Graph Your Way to Inspiration: Integrating Co-Author Graphs with Retrieval-Augmented Generation for Large Language Model Based Scientific Idea Generation
GYWI combines author knowledge graphs with retrieval-augmented generation to help LLMs generate more novel, feasible, and relevant scientific research ideas.
The Problem
LLMs can generate scientific ideas, but the results often lack controllable academic context and traceable inspiration pathways. Generated ideas may be generic or disconnected from real research trajectories.
The GYWI System
Author Knowledge Graphs: Build an author-centered knowledge graph construction method with inspiration source sampling algorithms to create an external knowledge base with academic context.
Hybrid Retrieval: Combine standard RAG with GraphRAG to retrieve both broad and deep knowledge, forming a hybrid context for the LLM.
Prompt Optimization: Use reinforcement learning principles to automatically guide LLMs in optimizing generated ideas based on the hybrid context.
Evaluation
Tested on arXiv papers (2018-2023) using GPT-4o, DeepSeek-V3, Qwen3-8B, and Gemini 2.5. Ideas evaluated across five dimensions:
| Dimension | Description |
|---|---|
| Novelty | How original the idea is |
| Feasibility | Can it actually be done |
| Clarity | How well-articulated |
| Relevance | Fit to the research area |
| Significance | Potential impact |
GYWI significantly outperforms mainstream LLMs across multiple metrics.
Source: arXiv:2602.22215