Agentic Physical AI for Nuclear Reactor Control

This paper proposes a novel approach to AI for physical systems, specifically nuclear reactor control, by introducing Agentic Physical AI. It argues that the prevailing paradigm of scaling general-purpose foundation models faces limitations in safety-critical control scenarios. The core idea is to prioritize physics-based validation over perceptual inference, leading to a domain-specific foundation model. The research demonstrates a significant reduction in execution-level variance and the emergence of stable control strategies through scaling the model and dataset. This work is significant because it addresses the limitations of existing AI approaches in safety-critical domains and offers a promising alternative based on physics-driven validation.