World models enable large language model (LLM) agents to anticipate action outcomes and make informed decisions. For efficient transfer to new environments that share interaction structure and dynamics, prior work encodes world dynamics as symbolic rules or executable programs; although compositional, they remain tied to environment‑specific objects. RoleWorld introduces reusable prototypes and environment‑specific parameter values as functional role abstractions to describe how objects behave during interactions. A teacher LLM extracts role prototypes and executable transition rules defined over role instances from interaction trajectories, and provides a skill that guides an actor LLM to incorporate role information into its decision making. In a target world without a teacher, the actor LLM uses this skill together with its current role assignments and parameter estimates, actively testing role‑instance hypotheses when uncertainty influences the decision. Feedback from each action refines role assignments and parameter estimates online, without updating LLM parameters and keeping transition rules fixed. Experiments on the MARS benchmark show that RoleWorld achieves robust generalization across environments with significantly altered object behavior, demonstrating that functional role abstractions are an effective component for knowledge transfer in world modeling.
Review