Running summarizing tools on a new article

Abstract: Abstract Large Language Models (LLMs) have achieved remarkable success in multilingual commonsense reasoning. However, their reliability drastically degrades when deployed in low-resource, domain-specific settings, often resulting in factual fabrications known as hallucinations. This vulnerability is especially acute for the Burmese language, due to its structural complexities, tokenization artifacts, and the dearth of high-quality localized datasets. In this paper, we present a systematic framework to quantify and mitigate hallucinations in Burmese domain-specific question answering. We introduce MM-AgriQA, a novel, manually annotated gold-standard dataset specifically curated for agricultural query answering under strict context boundaries. We evaluate multiple alignment and mitigation strategies, comparing fine-tuned low-resource models against vanilla commercial systems. Our empirical results demonstrate that incorporating advanced Retrieval-Augmented Generation (RAG) combined with strict syntactic prompt constraints reduces hallucination rates by up to 42.5% while preserving semantic alignment.