Towards a Balanced Metacognitive Model for Artificial Intelligence: A Hybrid and Hierarchical Architecture

Abstract

Metacognition, or the ability of a system to reason about its own cognitive processes, is increasingly recognized as an essential component for the development of robust, adaptable, and safe artificial intelligence (AI) systems. However, current computational approaches to metacognition remain fragmented, divided between classic symbolic architectures, which offer explicit control but can be brittle, and modern sub-symbolic models, which demonstrate flexible learning but lack transparency and reliable self-assessment mechanisms. This paper addresses this "Metacognitive Gap" by proposing a balanced path forward: a Hybrid and Hierarchical Metacognitive Architecture (HHMA). The HHMA is a novel framework that integrates principles from cognitive psychology and computer science to create a multi-level system. The lowest level of the architecture (Level 1) implements probabilistic monitoring, inspired by Bayesian models of human metacognition, to generate fast, continuous signals of confidence and uncertainty about the performance of the core cognitive system (Level 0). The highest level (Level 2) uses a symbolic reasoning engine, informed by a declarative model of the agent itself, to perform explicit diagnostics and exert strategic control. It is argued that this hybrid structure, which computationally implements dual-process theory, reconciles the strengths of competing paradigms. By combining sub-symbolic monitoring with symbolic control, the HHMA offers a promising path to improve the adaptability, explainability, and safety of AI, addressing some of the most pressing challenges in the field of assured autonomy and human-machine collaboration. 1. Hierarchical Separation of Concerns: It is recognized that different computational paradigms are better suited for different types of tasks. The HHMA assigns these tasks to distinct layers. Learning from high-dimensional, noisy data is relegated to sub-symbolic components, while deliberate reasoning and strategic decision-making are managed by symbolic components. 2. Probabilistic Monitoring: Uncertainty is an inescapable feature of cognition and