Code-as-World has an agent write executable physics rather than describe it
Posted 2026-08-27 with 37 upvotes on Hugging Face Daily Papers, this paper argues that vision-language models can recognize and explain physical events without holding explicit representations of object states, physical parameters and governing dynamics, which is what reasoning about interventions actually requires. Code-as-World represents a physical scene as executable code covering composition, dynamic evolution and visual appearance, and builds those representations from language descriptions or real video through an abductive loop where an agent proposes, executes, renders, verifies and refines a world hypothesis. The rendering step is what closes the loop, since the hypothesis is checked against observation rather than judged by another model.
↳ Follow the thread