RETRIEVAL-AUGMENTED GENERATION FOR SCIENTIFIC KNOWLEDGE SYNTHESIS: A REVIEW OF RETRIEVAL, QUESTION ANSWERING, SUMMARIZATION, AND EVALUATION

Abstract

The growth of literature is happening faster, making it hard for researchers to find the information they need. They have to look through a lot of things to get the evidence and put together what they find from different studies and make sure their knowledge is accurate. Large Language Models are good for natural language understanding and generation, but they depend on static knowledge, which leads to hallucinations, outdated information and limited transparency in the research domain. One way to solve this problem is the use of Retrieval Augmented Generation. It integrates external knowledge with LLMs to produce evidence for Scientific Questions and Document Summarization. It explores the current improvements in retrieval architectures, including adaptive, hierarchical, graph-based and self-improving strategies, with their applications in literature retrieval, multi-document reasoning and knowledge synthesis. This review discusses evaluation frameworks and benchmark datasets used to assess retrieval quality, citation grounding, answer reliability and faithfulness to the retrieved evidence. In addition, it also identifies current challenges related to retrieval scalability, long-context reasoning, multi-modal research, standardized data evaluation and computational efficiency. Unlike existing surveys that explore individual aspects of RAG, this review uses a unified framework that integrates retrieval architectures, research applications, and evaluation methodologies. It also provides a direction toward trustworthy, scalable and evidence-grounded knowledge systems.