Grounded world models in biological organisms and future embodied AI

Davide Nuzzi; Giovanni Pezzulo; Marco D'Alessandro; Paul Cisek; Riccardo Proietti; Roberto Bottini

REVIEW 3 major objections 3 minor 88 references

Grounded world models in biological organisms and future embodied AI

T0 review · 3 major / 3 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read The paper argues that living organisms build language on top of grounded world models acquired through sensorimotor interaction — the opposite of today's language-first embodied AI.

desk verdict A clear, honest perspective that makes a plausible grounding-first case but leaves the central necessity claim open — worth refereeing as a position piece, not as research. read the letter →

arxiv 2607.13560 v1 pith:JMEZYX5I submitted 2026-07-15 q-bio.NC cs.AI

Giovanni Pezzulo , Davide Nuzzi , Marco D'Alessandro , Riccardo Proietti , Roberto Bottini , Paul Cisek This is my paper

classification q-bio.NCcs.AI

keywords embodiedAIworldmodelsgroundingpredictiveprocessingactiveinferencecognitivemapsaffordanceslanguage

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 organizing principle of biological intelligence is the acquisition of grounded world models through autonomous, active interaction with the environment, and that language, reasoning, and social understanding are built on top of these models. Current embodied AI inverts this order by using linguistic regularities as the scaffold onto which other modalities are attached. The authors marshal five neural circuits—spatial/conceptual navigation, affordance perception, active exploration, allostatic control, and self–other distinction—as evidence that predictive sensorimotor circuits precede and support higher cognition. If correct, this means the dominant training regime for embodied AI is not just missing details but is structurally inverted. The paper does not claim the biological route is proven necessary; it explicitly leaves open whether active and passive approaches may converge.

What carries the argument

The carrying mechanism is the grounded world model: a generative, action-conditioned model of latent states and transitions, learned not from passive corpora but through continuous action–perception loops. The authors support this with five conserved neural circuits — grid/place cells for navigation in physical and conceptual spaces, dorsal-stream affordance competition, neuromodulatory control of exploration, allostatic–interoceptive regulation of needs and value, and corollary discharge for self–other distinction. Each circuit illustrates how intrinsic dynamics are preconfigured and then calibrated by action, providing the scaffold on which language and abstract thought are built.

What would settle it

A demonstration that a language-first embodied agent, given only text and static corpora (no interactive sensorimotor pretraining), acquires human-like understanding of physical fragility, social norms, and value alignment — or a brain-imaging study showing that grid-like codes and grounded semantic representations arise equivalently in animals that never acted in the world — would refute the necessity claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that in humans, language and communication emerge from world models — exactly the opposite of current embodied AI, where language-derived knowledge is often used as the scaffold for other forms of knowledge and reasoning. A grounded world model is defined as knowledge of relevant latent environmental states and action-dependent transitions, acquired through lived, open-ended sensorimotor interaction. The paper argues that intrinsic neural dynamics provide inductive priors, that action aligns these dynamics with the world, and that the same predictive and control circuits are later 'detached' and reused for planning, imagination, understanding others, and language

Load-bearing premise

The claim stands or falls on whether sensorimotor grounding is necessary — not merely one workable route — for human-like semantics, social alignment, and values, since the paper concedes that active and passive learning may converge on some domains such as space.

Editorial extensions

If this is right

If the paper is right, the language-grounding problem in embodied AI is a symptom of an inverted training order; meaning must come from sensorimotor experience before words.
Embodied AI should shift from passive pretraining to autonomous, intrinsically motivated, open-ended learning, with action as the alignment mechanism.
World models for planning, anticipation, and social inference would be learned before or alongside language rather than distilled from text.
Grid-like codes may provide a universal neural format for organizing both physical and abstract conceptual spaces, useful as an inductive bias.
Socially shared world models, built through interaction, may be the route to human-aligned values rather than post-hoc fine-tuning.

