Progress and Challenges in Addressing Hallucination Issues in Large Language Models

Abstract

The illusion problem of large language models refers to the phenomenon where the content generated by the model is inconsistent with the input information or objective facts, significantly limiting its reliability and safety. This paper systematically reviews the definition and classification of the illusion problem, including factual illusions and fidelity illusions, intrinsic and extrinsic illusions, as well as types of closed-domain and open-domain illusions. It also summarizes the current mainstream evaluation methods from three perspectives: data, models, and multi-task applications. Regarding mitigation strategies, the article proposes various technical paths at the data, model, and application levels, including data cleaning and augmentation, model architecture optimization, prompt engineering, and real-time retrieval-augmented generation, which effectively enhance the accuracy and consistency of the generated content. Future research should aim to establish a more refined evaluation system, promote collaborative optimization across multiple technologies, and achieve dynamic knowledge updates and lightweight deployment to strengthen the practicality and safety of large models in real-world scenarios.