AnyGoal: On World Model Generalization Beyond Expert Rollouts
arXiv cs.AI · 20 Oct 2026, 03:59 AEST
AnyGoal introduces a benchmark for evaluating world model generalization beyond expert rollouts. The study highlights the challenges current models face in recombining familiar state and action factors into valid dynamics.
Key facts
- The goal of a world model is to capture the dynamics of an environment.
- We study this relationship through the lens of compositional generalization: can a world model recombine familiar state and action factors into valid dynamics?
- Across multiple environments and world model families, we find that current models consistently exhibit failures under these recombinations.
Read aloud
AnyGoal introduces a benchmark for evaluating world model generalization beyond expert rollouts. The study highlights the challenges current models face in recombining familiar state and action factors into valid dynamics. The goal of a world model is to capture the dynamics of an environment. We study this relationship through the lens of compositional generalization: can a world model recombine familiar state and action factors into valid dynamics? Across multiple environments and world model families, we find that current models consistently exhibit failures under these recombinations.
Sources in this cluster
- AnyGoal: On World Model Generalization Beyond Expert Rollouts
arXiv cs.AI · 20 Oct 2026, 03:59 AEST