Pith. sign in

REVIEW 4 major objections 5 minor 77 references

Identifying Informative Environments for Cognition Parameter Inference via Bayesian Experimental Design

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The authors argue that the choice of planning environment in cognitive experiments should be optimized as a Bayesian experimental design problem, and that an amortized approximation can rank environments almost as well as exact Monte Carlo

desk verdict A clean BED-for-cognitive-experiments framework whose central empirical claim—that the amortized surrogate recovers MC-BED rankings—is under-supported: it ties a trivial depth heuristic on a tiny design space and never tests a held-out environment. read the letter →

arxiv 2607.28894 v2 pith:Q3RI4X7R submitted 2026-07-30 cs.AI stat.ML

Manisha Dubey , Rimvydas Rubavicius , N. Siddharth , Subramanian Ramamoorthy This is my paper
classification cs.AIstat.ML
keywords BayesianexperimentaldesigncognitivemodelinginverseplanningamortizedinferenceMouselab-MDPexpectedinformationgainenvironmentparameter
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the design of cognitive planning experiments—specifically, the choice of task environment—can be treated as a Bayesian experimental design (BED) problem, with the environment itself as the design variable. The authors build an exact Monte Carlo benchmark and an amortized neural surrogate that approximates both posterior inference over latent cognitive parameters and the expected information gain of each candidate environment. On Mouselab-MDP trees, the amortized surrogate reproduces the environment rankings of exact Monte Carlo BED while cutting computation drastically. The paper further argues that no single environment is uniformly best: expected information gain, posterior recovery, and information efficiency favor different designs, and the optimal environment depends on the true latent cognitive state. If right, this gives experimenters a principled way to select tasks that maximize information about hidden planning parameters.

What carries the argument

The central object is the expected information gain (EIG), defined as the expected reduction in posterior entropy over latent cognitive parameters θ from observing trajectories D generated in environment ξ. The authors also introduce an amortized posterior surrogate q_φ(θ | D, ξ), a neural network trained by distilling exact Bayesian posteriors computed on synthetic datasets, together with an auxiliary 'KL head' that directly regresses the exact posterior-to-prior KL. This KL head carries the design-ranking argument: it lets ABED evaluate environments at the cost of one forward pass, and it proves more stable than deriving KL from the amortized posterior itself. The Mouselab-MDP paradigm sup

What would settle it

Run ABED on a design space where informativeness is not monotonic with tree depth—asymmetric trees, varying reward priors, or different click costs—and compare its KL-head ranking to exact MC-BED on held-out environments. If Spearman correlation drops below roughly 0.8 or design regret becomes positive, the amortized ranking claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that environment selection for Bayesian inverse planning can be cast as choosing the design ξ that maximizes expected information gain, EIG(ξ) = H(θ) − E[ H(θ | D, ξ) ], and that this objective can be evaluated both exactly via Monte Carlo (MC-BED) and approximately via an amortized surrogate (ABED). The surrogate learns a neural posterior q_φ(θ | D, ξ) distilled from exact posteriors, plus an auxiliary head that directly regresses the exact posterior-to-prior KL. On a benchmark of Mouselab-MDP trees, ABED reproduces the MC-BED environment ranking (Spearman ρ around 0.9–1.0) with a single forward pass per candidate, and reveals that no single environment is unifo

Load-bearing premise

The claim that ABED reproduces exact BED rankings rests on the premise that the KL head, trained on 240–960 synthetic datasets, remains well-calibrated on held-out environments—an assumption the paper itself flags as open, especially since rankings in the benchmark are coarse because informativeness rises almost monotonically with tree depth.

Editorial extensions

If this is right

  • Experimenters can rank candidate task environments by expected information gain before collecting data, using ABED's single forward pass instead of repeated posterior inference.
  • Because no single environment is uniformly optimal, cognitive studies may need to select environments based on the specific inference objective or on the current posterior over parameters.
  • Adaptive experiment design—choosing the next environment based on an evolving posterior—becomes feasible with ABED, since the same model provides both posterior inference and design-value prediction.
  • In restricted design spaces where informativeness increases monotonically with depth, simple heuristics suffice, but ABED does not rely on such structural regularities and extends to richer design spaces.
  • The finding that design ranking is easier than full posterior recovery suggests that practitioners can trust the KL head for environment selection even when the posterior approximation is imperfect.

