CA+
Group Research Project
Feb - Sep 2024
Advisor:
Prof. Jiangtao Gong
Keywords:
Conversational agent, Psychological Counseling, Large Language Models
Current psychological counseling systems struggle with long-term client engagement due to limited multi-turn contextual understanding and evolving support. This CA+ framework introduces a cognition-augmented AI counselor that enhances long-term client engagement by integrating cognitive models, professional expertise, and personalized multi-turn dialogues for improved mental health support.
A longitudinal study validated that CA+ significantly improves client satisfaction, empathy, and engagement over baseline systems, with licensed counselors confirming its professionalism and practical potential for real-world mental health care applications.
[Preprint]
Yuanrong Tang, Yu Kang, Yifan WANG, Tianhong Wang, Lixiu Wu, Chen Zhong, Jiangtao Gong (2025).
Preparing for submission to ACM Transactions on Computer-Human Interaction (TOCHI)
My Contributions​
-
Responsible for literature review, analyzing over 2,000 papers and selecting 300+ references to inform the agent’s framework design.
-
Designed the conversational agent framework based on CBT processes.
-
Built structured database to enhance the AI’s ability to deliver precise and professional insights.
-
Supported user study design and performed data analysis.
Key Features
The framework comprises three primary components: ​​
-
Cognition Augmented Reasoning: The core of our AI, enhancing its ability to understand and respond to complex human interactions through situational awareness and multi-turn dialogue.
-
Dynamically Aligned Expertise: Ensures responses follow professional counseling standards, integrating psychological knowledge and structured reasoning for effective client support.
-
Multi-Turn Dialogic Engagement: Maintains engagement by tracking emotional states and personalizing interactions, ensuring a responsive and satisfying user experience.
Design Goals
To achieve this Cognition Augmented AI counselor, five interconnected design goals are proposed with detailed solutions.
Design Goal 1: Emulating Counselor Cognitive Pattern in Digital Interventions
The AI is designed to follow a structured approach similar to human counselors, providing appropriate responses based on the dynamic counseling process. This is implemented through Hierarchical Planning, which organizes the counseling process across multiple levels of abstraction, from overarching strategies to specific interventions. This framework integrates two complementary perspectives:
​
-
Top-Down Goal Decomposition begins with broad stages and progressively narrows to specific actions, such as goals, plans, steps, and answers. As goals are further refined, the framework develops session agendas, identifies agenda-related actions, and proposes follow-up steps for the client.
​
-
Bottom-Up Feedback and Adaptation employs a continuous feedback loop to refine higher-level components. If a step or action proves ineffective or unsuitable, the agent reports this information upward, enabling modifications to the session agenda and overall strategy.​
​
.png)
Design Goal 2: Conceptualization-Driven Implicit Client Profiling
This design goal builds a dynamic client profile that evolves with each interaction. Starting with Case Conceptualization, the AI compiles key client information into a structured memory. This memory guides Information Gathering through active listening and strategic questioning, continually refining the profile. The enriched understanding then informs Intervention Planning, allowing the AI to deliver increasingly personalized support based on the client’s needs and context.
.png)
Design Goal 3: Book-style Data Generation and Retrieval
This goal equips the AI with a structured knowledge base for quick access to professional counseling resources. Data Generation organizes professional materials into retrievable formats by extracting key content, tagging it by therapeutic function, and indexing it in a structured database. Data Retrieval then matches this knowledge to the client’s needs by using context-specific labels and keywords, ensuring relevant guidance is available in real time. This approach enhances the AI’s ability to deliver precise, professional insights rather than relying on general responses.
.png)
Design Goal 4: Adaptive Empathy and Personalization
This goal enables the AI to adjust responses based on the client’s short-term emotional states and long-term preferences.
​
-
Short-Term Adaptive Empathy: (Top of the figure) The AI analyzes emotional cues to adjust its support level dynamically. When the client is in a highly negative emotional state or struggles to express themselves, the agent prioritizes immediate comfort. It employs techniques like empathetic reflection, validation, and gentle supportive statements to provide relief and create a safe space. As the client’s emotional state improves, the agent transitions to deeper exploration, addressing underlying causes and working towards long-term solutions.
