Weakening in Real Time — The Association Between Artifact Reuse and Citation in Artifact-Dense Research, 2015-2024

Full paper. Companion replication package (data and code) is deposited separately and cited in the paper's Data and code availability section. Abstract. Research artifacts, above all public code repositories, are a primary vehicle through which computational research is used, while citation remains the standard proxy for research impact. Linking 20,529 arXiv papers published between 2015 and 2025 to their author-designated GitHub repositories, we measure reuse by annual fork counts and citation by annual citation counts, and estimate the rank association between the two flows in every combination of publication cohort and calendar year. The association weakens substantially, from roughly 0.45-0.50 in the late 2010s to roughly 0.25 by 2024. The variation lies on the calendar-period axis, not the publication-cohort axis: with period included the cohort coefficient is -0.0017 (p = 0.74), and the 2015 cohort declines from 0.49 to 0.04 over its own lifetime. Period, cohort and age are linearly dependent, so period and age decompositions are observationally equivalent and the data cannot select between them. Two candidate explanations fail. There is no discontinuity at the public release of general-purpose language models (-0.002, 95% CI [-0.061, +0.057]) and no gradient in field-level exposure to them (p = 0.82). Nor does the citation mix degrade: the share of substantively influential citations is flat across the decade. The decline is bounded. Enumerating the full sampling frame of 160,150 papers rather than drawing from it, there is no trend among works that are both substantially reused and substantially cited: -0.0031 per year (95% CI [-0.0156, +0.0094], n = 3,885 across ten cohorts), an interval that excludes the full-sample estimate. The test has 80% power against a slope of 0.0151. The decoupling is a fact about the many rather than the few.