Evaluating LLM Sensitivity and Response Calibration to Longitudinal Youth Wellbeing Contexts

Youth mental wellbeing concerns often emerge through changes in behavior, relationships, and life circumstances over time. As conversational AI systems increasingly retain information across interactions, safe responses may depend not only on what a young person says at the moment, but also on how models connect relevant prior information to the current interaction. We introduce an expert-informed evaluation framework for examining whether large language models (LLMs) use prior youth wellbeing history to interpret and respond to the same current message. Each paired case holds the current message constant while varying the prior history between lower- and higher-concern conditions. The framework spans four youth mental wellbeing categories and three developmental age bands. Preliminary probing across five frontier models reveals uneven linking across different types of prior information: models may recognize prior history but fail to connect broader changes in relationships, family circumstances, or social behavior to current wellbeing needs. Models may also overweigh ambiguous current language and escalate responses despite limited evidence of acute risk. These findings motivate larger-scale evaluation and technical validation of both contextual linking and response calibration.