pith. sign in

arxiv: 2604.11248 · v1 · submitted 2026-04-13 · 💻 cs.NE · cs.AI· cs.MA

Evolving Many Worlds: Towards Open-Ended Discovery in Petri Dish NCA via Population-Based Training

Uljad Berdica , Jakob Foerster , Frank Hutter , Arber Zela This is my paper

Pith reviewed 2026-05-10 15:14 UTC · model grok-4.3

classification 💻 cs.NE cs.AIcs.MA
keywords neural cellular automataopen-ended evolutionpopulation-based trainingself-organizationartificial lifeedge of chaosPetri dish NCA
0
0 comments X

The pith

Evolving populations of neural cellular automata with novelty and diversity rewards generates persistent lifelike patterns at the edge of chaos.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces PBT-NCA, a meta-evolutionary method that trains populations of Petri Dish Neural Cellular Automata under a composite objective rewarding historical behavioral novelty and current visual diversity while penalizing monocultures and dead states. This setup is intended to prevent the common collapse of such systems into frozen patterns or unstructured noise. The result is autonomous emergence of varied self-organizing behaviors that continue over long time scales. A sympathetic reader would care because sustained open-ended complexity arising purely from local interactions remains a central unsolved problem in artificial life.

Core claim

PBT-NCA evolves a population of PD-NCAs subject to a composite objective that rewards both historical behavioral novelty and contemporary visual diversity. By actively penalizing monocultures and dead states, the method drives the substrate to spontaneously generate emergent lifelike phenomena over extended horizons. These include highly regular coordinated periodic waves, spore-like scattering in which homogeneous groups eject cell-like clusters to colonize distant territories, and fluid shape-shifting macro-structures that migrate while maintaining stable outer boundaries enclosing highly active interiors. The system sustains effective complexity that is neither globally ordered nor random

What carries the argument

PBT-NCA, the population-based training algorithm applied to Petri Dish Neural Cellular Automata that uses a composite fitness function combining historical novelty, visual diversity, and penalties against uniform or extinct states to drive open-ended self-organization.

If this is right

  • The substrate autonomously discovers diverse morphological survival and self-organization strategies.
  • Coordinated periodic waves, spore-like scattering for colonization, and migrating shape-shifting structures emerge and persist.
  • Effective complexity is sustained in a state neither globally ordered nor globally random.
  • Multiple distinct behaviors coexist without one dominating the entire population.

Where Pith is reading between the lines

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

  • Similar evolutionary pressures could be tested on other differentiable multi-agent systems to check if open-endedness generalizes beyond this substrate.
  • The observed morphological strategies might serve as concrete targets for measuring progress toward open-ended artificial life.
  • Varying the relative weights of novelty versus diversity in the objective could reveal thresholds needed to keep dynamics at the edge of chaos.

Load-bearing premise

That a composite objective rewarding historical behavioral novelty and contemporary visual diversity, together with active penalization of monocultures and dead states, will reliably sustain effective complexity without the evolutionary process discovering loopholes that produce superficially diverse but ultimately uninteresting or unstable dynamics.

What would settle it

Long-term runs of PBT-NCA that show whether populations maintain high behavioral and visual diversity with ongoing lifelike dynamics or instead converge to repetitive frozen patterns, unstructured noise, or extinction despite the penalties.

Figures

Figures reproduced from arXiv: 2604.11248 by Arber Zela, Frank Hutter, Jakob Foerster, Uljad Berdica.

Figure 1
Figure 1. Figure 1: Overview of one PBT-NCA meta-iteration. Each [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: PBT-NCA evolving a population of 30 PD-NCA worlds, each with 7 NCA agents competing for territory. We plot the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Handcrafted behavior descriptor and novelty score. Each world is rolled out; the first [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual diversity score. At each timestep [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Shooters, Archipelago and Ant locomotion emerging from PBT-NCA (3 NCAs) at meta-iteration 20, 125 and 370. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Original PD-NCA (3 NCA agents) frames on itera [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Random Search (3 NCA agents) frames on meta [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Top figure: Mean novelty and composite score function over the number of meta-iterations. Bottom figure: Sampled hyperparameters over the number of meta-iterations for the best PBT-NCA (7 NCA agents). To match the compute budget of PBT-NCA, we use the same population size, P = 30, for RS. Each world was trained (rolled out) for a total of 6000 iterations (T = 500 meta￾iterations × Tworld = 12 iterations). … view at source ↗
Figure 10
Figure 10. Figure 10: Emergent dynamics at meta-iteration 145, 230, 310, 395 and 495 of PBT-NCA (7 NCA agents), demonstrating the [PITH_FULL_IMAGE:figures/full_fig_p007_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Rigid, circuit-like geometric patterns discovered with PBT-NCA (3 NCA agents) on the extended search space. We [PITH_FULL_IMAGE:figures/full_fig_p007_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Ecological Persistence (EP) and Effective Com [PITH_FULL_IMAGE:figures/full_fig_p008_12.png] view at source ↗
read the original abstract

