Deep-Ideation: Generating Novel Research Ideas with Evolving LLM Agents on Scientific Concept Network

ACL ARR 2026 January Submission10371 Authors

Keywords: Large Language Model, LLM Agent

Abstract: Novel research ideas play a critical role in advancing scientific inquiries. Recent advancements in Large Language Models (LLMs) have demonstrated their potential to generate novel research ideas by leveraging large-scale scientific literature. However, previous works face substantial challenges since they heavily rely on the textual content of scientific literature, but overlook the rich semantics and high-order connections embedded in scientific concept networks. To address these challenges, we propose Deep-Ideation, a powerful LLM agent that generates high-quality research ideas via iteratively searching for novel yet feasible combinations on scientific concept network. Our framework introduces an Explore-Expand-Evolve workflow to facilitate continuous dynamic interaction with the scientific concept network. This mechanism drives a structural evolution from concept accumulation to deep innovation, utilizing an Idea Stack to track the ideation trajectory and align it with overall research trends. Furthermore, we fine-tune a Critic Model on real-world reviewer feedback to align with expert standards, rigorously steering ideation toward greater novelty and feasibility. Extensive experiments demonstrate that Deep Ideation achieves publication-quality performance, with about 81.5\% of generated ideas surpassing the acceptance scores of top AI conferences. Furthermore, expert human evaluations corroborate these findings, confirming that Deep Ideation produces professional-grade research proposals. Our code is open-sourced for reproducibility at https://anonymous.4open.science/r/Deep_Ideation-E385.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: Autonomous agents, LLM agents

Languages Studied: English

Submission Number: 10371