PReMind-Bench for Personalized Vision-Language Safety
November 17, 2026
PReMind-Bench evaluates how vision-language assistants combine visual hazard recognition with personal context like medical conditions or allergies. The benchmark measures memory access, grounding, and the ability to abstain from inappropriate advice.
HOW THIS AFFECTS YOU
●
builderYou can test if your personal assistants are actually safe when they encounter physical hazards in a user's environment.
●
researcherThis allows you to study the interdependency between visual grounding and personal memory access.