TLDR: MoPHES is a novel framework that integrates mental state evaluation, conversational support, and professional treatment recommendations into a mobile application. It utilizes two fine-tuned MiniCPM4-0.5B LLMs deployed directly on mobile devices: one for assessing anxiety and depression severity, and another for empathetic multi-turn conversations. This on-device approach enhances user convenience and privacy. Experiments show MoPHES outperforms larger general-purpose LLMs and matches commercial models in mental health tasks, demonstrating the effectiveness of small, specialized LLMs for personalized psychological support.
The global landscape of mental health is facing a significant crisis, with nearly a billion people worldwide grappling with conditions like anxiety and depression. Traditional therapy methods struggle to meet this immense demand, often hindered by high costs, a shortage of professionals, and the persistent stigma surrounding mental illness. While artificial intelligence, particularly large language models (LLMs), offers a promising avenue for support, general-purpose LLMs often fall short in specialized mental health contexts due to a lack of specific training.
Addressing these challenges, a new framework called MoPHES (Mobile Psychological Health Evaluation and Support) has been introduced. This innovative system aims to provide comprehensive mental health care directly on mobile devices, integrating mental state evaluation, conversational support, and professional treatment recommendations. The core of MoPHES lies in its intelligent agent, which leverages two specially fine-tuned MiniCPM4-0.5B LLMs, each with a distinct role.
One of these specialized LLMs is trained on mental health conditions datasets to accurately assess a user’s mental state, predicting the severity of anxiety and depression. The other LLM is fine-tuned on multi-turn dialogues, enabling it to engage users in empathetic and supportive conversations. By combining these capabilities, MoPHES can offer more personalized support and relevant treatment recommendations, mimicking the nuanced approach of a human psychologist.
A crucial aspect of MoPHES is its on-device deployment. Both fine-tuned models run directly on mobile devices, enhancing user convenience and, more importantly, safeguarding user privacy by keeping sensitive data local. This is achieved using the llama.cpp framework, which quantifies model parameters to reduce file size and optimizes inference speed through techniques like dynamic batch size adjustment and intelligent context window management. For instance, testing on a Xiaomi 13 Ultra showed the dialogue model achieving an average inference speed of 17.3 tokens per second, with mental state assessments taking about 4.2 seconds.
The development of MoPHES involved meticulous data preparation. Researchers collected single-turn counseling Q&As from public sources like PsyQA and EmoLLM. To create the necessary training data, they used an AI model, GPT-4o-mini, to label users’ counseling questions with mental condition severities (anxiety and depression, each with four levels). The same AI model was also used to transform single-turn Q&As into multi-turn dialogues, ensuring a rich dataset for conversational training. These datasets were then used for supervised fine-tuning of the MiniCPM4-0.5B models.
Experiments demonstrated MoPHES’s remarkable performance. In detecting depression and anxiety severity, the fine-tuned MoPHES model significantly outperformed several larger general-purpose LLMs, including DeepSeek-R1-7B and Qwen2.5-7B, and achieved performance comparable to or even surpassing commercial models like Gemini-2.0-flash in anxiety prediction. In multi-turn dialogue generation, MoPHES also showed superior performance compared to GPT-4.1 and other domain-specific models, highlighting that a small, on-device model can excel in mental counseling when properly fine-tuned.
Further exploration revealed that incorporating a user’s mental state information into the dialogue generation process led to notably better performance on intrinsic metrics (like BLEU and ROUGE scores), indicating more relevant and context-aware responses. This capability allows the agent to adapt its support based on a continuous understanding of the user’s evolving mental health.
Also Read:
- Guiding AI: How Steering Vectors Enhance Mental Health Assessment in Language Models
- VERA-MH: A New Framework for Evaluating AI Chatbot Safety in Mental Health
MoPHES represents a significant step forward in mobile mental health support, offering an intelligent agent that combines predictive and conversational capabilities in a privacy-preserving, on-device format. While current iterations use two separate LLMs, future work aims to integrate these functions into a single model, further optimizing resource consumption. The research also plans to incorporate reinforcement learning techniques to align the models with user preferences and ethical guidelines, enhancing conversational naturalness and safety. You can read the full research paper here.
