Error-Aware Graph World Model stops long-horizon rollout divergence

World models that roll out learned dynamics for planning often fail when the environment is a graph—think agents, tools, skills, and dependencies. In these settings, a local prediction error can either stay local or spread through the graph, especially when edges are themselves predicted rather than fixed. Song and Cai formalize this problem with a unified Graph World Model (GWM) framework that handles both fixed-edge and dynamic-edge scenarios, introducing action nodes for node-, edge-, and graph-level decisions. They develop graph-valued rollout bounds that clearly separate error amplification coming from topology structure from amplification due to model inaccuracies, and they propose a joint node-edge operator for rollouts over evolving graphs.

Guided by this analysis, the authors introduce Error-Aware GWM, which applies spectral regularization to control spectral radius of the rollout operator, enforces rollout consistency across time steps, and weighs critical nodes that have outsized influence on downstream predictions. Experiments across synthetic topologies and heterogeneous agent-graph testbeds confirm that rollout error and planning regret grow with horizon, and that dynamic-edge training is necessary when the graph structure itself evolves. Error-Aware GWM prevents long-horizon divergence while maintaining prediction accuracy in those scenarios. The paper also clarifies that GWMs are most beneficial for dynamic graph rollout and agent planning; for static or sparse prediction tasks, specialized graph models remain superior.