Analyzing Rollout Error in Graph World Models for Planning

World models are crucial for AI planning, often by simulating future states through "rollouts." However, many real-world environments are best represented as graphs, where entities and their relationships change dynamically. This research delves into the unique challenges of rollout error in Graph World Models (GWMs), particularly how errors can propagate through graph structures over long planning horizons. The authors propose a comprehensive framework for GWMs that handles both static and evolving graph structures, including scenarios where edges themselves are predicted. They introduce novel mathematical bounds to quantify how errors amplify due to both the graph's topology and the model's inherent inaccuracies. Based on this analysis, they developed "Error-Aware GWM," an enhanced model incorporating spectral regularization, rollout consistency, and critical-node weighting. Experiments across various graph types demonstrate that rollout errors and planning failures typically increase with the planning horizon. The Error-Aware GWM significantly mitigates this long-horizon divergence while maintaining prediction accuracy. The findings suggest GWMs are particularly effective for dynamic graph rollouts and agent planning, though specialized models may still excel in static or sparse prediction tasks.