Code Plans, Diffusion Renders: Open-Ended Generative World Modeling

TL;DR - CoDeR is a generative world-modeling framework that encodes world rules and dynamics as executable code, then uses video generation models to render visual observations. This separation aims to support persistent, open-ended simulations beyond the temporal limits of conventional video world models.

• Coordinates five complementary roles to translate high-level concepts into structured rules, executable dynamics, and perceptual outputs.
• Maintains explicit state for long-term memory, autonomous world evolution, and interactions extending beyond the current observation.
• Supports persistent multi-agent scenarios in which multiple entities can act, interact, and evolve.
• The authors report state-of-the-art results across multiple evaluation settings and plan to release code and model weights.