Interactive world models must capture how multiple participants jointly change a shared environment while observing different parts of it. Competitive esports make this challenge concrete, as players cooperate and compete under conditions of incomplete information.
The Esports World Model Dataset, described in a recent arXiv preprint, is designed to support research in this area. The abstract notes that competitive games make the challenge concrete, with players cooperating and competing while fog of war limits visibility.
The dataset appears to target the specific problem of partial observability in multi-agent settings. However, the abstract is brief, and details on the dataset's composition, scale, and benchmark tasks are not included in the announcement.