Embodied Agent for Psychological Counseling
Research @ Institute for AI Industry Research, Tsinghua University
Supervised by Jiangtao Gong
Designing an LLM-Based Framework for Embodied Conversational Agents in Psychological Counseling
Link to the Article - https://arxiv.org/abs/2410.22041v1
Overview
I contributed to the development of Embodied Conversational Agents (ECAs), a novel framework designed to simulate authentic client behavior in psychological counseling scenarios. The framework integrates Cognitive Behavioral Therapy (CBT) principles and embodied cognition theories into a memory base, enabling AI agents to mimic real-world clients' thought processes, emotional responses, and social interactions with high fidelity.
Key Contributions
Framework Design and Theoretical Integration:
Defined Six Design Goals: Collaborated on formulating goals grounded in psychological counseling theories to guide the simulation framework's development. These included:
Comprehensive life-stage memory representation.
Simulation of cognitive processes following CBT principles.
Integration of perceptual memories to capture subjective emotions.
Modeling nuanced social interactions.
Ensuring consistency in data synthesis and memory retrieval.
Applied embodied cognition concepts to enhance authenticity in simulated client behaviors.
Memory Base Construction:
Designed mechanisms for generating core and intermediate beliefs based on real-world counseling data.
Developed a 4-phase memory simulation process, incorporating factual and perceptual memories aligned with CBT principles.
Structured client profiles into personal and social dimensions to simulate lifelike developmental trajectories and social interactions.
Dynamic Interaction System:
Implemented a context-driven memory retrieval mechanism to enable adaptive, realistic responses during simulated counseling sessions.
Enhanced dialogue generation by ensuring the semantic relevance of memories retrieved, supporting more nuanced and contextually accurate interactions.
Validation and Evaluation:
Conducted a pilot study in collaboration with licensed counselors to refine the framework through iterative feedback.
Benchmarked the ECAs framework against GPT-based and human-generated responses using metrics such as necessity, sufficiency, fidelity, and consistency.
Demonstrated significant improvements in simulation authenticity, emotional depth, and consistency.
Outcomes
Research Paper: Co-authored a publication detailing the ECAs framework, including methodology, design goals, and validation results. The work was featured in a peer-reviewed conference, showcasing its contributions to the field of psychological counseling simulations.
Dataset Creation: Generated a public dataset of detailed client profiles and embodied memories for further research in AI-driven mental health tools.
Impact: The ECAs framework outperformed baseline models, providing a scalable solution for counselor training and enhancing the realism of simulated therapeutic interactions.