Research Summary: The Shrinking Lifespan of LLMs in Science

Research from arXiv highlights that Large Language Models (LLMs) used in scientific contexts are exhibiting a shrinking lifespan and faster obsolescence. This study introduces 'time-to-peak' and 'lifespan' metrics to assess model relevance, finding that a model's release year is a stronger predictor of its longevity than its architecture, openness, or scale, with adoption trajectories following an inverted-U curve.

Why it matters

The rapid obsolescence of LLMs implies that organizations investing in or relying on these models for critical operations or research must account for shorter effective lifespans. This necessitates agile strategies for technology adoption, continuous evaluation, and planned transitions to newer models to maintain competitive advantage and operational relevance.

Key insights

• New metrics, 'time-to-peak' and 'lifespan,' have been introduced to quantify the obsolescence and scientific adoption trajectories of LLMs.
• Analysis of 62 LLMs across over 108,000 scientific papers (2019-2025) indicates that active adoption can be distinguished from background citation.
• A model's release year is a more significant predictor of its 'time-to-peak' and 'lifespan' than its technical characteristics such as architecture, openness, or scale.
• The adoption of LLMs in science generally follows an inverted-U curve, showing initial growth, a peak in relevance, and subsequent decline.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2604.07530