The Shrinking Lifespan of LLMs in Science

A new study by researcher Ana Trišović provides the first large-scale empirical analysis of how scientists adopt and abandon large language models (LLMs). Tracking 62 different models across over 108,000 scientific papers published between 2018 and 2025, the research introduces the concept of a 'scientific adoption curve'—an inverted-U pattern where usage rises after release, peaks, and then declines as newer models emerge.

Crucially, the study found this curve is compressing dramatically. Each additional year of release is associated with a 27% reduction in time-to-peak adoption, a statistically significant finding (p < 0.001) that holds even when controlling for model size and other factors. This means newer models reach their peak scientific relevance faster but also become obsolete more quickly.

Perhaps most surprisingly, release timing emerged as the dominant factor in predicting a model's lifecycle, explaining both time-to-peak and overall scientific lifespan more strongly than traditional attributes like architecture, openness, or scale. While model size and access modality (open vs. closed) still predict total adoption volume, when a model is released matters more than what's under the hood.

These findings complement technical 'scaling laws' with adoption-side regularities, suggesting the same forces driving rapid capability progress—faster iteration, more frequent releases—may be shortening the scientific relevance window. For researchers, this creates pressure to constantly update their toolkits and raises questions about citation practices in fast-moving fields.