Introduction
World models empower language-model agents to anticipate environmental dynamics and formulate plans prior to taking action. In text-based environments, agents must deduce symbolic action effects from serialized state descriptions; however, the influence of serialization structure on model performance has received limited attention. This study introduces HyperWorld, a systematic examination of state serialization techniques within learned textual world models.
Methodology
We compare raw observations against three symbolic serializations of the same ground-truth state: independent sentences, pairwise triples, and entity-centered hyperedge units that group multiple related facts around entities and relations. All variants employ an identical training objective: given a state and an action, the model must either forecast symbolic effects or classify the action as infeasible.
Key Findings
The experimental evaluation spans multiple model scales, data budgets, and both in-distribution and out-of-distribution test worlds. The principal results are:
- Hyperedge serialization delivers the most pronounced improvements for models of 0.5Bโ1.5B parameters, particularly under distribution shift.
- Larger models exhibit reduced performance gaps between serialization methods, with pairwise triples occasionally matching or surpassing hyperedges on in-distribution exact match metrics.
- Out-of-distribution performance: hyperedges achieve the highest factual F1 score and offer the best balance between feasibility detection and effect prediction for small-to-medium scale models.
- Downstream planning: when incorporated into greedy planning, the hyperedge world model attains the highest success rate among all tested representations.
Implications
These findings demonstrate that higher-order state organization serves as a straightforward yet effective inductive bias for learned symbolic world models, especially when model capacity is constrained or test environments diverge from training distributions. As of 2026, the integration of structured serialization methods like hyperedges is increasingly vital for building robust, generalizable AI agents capable of operating reliably in dynamic, text-based settings.
via ArXiv AI