DHCBT: a digital human psychological counseling system with a CBT multi-agent framework

Abstract

Academic stress is a major risk factor for mental health problems among college students. Existing online counseling systems often face three limitations: limited user trust in virtual interactions, insufficient adaptation to academic-stress scenarios, and inconsistent professional quality in single-agent counseling. To address these issues, we developed DHCBT, a digital human counseling system based on a hierarchical CBT multi-agent framework. Coordinated by a meta-agent and specialized expert agents, the framework structures counseling into two stages: establishment of an initial intervention framework and execution of the CBT core process. We evaluated the framework through benchmark testing on CBT-BENCH Level II, response-quality assessment on academic-stress counseling cases, and a controlled user study comparing DHCBT with a text-based chatbot. In benchmark testing, the framework showed strong performance on cognitive distortion and core-belief understanding tasks. In counseling-response evaluation, it outperformed baseline models in safety, adherence to CBT process, problem comprehension, and empathy. In the user study, DHCBT was associated with lower cognitive load and higher social presence, usability, and intention to continue use than the text-based control. These findings provide preliminary evidence that combining a digital human counseling interface with a CBT multi-agent framework is feasible for supporting college students with mild-to-moderate academic stress.