Architectural expressivity and comparison of AGI architectures

Establish a rigorous notion of expressivity for agent architectures in the category ArchAgents by defining comparison relations (such as simulations, expressivity preorders, or embeddings) that order architectures according to the classes of agents or behaviors they can express, and determine whether some architectures are universal or conservative extensions of others.

Background

The paper proposes a category-theoretic framework in which architectures are presented as hypergraph categories and concrete agents are their semantic interpretations. A key long-term goal is to compare architectures in a principled way beyond ad hoc benchmarks.

Within this framework, the authors highlight that understanding and ordering architectures by the behaviors they can express is essential for a unifying comparative theory of AGI architectures. They explicitly identify this as a central open problem and suggest directions such as defining simulations, expressivity preorders, and investigating universality or conservative extensions.