AI-Driven Automation Can Become the Foundation of Next-Era Science of Science (arxiv.org)

This paper argues that AI-driven automation can transform the Science of Science (SoS) by enabling large-scale, data-driven discovery of patterns and mechanisms that traditional tools—often limited to linear regression and rule-based simulations—fail to capture. The authors present a forward-looking perspective showing how modern AI techniques can automate the extraction of complex, multi-dimensional signals across publications, citations, collaborations and researcher behavior, and they demonstrate a preliminary multi-agent system that simulates research societies to reproduce real-world research patterns. The result is a blueprint for an automated SoS that can accelerate hypothesis generation, policy testing, and meta-scientific insight at scales previously infeasible. Technically, the paper contrasts classical statistical and agent-based approaches with AI’s capacity for scalable pattern recognition and simulation, while acknowledging important limitations—model validity, interpretability, data quality and biases—and outlining pathways to mitigate them. Suggested directions include integrating domain knowledge with learning-based models, improved evaluation and benchmarks for SoS tasks, and human–AI workflows to ensure robustness and ethical use. For AI/ML researchers this work signals rich opportunities: building specialized models and simulators for scholarly ecosystems, developing causal and interpretable tools for meta-science, and creating standardized datasets and evaluation protocols to turn SoS into a more predictive, automated discipline.