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

Recursive Code World Models. Building Complex Worlds through Recursive Scene Programs. By Zhiqi Li and two others, from the Georgia Institute of Technology. Published on the tenth of September. Its code repository is public and real: about thirty source files.

A code world model represents a world as an executable program. That is a useful representation, but it says nothing about how to build a complex world in it. Writing out an intricate 3D scene as code, from a picture, is a construction problem, and the representation alone does not solve it.

Recursive Code World Models, RCWM, is a framework for reconstructing a complex 3D world in code from a single reference image. It pairs two things. The first is a representation, the Recursive Scene Program, or RSP, which writes the executable world as compositional scene code. The second is a construction solver that calls itself recursively.

Every call of the solver follows the same complete process. Establish the whole. Recursively reconstruct the parts that are still unresolved. Then revisit the whole to refine how those parts compose. This global, local, global recursion gives fine-scale structures their own loops of perception and editing, while the geometry and relationships of the scene as a whole are preserved. Each part is treated as a smaller world in its own right, rebuilt with the same care as the scene that contains it, and then fitted back in.

Two mechanisms keep the levels consistent. Views aligned with the reference propagate a shared camera projection from one level to the next. And revisiting the parent handles what only appears after local refinement: boundaries, spatial relations, and errors shared between parts. Guiding all of it is a vision-language coding agent that compares the reference image directly with renders of the scene, and decides when to refine, when to descend a level, and when to return.

Across complex scenes, RCWM outperforms earlier code-based methods for reconstructing scenes from images. Ablation studies support the benefit of building recursively, and suggest that deeper calls can improve reconstruction at finer scales. Depth of recursion, on that evidence, is a lever: going further down helps with the smallest structures.

What RCWM offers is a principle of construction rather than a new representation alone. A complex world is built the way the solver builds every piece of it: look at the whole, work on the parts, then look at the whole again. Because the same process repeats at every depth, fine detail and overall layout are handled by one mechanism instead of two.