K. M., Todd, G., Wong, L., Zhang, C., Togelius, J., Weller, A., Chu, J., Griffiths, T., and Tenenbaum, J. (2025). Generation and evaluation in the human invention process through the lens of game design. In Proceedings of the Annual Meeting of the Cognitive Science So...

    2025

  11. [19]

    and Demiris, Y

    Cully, A. and Demiris, Y. (2017). Quality and diversity optimization: A unifying modular framework. IEEE Transactions on Evolutionary Computation , 22(2):245--259

    2017

  12. [20]

    B., Solar-Lezama, A., and Tavares, Z

    Das, R., Tenenbaum, J. B., Solar-Lezama, A., and Tavares, Z. (2023). Combining functional and automata synthesis to discover causal reactive programs. Proceedings of the ACM on Programming Languages , 7(POPL):1628--1658

    2023

  13. [21]

    M., and Lake, B

    Davidson, G., Todd, G., Togelius, J., Gureckis, T. M., and Lake, B. M. (2024). Goals as reward-producing programs. arXiv preprint arXiv:2405.13242

    2024 arXiv

  14. [22]

    Ding, J., Zhang, Y., Shang, Y., Zhang, Y., Zong, Z., Feng, J., Yuan, Y., Su, H., Li, N., Sukiennik, N., et al. (2024). Understanding world or predicting future? a comprehensive survey of world models. ACM Computing Surveys

    2024

  15. [23]

    L., and Efros, A

    Dubey, R., Agrawal, P., Pathak, D., Griffiths, T. L., and Efros, A. A. (2018). Investigating human priors for playing video games

    2018

  16. [24]

    Gelman, S. A. and Legare, C. H. (2011). Concepts and folk theories. Annual review of anthropology , 40(1):379--398

    2011

  17. [25]

    Geng, J., Chen, H., Arumugam, D., and Griffiths, T. L. (2025). Are large language models reliable ai scientists? assessing reverse-engineering of black-box systems

    2025

  18. [26]

    and Tenenbaum, J

    Gerstenberg, T. and Tenenbaum, J. B. (2017). Intuitive theories

    2017

  19. [27]

    E., Chang, B., Mitchener, L., Yiu, A., Szostkiewicz, C

    Ghareeb, A. E., Chang, B., Mitchener, L., Yiu, A., Szostkiewicz, C. J., Laurent, J. M., Razzak, M. T., White, A. D., Hinks, M. M., and Rodriques, S. G. (2025). Robin: A multi-agent system for automating scientific discovery

    2025

  20. [28]

    Gopnik, A. (1996). The scientist as child. Philosophy of science , 63(4):485--514

    1996

  21. [29]

    and Wellman, H

    Gopnik, A. and Wellman, H. M. (1992). Why the child's theory of mind really is a theory

    1992

  22. [30]

    and Wellman, H

    Gopnik, A. and Wellman, H. M. (2012). Reconstructing constructivism: causal models, bayesian learning mechanisms, and the theory theory. Psychological bulletin , 138(6):1085

    2012

  23. [31]

    Guertler, L., Cheng, B., Yu, S., Liu, B., Choshen, L., and Tan, C. (2025). Textarena. arXiv preprint arXiv:2504.11442

    2025 arXiv

  24. [32]

    and Schmidhuber, J

    Ha, D. and Schmidhuber, J. (2018). World models. arXiv preprint arXiv:1803.10122

    2018 arXiv

  25. [33]

    J., Wang, Z., Wang, D

    Hao, S., Gu, Y., Ma, H., Hong, J. J., Wang, Z., Wang, D. Z., and Hu, Z. (2023). Reasoning with language model is planning with world model. arXiv preprint arXiv:2305.14992

    2023 arXiv

  26. [34]

    D., Tenenbaum, J

    Henderson, L., Goodman, N. D., Tenenbaum, J. B., and Woodward, J. F. (2010). The structure and dynamics of scientific theories: A hierarchical bayesian perspective. Philosophy of Science , 77(2):172--200

    2010

  27. [35]

    Johnson-Laird, P. N. (1983). Mental models: Towards a cognitive science of language, inference, and consciousness . Number 6. Harvard University Press

    1983

  28. [36]

