A new research paper describes Khora, a world model that can simulate an arbitrary number of AI agents without needing to be retrained for each new headcount.
Most multi-agent world models are trained on a fixed number of agents, which locks the system into whatever population size it saw during training. Khora's researchers split the problem in two: a shared world state that evolves independently of how many agents are present, and a separate rendering step that queries that state to generate each agent's view. Because agent-specific views come from this population-agnostic rendering interface rather than from direct interaction between video streams, compute cost scales roughly linearly as more agents are added. The team also built a real-time interactive demo to show the approach running in an open-world setting.
The fixed-population assumption has been a quiet bottleneck for anyone trying to use world models for large-scale simulation, whether that means training robots, populating game worlds, or testing crowd behavior. If Khora's approach holds up, one trained model could scale from a handful of agents to a crowd without a costly retraining cycle every time the scenario changes.
That said, the paper's evidence is qualitative, not benchmarked against rival systems on hard numbers. Scaling that looks smooth in a demo video and scaling that survives a rigorous stress test are two different claims.