AnyGoal tests whether world models generalize past expert rollouts
Original title: AnyGoal: On World Model Generalization Beyond Expert Rollouts
The benchmark recombines in-support states and actions that expert trajectories rarely pair, and current models fail those cases.
AnyGoal is a benchmark of goal-shifted task variants from widely used world-model environments, released with an open-source library. Each task keeps individual state factors and actions inside training support while underrepresenting their combinations in expert trajectories. Across multiple environments and world-model families, the authors report consistent failures on these recombinations. The study compares dynamics and representation failures across families, interaction regimes, and metrics for goal-shifted prediction and planning.
Unconfirmed: The paper states current models consistently fail these recombinations; it does not name which families or environments in the abstract.
- The question is compositional generalization of learned dynamics beyond expert rollouts.
- Tasks are meant to be solvable if the underlying dynamics were learned.
- Failures are reported across model families, not a single architecture.
- The release includes both the benchmark and an open-source library.
From the original article
1 min read at arxiv.orgThe goal of a world model is to capture the dynamics of an environment. This has the great appeal that it allows agents to plan and simulate a course of actions with the help of a world model. For these uses to be reliable, however, the learned dynamics must behave well beyond the specific state-action combinations observed in expert rollouts.
By Heejeong Nam · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.