Video-based remote physiological measurement (RPM) can estimate vital signs like heart rate from ordinary camera footage, and vision-language models (VLMs) seem like a natural fit for end-to-end monitoring. But a new preprint on arXiv argues that current VLMs fail at this task, and it offers a diagnosis: the models lack what the authors call physiological grounding and numerical alignment.

The paper, posted as arXiv:2610.19052, suggests that VLMs do not reliably link visual cues—such as subtle skin color changes—to the underlying physiological processes they represent. At the same time, the models struggle to map continuous numerical values, like beats per minute, onto their language-based outputs. These two shortcomings, the authors argue, explain why VLMs underperform on RPM despite their strengths in other vision-language tasks.

The preprint frames the issue as a design problem rather than a simple scaling problem. It calls for rethinking how VLMs are trained and evaluated for physiological measurement, with more explicit attention to grounding visual features in biological reality and aligning outputs with precise numeric targets. The abstract does not include experimental results or proposed architectures, so the claims remain at the conceptual level for now.