Evaluating Reinforcement Learning Based Models for Test Time Enhancement in RAG
Vikas Kamra, Lakshya Gupta, Dhruv Arora, Ashwin Kumar Yadav · 2025
Traditional Retrieval Augmented Generation (RAG) approaches exhibit considerable potential by integrating exterior documents into the outputs of large language models. Despite these advances, most RAG methods rely heavily on dense vector retrieval techniques that capture only surface-level textual similarities, often failing to capture the complex, multihop relationships present in large corpora. This research introduces Dynamic Hierarchical Graph Enhanced - Retrieval Augmented Generation (DHGE-RAG), a unique framework that leverages dynamically constructed and hierarchically organized knowledge graphs to upgrade document retrieval and generation. With combining dynamic graph-enabled indexing, graph-directed multi-hop retrieval, and graph-boosted generation, the proposed approach provides improved retrieval accuracy, superior multi-hop reasoning, and a marked reduction in hallucinations in the generated outputs. The proposed approach evaluates DHGE-RAG on standard benchmarks such as HotpotQA, TriviaQA, and the CRAG benchmark, and reports significant improvements in F1-score and retrieval recall, alongside a substantial decrease in hallucination rates. This study discusses the challenges associated with scalability, multi-modal integration and outlines promising directions for future research in this emerging field.