Code world models encode dynamics with explicit states and executable rules. This work explores how to coordinate object representations, behavioral revisions, and visual memory within that formulation. Mallea is a system that builds, steers, and memorizes worlds directly from textual prompts.
A coding agent extracts interaction requirements from the prompt and jointly plans object structures, properties, and rules, favoring compact representations that satisfy those requirements. The planned rules drive world evolution; optional user actions can steer the process, while textual requests edit properties or extend capabilities without disturbing unaffected state.
Object texture memory retains reusable appearance only for white‑box‑defined objects and adapts it as those objects change. The program advances the world state, and a generative renderer combines the current structure with the remembered appearance to produce visual output. By organizing construction, intervention, and memory around the same set of objects, Mallea conditions each observation on the latest state and compatible appearance.
Project page: https://lxycopper.github.io/Mallea
Review