Reading between the lines

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

  • If the KL head's calibration holds beyond symmetric trees, the same recipe could rank environments in other simulator-based cognitive models—e.g., tasks with continuous parameters or multi-step trials—where exact posterior inference is intractable.
  • The per-state optimality table implies a participant-level adaptive design: start shallow, and only move to deeper environments when the posterior over H and K is uncertain. The paper does not test this, but its own results suggest it.
  • The consistent advantage of the directly supervised KL head over posterior-derived KL suggests a general lesson for amortized BED: regress the utility target explicitly rather than deriving it from the approximate posterior.
  • A direct test of the method's value would be to apply it to a design space where informativeness is not monotonic with depth; if the KL head still recovers the exact ranking, the approach would beat simple heuristics in settings where the benchmark currently cannot.
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

4 major / 5 minor

Summary. The paper proposes treating the choice of a cognitive experimental environment as a Bayesian experimental design (BED) problem. For Mouselab-MDP tree environments parameterized by depth and branching, it defines an expected information gain (EIG) objective over latent planning parameters (planning horizon H and working-memory K), computes a Monte Carlo BED benchmark using exact posterior inference over the discrete parameter space, and trains an amortized neural surrogate that outputs both an approximate posterior and an auxiliary scalar prediction of the posterior-to-prior KL. Experiments compare environment rankings from this amortized BED (ABED) with MC-BED and with structural/behavioral/supervised baselines. The paper reports that ABED closely matches MC-BED rankings at much lower cost, that no single environment is optimal across EIG, recoverability, and efficiency, and that the framework could support adaptive experiment design.

Significance. If the empirical claims held, this would be a useful step toward automating the design of cognitive experiments: the BED formulation is natural, the discrete-parameter Mouselab-MDP benchmark permits exact posterior evaluation, and the amortized surrogate would make design evaluation cheap. The paper is also unusually explicit about its limitations and about Monte Carlo and seed variability. However, the central evidence is currently too weak to support the generality claimed: the ranking tests use a design space in which a depth heuristic already achieves the same perfect ranking, and no held-out environment is ever evaluated. The contribution is therefore promising but needs substantial additional validation before it can be accepted as stated.

major comments (4)
  1. [Table 3; Supplementary 'Benchmark Design Space'] The central claim that ABED reproduces MC-BED rankings at low cost is not discriminated from a trivial structural heuristic in the reported experiments. Table 3 gives 'Largest tree' Spearman rho=1.000, best=d5b2, regret=0.000 — identical to ABED — and the Supplementary admits that 'expected information gain increases approximately with tree depth, making simple topology-based heuristics unexpectedly competitive.' Thus the reported agreement with MC-BED does not demonstrate that the learned KL head provides useful non-topological information. The comparison should be repeated on the expanded binary-ternary design space (and ideally on a space where depth/branching is not monotone with EIG), with the heuristic rankings reported for those same designs.
  2. [Conclusion; Supplementary 'Expanded Binary-Ternary Benchmark' and 'Limitations'] No held-out environment is tested. The KL head is trained on exact MC-BED targets for the same candidate environments (on separate datasets), so the reported Spearman correlations mostly measure how well the regressor fits environment-specific targets. The conclusion that ABED 'does not rely on structural regularities and naturally extends to settings where environment informativeness cannot be inferred from topology alone' is extrapolation. The paper's own Limitations section restricts the evaluation to symmetric trees. To support the main generalization claim, the authors should evaluate on environments excluded from training (e.g., asymmetric trees, different reward distributions/click costs) or perform leave-one-environment-out validation.
  3. [Table 1 vs. Table 5 and Table 8] There are large inconsistencies in the exact-inference results for the same planning-horizon setting. With H in {1,2,3}, fixed K=2, Table 1 reports d5b2 Ent. Red.=0.552, Min JS=0.183 and d6b2 Ent. Red.=0.632, Min JS=0.001; Supplementary Table 5 reports d5b2 0.753/0.394 and d6b2 0.861/0.338; Table 8 reports d5b2 0.893/0.420 and d6b2 0.935/0.007. Since RQ1 conclusions (e.g., 'd5b2 achieves highest worst-case separability') depend on these numbers, the authors must reconcile the tables and specify the trajectory counts and data splits underlying each.
  4. [Table 1; Supplementary Table 10] The planning-horizon MC-BED comparison has no uncertainty quantification, yet the top two designs differ by only 0.013 EIG units (d6b2=0.636, d5b2=0.623). In the joint benchmark, Supplementary Table 10 shows that a difference of 0.085 has a 95% CI of [0.009,0.161]; by comparison, 0.013 is almost certainly within Monte Carlo noise. The qualitative claim that d6b2 and d5b2 are the two most informative may be robust, but the specific ordering used in the text is not established. Report dataset-level CIs for the planning-horizon EIG estimates or avoid drawing conclusions from this ordering.
minor comments (5)
  1. [References] The entry 'Valentin, S.; Kleinegesse, S.; Bramley, N. R.; ... ???? Bayesian Optimal Experimental Design for Simulator Models of Cognition' has a placeholder year; complete it.
  2. [Table 2] The header repeats 'MAP HK' and interleaves exact and amortized columns in a way that is hard to parse; add subheadings for RQ1/RQ2/RQ3 or split into separate tables.
  3. [Figure 2 / Figure 3] Figure 2 (main) and Figure 3 (supplementary) appear to be the same calibration plot; if so, keep only one or clearly differentiate them.
  4. [Methodology] There is a typo: '8defines the ideal ...' should read 'Equation 8 defines ...'.
  5. [Reproducibility] No code or data availability statement is provided; given the emphasis on reproducibility and the seed-controlled protocol, a link would be helpful. Also, the term 'exact Monte Carlo BED' is slightly misleading: the posterior is exact, but the EIG is a Monte Carlo estimate; consider clarifying the terminology.

