Visualising the Attractor Landscape of Neural Cellular Automata

Alexander Mordvintsev; Harald Michael Ludwig; James Stovold; Mia-Katrin Kvalsund; Varun Sharma

At the whole-state level, Neural Cellular Automata learn simple behavioural manifolds; cell-level analysis reveals much more complex structure.

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 ↗

T0 review · grok-4.5

2026-07-12 22:25 UTC pith:4PV4OE5Q

load-bearing objection Abstract-only NCA interpretability note: useful macro/micro framing, but the attractor claim is unvalidated and the paper is not yet ready for a serious referee. the 2 major comments →

arxiv 2604.10639 v2 pith:4PV4OE5Q submitted 2026-04-12 cs.NE cs.ET

Visualising the Attractor Landscape of Neural Cellular Automata

James Stovold , Mia-Katrin Kvalsund , Harald Michael Ludwig , Varun Sharma , Alexander Mordvintsev This is my paper

classification cs.NE cs.ET

keywords neural cellular automatamanifold learningtopological data analysispersistent homologyautoencodersinterpretabilityattractor landscapeemergent behaviour

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.

Neural Cellular Automata (NCAs) can be trained to produce useful emergent behaviour, but what they have actually learnt remains hard to inspect. This paper argues that the right scale of analysis decides whether the black box opens easily. When each full grid state is treated as one data point, standard manifold-learning tools (PCA and autoencoders) and topological data analysis (persistent homology) recover a relatively simple behavioural manifold that can be visualised and interpreted. When the same tools are applied cell by cell, that manifold becomes highly complex and ordinary methods struggle. The practical claim is therefore scale-dependent: macroscopic views of NCA trajectories already give usable pictures of the attractor landscape the network has learnt, while microscopic views demand heavier machinery. If true, this supplies a concrete route to interpretability for NCAs as they move beyond toy Artificial Life models into real applications.

Core claim

When NCA analysis is performed at the macroscopic level (entire state as one data point), the underlying behavioural manifold is often simple and can be captured well by PCA, dense and sparse autoencoders, and persistent homology; at the microscopic (per-cell) level the manifold is highly complex and requires more sophisticated techniques.

What carries the argument

The behavioural manifold recovered by manifold learning (PCA, dense and sparse autoencoders) and topological data analysis (persistent homology) applied to NCA state trajectories, used to visualise the attractor landscape the NCA has learnt.

Load-bearing premise

That the chosen manifold-learning and topological tools applied to NCA states recover the true attractor landscape of what the model learnt, rather than artefacts of representation, sampling, or training dynamics.

What would settle it

For a trained NCA with a known simple target attractor, check whether macroscopic PCA or autoencoder embeddings recover that attractor topology while microscopic embeddings remain irreducibly complex; failure of the macroscopic recovery would falsify the claim.

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

If this is right

Macroscopic manifold analysis can serve as a practical interpretability tool for NCAs deployed outside toy models.
Designers can first inspect whole-state trajectories with simple linear methods before escalating to cell-level analysis.
Sparse autoencoders and persistent homology become the natural next tools when cell-level complexity must be resolved.
Training regimes that keep macroscopic manifolds simple may be preferred when post-hoc interpretation is required.

Where Pith is reading between the lines

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

The macro-simplicity result suggests NCAs may be compressing global dynamics into low-dimensional attractors even when local rules remain opaque, analogous to other emergent collective systems.
A natural follow-up is to test whether deliberately regularising the macroscopic manifold during training improves generalisation or controllability.
If the macro-micro gap holds across architectures, it could guide hybrid analysis pipelines that route global questions to cheap methods and local questions to heavy ones.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Abstract-only NCA interpretability note: useful macro/micro framing, but the attractor claim is unvalidated and the paper is not yet ready for a serious referee.

read the letter

This is an abstract-only look at arXiv:2604.10639, so treat everything below as provisional. The one thing worth knowing is the empirical contrast they report: whole-state NCA trajectories often sit on simple manifolds under PCA/autoencoders/persistent homology, while cell-level trajectories look highly complex. That macro-vs-micro distinction is the actual contribution relative to prior NCA work, and it is a practical observation people in the niche might use when deciding how to debug or visualise trained models.

What they do well is name a real problem—NCAs are useful but opaque—and apply a sensible toolkit (PCA, dense/sparse AEs, TDA) rather than inventing a new method. The framing is honest about “varying success.” Nothing in the abstract looks circular or self-contradictory; it is observational analysis of trained dynamics, not a fitted “prediction.”

The soft spot is load-bearing and currently unaddressed. The claim that these pipelines recover the attractor landscape of what the NCA has learnt, rather than sampling or representation artefacts, is asserted without validation criteria, controls (untrained/random NCAs, known simple systems), quantitative metrics, or ablations. From the abstract alone we also lack dataset description, figures, and any statement of free parameters for the manifold tools. The stress-test concern lands: without those checks the macro-simplicity / micro-complexity contrast could just be global correlation versus local variability. That does not make the paper bad; it makes the central result unproven at this stage.

Who it is for: people already working on NCA interpretability or ALife visualisation. A general ML reader will not get much. I would not bring the abstract to reading group, and I would not cite it yet. A serious editor should wait for the full paper with methods, controls, and numbers before spending referee time; on the abstract alone I would desk-reject or request the complete manuscript. If the full version ships those validations (and ideally code), the same framing becomes a useful, citable methods note. Right now it is a promising sketch, not a finished result.

