Pith. sign in

REVIEW 3 major objections 5 minor 47 references

Analogy making as amortised model construction

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 Analogy lets an agent build a new internal model by reusing solution-relevant structure from old ones, amortizing both model construction and planning.

desk verdict A clear workshop sketch that reframes analogy as partial MDP homomorphism with a module library; the synthesis is novel and honest, but the load-bearing amortization claim is not yet formally supported. read the letter →

arxiv 2507.16511 v1 pith:4ZP6M7Q7 submitted 2025-07-22 cs.LG cs.AI

David G. Nagy , Tingke Shen , Hanqi Zhou , Charley M. Wu , Peter Dayan This is my paper
classification cs.LGcs.AI
keywords analogyMarkovdecisionprocessesmodelconstructionamortizedinferencepartialhomomorphismsmodulelibraryreinforcementlearningcognitivemodeling
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

The paper tries to establish that flexible on-demand construction of internal models is made tractable by analogy. It formalizes an analogy as a partial homomorphism between Markov decision processes: a mapping of states and actions that preserves transition and reward structure over a relevant subset of the state-action space, but not necessarily everywhere. Past construals are decomposed into abstract modules that can be recombined to form new models, and each module carries a partial policy that can be transferred through the mapping. If this is right, the same machinery explains how a child understands an email password as a key and offers a design principle for reinforcement learning agents that generalize to novel situations without rebuilding their world model from scratch.

What carries the argument

The central object is the partial (S)MDP homomorphism, a structure-preserving mapping $\phi=(f,g)$ from one decision process to another that commutes with transition dynamics and preserves rewards, but only on a relevant subset of states and actions rather than globally. The paper uses this mapping as the formal backbone for analogy: it lets a source module act as an inductive bias for building the target construal, it allows policies and action abstractions to be transported across domains, and it supports composition because different modules can map to different pieces of a new situation. The companion mechanism is the module library: reusable fragments of state and action structure and associated partial policies, refined over time as the same abstraction is applied in increasingly varied contexts.

What would settle it

Run a controlled MDP experiment in which a source module shares solution-relevant transition and reward structure with a novel target while surface features differ; the framework predicts that agents with access to the module and a partial mapping should solve the target with measurably lower construction and planning cost than agents replanning from scratch, and that this advantage disappears when the shared structure is chosen outside the relevant subset. If the transferred policy is no better than a from-scratch baseline in the exact-homomorphism condition, the central claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that analogy making is amortized model construction. Faced with a novel situation, an agent chooses a construal $M$ that maximizes expected real-world reward under a policy that solves $M$, subject to the total cost of constructing and solving $M$ falling below a capacity limit. The reuse mechanism is a partial (S)MDP homomorphism $\phi=(f,g)$ from a source module to the target construal, preserving rewards and transition dynamics only over a relevant subset of state-action space; when the shared part is exact, an optimal policy for the source maps to an optimal policy for the target, and approximate matches have bounded losses. Because modules encode not only state and action abstractions but also partial solutions, composing them transfers planning effort as well as representation. The library of modules is grown by extracting consistently useful fragments from past construals, progressively abstracting them into reusable conceptual primitives.

Load-bearing premise

The load-bearing premise is that what makes an analogy useful is a mapping that preserves only the solution-relevant structure between old and new situations, and that an agent can find that relevant subset without already having solved the problem.

Editorial extensions

If this is right

  • An exact partial homomorphism guarantees that a policy optimal for the source module is optimal for the target construal, so shared abstract structure carries the solution, not just the representation.
  • When the homomorphism is only approximate, the performance loss from transferred policies is bounded through bounded-parameter MDPs, giving a formal sense in which good-enough analogies still help.
  • Composition of modules enables piecewise construal: a new office door can be handled by an apartment-door module while a projector is handled by a school-projector module, so unfamiliar situations can be assembled from familiar fragments.
  • Maintaining a library of abstract modules is an upfront investment that lowers the combined cost of construction and solution for future situations; as modules are reused in more contexts they become more abstract, eventually becoming conceptual primitives like door or key.
  • Mapping parts of the environment onto modules offers a functional account of affordance perception: perceiving what a situation affords is recognizing which module applies to it.