Reading between the lines

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

A testable prediction follows: language-first vision-language-action agents should underperform on genuinely physical and social tasks (e.g., handling a fragile cup) relative to agents trained with an interactive sensorimotor curriculum, holding compute constant — an experiment the paper describes in spirit but does not run.
The paper's five circuits imply a concrete architectural checklist for embodied AI: intrinsic dynamics as priors, efference copies for self-modeling, interoceptive valuation for goals, and an information-gain drive for exploration. Implementing these as inductive biases could be a faster path than scaling passive data.
If convergence between active and passive learning occurs for space but not for values or social norms, then the biological route may be essential specifically for alignment, not for all of cognition — an asymmetry the paper leaves open.
The 'hardware lottery' argument suggests that language-centric methods dominate partly because text is computationally cheap; building the sensorimotor training stack is the bottleneck, so the paper implicitly predicts a hardware/software shift toward interactive embodied training.

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.

Desk Editor's Note

read the letter

Colleague — this is a perspective piece, not a research result. That's the right frame from the start. The authors make a clear, readable case that biological intelligence builds grounded world models from sensorimotor interaction and only later attaches language, and that current embodied AI gets the order backward. The five examples — spatial/conceptual navigation, affordance competition, active perception, allostasis, corollary discharge — are accurate at the level they're summarized, and the figures are genuinely useful.

The best part, honestly, is the Discussion. The authors explicitly say they don't know whether the passive, language-first paradigm will hit fundamental limits, that active and passive approaches might converge on similar representations (space is their example), and that we don't know which biological details are essential vs. contingent. That directness earns credibility.

Soft spots, in proportion. The load-bearing premise is that the biological route is necessary or at least superior for human-like semantics, values, and social alignment. That premise is not established — and the paper's own concession about convergence undercuts it. If a language-first or purely engineering route can produce equivalent grounding, then the 'opposite' framing is a description of current training regimes, not a fundamental architectural fact. The paper also leans heavily on the authors' own active-inference/predictive-processing program; several load-bearing refs are self-citations or preprints (refs 2, 11, 15, 23, 30, 60, 86, 87). That's not a flaw per se for a position paper, but it means some specific claims aren't yet independently verified. And there's no concrete architecture, benchmark, or falsifiable prediction. Again, for a perspective that's acceptable, but it means the practical value is directional.

The central claim holds up as a plausible synthesis. The five neuroscience examples are consistent with it. But the paper does not — and cannot, given current evidence — close the necessity question. That's the honest bottom line.

Who benefits: researchers in embodied AI, cognitive robotics, and anyone arguing about what training regimes to adopt. It's a good discussion piece, and it deserves a serious referee — as a perspective/opinion piece, not as a research contribution. I'd probably cite it for the framing, and I'd discuss it in a reading group, but I wouldn't treat it as evidence for any specific empirical claim.

Referee Report

3 major / 3 minor

Summary. The paper presents a perspective on the relationship between biological and artificial world models. It argues that living organisms acquire grounded world models through active, autonomous interaction with the environment, and that these models provide the semantic foundation onto which language is attached. Current embodied AI, by contrast, typically uses language-derived knowledge as the scaffold for other modalities. The authors illustrate their thesis with five neuroscience examples: entorhinal–hippocampal navigation in physical and conceptual spaces; affordance-based perception; active perception and neuromodulatory control; allostatic interoceptive regulation; and corollary-discharge-based self/other distinction. They then discuss implications for future embodied AI, including social grounding and the inclusion of intrinsic goals. The paper is explicitly framed as an integrative review/position piece and contains no new experimental data.

Significance. If the thesis is correct, it would have a significant impact on the design of embodied AI, shifting from passive, language-centric pretraining toward interaction-driven, intrinsically motivated learning and social alignment. The manuscript's strengths are the accurate and accessible presentation of recent neuroscience findings (e.g., toroidal grid-cell topology, affordance competition, interoceptive/allostatic inference), the clear effort to link these to AI design, and an unusually honest Discussion that acknowledges the central uncertainty about whether the biological route is necessary. The paper's value is in proposing a research agenda rather than in providing definitive evidence.

major comments (3)

