AnyGoal: On World Model Generalization Beyond Expert Rollouts

AnyGoal. On World Model Generalization Beyond Expert Rollouts. By Heejeong Nam and four others, from Brown University, Cold Spring Harbor Laboratory, and two other institutions. Published on the nineteenth of October. Its code repository is worked out: about one hundred and ninety source files, with a readme and a licence.

The goal of a world model is to capture the dynamics of an environment, and the great appeal of that is planning. An agent can simulate a course of actions with the world model's help before committing to any of them. For those uses to be reliable, though, the learned dynamics must behave well beyond the specific combinations of state and action observed in expert rollouts. An agent planning toward a goal the experts never pursued will reach combinations the experts never produced. Despite that, the relationship between data coverage, dynamics generalization, and downstream planning performance remains underexplored. How much a model has seen, how well it generalizes, and how well it then plans have not been studied together.

The paper studies that relationship through the lens of compositional generalization. The question it poses is whether a world model can recombine familiar state factors and action factors into valid dynamics. Every individual piece has been seen during training. Only the way the pieces are put together is new.

AnyGoal is a benchmark of goal-shifted task variants, built from widely used world-model environments, and released together with an open-source library. Each task is designed to be within reach of a world model that has learned the underlying dynamics. Individual state factors and actions remain within their training support. What is underrepresented in the expert trajectories is their combinations. That design narrows what a failure can mean. Nothing on its own is unfamiliar, so a model that falls short has not managed to compose what it already knows.

Across multiple environments and world model families, current models consistently exhibit failures under these recombinations. The authors then examine those failures more closely, separating failures of the dynamics from failures of the representation, and comparing across model families, interaction regimes and evaluation metrics. The result is a map of where current world models generalize, and where challenges remain for goal-shifted prediction and for planning.

The benchmark asks a deliberately fair version of the generalization question: no unseen state, no unseen action, only a new combination. Current models still struggle with it. The authors offer the framework, the benchmark and the analysis together as a basis for evaluating progress toward world models that generalize to new planning tasks, which is what made world models appealing in the first place.