Clinical referral text overrides MRI evidence in multimodal large language models: glioma versus brain abscess

A retrospective single-center study included 60 contrast-enhanced T1-weighted brain MRIs (30 gliomas, 30 brain abscesses), with one axial slice taken per case and paired with five types of referral text, repeated 20 times each under three prompt instructions and submitted to GPT-4o, Claude Sonnet 5, Gemini 3.5 Flash and Qwen 3.6, yielding a total of 72,000 responses. The authors Beste, Bauer, Lopez and colleagues report that the text-driven span ranged from 0.79 to 1.00 (p<0.001 for all models), whereas the difference between confirmed glioma and abscess was only 0.01 to 0.10, and 0.25 for Gemini 3.5 Flash; under neutral text, the proportion of cases judged as tumor by the models ranged from 45% to 67%. The authors report that when instructed to ignore the text, the span for three models fell to 0.48 to 0.70, and for Gemini 3.5 Flash fell to 0.04, partly due to regression to a default tumor judgment. Model accuracy ranged from 0.52 to 0.73, while two blinded radiologist readers achieved 0.78 and 0.73.