LLMs have recently achieved impressive results on knowledge-intensive tasks through Retrieval-Augmented Generation (RAG) and structured reasoning techniques. However, current RAG systems still struggle with multi-step reasoning, unreliable chains, and poor integration of retrieved information. This project aims to develop and evaluate retrieval enhanced reasoning frameworks that improve performance and reliability of LLMs on complex inference tasks. The focus is on systems that combine efficient retrieval, hypergraph-based knowledge structures, and multistep reasoning pipelines, to create more interpretable and faithful reasoning processes.