Tag
GrandGuard introduces a comprehensive taxonomy, benchmark, and safeguards for elderly-specific risks in LLM chatbot interactions, finding that leading LLMs mishandle over 50% of such risks and proposing two safeguards achieving up to 96.2% detection accuracy.
VERA-MH is a clinically-validated evaluation framework to assess the safety of chatbots in mental health support, focusing on suicidal ideation risks. It uses role-played conversations and an LLM-as-a-Judge with a clinical rubric to evaluate responses.