SEAD NCA

Spatiotemporal Evolution with Attractor Dynamics

Following von Neumann's vision of cellular automata — exploring computation as a physical process.

Papers

On the Spatiotemporal Dynamics of Generalization in Neural Networks

Zichao Wei · Saarland University · 2026

Three physics-inspired postulates — locality, symmetry, and stability — derive the SEAD architecture (neural cellular automaton). Validates on parity (perfect length generalization), addition (L=16 → L=1,000,000 at 100%), and Rule 110.

On the Mirage of Long-Range Dependency, with an Application to Integer Multiplication

Zichao Wei · Saarland University · 2026

Long-range dependency in multiplication is a mirage of computational spacetime. A 321-parameter neural cellular automaton on a 2D outer-product grid achieves 683× length generalization. Includes an interactive browser demo.

On the Emergence of Syntax by Means of Local Interaction

Zichao Wei · Saarland University · 2026

An 18,658-parameter 2D neural cellular automaton, trained with 1-bit boundary supervision on arithmetic expressions, spontaneously develops internal structure aligned with the CKY parsing algorithm (Pearson r ≈ 0.71). Includes an interactive browser demo.