STELLAR
Steering-Vector Enhanced LLM Agents for Realistic Digital Twins in Mental Health
Joao Sedoc · New York University · United States
What they're working onSTELLAR (Steering-Vector Enhanced LLM Agents for Realistic Digital Twins in Mental Health) builds clinically faithful simulated patients for training diagnosticians and stress-testing mental-health AI. Evaluating these systems on real patients is often unsafe or impractical, and prompting alone yields simulated patients that lack clinical fidelity. Our foundational technology uses conceptors — mathematical profiles learned from real clinical interview data — to steer a language model's internal activations so it expresses symptoms of anxiety and depression realistically, adjustably, and repeatably, without retraining. On this base, we are developing digital-twin patient agents that reproduce how symptoms present in conversation, and expressive text-to-speech that gives spoken responses appropriate emotional tone. Validation combines psychometric benchmarks with clinical review.
AI approach
Large language models · LLM agents · steering vectors · digital twins
Mental health focus
Mental health research / simulation
Links