REVIEW 3 major objections 5 minor 1 cited by
Neural cellular automata: applications to biology and beyond classical AI
T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash
Pith's one-line read This review argues that neural cellular automata constitute a unifying, computationally lean paradigm spanning biological self-organization, distributed robotic control, and abstract reasoning, potentially enabling a new class of bioinspire
desk verdict A useful survey of NCAs in biology, but the flagship ARC-AGI-1 claim is carried on preprints the paper never verifies. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the neural cellular automaton (NCA): each cell on a discrete grid keeps a continuous vector state and updates it using a shared, small neural network that sees only the local neighborhood—normally a 3×3 convolution plus a dense layer, on the order of 10,000 parameters. Training is done by gradient descent or evolution on global losses, e.g., matching a target image or solving a maze, while stochastic updates and damage-masking during training encourage regeneration. The paper also highlights architectural extensions: private engram channels for memory and transfer, hierarchical stacked NCAs for multi-scale coupling, and criticality-pretrained cells. These make the NCA a
What would settle it
Independently reproduce the ARC-NCA and EngramNCA runs on a held-out subset of ARC-AGI-1 under the stated training budget and compare with a large language model baseline; if the performance gap is much smaller or the cost estimate is off by orders of magnitude, the strongest claim fails. Alternatively, show that NCA-trained morphologies cannot regenerate after damage when scaled beyond toy grid sizes, contradicting the claimed scalability.
Extended reading notes
Core claim
The core claim is that the NCA architecture—a spatial grid in which every cell runs the same feed-forward neural network update rule on local neighbor states—is not just a tool for pattern formation but a substrate for collective intelligence. The authors survey evidence that these systems self-assemble and repair morphologies, store and transfer genetic information through private cell-state channels, regenerate 3D machines, and even grow solutions to ARC-AGI-1 tasks from a few examples. They argue that the same iterative, locally constrained refinement that makes NCAs robust also links them to denoising diffusion models, and that the architecture's built-in multiscale competency makes it a
Load-bearing premise
The whole synthesis depends on the accuracy and representativeness of the primary results it cites—especially the reported ARC-AGI-1 performance of NCAs and the aging/regeneration simulations—since the paper reviews rather than reproduces them.
Editorial extensions
If this is right
- If NCAs remain competitive on ARC-AGI-1 at a fraction of the cost, abstraction and reasoning tasks may be approachable with much smaller, decentralized models.
- NCA-based models of regeneration and aging suggest that tissue-level goal-directedness, not just cellular damage, is a key variable in decline—implying targeted interventions can reactivate dormant regenerative potential.
- Since NCAs double as distributed controllers for soft and voxel-based robots, the same learned morphogenetic rules could give machines self-repair and adaptation capabilities.
- The structural parallel between NCA refinement and diffusion denoising suggests a path to hybrid generative models that keep spatial locality and parameter efficiency.
- A biology-inspired AI built on NCAs would be modular and interpretable with the tools of neuroscience and psychology, potentially easing alignment problems.
Reading between the lines
- An independent replication of the ARC-NCA and EngramNCA runs on a held-out ARC-AGI-1 subset would sharply test the strongest claim; the review does not audit those results itself.
- If the diffusion analogy is more than superficial, NCA training could benefit from explicit time-conditioning or noise-schedule curricula without giving up locality—a testable design change.
- The multiscale-competency perspective predicts that NCA-like systems will show better damage tolerance and transfer when trained with hierarchical or private-state architectures; one could measure this directly against vanilla NCAs.
- The toy-model caveat suggests the field needs benchmarks at organ-level complexity before the unification claim is secure.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This paper is a review/position paper on Neural Cellular Automata (NCA) in biology and AI. It surveys recent work on NCAs for morphogenesis, regeneration, aging, bioelectricity, evolution, the generative genome, molecular design, and robotics, and argues that NCAs instantiate a 'multiscale competency architecture' analogous to biological organization. It further claims that NCAs go beyond classical AI: they show robustness, generalization, decentralized control, and even competitive performance on ARC-AGI-1 abstraction/reasoning tasks (compared with ChatGPT-4.5) at a fraction of cost. The paper connects NCA dynamics to diffusion models and criticality, and concludes by advocating NCAs as a unifying 'computationally lean' paradigm for bio-inspired collective intelligence. The review is broad and cites extensive literature, but its flagship quantitative claims about ARC are not substantiated within the manuscript and rest on arXiv preprints [24,25]; the paper's own limitations section acknowledges NCAs are toy models with unsolved scaling and storage issues.
