A Retrieval-Augmented Multi-Agent Framework for Reducing Hallucinations in Large Language Models for Research Assistance

Read full text Summarise with VIRAT
Authors: Yashwant Singh
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding and generation, yet their propensity to produce factually incorrect or fabricated content—commonly termed 'hallucinations'—poses significant challenges in high-stakes research assistance applications. This paper introduces RAMAF (Retrieval-Augmented Multi-Agent Framework), a novel architecture that combines dense passage retrieval, cross-encoder re-ranking, and a coordinated ensemble of specialized verification agents to substantially reduce hallucination rates in LLM-driven research assistance systems. Our framework employs three cooperative agents: a Retrieval Agent responsible for sourcing evidence from curated knowledge bases, a Verification Agent that cross-checks candidate responses against retrieved documents, and a Synthesis Agent that merges evidence-grounded statements into coherent, citation-backed outputs. Experimental evaluation on the HaluEval, TruthfulQA, and PubMedQA benchmarks demonstrates that RAMAF achieves a hallucination rate of 4.1%—a 80.8% relative reduction compared to a GPT-4 baseline (21.4%) and a 72.3% reduction over single-agent RAG systems (14.8%). Factuality scores (FactScore) improve from 0.71 to 0.95, while ROUGE-L gains of 68.4% indicate significantly better alignment with reference answers. Keywords: Retrieval-Augmented Generation, Multi-Agent Systems, Hallucination Mitigation, Large Language Models, Research Assistance, Fact Verification, Dense Passage Retrieval

Fields of Study & Subjects

Topic Modeling Information Retrieval and Search Behavior Multimodal Machine Learning Applications

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN