Using citation networks to teach models how to generate research ideas

Existing methods for automating research idea generation rely on static literature retrieval or prompt engineering without exploiting how papers reference each other. This work proposes Graphs of Research (GoR), which constructs citation-evolution directed acyclic graphs from a seed paper's references, capturing relationships through citation position, frequency, predecessor links, and publication time. The authors built a dataset from five ML/NLP venues (498 training papers, ~7,600 references) and fine-tuned Qwen2.5-7B on structured prompts containing the citation graph and edge signals. In LLM-judge tournaments against GPT-4o baselines, GoR-SFT wins across head-to-head comparisons, suggesting that citation structure is a valuable supervision signal for idea generation.