Benchmarking Large Language Models on the Taiwan Neurology Board Examinations (2018–2024): A Comparative Evaluation of GPT-4o, GPT-o1, DeepSeek-V3, and DeepSeek-R1

Abstract

Background and Purpose: Neurology requires integration of clinical reasoning, imaging interpretation, and current knowledge, making it an ideal field for evaluating large language models (LLMs). Methods: Using 1715 questions from the Taiwan Neurology Board Examination (2018–2024), we assessed four LLMs—GPT-4o, GPT-o1, DeepSeek-V3, and DeepSeek-R1—across four formats: single-choice, multiple-choice, true–false, and image-based items. Results: GPT-o1 achieved the highest overall accuracy (83.86%) and demonstrated strong performance on cognitively demanding tasks (82.50% on true–false; 77.26% on image-based). DeepSeek-V3 scored lowest (65.62%) and showed the greatest variability. Statistical analyses confirmed significant inter-model differences (p < 0.01). Accuracy declined across all models in 2024, coinciding with shifts in question design. DeepSeek-R1 was further penalized by alignment-based refusals, resulting in up to 3.81% score loss. Conclusions: These results position the Taiwan Neurology Board Exam as a rigorous benchmark for LLM evaluation and underscore GPT-o1’s potential utility in neurology education and decision support.

Ähnliche Arbeiten