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

Researchers propose Graph2Idea, an AI framework that uses knowledge graphs to improve scientific idea generation by converting retrieved papers into structured knowledge relationships rather than flat text. The method demonstrates significant improvements in novelty, quality, and feasibility of generated research ideas compared to existing LLM-based approaches.

Graph2Idea addresses a fundamental limitation in how large language models currently approach scientific discovery: the reliance on unstructured textual evidence. While existing retrieval-augmented generation methods retrieve relevant papers, they present this information as disconnected text excerpts, making it difficult for LLMs to identify meaningful connections between problems, methodologies, and findings across different research works.

The framework's innovation lies in transforming retrieved papers into structured knowledge graphs that explicitly map relationships within scientific literature. By doing so, Graph2Idea creates what researchers call "compact graph-derived contexts" that retain relevant evidence while filtering out noise. This structured approach enables more precise knowledge recombination, a crucial capability for generating feasible and novel ideas.

The quantitative results demonstrate substantial improvements over baselines: novelty increased from 0.45 to 0.52, quality from 0.24 to 0.29, and feasibility from 0.22 to 0.28. These gains suggest that LLMs benefit significantly from explicit relational structure when synthesizing scientific knowledge. The two-stage generation process—first identifying promising directions, then synthesizing ideas—mirrors how human researchers approach ideation.

This work has broader implications for AI-assisted scientific research. As knowledge graphs become more sophisticated, they could enhance other knowledge-intensive tasks beyond idea generation, including literature synthesis, hypothesis formation, and experimental design. The approach demonstrates that simply scaling data or model size matters less than improving how information is structured and presented to AI systems.

• →Graph2Idea converts flat paper abstracts into structured knowledge graphs to improve research idea generation by LLMs.
• →The method achieves 15.6% improvement in novelty scores and 20.8% improvement in quality compared to baseline approaches.
• →Knowledge graph-structured contexts enable more explicit and traceable recombination of scientific evidence than flat text retrieval.
• →The framework addresses a critical gap in retrieval-augmented generation by making cross-paper relationships explicit and queryable.
• →Results suggest structuring information for AI systems can be as important as increasing model parameters or dataset size.