LLM hallucination study

• المؤلف الرئيسي: Potdar, Prateek Anish
• مؤلفون آخرون: Jun Zhao
• التنسيق: Final Year Project
• اللغة: English
• منشور في: Nanyang Technological University 2025
• الموضوعات: Computer and Information Science; LLM; Hallucination; RAG
• الوصول للمادة أونلاين: https://hdl.handle.net/10356/183825

Large Language Models (LLMs) exhibit impressive generative capabilities but often produce hallucinations—outputs that are factually incorrect, misleading, or entirely fabricated. These hallucinations pose significant challenges in high-stakes applications such as medical diagnosis, legal reasoning, financial analysis, and scientific research, where factual accuracy is critical. As LLMs become increasingly integrated into real-world systems, mitigating hallucinations is essential to ensuring reliable, trustworthy, and ethically sound AI deployments. Without effective strategies to reduce hallucinations, AI-generated content risks contributing to misinformation, undermining user trust, and limiting the adoption of LLMs in professional domains. My report investigates techniques to reduce hallucinations through systematic experimentation on Meta’s LLaMa model, a state-of-the-art open-source LLM. Specifically, I explored the impact of key generative parameters, including temperature scaling, top-k sampling, and retrieval- augmented generation, on factuality and coherence. These parameters play a crucial role in balancing response creativity and accuracy, directly influencing the probability of hallucinated content. By carefully tuning these hyperparameters and integrating external knowledge retrieval, I aimed to assess how different configurations affect the reliability of LLaMa’s generated responses. I systematically evaluated the effectiveness of these mitigation techniques using factuality scoring and response coherence analysis. Factuality was assessed by measuring the alignment of generated responses with authoritative sources, while coherence analysis examined the logical consistency and contextual appropriateness of outputs. Results from these experiments provide quantitative insights into the trade-offs between factual reliability, creativity, and response variability, offering practical guidelines for optimizing LLMs across different use cases.