Reading between the lines

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

  • Editorial inference: if this account is right, supplying a partial analogy should measurably cut model-construction cost only when the hinted mapping lands on a solution-relevant subset; a task where the hint targets surface-similar but irrelevant structure should show no such saving.
  • Editorial inference: the account implies that human one-shot transfer errors are systematic—agents should reuse a module when its relevant subset matches despite surface mismatch, and should fail to transfer when surface features match but relevant structure does not; this could be tested with synthetic MDPs and compared with human choices.
  • Editorial inference: because the framework leaves the choice of relevant subset open, a natural extension is to learn that selection itself as an amortized inference problem over the same library, which would convert the frame problem into an inductive-bias learning problem rather than a search over all details.
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 argues that the on-demand construction of internal MDPs—termed construals—can be amortized through analogy. Analogies are formalized as partial (S)MDP homomorphisms that preserve solution-relevant structure over a relevant subset of state-action space. The authors introduce an objective (Eq. 4) in which an agent chooses a model M maximizing expected real-world reward subject to constraints on both the solution cost Cs and the construction cost Cc. They propose that a library L of abstract, composable modules, derived from past construals, supports flexible recombination and policy transfer. The manuscript develops this framework through the door–key/email–password example, a module-composition example (chair on pedestal), and a three-stage computational decomposition (construal, solution, library refinement). It is explicitly labeled a sketch, with the Discussion listing open questions and conceding that much remains to be formalized. The paper contains no empirical evaluation and no new formal theorems; the main cited formal result is the optimal-policy transfer theorem of Ravindran and Barto for exact homomorphisms.

Significance. If the proposed framework were made precise, it would connect a substantial body of cognitive-science work on analogy and mental models to formal abstraction theory in RL, and it would articulate a plausible mechanism by which resource-limited agents could construct tractable internal models of novel situations. The paper has several strengths: it is grounded in prior formal work (Ravindran and Barto, 2002; 2003), it makes the resource-constrained construal objective explicit in Eq. 4, it connects to value-guided construals (Ho et al., 2022) and to library-learning methods such as DreamCoder and fragment grammars, and it is honest about its limitations. The library-learning decomposition into construal, solution, and library-refinement stages is a useful organizing device. However, the central claims about amortization and compositional transfer are not yet supported by formal guarantees or by any computational demonstration; the partial-homomorphism transfer theorem is cited from prior work that applies to exact homomorphisms, and the paper's own examples concede that composed modules need not transfer smoothly.

major comments (3)
  1. [Section 3.1, 'Mapping solutions through analogies'] The paper asserts that when part of the abstract structure is shared exactly, an optimal policy for the source MDP maps to an optimal policy for the target, citing Ravindran and Barto (2002). But the paper's own definition of an analogy is a partial MDP homomorphism, for which the preservation conditions hold only over a 'relevant subset' of state-action space (Section 3.1, paragraph 2). The cited theorem applies to exact homomorphisms over the full MDP. For a partial homomorphism, policy transfer is guaranteed only if the relevant subset contains every state-action pair reachable under the transferred policy and the target dynamics. The manuscript neither states this coverage condition nor proves a bounded-suboptimality version. Without such a condition, the amortization claim in Eq. 4—that reusing a source policy reduces planning cost while preserving adequate outcomes—is unsupported. Please provide a precise statement of the transfer guarantee for partial homomorphisms, including the required coverage assumption, or explicitly limit the claim to the exact-homomorphism case.
  2. [Section 3.2, 'Composing modules and adapting solutions'] The composition mechanism is presented as central to flexible reuse, but the paper explicitly concedes that 'there is no guarantee that these solutions will transfer smoothly to new configurations' and illustrates the failure with the chair-on-pedestal example. This concession is in tension with the claim that modular reuse supports the amortized objective in Eq. 4. Composition of partial homomorphisms is not automatically a homomorphism: even if each module preserves structure on its local fragment, interactions between fragments—such as the pedestal changing the affordance of the chair—can alter transition and reward structure so that the composed mapping fails to preserve solution-relevant structure. The manuscript does not define a composition operation on modules, nor does it give conditions under which the composition of two partial homomorphisms is again a partial homomorphism with a bounded loss of optimality. Please add a formal treatment of module composition, or reframe the composition section as an open problem and weaken the corresponding claim about amortized construction.
  3. [Section 2, Eq. 4] Eq. 4 is the formal centerpiece of the paper, but the terms Cc and Cs are left abstract, and no algorithm or complexity model is given for constructing M* or for selecting the relevant subset on which a partial homomorphism is defined. The stated goal is amortization—reducing future construction costs through reuse—yet the paper does not show, even in a simple worked example, that reuse of a module or analogy lowers Cc relative to constructing a model from scratch. As written, Eq. 4 restates the problem rather than providing a solution. Since the central claim is that analogy provides a tractable way to optimize this objective, the paper needs at least one of the following: a formal cost model with a demonstrated amortization effect, a small computational illustration, or an explicit statement that the tractability question remains open and is not addressed by the framework.
minor comments (5)
  1. [Section 2, Eq. 3 and Eq. 4] The notation R' and R'' is introduced without definition; please clarify whether these denote the same real-world reward function in both equations or different task-specific reward functions.
  2. [Section 3.1, paragraph 2] The phrase 'relevant subset of the state-action space' is used frequently but never defined formally. Please specify whether the subset is a subset of S × A, whether it depends on the source or target MDP, and how it could in principle be selected.
  3. [Section 3.1, 'Mapping solutions through analogies'] The reference to Bounded-parameter MDPs is made in one sentence with no definition or citation details; please provide a concrete statement of how approximate homomorphisms relate to bounded-parameter MDPs, or remove the claim.
  4. [Supplementary Materials, Figure S1] The main text refers to Figure S1 as illustrating library building as Bayesian inference, but the figure is only mentioned in the supplementary materials and is not described in the main text. Please either integrate a brief description or remove the reference.
  5. [Section 4, Discussion] The paper concludes with a strong statement that 'if modular reuse through analogy underlies the human ability...' but no testable predictions are offered. Adding even one distinguishing prediction—for example, about transfer failures when a relevant subset excludes reachable states—would strengthen the paper's scientific contribution.