The generation of sustained, open-ended complexity from local interactions remains a fundamental challenge in artificial life. Differentiable multi-agent systems, such as Petri Dish Neural Cellular Automata (PD-NCA), exhibit rich self-organization driven purely by spatial competition; however, they are highly sensitive to hyperparameters and frequently collapse into uninteresting patterns and dynamics, such as frozen equilibria or structureless noise. In this paper, we introduce PBT-NCA, a meta-evolutionary algorithm that evolves a population of PD-NCAs subject to a composite objective that rewards both historical behavioral novelty and contemporary visual diversity. Driven by this continuous evolutionary pressure, PBT-NCA spontaneously generates a plethora of emergent lifelike phenomena over extended horizons-a hallmark of true open-endedness. Strikingly, the substrate autonomously discovers diverse morphological survival and self-organization strategies. We observe highly regular, coordinated periodic waves; spore-like scattering where homogeneous groups eject cell-like clusters to colonize distant territories; and fluid, shape-shifting macro-structures that migrate across the substrate, maintaining stable outer boundaries that enclose highly active interiors. By actively penalizing monocultures and dead states, PBT-NCA sustains a state of effective complexity that is neither globally ordered nor globally random, operating persistently at the "edge of chaos".

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript presents PBT-NCA, a meta-evolutionary algorithm that evolves populations of Petri Dish Neural Cellular Automata (PD-NCA) using a composite objective rewarding historical behavioral novelty and contemporary visual diversity, along with penalties for monocultures and dead states. It reports that this leads to the autonomous discovery of diverse emergent phenomena including regular periodic waves, spore-like scattering for colonization, and fluid migrating macro-structures, which are presented as evidence of open-ended complexity sustained at the edge of chaos.

Significance. If the reported phenomena can be quantitatively validated, the work would be significant for the field of artificial life and evolutionary computation. It offers a practical method to overcome the hyperparameter sensitivity and collapse issues in PD-NCA by leveraging population-based training, potentially enabling more robust open-ended evolutionary systems. The empirical demonstration of self-discovered survival strategies adds to the literature on emergent complexity from local rules.

major comments (2)
  1. [Results] The central claims of open-ended discovery and persistent lifelike phenomena (e.g., waves, scattering, migrating structures) are supported only by qualitative observations and example images. No quantitative metrics, statistical tests, ablation studies, or controls (such as diversity measures over generations or comparisons to non-PBT baselines) are reported, which is necessary to distinguish sustained open-endedness from transient diversity.
  2. [Methods] The exact definition and implementation of the composite objective, including how historical behavioral novelty is measured and archived, the weights for each component, and the precise penalization of monocultures and dead states, are not provided in sufficient detail. This omission is load-bearing because it prevents assessment of whether the evolutionary process could exploit loopholes to produce superficially diverse but uninteresting dynamics.
minor comments (1)
  1. [Abstract] The abstract refers to 'extended horizons' without specifying the simulation timescales or number of evolutionary generations involved in the observations.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback and for acknowledging the potential significance of PBT-NCA for open-ended evolution. We address each major comment point by point below and have revised the manuscript to strengthen the presentation of results and methods.

read point-by-point responses

  1. Referee: [Results] The central claims of open-ended discovery and persistent lifelike phenomena (e.g., waves, scattering, migrating structures) are supported only by qualitative observations and example images. No quantitative metrics, statistical tests, ablation studies, or controls (such as diversity measures over generations or comparisons to non-PBT baselines) are reported, which is necessary to distinguish sustained open-endedness from transient diversity.

    Authors: We agree that quantitative support strengthens claims of sustained open-endedness. The original manuscript prioritized qualitative demonstration of autonomously discovered strategies (periodic waves, spore scattering, migrating macro-structures) as direct evidence of open-ended discovery from local rules. In the revised version we have added quantitative metrics: time-series plots of population-level behavioral diversity and novelty scores across generations, direct comparisons to non-PBT PD-NCA baselines showing faster collapse to low-complexity states, ablation studies removing individual objective components, and statistical summaries (means and variance) of persistence metrics over 10 independent runs. These additions appear in a new Results subsection and supplementary figures. revision: yes

  2. Referee: [Methods] The exact definition and implementation of the composite objective, including how historical behavioral novelty is measured and archived, the weights for each component, and the precise penalization of monocultures and dead states, are not provided in sufficient detail. This omission is load-bearing because it prevents assessment of whether the evolutionary process could exploit loopholes to produce superficially diverse but uninteresting dynamics.

