Automated Literature Review Generation System

Kechen Fang, Xinru Liang, Zijing Xu

Keywords: Automated Literature Review, Multi-Agent Systems, Large Language Models, Retrieval-Augmented Generation

Abstract: The exponential growth of academic literature poses a significant challenge to traditional manual review methods. To address this, we propose an automated literature review generation system leveraging Large Language Models (LLMs) within a multi-agent architecture. This approach decomposes the complex generation task into three specialized components: a Retrieval Agent ($A_R$), an Analysis Agent ($A_A$), and a Writing Agent ($A_W$). The Retrieval Agent gathers a comprehensive document set using citation-informed embeddings (e.g., SPECTER). The Analysis Agent performs structured knowledge distillation, transforming documents into a compact knowledge base ($K$) to overcome LLM context window limitations. Finally, the Writing Agent ($A_W$) drafts the review. This agent employs a Retrieval-Augmented Generation (RAG) process for factual grounding and includes a post-refinement loop with a reflection mechanism, enabling it to re-invoke retrieval for enhanced coverage and citation accuracy. This modular framework aims to significantly reduce hallucinations and improve the coherence, accuracy, and quality of automated academic reviews.

Submission Number: 23