CARE-Bench is a source-grounded benchmark evaluating sequential patient-facing triage as a four-label per-turn current-action task. It contains 500 cases and 1,059 evaluated patient-disclosure prefixes reconstructed from medical dialogue, consultation, and follow-up-question sources. Eleven models were evaluated on 269 held-out rounds under unprompted and minimally prompted open-ended protocols, using a fixed GPT-5.5 mapper to code responses. Unprompted macro-F1 ranged from 31.2 to 50.4; prompted macro-F1 ranged from 46.9 to 63.4. Prompted models often recommended care before needed clarification; when the correct action was to ask for more information, only 33.5% of prompted outputs preserved that step.
CARE-Bench evaluates how well LLMs perform patient-facing triage by classifying responses into four action labels. Testing 11 models on 269 rounds, unprompted macro-F1 scores were low (31.2–50.4). Prompting improved scores (46.9–63.4) but models still frequently skipped necessary clarification steps, recommending care prematurely. Only 33.5% of prompted outputs correctly asked for more information when that was the appropriate action, indicating that triage errors persist beyond simple prompting fixes.
The benchmark uses a fixed GPT-5.5 mapper to code open-ended model responses into a four-label action space, enabling standardized evaluation. The persistence of threshold errors after prompting suggests that current LLMs lack the nuanced reasoning required for safe triage, particularly in recognizing when to gather more information before recommending care.
As patient-facing medical LLMs and agents become more common for symptom triage, the findings highlight a critical safety gap: models may prematurely escalate or advise without sufficient patient history, posing risks in real-world deployment. This underscores the need for rigorous, domain-specific evaluation frameworks before clinical integration.
The benchmark provides a standardized method for healthcare AI developers to assess and improve triage safety, potentially reducing liability and building trust for clinical adoption. It also offers a competitive differentiator for companies aiming to deploy patient-facing LLMs responsibly.
Future work may focus on developing models with better clarification-seeking behaviors and integrating safety guardrails. Observing whether subsequent model releases improve on CARE-Bench metrics, especially in the 'ask for more information' category, will be a key signal. Regulatory bodies may also reference such benchmarks when setting standards for AI in healthcare.