From extractive to generative: multi-document summarization in the era of generative AI - advances, challenges, and emerging trends

Abstract

Multi-Document Summarization (MDS) has evolved from extractive techniques to sophisticated generative models, driven by advances in Large Language Models (LLMs), Retrieval-Augmented Generation (RAG), diffusion-based generation, and graph-enhanced architectures. While these methods have significantly improved the coherence, informativeness, and factuality of summaries, the growing diversity of approaches has introduced challenges in understanding their design trade-offs, domain applicability, and evaluation. This survey provides a comprehensive review of generative MDS models, offering a structured taxonomy that categorizes them into LLM-based, RAG-enhanced, graph-augmented, and diffusion-based families. We analyze each category’s architectural principles, capabilities, and limitations. In addition, we highlight benchmark datasets, evaluation protocols, and key application domains, including healthcare, legal, and scientific summarization. To move beyond descriptive comparison, we propose a unified analytical framework that decomposes generative MDS into three core dimensions: information aggregation, structural reasoning, and factual grounding. This framework organizes prior work, clarifies architectural trade-offs, and connects current limitations to actionable future research directions. Finally, we discuss open challenges such as hallucination, domain adaptation, and evaluation bottlenecks, and propose future research directions. This survey aims to guide researchers and practitioners in navigating the rapidly evolving landscape of trustworthy and scalable generative MDS.