Kholov Shavkat Yorovich, K.M. Bobobekov, Abdullozoda Husein Sadulo, Abdullozoda Hasan Sadulo

Annotation:

Hallucinations - factually incorrect yet confidently stated outputs - remain a critical challenge in large language models (LLMs). This paper proposes a hybrid pipeline that combines Retrieval-Augmented Generation (RAG) with a classical machinelearning verification module (SVM, logistic regression) for automatic hallucination detection and correction. The proposed architecture is evaluated on a corpus of university academic documents in English and Tajik. Experimental results show that the hybrid approach reduces the hallucination rate by 41 % compared to a plain LLM baseline and by 23 % compared to RAG without verification, achieving an F1-score of 0.90 on the English corpus and 0.85 on the Tajik corpus.

Keywords: LLM hallucinations, RAG, fact verification, classical machine learning, SVM, Tajik language, university documents, hallucination mitigation.