Circularity Check

1 steps flagged · score 3.0 of 10

The framework's central claim is partly self-definitional: 'analogy' is defined as preserving solution-relevant structure, but relevance is not specified independently of the solution it is meant to explain, so the amortisation benefit is built into the definition.

  1. self definitional [Section 1 (Introduction) and Section 3.1, 'Making a single analogy']
    "Specifically, we formalise potential analogies as mappings between MDPs that preserve solution-relevant structure (and in some cases the entire solution). ... 'Crucially, analogies typically involve partial mappings ..., preserving only aspects of structure deemed relevant in the current context. We refer to φ = (f (s), gs(a)) as a partial MDP homomorphism when the conditions are only required to hold over a relevant subset of the state-action space rather than globally over the entire MDP.'"

    The paper's central explanatory claim is that analogy enables amortised model construction by preserving solution-relevant structure. But the formal definition of an analogy is a mapping that preserves structure on a 'relevant subset', and the paper gives no independent, a priori specification of what makes a subset relevant. Relevance can only be read as 'whatever needs to be preserved for the solution to transfer'. Any successful transfer can then be declared to have preserved the relevant structure, and any failed mapping can be excluded as irrelevant. The amortisation benefit is therefore contained in the definition of analogy rather than derived from the formal framework.

full rationale

This is a conceptual workshop sketch with no numerical fits or empirical predictions, so the fitted-input-called-prediction and severe self-citation modes do not apply. The formal backbone is anchored in external prior work (Ravindran & Barto, 2002; 2003), and the paper does not claim their exact-homomorphism optimality theorem extends to its partial homomorphisms. The composition section explicitly concedes that 'there is no guarantee that these solutions will transfer smoothly to new configurations', and the discussion admits that the search problem is open, so those parts are stated limitations rather than hidden circular moves. The one circularity-adjacent defect is definitional: the central predicate 'solution-relevant structure' is never characterized independently of the solution or outcome, making the claim that analogies preserve solution-relevant structure and hence amortise construction a tautology at this stage. I therefore assign a score of 3 rather than 0-2; this is a partially self-definitional framework, not a circular numerical derivation.

