关键词: 智能体, 大模型, 心理咨询, 大学生

Abstract: Given the increasing complexity of mental health issues among university students and the significant limitations of traditional university counseling models in terms of coverage, response efficiency, and personalized support, this research pioneers an innovative application of artificial intelligence in the field of psychological services. We propose and construct a sophisticated multi-agent collaborative system that deeply integrates vertical domain expertise in psychology with the advanced capabilities of modern large language models. The fundamental objective is to leverage technological empowerment to build an intelligent, sustainable, and dynamically evolvable psychological support ecosystem tailored specifically for the university student population.
The system's architecture is built around a core of three functionally distinct psychological counseling agents—namely the Assessor, the Counselor, and the Supervisor—complemented by simulated student agents used for immersive training. This ensemble forms an organically collaborative “Assessment-Counseling-Supervision” tripartite framework. A key innovation lies in its operational mechanism, characterized by a dual-cycle model of “internal cycle training and external cycle service.” Within the secure, closed internal cycle, these agents engage in continuous high-fidelity interactions, rehearse complex scenarios, and refine intervention strategies, thereby accumulating rich, multi-school therapeutic experience alongside detailed personalized case profiles. This internal practice environment is crucial for honing capabilities. Subsequently, during the open external cycle, the system leverages the competencies and knowledge repositories solidified internally to deliver professional, precise, and timely assessment and intervention services to real university students seeking help. This design fundamentally facilitates a systematic enhancement of service quality across three critical dimensions: enhanced professionalism, improved personalization, and strengthened continuity.
Regarding professionalism, the system constructs its core competency through a rigorous process of deep knowledge fusion and continuous supervised iteration. First, a comprehensive corpus of structured psychological theories, empirical research findings, typical consultation dialogues, and standardized operational protocols is systematically infused into the foundational large language model. This process endows the Assessor and Counselor agents with profound domain-specific understanding and robust long-term memory, enabling them to formulate structured service strategies based on dynamic evaluations of the client's subjective state. Central to this dimension is the Supervisor agent, which acts as the perpetual quality assurance hub. It monitors consultation sessions in real-time, identifies potential deviations from best practices, and provides corrective feedback. This feedback is then directly integrated into the agents' memory modules, creating a closed-loop “practice-evaluation-learning-optimization” cycle. This iterative process, underpinned by a combination of non-parametric prompt engineering and parametric fine-tuning techniques, drives the system's autonomous evolution and ensures reliable, steady growth in its professional expertise and behavioral precision.
In terms of personalization, the system achieves nuanced individual adaptation through dynamic empathy modeling and extensive scenario-based training. The Counselor agent employs advanced empathetic planning algorithms to deeply understand, contextualize, and flexibly respond to the client's diverse emotional expressions and evolving real-time needs, allowing for dynamic adjustment of intervention tactics. To enhance its resilience in complex, open-ended real-world dialogues, the internal training phase utilizes the large model to generate a diverse array of virtual student agents. These agents embody varied personality traits, specific psychological concerns (from academic stress to interpersonal conflicts), and different cultural backgrounds, creating a rich tapestry of simulated interaction scenarios. Within these simulations, the Supervisor analyzes multi-modal cues (textual and paralinguistic information) to track the “client's” state, while the Counselor practices deploying integrated tools such as psychological assessment scales and cognitive-behavioral techniques. This layered, continuous, and scenario-driven training regimen ensures the system's strategies remain highly adaptive and effective, not only for common issues but also within unpredictable conversational flows.
Concerning continuity, the system ensures service consistency and longitudinal insight through a dedicated long-term memory architecture. Every critical element of the counseling interaction—key disclosures, assessment scores, implemented intervention plans, and their observed outcomes—is systematically recorded and stored in a structured, interlinked manner within the agents' long-term memory units. This creates a comprehensive, continuously evolving digital psychological profile for each user. Consequently, the system supports truly continuous care across multiple sessions, with every new interaction informed by the complete historical context. This capability facilitates long-term monitoring of a student's psychological trajectory, enables ongoing evaluation of intervention efficacy, and permits data-driven adjustments to support plans. It thereby transforms psychological support from a series of isolated events into a coherent, evolving, and adaptive therapeutic process.
In summary, the multi-agent collaborative framework and its dual-cycle operational model proposed in this study represent a significant advancement. By organically integrating deep domain knowledge, a real-time supervision mechanism, and a mechanism for continuous adaptive learning, it provides a comprehensive, forward-looking, and operationally viable technical blueprint for utilizing artificial intelligence to profoundly transform university mental health services. This solution holds promise for significantly enhancing the accessibility, personalization, and long-term intervention efficacy of psychological support. Furthermore, it lays a solid technical and practical foundation for developing the next generation of intelligent, prevention-oriented mental health platforms dedicated to fostering student well-being.

Key words: intelligent agents, large language models, psychological counseling, college students