Circularity Check

1 steps flagged · score 3.0 of 10

Partial circularity: the ABED ranking claim is a supervised regression onto the exact MC-BED KL target, not an independent prediction.

  1. fitted input called prediction [Methodology, 'Amortized Posterior Inference' (near Eq. 8); also Ablation and RQ3]
    "Furthermore, to support efficient Bayesian experimental design, an auxiliary prediction head is trained to regress the exact posterior-to-prior KL used for environment ranking."

    The auxiliary KL head's supervision is the exact posterior-to-prior KL for datasets from the same candidate environments; ABED then ranks environments by that head's predicted KL and reports that it 'recovers' the exact MC-BED ranking. The environment-level ranking is therefore a regression fit onto the very target whose ranking is claimed. The ablation makes this explicit: 'explicit KL supervision consistently recovers the exact BED ranking with zero regret,' while λ=0 gives regret 0.128. Held-out datasets make the fit nontrivial at dataset level, but no held-out environment is tested; the largest-tree heuristic also achieves the same perfect ranking. Thus the ranking 'prediction' largely reduces to reproducing its training target rather than predicting informativeness of new environments

full rationale

The overall derivation is not fully circular: the EIG objective (Eq. 1), the MC-BED benchmark, and the trade-off findings (RQ5, Table 4) are computed from the generative model and exact posteriors, independent of the amortized surrogate. The circular component is confined to the ABED ranking claim (RQ3): the auxiliary KL head is trained by regressing the exact MC-BED KL, so the observed agreement with MC-BED ranking is a supervised fit/calibration result on the training design space. This is a standard amortized-BED validation, but the paper presents it as 'ABED closely matches the environment rankings of exact Monte Carlo BED' (Abstract), which overstates the independence of the prediction. The paper itself flags external-validity limits: the Supplementary notes 'expected information gain increases approximately with tree depth, making simple topology-based heuristics unexpectedly competitive,' and the Limitations section states the evaluation is 'restricted to symmetric tree-structured Mouselab-MDP environments.' The conclusion that ABED 'naturally extends to settings where environment informativeness cannot be inferred from topology alone' is an extrapolation not supported by any held-out environment experiment. These are correctness/external-validity concerns rather than full circularity, but the ranking-recovery result is substantially weaker than an independent prediction. No load-bearing self-citation or uniqueness-theorem chain is present.

