Creating Local World Models using LLMs

Akeylimitation of state-of-the-art large language models is their lack of a consistent world model, which hinders their ability to perform unseen multi-hop reasoning tasks. This paper addresses this by extracting local world models from text into a systematic first-order logic framework, enabling structured reasoning. Focusing on the educational domain, we present a multi-step approach using Prolog to represent and reason with these models. Our method involves segmenting educational texts, generating Prolog definitions, and merging them into a comprehensive knowledge graph. Wesuccessfully extracted several small models and manually verified their accuracy, demonstrating the potential of this approach. While promising, our results are currently limited to small-scale models.