Technology Category
Machine Learning: Large Multimodal Models (LMMs)Humans and AI: Emotional IntelligenceNatural Language Processing: Conversational AI/Dialog Systems
An Offline Mobile Conversational Agent for Mental Health Support: Learning from Emotional Dialogues and Psychological Texts with Student-Centered Evaluation
🤖 AI Summary
To address pervasive issues of network dependency, privacy risks, and limited accessibility in digital mental health platforms, this paper introduces EmoSApp—the first fully offline smartphone dialogue application designed for emotional support. Built upon the LLaMA-3.2-1B-Instruct foundation model, EmoSApp employs domain-specific fine-tuning using a custom psychological Q&A dataset and achieves efficient on-device deployment via quantization and optimization with TorchTune and ExecuTorch. Its novelty lies in the synergistic integration of psychological text modeling and empathetic dialogue training, coupled with a student-centered multi-turn evaluation framework that enhances response empathy and conversational coherence. Experimental results demonstrate state-of-the-art performance across nine low-resource commonsense reasoning benchmarks; qualitative evaluation further confirms its practical utility in recommendation relevance, empathetic responsiveness, and interaction continuity.
Technology Category
Application Category
📝 Abstract
Mental health plays a crucial role in the overall well-being of an individual. In recent years, digital platforms have been increasingly used to expand mental health and emotional support. However, there are persistent challenges related to limited user accessibility, internet connectivity, and data privacy, which highlight the need for an offline, smartphone-based solution. To address these challenges, we propose EmoSApp (Emotional Support App): an entirely offline, smartphone-based conversational app designed for mental health and emotional support. The system leverages Large Language Models (LLMs), specifically fine-tuned, quantized and deployed using Torchtune and Executorch for resource-constrained devices, allowing all inferences to occur on the smartphone. To equip EmoSApp with robust domain expertise, we fine-tuned the LLaMA-3.2-1B-Instruct model on our custom curated ``Knowledge dataset'' of 14,582 mental-health QA pairs, along with the multi-turn conversational data.
Through qualitative human evaluation with the student population, we demonstrate that EmoSApp has the ability to respond coherently, empathetically, maintain interactive dialogue, and provide relevant suggestions to user's mental health problems. Additionally, quantitative evaluations on nine standard commonsense and reasoning benchmarks demonstrate the efficacy of our fine-tuned, quantized model in low-resource settings. By prioritizing on-device deployment and specialized domain adaptation, EmoSApp serves as a blueprint for future innovations in portable, secure, and highly tailored AI-driven mental health solutions.
Problem
Research questions and friction points this paper is trying to address.
Develops an offline smartphone app for mental health support
Addresses limited accessibility and data privacy concerns
Fine-tunes LLMs for empathetic, resource-efficient dialogues
Innovation
Methods, ideas, or system contributions that make the work stand out.
Offline smartphone-based conversational mental health app
Fine-tuned quantized LLM for resource-constrained devices
Domain adaptation using curated mental health datasets
🔎 Similar Papers
No similar papers found.
