CARE-Bench: Conversational models misdirect patients to the right level of care

Out of 500 patient triage cases, the macro-F1 score of the unguided models peaks at 31-50%, and threshold errors persist despite instructions.

Medical models and agents intended for the general public are increasingly answering questions about symptoms before any contact with a caregiver. The central safety question is not the diagnosis but the action to be recommended: consult urgently, make an appointment, manage yourself, or first ask a follow-up question. CARE-Bench formalizes this sequential triage as a four-label task evaluated at each round of dialogue.

The evaluation set includes 500 cases and 1,059 patient disclosure prefixes reconstructed from medical dialogues, consultations, and sources of follow-up questions. Eleven models were evaluated on 269 selected rounds, according to open protocols with and without minimum instruction.

Without instruction, the macro-F1 remains low, from 31.2 to 50.4. The addition of a setpoint improves ten out of eleven models, bringing the macro-F1 to between 46.9 and 63.4, but substantial threshold errors remain. The models thus guided frequently recommend a course of action to be taken before obtaining the necessary clarifications: when the right action was to ask for more information, only 33.5% of the responses preserved this step. For the authors, the persistence of these errors shows that consumer triage is not a simple prompt engineering problem and justifies an explicit evaluation of the moment of action before any deployment.