Newest papers

AnyGoal: On World Model Generalization Beyond Expert Rollouts

Heejeong Nam, Divyat Mahajan, Quentin Le Lidec, Randall Balestriero, David Alexander Klindt · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26844 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

El objetivo de un modelo mundial es capturar la dinámica de un entorno. Esto tiene el gran atractivo de que permite a los agentes planificar y simular un curso de acciones con la ayuda de un modelo mundial. Para que estos usos sean fiables, sin embargo, la dinámica aprendida debe comportarse mucho más allá de las combinaciones de estado-acción específicas observadas en los lanzamientos de expertos. A pesar de esto, la relación entre la cobertura de datos, la generalización de la dinámica y el desempeño de la planificación posterior sigue siendo inexplorada. Estudiamos esta relación a través de la lente de la generalización de la composición: ¿puede un modelo mundial recombinar los factores de estado y acción conocidos en dinámicas válidas? Para responder esto, introducimos AnyGoal, un benchmark de variantes de estudio de tareas desplazadas por

Machine translation (Meta M2M100); the English original follows.

Abstract

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. Despite this, the relationship between data coverage, dynamics generalization, and downstream planning performance remains underexplored. We study this relationship through the lens of compositional generalization: can a world model recombine familiar state and action factors into valid dynamics? To answer this, we introduce AnyGoal, a benchmark of goal-shifted task variants built from widely used world-model environments, 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, but their combinations are underrepresented in expert trajectories. Across multiple environments and world model families, we find that current models consistently exhibit failures under these recombinations. By examining dynamics and representation failures across model families, interaction regimes, and evaluation metrics, we study where current world models generalize and where challenges remain for goal-shifted prediction and planning. Together, this framework, benchmark, and analysis provide a basis for evaluating progress toward world models that generalize to new planning tasks.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Generalization, Task (project management), Action (physics), Artificial intelligence, Representation (politics)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange