Code as Worlds: Agentic Discovery of Physical World Representations

This method enables agents to discover executable world representations through code to improve physical reasoning capabilities. The approach focuses on transforming sensory or conceptual data into programmable environments for more accurate physical simulations.