Edit Knowledge, Not Just Facts via Multi-Step Reasoning over Background Stories

The paper introduces a novel training strategy for large language models that emphasizes knowledge updating through multi-step reasoning rather than mere memorization of facts. It proposes a framework where new information is integrated as coherent background stories, and models are trained using self-generated multi-hop questions that necessitate this new knowledge for task completion. Experimental results demonstrate that this approach significantly enhances the model's ability to utilize updated knowledge in reasoning tasks, which is crucial for practitioners aiming to build more adaptive AI systems.