Assumptions & free parameters 0 free parameters · 5 assumptions · 2 invented entities

The formal apparatus rests on MDP homomorphism theory from previous literature; the paper's own additions are the library and the partiality criterion. No free parameters are fitted. The main unmeasured assumptions are the existence of a module library and the computability of the relevant subset for partial homomorphisms.

assumptions (5)
  • domain assumption MDPs, including POMDPs, BAMDPs, and IPOMDPs, are an adequate formalism for human internal models used in planning.
    Footnote 1 states that MDP is used as a shorthand for a family of Markov decision process based models; the framework assumes these capture the relevant structure of human planning.
  • domain assumption The agent faces the on-demand construal objective in Eq. 4, with separable costs Cs(M) and Cc(M,W) and a single capacity limit Cmax.
    Section 2 introduces this as the formal challenge; the decomposition is convenient but not derived or measured.
  • ad hoc to paper Useful analogies can be formalized as partial (S)MDP homomorphisms preserving solution-relevant structure over a relevant subset of states and actions.
    Section 3.1 defines an analogy as a partial homomorphism; this is the central modeling move and is assumed, not derived.
  • ad hoc to paper Agents possess a library L of composable abstract modules and alternate between construal, solution, and library refinement stages.
    Section 3.2 and Figure S1 posit the library and the three-stage decomposition; it is presented as a proposal supported by related work in program induction, not by evidence in this paper.
  • standard math For exact MDP homomorphisms, an optimal policy maps to an optimal policy in the target MDP.
    The paper relies on this external theorem from Ravindran & Barto (2002) for solution transfer under analogy; it is cited but not reproved.
invented entities (2)
  • Library of abstract MDP modules
    purpose: Reusable fragments that compose into new construals and carry transferred policies
    A postulated cognitive and algorithmic construct; no data or falsifiable prediction in the paper establishes its existence. Related constructs appear in cited program induction work.
  • Partial MDP homomorphism as an analogy
    purpose: Formal vehicle for mapping source construals to target construals while preserving solution-relevant structure
    The mapping is the paper's central formalization; it is an analytical tool rather than an empirical entity, and the paper provides no independent testable constraints.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Analogy making as amortised model construction." pith.science (2026). https://pith.science/paper/4ZP6M7Q7

@misc{pith2026250716511,
  author       = {Pith},
  title        = {Pith review of: Analogy making as amortised model construction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4ZP6M7Q7}},
  note         = {Machine review of arXiv:2507.16511}
}
read the original abstract

Humans flexibly construct internal models to navigate novel situations. To be useful, these internal models must be sufficiently faithful to the environment that resource-limited planning leads to adequate outcomes; equally, they must be tractable to construct in the first place. We argue that analogy plays a central role in these processes, enabling agents to reuse solution-relevant structure from past experiences and amortise the computational costs of both model construction (construal) and planning. Formalising analogies as partial homomorphisms between Markov decision processes, we sketch a framework in which abstract modules, derived from previous construals, serve as composable building blocks for new ones. This modular reuse allows for flexible adaptation of policies and representations across domains with shared structural essence.

Figures

Figures reproduced from arXiv: 2507.16511 by the authors.

Figure 1
Figure 1. Library of abstract MDP modules for amortised model construal. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The design process of certain real world artefacts (e.g. vacuum-cleaning robots) resembles [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Analogy-guided model construction. a) The child draws an analogy [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Modular construction of model, and transfer of solutions through the analogy. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 39 canonical work pages

  1. [1]

    A theory of abstraction in reinforcement learning

    David Abel. A theory of abstraction in reinforcement learning. CoRR, 2022

    work page 2022

  2. [2]

    Synthesizing world models for bilevel planning

    Zergham Ahmed, Joshua B. Tenenbaum, Christopher J. Bates, and Samuel J. Gershman. Synthesizing world models for bilevel planning, March 2025. arXiv:2503.20124 [cs]

    work page Pith review arXiv 2025

  3. [3]

    Ankesh Anand, Jacob Walker, Yazhe Li, Eszter Vértes, Julian Schrittwieser, Sherjil Ozair, Théophane Weber, and Jessica B. Hamrick. Procedural Generalization by Planning with Self - Supervised World Models , November 2021. arXiv:2111.01587 [cs]

    arXiv 2021

  4. [4]

    Jacob J. W. Bakermans, Joseph Warren, James C. R. Whittington, and Timothy E. J. Behrens. Constructing future behavior in the hippocampal formation through composition and replay. Nature Neuroscience, pp.\ 1--12, March 2025. ISSN 1546-1726. Publisher: Nature Publishing Group

    work page 2025

  5. [5]

