Pith. sign in

REVIEW 4 major objections 4 minor 25 references

A context-gated dual-process learner claims to reproduce anchoring, priming, framing, and load-induced errors without any bias-specific tuning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A graph-based dual-process model with a learned gate claims to reproduce four cognitive biases in theory-of-mind tasks, but the bias effects are mostly learned from supervised labels rather than emergent.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection The biases are fitted, not emergent: frame/load are inputs to the trained gate, and the Figure 7 numbers don't match the model's own convex combination. the 4 major comments →

arxiv 2509.08705 v1 pith:FDCQHIJM submitted 2025-09-10 cs.AI

One Model, Two Minds: A Context-Gated Graph Learner that Recreates Human Biases

Shalima Binta Manir , Tim Oates This is my paper
classification cs.AI
keywords theory of minddual-process theorygraph convolutional networksmeta-learningcognitive biasesanchoringframing effectgating mechanism
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 human-like reasoning biases can emerge from a single architecture that arbitrates between fast, habitual inference and slow, deliberative meta-adaptation. It introduces OM2M, which couples a graph convolutional network (System 1) with a meta-learning controller (System 2) and a learned context gate that blends their outputs. The authors show that this model, trained only on scenario-appropriate beliefs, quantitatively reproduces anchoring, one-shot priming, framing shifts, and cognitive-load fatigue effects, and that it generalizes to unseen contexts where single-process baselines collapse. If correct, the paper would provide a mechanistic, emergent account of classic dual-process biases rather than one that hard-codes them.

Core claim

The central claim is that a fast-slow arbitration mechanism, implemented as a graph convolutional network for habitual inference, an MLP meta-controller that performs a gradient-like parameter update for deliberate correction, and a context-gated convex blend of the two, can reproduce hallmark human reasoning biases without any bias-specific tuning. In experiments, the gate rises to override habitual errors when evidence is strong, stays intermediate under ambiguity, collapses under cognitive load, and shifts with framing cues while all factual inputs remain constant. The model also achieves about 90 percent accuracy on held-out false-belief contexts, versus 30-50 percent for ablated single-

What carries the argument

The load-bearing mechanism is the contextual gate: a learned scalar g in (0,1) that weights the final logits as y = g * y_System2 + (1 - g) * y_System1. System 1 is a graph convolutional network with agent meta-vectors that produces fast habitual beliefs; System 2 is an MLP meta-controller that reads System 1's output, its flattened parameters, and the context vector, then predicts parameter deltas to produce a revised belief. The gate takes the same context vector—including cognitive load and frame scalars—and decides how much of the final response comes from each system.

Load-bearing premise

The training labels already encode scenario-appropriate beliefs, and the gate is fed the same frame and load cues on which the biases are measured, so the reproduced bias behavior could be learned association rather than an emergent property of fast-slow arbitration.

