New Paper Proposes CiteAnchor to Catch Hallucinated Citations

Citations that look plausible but point nowhere real have become a familiar failure mode of generative AI writing tools, and a new preprint tries to separate two questions that most detection methods blur together: is a citation valid, and can its source actually be found. The authors call the problem bibliographic drift. Records for the same work can differ across preprint versions, journal reprints and conference proceedings, so a system that only checks whether a cited title and author list look like a real paper will pass through citations that are internally consistent but unverifiable, and reject some genuine ones whose metadata happens to vary between databases.

The paper’s first contribution is CiteResolve, a benchmark built from 423 real-world hallucinated citations, each one annotated for whether a unique underlying source can be recovered at all. That annotation step matters because not every fabricated-looking citation is equally fixable: some can be traced to a real document once a search strategy accounts for the right source type, while others genuinely have no retrievable match.

The second contribution is CiteAnchor itself, a framework that adapts its retrieval and field-level verification to the kind of source being checked, rather than applying one matching strategy to every reference. It treats validity and source recoverability as separate judgments instead of collapsing them into a single plausibility score.

The numbers

On an existing hallucinated-citation detection benchmark, CiteAnchor reaches an F1 score of 0.841 when paired with the GPT-5.6 Luna backbone and 0.784 with an open-weight backbone, which the authors report as gains of 7.9 and 36.8 points over a baseline using the same backbones. On the authors’ own CiteResolve benchmark, the framework achieves 74.7% and 70.7% end-to-end task success with those same two backbones, which the paper describes as large improvements over web search baselines.

The authors have released their code and data, which is what lets the field check these numbers rather than take them on trust, the same gap the paper is trying to close for citations themselves.