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

Major artificial intelligence research topics do not grow gradually but instead advance through abrupt, detectable jumps, according to an analysis of 80,814 papers from five leading conferences published between 2017 and 2025 [1][2]. The study, posted to arXiv on June 11, 2026, examined accepted main-track papers from ACL, CVPR, ICLR, ICML, and NeurIPS [1][2]. Researchers found that significant topics often remain marginal for years before surging across multiple venues within a span of one to three years, a pattern they term “topical phase transitions” [1][2]. Large language models became the dominant cross-venue topic by 2025, while diffusion models rose with comparable abruptness [1][2]. Language-model methods also crossed into computer vision through vision-language models [1][2]. In contrast, reinforcement learning compounded smoothly, a trajectory the authors use to distinguish genuine phase transitions from ordinary growth [1][2]. The research team then tested whether a phase transition leaves a detectable footprint before it peaks. They defined an early-warning signature using four publication-dynamics criteria frozen on data from 2017 to 2021 [1][2]. When evaluated out of sample on transitions from 2023 to 2025, the signature achieved a precision of 27% and recall of 63% against a 13.5% base rate [1][2]. Applying the signature to 2025 data, the paper flags reasoning and test-time compute, agentic AI, multimodal large language models, retrieval-augmented generation, and world models as topics to monitor over the 2026–2028 period [1][2]. The source code for the analysis has been made publicly available on GitHub [1][2].