MTDiag: A Multi-Turn Diagnostic Dataset Towards Clinically Meaningful LLM Evaluation

Researchers released MTDiag, a multi-turn diagnostic dialogue dataset built from DDXPlus, MIMIC-IV, and published case reports, to evaluate large language models (LLMs) as diagnostic agents in dynamic clinical encounters. The dataset, normalized to UMLS concept identifiers and ICD-10 codes, includes a UserLM-8B-based utterance-generation pipeline and physician-validated cases, alongside new clinical knowledge-grounded metrics beyond diagnostic accuracy.

arXiv:2608.25085v1 Announce Type: new Abstract: Clinical diagnosis is fundamentally interactive and incremental, yet the dominant paradigm for evaluating Large Language Models LLMs in medicine remains static QA benchmarks or template-based dialogues. These benchmarks say little about whether a model can serve as a diagnostic agent in a dynamic clinical encounter, with LLMs showing significant accuracy and reliability degradation in multi-turn settings. To address this issue, we present MTDiag, a large multi-turn diagnostic dialogue dataset constructed from three heterogeneous sources: DDXPlus, MIMIC-IV, and published case reports AJCR , covering common ED presentations as well as long-tail rare and atypical conditions. All cases are normalized into a canonical schema anchored in the most comprehensive and widely-adopted medical knowledge bases UMLS concept identifiers, with ICD-10 diagnosis codes . We release the schema, a UserLM-8B-based utterance-generation pipeline, and the physician-validated dataset that converts structured clinical evidence into natural-language utterances. Importantly, we introduce and motivate clinical knowledge-grounded metrics for evaluating LLMs as diagnostic agents, beyond diagnostic accuracy, for the task of multi-turn differential diagnosis.