[Discussion] The paper's central contrast is weakened by its own concession: after stating that biological systems are organized in the opposite way to language-first embodied AI, the Discussion says 'It is also possible that active and passive learning approaches converge toward similar representations of certain domains, such as space, despite relying on very different learning processes.' If convergence at the level of semantics and values is possible, the 'exactly the opposite' claim (Introduction) is not a fundamental architectural fact but a contingent description of current training regimes. This is load-bearing because the prescriptive conclusion—that future embodied AI should adopt the biological route—depends on the biological route being necessary or at least superior. Please either provide evidence for non-convergence or reframe the claim as a testable hypothesis.
[Discussion] The assertion that 'the biological route has the advantage that phylogenetic development... leaves no conceptual gaps' is presented as a decisive advantage but is not supported by evidence or a mechanistic argument. It is unclear what 'no conceptual gaps' means in engineering terms, and the claim conflates phylogenetic history with normative necessity. Since this sentence is the main justification for why the biological route is preferable, it needs either a precise formulation (e.g., a specific prediction about learning dynamics) or to be explicitly labeled as a hypothesis.
[Introduction] The paper defines a world model narrowly as 'relevant latent states... and an action-dependent transition model,' but later broadens the term to include interoceptive/allostatic systems, emotions, and value systems (e.g., 'allostatic-interoceptive system can be understood as a form of world model'). This conceptual stretch makes it difficult to identify exactly which principles are being proposed for transfer to AI. The prescriptive discussion would be stronger if it distinguished between (i) predictive models of external environment dynamics, (ii) internal state regulation models, and (iii) value/alignment mechanisms, and stated which of these are candidates for AI design.

minor comments (3)

[References] Several load-bearing references are preprints or self-citations (e.g., refs 2, 11, 15, 23, 60, 80). While this is acceptable for a perspective, the manuscript should mark them clearly as author preprints and, where possible, cite peer-reviewed alternatives to avoid over-reliance on one research program.
[Figure 1] The caption states that the latent manifold is 'obtained by encoding this neural activity with a model trained to... decode the behavioral state.' It would be helpful to state explicitly that the behavioral annotations are derived from freely moving animals and that the manifold is an inference, not a direct observation.
[Introduction] The terms 'passive' and 'active' are used throughout but defined only implicitly. A one-sentence operational definition would improve clarity.

Circularity Check

0 steps flagged · score 2.0 of 10

Perspective paper with no formal derivation; self-citation density creates an independence caveat but not circularity.

full rationale

This is a review/perspective rather than a derivation: there are no equations, fitted parameters, or predictions whose values are forced by construction. The five circuit examples are presented as empirical illustrations (Kato et al.; Gardner et al.; Aronov et al.; Cisek), and each is described as illustrative rather than as an output of a derivation. The central claim that language is grounded in sensorimotor world models is a synthesis of embodied-cognition and predictive-processing theories, supported by external references (e.g., Barsalou, Glenberg, Gibson, Levinson) as well as the authors' own prior work. The authors explicitly disclaim that the necessity of the biological route is established: 'it is unclear whether the predominantly passive paradigm will eventually encounter fundamental limitations' and 'we do not yet know with confidence which aspects of biological world-model acquisition are fundamental and which are contingent products of evolution.' Because the key necessity premise is marked as an open question, it is not used as a closed input to a formal result. The self-citations (e.g., refs 2, 11, 15, 23, 30, 60, 77, 86, 87) appear in bridge arguments, but the main empirical content is externally grounded and the normative conclusion is explicitly hedged. No step was found where an equation reduces to its own inputs by construction, or where a fitted parameter is renamed as a prediction; the self-citation cluster is a citation-independence concern, not a circular reduction.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The paper makes no quantitative predictions and introduces no free parameters or invented entities. It rests on domain assumptions about brain organization and the nature of current AI training, several of which are supported by the authors' own prior work.

assumptions (4)

domain assumption Biological intelligence is organized so that grounded world models acquired through sensorimotor interaction provide the semantic scaffold to which language is attached.
This is the central premise of the paper, stated in the abstract and Introduction; it is assumed from embodied-cognition literature rather than derived from new data.
domain assumption Intrinsic neural dynamics are largely preconfigured and are aligned with the external world through action (e.g., grid-cell and hippocampal dynamics precede navigational experience).
Section 'Navigation in physical and conceptual spaces' relies on refs 17, 30, and 31, including author preprints, to support the claim that intrinsic dynamics precede experience.
domain assumption Current embodied AI (VLA models) is trained predominantly passively, with language as the dominant scaffold for other modalities.
Introduction and Discussion characterize the current field qualitatively; no quantitative comparison of training data or scaffold usage is provided.
ad hoc to paper Principles derived from biological grounded world models can be transferred to embodied AI as constraints or inductive biases.
Discussion's prescriptive agenda assumes that biological principles are transferable to engineered systems; the paper itself notes it is unclear which aspects are essential versus incidental.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Grounded world models in biological organisms and future embodied AI." pith.science (2026). https://pith.science/paper/JMEZYX5I

