Select, Read, and Write: A Multi-Agent Framework of Full-Text-based Related Work Generation

TL;DR

This work addresses the challenge of generating high-quality related work sections from full texts by introducing a memory-driven, three-agent framework (selector, reader, writer) that iteratively reads paper sections. It leverages two explicit graph structures—co-occurrence graph $G_{co}$ and citation graph $G_{ci}$—through graph-aware selectors to better capture inter-paper relationships and reading order. Extensive experiments on the OARelatedWork dataset show that full-text RWG models outperform abstract-based baselines, with Graph-Ci (citation-graph constrained reading) providing the strongest improvements across multiple base models and evaluation metrics, including human judgments. The approach highlights the practical potential of structured, graph-guided, memory-aware reading for automating scholarly writing, while acknowledging gaps to human performance and the need for automated reference retrieval in future work.

Abstract

Automatic related work generation (RWG) can save people's time and effort when writing a draft of related work section (RWS) for further revision. However, existing methods for RWG always suffer from shallow comprehension due to taking the limited portions of references papers as input and isolated explanation for each reference due to ineffective capturing the relationships among them. To address these issues, we focus on full-text-based RWG task and propose a novel multi-agent framework. Our framework consists of three agents: a selector that decides which section of the papers is going to read next, a reader that digests the selected section and updates a shared working memory, and a writer that generates RWS based on the final curated memory. To better capture the relationships among references, we also propose two graph-aware strategies for selector, enabling to optimize the reading order with constrains of the graph structure. Extensive experiments demonstrate that our framework consistently improves performance across three base models and various input configurations. The graph-aware selectors outperform alternative selectors, achieving state-of-the-art results. The code and data are available at https://github.com/1190200817/Full_Text_RWG.

Figures (11)

• Figure 1: Overview of our multi-agent framework. The framework comprises a selector, a reader, and a writer, which collaboratively read the papers and generate the related work section.
• Figure 2: Illustration of our graph-aware selector. (a) Under the constraints of the graph structure, the selector selects either to continue reading the current paper or jump to an adjacent paper. We design two types of graphs: a (b) citation graph and a (c) co-occurrence graph.
• Figure 3: Performance comparison of five different selector strategies across three base models: (a) average number of edges in graph-based metrics, (b) overall LLM-based evaluation.
• Figure 4: A case study comparing the RWS generated by GPT-4o and $\text{GPT-4o}_\text{ Graph-Ci}$. On the right, a co-occurrence graph for graph-based metrics is constructed from the generated RWS. Our $\text{GPT-4o}_\text{ Graph-Ci}$ model gives a more cohesive and interrelated explanation of the references, which is much easier for readers to follow. In contrast, GPT-4o fails to establish connections between references.
• Figure 5: The proportion of sections that are selected for reading by $\text{GPT-4o}_\text{ Graph-Ci}$.