Source-linked AI summary
HyperWorld: Hypergraph-Structured State Serialization Improves Learned Textual World Models
Yun-Jian Zhang, Chen-Wei Liang, Tian-Yi Zhang, Jian Ding, Yi-Lun Wu, Ao-Bo Li, Wei-Cong Su, Saifullah, Hong-Yu An, Mu-Jiang-Shan Wang
TL;DR
Textual world models must infer symbolic action effects from serialized states, but the role of serialization structure is underexplored. HyperWorld conducts a controlled comparison of information-equivalent raw, sentence, triple, and hyperedge renderings, finding that hyperedges are especially effective for smaller models and distribution shift, while also improving matched greedy-planning success.
Problem
Existing textual world models use flat observations or pairwise triples, leaving underexplored whether higher-order serialization structure improves learning when information content is fixed.
Method
HyperWorld derives four information-equivalent serializations of symbolic TextWorld states and fine-tunes the same small language-model world-model setup to predict effects or infeasibility.
Results
Hyperedges provide the strongest overall profile at 0.5B–1.5B, most OOD settings, and matched planning, while triples can lead on 3B IID exact match.
Takeaways & Limitations
Higher-order state organization is a simple inductive bias for learned symbolic world models, particularly with limited capacity or distribution shift.
Abstract
from arXiv · showhide
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. We present HyperWorld, a controlled study of state serialization for learned textual world models. We 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 use the same training objective: given a state and an action, predict symbolic effects or judge the action infeasible. Across model scales, data budgets, and in-distribution and out-of-distribution test worlds, hyperedge serialization gives the clearest gains for 0.5B--1.5B models and under distribution shift. Larger models reduce the gap, and pairwise triples can match or slightly exceed hyperedges on in-distribution exact match, but hyperedges achieve the strongest out-of-distribution fact F1 and the best small-to-medium scale trade-off between feasibility detection and effect prediction. In downstream greedy planning, the hyperedge world model also attains the highest success rate among the tested representations. These results show that higher-order state organization is a simple but effective inductive bias for learned symbolic world models, especially when model capacity is limited or test environments differ from training.
1 Introduction
HyperWorld studies whether organizing identical symbolic state information as higher-order hyperedges helps small language models learn textual world dynamics. It compares four information-equivalent serializations and evaluates both predictive dynamics and downstream planning.
- The paper asks whether hypergraph-structured serialization improves small language models’ learning of textual world dynamics when information content is held fixed.
- HyperWorld compares raw observations, independent sentences, pairwise triples, and entity-centred hyperedge units derived from the same ground-truth symbolic states.The testbed spans procedurally generated TextWorld games and IID and out-of-distribution splits.
- Hyperedge serialization improves effect prediction and OOD robustness at 0.5B–1.5B, while remaining competitive at 3B where triples are strong on IID metrics.
- The hyperedge world model achieves a higher live-game success rate than sentence- or triple-based world models under the same greedy search procedure.
- Figure 1 traces the pipeline from equivalent state renderings through LoRA fine-tuning, dynamics evaluation, and world-model-guided planning.
2 Related Work
Prior textual world models use flat text or pairwise relations, while HyperWorld examines whether hypergraphs preserve higher-order associations that those representations fragment. The work transfers hypergraph ideas from language systems to symbolic dynamics learning.
- Textual world-model research includes knowledge-graph transitions, memory graphs, graph-structured states, and object–attribute hierarchies for agent reasoning or reward prediction.
- Existing approaches rely on raw text or pairwise relations and do not examine higher-order serialization structure.
- Graph-theoretic robustness and diagnosability provide a related structural perspective, although those notions differ from learned state serialization.
- Hypergraphs represent n-ary associations that pairwise graphs fragment, with reported benefits in retrieval-augmented generation and multi-hop question answering.
- HyperWorld transfers this representational idea from knowledge retrieval to learning forward dynamics in text environments.
