Code World Model: Coding Agent as World Brain

TL;DR - Code World Model uses a coding agent to maintain executable, persistent world state while a conditioned video model renders the resulting scenes. This separation could enable more coherent, rule-consistent, and open-ended simulated worlds than models that learn dynamics solely from video.

• A language-model coding agent reasons about events, updates state, and generates executable code governing world evolution.
• A proxy representation encodes frame-level spatiotemporal constraints and is compiled into video that conditions the visual generator.
• The authors construct aligned proxy-observation datasets from gameplay and real-world videos.
• Fine-tuned MiniMax-H3 follows proxy specifications in simple interactive worlds while retaining detailed visuals and dynamics.