Tag
This paper proposes a System 1 decision model for medical multiple-choice questions that trains only small projection heads and a temperature on top of frozen language models, achieving up to 46.0% accuracy on MedMCQA as a cheap, fast baseline.