@misc{pith2026260713560,
  author       = {Pith},
  title        = {Pith review of: Grounded world models in biological organisms and future embodied AI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JMEZYX5I}},
  note         = {Machine review of arXiv:2607.13560}
}

read the original abstract

Recent advances in generative and embodied AI have been driven by large-scale predictive learning over multimodal data. However, the resulting systems remain largely based on passive training regimes where linguistic regularities create the scaffold onto which information from other modalities is attached. Conversely, neuroscience and cognitive science suggest that biological intelligence is organized in the opposite way, where grounded world models acquired through interaction with the environment provide the semantic scaffold to which language is attached. Here, we illustrate five examples of neural circuits supporting grounded world modelling, which underlie navigation in physical and conceptual spaces, affordance-based perception and interaction with objects, active perception and exploratory learning, allostatic control and emotion, and the distinction between self- and world-generated outcomes. These examples highlight several features largely missing from current embodied AI, including the role of intrinsic dynamics as a foundation for learning, the centrality of action in aligning these dynamics with the external world, the prominence of autonomous experience and open-ended learning over passive assimilation of externally provided data, and the fact that early predictive and control mechanisms scaffold higher cognitive abilities such as reasoning, conceptual navigation, planning, imagination, understanding others' minds, and communication. Finally, we discuss whether and how principles derived from biological systems may inform future embodied AI, including training regimes based on social interaction to construct world models that are not only grounded but also socially shared and aligned with human norms and values.

Figures

Figures reproduced from arXiv: 2607.13560 by the authors.

**Figure 1.** Whole-brain neural dynamics and a preconfigured behavioral manifold in C. elegans. A. Schematic of pan-neuronal calcium imaging in a physically immobilized nematode (top), and the corresponding raw activity matrix (bottom) showing normalized fluorescence (ΔF/F) simultaneously recorded from 109 neurons across the head ganglia and anterior ventral cord over the full recording session (data from Kato et al. 22). Becaus… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

88 extracted references · 6 canonical work pages

[1]

M., Ullman, T

Lake, B. M., Ullman, T. D., Tenenbaum, J. B. & Gershman, S. J. Building machines that learn and think like people. Behav Brain Sci 40, e253 (2017)

2017
[2]

& Friston, K

Pezzulo, G., Parr, T., Cisek, P., Clark, A. & Friston, K. Generating meaning: active inference and the scope and limits of passive AI. Trends in Cognitive Sciences https://doi.org/10.1016/j.tics.2023.10.002 (2023) doi:10.1016/j.tics.2023.10.002

doi:10.1016/j.tics.2023.10.002 2023
[3]

Taniguchi, T. et al. World Models and Predictive Coding for Cognitive and Developmental Robotics: Frontiers and Challenges. Advanced Robotics 37, (2023)

2023
[4]

A path towards autonomous machine intelligence

LeCun, Y. A path towards autonomous machine intelligence. Open Review 62, 1–62 (2022)

2022
[5]

& Lillicrap, T

Hafner, D., Pasukonis, J., Ba, J. & Lillicrap, T. Mastering diverse control tasks through world models. Nature 640, 647–653 (2025)

2025
[6]

& Schmidhuber, J

Ha, D. & Schmidhuber, J. World Models. arXiv:1803.10122 [cs, stat] https://doi.org/10.5281/zenodo.1207631 (2018) doi:10.5281/zenodo.1207631

arXiv 2018
[7]

Zador, A. et al. Catalyzing next-generation Artificial Intelligence through NeuroAI. Nat Commun 14, 1597 (2023)

2023
[8]

& Friston, K

Parr, T., Pezzulo, G. & Friston, K. J. Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. (MIT Press, Cambridge, MA, USA, 2022)

2022

Show all 88 references

[9]

Whatever next? Predictive brains, situated agents, and the future of cognitive science

Clark, A. Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences 36, 181–204 (2013)

2013
[10]

The Predictive Mind

Hohwy, J. The Predictive Mind. (Oxford University Press, 2013)

2013
[11]

