Topical Phase Transitions in Artificial Intelligence Research: Large-Scale Evidence and an Early-Warning Signature for Emerging Topics

The article presents a comprehensive analysis of 80,814 AI research papers from major conferences, revealing that significant AI topics experience abrupt growth through "topical phase transitions." Key findings indicate that large language models became the predominant topic by 2025, with diffusion models also emerging rapidly, while reinforcement learning displayed smoother growth patterns. The authors propose an early-warning signature based on publication dynamics to identify emerging topics, flagging areas such as reasoning, agentic AI, and multimodal LLMs for future monitoring, which could aid researchers in anticipating shifts in AI research focus.