Abstract
This study investigates how large language models (LLMs) respond to factual contradictions under varying prompt styles in both English and Kazakh—a low-resource language. We evaluate the performance of DeepSeek-R1-Distill-Qwen-7B, and GPT-4o-mini on a dataset of 100 question– context pairs with deliberately introduced factual distortions using Retrieval-Augmented Generation (RAG) and Zero-shot approaches across strict, standard, and weak prompt types. Results show that DeepSeek outperforms in English language, particularly in zero-shot settings, as well as it exhibits higher variability and sensitivity to prompt strength. GPT-4o-mini demonstrates more accurate performance across Kazakh datasets. Qualitative analysis reveals distinct model behaviors, such as blind acceptance of falsehoods, expression of uncertainty, and ethical refusals. The findings highlight the challenges LLMs face in contradiction detection and emphasize the need for robust solutions to improve factual reliability.