The Shrinking Lifespan of LLMs in Science

The study introduces metrics such as time-to-peak and lifespan to evaluate the obsolescence of 62 language models (LLMs) based on over 108,000 citing papers from 2019 to 2025. It finds that the release year of a model is a stronger predictor of its adoption trajectory than its architecture or scale, with each successive release year leading to a 27% shorter time-to-peak and a 23% shorter lifespan. This trend indicates that reliance on individual models may become increasingly risky for practitioners, impacting reproducibility and the need for frequent model updates.