MedRealMM benchmark reveals AI still triggers more safety errors than doctors
5,620 real Chinese medical consultations show frontier AI models fail safety-critical tests
MedRealMM is a new benchmark for evaluating multimodal large language models in real-world online medical consultation. Unlike previous benchmarks that relied on synthetic conversations or multiple-choice questions, MedRealMM uses authentic, de-identified patient-doctor interactions from a Chinese internet hospital spanning 64 clinical departments. It employs a Multimodal Clinical Challenge Point (MCCP) extraction framework to identify clinically demanding moments, converting them into next-response generation tasks with full text-image context. Each instance is paired with a case-specific rubric refined by physicians to reward clinically desirable behaviors and penalize unsafe or contradictory responses. The current release contains 5,620 real-world multimodal cases.
When tested across 19 general-purpose and medical-specialized LLMs (including text-only and multimodal systems), the results were sobering. Image information proved critical for reliable clinical performance, yet even the best frontier models remain below the quality of online physician responses. Frontier models could satisfy as many or more positive clinical criteria as doctors, but they also triggered more negative criteria—meaning unsafe, unsupported, or contradictory outputs. This indicates that safety-sensitive error avoidance remains a central bottleneck for deploying AI in real patient care. MedRealMM offers a realistic and reproducible benchmark to drive progress in multimodal medical reasoning, with the dataset publicly available on Hugging Face.
- 5,620 real-world multimodal cases from 64 clinical departments, using authentic patient-doctor interactions rather than synthetic data
- Evaluates 19 LLMs; image information is critical—multimodal models outperform text-only ones in clinical accuracy
- Frontier models match or exceed physicians on positive criteria but trigger more negative (unsafe) criteria, highlighting safety as key bottleneck
Why It Matters
Real-world medical AI must prioritize safety; MedRealMM sets a rigorous standard for evaluating multimodal clinical reasoning.