Abstract
Although large language models (LLMs) are achieving impressive performance under zero- and few-shot learning configurations, their reasoning capacities are still poorly understood. As a step in this direction, we present several experiments on multiple-choice question answering, a setting that allows us to evaluate the stability of the model under different prompting, the capacity to understand when none of the provided answers is correct, and to reason on specific answering strategies (e.g., recursively eliminate the worst answer). We use the Italian medical specialty tests yearly administered to admit medical doctors to specialties. Results show that a gpt-3.5-turbo model achieves excellent performance in the absolute score (an average of 108 out of 140) while still suffering in certain reasoning capacities, particularly in failing to understand when none of the provided answers is correct.