Medical VQA: candidate-expanding router reaches 92.23 per cent accuracy on mixed-format questions
In an arXiv preprint submitted on 1 September 2026, Hai-Dang Nguyen and Huy-Hieu Pham address a key problem in medical visual question answering: routing questions that arrive in different formats - text, image or report - to the right expert models determines the accuracy of mixed-format pipelines.
Their contribution is a new candidate-expanding router with permutation-stabilized expert ordering, which the authors report reached 92.23 per cent final accuracy across 1,403 retrospective cases. The work shows a practical path for improving mixed-format pipelines through routing and schema controls.
The limitations matter: this is a preprint and an internal retrospective analysis, and author-reported accuracy is not clinical safety or external validity. Published reporting evaluation is essential before any medical use.
