Hypergraph Serialization Boosts Textual World Model Performance

Large language models acting as agents in text-based environments rely on "world models" to predict outcomes and plan actions. A critical, yet underexplored, aspect of these models is how the environment's state information is structured or "serialized" for the model to process. This study, named HyperWorld, systematically investigates different serialization methods for textual world models. The researchers compared raw text observations with three symbolic representations: independent sentences, pairwise triples, and entity-centered hyperedge units. The hyperedge approach groups multiple related facts around specific entities and relationships, offering a more structured representation. Their findings indicate that hyperedge serialization provides notable performance gains, particularly for models with 0.5B to 1.5B parameters and when dealing with environments different from their training data. While larger models can reduce this performance gap, and pairwise triples sometimes match hyperedges on in-distribution exact matches, hyperedges consistently deliver superior out-of-distribution fact F1 scores and a better balance between feasibility detection and effect prediction for small-to-medium scale models. In practical planning tasks, the hyperedge world model also achieved the highest success rate, suggesting that organizing state information in a higher-order, structured manner is a simple yet powerful inductive bias for improving learned symbolic world models, especially under resource constraints or novel environments.