Cognito: A Metacognitive Neuro-Symbolic Architecture for Verifiable Artificial General Intelligence

Abstract

The field of Artificial Intelligence (AI) is in a "third summer," driven by the remarkable success of large-scale sub-symbolic models, yet simultaneously confronted by their inherent limitations in robustness, explainability, and systematic reasoning.1 Neuro-Symbolic AI (NeSy) has emerged as the prevailing paradigm to address these gaps, seeking to merge pattern-based learning with logical reasoning. However, this paper posits that current NeSy research, despite its progress, overlooks a critical element for genuine autonomy and reliability: metacognition. We introduce Cognito, a novel architecture that places a Metacognitive Controller at its core. This controller is designed to monitor, evaluate, and dynamically regulate the interaction between a high-efficiency sub-symbolic substrate and a verifiable symbolic reasoning engine. Cognito's main contributions are: (1) a formal model for metacognitive control in AI, addressing a quantified gap in the literature; (2) a path toward verifiability-by-design in complex AI systems using dependently typed programming languages; and (3) a foundational framework for the development of a more robust, adaptable, and trustworthy Artificial General Intelligence (AGI).