& Levin, M

Pezzulo, G. & Levin, M. Bootstrapping Life-Inspired Machine Intelligence: The Biological Route from Chemistry to Cognition and Creativity. Preprint at https://doi.org/10.48550/arXiv.2602.08079 (2026)

2026 doi arXiv
[12]

Barsalou, L. W. Grounded cognition. Annual Review of Psychology 59, 617–645 (2008)

2008
[13]

Glenberg, A. M. Embodiment as a Unifying Perspective for Psychology. Wiley Interdisciplinary Reviews 1, 586–596 (2010)

2010
[14]

& Moser, E

Buzsáki, G. & Moser, E. I. Memory, navigation and theta rhythm in the hippocampal-entorhinal system. Nat Neurosci 16, 130–138 (2013)

2013
[15]

Bottini, R. et al. Actionable cognitive maps for spatial and conceptual thinking. PsyArXiv Preprints (2026). 15

2026
[16]

& Castelfranchi, C

Pezzulo, G. & Castelfranchi, C. The Symbol Detachment Problem. Cognitive Processing 8, 115–131 (2007)

2007
[17]

The Brain from inside Out

Buzsáki, G. The Brain from inside Out. (Oxford University Press, 2019)

2019
[18]

Llinás, R. R. I of the Vortex: From Neurons to Self. vol. 50 (MIT press Cambridge, MA, 2001)

2001
[19]

Neural simulation of action: A unifying mechanism for motor cognition

Jeannerod, M. Neural simulation of action: A unifying mechanism for motor cognition. NeuroImage 14, S103–S109 (2001)

2001
[20]

Levinson, S. C. On the human ‘interaction engine’. in Roots of human sociality: Culture, cognition and interaction (eds Enfield, N. J. & Levinson, S. C.) 39–69 (Berg, Oxford, 2006)

2006
[21]

Pezzulo, G. et al. The body talks: Sensorimotor communication and its brain and kinematic signatures. Physics of Life Reviews 28, 1–21 (2019)

2019
[22]

Kato, S. et al. Global brain dynamics embed the motor command sequence of Caenorhabditis elegans. Cell 163, 656–669 (2015)

2015
[23]

& Pezzulo, G

Nuzzi, D., Mattia, M. & Pezzulo, G. Generating whole-brain neural activity and behavior through unified latent dynamics. 2026.06.05.730482 Preprint at https://doi.org/10.64898/2026.06.05.730482 (2026)

2026 doi
[24]

& Friston, K

Pezzulo, G., Parr, T. & Friston, K. The evolution of brain architectures for predictive coding and active inference. Philosophical Transactions of the Royal Society B 377, 20200531 (2022)

2022
[25]

& Corbetta, M

Pezzulo, G., Zorzi, M. & Corbetta, M. The secret life of predictive brains: what’s spontaneous activity for? Trends Cogn Sci 25, 730–743 (2021)

2021
[26]

Gardner, R. J. et al. Toroidal topology of population activity in grid cells. Nature 602, 123–128 (2022)

2022
[27]

& Fiete, I

Chaudhuri, R., Gerçek, B., Pandey, B., Peyrache, A. & Fiete, I. The intrinsic attractor manifold and population dynamics of a canonical cognitive circuit across waking and sleep. Nat Neurosci 22, 1512–1520 (2019)

2019
[28]

Redish, A. D. Vicarious trial and error. Nat Rev Neurosci 17, 147–159 (2016)

2016
[29]

Foster, D. J. Replay comes of age. Annual review of neuroscience 40, 581–602 (2017)

2017
[30]

Guardamagna, M. et al. Toroidal topology of grid-cell activity precedes spatial navigation during development. 2026.03.10.710908 Preprint at https://doi.org/10.64898/2026.03.10.710908 (2026)

2026 doi
[31]

& Tonegawa, S

Dragoi, G. & Tonegawa, S. Preplay of future place cell sequences by hippocampal cellular assemblies. Nature 469, 397–401 (2011)

2011
[32]

Moser, M.-B., Rowland, D. C. & Moser, E. I. Place cells, grid cells, and memory. Cold Spring Harb Perspect Biol 7, a021808 (2015)

2015
[33]

Shilnikov, A. L. & Maurer, A. P. The Art of Grid Fields: Geometry of Neuronal Time. Front. Neural Circuits 10, (2016)

