Answers are grounded in this paper and cite their sources.
Recent AI scientist systems show promise for autonomous ideation, yet existing multi-agent approaches often rely on rigid protocols and coarse novelty metrics. This frequently leads to expert dominance, communication bottlenecks, and incremental innovations. To address these limitations, we propose a multi-agent scientific ideation framework that combines a World Café structure for rotating group discussion, an adaptive turn-taking scheduler called MeshScheduler, and the Relative Novelty Density (RND) metric for fine-grained novelty assessment. Evaluated using LLaMA3.1 under regulated discussion costs, our framework achieves superior Overall Novelty (ON) and Contemporary Impact (CI) compared to strong baselines like AI Scientist, HypoGen, and VirSci. A cross-disciplinary case study on AI-driven heart-sound auscultation further shows that the framework can generate actionable research concepts for clinically relevant problems that require both medical and computational perspectives.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: