Use of Large Language Models for Automated Short-Answer Question Grading in Second-Cycle Medical Education: Performance and Consistency Evaluation Study
Related Experiment Video
Updated: Nov 4, 2026
Augmenting Large Language Models via Vector Embeddings to Improve Domain-Specific Responsiveness
Published on: December 6, 2024
Sabrina Sidali1,2, Pierre-Antoine Soret3,4, Mathilde Wagner5,6
1Cordeliers Research Center, Sorbonne University, Inserm, University of Paris, Functional Genomics of Solid Tumors Team, French National Cancer League-Accredited Team, Labex OncoImmunology, F-75006 Paris, Île-de-France, France.
Background:
Short-answer questions (SAQs) effectively assess higher-order reasoning skills in medical education but face challenges in scalability due to time-intensive manual grading and concerns about interrater reliability. Recent advances in large language models (LLMs) may offer a solution through automated assessment.
Objective:
This study evaluated the performance and consistency of LLMs-ChatGPT, Claude Sonnet 4, and Mistral Pro-compared with those of human graders for SAQ grading in hepatology and gastroenterology.
Methods:
We retrospectively analyzed SAQs from hepatology and gastroenterology examinations at Sorbonne University, Paris, France (2023-2025). Three human evaluators (2 hepatogastroenterologists and 1 radiologist) independently graded all responses twice, with the majority consensus score serving as the reference standard. Six LLMs were evaluated: ChatGPT (GPT-4o, GPT-4.5, GPT-4o3, and GPT-5), Claude Sonnet 4, and Mistral Pro. Each LLM graded all responses twice using standardized prompts, including questions, model answers, grading guidelines, and scoring rubrics. Agreement was assessed using quadratic weighted kappa (QWK), Cohen κ, the intraclass correlation coefficient (ICC), and the mean absolute error (MAE) for interhuman, intrahuman, LLM-versus-human, and intramodel comparisons.
Results:
A total of 902 student responses from 12 SAQs were included and analyzed. Interhuman grader agreement was almost perfect (QWK 0.90-0.92; ICC 0.90-0.92; MAE 0.05-0.07). Claude Sonnet 4 and GPT-4o3 demonstrated near-human-level performance when compared with the majority vote of human evaluators (Claude Sonnet 4: QWK 0.88-0.90; ICC 0.88-0.90; MAE 0.06-0.07 and GPT-4o3: QWK 0.82-0.89; ICC 0.82-0.89; MAE 0.07-0.09). GPT-5 showed substantial agreement (QWK 0.77-0.79). Mistral Pro and older ChatGPT versions performed poorly (QWK <0.63). Intramodel consistency was the highest for Claude Sonnet 4 (QWK 0.93), approaching the best human intrarater reliability. Agreement did not correlate with prompt design choices (number of examples or minimum word count).
Conclusions:
Advanced LLMs, particularly Claude Sonnet 4 and GPT-4o3, demonstrated near-human-level grading performance, supporting their potential use as assistive tools in SAQ grading. However, performance varied by model, and agreement alone does not ensure educational validity. Until broader multicenter validation is conducted across diverse domains, stakes, and question formats-along with safeguards for bias, privacy, and model drift-LLMs should augment rather than replace human faculty graders.