MedVL-XLRepE cuts non-English medical VQA gap by up to 6.33%

Most medical vision-language benchmarks live in English, and a paper headed to EMNLP 2026 argues that gap hides a scenario-specific pattern of failure the field has not measured. Jingbo Wang, Sendong Zhao, Haochun Wang, Bing Qin and Ting Liu, in the arXiv preprint, built a benchmark spanning eight languages across four scenarios that isolate distinct medical VQA capabilities, and tested it on both open- and closed-source large vision-language models.

The headline finding: "cross-lingual degradation is not uniform but highly scenario-dependent." Some capabilities carry across languages; others collapse. Their fix, MedVL-XLRepE, is described as "a training-free scenario-aware representation engineering method, leveraging LVLMs' superior English medical VQA capability to steer non-English representations toward their English counterparts at inference time." No fine-tuning, no extra data; the intervention runs at inference and pulls hidden states toward what they would look like if the question were asked in English.

The reported number: "Across three LVLMs and eight languages, MedVL-XLRepE consistently mitigates cross-lingual degradation, with gains of up to 6.33%." That figure is a ceiling, not an average. The abstract does not name which model or which language produced it, and it does not describe what the four scenarios contain. Two researchers on our tracker list shared the paper when it appeared.