What would settle it

Train the model with the frame and load scalars removed from the gate input, or with training labels counterbalanced so no context-dependent answer exists, then test whether anchoring, framing, and fatigue effects still appear; if they vanish, the biases were fitted to the provided cues rather than emerging from the arbitration mechanism.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, a single learned mechanism can switch between habitual and deliberative inference, explaining anchoring, framing, priming, and fatigue as byproducts of gate dynamics rather than separate hard-coded rules.
  • The architecture predicts that interventions on the gate (e.g., clamping it high or low) should directly increase or suppress specific biases, offering a concrete handle for debiasing AI decision systems.
  • The generalization result implies that meta-adaptation plus context gating is sufficient for compositional false-belief reasoning across novel agents and contexts, a capability single-pass neural ToM models lack.
  • The framing result shows that the same factual input can yield different outputs solely through a scalar cue, which has direct implications for how AI systems should be tested for robustness to presentation changes.
  • The load experiment shows that resource constraints selectively impair deliberative reasoning while sparing habitual responses, matching a key dual-process prediction and suggesting a way to model fatigue in deployed agents.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • An unstated extension: the gate value itself could serve as an interpretable measure of an AI's uncertainty or cognitive effort, usable as a confidence signal or as a trigger for human oversight.
  • The one-shot priming result suggests a natural test: if the working-memory module were given a decay rate, the model should reproduce the graded, exponentially decaying priming curves measured in human experiments, which the paper does not run.
  • A direct test of whether the biases are truly emergent would be to train on labels that contradict the hypothesized bias (e.g., encouraging System 2 engagement under negative frames); if the gate still shifts with framing, the effect is intrinsic to the arbitration, not just fitted to the labels.
  • The same fast-slow gating framework could be transferred to other social-cognition tasks such as intent prediction or moral judgment, where dual-process theories predict similar context-dependent override patterns.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes OM2M, a dual-process architecture for Theory-of-Mind reasoning in which a GCN-based System 1 provides fast, habitual belief inference, an MLP meta-controller (System 2) produces context-dependent parameter updates, and a learned scalar gate blends the two outputs as a function of context. The authors report that OM2M generalizes to held-out false-belief contexts and reproduces anchoring, priming, framing, and cognitive-load effects, claiming in the abstract and contributions that these biases emerge from fast-slow arbitration 'without any bias-specific tuning.' The manuscript includes ablation results, bias experiments, and pseudocode for training and inference.

Significance. If the central claim were established, a single neural architecture that both generalizes in false-belief tasks and quantitatively reproduces multiple dual-process biases would be a valuable contribution to cognitively grounded AI. The gating mechanism is interpretable and the false-belief generalization experiment is a sensible testbed. However, the paper's empirical protocol does not support the 'emergence without bias-specific tuning' claim: the gate and meta-controller are trained with supervised labels on diverse contexts that include the bias variables, and anchoring and priming are explicitly implanted. The manuscript also contains an internal numerical inconsistency in the framing results and overstates the generalization evidence, which rests on leave-one-out over eight hand-defined contexts. The strengths are the clear architectural idea and the attempt to connect to dual-process theory, but the load-bearing evidence is currently confounded.

major comments (4)
  1. [Training and Evaluation Protocol; Algorithm 3] The central claim that biases emerge 'without any bias-specific tuning' is contradicted by the training protocol. Algorithm 3 trains the meta-controller and gating network with cross-entropy on 'scenario-appropriate belief labels' over a diverse set of scenarios 'including ambiguous, surprising, and cognitively demanding variants,' and the context vector c supplied to both the controller and the gate includes cognitive-load and frame scalars. Thus any dependence of the final output or gate on load/frame can be learned directly from supervised labels; it is fitted behavior, not an emergent property of fast-slow arbitration. To support the emergence claim, the bias variables would need to be absent from training (or at least not supervised with bias-correlated labels) and the bias effects demonstrated at test time. As written, the abstract's 'without any bias-specific tuning' is not justif
  2. [Anchor and Priming Experiments; Algorithm 5] Anchoring and priming are explicitly constructed, not emergent. Anchoring is induced by repeatedly training System 1 on a canonical context and then presenting the anchor context 15 additional times to 'reinforce this bias.' Priming is implemented by a dedicated working-memory module that stores a transient override signal from System 2. These are precisely bias-specific manipulations. The paper can claim at most that a deliberately implanted bias can be overridden by the learned gate, not that OM2M 'autonomously exhibits' anchoring and priming without bias-specific tuning.
  3. [Results and Analysis: Framing Effect (Figure 7)] The reported blended probabilities are inconsistent with the model definition. In the 'Contextual Gate' section the final output is y = g · y_System2 + (1−g) · y_System1. With y_System1 = 0.01 and y_System2 = 1.00, the gate values 0.04, 0.13, and 0.24 yield blended P(Basket) = 0.0496, 0.1387, and 0.2476, not the reported 0.11, 0.92, and 1.00. Because the framing effect is quantified through these numbers, the quantitative evidence for the effect is internally unreliable; the authors need to clarify how the blended values were computed or correct the figure.
  4. [False-Belief Generalization Test; Table 1] The '90% accuracy on an unseen, rich-context split' claim is not supported by the experimental design. The leave-one-out protocol covers only eight hand-defined 3-bit context configurations (plus three agents), and the two-alternative output makes the space very small. With 90±20% across five seeds, the result is far weaker than the abstract's generalization claim, and the small discrete space leaves room for memorization rather than compositional reasoning. The authors should either present a larger, richer context space or substantially soften the claim.
minor comments (4)
  1. [Appendix (Algorithms 1-5)] The pseudocode is high-level and omits several details needed for reproducibility: number of GCN layers and hidden sizes, the MLP controller architecture and parameter-flattening scheme, the precise encoding of the context vector, dataset sizes, and training epochs. Please provide these details or point to released code.
  2. [Results and Analysis (Figures 3-7)] Figures 3-7 are referenced in the text but do not appear in the manuscript. If they were omitted during preparation, they should be included in a revised version; as submitted, the reader cannot verify the plotted trends.
  3. [Table 3] The 'Phase 1/2/3' labels are not defined in the table caption. Please rename them to 'anchor context,' 'conflicting context,' and 'ambiguous context' to match the text.
  4. [References] The citation 'Kipf 2016' for graph convolutional networks is normally Kipf and Welling (ICLR 2017); please update the reference.