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

The paper relies on several free parameters (η, H set, K set) and domain assumptions about the cognitive model. It does not introduce fundamentally new entities; rather, it applies existing techniques (BED, amortized inference) to a new problem. The latent cognitive parameters (H, K) are not invented for this paper but are standard in the Mouselab-MDP literature.

free parameters (3)
  • Softmax inverse temperature η = 2.0
    Chosen by hand to control action stochasticity in the softmax policy (Eq. 3). The value is not fitted but is a modeling choice.
  • Planning horizon H set = {1,2,3}
    Discrete set of planning horizons used in inference. The prior is uniform over this set. The choice is made by the authors to define the latent parameter space.
  • Working memory K set = {2,5}
    Discrete set of working-memory capacities used in joint inference. The prior is uniform over these values. The choice is made by the authors to define the latent parameter space.
assumptions (3)
  • domain assumption The cognitive model's planner evaluates actions using recursive lookahead with bounded memory, and behavior follows a softmax policy.
    This defines the generative model p(D|θ, ξ). The paper states this in the Behavioral Simulator section. It is a modeling assumption about human planning, not derived from first principles.
  • standard math The posterior over latent cognitive parameters is computable via log-space normalization over a finite parameter set.
    Bayes' rule applied to the discrete parameter space. This is a standard computational assumption.
  • domain assumption The amortized posterior surrogate qφ(θ|D, ξ) can be trained to approximate the exact posterior using distillation and supervised classification.
    This is a core assumption of the amortized inference approach. The paper provides empirical evidence but no theoretical guarantee that the surrogate can achieve sufficient accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Identifying Informative Environments for Cognition Parameter Inference via Bayesian Experimental Design." pith.science (2026). https://pith.science/paper/Q3RI4X7R

@misc{pith2026260728894,
  author       = {Pith},
  title        = {Pith review of: Identifying Informative Environments for Cognition Parameter Inference via Bayesian Experimental Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q3RI4X7R}},
  note         = {Machine review of arXiv:2607.28894}
}
read the original abstract

Computational cognitive modeling seeks to infer latent cognitive mechanisms underlying observed behavior. Bayesian inverse planning provides a principled framework for such inference, but its success depends critically on the experimental environment. Existing approaches typically treat environments as fixed, leaving open the question of which cognitive experiments are most informative for cognition parameter inference. We formulate the design of cognitive planning experiments as a Bayesian Experimental Design (BED) problem, treating the experimental environment as the design variable. We establish an exact Monte Carlo BED benchmark and introduce an amortized Bayesian experimental design framework for efficient posterior inference and design evaluation. Experiments on the Mouselab-MDP process-tracing paradigm show that amortized BED closely matches the environment rankings of exact Monte Carlo BED while substantially reducing computational cost. We further show that no single environment is uniformly optimal across cognitive inference objectives, revealing trade-offs between expected information gain, posterior recoverability, and information efficiency. These results provide a principled framework for designing informative cognitive experiments for Bayesian parameter inference.

Figures

Figures reproduced from arXiv: 2607.28894 by the authors.

Figure 1
Figure 1. Overview of the proposed pipeline for identifying [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Calibration of predicted posterior-to-prior KL across [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Calibration of predicted posterior-to-prior KL across candidate environments for joint inference of [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

77 extracted references · 8 linked inside Pith

  1. [1]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Estimating cognitive biases with attention-aware inverse planning , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  2. [2]

    Models of man, social and rational: Mathematical essays on rational human behavior in a social setting , volume=

    A behavioral model of rational choice , author=. Models of man, social and rational: Mathematical essays on rational human behavior in a social setting , volume=

  3. [3]

    Proceedings of the Annual Meeting of the Cognitive Science Society , volume=

    Goal inference as inverse planning , author=. Proceedings of the Annual Meeting of the Cognitive Science Society , volume=

  4. [4]

    Cognition , volume=

    Action understanding as inverse planning , author=. Cognition , volume=. 2009 , publisher=

    2009

  5. [5]

    and Lieder, Falk and Goodman, Noah D

    Griffiths, Thomas L. and Lieder, Falk and Goodman, Noah D. , title =. Topics in Cognitive Science , volume =. doi:https://doi.org/10.1111/tops.12142 , url =. https://onlinelibrary.wiley.com/doi/pdf/10.1111/tops.12142 , abstract =

    doi:10.1111/tops.12142

  6. [6]

    2024 , publisher=

    Bayesian models of cognition: Reverse engineering the mind , author=. 2024 , publisher=