3 Method
The method treats environments as partially observable symbolic systems and compares four lossless state serializers while holding training targets and hyperparameters fixed. A LoRA-fine-tuned decoder-only language model predicts action effects or infeasibility, then supports greedy planning.
- 3.1 Problem Setup: The environment state is represented as a set of ground facts, and admissible actions transform the state by adding and removing fact sets.
- 3.1 Problem Setup: The world model maps a serialized state and action to added and removed effects when admissible, or to INFEASIBLE otherwise.
- 3.2 Information-Equivalent State Serializations: Raw observations, independent sentences, pairwise triples, and entity-centred hyperedges render the same fact set with different grouping structures.
- 3.2 Information-Equivalent State Serializations: Hyper serialization groups facts into deterministic entity-centred n-ary units, covering player, objects, room connectivity, and remaining facts losslessly.
- 3.3 World-Model Learning: Transitions combine goal-directed trajectories with ε-random branches, while syntactically valid but inadmissible commands provide infeasibility negatives.
- 3.3 World-Model Learning: A pretrained decoder-only LM is LoRA-fine-tuned on a unified state–action–effects objective with identical targets, hyperparameters, and data across serializers.
- 3.4 World-Model-Guided Planning: The greedy planner scores imagined goal progress, feasibility, novelty, and model confidence, then executes the highest-scoring action.
4 Experiments
Experiments compare serialization formats across model scales, data budgets, test distributions, rollouts, and planning. Hyperedge grouping is strongest especially for OOD robustness and downstream success, while larger models and some IID metrics narrow its advantage.
- Setup: 410 TextWorld games produced train, validation, test-IID, and harder test-OOD splits with 8,375, 797, 1,058, and 2,001 transitions.Models were Qwen2.5-Instruct at 0.5B, 1.5B, and 3B, trained identically with LoRA.
- Main Results: At 1.5B, hyper reached OOD EM 0.914 and fact F1 0.939, exceeding triples by 7.6 and 6.6 points, respectively.Raw text reached only 0.603 OOD EM in this setting.
- Main Results: At 1.5B, raw achieved feasibility accuracy 0.965 IID / 0.944 OOD, while hyper reached 0.944 / 0.913 and outperformed flat or pairwise renderings on effect prediction.Sentences and triples fell to approximately 0.72 feasibility accuracy at 0.5B because joint preconditions were scattered across lines.
- Sample Efficiency: With 10% of training data, OOD EM was approximately 0.63–0.64 for all methods; at 25%, hyper reached 0.824 versus 0.703 for triples and 0.709 for sentences.At the same 25% budget, sentences exceeded hyper on IID EM, 0.885 versus 0.836.
- Multi-Step Rollouts: At horizon 5 on test-OOD, rollout state F1 was 0.989 for hyper, versus 0.980 for sentences and 0.983 for triples.In matched greedy planning on 30 test-IID games, hyper achieved 76.7% success versus 53.3% for sentences and 56.7% for triples, using 14.3 average steps versus 22.9–34.3.
- World-Model-Guided Planning: Removing the confidence term raised hyper planning success from 76.7% to 93.3%, indicating that the default confidence weight was miscalibrated.Calibrated confidence–planning integration remains future work.
5 Conclusion
HyperWorld presents a controlled study showing that state-serialization structure affects learned textual world models. Hyperedge grouping offers the strongest overall profile at smaller scales and under distribution shift, while achieving the highest planning success among matched models.
- HyperWorld studies how state-serialization structure affects learned textual world models.The study isolates serialization structure as the central experimental variable.
- Hyperedge grouping is not uniformly optimal, since triples match or exceed it on 3B IID exact match.
- Hyperedges offer the strongest overall profile at 0.5B–1.5B and the best OOD generalization in most settings.
- 76.7% planning success is achieved by the hyperedge model, versus 53–57% for flat and pairwise alternatives.
- At 1.5B, hyperedges yield up to 31 percentage-point OOD EM gains over raw text on TextWorld.