This paper introduces Graph World Models (GWMs) as a unified research paradigm, addressing the limitations of classical world models through structured graph representations. It proposes a novel taxonomy for GWMs based on relational inductive biases (spatial, physical, and logical) and discusses future research directions.

This paper introduces a novel world model that leverages Vector Symbolic Architecture (VSA) principles, specifically Fourier Holographic Reduced Representation (FHRR), to address the limitations of unstructured neural network-based models. By encoding states and actions as high-dimensional complex vectors and modeling transitions with element-wise multiplication, the proposed model achieves superior generalization, sample efficiency, and interpretability.