کنفرانس بین المللی محاسبات نرم

صفحه اصلی / ششمین کنفرانس بین المللی محاسبات نرم

When Autocomplete Lies - An Evaluation On LLM Hallucination

نویسندگان :

Reza Shabankhah¹ Amirhossein Moradi² Abdorreza Hesam mohseni³

1- دانشگاه گیلان - دانشکده فنی و مهندسی شرق گیلان 2- دانشکده فنی مهندسی شرق گیلان/گیلان 3- دانشکده فنی مهندسی شرق گیلان/گیلان

کلمات کلیدی :

Hallucination Detection،Retrieval-Augmented Generation،Model Trustworthiness،Factuality Verification,،Mitigation Strategies

چکیده :

Large Language Models (LLMs) have changed how we process natural language. Still, they have a major problem: hallucination. This means they can create believable text that is wrong or not based on evidence. This paper looks at this issue by sorting hallucinations into factual, logical, intrinsic, and extrinsic types. We look at the reasons why this happens. These reasons include the statistics used to train these models, weaknesses that allow data poisoning, and how tests reward guessing instead of showing uncertainty. The paper looks at ways to spot hallucinations, such as self-checking, checking against known information, and using neuro-symbolic methods. It also reviews ways to reduce hallucinations, including using retrieval-augmented generation, prompt engineering, parameter-efficient fine-tuning, and specific decoding techniques. Our study shows that while theory says we can't completely get rid of hallucinations, we can greatly reduce them by using many methods together. This will help create more reliable AI systems.