Generating interesting scientific ideas using knowledge graphs and LLMs: Evaluations with 100 research group leaders.

SCIMUSE generated 4,400+ personalized research ideas from a knowledge graph of 58 million papers, but expert interest was modest (mean 2.40/5; most common 1).

• Why it matters: The work addresses the difficulty of finding novel, impactful ideas amid rapidly expanding literature, and tests whether AI ideation can produce compelling suggestions across disciplines.
• What they did: Using SCIMUSE, the team evaluated ideas rated by 100+ research group leaders, comparing knowledge-graph concept-pair prompting against a titles-only GPT baseline and analyzing graph-feature controls.
• The result: Adding knowledge-graph concept pairs did not improve interest versus the baseline, and high-citation-predicted pairs showed a weak tendency toward lower interest; however, interest was predictable via a supervised neural network and an LLM zero-shot ranking.

AI summary of the abstract · may contain errors · not medical advice · read the paper.