arXiv Game World Models & Generation· Riccardo Andreoni·· 2026-10-14AI score29
Neurosymbolic learning of interpretable world models
Neurosymbolic Learning of Interpretable World Models
AI summary
Researchers propose the Concept-Grounded World Model (CGWM), a neurosymbolic design that learns interpretable concept representations from observations. It is built on knowledge of valid world states and their dynamics, expressed in temporal logic, and it frames planning as differentiable reasoning, optimizing the agent's actions by gradient ascent to meet a goal. In experiments in two game environments, CGWM's concept grounding is markedly more accurate, including over long horizons, and its planning is more robust both in and out of distribution.
Source: arXiv Game World Models & Generation · arxiv.org