HyperWorld study finds hypergraph-structured state serialization improves learned textual world models under distribution shift

World models enable language-model agents to predict environment dynamics and plan before acting, but in text environments the model must learn symbolic action effects from serialized state descriptions. The role of serialization structure in this process has been underexplored until now.

HyperWorld presents a controlled study comparing four state serialization approaches for learned textual world models: raw observations, independent sentences, pairwise triples, and entity-centered hyperedge units that group multiple related facts around entities and relations. All variants use the same training objective: given a state and an action, predict symbolic effects or judge the action infeasible.

Across model scales, data budgets, and both in-distribution and out-of-distribution test worlds, hyperedge serialization yields the clearest performance gains for models in the 0.5B–1.5B parameter range and under distribution shift. Larger models reduce the performance gap between serialization methods, and pairwise triples can match or slightly exceed hyperedges on in-distribution exact match metrics.

However, hyperedges achieve the strongest out-of-distribution fact F1 and the best small-to-medium scale trade-off between feasibility detection and effect prediction. In downstream greedy planning evaluations, the hyperedge world model also attains the highest success rate among the tested representations.

The results indicate that higher-order state organization—embodied here by hyperedge-based serialization—serves as a simple but effective inductive bias for learned symbolic world models, especially when model capacity is limited or test environments differ from training conditions.