Working Paper: Towards a Category-theoretic Comparative Framework for Artificial General Intelligence
| Source: ArXiv | Original article
A new working paper posted to arXiv (2603.28906v1, 29 March 2026) proposes the first systematic, category‑theoretic framework for comparing artificial general intelligence (AGI) architectures. Authored by Pablo de los Riscos, Fernando J. Corbacho and Michael A. Arbib, the manuscript argues that the field’s lack of a single formal definition hampers both scientific discourse and industry investment. Sections 3‑5 lay out three analytical layers—architectural, implementation and property‑based—each expressed as categorical objects and functors that map between design choices, hardware realizations and behavioural guarantees.
The proposal matters because AGI research is now a multi‑billion‑dollar race, yet progress is scattered across divergent models ranging from large‑scale transformer systems to neuromorphic and hybrid symbolic‑connectionist hybrids. A common mathematical language could make it possible to benchmark safety properties, scalability and alignment potential across these disparate approaches, reducing duplication and sharpening regulatory dialogue. Category theory’s track record in unifying concepts in machine learning and quantum computing suggests it can capture the compositional structure of cognition that many AGI blueprints implicitly rely on.
The next steps will test the framework against existing roadmaps such as the Mimosa multi‑agent system and the “first analyst” AI agents discussed earlier this month. Peer review, open‑source implementations on platforms like the CoLab repository, and citations in upcoming conference submissions will indicate whether the community adopts the formalism. If embraced, the framework could become a reference point for funding bodies, standards organisations and the next generation of AGI safety audits.
Sources
Back to AIPULSEN