Circularity Check

3 steps flagged

Bias 'emergence' is largely built in: anchoring is induced by overtraining, priming is a dedicated working-memory store, and frame/load enter as gate inputs trained on labeled contexts.

specific steps
  1. self definitional [Anchor and Priming Experiments, 'One-Shot Priming'; Results, 'Priming and Working Memory Test result']
    "A working memory module stores a transient override signal generated by System 2 in response to a priming context. The model is then probed with an ambiguous scenario; the priming effect lasts one trial unless refreshed. ... the probability spikes to 1.00, evidencing a full System 2 override through working memory. On the subsequent ambiguous probe, with no re-priming, the working memory is empty and the probability returns to baseline."

    The 'priming effect' is defined as storing an override signal in a working-memory module and reading it on the next trial. The observed spike-and-decay is the module's storage/retention semantics, not an emergent property of fast-slow arbitration. The paper presents this as reproducing a cognitive bias, but the result is guaranteed by construction: store signal -> use signal -> clear signal. The claim 'without any bias-specific tuning' is contradicted by this dedicated module, whose sole purpose is to create exactly this one-trial effect.

  2. fitted input called prediction [Anchor Bias Results, 'Inducing Anchoring']
    "System 1 was trained repeatedly on the canonical context where Sally is present and the toy is definitely in the box, converging to a low loss (0.236→0.004 over 60 epochs). ... After freezing System 1, the anchor context was presented 15 additional times to reinforce this bias."

    The anchor is deliberately induced by repeated training on a single context. The reported anchored response (P(Box)=0.99, gate=0.20) is simply the fitted output of the trained predictor on its training context. Calling this a reproduction of anchoring 'without any bias-specific tuning' is circular: the bias is the training. The subsequent 'override' is also a learned gate response on labeled conflicting/ambiguous contexts, so the whole anchoring demonstration reduces to the training protocol rather than to emergent dual-process arbitration.

  3. fitted input called prediction [Contextual Gate; Framing Effect Experiment; Framing Effect: Contextual Modulation Without Changing Facts]
    "the contextual gating network, which receives the current context vector—including cognitive load/fatigue and scenario framing variables—and produces a scalar g∈(0,1) ... Framing was encoded as a scalar in the context vector and supplied directly to the gating network. ... The key driver of the effect is the learnable gate, which rises sharply from 0.04 (negative) to 0.24 (positive)."

