Large Language Models (LLMs) exhibit remarkable generative capabilities but remain susceptible to hallucinations—outputs that are fluent yet inaccurate, ungrounded, or in-consistent with source material. This paper presents a method-oriented taxonomy of hallucination mitigation strategies in text-based Large Language Models (LLMs), encompassing six categories: Training and Learning Approaches, Architectural Modifications, Input / Prompt Optimization, Post-Generation Quality Control, Interpretability and Diagnostic Methods, and Agent-Based Orchestration. By synthesizing over 300 studies, we identify persistent challenges including the lack of standardized evaluation benchmarks, attribution difficulties in multi-method frameworks, computational trade-offs between accuracy and latency, and the vulnerability of retrieval-based methods to noisy or outdated sources. We highlight underexplored research directions such as knowledge-grounded fine-tuning strategies balancing factuality with creative utility; and hybrid retrieval–generation pipelines integrated with self-reflective reasoning agents. This taxonomy offers both a synthesis of current knowledge and a roadmap for advancing reliable, con-text-sensitive mitigation in high-stakes domains such as healthcare, law, and defense.
Kazlaris et al. (2025) studied this question.