Hallucinated citations—references that do not exist or are irrelevant—are a growing problem in academic publishing. A recent audit by GPTZero found 53 accepted papers at NeurIPS 2025 with such fabricated references, prompting venues to introduce automated screening tools for submissions.
These tools, which range from simple rule-based checks to large language model (LLM) verifiers, are emerging as a first line of defense. But their reliability is not yet well understood. The paper "HALLMARK: How to Diagnose LLM Citation Verifiers and When to Deploy Them" (arXiv:2607.18360) takes on this issue directly.
The authors propose a diagnostic framework to assess the strengths and weaknesses of existing citation verifiers. Rather than assuming any single checker is trustworthy, HALLMARK offers a way to systematically test them and decide when they are appropriate for use in a given workflow.
This matters because false positives or false negatives in citation screening can either waste researchers' time or let hallucinated references slip through. A principled diagnostic approach, as outlined in the paper, could help venues choose the right tool for the job.