Skip to content
Conference

Hallucination Detection and Reduction in Genai (LLMs)

Sep 2026 · 2026 Forth International Conference on Networks, Multimedia and Information Technology (NMITCON) · pp. 1-8 · 0 citations · 23 references

Abstract

Hallucinations are false or unverifiable outputs generated by Large Language Models (LLMs) that reduce their trustworthiness in critical applications such as healthcare, education, and research. This paper proposes a hybrid framework that unifies three complementary verification signals for the detection and reduction of hallucinations: SelfCheckGPT-based semantic consistency checking across multiple sampled responses, retrieval-based verification using Sentence-Transformer embeddings with FAISS similarity search, and structured knowledge graph validation. Unlike existing methods that rely on a single internal or external signal, the proposed framework aggregates all three into one weighted hallucination score and applies threshold-based classification, after which detected hallucinations are corrected using the most relevant retrieved evidence. The framework is evaluated on 200 diverse queries in a zero-resource setting using pseudo groundtruth labels derived from knowledge graph validation and retrieval agreement. At the moderate threshold, it attains a detection rate of 63.5% and an equal correction success rate. Compared with the SelfCheckGPT baseline, the hybrid model improves the area under the curve from 0.454 to 0.647, precision from 0.765 to 0.858, recall from 0.157 to 0.657, and F1-score from 0.260 to 0.744. The results confirm that combining internal and external verification yields more reliable and generalizable hallucination detection, offering a practical route toward trustworthy generative AI.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.