AI
New Detection System Flags AI-Fabricated Citations in Research Papers
Researchers develop a multi-agent framework that identifies fake references with 97% accuracy and distinguishes between genuine, borderline, and entirely fabricated citations.
Large language models have become commonplace in academic writing, but they frequently generate plausible-sounding citations that are actually invented—references that pass a casual glance but crumble under verification. A new computer science research paper tackles this growing problem with a structured detection system.
The Citation Hallucination Problem#
When language models produce citations, they often create references that sound authentic but don't actually exist in any bibliography. Traditional verification methods treat the result as a simple binary check: either the citation is found or it isn't. This approach leaves little room for nuance and provides auditors with minimal insight into why a citation fails to verify. As LLMs increasingly find their way into scientific publishing, the stakes for catching these fabrications grow.
CiteTracer: A Multi-Layered Verification Approach#
The researchers introduced a 12-category taxonomy that moves beyond binary judgment, sorting citations into three broader classes—Real, Potential, and Hallucinated—with finer-grained codes underneath. They then built CiteTracer, a cascading multi-agent system that operates in stages:
- Extracts structured citations from PDFs and BibTeX files
- Searches for evidence via local cache lookup, direct URL fetching, academic connector APIs, and web search
- Applies rule-based field matching (authors, titles, publication dates, venues)
- Escalates uncertain cases to specialist judges trained on each class
Performance and Real-World Testing#
On a synthetic benchmark of 2,450 controlled examples—built by starting with real citations and applying LLM mutations—CiteTracer achieved 97.1% overall accuracy. The system also performed well on each class separately, with F1 scores of 97.0 for real citations, 95.8 for borderline cases, and 98.5 for hallucinations.
The researchers tested the system against 957 genuine fabricated citations harvested from desk-rejected conference submissions and the ICLR 2026 conference. In this real-world scenario, CiteTracer caught 97.1% of fabrications without flagging uncertain cases for manual review.
