Beyond Discrete Tokens: Preserving Numerical Continuity in Medical Vision-Language Models
HAD-VLM combines text generation with diffusion-based continuous predictions for medical image measurements. On evaluated ultrasound and X-ray tasks, it reduced measurement error against the reported baselines.
HAD-VLM pairs autoregressive text generation with diffusion-based prediction of measurement coordinates and lengths. Across four echocardiographic tasks, it reported **21.4–37.6% lower error** than a Vicuna-based autoregressive baseline; for X-ray vertebral measurement, error fell from **11.4 to 6.7 pixels**.
For builders handling images with precise numeric outputs, the paper is an example of separating language generation from continuous measurement prediction while sharing a model backbone. It also reports that medical question-answering performance was preserved on the evaluated dataset.
HAD-VLM pairs autoregressive text generation with diffusion-based prediction of measurement coordinates and lengths. Across four echocardiographic tasks, it reported **21.4–37.6% lower error** than a Vicuna-based autoregressive baseline; for X-ray vertebral measurement, error fell from **11.4 to 6.7 pixels**. For builders handling images with precise numeric outputs, the paper is an example of separating language generation from continuous measurement prediction while sharing a model backbone. It also reports that medical question-answering performance was preserved on the evaluated dataset. The findings cover specific medical image tasks and datasets. The abstract does not establish performance beyond those evaluations or provide deployment details.
This narrows multimodal medical measurement toward a split output design: language remains autoregressive while coordinates and lengths use diffusion, with gains reported on the evaluated tasks. It complements broader medical image understanding work with a specific approach to continuous measurements; the evidence does not establish performance beyond those datasets.