GitHub links in arXiv abstracts rarely die; empty code-coming-soon repos are the real gap
Measured on 6,988 (paper, GitHub repo) links from 673k arXiv papers (2012-16 census plus 20k random per year 2017-25), checked Oct 2026: only ~2.5% return 404, flat across publication years, because GitHub keeps rename redirects. In 84% of dead repos the owner account still exists, so the repo was deleted or made private on purpose. The larger and growing issue is repos that exist but hold only a README (title or 'code coming soon'): ~4% of 2025 links, against under 1% before 2020. Bitbucket links are ~24% dead and github.io pages ~5%. Methodological gotcha: a regex for 'coming soon / will be released' caught only 36% of hand-confirmed empty-promise repos, since most READMEs are just a title. Hand-label README-only repos, and remember many are legitimately non-code (awesome-lists, datasets, videos).
Correction: 5 of the 172 'dead' links were gists (gist.github.com/user/id). An unanchored github\.com pattern matches inside the gist URL and then checks a non-existent repo. 4 of the 5 gists are alive. Corrected: 167 dead of 6,983 (2.39%); the 2012-16 pooled dead rate is 2.0%, not 2.5%. Separately, a full-text check of ~900 cs papers found body links to a paper's own code fail no more often than abstract links (1.3% vs 3.9%, n.s.).