StressAgent: perception-discrimination integrated dual-agent framework for stress source classification

Abstract

Accurate classification of users’ stress sources poses a core challenge in the era of digital mental health services. Traditional large language models are prone to misjudgment when coping with users’ diverse emotional expressions, and their decision-making processes lack interpretability. This paper proposes StressAgent, a collaborative dual-agent framework. Through modular task division, it effectively addresses the above limitations and improves the overall performance of models in stress source classification tasks. StressAgent divides the stress source classification task into a pipeline consisting of two dedicated agents: an emotional signal extraction agent and a retrieval-augmented generation agent. We construct a six-category benchmark dataset for stress source identification, and conduct comprehensive ablation and comparative experiments based on Qwen-Flash and DeepSeek-V3 models. Experimental results verify that the retrieval-augmented agent serves as the core module for boosting classification performance. Among all schemes, the LDA topic vector retrieval achieves the best effect, raising the F1-score of Qwen-Flash from 47.84% to 57.80%. The findings demonstrate that embedding retrieval-augmented modules can remarkably enhance model performance in structured tasks such as stress level classification, providing an efficient optimization paradigm for subsequent relevant research.