Abstract

Retrieval-Augmented Generation (RAG) has significantly enhanced largelanguage models (LLMs) in knowledge-intensive tasks by incorporating externalknowledge retrieval. However, existing RAG frameworks primarily rely onsemantic similarity and correlation-driven retrieval, limiting their ability todistinguish true causal relationships from spurious associations. This resultsin responses that may be factually grounded but fail to establishcause-and-effect mechanisms, leading to incomplete or misleading insights. Toaddress this issue, we introduce Causal Dynamic Feedback for AdaptiveRetrieval-Augmented Generation (CDF-RAG), a framework designed to improvecausal consistency, factual accuracy, and explainability in generativereasoning. CDF-RAG iteratively refines queries, retrieves structured causalgraphs, and enables multi-hop causal reasoning across interconnected knowledgesources. Additionally, it validates responses against causal pathways, ensuringlogically coherent and factually grounded outputs. We evaluate CDF-RAG on fourdiverse datasets, demonstrating its ability to improve response accuracy andcausal correctness over existing RAG-based methods. Our code is publiclyavailable at https://github.com/ elakhatibi/CDF-RAG.