MedPRESS: A Multi-turn Benchmark for Patient-Pressure-Induced Medical Sycophancy in LLMs

За 30 секунд

Что произошло

Researchers released MedPRESS, a benchmark testing whether medical LLMs maintain safe advice when patients pressure them across five-turn conversations.

Почему это важно

Engineers building or deploying medical chatbots need this to understand how models degrade under realistic patient interaction patterns.

На что обратить внимание

The benchmark tests 20 models but does not reveal which specific models fail worst, limiting direct actionability for practitioners.