Mitigating Hallucinations in Large Language Models: Taxonomies, Detection Frameworks, and Evaluation Paradigms
Year: 2023
Type: article
Abstract: Large Language Models (LLMs) have rapidly transitioned from experimental research artifacts to widely deployed socio-technical systems used in journalism, law, education, healthcare, and scientific research. Alongside their demonstrated fluency and versatility, these models exhibit a persistent and consequential limitation: hallucination, broadly understood as the generation of content that is syntactically plausible but factually incorrect, unverifiable, or internally inconsistent. High-profile incidents—including fabricated legal precedents, false accusations against real individuals, and erroneous scientific claims—have underscored the real-world risks associated with unmitigated hallucinations. This article provides a comprehensive, system-level examination of hallucinations in LLMs, synthesizing empirical case studies, theoretical taxonomies, detection methodologies, mitigation strategies, and evaluation frameworks. Drawing on recent academic literature and documented failures in deployed systems, the paper categorizes hallucinations into factual, logical, contextual, and self-contradictory forms, and analyzes their underlying mechanisms in model training, prompting, and deployment contexts. Particular emphasis is placed on black-box and zero-resource detection approaches, including SelfCheckGPT, Chain-of-Thought consistency methods, and alignment-based evaluation metrics such as GPTScore, GEval, and AlignScore. The discussion critically examines the limitations of current benchmarks and the epistemic challenges of evaluating truthfulness in generative systems. The article concludes by outlining open research challenges and ethical considerations, arguing that hallucination mitigation should be framed not as a problem of complete elimination, but as a continuous risk-management process integrating technical, institutional, and human-in-the-loop safeguards.
Authors: Malhotra, Arjun, Santoso, Rina Putri
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7219698700
