Streaming Multi-Agent Autoregressive Diffusion Model with World State Registers

Researchers introduced WorldWeaver, a streaming multi-agent autoregressive video diffusion model for generating coherent scenes across agents and views.

It uses learnable world state registers to carry persistent shared scene information and agent-specific status across chunks, helping maintain cross-agent context.

The model also adds a Mixture-of-Transformers design and trains with supervision from agent status, global views, and scene text.

In two-agent Minecraft experiments, WorldWeaver outperformed prior methods in consistency and visual quality, showing the value of explicit world-state modeling.