2016
[34]

& Tank, D

Aronov, D., Nevers, R. & Tank, D. W. Mapping of a non-spatial dimension by the hippocampal–entorhinal circuit. Nature 543, 719–722 (2017)

2017
[35]

Veselic, S. et al. A cognitive map for value-guided choice in the ventromedial prefrontal cortex. Cell 188, 3259–3273 (2025)

2025
[36]

A., Miller, D

Park, S. A., Miller, D. S. & Boorman, E. D. Inferences on a multidimensional social hierarchy use a grid-like code. Nature neuroscience 24, 1292–1301 (2021)

2021
[37]

Bellmund, J. L. S., Gärdenfors, P., Moser, E. I. & Doeller, C. F. Navigating cognition: Spatial codes for human thinking. Science 362, eaat6766 (2018)

2018
[38]

Behrens, T. E. J. et al. What Is a Cognitive Map? Organizing Knowledge for Flexible Behavior. Neuron 100, 490–509 (2018)

2018
[39]

& Doeller, C

Bottini, R. & Doeller, C. F. Knowledge across reference frames: Cognitive maps and image spaces. Trends in cognitive sciences 24, 606–619 (2020)

2020
[40]

time cells

Kraus, B. J., Robinson, R. J., White, J. A., Eichenbaum, H. & Hasselmo, M. E. Hippocampal “time cells”: time versus path integration. Neuron 78, 1090–1101 (2013)

2013
[41]

Bongioanni, A. et al. Activation and disruption of a neural mechanism for novel choice in monkeys. Nature 591, 270–274 (2021). 16

2021
[42]

Viganò, S., Bayramova, R., Doeller, C. F. & Bottini, R. Mental search of concepts is supported by egocentric vector representations and restructured grid maps. Nature Communications 14, 8132 (2023)

2023
[43]

Bao, X. et al. Grid-like neural representations support olfactory navigation of a two- dimensional odor space. Neuron 102, 1066–1075 (2019)

2019
[44]

& Bottini, R

Giari, G., Vignali, L., Xu, Y. & Bottini, R. MEG frequency tagging reveals a grid-like code during attentional movements. Cell Reports 42, (2023)

2023
[45]

F., Theves, S

Eperon, A., Doeller, C. F., Theves, S. & Bottini, R. Action information is integrated into entorhinal representations of conceptual space and is reflected in eye movements. Plos Biology 24, e3003755 (2026)

2026
[46]

Viganò, S., Bayramova, R., Doeller, C. F. & Bottini, R. Spontaneous eye movements reflect the representational geometries of conceptual spaces. Proceedings of the National Academy of Sciences 121, e2403858121 (2024)

2024
[47]

Gibson, J. J. The Ecological Approach to Visual Perception. (Lawrence Erlbaum Associates, Inc, Mahwah, NJ, 1979)

1979
[48]

Cortical mechanisms of action selection: the affordance competition hypothesis

Cisek, P. Cortical mechanisms of action selection: the affordance competition hypothesis. Phil. Trans. R. Soc. B. 362, 1585–1599 (2007)

2007
[49]

& Cisek, P

Pezzulo, G. & Cisek, P. Navigating the Affordance Landscape: Feedback Control as a Process Model of Behavior and Cognition. Trends in Cognitive Sciences https://doi.org/10.1016/j.tics.2016.03.013 (2016) doi:10.1016/j.tics.2016.03.013

2016 doi
[50]

Goodale, M. A. & Milner, A. D. Separate visual pathways for perception and action. Trends in neurosciences 15, 20–25 (1992)

1992
[51]

& Burkhalter, A

Wang, Q., Gao, E. & Burkhalter, A. Gateways of Ventral and Dorsal Streams in Mouse Visual Cortex. J Neurosci 31, 1905–1918 (2011)

1905
[52]

Finlay, B. L. & Uchiyama, R. Developmental mechanisms channeling cortical evolution. Trends Neurosci 38, 69–76 (2015)

2015
[53]

Evolution of behavioural control from chordates to primates

Cisek, P. Evolution of behavioural control from chordates to primates. Philosophical Transactions of the Royal Society B 377, 20200522 (2022)

2022
[54]