    Authors: We have expanded the Methods section with the requested implementation details. The revised text now includes the precise formulation of the composite objective (weighted sum of historical novelty via a fixed-size archive of behavior embeddings, contemporary visual diversity via average pairwise latent-space distances, and explicit penalties for monoculture homogeneity and dead-state quiescence), the exact weights used, the archiving mechanism (periodic sampling and embedding of population states), and pseudocode for the full PBT-NCA loop. These clarifications allow direct evaluation of potential loopholes and ensure reproducibility. revision: yes

Circularity Check

0 steps flagged

No significant circularity: empirical outcomes of defined evolutionary pressure

full rationale

The paper introduces PBT-NCA as a meta-evolutionary algorithm whose composite objective explicitly rewards behavioral novelty, visual diversity, and penalizes monocultures/dead states. All reported phenomena (periodic waves, spore-like scattering, shape-shifting structures) are presented as observed empirical results of running this system over extended horizons, not as quantities derived by construction from the objective or from prior self-citations. No equations, uniqueness theorems, or fitted-parameter predictions appear in the abstract or described approach that would reduce the central claim to its inputs. The derivation chain is therefore self-contained as an experimental demonstration rather than a deductive or definitional loop.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on standard evolutionary computation assumptions plus the domain-specific claim that novelty-diversity pressure prevents collapse in PD-NCA substrates.

free parameters (1)
  • weights in composite novelty-diversity objective
    The abstract refers to a composite objective without specifying how novelty and diversity terms are balanced; these weights are free parameters that must be chosen or tuned.
axioms (1)
  • domain assumption Penalizing monocultures and dead states combined with novelty rewards will sustain persistent complexity at the edge of chaos
    Invoked in the description of PBT-NCA's effect on the substrate dynamics.

pith-pipeline@v0.9.0 · 5541 in / 1282 out tokens · 34267 ms · 2026-05-10T15:14:18.817345+00:00 · methodology

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

6 extracted references · 6 canonical work pages

  1. [1]

    G., Hoffman, M

    Andrychowicz, M., Denil, M., Colmenarejo, S. G., Hoffman, M. W., Pfau, D., Schaul, T., Shillingford, B., and de Freitas, N. (2016). Learning to learn by gradient descent by gradient descent. In Proceedings of the 30th International Conference on Neural In- formation Processing Systems, NeurIPS’16, page 3988–3996, Red Hook, NY , USA. Curran Associates Inc....

    work page 2016

  2. [2]

    Baym, M., Lieberman, T

    MIT Press One Rogers Street, Cambridge, MA 02142-1209, USA journals-info . . . . Baym, M., Lieberman, T. D., Kelsic, E. D., Chait, R., Gross, R., Yelin, I., and Kishony, R. (2016). Spatiotemporal microbial evolution on antibiotic landscapes.Science, 353(6304):1147–

    work page 2016

  3. [3]

    Bedau, M. A. (1996). Measurement of evolutionary activity, teleol- ogy, and life. Berto, F. and Tagliabue, J. (2012). Cellular automata.Stanford Encyclopedia of Philosophy. Bruce, J., Dennis, M., Edwards, A., Parker-Holder, J., Shi, Y ., Hughes, E., Lai, M., Mavalankar, A., Steigerwald, R., Apps, C., et al. (2024). Genie: Generative interactive environmen...

    work page arXiv 1996

  4. [4]

    Chan, B. W.-C. (2020). Lenia and expanded universe.Artificial Life, ALIFE 2020: The 2020 Conference on Artificial Life:221–

    work page 2020

  5. [5]

    Chan, B. W.-C. (2023). Towards large-scale simulations of open- ended evolution in continuous cellular automata. InProceed- ings of the Companion Conference on Genetic and Evolution- ary Computation, GECCO ’23 Companion, page 127–130, New York, NY , USA. Association for Computing Machinery. Chen, T., Chen, X., Chen, W., Wang, Z., Heaton, H., Liu, J., and ...

    work page arXiv 2023

  6. [6]

    Wang, R., Lehman, J., Clune, J., and Stanley, K. O. (2019). Paired open-ended trailblazer (poet): Evolvable evolutionary chal- lenges and their solutions through a generative environment model. InProceedings of the Genetic and Evolutionary Com- putation Conference, pages 142–151. White, C., Safari, M., Sukthanker, R. S., Ru, B., Elsken, T., Zela, A., Dey,...

    work page arXiv 2019