Title: A Review of Hallucination Mitigation Techniques in Retrieval-Augmented Generation Systems
Authors: Emmanuel Eniola Ajala, Harouna Wendpanga Yann Christian Sankara, Serif Oyindamola Oyesiji, Stanley Nwakamma.
Volume: 10
Issue: 9
Pages: 162-176
Publication Date: 2026/09/28
Abstract:
Retrieval-augmented generation has become the default architecture for grounding large language models in external knowledge, and it is widely deployed precisely because it promises to reduce hallucination, the generation of fluent content unsupported by fact or source. The promise is real but partial: retrieval-augmented systems continue to hallucinate, and they do so through failure modes that differ from those of closed-book models, including retrieval of irrelevant or conflicting evidence, faithful generation from unfaithful sources, unfaithful generation from faithful sources, and fabricated attribution. This review synthesizes the literature on hallucination and its mitigation in retrieval-augmented generation systems. It first consolidates definitions and origins of hallucination in neural text generation, distinguishing factuality from faithfulness and intrinsic from extrinsic error. It then reviews the architectural foundations of retrieval augmentation, from dense retrieval and fusion-based readers through end-to-end retrieval-augmented models, and organizes documented failure points into a pipeline-aligned taxonomy spanning query formulation, retrieval, context construction, generation, and attribution. Mitigation research is examined at each stage: retrieval-side methods including query rewriting, reranking, and robustness training against irrelevant context; generation-side methods including grounded and contrastive decoding, self-verification, uncertainty estimation, and calibrated abstention; and pipeline-level methods including active and iterative retrieval, self-reflective generation, verification chains, and structured knowledge integration. The evaluation landscape is reviewed across faithfulness metrics, attribution measures, and RAG-specific benchmarks, followed by evidence from high-stakes domain deployments. The review closes with open problems in knowledge conflict resolution, long-context utilization, calibration, and evaluation validity, and with directions for future research.