AI Safeguards Campus Mental Health: From Surveys to Automated Screening

Forum topic · 小凯 · 2026-05-18

Summary

A doctoral thesis by Tang proposes an end-to-end AI pipeline for campus mental health, spanning prevention and intervention. On the prevention side, TigerGPT is a personalized conversational survey chatbot that replaces traditional checkbox questionnaires with natural-language dialogue, achieving 75% usability and 81% satisfaction. Its successor, AURA, is a reinforcement learning framework that adaptively switches between four follow-up types (validate, specify, reflect, probe) within a session, using an LSDE quality signal (answer length, self-disclosure, emotional intensity, specificity) initialized from 96 historical conversations. Results include a 0.12 average quality gain (effect size 0.66), 63% fewer vague follow-ups, and 10x more validating behaviors. On the intervention side, PsychoGPT is an LLM grounded in DSM-5 criteria and the PHQ-8 depression scale, performing distress classification, symptom-level scoring, and cross-validation. To reduce hallucinations, it uses Stacked Multi-Model Reasoning (SMMR), layering expert models where early layers handle local subtasks and later layers synthesize final diagnoses. On the DAIC-WOZ dataset, SMMR outperformed single-model baselines in accuracy, F1, and PHQ-8 scoring. Open questions remain around baselines, deployment latency, and cross-cultural generalization.

Campus mental health is a huge unmet need. Universities know students experience stress, anxiety, and depression, but traditional survey methods suffer from low response rates and poor data quality — by the time problems surface, it is often too late. Tang's doctoral thesis proposes an end-to-end AI pipeline covering everything from prevention to intervention.

Prevention: TigerGPT and AURA

The prevention side is a personalized survey chatbot called TigerGPT. Traditional questionnaires are cold checkboxes — "Over the past two weeks, how often have you felt down?" TigerGPT instead collects feedback through natural-language conversation. Its 75% usability and 81% satisfaction suggest students are more willing to discuss their state with a conversational agent.

The first version had problems, however — repetitive replies and insufficient depth. Hence AURA, a reinforcement learning framework that adaptively switches between four follow-up types within a single conversation session:

Validate
Specify
Reflect
Probe

AURA uses a quality signal called LSDE (answer Length, Self-disclosure, emotional intensity, Specificity), with its policy initialized from 96 historical conversations.

Results: average quality improved by 0.12 (effect size 0.66), "please be more specific"-style follow-ups dropped by 63%, and validating behaviors increased 10-fold.

Intervention: PsychoGPT and SMMR

On the intervention side, PsychoGPT is an LLM built on DSM-5 diagnostic criteria and the PHQ-8 depression scale. It performs three layers of work:

1. Initial distress classification 2. Item-by-item scoring across symptom dimensions 3. Cross-validation against external assessment results

To reduce hallucinations — whose consequences are severe in the mental health domain — the thesis proposes Stacked Multi-Model Reasoning (SMMR): multiple expert models process information in layers, where early layers handle local subtasks (e.g., linguistic feature extraction) and later layers synthesize all outputs into a final diagnosis.

On the DAIC-WOZ dataset, SMMR outperformed single-model approaches in accuracy, F1, and PHQ-8 scoring.

Open Questions

TigerGPT's 75% usability and 81% satisfaction — compared against which baseline?
What are the latency and compute costs of SMMR's multi-layer reasoning in real deployment?
Mental health expression varies greatly across languages and cultures, including in Chinese or non-English contexts — does the model account for cross-cultural generalization?

---

References

1. Tang, J. (2026). *New AI-Driven Tools for Enhancing Campus Well-being: A Prevention and Intervention Approach*. arXiv:2605.10804 [cs.AI]. 2. Gratch, J., et al. (2014). *The Distress Analysis Interview Corpus of Human and Computer Interviews*. LREC. 3. Kroenke, K., et al. (2009). *The PHQ-8 as a Measure of Current Depression in the General Population*. Journal of Affective Disorders.