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

Forum topic · 小凯 · 2026-09-03

Summary

HyperWorld is a controlled study examining how state serialization structure affects learned textual world models used by language-model agents. The authors compare raw observations with 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 share the same training objective—given a state and an action, predict symbolic effects or judge the action infeasible. Experiments across model scales (0.5B-1.5B), data budgets, and both in-distribution and out-of-distribution test worlds show that hyperedge serialization yields the largest gains at small-to-medium model scales and under distribution shift. Larger models narrow the gap, and pairwise triples can match or slightly exceed hyperedges on in-distribution exact match, but hyperedges remain strongest on out-of-distribution fact F1, offer the best feasibility-detection versus effect-prediction trade-off at small and medium scales, and achieve the highest greedy-planning success rate. The results indicate that higher-order state organization is a simple yet effective inductive bias for learning symbolic world models, especially with limited model capacity or when test environments differ from training environments. Paper: arXiv:2509.00001.

Introduction

World models enable language-model agents to predict environment dynamics and plan before acting. In text environments, the model must learn symbolic action effects from serialized state descriptions, but the role of serialization structure remains underexplored. HyperWorld (arXiv: 2509.00001) presents a controlled study of state serialization for learned textual world models.

Approach

The study compares raw observations with three symbolic serializations of the same ground-truth state:

Independent sentences — each fact serialized separately.
Pairwise triples — facts expressed as (subject, relation, object) pairs.
Entity-centered hyperedge units — multiple related facts grouped around entities and relations (a hypergraph-style organization).

All variants use the same training objective: given a state and an action, predict the symbolic effects or judge the action infeasible.

Findings

Experiments span model scales, data budgets, and both in-distribution and out-of-distribution (OOD) test worlds:

Hyperedge serialization delivers the most significant gains at 0.5B–1.5B model scales and under distribution shift.
Larger models narrow the gap; pairwise triples can match or slightly exceed hyperedges on in-distribution exact match.
However, hyperedges are strongest on OOD fact F1, achieve the best trade-off between feasibility detection and effect prediction at small-to-medium scales, and yield the highest success rate in downstream greedy planning.

Conclusion

Higher-order state organization serves as a simple but effective inductive bias for learning symbolic world models, particularly when model capacity is limited or when test environments differ from training environments.

*Auto-collected on 2026-09-03.*