    Structured agents for physical construction

    Victor Bapst, Alvaro Sanchez-Gonzalez, Carl Doersch, Kimberly Stachenfeld, Pushmeet Kohli, Peter Battaglia, and Jessica Hamrick. Structured agents for physical construction. In Proceedings of the 36th International Conference on Machine Learning , pp.\ 464--474. PMLR, May 2019. ISSN: 2640-3498

    work page 2019

  6. [6]

    Del rigor en la ciencia

    Jorge Luis Borges. Del rigor en la ciencia. El hacedor, 45, 1960

    work page 1960

  7. [7]

    The career of metaphor

    Brian F Bowdle and Dedre Gentner. The career of metaphor. Psychological review, 112 0 (1): 0 193, 2005

    work page 2005

  8. [8]

    Olausson, Lionel Wong, Gabriel Grand, Joshua B

    Matthew Bowers, Theo X. Olausson, Lionel Wong, Gabriel Grand, Joshua B. Tenenbaum, Kevin Ellis, and Armando Solar-Lezama. Top- Down Synthesis for Library Learning . Proceedings of the ACM on Programming Languages, 7 0 (POPL): 0 1182--1213, January 2023. ISSN 2475-1421

    work page 2023

Show all 47 references
  1. [9]

    u hler, Sarah Schw \

    Martin V Butz, Maximilian Mittenb \"u hler, Sarah Schw \"o bel, Asya Achimova, Christian Gumbsch, Sebastian Otte, and Stefan Kiebel. Contextualizing predictive minds. Neuroscience & Biobehavioral Reviews, 168: 0 105948, 2025

    2025

  2. [10]

    Gershman

    Ishita Dasgupta and Samuel J. Gershman. Memory as a Computational Resource . Trends in Cognitive Sciences, 25 0 (3): 0 240--251, March 2021. ISSN 1364-6613

    2021

  3. [11]

    Cognitive Wheels : the Frame Problem of AI

    D Dennett. Cognitive Wheels : the Frame Problem of AI . The philosophy of artificial intelligence, 1990

    1990

  4. [12]

    Carlos Diuk, Andre Cohen, and Michael L. Littman. An object-oriented representation for efficient reinforcement learning. In Proceedings of the 25th international conference on Machine learning - ICML '08 , pp.\ 240--247, Helsinki, Finland, 2008. ACM Press. ISBN 978-1-60558-205-4

    2008

  5. [13]

    Tenenbaum

    Kevin Ellis, Lionel Wong, Maxwell Nye, Mathias Sablé-Meyer, Luc Cary, Lore Anaya Pozo, Luke Hewitt, Armando Solar-Lezama, and Joshua B. Tenenbaum. DreamCoder : growing generalizable, interpretable knowledge with wake–sleep Bayesian program learning. Philosophical Transactions ...

    2023

  6. [14]

    Mental spaces: Aspects of meaning construction in natural language

    Gilles Fauconnier. Mental spaces: Aspects of meaning construction in natural language. Cambridge University Press, 1994

    1994

  7. [15]

    Structure-mapping: A theoretical framework for analogy

    Dedre Gentner. Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7 0 (2): 0 155--170, April 1983. ISSN 0364-0213

    1983

  8. [16]

    Analogy and Abstraction

    Dedre Gentner and Christian Hoyos. Analogy and Abstraction . Topics in Cognitive Science, 9 0 (3): 0 672--693, 2017. ISSN 1756-8765. \_eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1111/tops.12278

    2017 doi

  9. [17]

    Gershman and Noah D

    S. Gershman and Noah D. Goodman. Amortized Inference in Probabilistic Reasoning . Cognitive Science, 2014

    2014

  10. [18]

    James J. Gibson. The Theory of Affordances : (1979). In The People , Place , and Space Reader . Routledge, 2014. ISBN 978-1-315-81685-2. Num Pages: 5

    1979

  11. [19]

    A theory of abstraction

    Fausto Giunchiglia and Toby Walsh. A theory of abstraction. Artificial Intelligence, 57 0 (2): 0 323--389, October 1992. ISSN 0004-3702

    1992

  12. [20]

    Generalizing plans to new environments in relational MDPs

    Carlos Guestrin, Daphne Koller, Chris Gearhart, and Neal Kanodia. Generalizing plans to new environments in relational MDPs . In Proceedings of the 18th international joint conference on Artificial intelligence , IJCAI '03, pp.\ 1003--1010, San Francisco, CA, USA, August 2003....