Significance. If the reported ARC-AGI-1 results and the generalization-from-few-examples claims were verified, the paper would point to a substantial result: NCAs as a low-cost, decentralized alternative to large transformer models on abstract reasoning. The review also usefully consolidates the diverse NCA literature and gives credit to a self-contained list of applications and open problems, including criticality and hybrid evolution/diffusion training. However, as a review it contributes no new experiments or formal analysis; its significance therefore rests on the accuracy of cited prior work and on the coherence of the 'multiscale competency' framing. The authors are explicit about many limitations, which is commendable and should be retained.
major comments (3)
- [§4 (ARC-NCA paragraph) and Abstract] The sentence 'vanilla NCAs and EngramNCAs perform competitively on the ARC challenge compared to ChatGPT-4.5 – notably at the fraction of costs and computational resources' is a quantitative, load-bearing claim, but no numbers or protocols are given. The manuscript does not report solve rates, task subset, number of demonstrations, inference budget, or what 'competitive' means in this context; it only cites refs [24,25], both arXiv preprints. Likewise, 'can learn to generalize ... from a minimal set of two or three training examples' is stated without supporting evaluation. Since the Abstract names ARC-AGI-1 as the main evidence for NCAs 'beyond classical AI', this unverified external dependence must be resolved. At minimum, state the reported accuracies and evaluation conditions, or explicitly label the claim as reported by the cited preprints with appropriate caveats.
- [§5 vs. §6] The limitations section concedes NCAs 'represent toy-models for biological organization', that simulating realistic organ-level complexity at unicellular resolutions is 'currently infeasible', and that integrating molecular/genetic/biomechanical detail is difficult. The conclusion nevertheless calls NCAs 'a model of choice' for AI-oriented computational biology and the basis of a 'unifying computationally lean paradigm'. A review may advocate in spite of limitations, but the manuscript should specify the intended scope (e.g., conceptual models of collective dynamics rather than quantitative organ simulation) and explain why the identified gaps do not prevent the stated 'model of choice' status. As written, the conclusion outruns the evidence assembled in the paper.
- [§1, §4: multiscale competency architecture] The 'multiscale competency architecture' is the central interpretive frame, but it is never defined operationally. The reader cannot tell whether a given NCA 'has' a competency at a scale, how competencies are measured, or how hierarchical NCA layers (refs [10,11]) implement 'competency amplification' as opposed to a generic multi-scale model. Since the unifying claim depends on this equivalence, the authors should provide a concrete criterion or a worked example, mapping one NCA architecture cell-by-cell to homeostatic loops at each level. Otherwise the central thesis remains resistant to evaluation.
minor comments (5)
- [Section 2, equations] The mathematical notation is garbled in the provided text (e.g., 's!"#$=f/s!",{s%"}%∈𝒩(𝒾)2' and 'x!"#$=x!"+Δx!""'). Please check the typesetting and define all symbols consistently.
- [Abstract and throughout] Typos and OCR artifacts: 'is constraint to' should be 'is constrained to'; 'William Jame's' should be 'William James's'; reference list contains 'diWerentiable', 'parameter-eWicient', 'DiWusion' instead of 'Differentiable', 'parameter-efficient', 'Diffusion'.
- [Section 2] The sentence 'the NCA's grid stats serve as recurrent feedback signal' should read 'grid states'. Also, the claim that dynamics are 'differentiable across temporal state updates' should be reconciled with the asynchronous/stochastic updates described later in the same section.
- [Figure 1] Some panels (e.g., K, referencing ARC-AGI-1 tasks) are cited without a quantitative caption. Consider adding one line describing what is shown and the reported performance.
- [Section 5] The paragraph on modularity/interfacing is hard to parse: 'this limits modularity and compatibility across multiple NCAs that operate in the same environment but developed not necessarily compatible communication strategies'. Please rephrase for clarity.
Circularity Check
No significant circularity: this is a literature review that reports external results; the ARC-AGI-1 claim is a citation to primary preprints, not a fitted prediction or definitionally forced result.
full rationale
This paper is a review/position paper, not a derivation: it contains no equations that are fit to data and then re-predicted, and no parameter is defined in terms of the quantity it is used to explain. The central 'unifying paradigm' claim is an interpretive synthesis of existing NCA results, including work from the authors' own laboratory (e.g., refs 7, 9, 14, 17, 24, 26, 67). Self-citation is present but not circular in the technical sense: the cited items are separate primary studies with their own benchmarks, and key evidence such as the ARC-AGI-1 comparison is attributed to two preprints (refs 24 and 25), one of which is external to this author group; the claim is thus a reported literature result, not a conclusion manufactured by the review's own construction. The paper also explicitly acknowledges its limits, calling NCAs 'toy-models' (Section 5) and stating that mapping synthetic cell states to biological counterparts remains a 'critical missing link' (Section 5), which undercuts any reading that the framework is forced by definition. The ARC comparison's unverified status is a correctness/evidence concern, not a circularity concern.
Assumptions & free parameters
assumptions (3)
- domain assumption Biological organization is correctly described by a multiscale competency architecture in which nested agents maintain homeostatic goals.
- domain assumption The cited NCA results, notably ARC-AGI-1 performance (refs 24, 25) and aging/regeneration simulations (refs 14, 67), are accurate and representative.
- domain assumption Cells in biological tissues can be approximated as homogeneous units with vector-valued states updated by local rules.
