The study introduces the idea of true worlds—detailed simulations where past events shape future outcomes—to test model fidelity and shows that classical models suffer predictable average reward losses along reachable trajectories.

Classical digital twins struggle to retain historical information as complexity grows, leading to misinterpretations of critical actions like braking versus accelerating when pedestrians are present.

Researchers at Nanyang Technological University in Singapore find that classical world models struggle to accurately simulate environments with long-term dependencies, even in seemingly simple, classically defined settings.

The findings point to potential applications where quantum models could enhance fidelity for AI development and safer testing before real-world deployment.

The researchers used a true-worlds framework to compare classical and quantum models directly, without approximations, highlighting memory constraints in classical simulations.

A single qutrit, a three-level quantum system, can perfectly recreate identical scenarios and align virtual agent policies with real-world outcomes, outperforming classical models.

Quantum world models built with just one qutrit demonstrate flawless replication of environments with extended temporal dependence, indicating an irreducible quantum advantage in real-world alignment.

While digital twins have value, the study emphasizes that simply increasing memory or complexity of classical simulations does not inherently fix inaccuracies in history-dependent environments.