AGWM: World Models for Environments with Compositional Prerequisites

A new paper on arXiv (2605.06841) introduces Affordance-Grounded World Models (AGWM) to address limitations in model-based learning for interactive environments. Standard world models learn stationary transition functions that map states and actions to next states, but they often internalize correlations as causal rules while ignoring action preconditions. In environments where actions become executable only after prerequisites are met or become non-executable when prerequisites are destroyed—termed structure-changing events (SC events)—conventional models fail to determine action executability in multi-step predictions. AGWM aims to improve predictions by grounding models in affordances, enabling agents to simulate trajectories more accurately. The research focuses on environments with compositional prerequisites, where the set of available actions changes dynamically based on state conditions.