Referee Report

2 major / 2 minor

Summary. This paper applies manifold-learning methods (PCA, dense and sparse autoencoders) and topological data analysis (persistent homology) to Neural Cellular Automata in order to recover and visualise their behavioural attractor landscapes. The central empirical claim is that, when the entire NCA state is treated as a single data point (macroscopic analysis), the underlying manifold is often relatively simple and can be captured well, whereas when individual cell states are treated as data points (microscopic analysis) the manifold is highly complex and requires more sophisticated techniques. The work is motivated by the growing use of NCAs beyond toy Artificial Life settings and the corresponding need for interpretability of what trained NCAs have learnt.

Significance. If the macro-simplicity versus micro-complexity distinction is rigorously established, the paper would supply a practical guide for NCA interpretability: global behavioural questions may often be answerable with comparatively simple embeddings, while cell-level questions demand heavier machinery. The multi-technique pipeline and the explicit framing of NCAs as systems whose learnt dynamics should be opened rather than left opaque are strengths in principle. Significance remains provisional until quantitative metrics, controls, and full experimental detail are available to confirm that the pipelines recover genuine attractors rather than artefacts.

major comments (2)

[Abstract (central empirical claim)] The load-bearing claim that macroscopic behavioural manifolds are 'often quite simple' while microscopic ones are 'highly complex' rests on the unvalidated premise that PCA, dense/sparse autoencoders and persistent homology recover the true attractor landscape of the trained NCA rather than sampling, representation or training artefacts. The abstract supplies no validation criteria, ground-truth comparisons or controls (e.g. untrained/randomised NCAs, or known simple dynamical systems). Without those, the reported distinction could equally reflect global state correlations versus local cell variability, embedding-dimension choices, or trajectory sampling.
[Abstract (results phrasing)] Results are characterised only as 'with varying success' and 'often quite simple,' with no quantitative metrics, error bars, baselines, technique ablations or dataset description. For a central claim about manifold simplicity versus complexity, the absence of stated success criteria or comparative measures renders the claim non-falsifiable from the available text and prevents assessment of whether the pipelines actually support the macro/micro contrast.

minor comments (2)

[Abstract] Phrases such as 'with varying success' and 'often' are too vague for a results summary; once full results are present they should be replaced by concrete reconstruction errors, persistence statistics or success rates.
[Abstract] The abstract does not name the NCA tasks, architectures or training regimes studied; even a brief indication would help readers judge the scope of the claimed macro/micro distinction.

Circularity Check

0 steps flagged

No significant circularity: observational analysis of NCA dynamics with no derivation that reduces predictions to fitted inputs by construction.

full rationale

Only the abstract is available. It describes applying manifold-learning (PCA, dense/sparse autoencoders) and TDA (persistent homology) to trained NCA states and reporting that whole-state manifolds are often simple while cell-level manifolds are complex. There are no equations, no fitted parameters re-used as predictions, no uniqueness theorems, and no load-bearing self-citations that define the claimed result. The work is observational visualisation rather than a first-principles derivation; success is not equated by construction to the inputs of the analysis pipeline. Residual concerns about whether the pipelines recover true attractors versus artefacts are correctness/validation risks, not circularity. Score 0 with empty steps is the honest finding for this abstract-only review.

Axiom & Free-Parameter Ledger

2 free parameters · 3 axioms · 0 invented entities

Abstract-only: free parameters of the underlying NCA training and of the analysis models (autoencoder widths, persistence thresholds, PCA cutoffs) are not stated. Domain assumptions are standard for NCA and manifold learning. No new physical entities are invented; the 'behavioural manifold' is an analytical construct, not a postulated particle or force. Ledger is necessarily incomplete without methods sections.

free parameters (2)

NCA training and architecture hyperparameters
Abstract does not specify channel counts, update rules, loss terms, or training schedules; any quantitative manifold geometry depends on these choices.
Manifold-analysis hyperparameters (PCA rank, AE latent dim, persistence thresholds)
Success of capture of the manifold depends on these analysis knobs, which are not given in the abstract.

axioms (3)

domain assumption Trained NCA dynamics admit a meaningful low-dimensional behavioural manifold that manifold-learning and TDA tools can recover from sampled states.
Load-bearing for interpreting PCA/AE/PH outputs as the attractor landscape of what was learnt.
domain assumption Treating the full grid state vs individual cell states as data points is a valid macroscopic vs microscopic decomposition of NCA behaviour.
Defines the central macro/micro contrast reported in the abstract.
standard math Standard constructions of PCA, dense/sparse autoencoders, and persistent homology apply without domain-specific modification.
Background mathematical tools assumed available and correctly used.

pith-pipeline@v1.1.0-grok45 · 6127 in / 2502 out tokens · 23847 ms · 2026-07-12T22:25:39.088154+00:00 · methodology

0 comments

read the original abstract

As Neural Cellular Automata (NCAs) are increasingly applied outside of the toy models in Artificial Life, there is a pressing need to understand how they behave and to build appropriate routes to interpret what they have learnt. By their very nature, the benefits of training NCAs are balanced with a lack of interpretability: we can engineer emergent behaviour, but have limited ability to understand what has been learnt. In this paper, we apply a variety of techniques to pry open the NCA black box and glean some understanding of what it has learnt to do. We apply techniques from manifold learning (principal components analysis and both dense and sparse autoencoders) along with techniques from topological data analysis (persistent homology) to capture the NCA's underlying behavioural manifold, with varying success. Results show that when analysis is performed at a macroscopic level (i.e. taking the entire NCA state as a single data point), the underlying manifold is often quite simple and can be captured and analysed quite well. When analysis is performed at a microscopic level (i.e. taking the state of individual cells as a single data point), the manifold is highly complex and more complicated techniques are required in order to make sense of it.