    2024

  7. [7]

    The Annals of Mathematical Statistics , volume=

    On a measure of the information provided by an experiment , author=. The Annals of Mathematical Statistics , volume=. 1956 , publisher=

    1956

  8. [8]

    International conference on machine learning , pages=

    Deep adaptive design: Amortizing sequential bayesian experimental design , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

Show all 77 references
  1. [9]

    arXiv preprint arXiv:2502.08004 , year=

    Optimizing Likelihoods via Mutual Information: Bridging Simulation-Based Inference and Bayesian Optimal Experimental Design , author=. arXiv preprint arXiv:2502.08004 , year=

    arXiv

  2. [10]

    Statistical Science , volume=

    Modern Bayesian experimental design , author=. Statistical Science , volume=. 2024 , publisher=

    2024

  3. [11]

    2024 , school=

    Automated data acquisition via Bayesian experimental design , author=. 2024 , school=

    2024

  4. [12]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    Environment design for inverse reinforcement learning , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

  5. [13]

    Advances in Neural Information Processing Systems , volume=

    Active exploration for inverse reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=

  6. [14]

    Advances in Neural Information Processing Systems , volume=

    Amortized bayesian experimental design for decision-making , author=. Advances in Neural Information Processing Systems , volume=

  7. [15]

    Proceedings of the Fortieth Conference on Uncertainty in Artificial Intelligence , pages=

    Walking the values in Bayesian inverse reinforcement learning , author=. Proceedings of the Fortieth Conference on Uncertainty in Artificial Intelligence , pages=

  8. [16]

    Bayesian Optimal Experimental Design for Simulator Models of Cognition , author=

  9. [17]

    Elife , volume=

    Designing optimal behavioral experiments using machine learning , author=. Elife , volume=. 2024 , publisher=

    2024

  10. [18]

    Journal of Artificial Intelligence Research , volume=

    Amortized variational inference: A systematic review , author=. Journal of Artificial Intelligence Research , volume=

  11. [19]

    Proceedings of the National Academy of Sciences , volume=

    The frontier of simulation-based inference , author=. Proceedings of the National Academy of Sciences , volume=. 2020 , publisher=

    2020

  12. [20]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Design amortization for Bayesian optimal experimental design , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  13. [21]

    Wellcome Open Research , volume=

    Enhancing experimental design through Bayes factor design analysis: insights from multi-armed bandit tasks , author=. Wellcome Open Research , volume=. 2024 , publisher=

    2024

  14. [22]

    Proceedings of the National Academy of Sciences , volume=

    Automating the practice of science: Opportunities, challenges, and implications , author=. Proceedings of the National Academy of Sciences , volume=. 2025 , publisher=

    2025

  15. [23]

    2024 , publisher=

    Approximate Bayesian inference in and for computational models of cognition: a simulation-based perspective , author=. 2024 , publisher=

    2024

  16. [24]

    arXiv preprint arXiv:2606.26448 , year=

    Closing the Loop to Discover Psychological Theories with an Automated Cognitive Scientist , author=. arXiv preprint arXiv:2606.26448 , year=

    arXiv

  17. [25]

    arXiv preprint arXiv:2510.04698 , year=

    The Bayesian Origin of the Probability Weighting Function in Human Representation of Probabilities , author=. arXiv preprint arXiv:2510.04698 , year=

    arXiv

  18. [26]

    arXiv preprint arXiv:2507.03682 , year=

    Towards Machine Theory of Mind with Large Language Model-Augmented Inverse Planning , author=. arXiv preprint arXiv:2507.03682 , year=

    arXiv

  19. [27]

    bioRxiv , pages=

    Identifiability of bayesian models of cognition , author=. bioRxiv , pages=. 2025 , publisher=

    2025

  20. [28]

    Toward Information Theoretic Active Inverse Reinforcement Learning , author=

  21. [29]

    arXiv preprint arXiv:2603.29003 , year=

    Active Inference with People: a general approach to real-time adaptive experiments , author=. arXiv preprint arXiv:2603.29003 , year=

  22. [30]

    ICLR 2025 Workshop on Foundation Models in the Wild , year=

    Autotom: Automated bayesian inverse planning and model discovery for open-ended theory of mind , author=. ICLR 2025 Workshop on Foundation Models in the Wild , year=

