LLM-Guided Adaptive Question Selection for Psychiatric Intake
Researchers have formulated psychiatric intake as a question-selection problem and introduced a benchmark based on 655 clinician-authored questions and synthetic patient vignettes covering five behavioral conditions. They compared random questioning, a clinical baseline, and an LLM-guided adaptive policy across 300 interview sessions. The LLM-guided approach showed improved efficiency in gathering clinically relevant information. This work addresses the lack of infrastructure for conversational AI in psychiatric intake, a high-stakes, time-limited process.
Key facts
- Psychiatric intake is formulated as a question-selection problem.
- Benchmark uses 655 clinician-authored intake questions.
- Synthetic patient vignettes cover 5 behavioral conditions.
- Evaluation includes 300 interview sessions across 4 patients.
- Compared random questioning, clinical baseline, and LLM-guided adaptive policy.
- LLM-guided policy showed improved efficiency.
- Task addresses limited infrastructure for conversational AI in psychiatric intake.
- Published on arXiv with ID 2604.22067.
Entities
Institutions
- arXiv