Graph Your Way to Inspiration: Integrating Co-Author Graphs with Retrieval-Augmented Generation for Large Language Model Based Scientific Idea Generation

Researchers Pengzhen Xie and Huizhi Liang have introduced GYWI (Graph Your Way to Inspiration), a novel system designed to enhance the scientific ideation capabilities of Large Language Models (LLMs). The core innovation addresses a critical weakness: standard LLM-generated ideas often lack academic grounding and traceable inspiration. GYWI bridges this gap by constructing an external knowledge base from author-centered knowledge graphs and employing a hybrid retrieval mechanism that combines traditional RAG with GraphRAG. This provides LLMs with a rich, structured context that includes both the depth of specific papers and the breadth of collaborative networks.

The system's technical pipeline involves three key components: building co-author graphs from academic data, a hybrid retrieval system for context, and a reinforcement learning-optimized prompt strategy to guide the LLM. Evaluated on a dataset built from arXiv (2018-2023) and tested across models including GPT-4o, DeepSeek-V3, Qwen3-8B, and Gemini 2.5, GYWI demonstrated significant improvements over standalone LLMs. The comprehensive assessment, which included automatic metrics, LLM-based scoring, human evaluation, and semantic visualization, measured five dimensions: novelty, feasibility, clarity, relevance, and significance. The results indicate a major step toward making AI a more reliable and context-aware partner in early-stage scientific research and brainstorming.