Cite this review
Pith. "Pith review of Neural cellular automata: applications to biology and beyond classical AI." pith.science (2026). https://pith.science/paper/EZ5ILNMB
@misc{pith2026250911131,
author = {Pith},
title = {Pith review of: Neural cellular automata: applications to biology and beyond classical AI},
year = {2026},
howpublished = {\url{https://pith.science/paper/EZ5ILNMB}},
note = {Machine review of arXiv:2509.11131}
}
read the original abstract
Neural Cellular Automata (NCA) represent a powerful framework for modeling biological self-organization, extending classical rule-based systems with trainable, differentiable (or evolvable) update rules that capture the adaptive self-regulatory dynamics of living matter. By embedding Artificial Neural Networks (ANNs) as local decision-making centers and interaction rules between localized agents, NCA can simulate processes across molecular, cellular, tissue, and system-level scales, offering a multiscale competency architecture perspective on evolution, development, regeneration, aging, morphogenesis, and robotic control. These models not only reproduce biologically inspired target patterns but also generalize to novel conditions, demonstrating robustness to perturbations and the capacity for open-ended adaptation and reasoning. Given their immense success in recent developments, we here review current literature of NCAs that are relevant primarily for biological or bioengineering applications. Moreover, we emphasize that beyond biology, NCAs display robust and generalizing goal-directed dynamics without centralized control, e.g., in controlling or regenerating composite robotic morphologies or even on cutting-edge reasoning tasks such as ARC-AGI-1. In addition, the same principles of iterative state-refinement is reminiscent to modern generative Artificial Intelligence (AI), such as probabilistic diffusion models. Their governing self-regulatory behavior is constraint to fully localized interactions, yet their collective behavior scales into coordinated system-level outcomes. We thus argue that NCAs constitute a unifying computationally lean paradigm that not only bridges fundamental insights from multiscale biology with modern generative AI, but have the potential to design truly bio-inspired collective intelligence capable of hierarchical reasoning and control.
Forward citations
Cited by 1 Pith paper
-
Emulating Cosmic Structure Formation with a Lagrangian Neural Cellular Automaton
A ~4,000-parameter recurrent local network correcting the Zeldovich approximation reaches percent-level matter power-spectrum accuracy at k≲0.5 h/Mpc at z=0, matching larger U-Net emulators on Quijote N-body tests.
Reference graph
Works this paper leans on
-
[21]
Pontes-Filho, S., et al., A Unified Substrate for Body-Brain Co-evolution. CoRR, 2022. abs/2203.12066. 22. Sudhakaran, S., et al., Growing 3D Artefacts and Functional Machines with Neural Cellular Automata, in The 2021 Conference on Artificial Life. 2021, MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA journals-info …. p. 108. 23. Variengien, A.,...
arXiv 2022
-
[42]
IEEE Transactions on Cognitive and Developmental Systems, 2017
Nichele, S., et al., Ca-neat: evolved compositional pattern producing networks for cellular automata morphogenesis and replication. IEEE Transactions on Cognitive and Developmental Systems, 2017. 10(3): p. 687-700. 43. Wolfram, S., Cellular automata. Los Alamos Science, 1983: p. 09-01. 44. ElZayyat, M. and D. Mušić. Cellular Automata as a Tool in Artificia...
arXiv 2017
-
[65]
Gumuskaya, G., et al., Motile Living Biobots Self-Construct from Adult Human Somatic Progenitor Seed Cells. Adv Sci (Weinh), 2024. 11(4): p. e2303575. 66. Stovold, J. Neural cellular automata can respond to signals. MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA journals-info …. 67. Hansali, S., et al., The role of bioelectrical patterns in re...
arXiv 2024
-
[87]
Allostasis, homeostasis, and the costs of physiological adaptation, 2004
Sterling, P ., Principles of allostasis: optimal design, predictive regulation, pathophysiology, and rational. Allostasis, homeostasis, and the costs of physiological adaptation, 2004. 17. 88. Bongard, J. and M. Levin, Living things are not (20th century) machines: updating mechanism metaphors in light of the modern science of machine behavior. Frontiers ...
arXiv 2004
-
[109]
Pigozzi, F ., A. Goldstein, and M. Levin, Associative conditioning in gene regulatory network models increases integrative causal emergence. Commun Biol, 2025. 8(1): p. 1027. 110. Biswas, S., W. Clawson, and M. Levin Learning in Transcriptional Network Models: Computational Discovery of Pathway-Level Memory and EWective Interventions. International Journa...
-
[129]
Paolo, G., et al., TAG: A Decentralized Framework for Multi-Agent Hierarchical Reinforcement Learning. CoRR, 2025. abs/2502.15425. 130. Deshpande, R., et al., Engineering morphogenesis of cell clusters with diWerentiable programming. Nat Comput Sci, 2025. 131. Podell, D., et al., SDXL: Improving Latent DiWusion Models for High-Resolution Image Synthesis. ...
arXiv 2025
Reviewed August 4, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.