Human guided empathetic AI agent for mental health support leveraging reinforcement learning-enhanced retrieval-augmented generation

Abstract

Global mental health issues is increasing due to problems such as the social stigma around treatment, a long-neglected burdens of insufficient resources, and the rising tide of mental issues. Large language models (LLMs) can accelerate the development of comprehensive, extensive solutions that support mental health. However, the LLMs’ capability to generate and comprehend human-like conversations is one of the main challenges faced by psychiatric counselling. This work proposes a mental health counselling LLM-based conversational agent that relies on the integration of Retrieval Augmented Generation (RAG) and Reinforcement learning. RAG provides the proposed LLM-based conversational agent with contextually relevant and accurate responses through useful information extracted from a curated dataset of psychological questions and answers pooled from mental health forums. Reinforcement Learning Integrated reward Model trained with Human feedback has also been used in the proposed framework to ensure contractually of the responses generated with moral and human values. By setting up a reward mechanism that considers variables like user feedback and empathetic scores of responses, the proposed Conversational Agent learns to prioritize empathetic answers and the ones that are user preferable. With the utilization of reward-based training, the agent was able to show substantial improvements in response quality. Improved emotional alignment, steady training dynamics, decreased hallucination rates with responses having less distress and increased empathy values were the significant outcomes. The proposed methodology ensures that the conversational agent remains attentive to the emotional requirements of people seeking for mental health care and provide improved relevance and accuracy in its responses.