Graph World Models: Concepts, Taxonomy, and Future Directions formalizes a new research paradigm in artificial intelligence. While traditional world models—which help AI agents predict and plan by simulating environments—are powerful, they often struggle with background noise, the accumulation of prediction errors over time, and a limited ability to perform logical reasoning. This paper introduces "Graph World Models" (GWMs) as a solution, proposing that by representing environments as graphs of entities and interactions rather than flat data, AI can better understand and navigate complex virtual spaces. Moving Beyond Flat Data Classical world models typically process information as unstructured vectors, which can lead to "noise sensitivity," where the model wastes capacity on irrelevant background details. They also suffer from "error accumulation," where small mistakes in early predictions compound into large deviations during long-term simulations. GWMs address these issues by decomposing the environment into a graph structure, where nodes represent entities and edges represent their relationships. This structural approach allows the model to focus on meaningful interactions, effectively filtering out noise and providing a more stable framework for long-term planning. A New Taxonomy of Relational Biases The authors propose a three-layer taxonomy to categorize GWMs based on the "relational inductive biases" (RIB) they inject into the model. These biases act as structural priors that guide how the AI interprets the world: Graph as Connector (Spatial RIB): These models focus on navigation and connectivity, using graphs to map out landmarks and reachable paths in an environment. Graph as Simulator (Physical RIB): These models focus on dynamics, distilling complex physical laws—like collisions or friction—into rules that govern how objects interact over time. * Graph as Reasoner (Logical RIB): These models focus on higher-level cognition, using graphs to extract semantic rules and causal relationships to support complex decision-making and instruction following. Current Approaches and Design Principles The paper highlights how these categories function in practice. Connector models, for instance, often use graph retrieval to simplify search tasks, while simulator models use object-centric or system-centric approaches to predict how entities move and collide. By moving away from pixel-level modeling and toward these structured representations, researchers can create agents that are more robust to perspective changes and better at handling multi-agent collaboration. Challenges and Future Directions Despite the potential of GWMs, the field faces several open challenges. A primary concern is scalability; as exploration time increases, the number of nodes and edges in a graph can grow linearly, putting significant pressure on memory and storage. Additionally, most current models rely on deterministic frameworks, which struggle to capture the inherent randomness of the real physical world. Future research is needed to develop dynamic graph adaptation, probabilistic relational dynamics, and dedicated benchmarks to evaluate how well these models perform across different tasks and environments.