Quantum-Structured World Models Explore Predictive Latent Dynamics.

World models are designed to learn latent states that summarize interaction histories, evolve over time, and support prediction, simulation, or planning. Traditionally, these latent states are represented using classical vectors, probability distributions, or neural network activations. This research introduces a novel framework called Quantum-Structured World Models (QSWMs), which draws inspiration from quantum theory. QSWMs utilize mathematical structures such as complex-valued representations and density-matrix-like latents for predictive world modeling. The core investigation is whether these quantum-inspired structures provide beneficial inductive biases for learning and predicting latent dynamics. The framework establishes three foundational properties: classical inclusion, predictive sufficiency, and structured compactness. The paper instantiates and evaluates complex-valued and density-matrix-like QSWM variants on elementary cellular automata, comparing them against strong classical baselines. Initial results indicate promising local predictive potential for complex-valued QSWMs. However, the study also highlights limitations, particularly in long-horizon rollout and the performance of density-matrix variants, suggesting areas for future development.