    Because g=Gate(context) and frame is an explicit component of that context, and because Algorithm 3 trains the gate with cross-entropy on scenario-appropriate labels over diverse contexts, the frame-dependence of g is a learned input-output mapping. Reporting this as an emergent 'framing bias' is circular: the very input feature used to train the gate is the variable whose effect is then reported. Moreover, the quoted blended probabilities do not satisfy the paper's own convex-combination equation y=g·y2+(1−g)·y1 (e.g., negative frame: 0.04·1.00+0.96·0.01=0.05, not 0.11), so the quantitative evidence is also internally inconsistent.

full rationale

The paper's architecture itself—especially the false-belief leave-one-out generalization test—may have non-circular, held-out content: System 1 is pretrained, System 2 and the gate are trained on other contexts, and held-out contexts are evaluated. That part is not circular. However, the central advertised contribution, 'Without any bias-specific tuning, OM2M quantitatively reproduces anchoring, priming, framing, and load-induced errors,' is not supported by the described protocols in a way that would demonstrate emergence. Anchoring is explicitly induced by overtraining on one context plus 15 reinforcement exposures. Priming is implemented by an explicit working-memory module that stores and then clears a System 2 override; the one-shot spike is the module's definition. Framing is supplied as a scalar directly to the gating network that is trained on labeled contexts, so the measured shift in g with frame is the trained gate's response to that input feature, not an emergent bias. The load claim is additionally clouded by an internal contradiction: the Training and Evaluation Protocol says the gate and controller are trained 'on a diverse set of scenarios, including ambiguous, surprising, and cognitively demanding variants,' while the Cognitive Load Effects section says 'The model is trained only on low-load data, so all fatigue effects emerge during evaluation.' If the former is true, load sensitivity is fitted; if the latter is true, the gate has essentially no learned exposure to load and the reported smooth nonlinear drop is an extrapolation/initialization artifact, not a demonstrated emergent property. Either way, the paper's emergence claim is not established by its own methods. No load-bearing self-citation was found; all citations are to external cognitive-science and machine-learning work. The internal arithmetic inconsistency in Figure 7 further weakens the quantitative framing result. Overall, the bias-reproduction claims reduce, to a substantial degree, to explicit construction or fitted input-output behavior, so the circularity score is 7 rather than 0.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The central claims depend on fitted gate and controller weights and on synthetic labels. The gate is the mechanism that produces the reported cognitive biases, and its weights are fitted to the very context features, load and frame, that are then measured as causing the biases. No external benchmark or human data grounds the invented modules.

free parameters (4)
  • GCN weights (System 1) = learned, not reported
    Core weights of the habitual reasoner, trained on synthetic canonical contexts.
  • Meta-controller weights (System 2) = learned, not reported
    Predicts parameter deltas; trained with labels on diverse scenarios.
  • Context gate weights = learned, not reported
    Maps context features, including load and frame, to the scalar gate; central to reported bias effects.
  • Agent meta-vectors = learned, not reported
    Per-agent latent embedding appended to GCN output; part of System 1.
axioms (5)
  • domain assumption Graph convolution is a sufficient encoder for false-belief reasoning.
    The paper assumes node and edge structure over agents, objects, and locations captures the belief-inference problem; no comparison to non-graph encoders.
  • domain assumption The synthetic context labels are correct ground truth for beliefs.
    No human annotation or external benchmark is used; authors define the expected belief for each 3-bit context.
  • ad hoc to paper Dual-process theory maps to GCN plus meta-controller plus gate.
    This mapping is asserted in the contributions and background without derivation from cognitive data.
  • ad hoc to paper Test-time cognitive-load effects can be studied by training only at low load.
    The fatigue experiment relies on extrapolation of the gate network to unseen load values; this is not demonstrated.
  • standard math Gradient-based optimization and cross-entropy loss yield the intended behavior.
    Standard machine learning assumptions; not proven in the paper.
invented entities (3)
  • context-gated blending network (scalar gate) no independent evidence
    purpose: Outputs scalar g in (0,1) to blend System 1 and System 2 outputs.
    No external validation ties the gate values to human cognitive effort; the gate behavior is only observed in the same simulations that report the bias effects.
  • agent meta-vector no independent evidence
    purpose: Per-agent latent state representing persistent beliefs.
    Introduced as an extra learned embedding; no independent measurement or comparison to human agent-specific representations.
  • working memory module no independent evidence
    purpose: Stores a one-shot System 2 override signal for priming.
    Implementation is described only qualitatively; the one-shot spike in the priming experiment may be an artifact of the protocol rather than a general mechanism.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of One Model, Two Minds: A Context-Gated Graph Learner that Recreates Human Biases." pith.science (2026). https://pith.science/paper/FDCQHIJM

