The Shrinking Lifespan of LLMs in Science

Scaling laws describe how language model capabilities change with compute and data. We ask a complementary question: what governs how long a language model matters after it is released? We examine the adoption of large language models (LLMs) as a scientific tool over time, by classifying citations to 64 LLMs across over 108,000 papers (2018–2025) as either active model use or background reference. From these usage trajectories, we establish three empirical regularities. First, scientific adoption follows an inverted-U trajectory (rising after release, peaking, then declining as newer models displace it), a pattern we confirm with formal tests and term the scientific adoption curve. Second, this curve is compressing: each additional year of model release is associated with a 25\% reduction in time-to-peak scientific adoption ($p < 0.001$), robust to minimum-age thresholds and controls for model size. Third, temporal competitive dynamics dominate model characteristics as a predictor of lifecycle dynamics: release year explains both time-to-peak and lifespan more strongly than openness, architecture, size, or training methodology. Together, these findings complement capability scaling laws with adoption-side regularities, and suggest that the forces driving rapid capability progress may be the same forces compressing scientific relevance.