    2025

  23. [31]

    Inverse decision-making using neural amortized Bayesian actors , author=

  24. [32]

    Nature , volume=

    A foundation model to predict and capture human cognition , author=. Nature , volume=. 2025 , publisher=

    2025

  25. [33]

    Nature Human Behaviour , volume=

    Capturing the complexity of human strategic decision-making with machine learning , author=. Nature Human Behaviour , volume=. 2025 , publisher=

    2025

  26. [34]

    Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=

    Optimally designing games for behavioural research , author=. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume=

  27. [35]

    Science , volume=

    Using large-scale experiments and machine learning to discover theories of human decision-making , author=. Science , volume=. 2021 , publisher=

    2021

  28. [36]

    International Statistical Review , volume=

    A review of modern computational algorithms for Bayesian optimal design , author=. International Statistical Review , volume=. 2016 , publisher=

    2016

  29. [37]

    Advances in Neural Information Processing Systems , volume=

    Game design for eliciting distinguishable behavior , author=. Advances in Neural Information Processing Systems , volume=

  30. [38]

    arXiv preprint arXiv:2005.03899 , year=

    Amortized Bayesian inference for models of cognition , author=. arXiv preprint arXiv:2005.03899 , year=

    2005 arXiv

  31. [39]

    Occam's razor is insufficient to infer the preferences of irrational agents , booktitle =

    Stuart Armstrong and S. Occam's razor is insufficient to infer the preferences of irrational agents , booktitle =. 2018 , url =

    2018

  32. [40]

    Dragan , editor =

    Rohin Shah and Noah Gundotra and Pieter Abbeel and Anca D. Dragan , editor =. On the Feasibility of Learning, Rather than Assuming, Human Biases for Reward Inference , booktitle =. 2019 , url =

    2019

  33. [41]

    Proceedings of the 21st international joint conference on Artifical intelligence

    Plan recognition as planning , author=. Proceedings of the 21st international joint conference on Artifical intelligence. Morgan Kaufmann Publishers Inc , pages=

  34. [42]

    Behavioral and brain sciences , volume=

    Resource-rational analysis: Understanding human cognition as the optimal use of limited computational resources , author=. Behavioral and brain sciences , volume=. 2020 , publisher=

    2020

  35. [43]

    Proceedings of the annual meeting of the cognitive science society , volume=

    Bayesian theory of mind: Modeling joint belief-desire attribution , author=. Proceedings of the annual meeting of the cognitive science society , volume=

  36. [44]

    arXiv preprint arXiv:2506.09520 , year=

    How attention simplifies mental representations for planning , author=. arXiv preprint arXiv:2506.09520 , year=

    arXiv

  37. [45]

    Dietterich , title =

    Thomas G. Dietterich , title =. J. Artif. Intell. Res. , volume =. 2000 , url =. doi:10.1613/JAIR.639 , timestamp =

    2000 doi

  38. [46]

    Current opinion in neurobiology , volume=

    Hierarchical reinforcement learning and decision making , author=. Current opinion in neurobiology , volume=. 2012 , publisher=

    2012

  39. [47]

    Trends in cognitive sciences , volume=

    Pragmatic language interpretation as probabilistic inference , author=. Trends in cognitive sciences , volume=. 2016 , publisher=

    2016

  40. [48]

    Learning Options in Reinforcement Learning , booktitle =

    Martin Stolle and Doina Precup , editor =. Learning Options in Reinforcement Learning , booktitle =. 2002 , url =. doi:10.1007/3-540-45622-8\_16 , timestamp =

    2002 doi

  41. [49]

    Algorithms for decision making , author=

  42. [50]

    arXiv preprint arXiv:2506.06941 , year=

    The illusion of thinking: Understanding the strengths and limitations of reasoning models via the lens of problem complexity , author=. arXiv preprint arXiv:2506.06941 , year=

    arXiv

  43. [51]

    Advances in Neural Information Processing Systems , volume=

    Identifiability in inverse reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=

  44. [52]

    Environment Design for Inverse Reinforcement Learning , author=

  45. [53]

    Royal Society Open Science , volume=

    Adaptive planning depth in human problem-solving , author=. Royal Society Open Science , volume=. 2025 , publisher=

    2025

  46. [54]

