Graphs of Research: Citation Graphs for AI Idea Generation
A new method called Graphs of Research (GoR) uses citation evolution graphs to supervise large language models in generating research ideas. The approach extracts a 2-hop reference neighborhood for each seed paper, deriving relations from citation position, frequency, predecessor links, and publication time, organized into a directed acyclic graph (DAG). The pipeline draws data from five major ML/NLP venues, with 498/50/50 train/validation/test seed papers and approximately 7,600 cited references. The model used is Qwen2.5-7B-Instruct-1M. This work addresses the limitation of existing methods that rely on static retrieval or complex prompt engineering without leveraging structural relations among references.
Key facts
- GoR is a supervised fine-tuning method for research idea generation.
- It extracts a 2-hop reference neighborhood for each seed paper.
- Relations among references are derived from citation position, frequency, predecessor links, and publication time.
- References are organized into a paper-evolution directed acyclic graph (DAG).
- Data is drawn from five major ML/NLP venues.
- The dataset comprises 498/50/50 train/validation/test seed papers and approximately 7,600 cited references.
- The model used is Qwen2.5-7B-Instruct-1M.
- GoR aims to improve upon static retrieval and complex prompt engineering methods.
Entities
—