Quantum World Models, a new type of imagination?

Maria Helena: Researcher at ICTI Itaú and Brazilian Ambassador

World Models have become a central paradigm in model-based artificial intelligence, enabling agents to simulate future trajectories in compact latent representations. Since the foundational work “World Models” (2018), these architectures have been explored in reinforcement learning, robotics, and generative video models as mechanisms for artificial imagination, allowing machines to internally predict and evaluate possible futures before acting.

However, current world models operate entirely within the classical computational regime. Each imagined scenario corresponds to a distinct rollout: a sequence of latent states generated either sequentially or by stochastic sampling. This imposes a fundamental limitation: classical imagination scales linearly or polynomially with the number of trajectories considered. As the complexity of the environment grows, the number of plausible futures quickly becomes computationally prohibitive.

The new class of model — Quantum World Models (QWM) — leverages the principles of quantum information to extend artificial imagination beyond classical limits. The central observation is that, in quantum mechanics, the superposition principle allows a system to represent multiple possible futures simultaneously within a single quantum state. Instead of sampling or enumerating rollouts, a QWM would evolve a superposition of futures in Hilbert space, allowing the model to propagate many temporal possibilities in parallel.

Early indications that such architectures may be feasible appear in reports such as “Quantum Reinforcement Learning with Quantum World Model” associated with research groups in China and presented in IEEE venues. Although the full manuscript is not publicly available, this appears to be the first explicit attempt to formalize a quantum-enhanced world-model architecture, suggesting that this idea is beginning to emerge within the scientific community.

The conceptual advantage is straightforward. Let © denote the number of classically distinguishable future scenarios a world model can represent. A classical model must process each of these scenarios individually. A quantum system, however, can encode and evolve a state with up to (2^C) amplitudes simultaneously. In principle, this allows a QWM to explore an exponentially large space of futures through unitary evolution, with interference patterns encoding preferences, predictions, or optimal decision pathways.

Such a model would fundamentally change the nature of artificial imagination. Rather than generating trajectories one by one, the agent would manipulate amplitude-weighted futures whose constructive or destructive interference yields refined predictions. This is not simply a speed-up; it constitutes a qualitatively different computational substrate for simulation, planning, and reasoning.

While the practical realization of Quantum World Models faces significant challenges such as noisy quantum hardware, limitations in qubit count, and the need for quantum-compatible latent representations, the conceptual implications are substantial. If feasible, QWM could provide exponential advantages in reinforcement learning, long-horizon planning, robotic decision-making, and video or dynamics prediction. More broadly, they may open a path toward computational imagination inaccessible to classical machines.

This letter aims to introduce the concept of Quantum World Models to the research community and encourage formal investigation into their feasibility, mathematical structure, and potential applications. By merging world-model architectures with quantum information processing, we may unlock a new regime of artificial imagination that transcends classical limits.

References:

D. Ha and J. Schmidhuber, “World Models,” arXiv preprint arXiv:1803.10122, 2018. doi: 10.48550/arXiv.1803.10122.
keywords: generative models; world models; reinforcement learning; representation learning; latent spaces; unsupervised learning; environment modeling.

P. Zeng, Y. He, F. R. Yu and V. C. M. Leung, “Quantum Reinforcement Learning with Quantum World Model,” GLOBECOM 2023–2023 IEEE Global Communications Conference, Kuala Lumpur, Malaysia, 2023, pp. 01–06, doi: 10.1109/GLOBECOM54140.2023.10437803.
keywords: performance evaluation; quantum entanglement; simulation; reinforcement learning; global communication; integrated circuit modeling; quantum circuit; quantum information; Grover’s algorithm; amplitude encoding.