The DOI matches its Crossref record; 174 of the 174 references it cites were found in a registry or an index; 37 of the 75 sources its claims cite were checked for support.
Meta-learned models of cognition
Open accessAbstract
Psychologists and neuroscientists extensively rely on computational models for studying and analyzing the human mind. Traditionally, such computational models have been hand-designed by expert researchers. Two prominent examples are cognitive architectures and Bayesian models of cognition. Although the former requires the specification of a fixed set of computational structures and a definition of how these structures interact with each other, the latter necessitates the commitment to a particular prior and a likelihood function that - in combination with Bayes' rule - determine the model's behavior. In recent years, a new framework has established itself as a promising tool for building models of human cognition: the framework of meta-learning. In contrast to the previously mentioned model classes, meta-learned models acquire their inductive biases from experience, that is, by repeatedly interacting with an environment. However, a coherent research program around meta-learned models of cognition is still missing to date. The purpose of this article is to synthesize previous work in this field and establish such a research program. We accomplish this by pointing out that meta-learning can be used to construct Bayes-optimal learning algorithms, allowing us to draw strong connections to the rational analysis of cognition. We then discuss several advantages of the meta-learning framework over traditional methods and reexamine prior work in the context of these new insights.
Overview
CoChat Summary
TL;DR
Meta-learning can be used to construct Bayes-optimal learning algorithms, connecting it to rational analysis and extending the scope of cognitive theories beyond traditional Bayesian models. The paper synthesizes prior work and proposes a research program around meta-learned models of cognition, with four arguments for their advantages and a review of applications.
Key findings
- Meta-learning can produce Bayes-optimal learning algorithms; any behavioral phenomenon that can be explained by a Bayesian model can also be explained by a meta-learned model, linking meta-learning to rational analysis.
- Four advantages over traditional Bayesian methods: (1) approximately optimal learning when exact Bayesian inference is computationally intractable; (2) optimal learning when the prior or likelihood cannot be specified; (3) resource-rational models by manipulating algorithmic and computational complexity; (4) integration of neuroscientific insights via model architectures.
- Review examples: Dasgupta et al. (2020) with a single hidden unit captured conservatism bias and base rate neglect; Binz et al. (2022) discovered one-reason decision-making and equal weighting heuristics; Wang et al. (2016) showed a meta-learned model displaying model-based behavior despite pure model-free training; Lake (2019) showed compositional generalization.
- Future domain-general models require more diverse task distributions, natural language instructions, embodiment, and scalable architectures; e.g., Team et al. (2023) trained on over 10^40 possible tasks.
Methods
- Theoretical/conceptual synthesis and research program, not an empirical study; no human or animal sample is reported. It presents a simplified proof (Aitchison, 1975) that the posterior predictive distribution maximizes expected log-likelihood, and describes inner/outer-loop meta-learning with a ~30-line PyTorch reference implementation.
- Main analysis is argumentation and review: four arguments for meta-learning over Bayesian inference, a thematic review of prior psychology/neuroscience studies (heuristics/biases, language, inductive biases, model-based reasoning, exploration, cognitive control), and discussion of limitations and future directions.
- Running example uses insect lengths with assumed normal prior/likelihood; Figure 3 reports meta-learning loss across 30 runs and comparison to exact Bayesian inference.
Limitations
- Stated: meta-learning lacks interpretability/analytical solutions; training is intricate and time-consuming, requiring retraining when assumptions change; there is no guarantee a converged meta-learned model implements Bayes-optimal inference, with reported failures (e.g., Wang et al., 2021); it is not a replacement for Bayesian inference.
- Stated: it remains an empirical question which approximate inference method (meta-learning, variational inference, MCMC) better describes human learning.
- Evident from text: as a review/synthesis, the paper reports no new empirical tests of its proposed research program; many claims about future domain-general models are explicitly speculative.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
30 Valid · 10 Not valid · 26 Not checked · 69 Own results and interpretations
Every source the paper cites for this claim was checked and supports it.
In the paper
“For instance, Wang et al. (2018) presented observations supporting the hypothesis that prefrontal circuits may constitute a meta-reinforcement learning system.”
