It's just a multiple choice test with question prompts. This is the exact sort of thing an LLM should be very good at. This isn't chat gpt trying to do the job of an actual doctor, it would be quite abysmal at that. And even this multiple choice test had to be stacked in favor of chat gpt.

Because GPT models cannot interpret images, questions including imaging analysis, such as those related to ultrasound, electrocardiography, x-ray, magnetic resonance, computed tomography, and positron emission tomography/computed tomography imaging, were excluded.

Don't get me wrong though, I think there's some interesting ways AI can provide some useful assistive tools in medicine, especially tasks involving integrating large amounts of data. I think the authors use some misleading language though, saying things like AI "are performing at the standard we require from physicians," which would only be true if the job of a physician was filling out multiple choice tests.

I wonder why nobody seems capable of making a LLM that knows how to do research and cite real sources.

What would be much more useful is to provide a model with actual patient files and see what kills more people, doctors or models.

All these always do the same thing.

Researchers reduced [the task] to producing a plausible corpus of text, and then published the not-so-shocking results that the thing that is good at generating plausible text did a good job generating plausible text.

From the OP , buried deep in the methodology :

Because GPT models cannot interpret images, questions including imaging analysis, such as those related to ultrasound, electrocardiography, x-ray, magnetic resonance, computed tomography, and positron emission tomography/computed tomography imaging, were excluded.

Yet here's their conclusion :

The advancement from GPT-3.5 to GPT-4 marks a critical milestone in which LLMs achieved physician-level performance. These findings underscore the potential maturity of LLM technology, urging the medical community to explore its widespread applications.

It's literally always the same. They reduce a task such that chatgpt can do it then report that it can do to in the headline, with the caveats buried way later in the text.

Neat but I don't think LLMs are the way to go for these sort of things

This research has been done a lot of a times but I don't see the point of it. Exams are something I would expect LLMs, especially the higher end ones, to do well because of their nature. But it says next to nothing about how reliable the LLM as an actual doctor.

The 17th percentile in peds is not surprising. The model mixing it's training data with adults would absolutely kill someone.