Agentic Physical AI for Nuclear Reactor Control
A novel concept known as Agentic Physical AI has been introduced for managing nuclear reactors. The existing trend of enhancing general-purpose foundation models for universal multimodal reasoning encounters a significant obstacle at the control interface. Recent evaluations indicate that leading vision-language models attain merely 50–53% accuracy on fundamental quantitative physics tasks, functioning as rough estimators that maintain semantic coherence while breaching physical laws. This lack of input fidelity is not merely a scaling issue but a structural one: architectures focused on perception prioritize parameter-space imitation, while safety-critical control requires guarantees on the outcomes of actions taken. The proposed method features compact language models as Agentic Physical AI, where policy optimization relies on physics-based validation instead of perceptual inference. The study develops a 3-parameter model, illustrating a distinctly different approach to domain-specific foundation models.
Key facts
- Agentic Physical AI is proposed for nuclear reactor control.
- Frontier vision-language models achieve only 50–53% accuracy on basic quantitative physics tasks.
- The limitation is structural, not a scaling deficiency.
- Policy optimization is driven by physics-based validation.
- A 3-parameter model is trained.
- The approach targets domain-specific foundation models.
- Safety-critical control requires outcome-space guarantees.
- Perception-centric architectures optimize parameter-space imitation.
Entities
Institutions
- arXiv