What the sources say
- Supportedcites Wang et al. (2018) source
The Discussion says the authors propose a meta-reinforcement learning account of prefrontal cortex and dopamine function, and that their simulations account for a diverse range of observations. Simulation 1 specifically relates the model to observed prefrontal activity and behavior. The cited claim accurately characterizes this as support for a hypothesis, rather than proof of it. (LLM, conf=0.90)
hedgedstates a possible cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“We accomplish this by pointing out that meta-learning can be used to construct Bayes-optimal learning algorithms, allowing us to draw strong connections to the rational analysis of cognition.”
- Also stated in: § Meta-Learned Models of Cognition · plain
“we will make use of a recent result from the machine learning community showing that meta-learning can be used to construct Bayes-optimal learning algorithms (Mikulik, Delétang, et al., 2020; Ortega et al., 2019; Rabinowitz, 2019).”
What the sources say
- Supportedcites Mikulik, Delétang, et al., 2020 source
The Abstract (p. 1) says memory-based meta-learning might serve as a general technique for approximating Bayes-optimal agents. The Conclusions (p. 9) report that, in the tasks studied, trained RNN meta-learners converged to solutions that behaved virtually indistinguishably from Bayes-optimal agents. This supports the claim that meta-learning can be used for this purpose, without implying that it always succeeds. (LLM, conf=0.91)
- Supportedcites Ortega et al., 2019 source
Section B, “Bayes-Optimality” (p. 9), explicitly describes how to meta-learn a Bayes-optimal policy. The Conclusions (p. 12) also identify a meta-learning template for building Bayes-optimal agents, supporting the claim that meta-learning can be used to construct such learning algorithms. (LLM, conf=0.92)
- Partly supportedcites Rabinowitz, 2019 sourceExperimental — accuracy not yet measured
Section 3.2.6 (p. 7) says meta-learning’s inner loop should come to approximate amortised Bayes-optimal inference, and reports similar qualitative learning dynamics on in-distribution regression problems. That supports a connection between meta-learning and Bayes-optimal inference, but not the unqualified claim that it constructs Bayes-optimal learning algorithms. (LLM, conf=0.86)
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Also stated in: § Meta-Learned Models of Cognition · plain
Every source the paper cites for this claim was checked and supports it.
In the paper
“It is possible to cast many approximate inference schemes as resource-rational algorithms (A. N. Sanborn, 2017).”
What the sources say
- Supportedcites A. N. Sanborn, 2017 source
The Introduction (page not specified in our copy) describes approximation algorithms that produce good solutions while minimizing computational and memory demands, and says that using them to explain cognition is termed rational process modeling. It identifies sampling and variational methods as two types used to approximate probabilistic cognition, supporting the claim’s broad paraphrase. (LLM, conf=0.88)
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“in practice, it has been observed that meta-learning procedures similar to the one discussed here often lead to networks that closely approximate Bayesian inference (Mikulik, Delétang, et al., 2020; Rabinowitz, 2019).”
What the sources say
- Supportedcites Mikulik, Delétang, et al., 2020 source
The Abstract (p. 1) says memory-based meta-learning may be a general technique for numerically approximating Bayes-optimal agents. Section 4.2 (p. 5) reports that trained agents learned near-optimal Bayesian evidence integration, supporting the citing sentence’s characterization of similar procedures. (LLM, conf=0.90)
- Partly supportedcites Rabinowitz, 2019 sourceExperimental — accuracy not yet measured
Section 3.2.6 (p. 7) says meta-learning's inner loop should come to approximate amortised Bayes-optimal inference, and reports similar qualitative learning dynamics for an in-distribution regression task. Section 4.2.5 (p. 11) reports a similar qualitative match on another task. These passages support a connection, but do not establish that such procedures often produce networks that closely approximate Bayesian inference. (LLM, conf=0.86)
hedgedstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“There are also inference problems in which it is not possible to specify or compute the likelihood function.”
What the sources say
- Supportedcites Cranmer, Brehmer, & Louppe, 2020 source
The opening section (p. 1) says that the likelihood for an observation is typically intractable for complex simulators, in contrast to models where it can be explicitly calculated. The Significance Statement (p. 2) adds that for real-life simulators with large latent spaces, computing the likelihood integral explicitly is impossible. (LLM, conf=0.94)
- Not matchedcites Lueckmann et al., 2021
Couldn’t be matched to a paper in Crossref, so it can’t be checked.
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Supported
74 more claims · 15 with their cited sources checked
Create a free account to see all 135 claims and to verify the 1 reference still open.
