From Simulation to Diagnosis: Simulation-Based Reinforcement Learning for Empathetic Depression Diagnosis

ACL ARR 2026 January Submission5029 Authors

Keywords: Depression Detection, NLP for Social Good, Mental Health

Abstract: Given the rising prevalence of depression alongside persistently low diagnosis rates, Large Language Models (LLMs) offer an opportunity for accessible mental health assessment. However, clinical depression diagnosis is inherently a goal-oriented, sequential, and interactive process. Current approaches primarily rely on static supervised finetuning (SFT) using fixed depression-diagnosis dialogue datasets, which are ill-suited to the dynamic nature of real-world diagnosis interactions. Models trained on such static data often struggle to navigate the variability of patient behavior, failing to provide diagnostic accuracy. To overcome these limitations, we introduce SimRED (Simulation-based Reinforcement Learning for Empathetic Depression Diagnosis), a framework that trains depression diagnostic agents via reinforcement learning through extensive interactions with patient simulator. SimRED constructs a patient simulation environment derived from real-world dialogues. By interacting with these simulated patients, the diagnostic agent employs reinforcement learning to learn both diagnostic accuracy and empathetic expression. Experimental results demonstrate that SimRED significantly outperforms existing strong baselines in both diagnostic accuracy and the quality of empathetic expression.

Paper Type: Long

Research Area: Computational Social Science, Cultural Analytics, and NLP for Social Good

Research Area Keywords: AI for Social Good

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: Chinese

Submission Number: 5029