Recursive Code World Models: Building Complex Worlds through Recursive Scene Programs

Reconstructing complex 3D environments from a single reference image typically strains code-based world models, which struggle to balance micro-details with macroscopic coherence. Recursive Code World Models (RCWM) tackle this by pairing a compositional scene program with a solver that calls itself recursively. Following a global-local-global cycle, a vision-language coding agent establishes the overarching layout, recurses into unresolved components for fine-grained editing, and revisits parent structures to rectify boundary seams and relational errors.