Graziano, M. S. Ethological Action Maps: A Paradigm Shift for the Motor Cortex. Trends in cognitive sciences 20, 121–132 (2016)

2016
[55]

J., Battaglia-Mayer, A., Iannetti, G

Bufacchi, R. J., Battaglia-Mayer, A., Iannetti, G. D. & Caminiti, R. Cortico-spinal modularity in the parieto-frontal system: A new perspective on action control. Prog Neurobiol 231, 102537 (2023)

2023
[56]

Brooks, R. A. Intelligence without representation. Artificial Intelligence 47, 139–159 (1991)

1991
[57]

Mataric, M. J. Behaviour-based control: Examples from navigation, learning, and group behaviour. Journal of Experimental & Theoretical Artificial Intelligence 9, 323–336 (1997)

1997
[58]

& D’Ausilio, A

Fadiga, L., Craighero, L. & D’Ausilio, A. Broca’s area in language, action, and music. Ann N Y Acad Sci 1169, 448–458 (2009)

2009
[59]

& Sutton, R

Silver, D. & Sutton, R. S. Welcome to the era of experience. Google AI 1, 11 (2025)

2025
[60]

& Pezzulo, G

Proietti, R. & Pezzulo, G. An Integrative Theory of the Noradrenaline System: Meta-Control through Expected Information Gain. PsyArXiv (2025)

2025
[61]

Yu, A. J. & Dayan, P. Uncertainty, neuromodulation, and attention. Neuron 46, 681–692 (2005)

2005
[62]

Sara, S. J. The locus coeruleus and noradrenergic modulation of cognition. Nat Rev Neurosci 10, 211–223 (2009)

2009
[63]

Allostasis: a model of predictive regulation

Sterling, P. Allostasis: a model of predictive regulation. Physiology & behavior 106, 5–15 (2012)

2012
[64]

E., Quigley, K

Katsumi, Y., Theriault, J. E., Quigley, K. S. & Barrett, L. F. Allostasis as a core feature of hierarchical gradients in the human brain. Network Neuroscience 6, 1010–1031 (2022)

2022
[65]

Barrett, L. F. & Miller, E. K. Categorization is ‘baked’into the brain. Nature Reviews Neuroscience 1–22 (2026). 17

2026
[66]

Powers, W. T. Behavior: The Control of Perception. (Aldine, Hawthorne, NY, 1973)

1973
[67]

Maturana, H. R. & Varela, F. J. Autopoiesis and Cognition: The Realization of Living. (D. Reidel Pub., Dordrecht, Holland, 1980)

1980
[68]

Beyond the computer metaphor: Behavior as interaction

Cisek, P. Beyond the computer metaphor: Behavior as interaction. Journal of Consciousness Studies 6, 125–142 (1999)

1999
[69]

Tschantz, A., Seth, A. K. & Buckley, C. L. Learning action-oriented models through active inference. bioRxiv 764969 (2019) doi:10.1101/764969

2019 doi
[70]

Seth, A. K. & Friston, K. J. Active interoceptive inference and the emotional brain. Phil. Trans. R. Soc. B 371, 20160007 (2016)

2016
[71]

Barrett, L. F. & Simmons, W. K. Interoceptive predictions in the brain. Nat Rev Neurosci 16, 419–429 (2015)

2015
[72]

K., Suzuki, K

Seth, A. K., Suzuki, K. & Critchley, H. D. An Interoceptive Predictive Coding Model of Conscious Presence. Front Psychol 2, (2012)

2012
[73]

F., Quigley, K

Barrett, L. F., Quigley, K. S. & Hamilton, P. An active inference theory of allostasis and interoception in depression. Phil. Trans. R. Soc. B 371, 20160011 (2016)

2016
[74]

Barrett, L. F. The theory of constructed emotion: an active inference account of interoception and categorization. Soc Cogn Affect Neurosci 12, 1–23 (2017)

2017
[75]

Why do you fear the Bogeyman? An embodied predictive coding model of perceptual inference

Pezzulo, G. Why do you fear the Bogeyman? An embodied predictive coding model of perceptual inference. Cognitive, Affective, and Behavioral Neuroscience (2013)

2013
[76]

Crapse, T. B. & Sommer, M. A. Corollary discharge across the animal kingdom. Nat Rev Neurosci 9, 587–600 (2008)

2008
[77]

