Researchers tested 21 frontier large language models on 29 stepwise MSD Manual clinical vignettes and found that, although many models performed well on final diagnosis, they remained much weaker at ...