I work at a telemedicine company. We’ve benchmarked a few frontier LLMs on public medical imaging datasets. One test included high-quality and high-consensus otoscopic images. We didn’t anticipate the models to do well on something so niche, but what concerned us was how poorly calibrated the models were.

I know you can’t trust an LLM’s self-assessed “confidence” of a prediction, but I’ve found that confidence can at least be directionally correct for some tasks. For our benchmarks, however, confidence was poorly correlated. What’s worse is that binary classification models (“Do you see $diagnosis in this photo?”) highly influenced the LLM to confidently predict $diagnosis.

I’m concerned for those using LLMs for diagnostics, and getting confidently led to the wrong conclusion.

It makes a lot of sense if you understand how these models work but this was a cool read anyways and studies like this are impotent for curbing the unfortunate fever dream some folks seem to be collectively having about LLM omnipotence

I don’t understand how this is a different result than giving any LLM a task that is not completely grounded? I’ve observed this in coding tasks, if I forget to include a file referred to in the spec, the LLM will just hallucinate a version of it and my results suck. If I give it the file (and really, all the information I claimed it had access to), the task works fine. I fixed this in my pipeline with a prompt that does an extensive grounding analysis to determine if the assets I’m giving it are complete with respect to the spec (and that the spec is grounded as well, ie it doesn’t refer to something that is undefined).

I wonder if the above problem can be fixed similarly? Just ask the LLM to do a conservative grounding analysis before jumping to the main task?

The absolute only thing that matters is if they are provided an image what's the success rate.

But why should I care? If you demonstrated that a model can perform more accurate diagnoses than a doctor, but also it had this strange behavior when no image was presented, why should that deter me from using the model?