Generalization and the Rise of System-level Creativity in Science

A new study by researchers Hongbo Fang and James Evans, published on arXiv, uses citation network analysis from tens of millions of publications to reveal a fundamental shift in the nature of scientific production. The research identifies three distinct functional roles for scientific contributions: foundations (building core ideas), extensions (elaborating within a field), and generalizations (compressed, modular work reused across distant disciplines). The analysis shows a stark trend: while foundational and extensional work dominated post-war science, it has steadily declined since the early 1990s. In contrast, generalizations have risen sharply, indicating that the locus of innovation is moving from deep within disciplines to the interfaces between them.

This shift is not merely correlational. The researchers employed stacked difference-in-differences analyses, exploiting natural experiments like journals' transitions to online access and authors' adoption of large language models (LLMs), to provide causal evidence. The findings strongly suggest that digital knowledge infrastructure—including widespread online publishing and AI tools—is actively driving this reorganization. The paper argues that the much-discussed 'decline of disruption' in science is not a slowdown but a structural change in how breakthroughs occur. This creates a growing misalignment, as the current system for recognizing and rewarding scientists (e.g., through citations within narrow fields) is poorly suited for valuing the cross-disciplinary, generalizing work that is increasingly driving progress.