AnyGoal tests world-model generalization beyond expert rollouts
Original title: AnyGoal: On World Model Generalization Beyond Expert Rollouts
Current models fail when familiar state and action factors are recombined into underrepresented combinations.
AnyGoal is a benchmark of goal-shifted task variants built from widely used world-model environments, released with an open-source library. Individual state factors and actions stay inside training support, but their combinations are underrepresented in expert trajectories. Across environments and model families, the authors find consistent failures on these recombinations and examine where dynamics and representation generalization still break.
Unconfirmed: The reported generalization failures come from the authors' own evaluations.
- The benchmark targets compositional generalization for planning.
- Tasks are meant to be solvable if underlying dynamics were learned.
- Failures appear across multiple environments and world-model families.
- The framework is offered for tracking goal-shifted prediction and planning.
From the original article
1 min read at huggingface.co
The 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, Divyat Mahajan, Quentin Le Lidec · huggingface.co
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.