Abstract: [Purpose/Significance] This study aims to quantify the extent to which large language models have penetrated scientific knowledge production and to identify their structural characteristics, thereby providing insights for understanding the evolution of scientific research paradigms and the restructuring of academic norms. [Design/Methodology] Based on the abstract data from SSCI-indexed papers in the field of Information Science & Library Science from 2020 to 2025, we constructed a population-level statistical framework based on maximum likelihood estimation. After verifying the framework's stability and interpretability, we conducted a time-series trend analysis and multidimensional heterogeneity analysis of the proportion of large language model-generated content in academic texts. [Findings/Conclusion] The proportion of text generated by large language models in academic texts within the Information Science & Library Science field experienced a structural breakpoint in 2023, followed by an exponential surge in 2024, reaching 15.4% by the second half of 2025. The proportion of large language model-generated papers is significantly higher among papers with high academic impact and high interdisciplinarity diversity, while it is significantly lower among highly innovative papers. At the same time, large language model adoption has fostered linguistic and resource equity in scientific research. [Originality/Value] This study offers an interpretable perspective on measuring the use of large language models, providing theoretical value and practical guidance for reforming academic norms and evaluation systems, clarifying the boundaries of AI use, and preventing the homogenization of academic content.

Keywords: Large Language Models, Academic writing, Information Science & Library Science, Maximum likelihood estimation, Text generation detection