Agentic-Ideation Boosts Scientific Discovery with Efficient LLM Trajectories.

Scientific discovery heavily relies on effective ideation, and recent AI Scientist systems, particularly those leveraging Large Language Models (LLMs), show promise in automating this process. However, existing approaches often depend on rigid, pre-defined workflows, which limits their flexibility in exploring the vast scientific literature and complex reasoning required for research. This research introduces "Agentic-Ideation," a novel framework designed to address these limitations by enabling more flexible and efficient agentic LLMs for scientific ideation. Agentic-Ideation features an automated trajectory synthesis pipeline and a specialized agentic LLM. It defines a comprehensive tool space, including both external and cognitive tools. A key innovation is the Oracle-Guided Data Synthesis strategy, which uses a reference idea as guidance to efficiently reconstruct logical reasoning and tool invocation paths, transforming aimless trial-and-error into directed trajectory generation. The agent is then trained on these synthesized trajectories, with a masking strategy applied to tool execution results to focus on decision-making logic. Experimental results demonstrate that this method significantly outperforms state-of-the-art workflow-based baselines by 11.91% in overall quality and improves the sample efficiency of high-quality data synthesis by over 10 times.