& Pennartz, C

Brucklacher, M., Pezzulo, G., Mannella, F., Galati, G. & Pennartz, C. M. A. Learning to segment self-generated from externally caused optic flow through sensorimotor mismatch circuits. Neural Networks 181, 106716 (2025)

2025
[78]

Gibson, J. J. The perception of the visual world. (1950)

1950
[79]

D., Blakemore, S

Frith, C. D., Blakemore, S. J. & Wolpert, D. M. Abnormalities in the awareness and control of action. Philos Trans R Soc Lond B Biol Sci 355, 1771–1788 (2000)

2000
[80]

Raviv, H. et al. The First 1000 Days: An Agent-Based Model of Early Language Acquisition. 2026.03.28.715023 Preprint at https://doi.org/10.64898/2026.03.28.715023 (2026)

2026 doi
[81]

Resynthesizing behavior through phylogenetic refinement

Cisek, P. Resynthesizing behavior through phylogenetic refinement. Attention, Perception, & Psychophysics 81, 2265–2287 (2019)

2019
[82]

Barrett, L. F. How Emotions Are Made: The Secret Life of the Brain. (Houghton Mifflin Harcourt, 2017)

2017
[83]

The Extended Mirror Neuron Network Anatomy, Origin, and Functions

Bonini, L. The Extended Mirror Neuron Network Anatomy, Origin, and Functions. Neuroscientist 1073858415626400 (2016) doi:10.1177/1073858415626400

2016 doi
[84]

& Craighero, L

Rizzolatti, G. & Craighero, L. The mirror-neuron system. Annual Review of Neuroscience 27, 169–192 (2004)

2004
[85]

& Sinigaglia, C

Rizzolatti, G. & Sinigaglia, C. The functional role of the parieto-frontal mirror circuit: interpretations and misinterpretations. Nature Reviews Neuroscience 11, 264–274 (2010)

2010
[86]

& Friston, K

Pezzulo, G., Parr, T. & Friston, K. J. Shared worlds, shared minds: Strategies to develop physically and socially embedded AI. EMBO reports 1–6 (2025) doi:10.1038/s44319-025- 00549-8

2025 doi
[87]

Pezzulo, G. et al. A predictive processing framework for joint action and communication. Phys Life Rev 57, 61–78 (2026)

2026
[88]

The Hardware Lottery

Hooker, S. The Hardware Lottery. Preprint at https://doi.org/10.48550/arXiv.2009.06489 (2020)

2009 doi arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

M., Ullman, T

Lake, B. M., Ullman, T. D., Tenenbaum, J. B. & Gershman, S. J. Building machines that learn and think like people. Behav Brain Sci 40, e253 (2017)

2017

[2] [2]

& Friston, K

Pezzulo, G., Parr, T., Cisek, P., Clark, A. & Friston, K. Generating meaning: active inference and the scope and limits of passive AI. Trends in Cognitive Sciences https://doi.org/10.1016/j.tics.2023.10.002 (2023) doi:10.1016/j.tics.2023.10.002

doi:10.1016/j.tics.2023.10.002 2023

[3] [3]

Taniguchi, T. et al. World Models and Predictive Coding for Cognitive and Developmental Robotics: Frontiers and Challenges. Advanced Robotics 37, (2023)

2023

[4] [4]

A path towards autonomous machine intelligence

LeCun, Y. A path towards autonomous machine intelligence. Open Review 62, 1–62 (2022)

2022

[5] [5]

& Lillicrap, T

Hafner, D., Pasukonis, J., Ba, J. & Lillicrap, T. Mastering diverse control tasks through world models. Nature 640, 647–653 (2025)

2025

[6] [6]

& Schmidhuber, J

Ha, D. & Schmidhuber, J. World Models. arXiv:1803.10122 [cs, stat] https://doi.org/10.5281/zenodo.1207631 (2018) doi:10.5281/zenodo.1207631

arXiv 2018

[7] [7]

Zador, A. et al. Catalyzing next-generation Artificial Intelligence through NeuroAI. Nat Commun 14, 1597 (2023)

2023

[8] [8]

& Friston, K

Parr, T., Pezzulo, G. & Friston, K. J. Active Inference: The Free Energy Principle in Mind, Brain, and Behavior. (MIT Press, Cambridge, MA, USA, 2022)

2022