Determine whether emergent structure is caused by clone-and-mutate inheritance or network capacity

Determine whether the structured diversity produced by the self-replicating, non-uniform Neural Cellular Automata substrate arises from clone-and-mutate inheritance itself or from spare capacity introduced by using larger per-cell neural networks.

Background

Each cell in the substrate contains a small neural network that reads its depth-wise Moore neighborhood and produces the cell’s next state. The reported experiments use a hidden width of two, yielding a 44-parameter agent, which the paper presents as a minimal architecture capable of nonlinear mappings.

The authors explicitly leave unresolved whether the observed structured diversity is a consequence of the clone-and-mutate inheritance mechanism or whether it would instead be attributable to additional representational capacity in larger agents. They later note that persistence of the reported dynamics at hidden widths of four or eight was not tested.

References

Two hidden units are the smallest layer that still admits a non-linear, non-separable map from the 18-dimensional neighborhood to the next state, and the aim is to establish that clone-and-mutate inheritance alone produces structured diversity in the minimal setting; with a larger agent it would remain open whether the structure came from the mechanism or from spare capacity.

— Self-Replicating Neural Cellular Automata: Quantifying Emergent Phenotypic and Genotypic Diversity in an OpenEnded Substrate  (2609.19902 - Jain et al., 17 Sep 2026) in Section 3.1.2, Agent Architecture