RETRIEVAL-AUGMENTED GENERATION FOR SCIENTIFIC KNOWLEDGE DISCOVERY: IMPROVING FACTUALITY AND CITATION RELIABILITY

Abstract

The rapid growth of scientific publications has created a substantial challenge for researchers seeking to identify, synthesize, and verify relevant knowledge. Large Language Models (LLMs) provide powerful capabilities for scientific question answering and literature synthesis, but their reliance on parametric knowledge can result in hallucinated facts, outdated information, incomplete evidence, and unreliable citations. Retrieval-Augmented Generation (RAG) offers a promising solution by connecting generative models to external scientific literature during inference. However, conventional RAG systems may still retrieve irrelevant passages, overlook contradictory evidence, generate unsupported claims, or provide citations that do not adequately support the associated statements. This paper proposes a reliability-oriented RAG framework for scientific knowledge discovery that combines domain-specific retrieval, hybrid semantic and lexical search, cross-encoder reranking, evidence-aware generation, citation verification, and post-generation self-reflection. The framework treats factuality and citation reliability as separate but interconnected objectives. A structured evaluation methodology is proposed using retrieval metrics, answer faithfulness, factual correctness, citation precision, citation recall, citation completeness, and citation faithfulness. Recent research on scientific literature synthesis demonstrates that retrieval-based systems can substantially improve scientific answer quality and citation accuracy compared with language-model-only approaches. The proposed framework provides a practical architecture for developing transparent, evidence-grounded scientific assistants capable of supporting literature review, hypothesis exploration, methodological comparison, and knowledge discovery while reducing hallucination and citation errors.