PBPau Benito

Pau Benito

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Pau Benito has a total of 14 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Construct Validity of Frontier Large Language Models on 4 Spanish Health Professions Specialty-Access Examinations (MIR, EIR, PIR, and FIR): Cross-Sectional Benchmarking and Psychometric Study

    by Clemente García-Hidalgo, Manuel Carpio-Salmerón, Pau Benito, Carlos Carazo-Casas, Beatriz Carratalá, Jaime Campos Pavón

    Background: Large language models (LLMs) achieve high scores on health-professions examinations. Whether these scores support the same interpretations as human scores remains uncertain. Spain uses parallel multiple-choice examinations for entry to specialty training in medicine (Médico Interno Residente [MIR]), nursing (Enfermero Interno Residente [EIR]), psychology (Psicólogo Interno Residente [PIR]), and pharmacy (Farmacéutico Interno Residente [FIR]). Objective: Using the January 2026 MIR, EIR, PIR, and FIR examinations, this study aimed to quantify the correspondence between human and LLM item facility and to describe model accuracy, distractor selection, and performance associated with measured surface features. Methods: We evaluated 839 valid items using 12 configurations: GPT-5.2 at 3 reasoning efforts and Claude Opus, Sonnet, and Haiku 4.5 (Anthropic), each under neutral and expert prompts, with one run per configuration. The primary measure was the Spearman correlation (ρ) between human and machine facility on 828 complete paired items. We examined sensitivity to panel composition, scoring, pooling, and assumed reliability; compared distractor choices with a post hoc plausibility-weighted baseline; and evaluated a surface-feature classifier. Image analyses distinguished supplied from withheld images. An exploratory LLM-judge audit assessed generated explanations. Bootstrap CIs resample items or, for distractor concordance, error events; they exclude run-to-run uncertainty. Results: The highest pooled accuracy was 94.8% (795/839); all 12 configurations answered 668 of 828 paired items correctly (80.7%). Scores clustered near the ceiling among the strongest configurations. Human and machine facility correlated at 0.40 (95% CI 0.34-0.45), with estimates of 0.31 to 0.40 across panel definitions and 0.39 (95% CI 0.33-0.45) after exam-stratified pooling. The observed correlation was lower than predicted by a specified one-axis simulation with a uniform model advantage (0.63; predictive interval 0.59 to 0.66; P <.001). Among 672 model errors, dominant human-distractor selection differed from the plausibility-weighted baseline by –1.25 percentage points (95% CI –4.53 to 2.19), compatible with no difference. Surface-feature accuracy was 34.4% (289/839), similar to selecting the longest option (34.6%, 290/839). The adjusted association with model correctness was uncertain (odds ratio [OR] 0.87, 95% CI 0.72-1.04). Image-associated accuracy was lower both when images were supplied (OR 0.33, 95% CI 0.21-0.53) and when withheld (OR 0.11, 95% CI 0.08-0.16). Conclusions: In this single 2026 cohort, high LLM scores coexisted with limited separation among the strongest configurations and moderate correspondence with human item difficulty. These findings caution against interpreting model scores as equivalent evidence of human competence. They do not establish clinical performance or invalidate the examinations for human candidates.

  • Open Access

    Frontier Large Language Models on the 2026 Spanish MIR Examination: A Multimodal Cross-Sectional Evaluation.

    by Manuel Carpio Salmerón, Carlos Carazo-Casas, Pau Benito, Clemente Garcia, Jesús Alonso-Carrillo, Beatriz Carratalá, Georgios Kyriakos, Pablo González-Castro

