Analysis of LLM-generated citation trends on arXiv
Trends in LLM-Generated Citations on arXiv
AI summary
The proportion of references on arXiv suspected to be hallucinated by LLMs is showing an accelerating upward trend, with a noticeable intensification starting in early 2025, roughly coinciding with the release of Deep Research. The author used pdftotext++ and AnyStyle to extract citations from arXiv PDFs since 2020, marking references where "the title matches an existing arXiv paper but the author list differs" as candidate hallucinations, then used Claude Sonnet 4 to filter false positives. Currently, about 0.025% (approximately 1 in 4000) of references are suspected hallucinations, which the author believes is almost certainly a significant underestimate.
Related papers