Classification fidelity of large language models on endodontic items from the Turkish dental specialty examination

Large language models (LLMs) are increasingly used in dental education, but research has focused mainly on answer accuracy. Whether LLMs can identify the cognitive demand of examination items remains unclear. This study evaluated LLM classification fidelity for endodontic multiple-choice questions (MCQs) across three educational taxonomies and compared four publicly accessible models. Two hundred endodontic MCQs from the Turkish Dentistry Specialty Examination (DUS, 2012–2025) were classified by ChatGPT, Gemini, Kimi, and DeepSeek according to Revised Bloom’s Taxonomy, the SOLO Taxonomy, and Miller’s Pyramid. A three-layer expert consensus procedure comprising calibration training, procedural audit, and structured joint deliberation established the expert consensus reference standard. Agreement was assessed using weighted Cohen’s kappa (κw) with bootstrap 95% confidence intervals; classification performance was evaluated using accuracy and macro-averaged F1. Inter-model differences in item-level accuracy were tested using Cochran’s Q with Kendall’s W, followed by Holm–Bonferroni-corrected pairwise McNemar tests. For Revised Bloom’s Taxonomy, Gemini (κw=0.676) and DeepSeek (κw=0.667) achieved substantial agreement, while Kimi (κw=0.428) and ChatGPT (κw=0.365) performed lower. For SOLO, Gemini achieved substantial agreement (κw=0.622), while the remaining models showed moderate agreement. For Miller’s Pyramid, agreement did not exceed fair, with ChatGPT highest (κw=0.372). Inter-model differences were significant for all three taxonomies (all p<0.001). ChatGPT ranked lowest for Revised Bloom and SOLO but highest for Miller. Revised Bloom levels of expert consensus reference standard increased significantly across examination years (Spearman ρ=0.301, p<0.001). LLMs showed meaningful classification fidelity for Revised Bloom and SOLO taxonomies, particularly Gemini and DeepSeek, but performance for Miller’s Pyramid remained weak. Bloom, SOLO, and Miller should not be treated as interchangeable frameworks, and LLM selection should be tailored to the taxonomy and assessment purpose. Answer accuracy and taxonomy classification fidelity represent distinct capabilities that warrant independent evaluation.