Model Validation of Agentic AI Systems: A POMDP-Based Framework for Belief-State, Forecast, and Policy Validation

The paper introduces a POMDP-based model validation framework for agentic AI systems, addressing the limitations of existing methodologies that focus solely on predictive accuracy. It decomposes decision-making into components such as information, beliefs, forecasts, actions, and utility, allowing for independent validation of each. The framework is exemplified through a portfolio-management case study, demonstrating that latent-state inference significantly enhances decision quality, thereby providing a structured approach to model risk management in autonomous AI systems.