STELLAR – Steering-Vector Enhanced LLM Agents for Realistic Digital Twins in Mental Health

The scarcity of mental health training resources worldwide creates critical shortages in qualified health workers. Complex psychiatric presentations make it difficult for trainees to gain sufficient clinical exposure. To address these gaps, we propose STELLAR (Steering-Vector Enhanced LLM Agents for Realistic Digital Twins in Mental Health), AI-powered patient simulations using Large Language Models enhanced with Conceptors – advanced steering vectors enabling precise control over model outputs – to create realistic digital patient twins with controllable psychiatric symptoms for scalable clinician training. We leverage Conceptors to encode specific symptom constructs (Anxious/misery, Psychosis, Behavioral/externalizing, Fear) with scalable intensity and combinability, surpassing traditional prompting. We will develop Conceptors using the Philadelphia Neurodevelopmental Cohort, a community sample of 9,500+ individuals with comprehensive psychiatric assessments and diagnostic criteria. A key innovation is Conceptor shareability. Once validated, symptom-specific Conceptors can be distributed to other institutions, accelerating training and collaborative research. Primary applications include: (1) interdisciplinary clinical training through safe interactions with diverse patient presentations; (2) validation of language-based diagnostic tools against reproducible digital twin presentations before clinical deployment. This collaboration unites clinical expertise for symptom annotation with AI capabilities for Conceptor development. STELLAR will be tested by an experienced clinical team with global reach and diverse stakeholder input.