    PLOS Computational Biology , volume=

    Humans account for cognitive costs when finding shortcuts: An information-theoretic analysis of navigation , author=. PLOS Computational Biology , volume=. 2023 , publisher=

    2023

  47. [55]

    Psychological Research , volume=

    Human performance on visually presented traveling salesman problems , author=. Psychological Research , volume=. 2001 , publisher=

    2001

  48. [56]

    Nature , volume=

    Expertise increases planning depth in human gameplay , author=. Nature , volume=. 2023 , publisher=

    2023

  49. [57]

    40th Annual Cognitive Science Society Meeting (CogSci 2018) , pages=

    Adaptive planning in human search , author=. 40th Annual Cognitive Science Society Meeting (CogSci 2018) , pages=. 2019 , organization=

    2018

  50. [58]

    1959 , publisher=

    Individual choice behavior , author=. 1959 , publisher=

    1959

  51. [59]

    Advances in Neural Information Processing Systems , volume=

    Plans: Neuro-symbolic program learning from videos , author=. Advances in Neural Information Processing Systems , volume=

  52. [60]

    Models of working memory , author=

  53. [61]

    Elements of episodic memory , author=

  54. [62]

    Nature human behaviour , volume=

    Using games to understand the mind , author=. Nature human behaviour , volume=. 2024 , publisher=

    2024

  55. [63]

    bioRxiv , pages=

    Expert navigators deploy rational complexity-based decision precaching for large-scale real-world planning , author=. bioRxiv , pages=. 2024 , publisher=

    2024

  56. [64]

    arXiv preprint arXiv:2507.12821 , year=

    Assessing adaptive world models in machines with novel games , author=. arXiv preprint arXiv:2507.12821 , year=

    arXiv

  57. [65]

    Advances in neural information processing systems , volume=

    Generalization of reinforcement learners with working and episodic memory , author=. Advances in neural information processing systems , volume=

  58. [66]

    Advances in Neural Information Processing Systems , volume=

    Fast deep reinforcement learning using online adjustments from the past , author=. Advances in Neural Information Processing Systems , volume=

  59. [67]

    Varieties of attention , year=

    Changing views of attention and automaticity , author=. Varieties of attention , year=

  60. [68]

    Cognitive psychology , volume=

    The reviewing of object files: Object-specific integration of information , author=. Cognitive psychology , volume=. 1992 , publisher=

    1992

  61. [69]

    , author=

    The magical number seven, plus or minus two: Some limits on our capacity for processing information. , author=. Psychological review , volume=. 1956 , publisher=

    1956

  62. [70]

    Behavioral and brain sciences , volume=

    The magical number 4 in short-term memory: A reconsideration of mental storage capacity , author=. Behavioral and brain sciences , volume=. 2001 , publisher=

    2001

  63. [71]

    Nature communications , volume=

    Single neurons may encode simultaneous stimuli by switching between activity patterns , author=. Nature communications , volume=. 2018 , publisher=

    2018

  64. [72]

    Trends in cognitive sciences , volume=

    A rhythmic theory of attention , author=. Trends in cognitive sciences , volume=. 2019 , publisher=

    2019

  65. [73]

    Nature communications , volume=

    Representation of multiple objects in macaque category-selective areas , author=. Nature communications , volume=. 2018 , publisher=

    2018

  66. [74]

    bioRxiv , pages=

    Working Memory of Multi-Object Scenes in Primate Frontal Cortex , author=. bioRxiv , pages=. 2026 , publisher=

    2026

  67. [75]

    Current Opinion in Systems Biology , pages=

    Think before you fit: parameter identifiability, sensitivity and uncertaintyin systems biology models , author=. Current Opinion in Systems Biology , pages=. 2025 , publisher=

    2025

  68. [76]

    The 3rd Multidisciplinary Conference on Reinforcement Learning and Decision Making, Ann Arbor, MI , year=

    Mouselab-MDP: A new paradigm for tracing how people plan , author=. The 3rd Multidisciplinary Conference on Reinforcement Learning and Decision Making, Ann Arbor, MI , year=

  69. [77]

    Proceedings of the Annual Meeting of the Cognitive Science Society , volume=

    A resource-rational analysis of human planning , author=. Proceedings of the Annual Meeting of the Cognitive Science Society , volume=

Pith tools

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