Neural cellular automata reveal transient states in AI growth from single seed

A new study from researchers Hiroki Sato, Atsushi Masumori, and Takashi Ikegami (arXiv:2607.15726) dives deep into how Growing Neural Cellular Automata (GNCA) develop complex morphologies from a single seed cell using shared local rules. By tracing the full developmental trajectory of trained GNCA models, they discovered that morphological convergence does not happen incrementally. Instead, it proceeds non-monotonically through transient intermediate configurations. Channel-wise analysis revealed that hidden channels self-organize into modular groups in parallel with visible form development. Geometric analysis of the cell state space showed that cell states diversify within a low-dimensional, smooth manifold.

To further investigate cell development, the team applied community detection on an ε-neighbor network of cells. This analysis successfully extracted discrete cell types from the continuous state space, identifying transient cell-type communities during early development and stable, finer-grained types corresponding to spatially coherent regions of the mature morphology. The temporal coordination across multiple independent measures indicates that GNCA development is a reorganization of transient states rather than incremental refinement. These findings have implications for understanding self-organizing AI systems and could inspire new approaches to morphogenetic engineering and developmental AI.