Engineering the Reproducible Literature Review Section for Scholarly Publications and Grant Applications

Large language models (LLMs) have the potential to transform the synthesis of scientific knowledge. While literature review sections generated with the assistance of LLMs raise legitimate concerns due to limitations of the technology, researchers' interest in automation brings a rare opportunity to change scientific practice to increase the robustness and reproducibility of literature review sections. This position paper proposes a digital object called a reproducible literature review section containing a discourse graph and a bibliography in a computable format. By leveraging technologies including query-focused summarization with retrieval-augmented generation, discourse graphs, and scholarly big data infrastructure, the reproducible literature review section could address trust issues with human-generated literature review sections and LLM-generated text.