A Metacognitive Architecture for Fast and Slow AI Reasoning

Summary

The paper examines why recent AI advances remain largely narrow despite progress in algorithms, data, and computing power. The authors argue that studying human cognitive mechanisms could help AI systems acquire capabilities associated with broader intelligence. Drawing on Daniel Kahneman's distinction between fast and slow thinking, they propose a multi-agent architecture with two types of problem solvers. Fast, or System 1, agents respond using past experience and are intended for routine decisions. Slow, or System 2, agents are deliberately activated when a problem requires reasoning, search, or an optimal solution beyond the fast agent's expectations. A metacognitive mechanism is therefore responsible for deciding when additional deliberation is needed. Both agent types use a world model containing domain knowledge about their environment and a self model recording the system's past actions and the skills of its solvers. The paper presents this design as a way to study and incorporate human-like metacognitive capabilities into AI, rather than as a reported benchmark or deployed system.