PUPPET: Neural-Symbolic Standardized Patients for Mental Health

ACL ARR 2026 January Submission10657 Authors

Keywords: Large Language Models, Mental Health Training, Patient Simulation, Neuro-Symbolic AI, Virtual Standardized Patients

Abstract: The critical therapist shortage demands scalable training solutions. While Standardized Patients are the gold standard for such training, they remain scarce and costly. Current LLM-based approaches focus primarily on patient simulation to achieve conversational realism but fail to function as Virtual Standardized Patients regarding pedagogical rigor. Consequently, they lack faithful reactions to clinical errors and cannot provide explainable feedback. To bridge this gap, we propose PUPPET, the first neuro-symbolic Virtual Standardized Patient governed by an Observe-Think-Behave architecture. This approach enforces trainer-defined clinical logic to ensure deterministic state transitions while maintaining natural variability. Additionally, our PUPPET-TRAINER reveals the complete causal chain from intervention to response to facilitate transparent learning. Experiments across three distinct clinical scenarios confirm that our method consistently outperforms other baselines in both clinical faithfulness and pedagogical value.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: educational applications, mental health, healthcare applications, clinical NLP

Contribution Types: NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English, Chinese

Submission Number: 10657