A Persona-Based Collaborative Multi-Agent Framework for Scientific Idea Generation

Abstract

Automated research ideation systems have recently benefited from large language models and show strong capabilities in scientific reasoning; however, these systems often generate ideas that lack novelty and fail to effectively leverage targeted literature when operating without structured guidance. This reveals a gap in current ideation frameworks, which typically do not incorporate structured workflows, targeted retrieval, and collaborative reasoning processes. To address this, we propose a query-driven agentic research ideation pipeline that decomposes the scientific discovery workflow into four stages: query decomposition, agentic literature discovery through iterative retrieval and refinement, AI scientist persona construction, and multi-agent deliberation for idea generation and critique. Experimental evaluation against the VirSci baseline demonstrates that the proposed system produces more novel and higher-quality research ideas than single-agent approaches, with ablation studies showing that query decomposition, targeted retrieval, and structured critique each contribute significant improvements, resulting in higher precision while maintaining strong conceptual originality.