International Journal of Innovative Research in Computer and Communication Engineering
ISSN Approved Journal | Impact factor: 8.771 | ESTD: 2013 | Follows UGC CARE Journal Norms and Guidelines
| Monthly, Peer-Reviewed, Refereed, Scholarly, Multidisciplinary and Open Access Journal | High Impact Factor 8.771 (Calculated by Google Scholar and Semantic Scholar | AI-Powered Research Tool | Indexing in all Major Database & Metadata, Citation Generator | Digital Object Identifier (DOI) |
| TITLE | Addressing Hallucination in Large Language Models Through a Hybrid Detection-Verification Retrieval-Augmented Generation Framework |
|---|---|
| ABSTRACT | Advances in transformer-based language modeling have enabled machines to produce remarkably coherent and contextually relevant text across a wide range of tasks. Yet, these powerful systems carry a persistent flaw: they sometimes generate content that is plausible in tone but factually groundless — a phenomenon commonly referred to as hallucination. This shortcoming becomes particularly consequential in high-stakes domains such as clinical decision support, regulatory compliance, and legal advisory systems, where even minor inaccuracies can have serious downstream effects. Retrieval-Augmented Generation (RAG) has garnered considerable attention as a strategy to anchor language model outputs in externally verified knowledge, thereby mitigating the tendency toward fabrication. However, practical deployments of RAG expose a set of underexplored vulnerabilities: retrieved documents may introduce noise, generated content may conflict with retrieved evidence, and the pipeline often lacks a formal mechanism to detect or correct factual inconsistencies post-generation. This paper systematically examines five prominent research contributions addressing hallucination within the RAG paradigm and identifies recurring structural limitations across them. Building on this analysis, we introduce the Hybrid Detection-Verification RAG (HDV-RAG) framework — a multi-stage architecture that brings together semantic retrieval, cross-encoder re-ranking, hallucination classification with explanatory outputs, knowledge cross-verification, and Direct Preference Optimization (DPO)-driven fine-tuning into a single coherent pipeline. Empirical evaluation across multiple benchmarks demonstrates that HDV-RAG achieves measurable gains in factual consistency, response reliability, and BERTScore compared to standard RAG configurations. |
| AUTHOR | ANUSHA K, BHARGAVI NAIK, BHAVANA MOHANDAS SANGALAD, PROF. USHA K Students, Department of Computer Science and Engineering, Jain Institute of Technology, Davangere, Karnataka, India Professor, Department of Computer Science and Engineering, Jain Institute of Technology, Davangere, Karnataka, India |
| VOLUME | 184 |
| DOI | DOI: 10.15680/IJIRCCE.2026.1405015 |
| pdf/15_Addressing Hallucination in Large Language Models Through a Hybrid Detection-Verification Retrieval-Augmented Generation Framework.pdf | |
| KEYWORDS | |
| References | [1] J. Song, X. Wang, J. Zhu, Y. Wu, X. Cheng, R. Zhong, and C. Niu, "RAG-HAT: A Hallucination-Aware Tuning Pipeline for LLM in Retrieval-Augmented Generation," in Proc. 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP Industry Track), Miami, USA, Nov. 2024, pp. 1548–1558. [2] P. Béchard and O. Marquez Ayala, "Reducing Hallucination in Structured Outputs via Retrieval-Augmented Generation," in Proc. NAACL-HLT 2024 (Industry Track), 2024, pp. 228–238. [3] M. Y. Mohammed, S. A. Ali, S. K. Ali, A. A. Majeed, and E. H. Mohamed, "Aftina: Enhancing Stability and Preventing Hallucination in AI-Based Islamic Fatwa Generation Using LLMs and RAG," Neural Computing and Applications, vol. 37, pp. 20957–20982, 2025. [4] J. Kirchenbauer and C. Barns, "Hallucination Reduction in Large Language Models with Retrieval-Augmented Generation Using Wikipedia Knowledge," 2024. [5] P. Béchard and O. Marquez Ayala, "Reducing Hallucination in Structured Outputs via Retrieval-Augmented Generation," ServiceNow Research, 2024. |