    P., Littman, M

    Kaelbling, L. P., Littman, M. L., and Cassandra, A. R. (1998). Planning and acting in partially observable stochastic domains. Artificial intelligence , 101(1-2):99--134

    1998

  29. [37]

    H., Czechowski, K., Erhan, D., Finn, C., Kozakowski, P., Levine, S., et al

    Kaiser, L., Babaeizadeh, M., Milos, P., Osinski, B., Campbell, R. H., Czechowski, K., Erhan, D., Finn, C., Kozakowski, P., Levine, S., et al. (2019). Model-based reinforcement learning for atari. arXiv preprint arXiv:1903.00374

    2019 arXiv

  30. [38]

    M., Ullman, T

    Lake, B. M., Ullman, T. D., Tenenbaum, J. B., and Gershman, S. J. (2017). Building machines that learn and think like people. Behavioral and brain sciences , 40:e253

    2017

  31. [39]

    LeCun, Y. (2022). A path towards autonomous machine intelligence version 0.9. 2, 2022-06-27. Open Review , 62(1):1--62

    2022

  32. [40]

    K., Bau, D., Vi \'e gas, F., Pfister, H., and Wattenberg, M

    Li, K., Hopkins, A. K., Bau, D., Vi \'e gas, F., Pfister, H., and Wattenberg, M. (2023). Emergent world representations: Exploring a sequence model trained on a synthetic task. ICLR

    2023

  33. [41]

    and Xu, F

    Liu, R. and Xu, F. (2024). Violations of core object principles change adults’ behaviors in maze games. In Proceedings of the Annual Meeting of the Cognitive Science Society , volume 46

    2024

  34. [42]

    M., Broekens, J., Plaat, A., Jonker, C

    Moerland, T. M., Broekens, J., Plaat, A., Jonker, C. M., et al. (2023). Model-based reinforcement learning: A survey. Foundations and Trends in Machine Learning , 16(1):1--118

    2023

  35. [43]

    Newell, A. (1955). The chess machine: an example of dealing with a complex task by adaptation. In Proceedings of the March 1-3, 1955, western joint computer conference , pages 101--108

    1955

  36. [44]

    A., et al

    Newell, A., Simon, H. A., et al. (1972). Human problem solving , volume 104. Prentice-hall Englewood Cliffs, NJ

    1972

  37. [45]

    D., Togelius, J., and Lucas, S

    Perez-Liebana, D., Liu, J., Khalifa, A., Gaina, R. D., Togelius, J., and Lucas, S. M. (2019). General video game ai: A multitrack framework for evaluating agents, games, and content generation algorithms. IEEE Transactions on Games , 11(3):195--214

    2019

  38. [46]

    Plappert, M., Andrychowicz, M., Ray, A., McGrew, B., Baker, B., Powell, G., Schneider, J., Tobin, J., Chociej, M., Welinder, P., et al. (2018). Multi-goal reinforcement learning: Challenging robotics environments and request for research. arXiv preprint arXiv:1802.09464

    2018 arXiv

  39. [47]

    Pourcel, J., Colas, C., Molinaro, G., Oudeyer, P.-Y., and Teodorescu, L. (2023). Aces: Generating diverse programming puzzles with with autotelic generative models. arXiv preprint arXiv:2310.10692

    2023

  40. [48]

    K., Soros, L

    Pugh, J. K., Soros, L. B., and Stanley, K. O. (2016). Quality diversity: A new frontier for evolutionary computation. Frontiers in Robotics and AI , 3:40

    2016

  41. [49]

    Richens, J., Abel, D., Bellot, A., and Everitt, T. (2025). General agents need world models. arXiv preprint arXiv:2506.01622

    2025

  42. [50]

    Schaul, T. (2013). A video game description language for model-based or interactive learning. In 2013 IEEE Conference on Computational Inteligence in Games (CIG) , pages 1--8. IEEE

    2013

  43. [51]

    Schulz, L. (2012). The origins of inquiry: Inductive inference and exploration in early childhood. Trends in cognitive sciences , 16(7):382--389

    2012

  44. [52]

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

    1950

  45. [53]

    J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al

    Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al. (2016). Mastering the game of go with deep neural networks and tree search. nature , 529(7587):484--489

    2016

  46. [54]

