Multiplayer AI systems that model virtual worlds face a basic problem: each player sees the world from their own view, but all views must refer to the same underlying state. A new paper from arXiv introduces MultiWorldBench, a diagnostic benchmark built in Minecraft, to test whether independently controlled views actually describe one shared and persistent world.
The benchmark includes 495 case configurations, each designed to probe specific aspects of cross-view consistency. The authors argue that existing single-agent benchmarks are insufficient because they ignore the coordination required when multiple agents act in the same environment. MultiWorldBench aims to fill that gap by providing a controlled setting where inconsistencies between views can be measured directly.
Because the abstract is the only source available, the paper's exact evaluation metrics and baseline results are not described here. The contribution, as stated, is the benchmark itself: a concrete diagnostic tool for a problem that is central to multiplayer world models but has been hard to evaluate systematically.