HyperWorld: Hypergraph-Structured State Serialization Improves Learned Textual World Models

Language‑based agents can anticipate how an environment changes and plan ahead by using world models. In textual settings, the model must infer the symbolic effects of actions from state descriptions, yet it is unclear how the structure of that information influences performance. The study, called HyperWorld, investigates this by running a controlled experiment that compares several ways of serializing the same state: a raw representation, independent sentences, pairwise triples, and hyper‑entity units that group related facts. All approaches are trained with the same objective—predict the symbolic effects of an action or determine that the action is infeasible. Results, obtained with models ranging from 0.5 to 1.5 billion parameters under different data and distribution conditions, show that hyper‑entity serialization yields notable gains, especially for medium‑sized models facing out‑of‑distribution scenarios. Larger models narrow the gap between representations, and triples can match or slightly surpass hyper‑entities on in‑distribution tests, but hyper‑entities still lead in fact accuracy and feasibility detection when the domain shifts. In greedy planning trials, the hyper‑entity representation also achieves the highest success rate. These findings underscore that organizing state into higher‑order structures provides a simple, effective inductive bias for symbolic world models, particularly when model capacity is limited or the test environment differs from training, potentially leading to more robust and reliable AI agents.

Read the original article on arXiv cs.AI

This summary is an informational synthesis produced by dataqbs.com. All rights to the original content belong to its author and the cited media outlet. We act solely as curators of technology news and claim no authorship.