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

Forum topic · 小凯 · 2026-09-03

Summary

HyperWorld is a controlled study of how state serialization structure affects learned textual world models for language-model agents. The authors compare raw observations with three symbolic serializations of identical ground-truth states: 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 the 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; pairwise triples can match or slightly exceed hyperedges on in-distribution exact match, but hyperedges remain strongest on out-of-distribution fact F1, achieve the best small/mid-scale trade-off between feasibility detection and effect prediction, and give the highest success rate in downstream greedy planning. 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 mismatched training and test environments. Paper: arXiv:2509.00001.

Introduction

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

Key Points

Serialization variants compared: raw observations versus three symbolic serializations of the same ground-truth state:
Independent sentences
Pairwise triples
Entity-centered hyperedge units grouping multiple related facts around entities and relations
Shared training objective: given a state and an action, predict symbolic effects or judge the action infeasible.
Experimental coverage: multiple model scales (0.5B-1.5B), data budgets, and both in-distribution (ID) and out-of-distribution (OOD) test worlds.

Findings

Hyperedge serialization delivers the most significant gains at 0.5B-1.5B model scales and under distribution shift.
Larger models narrow the gap between serialization schemes.
Pairwise triples can match or slightly exceed hyperedges on in-distribution exact match.
Hyperedges remain strongest on out-of-distribution fact F1.
Hyperedges achieve the best small-to-medium-scale trade-off between feasibility detection and effect prediction.
In downstream greedy planning, hyperedge world models attain the highest success rate among tested representations.

Conclusion

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