Survey of Evidence-based Text Generation with LLMs

This research article presents a comprehensive survey of evidence-based text generation utilizing large language models (LLMs). The study systematically analyzes 134 papers and introduces a unified taxonomy to clarify the fragmented landscape of this field, which has been characterized by inconsistent terminology and isolated evaluation practices. The authors investigate 300 evaluation metrics across seven dimensions, focusing on approaches that incorporate citations, attribution, or quotations in text generation. The article also examines the distinctive characteristics and representative methods within this paradigm. Key contributions include the identification of emerging research trends, limitations, and future directions in evidence-based text generation. The findings underscore the importance of linking generated content to verifiable sources to enhance the reliability and trustworthiness of LLM outputs. Additionally, the study highlights the need for a consolidated understanding of the various approaches and methodologies currently in use, paving the way for future research in this rapidly evolving area.