AI Chatbots and Reinforcement Learning for Campus Mental Health
A recent dissertation presents a comprehensive framework aimed at enhancing campus well-being through both preventive measures and interventions. The authors created TigerGPT, a survey chatbot powered by large language models (LLMs), which recorded a usability score of 75% and a satisfaction rate of 81%. To overcome issues related to TigerGPT’s repetitive nature and shallow responses, they developed AURA, a reinforcement-learning system that modifies follow-up questions—validate, specify, reflect, probe—during interactions, guided by an LSDE quality signal (Length, Self-disclosure, Emotion, Specificity). AURA was trained on 96 previous conversations and demonstrated an average quality improvement of +0.12 (p=0.044, d=0). This research is available on arXiv with the identifier 2605.10804.
Key facts
- Dissertation addresses gaps in campus well-being monitoring and mental health risk detection.
- TigerGPT is a personalized survey chatbot using LLMs, achieving 75% usability and 81% satisfaction.
- AURA is a reinforcement-learning framework that adapts follow-up question types.
- AURA uses an LSDE quality signal (Length, Self-disclosure, Emotion, Specificity).
- AURA initialized from 96 prior conversations.
- AURA achieved +0.12 mean quality gain (p=0.044, d=0).
- Published on arXiv with identifier 2605.10804.
- Framework unifies prevention and intervention approaches.
Entities
Institutions
- arXiv