    2003

  13. [21]

    Mastering diverse control tasks through world models

    Danijar Hafner, Jurgis Pasukonis, Jimmy Ba, and Timothy Lillicrap. Mastering diverse control tasks through world models. Nature, pp.\ 1--7, 2025

    2025

  14. [22]

    Algebraic structure theory of sequential machines (prentice-hall international series in applied mathematics)

    Juris Hartmanis. Algebraic structure theory of sequential machines (prentice-hall international series in applied mathematics). Prentice-Hall, Inc., 1966

    1966

  15. [23]

    Ho, David Abel, Carlos G

    Mark K. Ho, David Abel, Carlos G. Correa, Michael L. Littman, Jonathan D. Cohen, and Thomas L. Griffiths. People construct simplified mental representations to plan. Nature, 606 0 (7912): 0 129--136, June 2022. ISSN 1476-4687. Number: 7912 Publisher: Nature Publishing Group

    2022

  16. [24]

    Hofstadter and Emmanuel Sander

    Douglas R. Hofstadter and Emmanuel Sander. Surfaces and Essences : Analogy as the Fuel and Fire of Thinking . Basic Books, April 2013. ISBN 978-0-465-02158-1. Google-Books-ID: XkQT5eTnurYC

    2013

  17. [25]

    Interplay of approximate planning strategies

    Quentin JM Huys, N \' all Lally, Paul Faulkner, Neir Eshel, Erich Seifritz, Samuel J Gershman, Peter Dayan, and Jonathan P Roiser. Interplay of approximate planning strategies. Proceedings of the National Academy of Sciences, 112 0 (10): 0 3098--3103, 2015

    2015

  18. [26]

    A Resource - Rational Approach to the Causal Frame Problem

    Thomas F Icard and Noah Goodman. A Resource - Rational Approach to the Causal Frame Problem . Proceedings of the 37th Annual Meeting of the Cognitive Science Society, pp.\ 6, 2015

    2015

  19. [27]

    Off-line replay maintains declarative memories in a model of hippocampal-neocortical interactions

    Szabolcs K \'a li and Peter Dayan. Off-line replay maintains declarative memories in a model of hippocampal-neocortical interactions. Nature neuroscience, 7 0 (3): 0 286--294, 2004

    2004

  20. [28]

    Metaphors We Live By

    George Lakoff and Mark Johnson. Metaphors We Live By . University of Chicago Press, December 2008. ISBN 978-0-226-47099-3. Google-Books-ID: r6nOYYtxzUoC

    2008

  21. [29]

    Hippocampal contributions to control: the third way

    M \'a t \'e Lengyel and Peter Dayan. Hippocampal contributions to control: the third way. Advances in neural information processing systems, 20, 2007

    2007

  22. [30]

    Nagy, Gerg o Orb \'a n, and Charley M

    David G. Nagy, Gerg o Orb \'a n, and Charley M. Wu. Adaptive compression as a unifying framework for episodic and semantic memory. Nature Reviews Psychology, pp.\ 1--15, June 2025. ISSN 2731-0574

    2025

  23. [31]

    Fragment Grammars : Exploring Computation and Reuse in Language

    Timothy J O’Donnell, Noah D Goodman, and Joshua B Tenenbaum. Fragment Grammars : Exploring Computation and Reuse in Language . 2009

    2009

  24. [32]

    Gershman

    Thomas Pouncy and Samuel J. Gershman. Inductive biases in theory-based reinforcement learning. Cognitive Psychology, 138: 0 101509, November 2022. ISSN 00100285

    2022

  25. [33]

    Gershman

    Thomas Pouncy, Pedro Tsividis, and Samuel J. Gershman. What Is the Model in Model ‐ Based Planning ? Cognitive Science, 45 0 (1), January 2021. ISSN 0364-0213, 1551-6709. doi:10.1111/cogs.12928. URL https://onlinelibrary.wiley.com/doi/10.1111/cogs.12928

    2021 doi

  26. [34]

    Balaraman Ravindran and Andrew G. Barto. Model Minimization in Hierarchical Reinforcement Learning . In Sven Koenig and Robert C. Holte (eds.), Abstraction, Reformulation , and Approximation , pp.\ 196--211, Berlin, Heidelberg, 2002. Springer. ISBN 978-3-540-45622-3

    2002

  27. [35]

    Balaraman Ravindran and Andrew G. Barto. SMDP homomorphisms: an algebraic approach to abstraction in semi- Markov decision processes. In Proceedings of the 18th international joint conference on Artificial intelligence , IJCAI '03, pp.\ 1011--1016, San Francisco, CA, USA, Augu...

