Reproduce “GRAFT: Graph-Distilled Generative Retrieval for Facet-Aware Scientific Literature Exploration” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Italo Luis da Silva, Hanqi Yan, Yujing Wang, Jiangnan Ye, Lin Gui, Yulan He
GRAFT, developed by King's College London, is a generative retrieval framework that explicitly models scientific paper relatedness across multiple semantic facets like problems, methods, and contributions, facilitating more interpretable literature exploration. The system recovers 91% of its dense-graph teacher's Recall@20 within its corpus and achieved superior Recall@20 of 0.548 on out-of-corpus queries, demonstrating its ability to discover non-obvious yet relevant connections.
Key insights
- Explicitly modeling heterogeneous relationships between scientific papers through semantic facets enhances interpretability and enables more nuanced, targeted literature exploration.
- Generative retrieval, when properly distilled from a structured graph, can effectively discover 'non-obvious' relevant papers beyond immediate semantic or citation neighborhoods, fostering exploratory discovery.
- Coverage-aware distillation and graph-grounded fusion mechanisms are essential for successfully distilling sparse, structured graph knowledge into a generative model, ensuring comprehensive learning and reliable output.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.