Working Paper: Towards a Category-theoretic Comparative Framework for Artificial General Intelligence

A team of researchers has published a forward-looking working paper titled 'Towards a Category-theoretic Comparative Framework for Artificial General Intelligence' on arXiv. Authored by Pablo de los Riscos, Fernando J. Corbacho, and Michael A. Arbib, the 37-page document addresses a critical gap in the AI field: the lack of a formal, mathematical foundation for comparing the myriad architectures vying to achieve AGI. The paper argues that while major tech companies invest billions, the field relies on empirical benchmarks without a unified language to structurally analyze systems like Reinforcement Learning (RL), Universal AI, and Active Inference.

The core proposal is to use Category Theory—a branch of abstract algebra—to model AGI architectures as 'Machines in a Category.' This approach provides a rigorous, algebraic framework to unambiguously expose the commonalities and differences between architectural paradigms. The authors present initial exercises applying this framework to RL, Causal RL, and Schema-based Learning (SBL), demonstrating how it can characterize architectural structure, agent-environment interaction, and behavioral development over time.

This work is positioned as the first step in a broader research program. The ultimate goal is to create a unified formal foundation that integrates architectural design, informational organization, and semantic properties of agents. The framework is intended to support the definition and assessment of both syntactic and semantic properties within explicitly characterized environments, moving the AGI conversation from empirical testing to principled, mathematical comparison and design.