Performance of GPT-5 and Gemini 2.5 Pro on the Orthopaedic In-Training Examination

Abstract

Background: Previous studies evaluating large language models (LLMs) on the Orthopaedic In-Training Examination (OITE) have primarily focused on earlier-generation models and single-pass accuracy. These investigations did not assess newer multimodal systems such as GPT-5 and Gemini 2.5 Pro, nor did they examine the reasoning quality underlying model responses or the consistency of outputs across repeated trials. As LLMs are increasingly used as educational tools, a more comprehensive evaluation framework is needed to assess not only correctness but also reliability and explanatory validity on specialty-specific, image-rich examinations. Methods: We conducted a controlled, parallel evaluation of GPT-5 and Gemini 2.5 Pro using 412 OITE-style questions from the 2023-2024 examination cycle obtained via an institutional AAOS ResStudy subscription. Primary outcomes included overall and subspecialty-specific accuracy. Secondary analyses evaluated explanatory quality, error-pattern classification, response consistency across repeated trials, and performance stratified by imaging burden. Paired accuracy was compared using McNemar's exact test. Results: Gemini 2.5 Pro demonstrated higher overall accuracy than GPT-5 on the 2023-2024 OITE question set (81.1% vs 76.0), with both models exceeding published PGY-5 resident benchmarks. Accuracy declined significantly with questions containing images (74.2% vs 71.6%). Subspecialty performance varied widely, with accuracy ranging from 42.9% to 94.1% for GPT-5 and from 57.1% to 95.8% for Gemini, and both models performing poorest in Hand and Wrist questions. Among incorrect responses, faulty reasoning accounted for 52.5% of GPT-5 errors, whereas stem misinterpretation was the predominant error for Gemini (43.6%). Incorrect or partially correct explanations accompanied 45.4% of GPT-5 and 41.7% of Gemini responses. Consistency testing showed high reproducibility (fully consistent responses: 88% for GPT-5 and 84% for Gemini), with all inconsistent outputs occurring in image-containing questions. Conclusions: GPT-5 and Gemini 2.5 Pro demonstrate strong performance on recent OITE content, exceeding prior LLM benchmarks; however, persistent limitations in multimodal reasoning, explanatory reliability, and response consistency indicate that high accuracy alone does not ensure dependable clinical reasoning, underscoring the need for cautious educational use.

Ähnliche Arbeiten