- Probably not a substantial claim. 28
- It may state two separate claims as one. 34
Citations
Cites
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404
PyTorch: An Imperative Style, High-Performance Deep Learning Library
Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al. (2019) · arXiv · doi:10.48550/arxiv.1912.01703 · Cited by 31,473
Multilayer feedforward networks are universal approximators
Hornik, K., Stinchcombe, M., & White, H. (1989) · Neural Networks · doi:10.1016/0893-6080(89)90020-8 · Cited by 21,728
Auto-Encoding Variational Bayes
Kingma, D. P., & Welling, M. (2013) · arXiv · doi:10.48550/arxiv.1312.6114 · Cited by 20,067
Stochastic relaxation, gibbs distributions, and the bayesian restoration of images.
Geman, S., & Geman, D. (1984) · IEEE Transactions on Pattern Analysis and Machine Intelligence · doi:10.1109/tpami.1984.4767596 · Cited by 17,949
No free lunch theorems for optimization
Wolpert, D. H., & Macready, W. G. (1997) · IEEE Transactions on Evolutionary Computation · doi:10.1109/4235.585893 · Cited by 14,237
Causality: Models, Reasoning, and Inference
Pearl, J. (2009) · doi:10.1017/cbo9780511803161 · Cited by 11,653
Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
Finn, C., Abbeel, P., & Levine, S. (2017) · arXiv · doi:10.48550/arxiv.1703.03400 · Cited by 8,512
The free-energy principle: a unified brain theory?
Friston, K. (2010) · Nature Reviews Neuroscience · doi:10.1038/nrn2787 · Cited by 8,120
Deep Learning Face Attributes in the Wild
Liu, Z., Luo, P., Wang, X., & Tang, X. (2015) · 2015 IEEE International Conference on Computer Vision (ICCV) · doi:10.1109/iccv.2015.425 · Cited by 7,747
Research works printed but not linked: 34
- Santoro, A. , Bartunov, S. , Botvinick, M. , Wierstra, D. , & Lillicrap, T. (2016). Meta-learning with memory-augmented neural networks. In International conference on machine learning (pp. 1842–1850).
- Lengyel, 2007, Hippocampal contributions to control: The third way, Advances in Neural Information Processing Systems, 20, 889
- Mitchell, 1997, Machine learning, 1
- Bishop, 2006, Pattern recognition and machine learning
- Riveland, 2022, Generalization in sensorimotor networks configured with natural language instructions, bioRxiv
- Mikulik, 2020, Advances in neural information processing systems, 33, 18691
- Savage, 1972, The foundations of statistics
- Duff, 2003, Optimal learning: Computational procedures for Bayes-adaptive Markov decision processes
- Chomsky, 2014, Aspects of the theory of syntax, 11
- Ritter, S. , Wang, J. , Kurth-Nelson, Z. , Jayakumar, S. , Blundell, C. , Pascanu, R. , & Botvinick, M. (2018). Been there, done that: Meta-learning with episodic recall. In International conference on machine learning (pp. 4354–4363).
- Dayan, 2000, Explaining away in weight space, Advances in Neural Information Processing Systems, 13, 430
- Schulze Buschoff, L. M. , Schulz, E. , & Binz, M. (2023). The acquisition of physical knowledge in generative neural networks. In Proceedings of the 40th international conference on machine learning (pp. 30321–30341).
- Vinyals, 2016, Matching networks for one shot learning, Advances in Neural Information Processing Systems, 29, 3637
- Cover, 1999, Elements of information theory
- Newell, 1972, Human problem solving, 104
- Snell, 2017, Prototypical networks for few-shot learning, Advances in Neural Information Processing Systems, 30, 4080
- Sancaktar, 2022, Advances in neural information processing systems, 24170
- Courville, 2008, Advances in neural information processing systems, 369
- Murphy, 2012, Machine learning: A probabilistic perspective
- Strouse, 2021, Collaborating with humans without human data, Advances in Neural Information Processing Systems, 34, 14502
- Rescorla, 1972, Current research and theory, 64
- Schlag, I. , Irie, K. , & Schmidhuber, J. (2021). Linear transformers are secretly fast weight programmers. In International conference on machine learning (pp. 9355–9366).
- Rumelhart, 1988, Parallel distributed processing, 1
- Schmidhuber, J. (1987). Evolutionary principles in self-referential learning, or on learning how to learn: The meta-meta-… hook . Unpublished doctoral dissertation, Technische Universität München.