@misc{pith2026250908705,
  author       = {Pith},
  title        = {Pith review of: One Model, Two Minds: A Context-Gated Graph Learner that Recreates Human Biases},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FDCQHIJM}},
  note         = {Machine review of arXiv:2509.08705}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We introduce a novel Theory of Mind (ToM) framework inspired by dual-process theories from cognitive science, integrating a fast, habitual graph-based reasoning system (System 1), implemented via graph convolutional networks (GCNs), and a slower, context-sensitive meta-adaptive learning system (System 2), driven by meta-learning techniques. Our model dynamically balances intuitive and deliberative reasoning through a learned context gate mechanism. We validate our architecture on canonical false-belief tasks and systematically explore its capacity to replicate hallmark cognitive biases associated with dual-process theory, including anchoring, cognitive-load fatigue, framing effects, and priming effects. Experimental results demonstrate that our dual-process approach closely mirrors human adaptive behavior, achieves robust generalization to unseen contexts, and elucidates cognitive mechanisms underlying reasoning biases. This work bridges artificial intelligence and cognitive theory, paving the way for AI systems exhibiting nuanced, human-like social cognition and adaptive decision-making capabilities.

Figures

Figures reproduced from arXiv: 2509.08705 by Shalima Binta Manir, Tim Oates.

Figure 1
Figure 1. Figure 1: OM2M model pipeline overview: The input graph [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Relational graph representation of the Sally-Anne [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of increasing cognitive load on ambiguous [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: One-shot priming effect: P(Basket) on the am￾biguous context spikes to 1.00 immediately after priming, then returns to the baseline value 0.00 on the next trial, mir￾roring fast, transient memory-limited priming observed in human cognition. Interpretation: This sequence demonstrates that the model’s working memory mechanism enables a targeted, single-use belief revision: only the first post-prime infer￾enc… view at source ↗
Figure 5
Figure 5. Figure 5: Ambiguous context decomposition: System 2 ac [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Easy context: Correct inference is robust to cog [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 5 linked inside Pith

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    L.; Jara-Ettinger, J.; Tenenbaum, J

    Baker, C. L.; Jara-Ettinger, J.; Tenenbaum, J. B.; and Saxe, R. R. 2017. Rational quantitative attribution of beliefs, desires and percepts in human mentalizing. In Nature Human Behaviour, volume 1, 0064. Nature Publishing Group

    2017

  4. [4]

    L.; Saxe, R.; and Tenenbaum, J

    Baker, C. L.; Saxe, R.; and Tenenbaum, J. B. 2011. Bayesian theory of mind: Modeling joint belief‐desire attribution. Proceedings of the Annual Meeting of the Cognitive Science Society, 33(33): 2469--2474

    2011

  5. [5]

    W.; Hamrick, J

    Battaglia, P. W.; Hamrick, J. B.; Bapst, V.; Sanchez-Gonzalez, A.; Zambaldi, V.; Malinowski, M.; Tacchetti, A.; Raposo, D.; Santoro, A.; Faulkner, R.; et al. 2018. Relational inductive biases, deep learning, and graph networks. arXiv preprint arXiv:1806.01261

    Pith/arXiv arXiv 2018

  6. [6]

    X.; Dabney, W.; Munos, R.; Momennejad, I.; Lithcap, Z

    Botvinick, M.; Wang, J. X.; Dabney, W.; Munos, R.; Momennejad, I.; Lithcap, Z. C.; et al. 2020. Deep reinforcement learning and its neuroscientific implications. Neuron, 107(4): 603--616

    2020

  7. [7]

