Reproduce “ResearchArcade: Graph Interface for Academic Tasks” on OpenResearch

About this paper

Jingjun Xu, Chongshan Lin, Haofei Yu, Tao Feng, Jiaxuan You

RESEARCHARCADE introduces a unified, graph-based data interface for developing machine learning models to support diverse academic tasks, integrating multi-source and multi-modal information from platforms like ArXiv and OpenReview. The framework effectively models relational data and dynamic evolution, demonstrating enhanced performance for various predictive and generative tasks by providing a standardized input for LLMs, GNNs, and other architectures.

Key insights

• A unified, graph-based data interface is effective for integrating diverse academic information, significantly simplifying the development of machine learning models for a wide range of academic tasks.
• Incorporating graph structures with multi-hop aggregation consistently provides performance improvements across predictive academic tasks, highlighting the critical role of relational context.
• The inclusion of multi-modal information (figures and tables) is crucial for enhancing model performance in both predictive and generative tasks, leading to richer contextual understanding.

Summary by alphaXiv.