    2003

  28. [36]

    Compositionality under time pressure

    Valerio Rubino, Mani Hamidi, Peter Dayan, and Charley M Wu. Compositionality under time pressure. In M. Goldwater, F. Anggoro, B. Hayes, and D. Ong (eds.), Proceedings of the 45th Annual Conference of the Cognitive Science Society , Sydney, Australia , 2023. Cognitive Science Society

    2023

  29. [37]

    The child as hacker

    Joshua S Rule, Joshua B Tenenbaum, and Steven T Piantadosi. The child as hacker. Trends in cognitive sciences, 24 0 (11): 0 900--915, 2020

    2020

  30. [38]

    Mastering atari, go, chess and shogi by planning with a learned model

    Julian Schrittwieser, Ioannis Antonoglou, Thomas Hubert, Karen Simonyan, Laurent Sifre, Simon Schmitt, Arthur Guez, Edward Lockhart, Demis Hassabis, Thore Graepel, et al. Mastering atari, go, chess and shogi by planning with a learned model. Nature, 588 0 (7839): 0 604--609, 2020

    2020

  31. [39]

    Human-level reinforcement learning through theory-based modeling, exploration, and planning

    Pedro A Tsividis, Joao Loula, Jake Burga, Nathan Foss, Andres Campero, Thomas Pouncy, Samuel J Gershman, and Joshua B Tenenbaum. Human-level reinforcement learning through theory-based modeling, exploration, and planning. arXiv preprint arXiv:2107.12544, 2021

    2021 arXiv

  32. [40]

    Wu, and Fiery A

    Natalia Vélez, Charley M. Wu, and Fiery A. Cushman. Representational exchange in social learning: Blurring the lines between the ritual and instrumental. Behavioral and Brain Sciences, 45: 0 e271, 2022

    2022

  33. [41]

    Lew, Noah D

    Lionel Wong, Gabriel Grand, Alexander K. Lew, Noah D. Goodman, Vikash K. Mansinghka, Jacob Andreas, and Joshua B. Tenenbaum. From Word Models to World Models : Translating from Natural Language to the Probabilistic Language of Thought , June 2023 a . arXiv:2306.12672 [cs]

    2023 arXiv

  34. [42]

    Siegel, Jiahai Feng, Noa Korneev, Joshua B

    Lionel Wong, Jiayuan Mao, Pratyusha Sharma, Zachary S. Siegel, Jiahai Feng, Noa Korneev, Joshua B. Tenenbaum, and Jacob Andreas. Learning adaptive planning representations with natural language guidance, December 2023 b . arXiv:2312.08566

    2023 arXiv

  35. [43]

    Representational exchange in human social learning

    Charley M Wu, Natalia V \'e lez, Fiery A Cushman, Irene Cogliati Dezza, Eric Schulz, and Charley M Wu. Representational exchange in human social learning. The Drive for Knowledge, pp.\ 169--192, 2022

    2022

  36. [44]

    Group coordination catalyzes individual and cultural intelligence

    Charley M Wu, Rick Dale, and Robert D Hawkins. Group coordination catalyzes individual and cultural intelligence. Open Mind, 8: 0 1037--1057, 2024

    2024

  37. [45]

    Lucas, and Neil R

    Bonan Zhao, Christopher G. Lucas, and Neil R. Bramley. A model of conceptual bootstrapping in human cognition. Nature Human Behaviour, pp.\ 1--12, October 2023. ISSN 2397-3374. Publisher: Nature Publishing Group

    2023

  38. [46]

    Nagy, and Charley M

    Hanqi Zhou, David G. Nagy, and Charley M. Wu. Harmonizing Program Induction with Rate - Distortion Theory , May 2024. arXiv:2405.05294 [cs, math, stat]

    2024 arXiv

  39. [47]

    Situation models in language comprehension and memory

    Rolf A Zwaan and Gabriel A Radvansky. Situation models in language comprehension and memory. Psychological bulletin, 123 0 (2): 0 162, 1998

    1998

Pith tools

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