​
-
Long-term Personalization: (Bottom of the figure) By analyzing emotional patterns and context, the AI adapts its communication style to align with the client’s preferences. It stores insights in preference records, allowing it to focus on relevant topics, proactively address triggers, and foster an engaging, personalized experience. This alignment strengthens the user-agent relationship over time.
.png)
Design Goal 5: Ecological Self
This design goal creates a relatable and trustworthy AI persona by blending the two components:
​
Generic Consultant Settings: Establish the agent’s professional foundation, including expertise, ethical adherence, clear consultant identity, and data-driven problem-solving.
​
Customized Character Settings: Add a personalized touch through unique language styles, character backgrounds, life experiences, and social networks. For example, the agent could take on the persona of Albert Ellis, a renowned psychologist, and incorporate self-disclosure by sharing experiences of overcoming social anxiety. This enables the agent to establish deeper relatability, foster empathy, and build trust with clients.
.png)
Evaluation
Study 1: Comparative Client Experience Study
In a three-day study with 24 participants, we compared our AI counseling system (CA+) to a baseline version without advanced features. Participants interacted with the systems daily and provided feedback through questionnaires and interviews. The results showed that CA+ significantly outperformed the baseline system in multiple areas, including empathy, personalization, and client engagement.
CA+ offered a superior user experience by tailoring responses based on client needs, remembering past interactions, and adapting advice over time. Participants felt more comfortable, supported, and willing to share personal insights with CA+. They praised CA+ for delivering empathetic and practical guidance, enhancing both emotional support and actionable advice. The study demonstrates CA+’s effectiveness in creating a more engaging and personalized counseling experience.
.png)
Study 2: Certified Counselor's Quality Assessment
In this study, two licensed counseling psychologists evaluated transcripts of interactions from Study 1 to assess the CA+ system’s counseling quality, cognitive abilities, and professionalism. The experts noted that CA+ displayed a significant improvement over the baseline system, engaging in a "reciprocal dialogue" that fostered deeper client connections. CA+ effectively maintained multi-turn memory, recalled past conversations, and provided structured session summaries with actionable tasks, enhancing continuity and client engagement.
The experts praised CA+ for demonstrating core counseling techniques like active listening, empathy, and positive reinforcement. It guided clients in focusing on positive resources and maintaining boundaries, which contributed to effective, client-centered conversations. Counselors highlighted CA+’s ability to encourage clients to take positive actions, improving self-expression and communication with partners.
Conclusion
This research presents CA+, a Cognition Augmented Counselor Agent framework designed to address key challenges in AI-based mental health counseling. By enhancing contextual understanding, professionalism, and personalization, CA+ boosts client engagement and satisfaction in AI-driven therapy. Empirical studies confirm its effectiveness in upholding counseling standards and fostering positive client outcomes. CA+ advances AI counseling capabilities and offers a foundation for future innovations to improve global access to quality mental health care, addressing the shortage of mental health professionals.
Self-Reflection
Limitations. The evaluation relied on a basic ChatGPT-4.0 model as the baseline, which, while highlighting CA+’s advancements in engagement, empathy, and professionalism, lacked the domain-specific tuning and contextual awareness often present in advanced conversational AI systems designed for mental health. Additionally, the system primarily focused on textual inputs, overlooking the potential of integrating multimodal signals such as tone of voice, facial expressions, and body language, which are critical for understanding and responding to client emotions.
​
Furthermore, while the study demonstrated short-term improvements in client engagement, it did not assess CA+’s long-term efficacy in sustaining mental health outcomes. The absence of a longitudinal component limits understanding of its ability to foster sustained behavioral changes, track mental health improvements, and address potential risks such as client over-reliance on the system.
​​
​
Future Work. To enhance client engagement and emotional connection, developing a humor-aware agent could be a valuable step. This would involve implementing humor generation techniques and training the model to balance humor with empathy, leveraging insights into therapeutic humor’s effectiveness. Future iterations could also incorporate multimodal interaction capabilities, including voice recognition, facial emotion detection, and haptic feedback, to enable deeper personalization and create a more immersive and human-like client experience.