- Paszke, 2019, Advances in neural information processing systems 32, 8024
- Hinton, 1993, Autoencoders, minimum description length and Helmholtz free energy, Advances in Neural Information Processing Systems, 6, 3
- Kruschke, 1990, Alcove: A connectionist model of human category learning, Advances in Neural Information Processing Systems, 3, 649
- Lueckmann, J.-M. , Boelts, J. , Greenberg, D. , Goncalves, P. , & Macke, J. (2021). Benchmarking simulation-based inference. In International conference on artificial intelligence and statistics (pp. 343–351).
- Czerlinski, 1999, Simple heuristics that make us smart, 97
- Brown, 2020, Language models are few-shot learners, Advances in Neural Information Processing Systems, 33, 1877
- Brändle, 2022b, Intrinsically motivated exploration as empowerment, PsyArXiv
- Kolmogorov, 1965, Three approaches to the quantitative definition of information, Problems of Information Transmission, 1, 1
- Jagadish, 2023, Zero-shot compositional reinforcement learning in humans
- Kumar, S. , Dasgupta, I. , Cohen, J. , Daw, N. , & Griffiths, T. (2020b). Meta-learning of structured task distributions in humans and machines. In International conference on learning representations.
Cited by
Human-like systematic generalization through a meta-learning neural network
Lake, B. M., & Baroni, M. (2023) · Nature · doi:10.1038/s41586-023-06668-3 · Cited by 230
Theory Is All You Need: AI, Human Cognition, and Causal Reasoning
(2024) · doi:10.1287/stsc.2024.0189 · Cited by 151
Building machines that learn and think with people
Collins, K. M., Sucholutsky, I., Bhatt, U., Chandra, K., Wong, L., Lee, M., Zhang, C. E., Zhi‐Xuan, T., Ho, M. K., Mansinghka, V. K., et al. (2024) · Nature Human Behaviour · doi:10.1038/s41562-024-01991-9 · Cited by 216
Technological folie à deux: feedback loops between AI chatbots and mental health
Dohnány, S., Kurth-Nelson, Z., Spens, E., Luettgau, L., Reid, A., Gabriel, I., Summerfield, C., Shanahan, M., & Nour, M. M. (2026) · Nature Mental Health · doi:10.1038/s44220-026-00595-8 · Cited by 51
Understanding the development of reward learning through the lens of meta-learning
Nussenbaum, K., & Hartley, C. A. (2024) · Nature Reviews Psychology · doi:10.1038/s44159-024-00304-1 · Cited by 42
Metadata
- Authors
- Binz, M., Dasgupta, I., Jagadish, A. K., Botvinick, M., Wang, J. X., & Schulz, E.
- Affiliations
- Max Planck Institute for Biological Cybernetics (DE)
- Google DeepMind (United Kingdom) (GB)
- Corresponding author
- Binz, Marcel
- Venue
- Behavioral and Brain Sciences
- Publisher
- Cambridge University Press (CUP)
- Published
- 2023-11-23
- Volume
- 47
- Pages
- e147-e147
- ISSN
- 0140-525X (ISSN-L)
- 1469-1825
- OpenAlex
- W4388929313
- Cited by
- 38
- References
- 106
- Citations by year
- 2026: 8 · 2025: 16 · 2024: 12 · 2023: 2
- FWCI
- 1.42
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- Behavioral and Brain Sciences (publishedVersion)
- PubMed (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- arXiv (Cornell University) (cc-by)
- Funders
- Volkswagen Foundation
- Topics
- Life Sciences › Neuroscience › Cognitive Neuroscience › Cognitive Science and Education Research
- Physical Sciences › Computer Science › Artificial Intelligence › Bayesian Modeling and Causal Inference
- Physical Sciences › Computer Science › Artificial Intelligence › Psychiatry, Mental Health, Neuroscience
- Keywords
- meta-learning, cognitive models, cognitive architecture, Bayesian models of cognition, inductive bias
- MeSH terms
- Algorithms, Bayes Theorem, Cognition, Humans, Learning, Models, Psychological
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Meta-learned models of cognition. https://app.cochat.ai/papers/10.1017/s0140525x23003266
- The paper
- Binz, M., Dasgupta, I., Jagadish, A. K., Botvinick, M., Wang, J. X., & Schulz, E. (2024). Meta-learned models of cognition. Behavioral and Brain Sciences. https://doi.org/10.1017/s0140525x23003266
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.