Graph2Idea:Retrieval-Augmented Scientific Idea Generation with Graph-Structured Contexts

A new framework called Graph2Idea aims to improve how large language models generate scientific research ideas by replacing flat text summaries with structured knowledge graphs, according to a paper posted to the arXiv preprint server [1]. The framework retrieves papers on a given topic, converts them into structured knowledge triples, and builds a target-centered knowledge graph to make relationships across papers explicit [1]. It then extracts compact, graph-derived contexts that retain relational evidence while reducing noisy text [1]. A two-stage generation process first identifies promising research directions and then guides the LLM to synthesize candidate ideas from the graph-grounded evidence [1]. In benchmark tests, Graph2Idea outperformed representative baselines under an automatic evaluation protocol [1]. Compared with the strongest baseline scores, the framework improved Novelty from 0.45 to 0.52, Quality from 0.24 to 0.29, and Feasibility from 0.22 to 0.28 [1]. The authors argue that graph-structured evidence enables more explicit, compact, and traceable recombination of prior scientific knowledge [2]. Large language models, which are neural networks trained on vast amounts of text for generation and analysis tasks, have been increasingly applied to scientific workflows [8]. However, when used for idea generation, they typically receive retrieved literature as flat text — titles, abstracts, or summaries — which can contain redundant or weakly relevant information and obscure cross-paper relations among problems, methods, and findings [1]. The paper was submitted to arXiv on 8 June 2026 [1]. arXiv, founded in 1991, is an open-access repository of electronic preprints that has grown to host more than two million articles, with a submission rate of roughly 24,000 papers per month as of late 2024 [6]. The repository is not peer-reviewed; papers are approved for posting after moderation [6]. The platform also hosts arXivLabs, a framework that allows community collaborators to develop and share experimental tools directly on article pages, such as bibliographic explorers and literature-mapping services [5].