Active Inference World Models: From Embodied Control to General-Purpose Adaptive Intelligence

World models have become a central concept in artificial intelligence, robotics and cognitive science because adaptive agents must infer hidden causes, predict future states and use internal models to guide action under uncertainty. However, many contemporary worldmodel approaches treat perception, planning, control, uncertainty and goal-directed behaviour as partially separable components. This paper argues that active inference provides a principled framework for building world models that are intrinsically embodied, action-oriented and uncertainty-sensitive. Under active inference, an agent does not first learn a model of the world and then use it for control. Instead perception, action, policy selection and learning are coupled through the minimisation of variational and expected free energy. Here we develop the concept of active inference world models as generative architectures that integrate sensory inference, latent dynamics, semantic priors, epistemic exploration and preference-guided action. Building on recent work on embodied active inference control, we show how a ray-cast PyBullet agent can be extended from reactive sensorimotor coordination toward a structured belief system over rooms, objects, candidate viewpoints and hidden object locations. In this setting, the agent uses semantic priors to guide search, update its beliefs through positive and negative evidence and select actions that balance pragmatic goal fulfilment against epistemic uncertainty reduction. We argue that this provides a concrete route from embodied control toward more general forms of adaptive intelligence, in which agents do not merely predict the world, but actively interrogate, model and transform it through action.