Orchestrated Reality: From Role-Play to Living, Playable Game Worlds -- LLM-Driven World Simulation as a Parameterized-Action POMDP

The article introduces a framework called "orchestrated reality," which utilizes a Parameterized-Action POMDP (Partially Observable Markov Decision Process) to create LLM-driven game worlds that integrate narrative and state management. This approach enables a single orchestration agent to maintain a persistent game state represented as a tree of canonical JSON entities, with actions defined by a discrete intent and structured parameters. The significance lies in its potential to enhance the autonomy and interactivity of game engines, allowing for more complex, player-driven narratives in sandbox and open-world environments, while addressing the limitations of current LLMs in sustaining validated game states.