Hallucination Detection and Correction in Medical VLMs via Counter-Evidence Verification

A new framework called Co}unter-Evidence Verification (CoEV) can detect and correct hallucinations in medical vision-language models without retraining, according to a paper posted on arXiv [1]. The method performs bidirectional verification between textual assertions and visual evidence to flag unsupported claims [2]. The paper, submitted on 17 June 2026, addresses a persistent reliability problem for vision-language models (VLMs) in medical diagnosis: hallucinations that undermine clinician trust [1]. Existing detection approaches mainly focus on identifying factual inconsistencies between generated text and reference data, but they rarely verify whether a model's visual attention corresponds to real evidence supporting the text [2]. CoEV closes that gap by testing whether each statement is backed by its corresponding evidence region and assigning it to a four-quadrant diagnostic map that captures both text factuality and visual grounding [2]. The framework is training-free and plug-and-play, meaning it can be applied to existing models without additional training [2]. On hallucination detection, CoEV improved average PR-AUC by 3.0% and ROC-AUC by 3.9% absolute points over existing methods, with gains of up to 18.5% in specific visual question-answering scenarios [2]. For hallucination correction, the framework boosted Micro-F1 by up to 12.5% and reduced hallucination rates by over 11.9% on medical report generation tasks [2]. The authors also report improved accuracy on medical VQA [2]. The paper's authors state the code will be released upon acceptance [1]. The research appears on arXiv, an open-access repository of electronic preprints that is moderated but not peer-reviewed [6]. Founded in 1991, arXiv passed two million articles by the end of 2021 and now receives roughly 24,000 submissions per month [6]. The platform also hosts arXivLabs, a framework launched in 2020 that allows community collaborators to build experimental tools on top of the repository, such as bibliographic explorers and code finders [4]. arXivLabs projects operate under guidelines that require partners to share arXiv's values of openness, community, excellence, and user data privacy [4]. The CoEV paper's abstract page includes links to several such tools, including Bibliographic Explorer and Connected Papers [1].