HyperWorld：超图结构状态序列化提升文本世界模型学习

AI 摘要

HyperWorld 研究通过对比四种信息等价的文本状态序列化方式（原始观察、独立句子、成对三元组、超边单元），发现超边序列化能显著提升小语言模型（0.5B-1.5B）在文本世界中的效果预测和分布外鲁棒性，并在规划任务中取得最高成功率。该研究为学习符号世界模型提供了高阶结构作为廉价归纳偏置的证据。

正文节选

HyperWorld: Hypergraph-Structured State Serialization Improves Learned Textual World Models Abstract World models, which predict how an environment evolves under actions, are increasingly used to equip language-model agents with the ability to plan before acting. In text environments, a world model must learn symbolic dynamics from serialized descriptions of the state, yet how the structure of this serialization affects learning remains largely unexamined: prior work serializes states either as

发布时间：2026-09-02 12:00

抓取时间：2026-09-02 12:08

来源机构：arXiv