Evaluation of LLM Hallucination Mitigation Strategies

The 6th International Artifical Intelligence and Data Science Congress, Nevşehir, Türkiye, 20 - 22 Nisan 2026, ss.245-248, (Tam Metin Bildiri)

Abstract— Large Language Models (LLMs) demonstrate

strong performance on a wide range of factual question-answering

tasks; however, accurately evaluating their tendency to produce

hallucinated or misconception-driven responses remains an open

challenge. This issue reduces trust in LLM-based systems,

particularly in critical domains such as healthcare, law, and

scientific writing. In this study, we evaluate several hallucination

mitigation strategies including retrieval-augmented generation

(RAG), citation-aware prompting, and self-consistency. We

construct a curated benchmark of questions paired with verified

correct answers, common incorrect beliefs, and authoritative

reference sources. The evaluation framework measures factual

accuracy, and response quality. We further analyze the impact of

self-consistency, citation-aware prompting, and reference-text

completeness on hallucination detection. Experimental findings

show that while all mitigation strategies improve performance

compared to the straightforward questioning, approaches that

explicitly enforce grounding—such as citation-aware

prompting—are the most effective in reducing hallucinations and

improving response quality. The results highlight systematic

failure modes where models reproduce common