    Spelke, E. S. and Kinzler, K. D. (2007). Core knowledge. Developmental science , 10(1):89--96

    2007

  47. [55]

    B., Griffiths, T

    Tenenbaum, J. B., Griffiths, T. L., and Kemp, C. (2006). Theory-based bayesian models of inductive learning and reasoning. Trends in cognitive sciences , 10(7):309--318

    2006

  48. [56]

    B., Kemp, C., Griffiths, T

    Tenenbaum, J. B., Kemp, C., Griffiths, T. L., and Goodman, N. D. (2011). How to grow a mind: Statistics, structure, and abstraction. science , 331(6022):1279--1285

    2011

  49. [57]

    G., Stephenson, M., Piette, \'E ., Soemers, D

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

    2024

  50. [58]

    A., Loula, J., Burga, J., Foss, N., Campero, A., Pouncy, T., Gershman, S

    Tsividis, P. A., Loula, J., Burga, J., Foss, N., Campero, A., Pouncy, T., Gershman, S. J., and Tenenbaum, J. B. (2021). Human-level reinforcement learning through theory-based modeling, exploration, and planning. arXiv preprint arXiv:2107.12544

    2021 arXiv

  51. [59]

    Ullman, T. D. and Tenenbaum, J. B. (2020). Bayesian models of conceptual development: Learning as building models of the world. Annual Review of Developmental Psychology , 2(1):533--558

    2020

  52. [60]

    Vafa, K., Chen, J., Rambachan, A., Kleinberg, J., and Mullainathan, S. (2024). Evaluating the world model implicit in a generative model. Advances in Neural Information Processing Systems , 37:26941--26975

    2024

  53. [61]

    Van Hasselt, H., Guez, A., and Silver, D. (2016). Deep reinforcement learning with double q-learning. In Proceedings of the AAAI conference on artificial intelligence , volume 30

    2016

  54. [62]

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

    2023

  55. [63]

    M., Gershman, S

    V \'e lez, N., Wu, C. M., Gershman, S. J., and Schulz, E. (2024). The rise and fall of technological development in virtual communities

    2024

  56. [64]

    Verma, V., Huang, D., Chen, W., Klein, D., and Tomlin, N. (2025). Measuring general intelligence with generated games. arXiv preprint arXiv:2505.07215

    2025 arXiv

  57. [65]

    M., Mathieu, M., Dudzik, A., Chung, J., Choi, D

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

    2019

  58. [66]

    Wang, H., Fu, T., Du, Y., Gao, W., Huang, K., Liu, Z., Chandak, P., Liu, S., Van Katwyk, P., Deac, A., et al. (2023). Scientific discovery in the age of artificial intelligence. Nature , 620(7972):47--60

    2023

  59. [67]

    Yannakakis, G. N. and Togelius, J. (2018). Artificial intelligence and games , volume 2. Springer

    2018

  60. [68]

    M., and Lobo, J

    Youn, H., Strumsky, D., Bettencourt, L. M., and Lobo, J. (2015). Invention as a combinatorial process: evidence from us patents. Journal of the Royal Society interface , 12(106):20150272

    2015

  61. [69]

    Yu, T., Quillen, D., He, Z., Julian, R., Hausman, K., Finn, C., and Levine, S. (2020). Meta-world: A benchmark and evaluation for multi-task and meta reinforcement learning. In Conference on robot learning , pages 1094--1100. PMLR

    2020

  62. [70]

    E., Collins, K

    Zhang, C. E., Collins, K. M., Wong, L., Barba, M., Weller, A., and Tenenbaum, J. B. (2024). People use fast, goal-directed simulation to reason about novel games. arXiv preprint arXiv:2407.14095

    2024 arXiv

  63. [71]

    Zhao, B., V \'e lez, N., and Griffiths, T. (2024). A rational model of innovation by recombination. In Proceedings of the Annual Meeting of the Cognitive Science Society , volume 46

    2024

  64. [72]

    Zhu, Z., Wang, X., Zhao, W., Min, C., Deng, N., Dou, M., Wang, Y., Shi, B., Wang, K., Zhang, C., et al. (2024). Is sora a world simulator? a comprehensive survey on general world models and beyond. arXiv preprint arXiv:2405.03520

    2024

Pith tools

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