Where Do Apparent LLM Clinical Triage Failures Arise? Localizing the Multiple-Choice Format Effect
The study investigated why LLMs perform worse in clinical triage when forced to choose from multiple-choice options compared to free-text responses. The core finding is that this penalty isn't due to poor clinical reasoning but an output format effect. Medical content is preserved, but the model shifts its focus at the final decision point from medical features to scaffold features, leading to miscalibration in choosing between adjacent acuity levels.