Co-Evolving Structured Knowledge and Reasoning in Language Models

Retrieval-augmented methods improve factual accuracy by grounding language models in external knowledge, but retrieving over unstructured text chunks often introduces irrelevant context and offers limited control over what information is accessed. Structured knowledge bases offer a more controllable alternative, yet they are expensive to construct, and reasoning over a static knowledge base is often brittle and difficult to generalize. To address these limitations, we propose \textit{KBevo}: a \textit{co-evolving} framework that jointly learns to construct a structured knowledge base and reason over it for knowledge-intensive question answering. By optimizing both components with verifiable QA outcome rewards, our method enables reasoning success to directly improve the quality of the constructed knowledge base. This leads to more complete and consistent knowledge structures, while also improving compositional factual reasoning and controllability.