    Evans, J. S. B. 2008. Dual-processing accounts of reasoning, judgment, and social cognition. Annu. Rev. Psychol., 59(1): 255--278

    2008

  8. [8]

    Finn, C.; Abbeel, P.; and Levine, S. 2017. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, 1126--1135. PMLR

    2017

  9. [9]

    Grant, E.; Finn, C.; Levine, S.; Darrell, T.; and Griffiths, T. L. 2018. Recasting gradient-based meta-learning as hierarchical bayes. In International Conference on Learning Representations

    2018

  10. [10]

    D.; Hosseini, S.; Chandra, V.; and Lago, P

    Hosseinpanah, A.; Lin, H.; Lane, N. D.; Hosseini, S.; Chandra, V.; and Lago, P. 2023. Social reasoning in graph neural networks. Frontiers in Artificial Intelligence, 6: 1245230

    2023

  11. [11]

    Kahneman, D. 2003. Maps of bounded rationality: Psychology for behavioral economics. American economic review, 93(5): 1449--1475

    2003

  12. [12]

    Kahneman, D. 2011. Thinking, Fast and Slow. New York, NY: Farrar, Straus and Giroux. ISBN 978-0-374-53355-7

    2011

  13. [13]

    Kipf, T. 2016. Semi-Supervised Classification with Graph Convolutional Networks. arXiv preprint arXiv:1609.02907

    Pith/arXiv arXiv 2016

  14. [15]

    Kosinski, M. 2024. Evaluating large language models in theory of mind tasks. Proceedings of the National Academy of Sciences, 121(45): e2405460121

    2024

  15. [16]

    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

  16. [17]

    Mi, F.; Xue, R.; Shi, W.; Zhu, Y.; Wang, Y.; Ong, Y.-S.; and Sun, A. 2023. Emergent theory of mind in large language models via prompting. In arXiv preprint arXiv:2302.02083

    Pith/arXiv arXiv 2023

  17. [18]

    Milli, S.; Lieder, F.; and Griffiths, T. L. 2021. A rational reinterpretation of dual-process theories. Cognition, 217: 104881

    2021

  18. [19]

    Nichol, A.; Achiam, J.; and Schulman, J. 2018. On first-order meta-learning algorithms. In arXiv preprint arXiv:1803.02999

    Pith/arXiv arXiv 2018

  19. [20]

    Premack, D.; and Woodruff, G. 1978. Does the chimpanzee have a theory of mind? Behavioral and Brain Sciences, 1(4): 515--526

    1978

  20. [21]

    A.; and Botvinick, M

    Rabinowitz, N.; Perbet, F.; Song, F.; Zhang, C.; Eslami, S. A.; and Botvinick, M. 2018. Machine theory of mind. In International conference on machine learning, 4218--4227. PMLR

    2018

  21. [22]

    E.; and West, R

    Stanovich, K. E.; and West, R. F. 2000 a . Advancing the rationality debate. Behavioral and brain sciences, 23(5): 701--717

    2000

  22. [23]

    E.; and West, R

    Stanovich, K. E.; and West, R. F. 2000 b . Individual differences in reasoning: Implications for the rationality debate? Behavioral and Brain Sciences, 23(5): 645--665

    2000

  23. [24]

    Swaby, L.; Stewart, M.; Harrold, D.; Willis, C.; and Palmer, G. 2024. Machine Theory of Mind for Autonomous Cyber-Defence. arXiv preprint arXiv:2412.04367

    Pith/arXiv arXiv 2024

  24. [25]

    Tversky, A.; and Kahneman, D. 1974. Judgment under Uncertainty: Heuristics and Biases: Biases in judgments reveal some heuristics of thinking under uncertainty. science, 185(4157): 1124--1131

    1974

  25. [26]

    D.; Spelke, E

    Ullman, T. D.; Spelke, E. S.; Battaglia, P. W.; and Tenenbaum, J. B. 2017. Mind games: Game engines as an architecture for intuitive physics and psychology. Trends in Cognitive Sciences, 21(9): 649--665

    2017

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.