Citation Hallucination in LLMs
Citation hallucination is the failure mode in which a large language model produces a confident-looking bibliographic reference — author, title, venue, year, sometimes a DOI — that does not correspond to any real publication. Surveys across models and domains report hallucination rates from roughly 11% to over 90%, with author names the most error-prone field, and recent studies trace measurable contamination of real arXiv papers with fabricated citations.
Citation hallucination is the specific LLM hallucination pattern in which a model emits a plausibly formatted reference — an author list, title, journal or conference, year, sometimes a DOI or URL — that does not match any real work. Because academic and legal writing depends on verifiable references, this failure mode has been studied more closely than generic factual hallucination. Mechanistically, hallucinated citations are an artifact of how base models generate text. The weights have absorbed the surface patterns of citation strings ('Smith, J. et al. (2019). *Title*. Journal of X, 42(3), 117–134.') without storing a reliable index of which strings correspond to actual papers. When asked for a citation the model continues the pattern, producing a string that fits the distribution of real references. Studies have shown that no major model spontaneously emits citations when not asked, suggesting the failure is prompt-induced rather than intrinsic. Empirical surveys published in 2025-2026 give a range of measured hallucination rates. Across state-of-the-art models, rates of fabricated or invalid references span roughly 11% to 95%, with strong dependence on model, domain (biomedical and legal are worst), and prompt framing. Author names are consistently the most error-prone bibliographic field, followed by venue, title, and year. One audit of 2.2 million citations in 56,000+ papers at top AI/ML and security venues from 2020-2025 found that 1.07% of papers contained invalid or fabricated citations, with a steep increase in 2025 traced to LLM-assisted writing. Mitigations cluster into three families. Retrieval-grounded generation (RAG (Retrieval-Augmented Generation): How LLMs Access External Knowledge) replaces model-imagined citations with references to documents actually fetched at query time, which is the approach used by Perplexity AI and the Anthropic Claude Citations API. Multi-model consensus filters — accepting a citation only when three or more independent models produce it — have been reported at around 95% accuracy in detecting fabrications. Post-hoc verification pipelines query bibliographic databases (Crossref, Semantic Scholar, Westlaw) to check whether each cited reference resolves to a real record. None of these eliminates the problem, but together they shrink the surface area where confident-looking but nonexistent references can reach a published artifact.