Rendimiento de los modelos de lenguaje grandes en pruebas de aptitud cognitiva: una evaluación de múltiples...

Large language models (LLMs) have demonstrated high performance on knowledge-based medical examinations but their capabilities on cognitive aptitude tests emphasizing reasoning and abstraction remain underexplored. The Test for Medical Studies (TMS), a German medical school admission test, provides a standardized framework to examine these capabilities. This study aimed to evaluate the performance and consistency of multiple LLMs on text-based and visual-analytic TMS items.

Eight contemporary LLMs, comprising proprietary and open-source systems, were evaluated using a multi-run design on standardized TMS items spanning text-based and visual-analytic cognitive domains.

Mean accuracy remained substantially below levels typically reported for knowledge-based medical examinations, with marked performance differences between text-based and visual-analytic subtests. Open-source models performed competitively compared with proprietary systems. Inter-run reliability was heterogeneous, indicating notable variability across repeated evaluations.

Current LLMs show limited and domain-dependent performance on cognitive aptitude tasks relevant to medical school admission. High accuracy on knowledge-based examinations does not translate into stable performance on aptitude tests emphasizing fluid intelligence. The observed modality-dependent performance patterns and inter-run variability highlight the importance of differentiated, multi-run evaluation strategies when assessing LLMs for applications in medical education.