    Introduction: Large language model-based AI systems are increasingly used in medical education, but their educational value in multilingual, high-stakes examination settings remains insufficiently defined. Objective: To assess the accuracy of latest-generation AI systems on the official 2026 Spanish Médico Interno Residente (MIR) examination and compare their performance with the Healthcademia chatbot, AMIR faculty experts, and average candidates. Materials and Methods: A cross-sectional quantitative evaluation analyzed all 200 valid questions from the official MIR 2026 Version 0 booklet. Performance was assessed using automated pipelines for text-only items and manual multimodal input for the 24 image-associated questions. Two instruction conditions, neutral and expert-role, were tested, with additional effort-level stratification for GPT-5.2. Accuracy was measured against the final official answer key. Human comparison groups included AMIR faculty experts and the average MIR candidate. Results: Several latest-generation AI systems showed very high accuracy on the MIR 2026 examination. The highest-performing configuration, GPT-5.2 high effort without expert-role instruction, achieved 199/200 correct answers (99.5%). Gemini 3 Flash with expert-role instruction achieved 198/200 (99.0%). AMIR faculty experts achieved 194/200 (97.0%), whereas the average candidate achieved 131/200 (65.7%). On image-associated questions, several multimodal configurations achieved 24/24 correct answers when the corresponding image was provided. Conclusions: Under the conditions evaluated in this study, several AI configurations achieved near-perfect accuracy on a complete, high-stakes national medical licensing examination. These findings support licensing-style examinations as benchmarks for educational AI and suggest potential use in supervised feedback and self-assessment. Performance on multiple-choice items should not be interpreted as evidence of autonomous clinical reasoning.


  • Top Cited Papers
    Open Access

    GPT-4o and OpenAI o1 Performance on the 2024 Spanish Competitive Medical Specialty Access Examination: Cross-Sectional Quantitative Evaluation Study

    by Pau Benito, Mikel Isla-Jover, Pablo González-Castro, Pedro José Fernández Esparcia, Manuel Carpio, Iván Blay-Simón, Pablo Gutiérrez-Bedia, Maria J Lapastora, Beatriz Carratalá, Carlos Carazo-Casas

    Background: In recent years, generative artificial intelligence and large language models (LLMs) have rapidly advanced, offering significant potential to transform medical education. Several studies have evaluated the performance of chatbots on multiple-choice medical examinations. Objective: The study aims to assess the performance of two LLMs—GPT-4o and OpenAI o1—on the Médico Interno Residente (MIR) 2024 examination, the Spanish national medical test that determines eligibility for competitive medical specialist training positions. Methods: A total of 176 questions from the MIR 2024 examination were analyzed. Each question was presented individually to the chatbots to ensure independence and prevent memory retention bias. No additional prompts were introduced to minimize potential bias. For each LLM, response consistency under verification prompting was assessed by systematically asking, “Are you sure?” after each response. Accuracy was defined as the percentage of correct responses compared to the official answers provided by the Spanish Ministry of Health. It was assessed for GPT-4o, OpenAI o1, and, as a benchmark, for a consensus of medical specialists and for the average MIR candidate. Subanalyses included performance across different medical subjects, question difficulty (quintiles based on the percentage of examinees correctly answering each question), and question types (clinical cases vs theoretical questions; positive vs negative questions). Results: Overall accuracy was 89.8% (158/176) for GPT-4o and 90% (160/176) after verification prompting, 92.6% (163/176) for OpenAI o1 and 93.2% (164/176) after verification prompting, 94.3% (166/176) for the consensus of medical specialists, and 56.6% (100/176) for the average MIR candidate. Both LLMs and the consensus of medical specialists outperformed the average MIR candidate across all 20 medical subjects analyzed, with ≥80% LLMs’ accuracy in most domains. A performance gradient was observed: LLMs’ accuracy gradually declined as question difficulty increased. Slightly higher accuracy was observed for clinical cases compared to theoretical questions, as well as for positive questions compared to negative ones. Both models demonstrated high response consistency, with near-perfect agreement between initial responses and those after the verification prompting. Conclusions: These findings highlight the excellent performance of GPT-4o and OpenAI o1 on the MIR 2024 examination, demonstrating consistent accuracy across medical subjects and question types. The integration of LLMs into medical education presents promising opportunities and is likely to reshape how students prepare for licensing examinations and change our understanding of medical education. Further research should explore how the wording, language, prompting techniques, and image-based questions can influence LLMs’ accuracy, as well as evaluate the performance of emerging artificial intelligence models in similar assessments.